MEASURE THEORY

Volume 3
D.H.Fremlin
By the same author:
Topological Riesz Spaces and Measure Theory, Cambridge University Press, 1974.
Consequences of Martins Axiom, Cambridge University Press, 1982.
Companions to the present volume:
Measure Theory, vol. 1, Torres Fremlin, 2000.
Measure Theory, vol. 2, Torres Fremlin, 2001.
First printing May 2002
MEASURE THEORY
Volume 3
Measure Algebras
D.H.Fremlin
Reader in Mathematics, University of Essex
Dedicated by the Author
to the Publisher
This book may be ordered from the publisher at the address below. For price and means of pay-
ment see the authors Web page http://www.essex.ac.uk/maths/staff/fremlin/mtsales.htm,
or enquire from [email protected].
First published in 2002
by Torres Fremlin, 25 Ireton Road, Colchester CO3 3AT, England
c _ D.H.Fremlin 2002
The right of D.H.Fremlin to be identied as author of this work has been asserted in accordance with
the Copyright, Designs and Patents Act 1988. This work is issued under the terms of the Design Science
License as published in http://dsl.org/copyleft/dsl.txt. For the source les see http://www.essex.
ac.uk/maths/staff/fremlin/mt3.2002/readme.txt.
Library of Congress classication QA312.F72
AMS 2000 classication 28A99
ISBN 0-9538129-3-6
Typeset by /
/
o-T
E
X
Printed in England by Biddles Short Run Books, Kings Lynn
5
Contents
General Introduction 10
Introduction to Volume 3 11
Chapter 31: Boolean algebras
Introduction 13
311 Boolean algebras 13
Boolean rings and algebras; ideals and ring homomorphisms to Z
2
; Stones theorem; the operations
, , ., \ and the relation ; topology of the Stone space; Boolean algebras as complemented
distributive lattices.
312 Homomorphisms 21
Subalgebras; ideals; Boolean homomorphisms; the ordering determines the ring structure; quotient al-
gebras; extension of homomorphisms; homomorphisms and Stone spaces.
313 Order-continuity 30
General distributive laws; order-closed sets; order-closures; Monotone Class Theorem; order-preserving
functions; order-continuity; order-dense sets; order-continuous Boolean homomorphisms; regularly em-
bedded subalgebras.
314 Order-completeness 40
Dedekind completeness and -completeness; quotients, subalgebras, principal ideals; order-continuous
homomorphisms; extension of homomorphisms; Loomis-Sikorski representation of a -complete algebra
as a quotient of a -algebra of sets; regular open algebras; Stone spaces; Dedekind completion of a
Boolean algebra.
315 Products and free products 51
Simple product of Boolean algebras; free product of Boolean algebras; algebras of sets and their quotients.
316 Further topics 59
The countable chain condition; weak (, )-distributivity; Stone spaces; atomic and atomless Boolean
algebras.
Chapter 32: Measure algebras
Introduction 68
321 Measure algebras 68
Measure algebras; elementary properties; the measure algebra of a measure space; Stone spaces.
322 Taxonomy of measure algebras 72
Totally nite, -nite, semi-nite and localizable measure algebras; relation to corresponding types of
measure space; completions and c.l.d. versions of measures; semi-nite measure algebras are weakly
(, )-distributive; subspace measures; simple products of measure algebras; Stone spaces of localizable
measure algebras; localizations of semi-nite measure algebras.
323 The topology of a measure algebra 82
Dening a topology and uniformity on a measure algebra; continuity of algebraic operations; order-closed
sets; Hausdor and metrizable topologies, complete uniformities; closed subalgebras.
324 Homomorphisms 88
Homomorphisms induced by measurable functions; order-continuous and continuous homomorphisms;
the topology of a semi-nite measure algebra is determined by the algebraic structure; measure-preserving
homomorphisms.
325 Free products and product measures 95
The measure algebra of a product measure; the localizable measure algebra free product of two semi-nite
measure algebras; the measure algebra of a product of probability measures; the probability algebra free
product of probability algebras; factorizing through subproducts.
326 Additive functionals on Boolean algebras 104
Additive, countably additive and completely additive functionals; Jordan decomposition; Hahn decom-
position.
327 Additive functionals on measure algebras 116
Absolutely continuous and continuous additive functionals; Radon-Nikod ym theorem; the standard ex-
tension of a continuous additive functional on a closed subalgebra.
6
Chapter 33: Maharams theorem
Introduction 123
331 Classication of homogeneous measure algebras 123
Relatively atomless algebras; one-step extension of measure-preserving homomorphisms; Maharam type
of a measure algebra; Maharam-type-homogeneous probability algebras of the same Maharam type are
isomorphic; the measure algebra of 0, 1

is homogeneous.
332 Classication of localizable measure algebras 132
Any localizable measure algebra is isomorphic to a simple product of homogeneous totally nite algebras;
exact description of isomorphism types; closed subalgebras.
333 Closed subalgebras 142
Relative Maharam types; extension of measure-preserving Boolean homomorphisms; complete classi-
cation of closed subalgebras of probability algebras as triples (A, , C).
334 Products 158
Maharam types of product measures.
Chapter 34: Liftings
Introduction 162
341 The lifting theorem 162
Liftings and lower densities; strictly localizable spaces have lower densities; construction of a lifting from
a density; complete strictly localizable spaces have liftings; liftings and Stone spaces.
342 Compact measure spaces 174
Inner regular measures; compact classes; compact and locally compact measures; perfect measures.
343 Realization of homomorphisms 182
Representing homomorphisms between measure algebras by functions; possible when target measure
space is locally compact; countably separated measures and uniqueness of representing functions; the
split interval; perfect measures.
344 Realization of automorphisms 190
Simultaneously representing countable groups of automorphisms of measure algebras by functions
countably separated measure spaces, measures on 0, 1
I
; characterization of Lebesgue measure as a
measure space; strong homogeneity of usual measure on 0, 1
I
.
345 Translation-invariant liftings 200
Translation-invariant liftings on R
r
and 0, 1
I
; there is no t.-i. Borel lifting on R.
346 Consistent liftings 208
Liftings of product measures which respect the product structure; translation-invariant liftings on 0, 1
I
;
products of Maharam-type-homogeneous probability spaces; lower densities respecting product struc-
tures; consistent liftings; the Stone space of Lebesgue measure.
Chapter 35: Riesz spaces
Introduction 219
351 Partially ordered linear spaces 219
Partially ordered linear spaces; positive cones; suprema and inma; positive linear operators; order-
continuous linear operators; Riesz homomorphisms; quotient spaces; reduced powers; representation of
p.o.l.ss as subspaces of reduced powers of R; Archimedean spaces.
352 Riesz spaces 226
Riesz spaces; identities; general distributive laws; Riesz homomorphisms; Riesz subspaces; order-dense
subspaces and order-continuous operators; bands; the algebra of complemented bands; the algebra of
projection bands; principal bands; f-algebras.
353 Archimedean and Dedekind complete Riesz spaces 237
Order-dense subspaces; bands; Dedekind ()-complete spaces; order-closed Riesz subspaces; order units;
f-algebras.
354 Banach lattices 246
Riesz norms; Fatou norms; the Levi property; order-continuous norms; order-unit norms; M-spaces; are
isomorphic to C(X) for compact Hausdor X; L-spaces; uniform integrability in L-spaces.
355 Spaces of linear operators 255
Order-bounded linear operators; the space L

(U; V ); order-continuous operators; extension of order-

continuous operators; the space L

(U; V ); order-continuous norms.

7
356 Dual spaces 264
The spaces U

, U

, U

; biduals, embeddings UV

where V U

; perfect Riesz spaces; L- and

M-spaces; uniformly integrable sets in the dual of an M-space; relative weak compactness in L-spaces.
Chapter 36: Function spaces
Introduction 275
361 S 276
Additive functions on Boolean rings; the space S(A); universal mapping theorems for linear operators
on S; the map T

: S(A)S(B) induced by a ring homomorphism : A B; projection bands in S(A);

identifying S(A) when A is a quotient of an algebra of sets.
362 S

287
Bounded additive functionals on A identied with order-bounded linear functionals on S(A); the L-
space S

and its bands; countably additive, completely additive, absolutely continuous and continuous
functionals; uniform integrability in S

.
363 L

298
The space L

(A), as an M-space and f-algebra; universal mapping theorems for linear operators on L

;
T

: L

(A)L

(B); representing L

when A is a quotient of an algebra of sets; integrals with respect

to nitely additive functionals; projection bands in L

; (L

and its bands; Dedekind completeness of

A and L

; representing -complete M-spaces; the generalized Hahn-Banach theorem; the Banach-Ulam

problem.
364 L
0
314
The space L
0
(A); representing L
0
when A is a quotient of a -algebra of sets; algebraic operations on
L
0
; action of Borel measurable functions on L
0
; identifying L
0
(A) with L
0
() when A is a measure
algebra; embedding S and L

in L
0
; suprema and inma in L
0
; Dedekind completeness in A and L
0
;
multiplication in L
0
; projection bands; T

: L
0
(A)L
0
(B); simple products; *regular open algebras;
*the space C

(X).
365 L
1
335
The space L
1
(A, ); identication with L
1
();
_
a
u; the Radon-Nikod ym theorem again;
_
wT

ud =
_
ud ; the duality between L
1
and L

; additive functions on A and linear operators on L

1
; T

:
L
1
(A, )L
1
(B, ) and P

: L
1
(B, )L
1
(A, ); conditional expectations; bands in L
1
; varying .
366 L
p
351
The spaces L
p
(A, ); identication with L
p
(); L
q
as the dual of L
p
; the spaces M
0
and M
1,0
; T

:
M
0
(A, )M
0
(B, ) and P

: M
1,0
(B, )M
1,0
(A, ); conditional expectations; the case p = 2.
367 Convergence in measure 360
Order*-convergence of sequences in lattices; in Riesz spaces; in Banach lattices; in quotients of spaces
of measurable functions; in C(X); Lebesgues Dominated Convergence Theorem and Doobs Martingale
Theorem; convergence in measure in L
0
(A); and pointwise convergence; dened by the Riesz space struc-
ture; positive linear operators on L
0
; convergence in measure and the canonical projection (L
1
)

L
1
.
368 Embedding Riesz spaces in L
0
375
Extension of order-continuous Riesz homomorphisms into L
0
; representation of Archimedean Riesz
spaces as subspaces of L
0
; Dedekind completion of Riesz spaces; characterizing L
0
spaces as Riesz
spaces; weakly (, )-distributive Riesz spaces.
369 Banach function spaces 386
Riesz spaces separated by their order-continuous duals; representing U

when UL
0
; Kakutanis repre-
sentation of L-spaces as L
1
spaces; extended Fatou norms; associate norms; L

= (L

; Fatou norms
and convergence in measure; M
,1
and M
1,
, | |
,1
and | |
1,
; L

1
+L

2
.
Chapter 37: Linear operators between function spaces
Introduction 402
371 The Chacon-Krengel theorem 402
L

(U; V ) = L

(U; V ) = B(U; V ) for L-spaces U and V ; the class T

(0)
,
of | |
1
-decreasing, | |

-decreasing
linear operators from M
1,0
(A, ) to M
1,0
(B, ).
372 The ergodic theorem 407
The Maximal Ergodic Theorem and the Ergodic Theorem for operators in T
(0)
,
; for inverse-measure-
preserving functions : XX; limit operators as conditional expectations; applications to continued
fractions; mixing and ergodic transformations.
8
373 Decreasing rearrangements 422
The classes T , T

; the space M
0,
; decreasing rearrangements u

; |u

|
p
= |u|
p
;
_
[Tuv[
_
u

if T T ; the very weak operator topology and compactness of T ; v is expressible as Tu, where T T ,
i
_
t
0
v

_
t
0
u

for every t; nding T such that

_
Tuv =
_
u

; the adjoint operator from T

(0)
,
to
T
(0)
,
.
374 Rearrangement-invariant spaces 440
T -invariant subspaces of M
1,
, and T -invariant extended Fatou norms; relating T -invariant norms
on dierent spaces; rearrangement-invariant sets and norms; when rearrangement-invariance implies
T -invariance.
375 Kwapiens theorem 451
Linear operators on L
0
spaces; if B is measurable, a positive linear operator from L
0
(A) to L
0
(B) can
be assembled from Riesz homomorphisms.
376 Integral operators 457
Kernel operators; free products of measure algebras and tensor products of L
0
spaces; tensor products
of L
1
spaces; abstract integral operators (i) as a band in L

(U; V ) (ii) represented by kernels belonging

to L
0
(A

B) (iii) as operators converting weakly convergent sequences into order*-convergent sequences;

operators into M-spaces or out of L-spaces; disintegrations.
Chapter 38: Automorphism groups
Introduction 477
381 Automorphism groups of Boolean algebras 477
Assembling an automorphism; elements supporting an automorphism; cyclic automorphisms; exchanging
involutions; the support of an automorphism; full subgroups of Aut A; expressing an automorphism as a
product of involutions; subgroups of Aut A with many involutions; normal subgroups of full groups with
many involutions; simple groups.
382 Automorphism groups of measure algebras 488
Measure-preserving automorphisms as products of involutions; normal subgroups of Aut A and Aut

A.
383 Outer automorphisms 492
If G Aut A, H Aut B have many involutions, any isomorphism between G to H arises from an
isomorphism between A and B; if A is nowhere rigid, Aut A has no outer automorphisms; applications
to localizable measure algebras.
384 Entropy 502
Entropy of a partition of unity in a probability algebra; conditional entropy; entropy of a measure-
preserving homomorphism; calculation of entropy (Kolmogorov-Sina theorem); Bernoulli shifts; isomor-
phic homomorphisms and conjugacy classes in Aut

A; almost isomorphic inverse-measure-preserving
functions.
385 More about entropy 515
Periodic and aperiodic parts of an endomorphism; the Halmos-Rokhlin-Kakutani lemma; the Shannon-
McMillan-Breiman theorem; various lemmas.
386 Ornsteins theorem 528
Bernoulli partitions; nding Bernoulli partitions with elements of given measure (Sinas theorem); ad-
justing Bernoulli partitions; Ornsteins theorem (Bernoulli shifts of the same nite entropy are isomor-
phic); Ornsteins and Sinas theorems in the case of innite entropy.
387 Dyes theorem 551
Full subgroups of Aut A; and orbits of inverse-measure-preserving functions; induced automorphisms of
principal ideals; von Neumann transformations; von Neumann transformations generating a given full
subgroup; classication of full subgroups generated by a single automorphism.
Chapter 39: Measurable algebras
Introduction 567
391 Kelleys theorem 567
Measurable algebras; strictly positive additive functionals and weak (, )-distributivity; additive func-
tionals subordinate to or dominating a given functional; intersection numbers; existence of strictly
positive additive functionals; -linked Boolean algebras; Gaifmans example.
392 Submeasures 575
Submeasures; exhaustive, uniformly exhaustive and Maharam submeasures; the Kalton-Roberts theorem
(a strictly positive uniformly exhaustive submeasure provides a strictly positive additive functional).
9
393 The Control Measure Problem 579
Forms of the Control Measure Problem; strictly positive Maharam submeasures, exhaustive submeasures,
exhaustive submeasures on the clopen algebra of 0, 1
N
; absolute continuity of submeasures; countable
atomless Boolean algebras; topologies on Boolean algebras; topologies on L
0
spaces; vector measures;
examples of submeasures.
394 Kawadas theorem 595
Full local semigroups; -equidecomposability; fully non-paradoxical subgroups of Aut A; ]b : a| and
b : a|; invariant additive functions from A to L

(C), where C is the xed-point algebra of a group;

invariant additive functionals and measures; ergodic fully non-paradoxical groups.
395 The Hajian-Ito theorem 610
Invariant measures on measurable algebras; weakly wandering elements.
Appendix to Volume 3
Introduction 613
3A1 Set theory 613
Calculation of cardinalities; conal sets, conalities; notes on the use of Zorns Lemma; the natural
numbers as nite ordinals; lattice homomorphisms; the Marriage Lemma.
3A2 Rings 615
Rings; subrings, ideals, homomorphisms, quotient rings, the First Isomorphism Theorem; products.
3A3 General topology 618
Hausdor, regular, completely regular, zero-dimensional, extremally disconnected, compact and locally
compact spaces; continuous functions; dense subsets; meager sets; Baires theorem for locally compact
spaces; products; Tychonos theorem; the usual topologies on 0, 1
I
, R
I
; cluster points of lters;
topology bases; uniform convergence of sequences of functions; one-point compactications.
3A4 Uniformities 622
Uniform spaces; and pseudometrics; uniform continuity; subspaces; product uniformities; Cauchy lters
and completeness; extending uniformly continuous functions; completions.
3A5 Normed spaces 625
The Hahn-Banach theorem in analytic and geometric forms; cones and convex sets; weak and weak*
topologies; reexive spaces; Uniform Boundedness Theorem; completions; normed algebras; compact
linear operators; Hilbert spaces.
3A6 Groups 628
Involutions; inner and outer automorphisms; normal subgroups.
References for Volume 3 629
Index to Volumes 1-3
Principal topics and results 633
General index 641
10
General introduction In this treatise I aim to give a comprehensive description of modern abstract measure
theory, with some indication of its principal applications. The rst two volumes are set at an introductory
level; they are intended for students with a solid grounding in the concepts of real analysis, but possibly with
rather limited detailed knowledge. As the book proceeds, the level of sophistication and expertise demanded
will increase; thus for the volume on topological measure spaces, familiarity with general topology will be
assumed. The emphasis throughout is on the mathematical ideas involved, which in this subject are mostly
to be found in the details of the proofs.
My intention is that the book should be usable both as a rst introduction to the subject and as a reference
work. For the sake of the rst aim, I try to limit the ideas of the early volumes to those which are really
essential to the development of the basic theorems. For the sake of the second aim, I try to express these ideas
in their full natural generality, and in particular I take care to avoid suggesting any unnecessary restrictions
in their applicability. Of course these principles are to to some extent contradictory. Nevertheless, I nd that
most of the time they are very nearly reconcilable, provided that I indulge in a certain degree of repetition.
For instance, right at the beginning, the puzzle arises: should one develop Lebesgue measure rst on the
real line, and then in spaces of higher dimension, or should one go straight to the multidimensional case? I
believe that there is no single correct answer to this question. Most students will nd the one-dimensional
case easier, and it therefore seems more appropriate for a rst introduction, since even in that case the
technical problems can be daunting. But certainly every student of measure theory must at a fairly early
stage come to terms with Lebesgue area and volume as well as length; and with the correct formulations, the
multidimensional case diers from the one-dimensional case only in a denition and a (substantial) lemma.
So what I have done is to write them both out (114-115). In the same spirit, I have been uninhibited,
when setting out exercises, by the fact that many of the results I invite students to look for will appear in
later chapters; I believe that throughout mathematics one has a better chance of understanding a theorem
if one has previously attempted something similar alone.
As I write this Introduction (December 2001), the plan of the work is as follows:
Volume 1: The Irreducible Minimum
Volume 2: Broad Foundations
Volume 3: Measure Algebras
Volume 4: Topological Measure Spaces
Volume 5: Set-theoretic Measure Theory.
Volume 1 is intended for those with no prior knowledge of measure theory, but competent in the elementary
techniques of real analysis. I hope that it will be found useful by undergraduates meeting Lebesgue measure
for the rst time. Volume 2 aims to lay out some of the fundamental results of pure measure theory
(the Radon-Nikod ym theorem, Fubinis theorem), but also gives short introductions to some of the most
important applications of measure theory (probability theory, Fourier analysis). While I should like to
believe that most of it is written at a level accessible to anyone who has mastered the contents of Volume 1,
I should not myself have the courage to try to cover it in an undergraduate course, though I would certainly
attempt to include some parts of it. Volumes 3 and 4 are set at a rather higher level, suitable to postgraduate
courses; while Volume 5 will assume a wide-ranging competence over large parts of analysis and set theory.
There is a disclaimer which I ought to make in a place where you might see it in time to avoid paying for
this book. I make no attempt to describe the history of the subject. This is not because I think the history
uninteresting or unimportant; rather, it is because I have no condence of saying anything which would not
be seriously misleading. Indeed I have very little condence in anything I have ever read concerning the
history of ideas. So while I am happy to honour the names of Lebesgue and Kolmogorov and Maharam in
more or less appropriate places, and I try to include in the bibliographies the works which I have myself
consulted, I leave any consideration of the details to those bolder and better qualied than myself.
The work as a whole is not yet complete; and when it is nished, it will undoubtedly be too long
to be printed as a single volume in any reasonable format. I am therefore publishing it one part at a
time. However, drafts of most of the rest are available on the Internet; see http://www.essex.ac.uk/
maths/staff/fremlin/mt.htm for detailed instructions. For the time being, at least, printing will be in
short runs. I hope that readers will be energetic in commenting on errors and omissions, since it should be
possible to correct these relatively promptly. An inevitable consequence of this is that paragraph references
may go out of date rather quickly. I shall be most attered if anyone chooses to rely on this book as a source
Introduction to Volume 3 11
for basic material; and I am willing to attempt to maintain a concordance to such references, indicating
where migratory results have come to rest for the moment, if authors will supply me with copies of papers
which use them.
I mention some minor points concerning the layout of the material. Most sections conclude with lists of
basic exercises and further exercises, which I hope will be generally instructive and occasionally enter-
taining. How many of these you should attempt must be for you and your teacher, if any, to decide, as no
two students will have quite the same needs. I mark with a >>> those which seem to me to be particularly
important. But while you may not need to write out solutions to all the basic exercises, if you are in any
doubt as to your capacity to do so you should take this as a warning to slow down a bit. The further
exercises are unbounded in diculty, and are unied only by a presumption that each has at least one
solution based on ideas already introduced. Occasionally I add a nal problem, a question to which I do
not know the answer and which seems to arise naturally in the course of the work.
The impulse to write this book is in large part a desire to present a unied account of the subject.
Cross-references are correspondingly abundant and wide-ranging. In order to be able to refer freely across
the whole text, I have chosen a reference system which gives the same code name to a paragraph wherever
it is being called from. Thus 132E is the fth paragraph in the second section of the third chapter of
Volume 1, and is referred to by that name throughout. Let me emphasize that cross-references are supposed
to help the reader, not distract her. Do not take the interpolation (121A) as an instruction, or even a
recommendation, to lift Volume 1 o the shelf and hunt for 121. If you are happy with an argument as it
stands, independently of the reference, then carry on. If, however, I seem to have made rather a large jump,
or the notation has suddenly become opaque, local cross-references may help you to ll in the gaps.
Each volume will have an appendix of useful facts, in which I set out material which is called on
somewhere in that volume, and which I do not feel I can take for granted. Typically the arrangement of
material in these appendices is directed very narrowly at the particular applications I have in mind, and is
unlikely to be a satisfactory substitute for conventional treatments of the topics touched on. Moreover, the
ideas may well be needed only on rare and isolated occasions. So as a rule I recommend you to ignore the
appendices until you have some direct reason to suppose that a fragment may be useful to you.
During the extended gestation of this project I have been helped by many people, and I hope that my
friends and colleagues will be pleased when they recognise their ideas scattered through the pages below.
But I am especially grateful to those who have taken the trouble to read through earlier drafts and comment
on obscurities and errors.
Introduction to Volume 3
One of the rst things one learns, as a student of measure theory, is that sets of measure zero are
frequently negligible in the straightforward sense that they can safely be ignored. This is not quite a
universal principle, and one of my purposes in writing this treatise is to call attention to the exceptional
cases in which negligible sets are important. But very large parts of the theory, including some of the topics
already treated in Volume 2, can be expressed in an appropriately abstract language in which negligible sets
have been factored out. This is what the present volume is about. A measure algebra is a quotient of an
algebra of measurable sets by an ideal of negligible sets; that is, the elements of the measure algebra are
equivalence classes of measurable sets. At the cost of an extra layer of abstraction, we obtain a language
which can give concise and elegant expression to a substantial proportion of the ideas of measure theory,
and which oers insights almost everywhere in the subject.
It is here that I embark wholeheartedly on pure measure theory. I think it is fair to say that the
applications of measure theory to other branches of mathematics are more often through measure spaces
rather than measure algebras. Certainly there will be in this volume many theorems of wide importance
outside measure theory; but typically their usefulness will be in forms translated back into the language of
the rst two volumes. But it is also fair to say that the language of measure algebras is the only reasonable
way to discuss large parts of a subject which, as pure mathematics, can bear comparison with any.
In the structure of this volume I can distinguish seven working and two accessory chapters. The
accessory chapters are 31 and 35. In these I develop the theories of Boolean algebras and Riesz spaces (=
vector lattices) which are needed later. As in Volume 2 you have a certain amount of choice in the order in
12 Introduction to Volume 3
which you take the material. Everything except Chapter 35 depends on Chapter 31, and everything except
Chapters 31 and 35 depends on Chapter 32. Chapters 33, 34 and 36 can be taken in any order, but Chapter
36 relies on Chapter 35. (I do not mean that Chapter 33 is never referred to in Chapter 34, nor even that
no results from Chapter 33 are relied on in the later chapters. What I mean is that their most important
ideas are accessible without learning the material of Chapter 33 properly.) Chapter 37 depends on Chapters
35 and 36. Chapter 38 would be dicult to make sense of without some notion of what has been done in
Chapter 33. Chapter 39 uses fragments of Chapters 35 and 36.
The rst half of the volume follows almost the only line permitted by the structure of the subject. If we
are going to study measure algebras at all, we must know the relevant facts about Boolean algebras (Chapter
31) and how to translate what we know about measure spaces into the new language (Chapter 32). Then
we must get a proper grip on the two most important theorems: Maharams theorem on the classication of
measure algebras (Chapter 33) and the von Neumann-Maharam lifting theorem (Chapter 34). Since I am
now writing for readers who are committed I hope, happily committed to learning as much as they can
about the subject, I take the space to push these ideas as far as they can easily go, giving a full classication
of closed subalgebras of probability algebras, for instance (333), and investigating special types of lifting
(345-346). I mention here three sections interpolated into Chapter 34 (342-344) which attack a subtle
and important question: when can we expect homomorphisms between measure algebras to be realizable in
terms of transformations between measure spaces, as discussed briey in 235 and elsewhere.
Chapters 36 and 37 are devoted to re-working the ideas of Chapter 24 on function spaces in the more
abstract context now available, and relating them to the general Riesz spaces of Chapter 35. I am concerned
here not to develop new structures, nor even to prove striking new theorems, but rather to oer new ways
of looking at the old ones. Only in the Ergodic Theorem (372) do I come to a really important new result.
Chapter 38 looks at two questions, both obvious ones to ask if you have been trained in twentieth-century
pure mathematics: what does the automorphism group of a measure algebra look like, and inside such an
automorphism group, what do the conjugacy classes look like? (The second question is a fancy way of asking
how to decide, given two automorphisms of one of the structures considered in this volume, whether they
are really dierent, or just copies of each other obtained by looking at the structure a dierent way up.)
Finally, in Chapter 39, I discuss what is known about the question of which Boolean algebras can appear
as measure algebras.
Concerning the prerequisites for this volume, we certainly do not need everything in Volume 2. The
important chapters there are 21, 23, 24, 25 and 27. If you are approaching this volume without having
read the earlier parts of this treatise, you will need the Radon-Nikod ym theorem and product measures
(of arbitrary families of probability spaces), for Maharams theorem; a simple version of the martingale
theorem, for the lifting theorem; and an acquaintance with L
p
spaces (particularly, with L
0
spaces) for
Chapter 36. But I would recommend the results-only versions of Volumes 1 and 2 in case some reference
is totally obscure. Outside measure theory, I call on quite a lot of terms from general topology, but none
of the ideas needed are dicult (Baires and Tychonos theorems are the deepest); they are sketched in
3A3 and 3A4. We do need some functional analysis for Chapters 36 and 39, but very little more than was
already used in Volume 2, except that I now call on versions of the Hahn-Banach theorem (3A5).
In this volume I assume that readers have substantial experience in both real and abstract analysis, and
I make few concessions which would not be appropriate when addressing active researchers, except that
perhaps I am a little gentler when calling on ideas from set theory and general topology than I should be
with my own colleagues, and I continue to include all the easiest exercises I can think of. I do maintain
my practice of giving proofs in very full detail, not so much because I am trying to make them easier, but
because one of my purposes here is to provide a complete account of the ideas of the subject. I hope that
the result will be accessible to most doctoral students who are studying topics in, or depending on, measure
theory.
311B Boolean algebras 13
Chapter 31
Boolean algebras
The theory of measure algebras naturally depends on certain parts of the general theory of Boolean
algebras. In this chapter I collect those results which will be useful later. Since many students encounter
the formal notion of Boolean algebra for the rst time in this context, I start at the beginning; and indeed I
include in the Appendix (3A2) a brief account of the necessary part of the theory of rings, as not everyone
will have had time for this bit of abstract algebra in an undergraduate course. But unless you nd the
algebraic theory of Boolean algebras so interesting that you wish to study it for its own sake in which case
you should perhaps turn to Sikorski 64 or Koppelberg 89 I do not think it would be very sensible to
read the whole of this chapter before proceeding to the main work of the volume in Chapter 32. Probably
311 is necessary to get an idea of what a Boolean algebra looks like, and a glance at the statements of
the theorems in 312 would be useful, but the later sections can wait until you have need of them, on the
understanding that apparently innocent formal manipulations may depend on concepts which take some
time to master. I hope that the cross-references will be suciently well-targeted to make it possible to read
this material in parallel with its applications.
311 Boolean algebras
In this section I try to give a sucient notion of the character of abstract Boolean algebras to make the
calculations which will appear on almost every page of this volume seem both elementary and natural. The
principal result is of course M.H.Stones theorem: every Boolean algebra can be expressed as an algebra of
sets (311E). So the section divides naturally into the rst part, proving Stones theorem, and the second,
consisting of elementary consequences of the theorem and a little practice in using the insights it oers.
311A Denitions (a) A Boolean ring is a ring (A, +, .) in which a
2
= a for every a A.
(b) A Boolean algebra is a Boolean ring A with a multiplicative identity 1 = 1
A
; I allow 1 = 0 in this
context.
Remark For notes on those parts of the elementary theory of rings which we shall need, see 3A2.
I hope that the rather arbitrary use of the word algebra here will give no diculties; it gives me the
freedom to insist that the ring 0 should be accepted as a Boolean algebra.
311B Examples (a) For any set X, (TX, , ) is a Boolean algebra; its zero is and its multiplicative
identity is X. PPP We have to check the following, which are all easily established, using Venn diagrams or
otherwise:
AB X for all A, B X,
(AB)C = A(BC) for all A, B, C X,
so that (TX, ) is a semigroup;
A = A = A for every A X,
so that is the identity in (TX, );
AA = for every A X,
so that every element of TX is its own inverse in (TX, ), and (TX, ) is a group;
AB = BA for all A, B X,
so that (TX, ) is an abelian group;
A B X for all A, B X,
(A B) C = A (B C) for all A, B, C X,
so that (TX, ) is a semigroup;
A (BC) = (A B)(A C), (AB) C = (A C)(B C) for all A, B, C X,
so that (TX, , ) is a ring;
14 Boolean algebras 311B
A A = A for every A X,
so that (TX, , ) is a Boolean ring;
A X = X A = A for every A X,
so that (TX, , ) is a Boolean algebra and X is its identity. QQQ
(b) Recall that an algebra of subsets of X (136E) is a family TX such that , X E for
every E , and E F for all E, F . In this case (, , ) is a Boolean algebra with zero and
identity X. PPP If E, F , then
E F = X ((X E) (X F)) ,
EF = (E (X F)) (F (X E)) .
Because and X = X both belong to , we can work through the identities in (a) above to see that ,
like TX, is a Boolean algebra. QQQ
(c) Consider the ring Z
2
= 0, 1, with its ring operations +
2
, given by setting
0 +
2
0 = 1 +
2
1 = 0, 0 +
2
1 = 1 +
2
0 = 1,
0 0 = 0 1 = 1 0 = 0, 1 1 = 1.
I leave it to you to check, if you have not seen it before, that this is a ring. Because 0 0 = 0 and 1 1 = 1,
it is a Boolean algebra.
311C Proposition Let A be a Boolean ring.
(a) a +a = 0, that is, a = a, for every a A.
(b) ab = ba for all a, b A.
proof (a) If a A, then
a +a = (a +a)(a +a) = a
2
+a
2
+a
2
+a
2
= a +a +a +a,
so we must have 0 = a +a.
(b) Now for any a, b A,
a +b = (a +b)(a +b) = a
2
+ab +ba +b
2
= a +ab +ba +b,
so
0 = ab +ba = ab +ab
and ab = ba.
311D Lemma Let A be a Boolean ring, I an ideal of A (3A2E), and a A I. Then there is a ring
homomorphism : A Z
2
such that a = 1 and d = 0 for every d I.
proof (a) Let 1 be the family of those ideals J of A which include I and do not contain a. Then 1 has a
maximal element K say. PPP Apply Zorns lemma. Since I 1, 1 ,= . If is a non-empty totally ordered
subset of 1, then set J

=

. If b, c J

and d A, then there are J

1
, J
2
such that b J
1
and
c J
2
; now J = J
1
J
2
is equal to one of J
1
, J
2
, so belongs to , and 0, b + c, bd all belong to J, so all
belong to J

. Thus J

A; of course I J

and a / J

, so J

1 and is an upper bound for in 1. As

is arbitrary, the hypotheses of Zorns lemma are satised and 1 has a maximal element. QQQ
(b) For b A set K
b
= d : d A, bd K. The following are easy to check:
(i) K K
b
for every b A, because K is an ideal.
(ii) K
b
A for every b A. PPP 0 K K
b
. If d, d
t
K
b
and c A then
b(d +d
t
) = bd +bd
t
, b(dc) = (bd)c
belong to K, so d +d
t
, dc K
b
. QQQ
(iii) If b A and a / K
b
, then K
b
1 so K
b
= K.
311Fc Boolean algebras 15
(iv) Now a
2
= a / K, so a / K
a
and K
a
= K.
(v) If b A K then b / K
a
, that is, ba = ab / K, and a / K
b
; consequently K
b
= K.
(vi) If b, c A K then c / K
b
so bc / K.
(vii) If b, c A K then
bc(b +c) = b
2
c +bc
2
= bc +bc = 0 K,
so b +c K
bc
. By (vi) and (v), K
bc
= K so b +c K.
(c) Now dene : A Z
2
by setting d = 0 if d K, d = 1 if d A K. Then is a ring
homomorphism. PPP
(i) If b, c K then b +c, bc K so
(b +c) = 0 = b +
2
c, (bc) = 0 = b c.
(ii) If b K, c A K then
c = (b +b) +c = b + (b +c) / K
so b +c / K, while bc K, so
(b +c) = 1 = b +
2
c, (bc) = 0 = b c.
(iii) Similarly,
(b +c) = 1 = b +
2
c, (bc) = 0 = b c
if b A K and c K.
(iv) If b, c A K, then by (b-vi) and (b-vii) we have b +c K, bc / K so
(b +c) = 0 = b +
2
c, (bc) = 1 = b c.
Thus is a ring homomorphism. QQQ
(d) Finally, if d I then d K so d = 0; and a = 1 because a / K.
311E M.H.Stones Theorem: rst form Let A be any Boolean ring, and let Z be the set of ring
homomorphisms from A onto Z
2
. Then we have an injective ring homomorphism a a : A TZ, setting
a = z : z Z, z(a) = 1. If A is a Boolean algebra, then

1
A
= Z.
proof (a) If a, b A, then

a+b = z : z(a+b) = 1 = z : z(a) +

2
z(b) = 1 = z : z(a), z(b) = 0, 1 = a

b,

ab = z : z(ab) = 1 = z : z(a)z(b) = 1 = z : z(a) = z(b) = 1 = a

b.
Thus a a is a ring homomorphism.
(b) If a A and a ,= 0, then by 311D, with I = 0, there is a z Z such that z(a) = 1, that is, z a;
so that a ,= . This shows that the kernel of a a is 0, so that the homomorphism is injective (3A2Db).
(c) If A is a Boolean algebra, and z Z, then there is some a A such that z(a) = 1, so that
z(1
A
)z(a) = z(1
A
a) ,= 0 and z(1
A
) ,= 0; thus

1
A
= Z.
311F Remarks (a) For any Boolean ring A, I will say that the Stone space of A is the set Z of non-zero
ring homomorphisms from A to Z
2
, and the canonical map a a : A TZ is the Stone representation.
(b) Because the map a a : A TZ is an injective ring homomorphism, A is isomorphic, as Boolean
ring, to its image c = a : a A, which is a subring of TZ. Thus the Boolean rings TX of 311Ba are
leading examples in a very strong sense.
(c) I have taken the set Z of the Stone representation to be actually the set of homomorphisms from A
onto Z
2
. Of course we could equally well take any set which is in a natural one-to-one correspondence with
Z; a popular choice is the set of maximal ideals of A, since a subset of A is a maximal ideal i it is the
kernel of a member of Z, which is then uniquely dened.
16 Boolean algebras 311G
311G The operations , \ , . on a Boolean ring Let A be a Boolean ring.
(a) Using the Stone representation, we can see that the elementary operations , , , of set theory
all correspond to operations on A. If we set
a b = a +b +ab, a b = ab, a \ b = a +ab, a .b = a +b
for a, b A, then we see that

ab = a

b(a

b) = a

b,

ab = a

b,

a\b = a

b,

a.b = a

b.
Consequently all the familiar rules for manipulation of , , etc. will apply also to , , and we shall have,
for instance,
a (b c) = (a b) (a c), a (b c) = (a b) (a c)
for any members a, b, c of any Boolean ring A.
(b) Still importing terminology from elementary set theory, I will say that a set A A is disjoint if
a b = 0, that is, ab = 0, for all distinct a, b A; and that an indexed family a
i
)
iI
in A is disjoint if
a
i
a
j
= 0 for all distinct i, j I. (Just as I allow to be a member of a disjoint family of sets, I allow
0 A or a
i
= 0 in the present context.)
(c) A partition of unity in A will be either a disjoint set C A such that there is no non-zero a A
such that a c = 0 for every c C or a disjoint family c
i
)
iI
in A such that there is no non-zero a A
such that a c
i
= 0 for every i I. (In the rst case I allow 0 C, and in the second I allow c
i
= 0.)
(d) If C and D are two partitions of unity, I say that C renes D if for every c C there is a d D
such that cd = d. Note that if C renes D and D renes E then C renes E. PPP If c C, there is a d D
such that cd = c; now there is an e E such that de = d; in this case,
ce = (cd)e = c(de) = cd = c;
as c is arbitrary, C renes E. QQQ
311H The order structure of a Boolean ring Again treating a Boolean ring A as an algebra of
sets, we have a natural ordering on it, setting a b if ab = a, so that a b i a

b. This translation makes

it obvious that is a partial ordering on A, with least element 0, and with greatest element 1 i A is a
Boolean algebra. Moreover, A is a lattice (denition: 2A1Ad), with a b = supa, b, a b = infa, b for
all a, b A. Generally, for a
0
, . . . , a
n
A,
sup
in
a
i
= a
0
. . . a
n
, inf
in
a
i
= a
0
. . . a
n
;
suprema and inma of nite subsets A correspond to unions and intersections of the corresponding families
in the Stone space. (But suprema and inma of innite subsets of A are a very dierent matter; see 313
below.)
It may be obvious, but it is nevertheless vital to recognise that when A is a ring of sets then agrees
with .
311I The topology of a Stone space: Theorem Let Z be the Stone space of a Boolean ring A, and
let T be
G : G Z and for every z G there is an a A such that z a G.
Then T is a topology on Z, under which Z is a locally compact zero-dimensional Hausdor space, and
c = a : a A is precisely the set of compact open subsets of Z. A is a Boolean algebra i Z is compact.
311J Boolean algebras 17
proof (a) Because c is closed under , and

c = Z (recall that Z is the set of surjective homomorphisms

from A to Z
2
, so that every z Z is somewhere non-zero and belongs to some a), c is a topology base, and
T is a topology.
(b) T is Hausdor. PPP Take any distinct z, w Z. Then there is an a A such that z(a) ,= w(a);
let us take it that z(a) = 1, w(a) = 0. There is also a b A such that w(b) = 1, so that w(b + ab) =
w(b) +
2
w(a)w(b) = 1 and w (b +ab); also
a(b +ab) = ab +a
2
b = ab +ab = 0,
so
a (b +ab) = (a(b +ab)) =

0 = ,
and a, (b +ab) are disjoint members of T containing z, w respectively. QQQ
(c) If a A then a is compact. PPP Let T be an ultralter on Z containing a. For each b A,
z
0
(b) = lim
zJ
z(b) must be dened in Z
2
, since one of the sets z : z(b) = 0, z : z(b) = 1 must belong
to T. If b, c A, then the set
F = z : z(b) = z
0
(b), z(c) = z
0
(c), z(b +c) = z
0
(b +c), z(bc) = z
0
(bc)
belongs to T, so is not empty; take any z
1
F; then
z
0
(b +c) = z
1
(b +c) = z
1
(b) +
2
z
1
(c) = z
0
(b) +
2
z
0
(c),
z
0
(bc) = z
1
(bc) = z
1
(b)z
1
(c) = z
0
(b)z
0
(c).
As b, c are arbitrary, z
0
: A Z
2
is a ring homomorphism. Also z
0
(a) = 1, because a T, so z
0
a.
Now let G be any open subset of Z containing z
0
; then there is a b A such that z
0


b G; since
lim
zJ
z(b) = z
0
(b) = 1, we must have

b = z : z(b) = 1 T and G T. Thus T converges to z

0
. As T
is arbitrary, a is compact (2A3R). QQQ
(d) This shows that a is a compact open set for every a A. Moreover, since every point of Z belongs to
some a, every point of Z has a compact neighbourhood, and Z is locally compact. Every a is closed (because
it is compact, or otherwise), so c is a base for T consisting of open-and-closed sets, and T is zero-dimensional.
(e) Now suppose that E Z is an open compact set. If E = then E =

0. Otherwise, set
( = a : a A, a E.
Then ( is a family of open subsets of Z and

( = E, because E is open. But E is also compact, so there
is a nite (
0
( such that E =

(
0
. Express (
0
as a
0
, . . . , a
n
. Then
E = a
0
. . . a
n
= (a
0
. . . a
n
).
This shows that every compact open subset of Z is of the form a for some a A.
(f ) Finally, if A is a Boolean algebra then Z =

1 is compact, by (c); while if Z is compact then (e) tells

us that Z = a for some a A, and of course this a must be a multiplicative identity for A, so that A is a
Boolean algebra.
311J We have a kind of converse of Stones theorem.
Proposition Let X be a locally compact zero-dimensional Hausdor space. Then the set A of open-and-
compact subsets of X is a subring of TZ. If Z is the Stone space of A, there is a unique homeomorphism
: Z X such that a =
1
[a] for every a A.
proof (a) Because X is Hausdor, all its compact sets are closed, so every member of A is closed. Conse-
quently a b, a b, a b and ab belong to A for all a, b A, and A is a subring of TX.
It will be helpful to know that A is a base for the topology of X. PPP If G X is open and x G,
then (because X is locally compact) there is a compact set K X such that x int K; now (because X is
zero-dimensional) there is an open-and-closed set a X such that x a G int K; because a is a closed
subset of a compact subset of X, it is compact, and belongs to A, while x a G. QQQ
18 Boolean algebras 311J
(b) Let R Z X be the relation
(z, x): for every a A, x a z(a) = 1.
Then R is the graph of a bijective function : Z X.
PPP (i) If z Z and x, x
t
X are distinct, then, because X is Hausdor, there is an open set G X
containing x and not containing x
t
; because A is a base for the topology of X, there is an a A such that
x a G, so that x
t
/ a. Now either z(a) = 1 and (z, x
t
) / R, or z(a) = 0 and (z, x) / R. Thus R is the
graph of a function with domain included in Z and taking values in X.
(ii) If z Z, there is an a
0
A such that z(a
0
) = 1. Consider / = a : z(a) = 1. This is a family
of closed subsets of X containing the compact set a
0
, and a b / for all a, b /. So

/ is not empty
(3A3Db); take x

/. Then x a whenever z(a) = 1. On the other hand, if z(a) = 0, then

z(a
0
a) = z(a
0
(a a
0
)) = z(a
0
) +
2
z(a
0
)z(a) = 1,
so x a
0
a and x / a. Thus (z, x) R and (z) = x is dened. As z is arbitrary, the domain of is the
whole of Z.
(iii) If x X, dene z : A Z
2
by setting z(a) = 1 if x a, 0 otherwise. It is elementary to check
that z is a ring homomorphism form A to Z
2
. To see that it takes the value 1, note that because A is a
base for the topology of X there is an a A such that x a, so that z(a) = 1. So z Z, and of course
(z, x) R. As x is arbitrary, is surjective.
(iv) If z, z
t
Z and (z) = (z
t
), then, for any a A,
z(a) = 1 (z) a (z
t
) a z
t
(a) = 1,
so z = z
t
. Thus is injective. QQQ
(c) For any a A,

1
[a] = z : (z) a = z : z(a) = 1 = a.
It follows that is a homeomorphism. PPP (i) If G X is open, then (because A is a base for the topology
of X) G =

a : a A, a G and

1
[G] =

1
[a] : a A, a G =

a : a A, a G
is an open subset of Z. As G is arbitrary, is continuous. (ii) On the other hand, if G X and
1
[G] is
open, then
1
[G] is of the form

a,
a for some / A, so that G =

/ is an open set in X. Accordingly

is a homeomorphism. QQQ
(d) Finally, I must check the uniqueness of . But of course if

: Z X is any function such that

1
[a] = a for every a A, then the graph of

must be R, so

= .
311K Remark Thus we have a correspondence between Boolean rings and zero-dimensional locally
compact Hausdor spaces which is (up to isomorphism, on the one hand, and homeomorphism, on the
other) one-to-one. Every property of Boolean rings which we study will necessarily correspond to some
property of zero-dimensional locally compact Hausdor spaces.
311L Complemented distributive lattices I have introduced Boolean algebras through the theory
of rings; this seems to be the quickest route to them from an ordinary undergraduate course in abstract
algebra. However there are alternative approaches, taking the order structure rather than the algebraic
operations as fundamental, and for the sake of an application in Chapter 35 I give the details of one of these.
Proposition Let A be a lattice such that
(i) (a b) c = (a c) (b c) for all a, b, c A;
(ii) there is a bijection a a
t
: A A which is order-reversing, that is, a b i b
t
a
t
, and
such that a
tt
= a for every a;
(iii) A has a least element 0 and a a
t
= 0 for every a A.
Then A has a Boolean algebra structure for which a b i a b.
311X Boolean algebras 19
proof (a) Write 1 for 0
t
; if a A, then a
t
0 so a = a
tt
0
t
= 1, and 1 is the greatest element of A.
If a, b A then, because
t
is an order-reversing bijection, a
t
b
t
= (a b)
t
. PPP For c A,
a
t
b
t
c a
t
c &b
t
c c
t
a &c
t
b
c
t
a b (a b)
t
c. QQQ
Similarly, a
t
b
t
= (a b)
t
. If a, b, c A then
(a b) c = ((a
t
b
t
) c
t
)
t
= ((a
t
c
t
) (b
t
c
t
))
t
= (a c) (b c).
(b) Dene addition and multiplication on A by setting
a +b = (a b
t
) (a
t
b), ab = a b
for a, b A.
(c)(i) If a, b A then
(a +b)
t
= (a
t
b) (a b
t
) = (a
t
a) (a
t
b
t
) (b a) (b b
t
)
= 0 (a
t
b
t
) (b a) = (a
t
b
t
) (a b).
So if a, b, c A then
(a +b) +c = ((a +b) c
t
) ((a +b)
t
c)
= (((a b
t
) (a
t
b)) c
t
) (((a
t
b
t
) (a b)) c)
= (a b
t
c
t
) (a
t
b c
t
) (a
t
b
t
c) (a b c);
as this last formula is symmetric in a, b and c, it is also equal to a + (b +c). Thus addition is associative.
(ii) For any a A,
a + 0 = 0 +a = (a
t
0) (a 0
t
) = 0 (a 1) = a,
so 0 is the additive identity of A. Also
a +a = (a a
t
) (a
t
a) = 0 0 = 0
so each element of A is its own additive inverse, and (A, +) is a group. It is abelian because , are
commutative.
(d) Because is associative and commutative, (A, ) is a commutative semigroup; also 1 is its identity,
because a 1 = a for every a A. As for the distributive law in A,
ab +ac = (a b (a c)
t
) ((a b)
t
a c)
= (a b (a
t
c
t
)) ((a
t
b
t
) a c)
= (a b a
t
) (a b c
t
) (a
t
a c) (b
t
a c)
= (a b c
t
) (b
t
a c)
= a ((b c
t
) (b
t
c)) = a(b +c)
for all a, b, c A. Thus (A, +, ) is a ring; because a a = a for every a, it is a Boolean ring.
(e) For a, b A,
a b ab = a a b = a a b,
so the order relations of A coincide.
Remark It is the case that the Boolean algebra structure of A is uniquely determined by its order structure,
but I delay the proof to the next section (312L).
311X Basic exercises (a) Let A
0
, . . . , A
n
be sets. Show that
A
0
. . . A
n
= x : #(i : i n, x A
i
) is odd.
20 Boolean algebras 311Xb
(b) Let X be a set, and TX. Show that the following are equiveridical: (i) is an algebra of subsets
of X; (ii) is a subring of TX (that is, contains and is closed under and ) and contains X; (iii) ,
X E for every E , and E F for all E, F .
(c) Let A be any Boolean ring. Let a a
t
be any bijection between A and a set B disjoint from A. Set
B = A B, and extend the addition and multiplication of A to form binary operations on B by using the
formulae
a +b
t
= a
t
+b = (a +b)
t
, a
t
+b
t
= a +b,
a
t
b = b +ab, ab
t
= a +ab, a
t
b
t
= (a +b +ab)
t
.
Show that B is a Boolean algebra and that A is an ideal in B.
>>>(d) Let A be a Boolean ring, and K a nite subset of A. Show that the subring of A generated by K
has at most 2
2
#(K)
members, being the set of sums of products of members of K.
>>>(e) Show that any nite Boolean ring is isomorphic to TX for some nite set X (and, in particular, is
a Boolean algebra).
(f ) Let A be any Boolean ring. Show that
a (b c) = (a b) (a c), a (b c) = (a b) (a c)
for all a, b, c A directly from the denitions in 311G, without using Stones theorem.
>>>(g) Let A be any Boolean ring. Show that if we regard the Stone space Z of A as a subset of 0, 1
A
,
then the topology of Z (311I) is just the subspace topology induced by the ordinary product topology of
0, 1
A
.
(h) Let I be any set, and set X = 0, 1
I
with its usual topology (3A3K). Show that for a subset E of
X the following are equiveridical: (i) E is open-and-compact; (ii) E is determined by coordinates in a nite
subset of I (denition: 254M); (iii) E belongs to the algebra of subsets of X generated by E
i
: i I,
where E
i
= x : x(i) = 1 for each i.
(i) Let (A, ) be a lattice such that () A has a least element 0 and a greatest element 1 () for every
a, b, c A, a (b c) = (a b) (a c) and a (b c) = (a b) (a c) () for every a A there is an
a
t
A such that a a
t
= 1 and a a
t
= 0. Show that there is a Boolean algebra structure on A for which
agrees with .
311Y Further exercises (a) Let A be a Boolean ring, and B the Boolean algebra constructed by the
method of 311Xc. Show that the Stone space of B can be identied with the one-point compactication
(3A3O) of the Stone space of A.
(b) Let (A, , , 0, 1) be such that (i) (A, ) is a commutative semigroup with identity 0 (ii) (A, ) is a
commutative semigroup with identity 1 (iii) a (b c) = (a b) (a c), a (b c) = (a b) (a c) for all
a, b, c A (iv) a a = a a = a for every a A (v) for every a A there is an a
t
A such that a a
t
= 1,
a a
t
= 0. Show that there is a Boolean algebra structure on A for which = , = .
311 Notes and comments My aim in this section has been to get as quickly as possible to Stones theorem,
since this is surely the best route to a picture of general Boolean algebras; they are isomorphic to algebras
of sets. This means that all their elementary algebraic properties indeed, all their rst-order properties
can be eectively studied in the context of elementary set theory. In 311G-311H I describe a few of the
ways in which the Stone representation suggests algebraic properties of Boolean algebras.
You should not, however, come too readily to the conclusion that Boolean algebras will never be able to
surprise you. In this book, in particular, we shall need to work a good deal with suprema and inma of
innite sets in Boolean algebras, for the ordering of 311H; and even though this corresponds to the ordering
of ordinary sets, we nd that (supA) is suciently dierent from

aA
a to need new kinds of intuition.
312B Homomorphisms 21
(The point is that

aA
a is an open set in the Stone space, but need not be compact if A is innite, so
may well be smaller than (sup A), even when supA is dened in A.) There is also the fact that Stones
theorem depends crucially on a fairly strong form of the axiom of choice (employed through Zorns Lemma
in the argument of 311D). Of course I shall be using the axiom of choice without scruple throughout this
volume. But it should be clear that such results as 312B-312C in the next section cannot possibly need the
axiom of choice for their proofs, and that to use Stones theorem in such a context is slightly misleading.
Nevertheless, it is so useful to be able to regard a Boolean algebra as an algebra of sets especially when
dealing with only nitely many elements of the algebra at a time that henceforth I will almost always use
the symbols ., for the addition and multiplication of a Boolean ring, and will use , \ , without
further comment, just as if I were considering , and in the Stone space. (In 311Gb I have given a
denition of disjointness in a Boolean algebra based on the same idea.) Even without the axiom of choice
this approach can be justied, once we have observed that nitely-generated Boolean algebras are nite
(311Xd), since relatively elementary methods show that any nite Boolean algebra is isomorphic to TX for
some nite set X.
I have taken a Boolean algebra to be a particular kind of commutative ring with identity. Of course
there are other approaches. If we wish to think of the order relation as primary, then 311L and 311Xi are
reasonably natural. Other descriptions can be based on a list of the properties of the binary operations ,
and the complementation operation a a
t
= 1 \ a, as in 311Yb. I give extra space to 311L only because
this is well adapted to an application in 352Q below.
312 Homomorphisms
I continue the theory of Boolean algebras with a section on subalgebras, ideals and homomorphisms.
From now on, I will relegate Boolean rings which are not algebras to the exercises; I think there is no need
to set out descriptions of the triing modications necessary to deal with the extra generality. The rst part
of the section (312A-312K) concerns the translation of the basic concepts of ring theory into the language
which I propose to use for Boolean algebras. 312L shows that the order relation on a Boolean algebra denes
the algebraic structure, and in 312M-312N I give a fundamental result on the extension of homomorphisms.
I end the section with results relating the previous ideas to the Stone representation of a Boolean algebra
(312O-312S).
312A Subalgebras Let A be a Boolean algebra. I will use the phrase subalgebra of A to mean a
subring of A containing its multiplicative identity 1 = 1
A
.
312B Proposition Let A be a Boolean algebra, and B a subset of A. Then the following are equiv-
eridical, that is, if one is true so are the others:
(i) B is a subalgebra of A;
(ii) 0 B, a b B for all a, b B, and 1 \ a B for all a B;
(iii) B ,= , a b B for all a, b B, and 1 \ a B for all a B.
proof (a)(i)(iii) If B is a subalgebra of A, and a, b B, then of course we shall have
0, 1 B, so B ,= ,
a b B, 1 \ a = 1 .a B.
(b)(iii)(ii) If (iii) is true, then there is some b
0
B; now 1 \ b
0
B, so
0 = b
0
(1 \ b
0
) B.
If a, b B, then
a b = 1 \ ((1 \ a) (1 \ b)) B.
So (ii) is true.
22 Boolean algebras 312B
(c)(ii)(i) If (ii) is true, then for any a, b B,
a b = 1 \ ((1 \ a) (1 \ b)) B,
a .b = (a (1 \ b)) (b (1 \ a)) B,
so (because also 0 B) B is a subring of A, and
1 = 1 \ 0 B,
so B is a subalgebra.
Remark Thus an algebra of subsets of a set X, as dened in 136E or 311Bb, is just a subalgebra of the
Boolean algebra TX.
312C Ideals in Boolean algebras: Proposition If A is a Boolean algebra, a set I A is an ideal of
A i 0 I, a b I for all a, b I, and a I whenever b I and a b.
proof (a) Suppose that I is an ideal. Then of course 0 I. If a, b I then a b I so a b =
(a .b) .(a b) I. If b I and a b then a = a b I.
(b) Now suppose that I satises the conditions proposed. If a, b I then
a .b a b I
so a .b I, while of course a = a I, and also 0 I, by hypothesis; thus I is a subgroup of (A, .).
Finally, if a I and b A then
a b a I,
so b a = a b I; thus I is an ideal.
Remark Thus what I have called an ideal of subsets of X in 232Xc is just an ideal in the Boolean algebra
TX.
312D Principal ideals Of course, while an ideal I in a Boolean algebra A is necessarily a subring, it
is not as a rule a subalgebra, except in the special case I = A. But if we say that a principal ideal of A is
the ideal A
a
generated by a single element a of A, we have a special phenomenon.
312E Proposition Let A be a Boolean algebra, and a any element of A. Then the principal ideal A
a
of A generated by a is just b : b A, b a, and (with the inherited operations A
a
A
a
, .A
a
A
a
)
is a Boolean algebra in its own right, with multiplicative identity a.
proof b a i b a = a, so that
A
a
= b : b a = b a : b A
is an ideal of A, and of course it is the smallest ideal of A containing a. Being an ideal, it is a subring; the
idempotent relation b b = b is inherited from A, so it is a Boolean ring; and a is plainly its multiplicative
identity.
312F Boolean homomorphisms Now suppose that A and B are two Boolean algebras. I will use the
phrase Boolean homomorphism to mean a function : A B which is a ring homomorphism (that is,
(a .b) = a .b, (a b) = a b for all a, b A) which is uniferent, that is, (1
A
) = 1
B
.
312G Proposition Let A, B and C be Boolean algebras.
(a) If : A B is a Boolean homomorphism, then [A] is a subalgebra of B.
(b) If : A B and : B C are Boolean homomorphisms, then : A C is a Boolean
homomorphism.
(c) If : A B is a bijective Boolean homomorphism, then
1
: B A is a Boolean homomorphism.
proof These are all immediate consequences of the corresponding results for ring homomorphisms (3A2D).
312K Homomorphisms 23
312H Proposition Let A and B be Boolean algebras, and : A B a function. Then the following
are equiveridical:
(i) is a Boolean homomorphism;
(ii) (a b) = a b and (1
A
\ a) = 1
B
\ a for all a, b A;
(iii) (a b) = a b and (1
A
\ a) = 1
B
\ a for all a, b A;
(iv) (a b) = a b and a b = 0
B
whenever a, b A and a b = 0
A
, and (1
A
) = 1
B
.
proof (i)(iv) If is a Boolean homomorphism then of course (1
A
) = 1
B
; also, given that a b = 0 in A,
a b = (a b) = (0
A
) = 0
B
,
(a b) = (a .b) = a .b = a b.
(iv)(iii) Assume (iv), and take a, b A. Then
a = (a b) (a \ b), b = (a b) (b \ a),
so
(a b) = a (b \ a) = (a b) (a \ b) (b \ a) = a b.
Taking b = 1 \ a, we must have
1
B
= (1
A
) = a (1
A
\ a), 0
B
= a (1
A
\ a),
so (1
A
a) = 1
B
\ a. Thus (iii) is true.
(iii)(ii) If (iii) is true and a, b A, then
(a b) = (1
A
\ ((1
A
\ a) (1
A
\ b)))
= 1
B
\ ((1
B
\ a) (1
B
\ b))) = a b.
So (ii) is true.
(ii)(i) If (ii) is true, then
(a .b) = ((1
A
\ ((1
A
\ a) (1
A
\ b)) (1
A
\ (a b)))
= (1
B
\ ((1
B
\ a) (1
B
\ b)) (1
B
\ (a b))) = a .b
for all a, b A, so is a ring homomorphism; and now
(1
A
) = (1
A
\ 0
A
) = 1
B
\ (0
A
) = 1
B
\ 0
B
= 1
B
,
so that is a Boolean homomorphism.
312I Proposition If A, B are Boolean algebras and : A B is a Boolean homomorphism, then
a b whenever a b in A.
proof
a b = a b = a = a b = a = a b.
312J Proposition Let A be a Boolean algebra, and a any member of A. Then the map b a b is a
surjective Boolean homomorphism from A onto the principal ideal A
a
generated by a.
proof This is an elementary verication.
312K Quotient algebras: Proposition Let A be a Boolean algebra and I an ideal of A. Then the
quotient ring A/I (3A2F) is a Boolean algebra, and the canonical map a a

: A A/I is a Boolean
homomorphism, so that
(a .b)

= a

.b

, (a b)

= a

, (a b)

= a

, (a \ b)

= a

\ b

for all a, b A.
24 Boolean algebras 312K
(b) The order relation on A/I is dened by the formula
a

a \ b I.
For any a A,
u : u a

= b

: b a.
proof (a) Of course the map a a

= a .b : b I is a ring homomorphism (3A2F). Because

(a

)
2
= (a
2
)

= a

for every a A, A/I is a Boolean ring; because 1

is a multiplicative identity, it is a Boolean algebra, and

a a

is a Boolean homomorphism. The formulae given are now elementary.

(b) We have
a

\ b

= 0 a \ b I.
Now
u : u a

= u a

: u A/I = (b a)

: b A = b

: b a.
312L The above results are both repetitive and nearly trivial. Now I come to something with a little
more meat to it.
Proposition If A and B are Boolean algebras and : A B is a bijection such that a b whenever
a b, then is a Boolean algebra isomorphism.
proof (a) Because is surjective, there must be c
0
, c
1
A such that c
0
= 0
B
, c
1
= 1
B
; now (0
A
) c
0
,
c
1
(1
A
), so we must have (0
A
) = 0
B
, (1
A
) = 1
B
.
(b) If a A, then a (1
A
\ a) = 1
B
. PPP There is a c A such that c = 1
B
\ (a (1
A
\ a)). Now
(c a) c a = 0
B
, (c \ a) c (1
A
\ a) = 0
B
;
as is injective, c a = c \ a = 0
A
and c = 0
A
, c = 0
B
, a (1
A
\ a) = 1
B
. QQQ
(c) If a A, then a (1
A
\ a) = 0
B
. PPP It may be clear to you that this is just a dual form of (b). If not,
I repeat the argument in the form now appropriate. There is a c A such that c = 1
B
\ (a (1
A
\ a)).
Now
(c a) c a = 1
B
, (c (1
A
\ a)) c (1
A
\ a) = 1
B
;
as is injective, c a = c (1
A
\ a) = 1
A
and c = 1
A
, c = 1
B
, a (1
A
\ a) = 0
B
. QQQ
(d) Putting (b) and (c) together, we have (1
A
\ a) = 1
B
\ a for every a A. Now (a b) = a b
for every a, b A. PPP Surely a b (a b). Let c A be such that c = (a b) \ (a b). Then
(c a) c a = 0
B
, (c b) c b = 0
B
,
so c a = c b = 0 and c 1
A
\ (a b); accordingly
c (1
A
\ (a b)) = 1
B
\ (a b);
as also c (a b), c = 0
B
and (a b) = a b. QQQ
(e) So the conditions of 312H(iii) are satised and is a Boolean homomorphism; being bijective, it is
an isomorphism.
312M I turn next to a fundamental lemma on the construction of homomorphisms. We need to start
with a proper description of a certain type of subalgebra.
Lemma Let A be a Boolean algebra, and A
0
a subalgebra of A; let c be any member of A. Then
A
1
= (a c) (b \ c) : a, b A
0

is a subalgebra of A; it is the subalgebra of A generated by A

0
c.
312N Homomorphisms 25
proof We have to check the following:
a = (a c) (a \ c) A
1
for every a A
0
, so A
0
A
1
; in particular, 0 A
1
.
1 \ ((a c) (b \ c)) = ((1 \ a) c) ((1 \ b) \ c) A
1
for all a, b A
0
, so 1 \ d A
1
for every d A
1
.
(a c) (b \ c) (a
t
c) (b
t
\ c) = ((a a
t
) c) ((b b
t
) \ c) A
1
for all a, b, a
t
, b
t
A
0
, so d d
t
A
1
for all d, d
t
A
1
. Thus A
1
is a subalgebra of A (using 312B).
c = (1 c) (0 \ c) A
1
,
so A
1
includes A
0
c; and nally it is clear that any subalgebra of A including A
0
c, being closed under
, and complementation, must include A
1
, so that A
1
is the subalgebra of A generated by A
0
c.
312N Lemma Let A and B be Boolean algebras, A
0
a subalgebra of A, : A
0
B a Boolean
homomorphism, and c A. If v B is such that a v b whenever a, b A
0
and a c b, then there
is a unique Boolean homomorphism
1
from the subalgebra A
1
of A generated by A
0
c such that
1
extends and
1
c = v.
proof (a) The basic fact we need to know is that if a, a
t
, b, b
t
A
0
and
(a c) (b \ c) = d = (a
t
c) (b
t
\ c),
then
(a v) (b \ v) = (a
t
v) (b
t
\ v).
PPP We have
a c = d c = a
t
c.
Accordingly (a .a
t
) c = 0 and c 1 \ (a .a
t
). Consequently (since a .a
t
surely belongs to A
0
)
v (1 \ (a .a
t
)) = 1 \ (a .a
t
),
and
a v = a
t
v.
Similarly,
b \ c = d \ c = b
t
\ c,
so
(b .b
t
) \ c = 0, b .b
t
c, (b .b
t
) v
and
b \ v = b
t
\ v.
Putting these together, we have the result. QQQ
(b) Consequently, we have a function
1
dened by writing

1
((a c) (b \ c)) = (a v) (b \ c)
for all a, b A
0
; and 312M tells us that the domain of
1
is just A
1
. Now
1
is a Boolean homomorphism.
PPP This amounts to running through the proof of 312M again.
(i) If a, b A
0
, then

1
(1 \ ((a c) (b \ c))) =
1
(((1 \ a) c) ((1 \ b) \ c))
= ((1 \ a) v) ((1 \ b) \ v)
= ((1 \ a) v) ((1 \ b) \ v)
= 1 \ ((a v) (b \ v)) = 1 \
1
((a c) (b \ c)).
26 Boolean algebras 312N
So
1
(1 \ d) = 1 \
1
d for every d A
1
.
(ii) If a, b, a
t
, b
t
A
0
, then

1
((a c) (b \ c) (a
t
c) (b
t
\ c)) =
1
(((a a
t
) c) ((b b
t
) \ c))
= ((a a
t
) v) ((b b
t
) \ v)
= ((a a
t
) v) ((b b
t
) \ v)
= (a v) (b \ v) (a
t
v) (b
t
\ v)
=
1
((a c) (b \ c))
1
((a
t
v) (b
t
\ v)).
So
1
(d d
t
) =
1
d
1
d
t
for all d, d
t
A
1
.
By 312H(iii),
1
is a Boolean homomorphism. QQQ
(c) If a A
0
, then

1
a =
1
((a c) (a \ c)) = (a v) (a \ v) = a,
so
1
extends . As for the action of
1
on c,

1
c =
1
((1 c) (0 \ c)) = (1 v) (0 \ v) = (1 v) (0 \ v) = v,
as required.
(d) Finally, the formula of (b) is the only possible denition for any Boolean homomorphism from A
1
to
B which will extend and take c to v. So
1
is unique.
312O Homomorphisms and Stone spaces Because the Stone space Z of a Boolean algebra A (311E)
can be constructed explicitly from the algebraic structure of A, it must in principle be possible to describe
any feature of the Boolean structure of A in terms of Z. In the next few paragraphs I work through the
most important identications.
Proposition Let A be a Boolean algebra, and Z its Stone space; write a Z for the open-and-closed set
corresponding to a A. Then there is a one-to-one correspondence between ideals I of A and open sets
G Z, given by the formulae
G =

aI
a, I = a : a G.
proof (a) For any ideal I A, set H(I) =

aI
a; then H(I) is a union of open subsets of Z, so is open.
For any open set G Z, set J(G) = a : a A, a G; then J(G) satises the conditions of 312C, so is
an ideal of A.
(b) If I A, then J(H(I)) = I. PPP (i) If a I, then a H(I) so a J(H(I)). (ii) If a J(H(I)), then
a H(I) =

bI

b. Because a is compact and all the

b are open, there must be nitely many b

0
, . . . , b
n
I
such that a

b
0
. . .

b
n
. But now a b
0
. . . b
n
I, so a I. QQQ
(c) If G Z is open, then H(J(G)) = G. PPP (i) If z G, then (because a : a A is a base for the
topology of Z) there is an a A such that z a G; now a J(G) and z H(J(G)). (ii) If z H(J(G)),
there is an a J(G) such that z a; now a G, so z G. QQQ
This shows that the maps G J(G), I H(I) are two halves of a one-to-one correspondence, as
required.
312P Theorem Let A, B be Boolean algebras, with Stone spaces Z, W; write a Z,

b W for the
open-and-closed sets corresponding to a A, b B. Then we have a one-to-one correspondence between
Boolean homomorphisms : A B and continuous functions : W Z, given by the formula
a = b
1
[a] =

b.
proof (a) Recall that I have constructed Z, W as the sets of Boolean homomorphisms from A, B to
Z
2
(311E). So if : A B is any Boolean homomorphism, and w W,

(w) = w is a Boolean
homomorphism from A to Z
2
(312Gb), and belongs to Z. Now
1

[a] = a for every a A. PPP

312R Homomorphisms 27

[a] = w :

(w) a = w : w a = w : w(a) = 1 = w : w a. QQQ

Consequently

is continuous. PPP Let G be any open subset of Z. Then G =

a : a G, so

[G] =

[a] : a G =

a : a G
is open. As G is arbitrary,

is continuous. QQQ
(b) If : W Z is continuous, then for any a A the set
1
[a] must be an open-and-closed set in
W; consequently there is a unique member of B, call it

a, such that
1
[a] =

a. Observe that, for any

w W and a A,
w(

a) = 1 w

a (w) a ((w))(a) = 1,
so (w) = w

.
Now

is a Boolean homomorphism. PPP (i) If a, b A then

(a b) =
1
[(a b)] =
1
[a

b] =
1
[a]
1
[

b] =

b = (

b),
so

(a b) =

b. (ii) If a A, then

(1 \ a) =
1
[(1 \ a)] =
1
[Z a] = W
1
[a] = W

a = (1 \

a),
so

(1 \ a) = 1 \

a. (iii) By 312H,

is a Boolean homomorphism. QQQ

(c) For any Boolean homomorphism : A B, =

. PPP For a A,
(

a) =
1

[a] = a,
so

a = a. QQQ
(d) For any continuous function : W Z, =

. PPP For any w W,

(w) = w

= (w). QQQ
(e) Thus

are the two halves of a one-to-one correspondence, as required.

312Q Theorem Let A, B, C be Boolean algebras, with Stone spaces Z, W and V . Let : A B
and : B C be Boolean homomorphisms, with corresponding continuous functions : W Z and
: V W. Then the Boolean homomorphism : A C corresponds to the continuous function
: V Z.
proof For any a A,

a = ((a)) =
1
[ a] =
1
[
1
[a]] = ()
1
[a].
312R Proposition Let A and B be Boolean algebras, with Stone spaces Z and W, and : A B a
Boolean homomorphism, with associated continuous function : W Z. Then
(a) is injective i is surjective;
(b) is surjective i is injective.
proof (a) If a A, then
a [W] = (w) / a for every w W
((w))(a) = 0 for every w W
w(a) = 0 for every w W
a = 0.
Now W is compact, so [W] is also compact, therefore closed, and
28 Boolean algebras 312R
is not surjective Z [W] ,=
there is a non-zero a A such that a Z [W]
there is a non-zero a A such that a = 0
is not injective
(3A2Db).
(b)(i) If is surjective and w, w
t
are distinct members of W, then there is a b B such that w

b
and w
t
/

b. Now b = a for some a A, so (w) a and (w

t
) / a, and (w) ,= (w
t
). As w and w
t
are
arbitrary, is injective.
(ii) If is injective and b B, then K = [

b], L = [W

b] are disjoint compact subsets of Z. Consider

I = a : a A, L a = . Then

aI
a = Z L K. Because K is compact and every a is open, there
is a nite family a
0
, . . . , a
n
I such that K a
0
. . . a
n
. Set a = a
0
. . . a
n
. Then a = a
0
. . . a
n
includes K and is disjoint from L. So a =
1
[a] includes

b and is disjoint from W

b; that is, a =

b and
a = b. As b is arbitrary, is surjective.
312S Principal ideals If A is a Boolean algebra and a A, we have a natural surjective Boolean
homomorphism b b a : A A
a
, the principal ideal generated by a (312J). Writing Z for the Stone space
of A and Z
a
for the Stone space of A
a
, this homomorphism must correspond to an injective continuous
function : Z
a
Z (312Rb). Because Z
a
is compact and Z is Hausdor, must be a homeomorphism
between Z
a
and its image [Z
a
] Z (3A3Dd). To identify [Z
a
], note that it is compact, therefore closed,
and that
Z [Z
a
] =
_

b : b A,

b [Z
a
] =
=
_

b :
1
[

b] = =
_

b : b a = 0 = Z a,
so that [Z
a
] = a. It is therefore natural to identify Z
a
with the open-and-closed set a Z.
312X Basic exercises (a) Let A be a Boolean ring, and B a subset of A. Show that B is a subring
of A i 0 B and a b, a \ b B for all a, b B.
(b) Let A be a Boolean algebra and B a subset of A. Show that B is a subalgebra of A i 1 B and
a \ b B for all a, b B.
(c) Let A be a Boolean algebra. Suppose that I A A are such that 1 A, a b I for all a, b I
and a \ b A whenever a, b A and b a. Show that A includes the subalgebra of A generated by I.
(Hint: 136Xf.)
(d) Show that if A is a Boolean ring, a set I A is an ideal of A i 0 I, a b I for all a, b I, and
a I whenever b I and a b.
(e) Let A and B be Boolean algebras, and : A B a function such that (i) (a) (b) whenever a b
(ii) (a) (b) = 0
B
whenever a b = 0
A
(iii) (a) (b) (c) = 1
B
whenever a b c = 1
A
. Show that
is a Boolean homomorphism.
(f ) Let A be a Boolean ring, and a any member of A. Show that the map b ab is a ring homomorphism
from A onto the principal ideal A
a
generated by a.
(g) Let A
1
and A
2
be Boolean rings, and let B
1
, B
2
be the Boolean algebras constructed from them
by the method of 311Xc. Show that any ring homomorphism from A
1
to A
2
has a unique extension to a
Boolean homomorphism from B
1
to B
2
.
312 Notes Homomorphisms 29
(h) Let A and B be Boolean rings, A
0
a subalgebra of A, : A
0
B a ring homomorphism, and c A.
Show that if v B is such that a \ v = b v = 0 whenever a, b A
0
and a \ c = b c = 0, then there is
a unique ring homomorphism
1
from the subring A
1
of A generated by A
0
c such that
1
extends
0
and
1
c = v.
(i) Let A be a Boolean ring, and Z its Stone space. Show that there is a one-to-one correspondence
between ideals I of A and open sets G Z, given by the formulae G =

aI
a, I = a : a G.
(j) Let A be a Boolean algebra, and suppose that A is the subalgebra of itself generated by A
0
c,
where A
0
is a subalgebra of A and c A. Let Z be the Stone space of A and Z
0
the Stone space of A
0
. Let
: Z Z
0
be the continuous surjection corresponding to the embedding of A
0
in A. Show that c and
Z c are injective.
Now let B be another Boolean algebra, with Stone space W, and : A
0
B a Boolean homomorphism,
with corresponding function : W Z
0
. Show that there is a continuous function
1
: W Z such that

1
= i there is an open-and-closed set V W such that [V ] [c] and [W V ] [Z c].
(k) Let A be a Boolean algebra, with Stone space Z, and I an ideal of A, corresponding to an open set
G Z. Show that the Stone space of the quotient algebra A/I may be identied with Z G.
312Y Further exercises (a) Find a function : T0, 1, 2 Z
2
such that (1 \ a) = 1 \ a for every
a T0, 1, 2 and (a) (b) whenever a b, but is not a Boolean homomorphism.
(b) Let A be the Boolean ring of nite subsets of N. Show that there is a bijection : A A such that
a b whenever a b but is not a ring homomorphism.
(c) Let A, B be Boolean rings, with Stone spaces Z, W. Show that we have a one-to-one correspondence
between ring homomorphisms : A Band continuous functions : H Z, where H W is an open set,
such that
1
[K] is compact for every compact set K Z, given by the formula a = b
1
[a] =

b.
(d) Let A, B, C be Boolean rings, with Stone spaces Z, W and V . Let : A B and : B C be ring
homomorphisms, with corresponding continuous functions : H Z and : G W. Show that the ring
homomorphism : A C corresponds to the continuous function :
1
[H] Z.
(e) Let A and B be Boolean rings, with Stone spaces Z and W, and : A B a ring homomorphism,
with associated continuous function : H Z. Show that is injective i [H] is dense in Z, and that
is surjective i is injective and H = W.
(f ) Let A be a Boolean ring and a A. Show that the Stone space of the principal ideal A
a
of A generated
by a can be identied with the compact open set a in the Stone space of A. Show that the identity map is
a ring homomorphism from A
a
to A, and corresponds to the identity function on a.
312 Notes and comments The denitions of subalgebra and Boolean homomorphism (312A, 312F),
like that of Boolean algebra, are a trie arbitrary, but will be a convenient way of mandating appropriate
treatment of multiplicative identities. I run through the work of 312A-312J essentially for completeness;
once you are familiar with Boolean algebras, they should all seem obvious. 312L has a little bit more to it.
It shows that the order structure of a Boolean algebra denes the ring structure, in a fairly strong sense.
I call 312N a lemma, but actually it is the most important result in this section; it is the basic tool we have
for extending a homomorphism from a subalgebra to a slightly larger one, and with Zorns Lemma (another
lemma which deserves a capital L) will provide us with general methods of constructing homomorphisms.
In 312O-312S I describe the basic relationships between the Boolean homomorphisms and continuous
functions on Stone spaces. 312P-312Q show that, in the language of category theory, the Stone representation
provides a contravariant functor from the category of Boolean algebras with Boolean homomorphisms to the
category of topological spaces with continuous functions. Using 311I-311J, we know exactly which topological
spaces appear, the zero-dimensional compact Hausdor spaces; and we know also that the functor is faithful,
that is, that we can recover Boolean algebras and homomorphisms from the corresponding topological spaces
30 Boolean algebras 312 Notes
and continuous functions. There is an agreeable duality in 312R. All of this can be done for Boolean rings,
but there are some extra complications (312Yc-312Yf).
To my mind, the very essence of the theory of Boolean algebras is the fact that they are abstract rings,
but at the same time can be thought of locally as algebras of sets. Consequently we can bring two
quite separate kinds of intuition to bear. 312N gives an example of a ring-theoretic problem, concerning the
extension of homomorphisms, which has a resolution in terms of the order relation, a concept most naturally
described in terms of algebras-of-sets. It is very much a matter of taste and habit, but I myself nd that a
Boolean homomorphism is easiest to think of in terms of its action on nite subalgebras, which are directly
representable as TX for some nite X (311Xe); the corresponding continuous map between Stone spaces is
less helpful. I oer 312Xj, the Stone-space version of 312N, for you to test your own intuitions on.
313 Order-continuous homomorphisms
Because a Boolean algebra has a natural partial order (311H), we have corresponding notions of upper
bounds, lower bounds, suprema and inma. These are particularly important in the Boolean algebras
arising in measure theory, and the innitary operations sup and inf require rather more care than the
basic binary operations , , because intuitions from elementary set theory are sometimes misleading.
I therefore take a section to work through the most important properties of these operations, together with
the homomorphisms which preserve them.
313A Relative complementation: Proposition Let A be a Boolean algebra, e a member of A, and
A a non-empty subset of A.
(a) If supA is dened in A, then infe \ a : a A is dened and equal to e supA.
(b) If inf A is dened in A, then supe \ a : a A is dened and equal to e inf A.
proof (a) Writing a
0
for supA, we have e \ a
0
e \ a for every a A, so e \ a
0
is a lower bound for
C = e \ a : a A. Now suppose that c is any lower bound for C. Then (because A is not empty) c e,
and
a = (a \ e) (e \ (e \ a)) (a
0
\ e) (e \ c)
for every a A. Consequently a
0
(a
0
\ e) (e \ c) is disjoint from c and
c = c e e \ a
0
.
Accordingly e \ a
0
is the greatest lower bound of C, as claimed.
(b) This time set a
0
= inf A, C = e \ a : a A. As before, e \ a
0
is surely an upper bound for C. If c
is any upper bound for C, then
e \ c e \ (e \ a) = e a a
for every a A, so e \ c a
0
and e \ a
0
c. As c is arbitrary, e \ a
0
is indeed the least upper bound of C.
Remark In the arguments above I repeatedly encourage you to treat , , \ , as if they were the
corresponding operations and relation of basic set theory. This is perfectly safe so long as we take care that
every manipulation so justied has only nitely many elements of the Boolean algebra in hand at once.
313B General distributive laws: Proposition Let A be a Boolean algebra.
(a) If e A and A A is a non-empty set such that supA is dened in A, then supe a : a A is
dened and equal to e supA.
(b) If e A and A A is a non-empty set such that inf A is dened in A, then infe a : a A is
dened and equal to e inf A.
(c) Suppose that A, B A are non-empty and supA, sup B are dened in A. Then supa b : a A, b
B is dened and is equal to supA supB.
(d) Suppose that A, B A are non-empty and inf A, inf B are dened in A. Then infa b : a A, b B
is dened and is equal to inf A inf B.
313C Order-continuous homomorphisms 31
proof (a) Set
B = e \ a : a A, C = e \ b : b B = e a : a A.
Using 313A, we have
inf B = e \ supA, sup C = e \ inf B = e supA,
as required.
(b) Set a
0
= inf A, B = e a : a A. Then e a
0
e a for every a A, so e a
0
is a lower bound
for B. If c is any lower bound for B, then c \ e a for every a A, so c \ e a
0
and c e a
0
; thus e a
0
is the greatest lower bound for B, as claimed.
(c) By (a), we have
a supB = sup
bB
a b
for every a A, so
sup
aA,bB
a b = sup
aA
(a supB) = sup A supB,
using (a) again.
(d) Similarly, using (b) twice,
inf
aA,bB
a b = inf
aA
(a inf B) = inf A inf B.
313C As always, it is worth developing a representation of the concepts of sup and inf in terms of
Stone spaces.
Proposition Let A be a Boolean algebra, and Z its Stone space; for a A write a for the corresponding
open-and-closed subset of Z.
(a) If A A and a
0
A then a
0
= supA in A i a
0
=

aA
a.
(b) If A A is non-empty and a
0
A then a
0
= inf A in A i a
0
= int

aA
a.
(c) If A A is non-empty then inf A = 0 in A i

aA
a is nowhere dense in Z.
proof (a) For any b A,
b is an upper bound for A a

b for every a A

_
aA
a

b
_
aA
a

b
because

b is certainly closed in Z. It follows at once that if a

0
is actually equal to

aA
a then a
0
must be
the least upper bound of A in A. On the other hand, if a
0
= supA, then

aA
a a
0
. ??? If a
0
,=

aA
a,
then a
0

aA
a is a non-empty open set in Z, so includes

b for some non-zero b A; now a a

0

b, so
a a
0
\ b for every a A, and a
0
\ b is an upper bound for A strictly less than a
0
. XXX Thus a
0
must be
exactly

aA
a.
(b) Take complements: setting a
1
= 1 \ a
0
, we have
a
0
= inf A a
1
= sup
aA
1 \ a
(by 313A)
a
1
=
_
aA
Z a
a
0
= Z
_
aA
Z a = int

aA
a.
(c) Since

aA
a is surely a closed set, it is nowhere dense i it has empty interior, that is, i 0 = inf A.
32 Boolean algebras 313D
313D I started the section with the results above because they are easily stated and of great importance.
But I must now turn to some new denitions, and I think it may help to clarify the ideas involved if I give
them in their own natural context, even though this is far more general than we have any immediate need
for here.
Denitions Let P be a partially ordered set and C a subset of P.
(a) C is order-closed if supA C whenever A is a non-empty upwards-directed subset of C such that
supA is dened in P, and inf A C whenever A is a non-empty downwards-directed subset of C such that
inf A is dened in P.
(b) C is sequentially order-closed if sup
nN
p
n
C whenever p
n
)
nN
is a non-decreasing sequence in
C such that sup
nN
p
n
is dened in P, and inf
nN
p
n
C whenever p
n
)
nN
is a non-increasing sequence in
C such that inf
nN
p
n
is dened in P.
Remark I hope it is obvious that an order-closed set is sequentially order-closed.
313E Order-closed subalgebras and ideals Of course, in the very special cases of a subalgebra
or ideal of a Boolean algebra, the concepts order-closed and sequentially order-closed have expressions
simpler than those in 313D. I spell them out.
(a) Let B be a subalgebra of a Boolean algebra A.
(i) The following are equiveridical:
() B is order-closed in A;
() supB B whenever B B and supB is dened in A;
(
t
) inf B B whenever B B and inf B is dened in A;
() supB B whenever B B is non-empty and upwards-directed and supB is dened in A;
(
t
) inf B B whenever B B is non-empty and downwards-directed and inf B is dened in A.
PPP Of course () (). If () is true and B B is any set with a supremum in A, then B
t
=
0 b
0
. . . b
n
: b
0
, . . . , b
n
B is a non-empty upwards-directed set with the same upper bounds as
B, so supB = sup B
t
B. Thus () () and (), () are equiveridical. Next, if () is true and B B is
a set with an inmum in A, then B
t
= 1 \ b : b B B and sup B
t
= 1 \ inf B is dened, so sup B
t
and
inf B belong to B . Thus () (
t
). In the same way, (
t
) (
t
) () and (), (
t
), (), (
t
) are all
equiveridical. But since we also have () ()&(
t
), () is equiveridical with the others. QQQ
Replacing the sets B above by sequences, the same arguments provide conditions for B to be sequentially
order-closed, as follows.
(ii) The following are equiveridical:
() B is sequentially order-closed in A;
() sup
nN
b
n
B whenever b
n
)
nN
is a sequence in B and sup
nN
b
n
is dened in A;
(
t
) inf
nN
b
n
B whenever b
n
)
nN
is a sequence in B and inf
nN
b
n
is dened in A;
() sup
nN
b
n
B whenever b
n
)
nN
is a non-decreasing sequence in B and sup
nN
b
n
is dened in
A;
(
t
) inf
nN
b
n
B whenever b
n
)
nN
is a non-increasing sequence in B and inf
nN
b
n
is dened in
A.
(b) Now suppose that I is an ideal of A. Then if A I is non-empty all lower bounds of A necessarily
belong to I; so that
I is order-closed i supA I whenever A I is non-empty, upwards-directed and has a
supremum in A;
I is sequentially order-closed i sup
nN
a
n
I whenever a
n
)
nN
is a non-decreasing sequence
in I with a supremum in A.
Moreover, because I is closed under ,
I is order-closed i sup A I whenever A I has a supremum in A;
I is sequentially order-closed i sup
nN
a
n
I whenever a
n
)
nN
is a sequence in I with a
supremum in A.
313G Order-continuous homomorphisms 33
(c) If A = TX is a power set, then a sequentially order-closed subalgebra of A is just a -algebra of sets,
while a sequentially order-closed ideal of A is a what I have called a -ideal of sets (112Db). If A is itself a
-algebra of sets, then a sequentially order-closed subalgebra of A is a -subalgebra in the sense of 233A.
Accordingly I will normally use the phrases -subalgebra, -ideal for sequentially order-closed subal-
gebras and ideals of Boolean algebras.
313F Order-closures and generated sets (a) It is an immediate consequence of the denitions that
(i) if o is any non-empty family of subalgebras of a Boolean algebra A, then

o is a subalgebra of A;
(ii) if T is any non-empty family of order-closed subsets of a partially ordered set P, then

T is an
order-closed subset of P;
(iii) if T is any non-empty family of sequentially order-closed subsets of a partially ordered set P, then

T is a sequentially order-closed subset of P.

(b) Consequently, given any Boolean algebra A and a subset B of A, we have a smallest subalgebra B of
A including B, being the intersection of all the subalgebras of A which include B; a smallest -subalgebra
B

of A including B, being the intersection of all the -subalgebras of A which include B; and a smallest
order-closed subalgebra B

of A including B, being the intersection of all the order-closed subalgebras of

A which include B. We call B, B

and B

the subalgebra, -subalgebra and order-closed subalgebra

generated by B. (I shall return to this in 331E.)
(c) If A is a Boolean algebra and B any subalgebra of A, then the smallest order-closed subset B of A
which includes B is again a subalgebra of A (so is the order-closed subalgebra of A generated by B). PPP (i)
The set b : 1 \ b B is order-closed (use 313A) and includes B, so includes B; thus 1 \ b B for every
b B. (ii) If c B, the set b : b c B is order-closed (use 313Bb) and includes B, so includes B; thus
b c B whenever b B and c B. (iii) If c B, the set b : b c B is order-closed and includes B
(by (ii)), so includes B; thus b c B whenever b, c B. (iv) By 312B, B is a subalgebra of A. QQQ
313G This is a convenient moment at which to spell out an abstract version of the Monotone Class
Theorem (136B).
Lemma Let A be a Boolean algebra.
(a) Suppose that 1 I A A and that
a b I for all a, b I,
b \ a A whenever a, b A and a b.
Then A includes the subalgebra of A generated by I.
(b) If moreover sup
nN
a
n
A for every non-decreasing sequence a
n
)
nN
in A with a supremum in A,
then A includes the -subalgebra of A generated by I.
(c) And if supC A whenever C A is an upwards-directed set with a supremum in A, then A includes
the order-closed subalgebra of A generated by I.
proof (a)(i) Let P be the family of all sets J such that I J A and a b J for all a, b J. Then
I P and if Q P is upwards-directed and not empty,

Q P. By Zorns Lemma, P has a maximal
element B.
(ii) Now
B = c : c A, c b A for every b B.
PPP If c B, then of course c b B A for every b B, because B P. If c A B, consider
J = B c b : b B.
Then c = c 1 J so J properly includes B and cannot belong to P. On the other hand, if b
1
, b
2
B,
b
1
b
2
B J, (c b
1
) b
2
= b
1
(c b
2
) = (c b
1
) (c b
2
) = c (b
1
b
2
) J,
so c
1
c
2
J for all c
1
, c
2
J; and of course I B J. So J cannot be a subset of A, and there must be
a b B such that c b / A. QQQ
34 Boolean algebras 313G
(iii) Consequently c \ b B whenever b, c B and b c. PPP If a B, then b a, c a B A and
b a c a, so
(c \ b) a = (c a) \ (b a) A
by the hypothesis on A. By (ii), c b B. QQQ
(iv) It follows that B is a subalgebra of A. PPP If b B, then
b 1 I B,
so 1 \ b B. If a, b B, then
a b = 1 \ ((1 \ a) (1 \ b)) B.
0 = 1 \ 1 B, so that the conditions of 312B(ii) are satised. QQQ
Now the subalgebra of A generated by I is included in B and therefore in A, as required.
(b) Now suppose that sup
nN
a
n
belongs to A whenever a
n
)
nN
is a non-decreasing sequence in A with
a supremum in A. Then B, as dened in part (a) of the proof, is a -subalgebra of A. PPP Let b
n
)
nN
be a
non-decreasing sequence in B with a supremum c in A. Then for any b B, b
n
b)
nN
is a non-decreasing
sequence in A with a supremum c b in A (313Ba). So c b A. As b is arbitrary, c B, by the criterion
in (a-ii) above. As b
n
)
nN
is arbitrary, B is a -subalgebra, by 313Ea. QQQ
Accordingly the -subalgebra of A generated by I is included in B and therefore in A.
(c) Finally, if supC A whenever C is a non-empty upwards-directed subset of A with a least upper
bound in A, B is order-closed. PPP Let C B be a non-empty upwards-directed set with a supremum c in
A. Then for any b B, c b : c C is a non-empty upwards-directed set in A with supremum c b in A.
So c b A. As b is arbitrary, c B. As C is arbitrary, B is order-closed in A. QQQ
Accordingly the order-closed subalgebra of A generated by I is included in B and therefore in A.
313H Denitions It is worth distinguishing various types of supremum- and inmum-preserving func-
tion. Once again, I do this in almost the widest possible context. Let P and Q be two partially ordered sets,
and : P Q an order-preserving function, that is, a function such that (p) (q) in Q whenever
p q in P.
(a) I say that is order-continuous if (i) (supA) = sup
pA
(p) whenever A is a non-empty upwards-
directed subset of P and supA is dened in P (ii) (inf A) = inf
pA
(p) whenever A is a non-empty
downwards-directed subset of P and inf A is dened in P.
(b) I say that is sequentially order-continuous or -order-continuous if (i) (p) = sup
nN
(p
n
)
whenever p
n
)
nN
is a non-decreasing sequence in P and p = sup
nN
p
n
in P (ii) (p) = inf
nN
(p
n
)
whenever p
n
)
nN
is a non-increasing sequence in P and p = inf
nN
p
n
in P.
Remark You may feel that one of the equivalent formulations in Proposition 313Lb gives a clearer idea of
what is really being demanded of in the ordinary cases we shall be looking at.
313I Proposition Let P, Q and R be partially ordered sets, and : P Q, : Q R order-preserving
functions.
(a) : P R is order-preserving.
(b) If and are order-continuous, so is .
(c) If and are sequentially order-continuous, so is .
(d) is order-continuous i
1
[B] is order-closed for every order-closed B Q.
proof (a)-(c) I think the only point that needs remarking is that if A P is upwards-directed, then
[A] Q is upwards-directed, because is order-preserving. So if sup A is dened in P and , are
order-continuous, we shall have
((supA)) = (sup[A]) = sup[[A]].
(d)(i) Suppose that is order-continuous and that B Q is order-closed. Let A
1
[B] be a non-
empty upwards-directed set with supremum p P. Then [A] B is non-empty and upwards-directed,
because is order-preserving, and (p) = sup [A] because is order-continuous. Because B is order-closed,
313L Order-continuous homomorphisms 35
(p) B and p
1
[B]. Similarly, if A
1
[B] is non-empty and downwards-directed, and inf A is
dened in P, then (inf A) = inf [A] B and inf A
1
[B]. Thus
1
[B] is order-closed; as B is
arbitrary, satises the condition.
(ii) Now suppose that
1
[B] is order-closed in P whenever B Q is order-closed in Q. Let A P
be a non-empty upwards-directed subset of P with a supremum p P. Then (p) is an upper bound of
[A]. Let q be any upper bound of [A] in Q. Consider B = r : r q; then B Q is upwards-directed
and order-closed, so
1
[B] is order-closed. Also A
1
[B] is non-empty and upwards-directed and has
supremum p, so p
1
[B] and (p) B, that is, (p) q. As q is arbitrary, (p) = sup [A]. Similarly,
(inf A) = inf [A] whenever A P is non-empty, downwards-directed and has an inmum in P; so is
order-continuous.
313J It is useful to introduce here the following notion.
Denition Let A be a Boolean algebra. A set D A is order-dense if for every non-zero a A there is
a non-zero d D such that d a.
Remark Many authors use the simple word dense where I have insisted on the phrase order-dense. In
the work of this treatise it will be important to distinguish clearly between this concept of dense set and
the topological concept (2A3U); typically, in those contexts in which both appear, an order-dense set can
be in some sense much smaller than a topologically dense set.
313K Lemma If A is a Boolean algebra and D A is order-dense, then for any a A there is a
disjoint C D such that supC = a; in particular, a = supd : d D, d a and there is a partition of
unity C D.
proof Set D
a
= d : d D, d a. Applying Zorns lemma to the family ( of disjoint sets C D
a
, we
have a maximal C (. Now if b A and b , a, there is a d D such that 0 ,= d a \ b. Because C is
maximal, there must be a c C such that c d ,= 0, so that c , b. Turning this round, any upper bound of
C must include a, so that a = supC. It follows at once that a = sup D
a
.
Taking a = 1 we obtain a partition of unity included in D.
313L Proposition Let A and B be Boolean algebras and : A B a Boolean homomorphism.
(a) is order-preserving.
(b) The following are equiveridical:
(i) is order-continuous;
(ii) whenever A A is non-empty and downwards-directed and inf A = 0 in A, then inf [A] = 0 in B;
(iii) whenever A A is non-empty and upwards-directed and supA = 1 in A, then sup [A] = 1 in B;
(iv) whenever A A and sup A is dened in A, then (supA) = sup [A] in B;
(v) whenever A A and inf A is dened in A, then (inf A) = inf [A] in B;
(vi) whenever C A is a partition of unity, then [C] is a partition of unity in B.
(c) The following are equiveridical:
(i) is sequentially order-continuous;
(ii) whenever a
n
)
nN
is a non-increasing sequence in A and inf
nN
a
n
= 0 in A, then inf
nN
a
n
= 0
in B;
(iii) whenever A A is countable and supA is dened in A, then (supA) = sup [A] in B;
(iv) whenever A A is countable and inf A is dened in A, then (inf A) = inf [A] in B;
(v) whenever C A is a countable partition of unity, then [C] is a partition of unity in B.
proof (a) This is 312I.
(b)(i)(ii) is trivial, as 0 = 0.
(ii)(iv) Assume (ii), and let A be any subset of A such that c = supA is dened in A. If A = ,
then c = 0 and sup[A] = 0 = c. Otherwise, set
A
t
= a
0
. . . a
n
: a
0
, . . . , a
n
A, C = c a : a A
t
.
36 Boolean algebras 313L
Then A
t
is upwards-directed and has the same upper bounds as A, so c = sup A
t
and 0 = inf C, by 313Aa.
Also C is downwards-directed, so inf [C] = 0 in B. But now
[C] = c \ a : a A
t
= c \ b : b [A
t
],
[A
t
] = a
0
. . . a
n
: a
0
, . . . , a
n
A = b
0
. . . b
n
: b
0
, . . . , b
n
[A],
because is a Boolean homomorphism. Again using 313Aa and the fact that b c for every b [A
t
], we
get
c = sup [A
t
] = sup[A].
As A is arbitrary, (iv) is satised.
(iv)(v) If A A and c = inf A is dened in A, then 1 \ c = sup
aA
1 \ a, so
c = 1 \ (1 \ c) = 1 \ sup
aA
(1 \ a) = inf
aA
1 \ (1 \ a) = inf
aA
a.
(v)(ii) is trivial, because 0 = 0.
(iv)(iii) is similarly trivial.
(iii)(vi) Assume (iii), and let C be a partition of unity in A. Then C
t
= c
0
. . . c
n
: c
0
, . . . , c
n

C is upwards-directed and has supremum 1, so sup[C
t
] = 1. But (because is a Boolean homomorphism)
[C] and [C
t
] have the same upper bounds, so sup [C] = 1, as required.
(vi)(ii) Assume (vi), and let A A be a set with inmum 0. Set
D = d : d A, a A, d a = 0.
Then D is order-dense in A. PPP If e A\ 0, then there is an a A such that e , a, so that e \ a is a
non-zero member of D included in e. QQQ Consequently there is a partition of unity C D, by 313K. But
now if b is any lower bound for [A] in B, we must have b d = 0 for every d D, so c 1 \ b for every
c C, and 1 \ b = 1, b = 0. Thus inf [A] = 0. As A is arbitrary, (ii) is satised.
(v)&(iv)(i) is trivial.
(c) We can use nearly identical arguments, remembering only to interpolate the word countable from
time to time. I spell out the new version of (ii)(iii), even though it requires no more than an adaptation
of the language. Assume (ii), and let A be a countable subset of A with a supremum c A. If A = , then
c = 0 so c = 0 = sup[A]. Otherwise, let a
n
)
nN
be a sequence running over A; set a
t
n
= a
0
. . . a
n
and
c
n
= c \ a
t
n
for each n. Then a
t
n
)
nN
is non-decreasing, with supremum c, and c
n
)
nN
is non-increasing,
with inmum 0; so inf
nN
c
n
= 0 and
sup
nN
a
n
= sup
nN
a
t
n
= c.
For (v)(ii), however, a dierent idea is involved. Assume (v), and suppose that a
n
)
nN
is a non-
increasing sequence in A with inmum 0. Set c
0
= 1 \ a
0
, c
n
= a
n1
\ a
n
for n 1; then C = c
n
: n N
is a partition of unity in A (because if c c
n
= 0 for every n, then c a
n
for every n), so [C] is a partition
of unity in B. Now if b a
n
for every n, b c
n
for every n, so b = 0; thus inf
nN
a
n
= 0. As a
n
)
nN
is
arbitrary, (ii) is satised.
313M The following result is perfectly elementary, but it will save a moment later on to have it spelt
out.
Lemma Let A and B be Boolean algebras and : A B an order-continuous Boolean homomorphism.
(a) If D is an order-closed subalgebra of B, then
1
[D] is an order-closed subalgebra of A.
(b) If C is the order-closed subalgebra of A generated by C A, then the order-closed subalgebra D of
B generated by [C] includes [C].
(c) Now suppose that is surjective and that C A is such that the order-closed subalgebra of A
generated by C is A itself. Then the order-closed subalgebra of B generated by [C] is B.
proof (a) Setting C =
1
[D]: if a, a
t
C then (a b) = a b, (a .a
t
) = a .a
t
D, so a a
t
,
a .a
t
C; 1 = 1 D so 1 C; thus C is a subalgebra of A. By 313Id, C is order-closed.
313R Order-continuous homomorphisms 37
(b) By (a),
1
[D] is an order-closed subalgebra of A. It includes C so includes C, and [C] D.
(c) In the language of (b), we have C = A, so D must be B.
313N Denition The phrase regular embedding is sometimes used to mean an injective order-
continuous Boolean homomorphism; a subalgebra B of a Boolean algebra A is said to be regularly em-
bedded in A if the identity map from Bto A is order-continuous, that is, if whenever b Bis the supremum
(in B) of B B, then b is also the supremum in A of B; and similarly for inma. One important case is
when B is order-dense (313O); another is in 314G-314H below.
313O Proposition Let A be a Boolean algebra and B an order-dense subalgebra of A. Then B is
regularly embedded in A. In particular, if B B and c B then c = supB in B i c = supB in A.
proof I have to show that the identity homomorphism : B A is order-continuous. ??? Suppose, if
possible, otherwise. By 313L(b-ii), there is a non-empty set B B such that inf B = 0 in B but B = [B]
has a non-zero lower bound a A. In this case, however (because B is order-dense) there is a non-zero
d B with d a, in which case d is a non-zero lower bound for B in B. XXX
313P The most important use of these ideas to us concerns quotient algebras (313Q); I approach by
means of a supercially more general result.
Theorem Let A and B be Boolean algebras and : A B a Boolean homomorphism with kernel I.
(a)(i) If is order-continuous then I is order-closed.
(ii) If [A] is regularly embedded in B and I is order-closed then is order-continuous.
(b)(i) If is sequentially order-continuous then I is a -ideal.
(ii) If [A] is regularly embedded in B and I is a -ideal then is sequentially order-continuous.
proof (a)(i) If A I is upwards-directed and has a supremum c A, then c = sup [A] = 0, so c I. As
remarked in 313Eb, this shows that I is order-closed.
(ii) We are supposing that the identity map from [A] to B is order-continuous, so it will be enough to
show that is order-continuous when regarded as a map from A to [A]. Suppose that A A is non-empty
and downwards-directed and that inf A = 0. ??? Suppose, if possible, that 0 is not the greatest lower bound
of [A] in [A]. Then there is a c A such that 0 ,= c a for every a A. Now
(c \ a) = c \ a = 0
for every a A, so c \ a I for every a A. The set C = c \ a : a A is upwards-directed and has
supremum c; because I is order-closed, c = supC I, and c = 0, contradicting the specication of c. XXX
Thus inf [A] = 0 in either [A] or B. As A is arbitrary, is order-continuous, by the criterion (ii) of 313Lb.
(b) Argue in the same way, replacing each set A by a sequence.
313Q Corollary Let A be a Boolean algebra and I an ideal of A; write for the canonical map from
A to A/I.
(a) is order-continuous i I is order-closed.
(b) is sequentially order-continuous i I is a -ideal.
proof [A] = A/I is surely regularly embedded in A/I.
313R For order-continuous homomorphisms, at least, there is an elegant characterization in terms of
Stone spaces.
Proposition Let A and B be Boolean algebras, and : A B a Boolean homomorphism. Let Z and W
be their Stone spaces, and : W Z the corresponding continuous function (312P). Then the following
are equiveridical:
(i) is order-continuous;
(ii)
1
[M] is nowhere dense in W for every nowhere dense set M Z;
(iii) int [H] ,= for every non-empty open set H W.
38 Boolean algebras 313R
proof (a)(i)(iii) Suppose that is order-continuous. ??? Suppose, if possible, that H W is a non-empty
open set and int [H] = . Let b B 0 be such that

b H. Then [

b] has empty interior; but also it

is a closed set, so its complement is dense. Set A = a : a A, a [

b] = . Then

aA
a = Z [

b] is a
dense open set, so supA = 1 in A (313Ca). Because is order-continuous, sup[A] = 1 in B (313L(b-iii)),
and there is an a A such that a b ,= 0. But this means that

b
1
[a] ,= and [

b] a ,= , contrary
to the denition of A. XXX
Thus there is no such set H, and (iii) is true.
(b)(iii)(ii) Now assume (iii). If M Z is nowhere dense, set N =
1
[M], so that N W is a closed
set. If H = int N, then int [H] int M = , so (iii) tells us that H is empty; thus N and
1
[M] are
nowhere dense, as required by (ii).
(c)(ii)(i) Assume (ii), and let A A be a non-empty set such that inf A = 0 in A. Then M =

aA
a
has empty interior in Z (313Cb), so (being closed) is nowhere dense, and
1
[M] is also nowhere dense. If
b B 0, then

b ,
1
[M] =

aA

1
[a] =

aA
a,
so b is not a lower bound for [A]. This shows that inf [A] = 0 in B. As A is arbitrary, is order-continuous
(313L(b-ii)).
313X Basic exercises (a) Use 313C to give alternative proofs of 313A and 313B.
(b) Let P be a partially ordered set. Show that there is a topology on P for which the closed sets are
just the order-closed sets.
(c) Let P be a partially ordered set, Q P an order-closed set, and R a subset of Q which is order-closed
in Q when Q is given the partial ordering induced by that of P. Show that R is order-closed in P.
>>>(d) Let A be a Boolean algebra. Suppose that 1 I A and that a b I for all a, b I. (i) Let
B be the intersection of all those subsets A of A such that I A and b \ a A whenever a, b A and
a b. Show that B is a subalgebra of A. (ii) Let B

be the intersection of all those subsets A of A such

that I A, b \ a A whenever a, b A and a b and sup
nN
b
n
A whenever b
n
)
nN
is a non-decreasing
sequence in A with a supremum in A. Show that B

is a -subalgebra of A. (iii) Let B

be the intersection
of all those subsets A of A such that I A, b \ a A whenever a, b A and a b and sup B A whenever
B is a non-empty upwards-directed subset of A with a supremum in A. Show that B

is an order-closed
subalgebra of A. (iv) Hence give a proof of 313G not relying on Zorns Lemma or any other use of the axiom
of choice.
(e) Let A be a Boolean algebra, and Ba subalgebra of A. Let B

be the smallest sequentially order-closed

subset of A including B. Show that B

is a subalgebra of A.
>>>(f ) Let X be a set, and / a subset of TX. Show that / is an order-closed subalgebra of TX i it is of
the form f
1
[F] : F Y for some set Y , function f : X Y .
(g) Let P and Q be partially ordered sets, and : P Q an order-preserving function. Show that is
sequentially order-continuous i
1
[C] is sequentially order-closed in A for every sequentially order-closed
C B.
(h) For partially ordered sets P and Q, let us call a function : P Q monotonic if it is either
order-preserving or order-reversing. State and prove denitions and results corresponding to 313H, 313I and
313Xg for general monotonic functions.
>>>(i) Let A be a Boolean algebra. Show that the operations (a, b) a b and (a, b) a b are order-
continuous operations from AA to A, if we give AA the product partial order, saying that (a, b) (a
t
, b
t
)
i a a
t
and b b
t
.
313Yf Order-continuous homomorphisms 39
(j) Let A be a Boolean algebra. Show that if a subalgebra of A is order-dense then it is dense in the
topology of 313Xb.
>>>(k) Let A be a Boolean algebra and A A any disjoint set. Show that there is a partition of unity in
A including A.
>>>(l) Let A, B be Boolean algebras and
1
,
2
: A B two order-continuous Boolean homomorphisms.
Show that a :
1
a =
2
a is an order-closed subalgebra of A.
(m) Let A and B be Boolean algebras and
1
,
2
: A B two Boolean homomorphisms. Suppose that

1
and
2
agree on some order-dense subset of A, and that one of them is order-continuous. Show that they
are equal. (Hint: if
1
is order-continuous,
2
a
1
a for every a.)
(n) Let A and B be Boolean algebras, A
0
an order-dense subalgebra of A, and : A B a Boolean
homomorphism. Show that is order-continuous i A
0
: A
0
B is order-continuous.
>>>(o) Let A be a Boolean algebra. For A A set A

= b : a b = 0 a A. (i) Show that A

is an
order-closed ideal of A. (ii) Show that a set A A is an order-closed ideal of A i A = A

. (iii) Show
that if I A is an order-closed ideal then a

: a I

is an order-dense ideal in the quotient algebra A/I.

(p) Let A and B be Boolean algebras, with Stone spaces Z and W; let : A B be a Boolean
homomorphism, and : W Z the corresponding continuous function. Show that the following are
equiveridical: (i) is order-continuous; (ii) int
1
[F] =
1
[int F] for every closed F Z (iii)
1
[G] =

1
[G] for every open G Z.
313Y Further exercises (a) Prove 313A-313C for general Boolean rings.
(b) Let A and B be Boolean algebras, with Stone spaces Z and W, and : A B a Boolean homo-
morphism, with associated continuous function : W Z. Show that is sequentially order-continuous
i
1
[M] is nowhere dense for every nowhere dense zero set M Z.
(c) Let P be any partially ordered set, and let T be the topology of 313Xb. (i) Show that a sequence
p
n
)
nN
in P is T-convergent to p P i every subsequence of p
n
)
nN
has a monotonic sub-subsequence
with supremum or inmum equal to p. (ii) Show that a subset A of P is sequentially order-closed, in the
sense of 313Db, i the T-limit of any T-convergent sequence in A belongs to A. (iii) Suppose that A is an
upwards-directed subset of P with supremum p
0
P. For a A set F
a
= p : a p A, and let T
be the lter on P generated by F
a
: a A. Show that T p
0
for T. (iv) Show that if Q is another
partially ordered set, endowed with a topology S in the same way, then a monotonic function : P Q is
order-continuous i it is continuous for the topologies T and S, and is sequentially order-continuous i it is
sequentially continuous for these topologies.
(d) Let U be a Banach lattice (242G, 354Ab). Show that its norm is order-continuous in the sense of
242Yc (or 354Dc) i its restriction to u : u 0 is order-continuous in the sense of 313Ha.
(e) Let A and B be Boolean algebras with Stone spaces Z and W respectively, : A B a Boolean
homomorphism and : W Z the corresponding continuous function. Show that [A] is order-dense in B
i is irreducible, that is, [F] ,= [W] for any proper closed subset F of W.
(f ) Let A and B be Boolean algebras with Stone spaces Z and W respectively, : A B a Boolean
homomorphism and : W Z the corresponding continuous function. Show that the following are
equiveridical: (i) is injective and order-continuous; (ii) for M Z, M is nowhere dense i
1
[M] is
nowhere dense.
40 Boolean algebras 313 Notes
313 Notes and comments I give elementary proofs of 313A-313B because I believe that they help to
exhibit the relevant aspects of the structure of Boolean algebras; but various abbreviations are possible,
notably if we allow ourselves to use the Stone representation (313Xa). 313A and 313Ba-b can be expressed
by saying that the Boolean operations , and \ are (separately) order-continuous. Of course, \ is order-
reversing, rather than order-preserving, in the second variable; but the natural symmetry in the concept of
partial order means that the ideas behind 313H-313I can be applied equally well to order-reversing functions
(313Xh). In fact, and can be regarded as order-continuous functions on the product space (313Bc-d,
313Xi). Clearly 313Bc-d can be extended into forms valid for any nite sequence A
0
, . . . , A
n
of subsets of
A in place of A, B. But if we seek to go to innitely many subsets of A we nd ourselves saying something
new; see 316G-316J below.
Proposition 313C, and its companions 313R, 313Xp and 313Yb, are worth studying not only as a useful
technique, but also in order to understand the dierence between supA, where A is a set in a Boolean
algebra, and

/, where / is a family of sets. Somehow sup A can be larger, and inf A smaller, than ones
rst intuition might suggest, corresponding to the fact that not every subset of the Stone space corresponds
to an element of the Boolean algebra.
I should like to use the words order-closed and sequentially order-closed to mean closed, or sequentially
closed, for some more or less canonical topology. The diculty is that while a great many topologies can be
dened from a partial order (one is described in 313Xb and 313Yc, and another in 367Yc), none of them has
such pre-eminence that it can be called the order-topology. Accordingly there is a degree of arbitrariness
in the language I use here. Nevertheless (sequentially) order-closed subalgebras and ideals are of such
importance that they seem to deserve a concise denotation. The same remarks apply to (sequential) order-
continuity. Concerning the term order-dense in 313J, this has little to do with density in any topological
sense, but the word dense, at least, is established in this context.
With all these denitions, there is a good deal of scope for possible interrelations. The most important
to us is 313Q, which will be used repeatedly (typically, with A an algebra of sets), but I think it is worth
having the expanded version in 313P available.
I take the opportunity to present an abstract form of an important lemma on -algebras generated by
families closed under (136B, 313Gb). This time round I use the Zorns Lemma argument in the text
and suggest the alternative, elementary method in the exercises (313Xd). The two methods are opposing
extremes in the sense that the Zorns Lemma argument looks for maximal subalgebras included in A (which
are not unique, and have to be picked out using the axiom of choice) and the other approach seeks minimal
subalgebras including I (which are uniquely dened, and can be described without the axiom of choice).
Note that the concept of order-closed algebra of sets is not particularly useful; there are too few order-
closed subalgebras of TX and they are of too simple a form (313Xf). It is in abstract Boolean algebras that
the idea becomes important. In the most important partially ordered sets of measure theory, the sequentially
order-closed sets are the same as the order-closed sets (see, for instance, 316Fb below), and most of the
important order-closed subalgebras dealt with in this chapter can be thought of as -subalgebras which are
order-closed because they happen to lie in the right kind of algebra.
314 Order-completeness
The results of 313 are valid in all Boolean algebras, but of course are of most value when many suprema
and inma exist. I now set out the most useful denitions which guarantee the existence of suprema and
inma (314A) and work through their elementary relationships with the concepts introduced so far (314C-
314J). I then embark on the principal theorems concerning order-complete Boolean algebras: the extension
theorem for homomorphisms to a Dedekind complete algebra (314K), the Loomis-Sikorski representation of
a Dedekind -complete algebra as a quotient of a -algebra of sets (314M), the characterization of Dedekind
complete algebras in terms of their Stone spaces (314S), and the idea of Dedekind completion of a Boolean
algebra (314T-314U). On the way I describe regular open algebras (314O-314Q).
314Be Order-completeness 41
314A Denitions Let P be a partially ordered set.
(a) P is Dedekind complete, or order-complete, or conditionally complete if every non-empty
subset of P with an upper bound has a least upper bound.
(b) P is Dedekind -complete, or -order-complete, if (i) every countable non-empty subset of P
with an upper bound has a least upper bound (ii) every countable non-empty subset of P with a lower
bound has a greatest lower bound.
314B Remarks (a) I give these denitions in the widest possible generality because they are in fact
of great interest for general partially ordered sets, even though for the moment we shall be concerned only
with Boolean algebras. Indeed I have already presented the same idea in the context of Riesz spaces (241F).
(b) You will observe that the denition in (a) of 314A is asymmetric, unlike that in (b). This is because
the inverted form of the denition is equivalent to that given; that is, P is Dedekind complete (on the
denition 314Aa) i every non-empty subset of P with a lower bound has a greatest lower bound. PPP (i)
Suppose that P is Dedekind complete, and that B P is non-empty and bounded below. Let A be the set
of lower bounds for B. Then A has at least one upper bound (since any member of B is an upper bound
for A) and is not empty; so a
0
= sup A is dened. Now if b B, b is an upper bound for A, so a
0
b; thus
a
0
A and must be the greatest member of A, that is, the greatest lower bound of B. (ii) Similarly, if every
non-empty subset of P with a lower bound has a greatest lower bound, P is Dedekind complete. QQQ
(c) In the special case of Boolean algebras, we do not need both halves of the denition 314Ab; in fact
we have, for any Boolean algebra A,
A is Dedekind -complete
every non-empty countable subset of A has a least upper bound
every non-empty countable subset of A has a greatest lower bound.
PPP Because A has a least element 0 and a greatest element 1, every subset of A has upper and lower
bounds; so the two one-sided conditions together are equivalent to saying that A is Dedekind -complete.
I therefore have to show that they are equiveridical. Now if A A is a non-empty countable set, so is
B = 1 \ a : a A, and
inf A = 1 \ supB, sup A = 1 \ inf B
whenever the right-hand-sides are dened (313A). So if the existence of a supremum (resp. inmum) of B
is guaranteed, so is the existence of an inmum (resp. supremum) of A. QQQ
The real point here is of course that (A, ) is isomorphic to (A, ).
(d) Most specialists in Boolean algebra speak of complete, or -complete, Boolean algebras. I prefer
the longer phrases Dedekind complete and Dedekind -complete because we shall be studying metrics on
Boolean algebras and shall need the notion of metric completeness as well as that of order-completeness.
(e) I have had to make some rather arbitrary choices in the denition here. The principal examples of
partially ordered set to which we shall apply these denitions are Boolean algebras and Riesz spaces, which
are all lattices. Consequently it is not possible to distinguish in these contexts between the property of
Dedekind completeness, as dened above, and the weaker property, which we might call monotone order-
completeness,
(i) whenever A P is non-empty, upwards-directed and bounded above then A has a least
upper bound in P (ii) whenever A P is non-empty, downwards-directed and bounded below
then A has a greatest lower bound in P.
(See 314Xa below. Monotone order-completeness is the property involved in 314Ya, for instance.) Never-
theless I am prepared to say, on the basis of my own experience of working with other partially ordered sets,
that Dedekind completeness, as I have dened it, is at least of sucient importance to deserve a name.
Note that it does not imply that P is a lattice, since it allows two elements of P to have no common upper
bound.
42 Boolean algebras 314Bf
(f ) The phrase complete lattice is sometimes used to mean a Dedekind complete lattice with greatest
and least elements; equivalently, a Dedekind complete partially ordered set with greatest and least elements.
Thus a Dedekind complete Boolean algebra is a complete lattice in this sense, but R is not.
(g) The most important Dedekind complete Boolean algebras (at least from the point of view of measure
theory) are the measure algebras of the next chapter. I shall not pause here to give other examples, but
will proceed directly with the general theory.
314C Proposition Let A be a Dedekind -complete Boolean algebra and I a -ideal of A. Then the
quotient Boolean algebra A/I is Dedekind -complete.
proof I use the description in 314Bc. Let B A/I be a non-empty countable set. For each u B, choose
an a
u
A such that u = a

u
= a
u
+I. Then c = sup
uB
a
u
is dened in A; consider v = c

in A/I. Because
the map a a

is sequentially order-continuous (313Qb), v = supB. As B is arbitrary, A/I is Dedekind

-complete.
314D Corollary Let X be a set, a -algebra of subsets of X, and 1 a -ideal of subsets of X. Then
1 is a -ideal of the Boolean algebra , and / 1 is Dedekind -complete.
proof Of course is Dedekind -complete, because if E
n
)
nN
is any sequence in then

nN
E
n
is
the least upper bound of E
n
: n N in . It is also easy to see that 1 is a -ideal of , since
F

nN
E
n
1 whenever F and E
n
)
nN
is a sequence in 1. So 314C gives the result.
314E Proposition Let A be a Boolean algebra.
(a) If A is Dedekind complete, then all its order-closed subalgebras and principal ideals are Dedekind
complete.
(b) If A is Dedekind -complete, then all its -subalgebras and principal ideals are Dedekind -complete.
proof All we need to note is that if C is either an order-closed subalgebra or a principal ideal of A, and
B C is such that b = sup B is dened in A, then b C (see 313E(a-i-)), so b is still the supremum of B
in C; while the same is true if C is a -subalgebra and B C is countable, using 313E(a-ii-).
314F I spell out some further connexions between the concepts order-closed set, order-continuous
function and Dedekind complete Boolean algebra which are elementary without being quite transparent.
Proposition Let A and B be Boolean algebras and : A B a Boolean homomorphism.
(a)(i) If A is Dedekind complete and is order-continuous, then [A] is order-closed in B.
(ii) If B is Dedekind complete and is injective and [A] is order-closed then is order-continuous.
(b)(i) If A is Dedekind -complete and is sequentially order-continuous, then [A] is a -subalgebra of
B.
(ii) If Bis Dedekind -complete and is injective and [A] is a -subalgebra of Bthen is sequentially
order-continuous.
proof (a)(i) If B [A], then a
0
= sup(
1
[B]) is dened in A; now
a
0
= sup([
1
[B]]) = supB
in B (313L(b-iv)), and of course a
0
[A]. By 313E(a-i-) again, this is enough to show that [A] is
order-closed in B.
(ii) Suppose that A A and inf A = 0 in A. Then [A] has an inmum b
0
in B, which belongs to
[A] because [A] is an order-closed subalgebra of B (313E(a-i-
t
)). Now if a
0
A is such that a
0
= b
0
,
we have
(a a
0
) = a a
0
= a
for every a A, so (because is injective) a a
0
= a
0
and a
0
a for every a A. But this means that
a
0
= 0 and b
0
= 0 = 0. As A is arbitrary, is order-continuous (313L(b-ii)).
(b) Use the same arguments, but with sequences in place of the sets B, A above.
314J Order-completeness 43
314G Corollary (a) If A is a Dedekind complete Boolean algebra and B is an order-closed subalgebra
of A, then B is regularly embedded in A (denition: 313N).
(b) Let A be a Dedekind complete Boolean algebra, B a Boolean algebra and : A B an order-
continuous Boolean homomorphism. If C A and C is the order-closed subalgebra of A generated by C,
then [C] is the order-closed subalgebra of B generated by [C].
proof (a) Apply 314F(a-ii) to the identity map from B to A.
(b) Let D be the order-closed subalgebra of B generated by [C]. By 313Mb, [C] D. On the other
hand, the identity homomorphism : C A is order-continuous, by (a), so : C B is order-continuous,
and [C] = [C] is order-closed in B, by 314F(a-i). But since [C] is surely included in [C], D is also
included in [C]. Accordingly [C] = D, as claimed.
314H Corollary Let A be a Boolean algebra and B a subalgebra of A.
(a) If A is Dedekind complete, then B is order-closed i it is Dedekind complete in itself and is regularly
embedded in A.
(b) If A is Dedekind -complete, then B is a -subalgebra i it is Dedekind -complete in itself and the
identity map from B to A is sequentially order-continuous.
proof Put 314E and 314F together.
314I Corollary (a) If A is a Boolean algebra and Bis an order-dense subalgebra of A which is Dedekind
complete in itself, then B = A.
(b) If A is a Dedekind complete Boolean algebra, B is a Boolean algebra, : A B is an injective
Boolean homomorphism and [A] is order-dense in B, then is an isomorphism.
proof (a) Being order-dense, B is regularly embedded in A (313O), so this is a special case of 314Ha.
(b) Because [A] is order-dense, it is regularly embedded in B; also, the kernel of is 0, which is
surely order-closed in A, so 313P(a-ii) tells us that is order-continuous. By 314F(a-i), [A] is order-closed
in B; being order-dense, it must be the whole of B (313K). Thus is surjective; being injective, it is an
isomorphism.
314J When we come to applications of the extension procedure in 312N, the following will sometimes
be needed.
Lemma Let A be a Boolean algebra and A
0
a subalgebra of A. Take any c A, and set
A
1
= (a c) (b \ c) : a, b A
0
,
the subalgebra of A generated by A
0
c (312M).
(a) Suppose that A is Dedekind complete. If A
0
is order-closed in A, so is A
1
.
(b) Suppose that A is Dedekind -complete. If A
0
is a -subalgebra of A, so is A
1
.
proof (a) Let D be any subset of A
1
. Set
E = e : e A, there is some d D such that e d,
A = a : a A
0
, a c E, B = b : b A
0
, b \ c E.
Because A is Dedekind complete, a

= sup A and b

= sup B are dened in A; because A

0
is order-closed,
both belong to A
0
, so d

= (a

c) (b

\ c) belongs to A
1
.
Now if d D, it is expressible as (a c) (b \ c) for some a, b A
0
; since a A and b B, we have
a a

and b b

, so d d

. Thus d

is an upper bound for D. On the other hand, if d

t
is any other upper
bound for D in A, it is also an upper bound for E, so we must have
a

c = sup
aA
a c d
t
, b

\ c = sup
bB
b \ c d
t
,
and d

d
t
. Thus d

= supD. This shows that the supremum of any subset of A

1
belongs to A
1
, so that
A
1
is order-closed.
(b) The argument is the same, except that we replace D by a sequence d
n
)
nN
, and A, B by sequences
a
n
)
nN
, b
n
)
nN
in A
0
such that d
n
= (a
n
c) (b
n
\ c) for every n.
44 Boolean algebras 314K
314K Extension of homomorphisms The following is one of the most striking properties of Dedekind
complete Boolean algebras.
Theorem Let A be a Boolean algebra and B a Dedekind complete Boolean algebra. Let A
0
be a Boolean
subalgebra of A and
0
: A
0
B a Boolean homomorphism. Then there is a Boolean homomorphism

1
: A B extending
0
.
proof (a) Let P be the set of all Boolean homomorphisms such that dom is a Boolean subalgebra of
A including A
0
and extends
0
. Identify each member of P with its graph, which is a subset of A B,
and order P by inclusion, so that means just that extends . Then any non-empty totally ordered
subset Q of P has an upper bound in P. PPP Let

be the simple union of these graphs. (i) If (a, b) and

(a, b
t
) both belong to

, then there are ,

t
Q such that a = b,
t
a = b
t
; now either
t
or
t
;
in either case, =
t
Q, so that
b = a = a =
t
a = b
t
.
This shows that

is a function. (ii) Because Q ,= ,

dom
0
dom dom

for some Q; thus

extends
0
(and, in particular, 0 dom

). (iii) Now suppose that a, a

t
dom(

).
Then there are ,
t
Q such that a dom, a
t
dom
t
; once again, =
t
Q, so that a, a
t
dom,
and
a a
t
dom dom

, 1 \ a dom dom

(a a
t
) = (a a
t
) = a a
t
=

a
t
,

(1 \ a) = (1 \ a) = 1 \ a = 1 \

a.
(iv) This shows that dom

is a subalgebra of A and that

is a Boolean homomorphism, that is, that

P; and of course

is an upper bound for Q in P. QQQ

(b) By Zorns Lemma, P has a maximal element
1
say.
??? Suppose, if possible, that A
1
= dom
1
is not the whole of A; take c A A
1
. Set A = a : a
A
1
, a c. Because B is Dedekind complete, d = sup
1
[A] is dened in B. If a
t
A and c a
t
, then of
course a a
t
and
1
a
1
a
t
whenever a A, so that
1
a
t
is an upper bound for
1
[A], and d
1
a
t
.
But this means that there is an extension of
1
to a Boolean homomorphism on the Boolean subalgebra
of A generated by A
1
c (312N). And this must be a member of P properly extending
1
, which is
supposed to be maximal. XXX
Thus dom
1
= A and
1
is an extension of
0
to A, as required.
314L The Loomis-Sikorski representation of a Dedekind -complete Boolean algebra The
construction in 314D is not only the commonest way in which new Dedekind -complete Boolean algebras
appear, but is adequate to describe them all. I start with an elementary general fact.
Lemma Let X be any topological space, and write / for the family of meager subsets of X. Then / is a
-ideal of subsets of X.
proof The point is that if A X is nowhere dense, so is every subset of A; this is obvious, since if B A
then B A so int B int A = . So if B A /, let A
n
)
nN
be a sequence of nowhere dense sets with
union A; then B A
n
)
nN
is a sequence of nowhere dense sets with union B, so B /. If A
n
)
nN
is a
sequence in / with union A, then for each n we may choose a sequence A
nm
)
mN
of nowhere dense sets
with union A
n
; then the countable family A
nm
)
n,mN
may be re-indexed as a sequence of nowhere dense
sets with union A, so A /. Finally, is nowhere dense, so belongs to /.
314M Theorem Let A be a Dedekind -complete Boolean algebra, and Z its Stone space. Let c be
the algebra of open-and-closed subsets of Z, and / the -ideal of meager subsets of Z. Then = EA :
E c, A / is a -algebra of subsets of Z, / is a -ideal of , and A is isomorphic, as Boolean algebra,
to //.
314P Order-completeness 45
proof (a) I start by showing that is a -algebra. PPP Of course = . If F , express it as
EA where E c, A /; then Z F = (Z E)A .
If F
n
)
nN
is a sequence in , express each F
n
as E
n
A
n
, where E
n
c and A
n
/. Now each E
n
is expressible as a
n
, where a
n
A. Because A is Dedekind -complete, a = sup
nN
a
n
is dened in A. Set
E = a c. By 313Ca, E =

nN
E
n
, so the closed set E

nN
E
n
has empty interior and is nowhere
dense. Accordingly, setting A = E

nN
F
n
, we have
A (E

nN
E
n
)

nN
A
n
/,
so that

nN
F
n
= EA . Thus is closed under countable unions and is a -algebra. QQQ
Evidently / , because c.
(b) For each F , there is exactly one E c such that FE /. PPP There is surely some E c
such that F is expressible as EA where A /, so that FE = A /. If E
t
is any other member of
c, then E
t
E is a non-empty open set in X, while E
t
E A (FE
t
); by Baires theorem for compact
Hausdor spaces (3A3G), A (FE
t
) / / and FE
t
/ /. Thus E is unique. QQQ
(c) Consequently the maps E E

: c // and a a

: A // are bijections. But since they

are also Boolean homomorphisms, they are isomorphisms, and A

= //, as claimed.
314N Corollary A Boolean algebra A is Dedekind -complete i it is isomorphic to a quotient /1
where is a -algebra of sets and 1 is a -ideal of .
proof Put 314D and 314M together.
314O Regular open algebras For Boolean algebras which are Dedekind complete in the full sense,
there is another general method of representing them, which leads to further very interesting ideas.
Denition Let X be a topological space. A regular open set in X is an open set G X such that
G = int G.
Note that if F X is any closed set, then G = int F is a regular open set, because G G F so
G int G int F = G
and G = int G.
314P Theorem Let X be any topological space, and write G for the set of regular open sets in X.
Then G is a Dedekind complete Boolean algebra, with 1
G
= X and 0
G
= , and with Boolean operations
given by
G
G
H = G H, G.
G
H = int GH, G
G
H = int G H, G\
G
H = G H,
with Boolean ordering given by
G
G
H G H,
and with suprema and inma given by
supH = int

H, inf H = int

H = int

H
for all non-empty H G.
Remark I use the expressions

G

G
.
G
\
G

G
in case the distinction between

and
. \
is insuciently marked.
46 Boolean algebras 314P
proof I base the proof on the study of an auxiliary algebra of sets which involves some of the ideas already
used in 314M.
(a) Let 1 be the family of nowhere dense subsets of X. Then 1 is an ideal of subsets of X. PPP Of course
1. If A B 1 then int A int B = . If A, B 1 and G is a non-empty open set, then G A is
a non-empty open set and (G A) B is non-empty; accordingly G cannot be a subset of A B = A B.
This shows that int A B = , so that A B 1. QQQ
(b) For any set A X, write A for the boundary of A, that is, A int A. Set
= E : E X, E 1.
The is an algebra of subsets of X. PPP (i) = 1 so . (ii) If A, B X, then A B = A B,
while int(A B) int A int B; so (A B) A B. So if E, F , (E F) E F 1 and
E F . (iii) If A X, then
(X A) = X A int(X A) = (X int A) (X A) = A int A = A.
So if E , (X E) = E 1 and X E . QQQ
If A 1, then of course A = A 1, so A ; accordingly 1 is an ideal in the Boolean algebra , and
we can form the quotient /1.
It will be helpful to note that every open set belongs to , since if G is open then G = G G cannot
include any non-empty open set (since any open set meeting G must meet G).
(c) For each E , set V
E
= int E; then V
E
is the unique member of G such that EV
E
1. PPP (i)
Being the interior of a closed set, V
E
G. Since int E V
E
E, EV
E
E 1. (ii) If G G is such
that EG 1, then
G V
E
G V
E
(GE) (V
E
E) 1,
so G V
E
, being open, must be actually empty, and G V
E
; but this means that G int V
E
= V
E
.
Similarly, V
E
G and V
E
= G. This shows that V
E
is unique. QQQ
(d) It follows that the map G G

: G /1 is a bijection, and we have a Boolean algebra structure

on G dened by the Boolean algebra structure of /1. What this means is that for each of the binary
Boolean operations
G
, .
G
,
G
, \
G
and for G, H G we must have G
G
H = int G H, writing
G
for
the operation on the algebra G and for the corresponding operation on or TX.
(e) Before working through the identications, it will be helpful to observe that if H is any non-empty
subset of G, then int

H = int

H. PPP Set G = int

H. For every H H, G H so G int H = H; thus

G int

H int

H = G,
so G = int

H. QQQ Consequently int

H, being the interior of a closed set, belongs to G.

(f )(i) If G, H G then their intersection in the algebra G is
G
G
H = int G H = int(G H) = G H,
using (d) for the rst equality and (e) for the second.
(ii) Of course X G and X

= 1
/1
, so X = 1
G
.
(iii) If G G then its complement 1
G
\
G
G in G is
int X G = int(X G) = X G.
(iv) If G, H G, then the relative complement in G is
G\
G
H = G
G
(1
G
\
G
H) = G (X H) = G H = int(G H).
(v) If G, H G, then G
G
H = int G H and G.
G
H = int GH, by the remarks in (d).
(g) We must note that for G, H G,
G
G
H G
G
H = G G H = G G H;
*314R Order-completeness 47
that is, the ordering of the Boolean algebra G is just the partial ordering induced on G by the Boolean
ordering of TX or .
(h) If H is any non-empty subset of G, consider G
0
= int

H and G
1
= int

H.
G
0
= inf H in G. PPP By (e), G
0
G. Of course G
0
H for every H H, so G
0
is a lower bound for
H. If G is any lower bound for H in G, then G H for every H H, so G

H; but also G is open, so

G int

H = G
0
. Thus G
0
is the greatest lower bound for H. QQQ
G
1
= supH in G. PPP Being the interior of a closed set, G
1
G, and of course
H = int H int

H = G
1
for every H H, so G
1
is an upper bound for H in G. If G is any upper bound for H in G, then
G = int G int

H = G
1
;
thus G
1
is the least upper bound for H in G. QQQ
This shows that every non-empty H G has a supremum and an inmum in G; consequently G is
Dedekind complete, and the proof is nished.
314Q Remarks (a) G is called the regular open algebra of the topological space X.
(b) Note that the map E V
E
: G of part (c) of the proof above is a Boolean homomorphism, if G
is given its Boolean algebra structure. Its kernel is of course 1; the induced map E

V
E
: /1 G is
just the inverse of the isomorphism G G

: G /1.
*314R I interpolate a lemma corresponding to 313R.
Lemma Let X and Y be topological spaces, and f : X Y a continuous function such that f
1
[M]
is nowhere dense in X for every nowhere dense M Y . Then we have an order-continuous Boolean
homomorphism from the regular open algebra RO(Y ) of Y to the regular open algebra RO(X) of X
dened by setting H = int f
1
[H] for every H RO(Y ).
proof (a) By the remark in 314O, the formula for H always denes a member of RO(X); and of course
is order-preserving.
Observe that if H RO(Y ), then f
1
[H] is open, so f
1
[H] H. It will be convenient to note straight
away that if V Y is a dense open set then f
1
[V ] is dense in X. PPP M = Y V is nowhere dense, so
f
1
[M] is nowhere dense and its complement f
1
[V ] is dense. QQQ
(b) If H
1
, H
2
RO(Y ) then (H
1
H
2
) = H
1
H
2
. PPP Because is order-preserving, (H
1
H
2
)
H
1
H
2
. ??? Suppose, if possible, that they are not equal. Then (because (H
1
H
2
) is a regular open
set) G = H
1
H
2
(H
1
H
2
) is non-empty. Set M = f[G]. Then f
1
[M] G is not nowhere dense,
so H = int M must be non-empty. Now G H
1
f
1
[H
1
], so
f[G] f[f
1
[H
1
]] f[f
1
[H
1
]] H
1
,
so M H
1
and H int H
1
= H
1
. Similarly, H H
2
, and f
1
[H] f
1
[H
1
H
2
] (H
1
H
2
). But also
H f[G] is not empty, so
,= G f
1
[H] G (H
1
H
2
),
which is impossible. XXXQQQ
(c) If H RO(Y ) and H
t
= Y H is its complement in RO(Y ) then H
t
= X H is the complement
of H in RO(X). PPP By (b), H and H
t
are disjoint. Now H H
t
is a dense open subset of Y , so
H H
t
f
1
[H] f
1
[H
t
] = f
1
[H H
t
]
is dense in X, and the regular open set H
t
must include the complement of H in RO(X). QQQ
Putting (b) and (c) together, we see that the conditions of 312H(ii) are satised, so that is a Boolean
homomorphism.
(d) To see that it is order-continuous, let H RO(Y ) be a non-empty set with supremum Y . Then
H
0
=

H is a dense open subset of Y (see the formula in 314P). So

48 Boolean algebras *314R

H1
H

H1
f
1
[H] = f
1
[H
0
]
is dense in X, and sup
H1
H = X in RO(X). By 313L(b-iii), is order-continuous.
314S It is now easy to characterize the Stone spaces of Dedekind complete Boolean algebras.
Theorem Let A be a Boolean algebra, and Z its Stone space; write c for the algebra of open-and-closed
subsets of Z, and G for the regular open algebra of Z. Then the following are equiveridical:
(i) A is Dedekind complete;
(ii) Z is extremally disconnected (denition: 3A3Ae);
(iii) c = G.
proof (i)(ii) If A is Dedekind complete, let G be any open set in Z. Set A = a : a A, a G,
a
0
= supA. Then G =

a : a A, because c is a base for the topology of Z, so a
0
= G, by 313Ca.
Consequently G is open. As G is arbitrary, Z is extremally disconnected.
(ii)(iii) If E c, then of course E = E = int E, so E is a regular open set. Thus c G. On the
other hand, suppose that G Z is a regular open set. Because Z is extremally disconnected, G is open; so
G = int G = G is open-and-closed, and belongs to c. Thus c = G.
(iii)(i) Since G is Dedekind complete (314P), c and A are also Dedekind complete Boolean algebras.
Remark Note that if the conditions above are satised, either 312L or the formulae in 314P show that the
Boolean structures of c and G are identical.
314T I come now to a construction of great importance, both as a foundation for further constructions
and as a source of insight into the nature of Dedekind completeness.
Theorem Let A be a Boolean algebra, with Stone space Z; for a A let a be the corresponding open-and-
closed subset of A. Let

A be the regular open algebra of Z (314P).
(a) The map a a is an injective order-continuous Boolean homomorphism from A onto an order-dense
subalgebra of

A.
(b) If B is any Dedekind complete Boolean algebra and : A B is an order-continuous Boolean
homomorphism, there is a unique order-continuous Boolean homomorphism
1
:

A B such that
1
a = a
for every a A.
proof (a)(i) Setting c = a : a A, every member of c is open-and-closed, so is surely equal to the
interior of its closure, and is a regular open set; thus a

A for every a A. The formulae in 314P tell us
that if a, b A, then a

b, taken in

A, is just the set-theoretic intersection a

b = (a b) ; while 1 \ a,
taken in

A, is
Z a = Z a = (1 \ a) .
And of course

0 = is the zero of

A. Thus the map a a : A

A preserves and complementation, so
is a Boolean homomorphism (312H). Of course it is injective.
(ii) If A A is non-empty and inf A = 0, then

aA
a is nowhere dense in Z (313Cc), so
infa : a A = int(

aA
a) =
(314P again). As A is arbitrary, the map a a : A

A is order-continuous.
(iii) If G

A is not empty, then there is a non-empty member of c included in it, by the denition of
the topology of Z (311I). So c is an order-dense subalgebra of

A.
(b) Now suppose that B is a Dedekind complete Boolean algebra and : A B is an order-continuous
Boolean homomorphism. Write a = a for a A, so that : A

A is an isomorphism between A and the
order-dense subalgebra c of

A. Accordingly
1
: c B is an order-continuous Boolean homomorphism,
being the composition of the order-continuous Boolean homomorphisms and
1
. By 314K, it has an
extension to a Boolean homomorphism
1
:

A B, and
1
= , that is,
1
a = a for every a A. Now

1
is order-continuous. PPP Suppose that H

A has supremum 1 in

A. Set
314Xc Order-completeness 49
H
t
= E : E c, E H for some H H.
Because c is order-dense in

A,
H = sup
EL,EH
E = sup
E1

,EH
E
for every H H (313K), and sup H
t
= 1 in

A. It follows at once that supH
t
= 1 in c, so sup
1
[H
t
] =
sup(
1
)[H
t
] = 1. Since any upper bound for
1
[H] must also be an upper bound for
1
[H
t
], sup
1
[H] = 1
in B. As H is arbitrary,
1
is order-continuous (313L(b-iii)). QQQ
If
t
1
:

A B is any other Boolean homomorphism such that
t
1
a = a for every a A, then
1
and
t
1
agree on c, and the argument just above shows that
t
1
is also order-continuous. But if G

A, G is the
supremum (in

A) of T = E : E c, E G, so

t
1
G = sup
EJ

t
1
E = sup
EJ

1
E =
1
G.
As G is arbitrary,
t
1
=
1
. Thus
1
is unique.
314U The Dedekind completion of a Boolean algebra For any Boolean algebra A, I will say that
the Boolean algebra

A constructed in 314T is the Dedekind completion of A.
When using this concept I shall frequently suppress the distinction between a A and a

A, and treat
A as itself an order-dense subalgebra of

A.
314V Upper envelopes (a) Let A be a Boolean algebra, and C a subalgebra of A. For a A, write
upr(a, C) = infc : c C, a c
if the inmum is dened in C. (The most important cases are when A is Dedekind complete and C is order-
closed in A, so that C is Dedekind complete (314E) and upr(a, C) is dened for every a A; but others also
arise.)
(b) If A A is such that upr(a, C) is dened for every a A, a
0
= supA is dened in A and c
0
=
sup
aA
upr(a, C) is dened in C, then c
0
= upr(a
0
, C). PPP If c C then
c
0
c upr(a, C) c for every a A
a c for every a A a
0
c. QQQ
In particular, upr(a a
t
, C) = upr(a, C) upr(a
t
, C) whenever the right-hand side is dened.
(c) If a A is such that upr(a, C) is dened, then upr(a c, C) = c upr(a, C) for every c C. PPP For
c
t
C,
a c c
t
a c
t
(1 \ c)
upr(a, C) c
t
(1 \ c) c upr(a, C) c
t
. QQQ
314X Basic exercises (a) Let A be a Boolean algebra. (i) Show that the following are equiveridical:
() A is Dedekind complete () every non-empty upwards-directed subset of A with an upper bound has a
least upper bound () every non-empty downwards-directed subset of A with a lower bound has a greatest
lower bound. (ii) Show that the following are equiveridical: () A is Dedekind -complete () every non-
decreasing sequence in A with an upper bound has a least upper bound () every non-increasing sequence
in A with a lower bound has a greatest lower bound.
(b) Let A be a Boolean algebra. Show that any principal ideal of A is order-closed. Show that A is
Dedekind complete i every order-closed ideal is principal.
>>>(c) Let A be a Dedekind -complete Boolean algebra, Ba Boolean algebra and : A Ba sequentially
order-continuous Boolean homomorphism. If C A and C is the -subalgebra of A generated by C, show
that [C] is the -subalgebra of B generated by [C].
50 Boolean algebras 314Xd
(d) Let A be a Dedekind complete Boolean algebra, B an order-closed subalgebra of A, and a A; let
A
a
be the principal ideal of A generated by a. Show that a b : b B is an order-closed subalgebra of
A
a
.
(e) Find a proof of 314Tb which does not appeal to 314K.
>>>(f ) Let A be a Dedekind complete Boolean algebra and B an order-dense subalgebra of A. Show that
A is isomorphic to the Dedekind completion of B.
(g) Let A be a Dedekind complete Boolean algebra and C an order-closed subalgebra of A. Show that an
element a of A belongs to C i upr(1 \ a, C) = 1 \ upr(a, C) i upr(1 \ a, C) upr(a, C) = 0.
>>>(h) Let A be a Dedekind complete Boolean algebra, C an order-closed subalgebra of A, and a
0
A,
c
0
C. Show that the following are equiveridical: (i) there is a Boolean homomorphism : A C such
that c = c for every c C and a
0
= c
0
(ii) 1 \ upr(1 \ a
0
, C) c
0
upr(a
0
, C).
314Y Further exercises (a) Let P be a Dedekind complete partially ordered set. Show that a set
Q P is order-closed i supR, inf R belong to Q whenever R Q is a totally ordered subset of Q with
upper and lower bounds in P. (Hint: show by induction on that if A Q is upwards-directed and bounded
above and #(A) then sup A Q.)
(b) Let P be a lattice. Show that P is Dedekind complete i every non-empty totally ordered subset of
P with an upper bound in P has a least upper bound in P. (Hint: if A P is non-empty and bounded
below in P, let B be the set of lower bounds of A and use Zorns Lemma to nd a maximal element of B.)
(c) Give an example of a Boolean algebra A with an order-closed subalgebra A
0
and an element c such
that the subalgebra generated by A
0
c is not order-closed.
(d) Let X be any topological space. Let / be the -ideal of meager subsets of X, and set

B = GA : G X is open, A /.
(i) Show that

B is a -algebra of subsets of X, and that

B// is Dedekind complete. (Members of

B are said
to be the subsets of X with the Baire property;

B is the Baire property algebra of X.) (ii) Show that
if A X and

G : G X is open, AG

B is dense, then A

B. (iii) Show that there is a largest open
set V /. (iv) Let G be the regular open algebra of X. Show that the map G G

is an order-continuous
Boolean homomorphism from G onto

B//, so induces a Boolean isomorphism between the principal ideal
of G generated by X V and

B//. (

B// is the category algebra of X; it is a Dedekind complete

Boolean algebra. X is called a Baire space if V = ; in this case G

=

B//.)
(e) Let A be a Dedekind -complete Boolean algebra, and a
n
)
nN
any sequence in A. For n N set
E
n
= x : x 0, 1
N
, x(n) = 1, and let B be the -algebra of subsets of 0, 1
N
generated by E
n
: n N.
(B is the Borel -algebra of 0, 1
N
; see 4A2Wd.) Show that there is a unique sequentially order-continuous
Boolean homomorphism : B A such that (E
n
) = a
n
for every n N. (Hint: dene a suitable function
from the Stone space Z of A to 0, 1
N
, and consider E : E 0, 1
N
,
1
[E] has the Baire property in
Z.) Show that [B] is the -subalgebra of A generated by a
n
: n N.
(f ) Let A be a Boolean algebra, and Z its Stone space. Show that A is Dedekind -complete i G is open
whenever G is a cozero set in Z. (Such spaces are called basically disconnected or quasi-Stonian.)
(g) Let A, B be Dedekind complete Boolean algebras and D A 0 an order-dense set. Suppose that
: D B is such that (i) [D] is order-dense in B (ii) for all d, d
t
D, d d
t
= 0 i d d
t
= 0. Show
that has a unique extension to a Boolean isomorphism from A to B.
(h) Let A be any Boolean algebra. Let be the family of order-closed ideals in A. Show that (i)
is a Dedekind complete Boolean algebra with operations dened by the formulae I J = I J, 1 \ J =
a : a b = 0 for every b J (ii) the map a A
a
, the principal ideal generated by a, is an injective
order-continuous Boolean homomorphism from A onto an order-dense subalgebra of (iii) is isomorphic
to the Dedekind completion of A.
315Ab Products and free products 51
314 Notes and comments At the risk of being tiresomely long-winded, I have taken the trouble to spell out
a large proportion of the results in this section and the last in their sequential as well as their unrestricted
forms. The point is that while (in my view) the underlying ideas are most clearly and dramatically expressed
in terms of order-closed sets, order-continuous functions and Dedekind complete algebras, a large proportion
of the applications in measure theory deal with sequentially order-closed sets, sequentially order-continuous
functions and Dedekind -complete algebras. As a matter of simple technique, therefore, it is necessary
to master both, and for the sake of later reference I generally give the statements of both versions in full.
Perhaps the points to look at most keenly are just those where there is a dierence in the ideas involved, as
in 314Bb, or in which there is only one version given, as in 314M and 314T.
If you have seen the Hahn-Banach theorem (3A5A), it may have been recalled to your mind by Theorem
314K; in both cases we use an order relation and a bit of algebra to make a single step towards an extension
of a function, and Zorns lemma to turn this into the extension we seek. A good part of this section has
turned out to be on the borderland between the theory of Boolean algebra and general topology; naturally
enough, since (as always with the general theory of Boolean algebra) one of our rst concerns is to establish
connexions between algebras and their Stone spaces.
I think 314T is the rst substantial universal mapping theorem in this volume; it is by no means the
last. The idea of the construction

A is not just that we obtain a Dedekind complete Boolean algebra in
which A is embedded as an order-dense subalgebra, but that we simultaneously obtain a theorem on the
canonical extension to

A of order-continuous Boolean homomorphisms dened on A. This characterization
is enough to dene the pair (

A, a a) up to isomorphism, so the exact method of construction of

A becomes
of secondary importance. The one used in 314T is very natural (at least, if we believe in Stone spaces), but
there are others (see 314Yh), with dierent virtues.
314K and 314T both describe circumstances in which we can nd extensions of Boolean homomorphisms.
Clearly such results are fundamental in the theory of Boolean algebras, but I shall not attempt any systematic
presentation here. 314Ye can also be regarded as belonging to this family of ideas.
315 Products and free products
I describe here two algebraic constructions of fundamental importance. They are very dierent in charac-
ter, indeed may be regarded as opposites, despite the common use of the word product. The rst part of the
section (315A-315G) deals with the easier construction, the simple product; the second part (315H-315P)
with the free product.
315A Products of Boolean algebras (a) Let A
i
)
iI
be any family of Boolean algebras. Set A =

iI
A
i
, with the natural ring structure
a .b = a(i) .b(i))
iI
,
a b = a(i) b(i))
iI
for a, b A. Then A is a ring (3A2H); it is a Boolean ring because
a a = a(i) a(i))
iI
= a
for every a A; and it is a Boolean algebra because if we set 1
A
= 1
A
i
)
iI
, then 1
A
a = a for every a A.
I will call A the simple product of the family A
i
)
iI
.
I should perhaps remark that when I = then A becomes , to be interpreted as the singleton Boolean
algebra.
(b) The Boolean operations on A are now dened by the formulae
a b = a(i) b(i))
iI
, a \ b = a(i) \ b(i))
iI
for all a, b A.
52 Boolean algebras 315B
315B Theorem Let A
i
)
iI
be a family of Boolean algebras, and A their simple product.
(a) The maps a
i
(a) = a(i) : A A
i
are all Boolean homomorphisms.
(b) If Bis any other Boolean algebra, then a map : B A is a Boolean homomorphism i
i
: B A
i
is a Boolean homomorphism for every i I.
proof Verication of these facts amounts just to applying the denitions with attention.
315C Products of partially ordered sets (a) It is perhaps worth spelling out the following elemen-
tary denition. If P
i
)
iI
is any family of partially ordered sets, its product is the set P =

iI
P
i
ordered
by saying that p q i p(i) q(i) for every i I; it is easy to check that P is now a partially ordered set.
(b) The point is that if A is the simple product of a family A
i
)
iI
of Boolean algebras, then the ordering
of A is just the product partial order:
a b a b = a a(i) b(i) = a(i) i I a(i) b(i) i I.
Now we have the following elementary, but extremely useful, general facts about products of partially
ordered sets.
315D Proposition Let P
i
)
iI
be a family of non-empty partially ordered sets with product P.
(a) For any non-empty set A P and q P,
(i) supA = q in P i sup
pA
p(i) = q(i) in P
i
for every i I,
(ii) inf A = q in P i inf
pA
p(i) = q(i) in P
i
for every i I.
(b) The coordinate maps p
i
(p) = p(i) : P P
i
are all order-preserving and order-continuous.
(c) For any partially ordered set Q and function : Q P, is order-preserving i
i
is order-preserving
for every i I.
(d) For any partially ordered set Q and order-preserving function : Q P,
(i) is order-continuous i
i
is order-continuous for every i,
(ii) is sequentially order-continuous i
i
is sequentially order-continuous for every i.
(e)(i) P is Dedekind complete i every P
i
is Dedekind complete.
(ii) P is Dedekind -complete i every P
i
is Dedekind -complete.
proof All these are elementary verications. Of course parts (b), (d) and (e) rely on (a).
315E Factor algebras as principal ideals Because Boolean algebras have least elements, we have a
second type of canonical map associated with their products. If A
i
)
iI
is a family of Boolean algebras with
simple product A, dene
i
: A
i
A by setting (
i
a)(i) = a, (
i
a)(j) = 0
A
j
if i I, a A
i
and j I i.
Each
i
is a ring homomorphism, and is a Boolean isomorphism between A
i
and the principal ideal of A
generated by
i
(1
A
i
). The family
i
(1
A
i
))
iI
is a partition of unity in A.
Associated with these embeddings is the following important result.
315F Proposition Let A be a Boolean algebra and e
i
)
iI
a partition of unity in A. Suppose
either (i) that I is nite
or (ii) that I is countable and A is Dedekind -complete
or (iii) that A is Dedekind complete.
Then the map a a e
i
)
iI
is a Boolean isomorphism between A and

iI
A
e
i
, writing A
e
i
for the
principal ideal of A generated by e
i
for each i.
proof The given map is a Boolean homomorphism because each of the maps a a e
i
: A A
e
i
is (312J).
It is injective because sup
iI
e
i
= 1, so if a A 0 there is an i such that a e
i
,= 0. It is surjective
because e
i
)
iI
is disjoint and if c

iI
A
e
i
then a = sup
iI
c(i) is dened in A and
a e
j
= sup
iI
c(i) e
j
= c(j)
for every j I (using 313Ba). The three alternative versions of the hypotheses of this proposition are
designed to ensure that the supremum is always well-dened in A.
315I Products and free products 53
315G Algebras of sets and their quotients The Boolean algebras of measure theory are mostly
presented as algebras of sets or quotients of algebras of sets, so it is perhaps worth spelling out the ways in
which the product construction applies to such algebras.
Proposition Let X
i
)
iI
be a family of sets, and
i
an algebra of subsets of X
i
for each i.
(a) The simple product

iI

i
may be identied with the algebra
= E : E X, x : (x, i) E
i
for every i I
of subsets of X = (x, i) : i I, x X
i
, with the canonical homomorphisms
i
:
i
being given by

i
E = x : (x, i) E
for each E .
(b) Now suppose that
i
is an ideal of
i
for each i. Then

iI

i
/
i
may be identied with /, where
= E : E , x : (x, i) E
i
for every i I,
and the canonical homomorphisms
i
: /
i
/
i
are given by the formula
i
(E

) = (
i
E)

for every
E .
proof (a) It is easy to check that is a subalgebra of TX, and that the map E
i
E)
iI
:

iI

i
is a Boolean isomorphism.
(b) Again, it is easy to check that is an ideal of , that the proposed formula for
i
does indeed dene
a map from / to
i
/
i
, and that E

i
E

)
iI
is an isomorphism between / and

iI

i
/
i
.
315H Free products I come now to the second construction of this section.
(a) Denition Let A
i
)
iI
be a family of Boolean algebras. For each i I, let Z
i
be the Stone space
of A
i
. Set Z =

iI
Z
i
, with the product topology. Then the free product of A
i
)
iI
is the algebra A of
open-and-closed sets in Z; I will denote it by

iI
A
i
.
(b) For i I and a A
i
, the set a Z
i
representing a is an open-and-closed subset of Z
i
; because
z z(i) : Z Z
i
is continuous,
i
(a) = z : z(i) a is open-and-closed, so belongs to A. In this context
I will call
i
: A
i
A the canonical map.
(c) The topological space Z may be identied with the Stone space of the Boolean algebra A. PPP By
Tychonos theorem (3A3J), Z is compact. If z Z and G is an open subset of Z containing z, then there
are J, G
j
)
iJ
such that J is a nite subset of I, G
j
is an open subset of Z
j
for each j J, and
z w : w Z, w(j) G
j
for every j J G.
Because each Z
j
is zero-dimensional, we can nd an open-and-closed set E
j
Z
j
such that z(j) E
j
G
j
.
Now
E = Z

jJ
w : w(j) E
j

is a nite intersection of open-and-closed subsets of Z, so is open-and-closed; and z E G. As z and G

are arbitrary, Z is zero-dimensional. Finally, Z, being the product of Hausdor spaces, is Hausdor. So the
result follows from 311J. QQQ
315I Theorem Let A
i
)
iI
be a family of Boolean algebras, with free product A.
(a) The canonical map
i
: A
i
A is a Boolean homomorphism for every i I.
(b) For any Boolean algebra B and any family
i
)
iI
such that
i
is a Boolean homomorphism from A
i
to B for every i, there is a unique Boolean homomorphism : A B such that
i
=
i
for each i.
proof These are both consequences of 312P-312Q. As in 315H, write Z
i
for the Stone space of A, and Z for

iI
Z
i
, identied with the Stone space of A, as observed in 315Hc. The maps
i
: A
i
A are dened as
the homomorphisms corresponding to the continuous maps z
i
(z) = z(i) : Z Z
i
, so (a) is surely true.
Now suppose that we are given a Boolean homomorphism
i
: A
i
B for each i I. Let W be the
Stone space of B, and let

i
: W Z
i
be the continuous function corresponding to
i
. By 3A3Ib, the map
w

(w) =

i
(w))
iI
: W Z is continuous, and corresponds to a Boolean homomorphism : A B;
54 Boolean algebras 315I
because

i
=
i

,
i
=
i
for each i. Moreover, is the only Boolean homomorphism with this property,
because if : A B is a Boolean homomorphism such that
i
=
i
for every i, then corresponds to a
continuous map

: W Z, and we must have
i

=

i
for each i, so that

=

and = . This proves
(b).
315J Of course 315I is the dening property of the free product (see 315Xg below). I list a few further
basic facts.
Proposition Let A
i
)
iI
be a family of Boolean algebras, and A their free product; write
i
: A
i
A for
the canonical maps.
(a) A is the subalgebra of itself generated by

iI

i
[A
i
].
(b) Write C for the set of those members of A expressible in the form inf
jJ

j
(a
j
), where J I is nite
and a
j
A
j
for every j. Then every member of A is expressible as the supremum of a disjoint nite subset
of C. In particular, C is order-dense in A.
(c) Every
i
is order-continuous.
(d) A = 0
A
i there is some i I such that A
i
= 0
A
i
.
(e) Now suppose that A
i
,= 0
A
i
for every i I.
(i)
i
is injective for every i I.
(ii) If J I is nite and a
j
is a non-zero member of A
j
for each j J, then inf
jJ

j
(a
j
) ,= 0.
(iii) If i, j are distinct members of I and a A
i
, b A
j
, then
i
(a) =
j
(b) i either a = 0
A
i
and
b = 0
A
j
or a = 1
A
i
and b = 1
A
j
.
proof As usual, write Z
i
for the Stone space of A
i
, and Z =

iI
Z
i
, identied with the Stone space of A
(315Hc).
(a) Write A
t
for the subalgebra of A generated by

iI

i
[A
i
]. Then
i
: A
i
A
t
is a Boolean homo-
morphism for each i, so by 315Ib there is a Boolean homomorphism : A A
t
such that
i
=
i
for each
i. Now, regarding as a Boolean homomorphism from A to itself, the uniqueness assertion of 315Ib (with
B = A) shows that must be the identity, so that A
t
= A.
(b) Write T for the set of nite partitions of unity in A consisting of members of C, and A for the set of
members of A expressible in the form sup D
t
where D
t
is a subset of a member of T. Then A is a subalgebra
of A. PPP (i) 1
A
C (set J = in the denition of members of C) so 1
A
T and 0
A
, 1
A
A. (ii) Note that
if c, d C then c d C. (iii) If a, b A, express them as supD
t
, sup E
t
where D
t
D T, E
t
E T.
Then
F = d e : d D, e E T,
so
1
A
\ a = sup D D
t
A,
a b = supf : f F, f a b A. QQQ
Also,
i
[A
i
] A for each i I. PPP If a A
i
, then
i
(a),
i
(1
A
i
\ a) T, so
i
(a) A. QQQ
So (a) tells us that A = A, and every member of A is a nite disjoint union of members of C.
(c) If i I and A A
i
and inf A = 0 in A
i
, take any non-zero c A. By (b), we can nd a nite J I
and a family a
j
)
jJ
such that c
t
= inf
jJ

j
(a
j
) c and c
t
,= 0. Regarding c
t
as a subset of Z, we have
a point z c
t
. Adding i to J and setting a
i
= 1
A
i
if necessary, we may suppose that i J. Now c
t
,= 0
A
so a
i
,= 0
A
i
and there is an a A such that a
i
, a, so there is a t a
i
a. In this case, setting w(i) = t,
w(j) = z(j) for j ,= i, we have w c
t

i
(a), and c
t
, c are not included in
i
(a). As c is arbitrary, this shows
that inf
i
[A] = 0. As A is arbitrary,
i
is order-continuous.
(d) The point is that A = 0
A
i Z = , which is so i some Z
i
is empty.
(e)(i) Because no Z
i
is empty, all the coordinate maps from Z to Z
i
are surjective, so the corresponding
homomorphisms
i
are injective (312Ra).
(ii) Because J is nite,
315L Products and free products 55
inf
jJ

j
(a
j
) = z : z Z, z(j) a
j
for every j J
is not empty.
(iii) If
i
(a) =
j
(b) = 0
A
then (using (i)) a = 0
A
i
and b = 0
A
j
; if
i
(a) =
j
(b) = 1
A
then a = 1
A
i
and
b = 1
A
j
. ??? If
i
(a) =
j
(b) A 0
A
, 1
A
, then there are t a and u Z
j

b. Now there is a z Z such

that z(i) = t and z(j) = u, so that z
i
(a)
j
(b). XXX
315K Proposition Let A
i
)
iI
be any family of Boolean algebras, and J
k
)
kK
any partition of I.
Then the free product A of A
i
)
iI
is isomorphic to the free product B of B
k
)
kK
, where each B
k
is the
free product of A
i
)
iJ
k
.
proof Write
i
: A
i
A,
t
i
: A B
k
and
k
: B
k
B for the canonical maps when k K, i J
k
. Then
the homomorphisms
k

t
i
: A
i
Bcorrespond to a homomorphism : A Bsuch that
i
=
k

t
i
whenever
i J
k
. Next, for each k, the homomorphisms
i
: A
i
A, for i J
k
, correspond to a homomorphism

k
: B
k
A such that
k

t
i
=
i
for i J
k
; and the family
k
)
kK
corresponds to a homomorphism
: B A such that
k
=
k
for k K. Consequently

i
=
k

t
i
=
k

t
i
=
i
whenever k K, i J
k
. Once again using the uniqueness assertion in 315Ib, is the identity homomor-
phism on A. On the other hand, if we look at : B B, then we see that

t
i
=
k

t
i
=
i
=
k

t
i
whenever k K, i J
k
. Now, for given k, b : b B
k
,
k
b =
k
b is a subalgebra of B
k
including

iJ
k

t
i
[A
i
], and must be the whole of B
k
, by 315Ja. So b : b B, b = b is a subalgebra of B including

kK

k
[B
k
], and is the whole of B. Thus is the identity on B and , are the two halves of an
isomorphism between A and B.
315L Algebras of sets and their quotients Once again I devote a paragraph to spelling out the
application of the construction to the algebras most important to us.
Proposition Let X
i
)
iI
be a family of sets, and
i
an algebra of subsets of X
i
for each i.
(a) The free product

iI

i
may be identied with the algebra of subsets of X =

iI
X
i
generated
by the set
i
(E) : i I, E
i
, where
i
(E) = x : x X, x(i) E.
(b) Now suppose that
i
is an ideal of
i
for each i. Then

iI

i
/
i
may be identied with /, where
is the ideal of generated by
i
(E) : i I, E
i
; the corresponding canonical maps
i
:
i
/
i
/
being dened by the formula
i
(E

) = (
i
(E))

for i I, E
i
.
proof I start by proving (b) in detail; the argument for (a) is then easy to extract. Write A
i
=
i
/
i
,
A = /.
(i) Fix i I for the moment. By the denition of ,
i
(E) for E
i
, and it is easy to check that

i
:
i
is a Boolean homomorphism. Again, because
i
(E) whenever E
i
, the kernel of the
homomorphism E (
i
(E))

:
i
A includes
i
, so the formula for
i
denes a homomorphism from A
i
to A.
Now let C =

iI
A
i
be the free product, and write
t
i
: A
i
C for the canonical homomorphisms. By
315I, there is a Boolean homomorphism : C A such that
t
i
=
i
for each i. The set
E : E , E

[C]
is a subalgebra of including
i
[
i
] for every i, so is itself, and is surjective.
(ii) We need a simple description of the ideal , as follows: a set E belongs to i there are a nite
K I and a family F
k
)
kK
such that F
k

k
for each k and E

kK

k
(F
k
). For evidently such sets
have to belong to , since the
k
(F
k
) will be in , while the family of all these sets is an ideal containing

i
(F) whenever i I, F
i
.
(iii) Now we can see that : C A is injective. PPP Take any non-zero c C. By 315Jb, we can nd a
nite J I and a family a
j
)
jJ
in

jJ
A
j
such that 0 ,= inf
jJ

t
j
a
j
c. Express each a
j
as E

j
, where
E
j

j
, and consider E = X

jJ

j
(E
j
) . Then
56 Boolean algebras 315L
E

= inf
jJ

j
a
j
= (inf
jJ

t
j
a
j
) (c).
Also, because
t
j
a
j
,= 0, E
j
/
j
for each j. But it follows that E / , because if K I is nite and
F
k

k
for each k K, set E
i
= X
i
for i I J, F
i
= for i I K; then there is an x X such that
x(i) E
i
F
i
for each i I, so that x E

kK
F
k
. By the criterion of (ii), E / . So
0 ,= E

(c).
As c is arbitrary, the kernel of is 0, and is injective. QQQ
So : C A is the required isomorphism.
(iv) This proves (b). Reading through the arguments above, it is easy to see the simplications which
compose a proof of (a), reading
i
for A
i
and for
i
.
315M Notation Free products are suciently surprising that I think it worth taking a moment to look
at a pair of examples relevant to the kinds of application I wish to make of the concept in the next chapter.
First let me introduce a somewhat more direct notation which seems appropriate for the free product of
nitely many factors. If A and B are two Boolean algebras, I write A B for their free product, and for
a A, b B I write a b for
1
(a)
2
(b), where
1
: A A B,
2
: B A B are the canonical
maps. Observe that (a
1
b
1
) (a
2
b
2
) = (a
1
a
2
) (b
1
b
2
), and that the maps a a b
0
, b a
0
b
are always ring homomorphisms. Now 315J(e-ii) tells us that a b = 0 only when one of a, b is 0. In the
context of 315L, we can identify E F with E F for E
1
, F
2
, and E

with (E F)

.
315N Lemma Let A, B be Boolean algebras.
(a) Any element of AB is expressible as sup
iI
a
i
b
i
where a
i
)
iI
is a nite partition of unity in A.
(b) If c A B is non-zero there are non-zero a A, b B such that a b c.
proof (a) Let C be the set of elements of A B representable in this form. Then C is a subalgebra of
A B. PPP (i) If a
i
)
iI
, a
t
j
)
jJ
are nite partitions of unity in A, and b
i
, b
t
j
members of B for i I and
j J, the a
i
a
t
j
)
iI,jJ
is a partition of unity in A, and
(sup
iI
a
i
b
i
) (sup
jJ
a
t
j
b
t
j
) = sup
iI,jJ
(a
i
b
i
) (a
t
j
b
t
j
)
= sup
iI,jJ
(a
i
a
t
j
) (b
i
b
t
j
) C.
So c c
t
C for all c, c
t
C. (ii) If a
i
)
iI
is a nite partition of unity in A and b
i
B for each i, then
1 \ sup
iI
a
i
b
i
= (sup
iI
a
i
1) \ (sup
iI
a
i
b
i
) = sup
iI
a
i
(1 \ b
i
) C.
Thus 1 \ c C for every c C. QQQ
Since a 1 = (a 1) ((1 \ a) 0) and 1 b belong to C for every a A, b B, C must be the whole
of A B, by 315Ja.
(b) Now this follows at once, just as in 315Jb.
315O Example A = TNTN is not Dedekind -complete. PPP Consider A = n n : n N A.
??? If A has a least upper bound c in A, then c is expressible as a supremum sup
jk
a
j
b
j
, by 315Jb. Because
k is nite, there must be distinct m, n such that j : m a
j
= j : n a
j
. Now n n c, so there
is a j k such that
(a
j
n) (b
j
n) = (n n) (a
j
b
j
) ,= 0,
so that neither a
j
n nor b
j
n is empty, that is, n a
j
b
j
. But this means that m a
j
, so that
(a
j
b
j
) (m n) = (a
j
m) (b
j
n) ,= 0,
and c (m n) ,= 0, even though a (m n) = 0 for every a A. XXX Thus we have found a
countable subset of A with no supremum in A, and A is not Dedekind -complete. QQQ
315P Example Now let A be any non-trivial atomless Boolean algebra, and B the free product AA.
Then the identity homomorphism from A to itself induces a homomorphism : B A given by setting
315Xj Products and free products 57
(a b) = a b for every a, b A. The point I wish to make is that is not order-continuous. PPP Let C
be the set a b : a, b A, a b = 0. Then (c) = 0
A
for every c C. If d B is non-zero, then by
315Nb there are non-zero a, b A such that ab d; now, because A is atomless, there is a non-zero a
t
a
such that a \ a
t
,= 0. At least one of b \ a
t
, b \ (a \ a
t
) is non-zero; suppose the former. Then a
t
(b \ a
t
) is a
non-zero member of C included in d. As d is arbitrary, this shows that supC = 1
B
. So
sup
cC
(c) = 0
A
,= 1
A
= (supC),
and is not order-continuous. QQQ
Thus the free product (unlike the product, see 315Dd) does not respect order-continuity.
315X Basic exercises (a) Let A
i
)
iI
be any family of Boolean algebras, with simple product A, and

i
: A A
i
the coordinate homomorphisms. Suppose we have another Boolean algebra A
t
, with homomor-
phisms
t
i
: A
t
A
i
, such that for every Boolean algebra B and every family
i
)
iI
of homomorphisms
from B to the A
i
there is a unique homomorphism : B A
t
such that
i
=
t
i
for every i. Show that
there is a unique isomorphism : A A
t
such that
t
i
=
i
for every i I.
(b) Let P
i
)
iI
be a family of non-empty partially ordered sets, with product partially ordered set P.
Show that P is a lattice i every P
i
is a lattice, and that in this case it is the product lattice in the sense
that p q = p(i) q(i))
iI
, p q = p(i) q(i))
iI
for all p, q P.
(c) Let A
i
)
iI
be a family of Boolean algebras with simple product A. For each i I let Z
i
be the Stone
space of A
i
, and let Z be the Stone space of A. Show that the coordinate maps from A onto A
i
induce
homeomorphisms between the Z
i
and open-and-closed subsets Z

i
of Z. Show that Z

i
)
iI
is disjoint. Show
that

iI
Z

i
is dense in Z, and is equal to Z i i : A
i
,= 0 is nite.
(d) Let A
i
)
iI
be a family of Boolean algebras, with simple product A. Suppose that for each i I we
are given an ideal I
i
of A
i
. Show that I =

iI
I
i
is an ideal of A, and that A/I may be identied, as
Boolean algebra, with

iI
A
i
/I
i
.
(e) Let X
i
)
iI
be any family of topological spaces. Let X be their disjoint union (x, i) : i I, x X
i
,
with the disjoint union topology; that is, a set G X is open in X i x : (x, i) G is open in X
i
for
every i I. (i) Show that the algebra of open-and-closed subsets of X can be identied, as Boolean algebra,
with the simple product of the algebras of open-and-closed sets of the X
i
. (ii) Show that the regular open
algebra of X can be identied, as Boolean algebra, with the simple product of the regular open algebras of
the X
i
.
(f ) Show that the topological product of any family of zero-dimensional spaces is zero-dimensional.
(g) Let A
i
)
iI
be any family of Boolean algebras, with free product A, and
i
: A
i
A the canonical
homomorphisms. Suppose we have another Boolean algebra A
t
, with homomorphisms
t
i
: A
i
A
t
, such
that for every Boolean algebra B and every family
i
)
iI
of homomorphisms from the A
i
to B there is a
unique homomorphism : A
t
B such that
i
=
t
i
for every i. Show that there is a unique isomorphism
: A A
t
such that
t
i
=
i
for every i I.
(h) Let I be any set, and let A be the algebra of open-and-closed sets of 0, 1
I
; for each i I set
a
i
= x : x 0, 1
I
, x(i) = 1 A. Show that for any Boolean algebra B, any family b
i
)
iI
in B there is
a unique Boolean homomorphism : A B such that (a
i
) = b
i
for every i I.
(i) Let A
i
)
iI
, B
j
)
jJ
be two families of Boolean algebras. Show that there is a natural injective
homomorphism :

iI
A
i

jJ
B
j

iI,jJ
A
i
B
j
dened by saying that
(a b) = a(i) b(j))
iI,jJ
for a

iI
A
i
, b

jJ
B
j
. Show that is surjective if I and J are nite.
(j) Let J(i))
iI
be a family of sets, with product Q =

iI
J(i). Let A
ij
)
iI,jJ(i)
be a family of
Boolean algebras. Describe a natural injective homomorphism :

iI

jJ(i)
A
ij

qQ

iI
A
i,q(i)
.
58 Boolean algebras 315Xk
(k) Let A and B be Boolean algebras with partitions of unity a
i
)
iI
, b
j
)
jJ
. Show that a
i
b
j
)
iI,jJ
is a partition of unity in A B.
(l) Let A and B be Boolean algebras and a A, b B. Write A
a
, B
b
for the corresponding principal
ideals. Show that there is a canonical isomorphism between A
a
B
b
and the principal ideal of A B
generated by a b.
(m) Let A
i
)
iI
be any family of Boolean algebras, with free product

iI
A
i
, and
i
: A
i
A the
canonical maps. Show that
i
[A
i
] is an order-closed subalgebra of A for every i.
(n) Let A be a Boolean algebra. Let us say that a family A
i
)
iI
of subalgebras of A is Boolean-
independent if inf
jJ
a
j
,= 0 whenever J I is nite and a
j
A
j
0 for every j J. Show that in this
case the subalgebra of A generated by

iI
A
i
is isomorphic to the free product

iI
A
i
.
(o) Let A
i
)
iI
and B
i
)
iI
be two families of Boolean algebras, and suppose that for each i I we are
given a Boolean homomorphism
i
: A
i
B
i
with kernel K
i
A
i
. Show that the
i
induce a Boolean
homomorphism :

iI
A
i

iI
B
i
with kernel generated by the images of the K
i
. Show that if every

i
is surjective, so is .
(p) Let A
i
)
iI
be any family of non-trivial Boolean algebras. Show that if J I and B
j
is a subalgebra
of A
j
for each j J, then

jJ
B
j
is canonically embedded as a subalgebra of

iI
A
i
.
(q) Let A and Bbe Boolean algebras, neither 0. Show that any element of ABis uniquely expressible
as sup
iI
a
i
b
i
where a
i
)
iI
is a partition of unity in A, with no a
i
equal to 0, and b
i
,= b
j
in B for i ,= j.
315Y Further exercises (a) Let A
i
)
iI
and B
i
)
iI
be two families of Boolean algebras, and suppose
that we are given Boolean homomorphisms
i
: A
i
B
i
for each i; let :

iI
A
i


iI
B
i
be the
induced homomorphism. (i) Show that if every
i
is order-continuous, so is . (ii) Show that if every
i
is
sequentially order-continuous, so is .
(b) Let Z
i
)
iI
be any family of topological spaces with product Z. For i I, z Z set
i
(z) = z(i).
Show that if M Z
i
is nowhere dense in Z
i
then
1
i
[M] is nowhere dense in Z. Use this to prove 315Jc.
(c) Let A
i
)
iI
be a family of Boolean algebras, and suppose that we are given subalgebras B
i
of A
i
for
each i; set A =

iI
A
i
and B =

iI
B
i
, and let : B A be the homomorphism induced by the
embeddings B
i
A
i
. (i) Show that if every B
i
is order-closed in A
i
, then [B] is order-closed in A. (ii)
Show that if every B
i
is a -subalgebra of A
i
, then [B] is a -subalgebra in A.
(d) Let X
i
)
iI
be a family of topological spaces, with product X. Let G
i
, G be the corresponding regular
open algebras. Show that G can be identied with the Dedekind completion of

iI
G
i
.
(e) Use the ideas of 315Xh and 315L to give an alternative construction of free product, for which 315I
and 315J(e-ii) are true, which does not depend on the concept of Stone space nor on any other use of the
axiom of choice. (Hint: show that for any Boolean algebra A there is a canonical surjection from the algebra
c
A
onto A, where c
J
is the algebra of subsets of 0, 1
J
generated by sets of the form x : x(j) = 1; show
that for such algebras c
J
, at least, the method of 315H-315I can be used; now apply the method of 315L to
describe

iI
A
i
as a quotient of c
J
where J = (a, i) : i I, a A
i
. Finally check that if no A
i
is trivial,
then nor is the free product.)
(f ) Let A and B be Boolean algebras. Show that A B is Dedekind complete i either A = 0 or
B = 0 or A is nite and B is Dedekind complete or B is nite and A is Dedekind complete.
(g) Let P
i
)
iI
be any family of partially ordered spaces. (i) Give a construction of a partially ordered
space P, together with a family of order-preserving maps
i
: P
i
P, such that whenever Q is a partially
ordered set and
i
: P
i
Q is order-preserving for every i I, there is a unique order-preserving map
: P Q such that
i
=
i
for every i. (ii) Show that will be order-continuous i every
i
is. (iii)
Show that P will be Dedekind complete i every P
i
is, but (except in trivial cases) is not a lattice.
316Ab Further topics 59
315 Notes and comments In this section I nd myself asking for slightly more sophisticated algebra than
seems necessary elsewhere. The point is that simple products and free products are best regarded as dened
by the properties described in 315B and 315I. That is, it is sometimes right to think of a simple product
of a family A
i
)
iI
of Boolean algebras as being a structure (A,
i
)
iI
) where A is a Boolean algebra,

i
: A A
i
is a homomorphism for every i I, and every family of homomorphisms from a Boolean algebra
B to the A
i
can be uniquely represented by a single homomorphism from B to A. Similarly, reversing the
direction of the homomorphisms, we can speak of a free product (it would be natural to say coproduct)
(A,
i
)
iI
) of A
i
)
iI
. On such denitions, it is elementary that any two simple products, or free products,
are isomorphic in the obvious sense (315Xa, 315Xg), and very general arguments from abstract algebra, not
restricted to Boolean algebras (see Bourbaki 68, IV.3.2), show that they exist. (But in order to prove such
basic facts as that the
i
are surjective, or that the
i
are, except when the construction collapses altogether,
injective, we do of course have to look at the special properties of Boolean algebras.) Now in the case of
simple products, the Cartesian product construction is so direct and so familiar that there seems no need
to trouble our imaginations with any other. But in the case of free products, things are more complicated.
I have given primacy to the construction in terms of Stone spaces because I believe that this is the fastest
route to eective mental pictures. But in some ways this approach seems to be inappropriate. If you take
what in my view is a tenable position, and say that a Boolean algebra is best regarded as the limit of its
nite subalgebras, then you might prefer a construction of a free product as a limit of free products of nitely
many nite subalgebras. Or you might feel that it is wrong to rely on the axiom of choice to prove a result
which certainly does not need it (see 315Ye).
Because I believe that the universal mapping theorem 315I is the right basis for the study of free products,
I am naturally led to use it as the starting point for proofs of theorems about free products, as in 315K.
But 315J(e-ii) seems to lie deeper. (Note, for instance, that in 315L we do need the axiom of choice, in part
(c) of the proof, since without it the product

iI
X
i
could be empty.)
Both simple product and free product are essentially algebraic constructions involving the category of
Boolean algebras and Boolean homomorphisms, and any relationships with such concepts as order-continuity
must be regarded as accidental. 315Cb and 315D show that simple products behave very straightforwardly
when the homomorphisms involved are order-continuous. 315P, 315Xm and 315Ya-315Yc show that free
products are much more complex and subtle.
For nite products, we have a kind of distributivity; (AB) C can be identied with (AC) (BC)
(315Xi, 315Xj). There are contexts in which this makes it seem more natural to write A B in place
of A B, and indeed I have already spoken of a direct sum of measure spaces (214K) in terms which
correspond closely to the simple product of algebras of sets described in 315Ga. Generally, the simple
product corresponds to disjoint unions of Stone spaces (315Xc) and the free product to products of Stone
spaces. But the simple product is indeed the product Boolean algebra, in the ordinary category sense;
the universal mapping theorem 315B is exactly of the type we expect from products of topological spaces
(3A3Ib) or partially ordered sets (315Dc), etc. It is the free product which is special to Boolean algebras.
The nearest analogy that I know of elsewhere is with the concept of tensor product of linear spaces (cf.
253).
316 Further topics
I introduce two special properties of Boolean algebras which will be of great importance in the rest of
this volume: the countable chain condition (316A-316F) and weak (, )-distributivity (316G-316K). I end
the section with brief notes on atoms in Boolean algebras (316L-316M).
316A Denitions (a) A Boolean algebra A is ccc, or satises the countable chain condition, if
every disjoint subset of A is countable.
(b) A topological space X is ccc, or satises the countable chain condition, or has Souslins prop-
erty, if every disjoint collection of open sets in X is countable.
60 Boolean algebras 316B
316B Theorem A Boolean algebra A is ccc i its Stone space Z is ccc.
proof (a) If A is ccc and ( is a disjoint family of open sets in Z, then for each G (
t
= ( we can nd
a non-zero a
G
A such that the corresponding open-and-closed set a
G
is included in G. Now a
G
: G (
t

is a disjoint family in A, so is countable; since a

G
,= a
H
for distinct G, H (
t
, (
t
and ( must be countable.
As ( is arbitrary, Z is ccc.
(b) If Z is ccc and A A is disjoint, then a : a A is a disjoint family of open subsets of Z, so must
be countable, and A is countable. As A is arbitrary, A is ccc.
316C Proposition Let A be a Dedekind -complete Boolean algebra and 1 a -ideal of A. Then the
quotient algebra B = A/1 is ccc i every disjoint family in A 1 is countable.
proof (a) Suppose that B is ccc and that A is a disjoint family in A 1. Then a

: a A is a disjoint
family in B, therefore countable, and a

,= b

when a, b are distinct members of A; so A is countable.

(b) Now suppose that B is not ccc. Then there is an uncountable disjoint set B B. Of course B 0
is still uncountable, so may be enumerated as b

)
<
, where is an uncountable cardinal (2A1K), so that

1
. For each <
1
, choose a

A such that a

= b

. Of course a

/ 1. If < <
1
, then b

= 0,
so a

1. Because <
1
, it is countable; because 1 is a -ideal, and A is Dedekind -complete,
d

= sup
<
a

1,
so that
c

= a

A 1.
But now of course
c

= 0
whenever < <
1
, so c

: <
1
is an uncountable disjoint family in A 1.
Remark An ideal 1 satisfying the conditions of this proposition is said to be
1
-saturated in A.
316D Corollary Let X be a set, a -algebra of subsets of X, and 1 a -ideal of A. Then the quotient
algebra /1 is ccc i every disjoint family in 1 is countable.
316E Proposition Let A be a ccc Boolean algebra. Then for any A A there is a countable B A
such that B has the same upper and lower bounds as A.
proof (a) Set
D =

aA
d : d a.
Applying Zorns lemma to the family ( of disjoint subsets of D, we have a maximal C
0
(. For each c C
0
choose a b
c
A such that c b
c
, and set B
0
= b
c
: c C
0
. Because A is ccc, C
0
is countable, so B
0
is also
countable. ??? If there is an upper bound e for B
0
which is not an upper bound for A, take a A such that
c
t
= a \ e ,= 0; then c
t
D and c
t
c = c
t
b
c
= 0 for every c C
0
, so C
0
c
t
(; but C
0
was supposed
to be maximal in (. XXX Thus every upper bound for B
0
is also an upper bound for A.
(b) Similarly, there is a countable set B
t
1
A
t
= 1 \ a : a A such that every upper bound for B
t
1
is
an upper bound for A
t
. Set B
1
= 1 \ b : b B
t
1
; then B
1
is a countable subset of A and every lower bound
for B
1
is a lower bound for A. Try B = B
0
B
1
. Then B is a countable subset of A and every upper (resp.
lower) bound for B is an upper (resp. lower) bound for A; so that B must have exactly the same upper and
lower bounds as A has.
316F Corollary Let A be a ccc Boolean algebra.
(a) If A is Dedekind -complete it is Dedekind complete.
(b) If A A is sequentially order-closed it is order-closed.
(c) If Q is any partially ordered set and : A Q is a sequentially order-continuous order-preserving
function, it is order-continuous.
316Hb Further topics 61
(d) If B is another Boolean algebra and : A B is a sequentially order-continuous Boolean homomor-
phism, it is order-continuous.
proof (a) If A is any subset of A, let B A be a countable set with the same upper bounds as A; then
supB is dened in A and must be supA.
(b) Suppose that B A is non-empty and upwards-directed and has a supremum a in A. Then there is
a non-empty countable set C B with the same upper bounds as B. Let c
n
)
nN
be a sequence running
over C. Because B is upwards-directed, we can choose b
n
)
nN
inductively such that
b
0
= c
0
, b
n+1
B, b
n+1
b
n
c
n+1
for every n N.
Now any upper bound for b
n
: n N must also be an upper bound for c
n
: n N = C, so is an upper
bound for the whole set B. But this means that a = sup
nN
b
n
. As b
n
)
nN
is a non-decreasing sequence in
A, and A is sequentially order-closed, a A.
In the same way, if B A is downwards-directed and has an inmum in A, this is also the inmum of
some non-increasing sequence in B, so must belong to A. Thus A is order-closed.
(c)(i) Suppose that A A is a non-empty upwards-directed set with supremum a
0
A. As in (b), there
is a non-decreasing sequence c
n
)
nN
in A with supremum a
0
. Because is sequentially order-continuous,
a
0
= sup
nN
c
n
in Q. But this means that a
0
must be the least upper bound of [A].
(ii) Similarly, if A A is a non-empty downwards-directed set with inmum a
0
, there is a non-increasing
sequence c
n
)
nN
in A with inmum a
0
, so that
inf [A] = inf
nN
c
n
= a
0
.
Putting this together with (i), we see that is order-continuous, as claimed.
(d) This is a special case of (c).
316G Denition Let A be a Boolean algebra. We say that A is weakly (, )-distributive if
whenever A
n
)
nN
is a sequence of non-empty downwards-directed subsets of A all with inmum 0 in A,
there is a set A A, also with inmum 0, such that for every a A, n N there is an a
t
A
n
such that
a
t
a.
316H Remarks (a) Note that for any Boolean algebra A, and any sequence A
n
)
nN
of non-empty
downwards-directed subsets of A, the set
B = b : b A, n N a
t
A
n
such that a
t
b
is downwards-directed. PPP If b
1
, b
2
B and n N, there are b
t
1
, b
t
2
A
n
such that b
1
b
t
1
and b
2
b
t
2
; now
there is a b
t
A
n
such that b
t
b
t
1
b
t
2
, and b
1
b
2
b
t
. As n is arbitrary, b
1
b
2
B; as b
1
and b
2
are
arbitrary, B is downwards-directed. QQQ Also, B is not empty (as 1 B). If A is weakly (, )-distributive
and inf A
n
= 0 for every n, then there is an A B such that inf A = 0, so in this case inf B = 0.
Thus a Boolean algebra A is weakly (, )-distributive i
whenever A
n
)
nN
is a sequence of non-empty downwards-directed subsets of A all with in-
mum 0 in A, there is a downwards-directed set A A, also with inmum 0, such that for every
a A, n N there is an a
t
A
n
such that a
t
a.
(b) I have expressed the denition above in terms which are adapted to general Boolean algebras. More
often than not, however, the algebras we are concerned with will be Dedekind -complete, and in this case
the condition
for every a A, n N there is an a
t
A
n
such that a
t
a
can be replaced by
for every a A there is a sequence a
n
)
nN
such that a
n
A
n
for every n and sup
nN
a
n
a;
so that the whole conclusion
there is a set A A, also with inmum 0, such that for every a A, n N there is an a
t
A
n
such that a
n
a
62 Boolean algebras 316Hb
can be replaced by
infsup
nN
a
n
: a
n
A
n
for every n N = 0.
Note that the set
B
t
= sup
nN
a
n
: a
n
A
n
for every n N
is also downwards-directed.
316I As usual, a characterization of the property in terms of the Stone spaces is extremely valuable.
Theorem Let A be a Boolean algebra, and Z its Stone space. Then A is weakly (, )-distributive i every
meager set in Z is nowhere dense.
proof (a) Suppose that A is weakly (, )-distributive and that M is a meager subset of Z. Then M
can be expressed as

nN
M
n
where each M
n
is nowhere dense. Set B
n
= b : b A,

b M
n
= .
Then

bB
n

b = Z M
n
is dense in Z, so supB
n
= 1 in A (313Ca). Set A
n
= 1 \ b : b B
n
; then
inf A
n
= 0 (313Aa). Also B
n
is upwards-directed (indeed, closed under ), so A
n
is downwards-directed.
Let A A be a set with inmum 0 such that every member of A includes members of every A
n
, and set
B = 1 \ a : a A. Then supB = 1 and every member of B is included in members of every B
n
, so that

b M
n
= for every b B, n N. Now

bB

b is a dense open subset of Z (313Ca again) disjoint from

M, so M is nowhere dense, as claimed.
(b) Suppose that every meager set in Z is nowhere dense, and that A
n
)
nN
is a sequence of non-empty
downwards-directed sets in A, all with inmum 0. Then M
n
=

aA
n
a is nowhere dense for each n (313Cc),
so M =

nN
M
n
is meager, therefore nowhere dense. Set B = b : b A,

b M = , A = 1 \ b : b B;
then sup B = 1 and inf A = 0. If c A, n N then c M M
n
, so
(Z c)

aA
n
(Z a) = M
n
c = ;
because Z c is compact and Z a is open for every a A
n
, there must be nitely many a
0
, . . . , a
k
A
n
such that c

ik
a
i
, that is, c inf
ik
a
i
. Because A
n
is downwards-directed, there must be an a
t
A
n
such that a
t
a
i
for every i k, so that c a
t
. But this is exactly what is required of A in the denition
316G. As A
n
)
nN
is arbitrary, A is weakly (, )-distributive.
316J It will be convenient later to have spelt out an elementary inversion of the denition in 316G.
Proposition Let A be a Dedekind -complete weakly (, )-distributive Boolean algebra. If A
n
)
nN
is a sequence of non-empty subsets of A such that c
n
= sup A
n
is dened in A for each n N, then
inf
nN
c
n
= supinf
nN
a
n
: a
n
A
n
for every n N.
proof Set c = inf
nN
c
n
, A = inf
nN
a
n
: a
n
A
n
for every n N. If a
n
A
n
for every n N, then
inf
nN
a
n
c, just because a
n
c
n
for every n; thus c is an upper bound for A. On the other hand, if
0 ,= d c, d = supa d : a A
n
, so inf
aA
n
d \ a = 0, for every n (313Ba, 313Aa). Because A is weakly
(, )-distributive, there must therefore be a sequence a
n
)
nN
such that a
n
A
n
for every n N and
inf
nN
d \ a
n
,= d, that is, d sup
nN
a
n
,= 0. This shows that c \ d cannot be and upper bound for A; as d
is arbitrary, c = supA, as claimed.
316K The regular open algebra of R For examples of weakly (, )-distributive algebras, I think we
can wait for Chapter 32 (see also 392I). But the standard example of an algebra which is not weakly (, )-
distributive is of such importance that (even though it has nothing to do with measure theory, narrowly
dened) I think it right to describe it here.
Proposition The algebra G of regular open subsets of R (314O) is not weakly (, )-distributive.
proof Enumerate Q as q
n
)
nN
. For each n N, set
A
n
= G : G G, q
i
G for every i n.
Then A
n
is downwards-directed, and
inf A
n
= int

A
n
= intq
i
: i n = .
316Xj Further topics 63
But if A G is such that
for every n N, G A there is an H A
n
such that H G,
then we must have Q G for every G A, so that
R = int Q int G = G
for every G A, and A R; which means that inf A ,= in G, and 316G cannot be satised.
316L Atoms in Boolean algebras (a) If A is a Boolean algebra, an atom in A is a non-zero a A
such that the only elements included in a are 0 and a.
(b) A Boolean algebra is atomless if it has no atoms.
(c) A Boolean algebra is purely atomic if every non-zero element includes an atom.
316M Proposition Let A be a Boolean algebra, with Stone space Z.
(a) There is a one-to-one correspondence between atoms a of A and isolated points z Z, given by the
formula a = z.
(b) A is atomless i Z has no isolated points.
(c) A is purely atomic i the isolated points of Z form a dense subset of Z.
proof (a)(i) If z is an isolated point in Z, then z is an open-and-closed subset of Z, so is of the form a
for some a A; now if b a,

b must be either or z, so b must be either a or 0, and a is an atom.

(ii) If a A and a has two points z and w, then (because Z is Hausdor, 311I) there is an open set G
containing z but not w. Now there is a c A such that z c G, so that a c must be dierent from both
0 and a, and a is not an atom.
(b) This follows immediately from (a).
(c) From (a) we see that A is purely atomic i a contains an isolated point for every non-zero a A; since
every non-empty open set in Z includes a non-empty set of the form a, this happens i every non-empty
open set in Z contains an isolated point, that is, i the set of isolated points is dense.
316X Basic exercises >>>(a) Show that any subalgebra of a ccc Boolean algebra is ccc.
(b) Show that any principal ideal of a ccc Boolean algebra is ccc.
(c) Let A
i
)
iI
be a family of Boolean algebras, with simple product A. Show that A is ccc i every A
i
is ccc and i : A
i
,= 0 is countable.
>>>(d) Let X be a separable topological space. Show that X is ccc.
>>>(e) Show that the regular open algebra of a topological space X is ccc i X is ccc, so that, in particular,
the regular open algebra of R is ccc.
(f ) Show that if A is a Boolean algebra and B is an order-dense subalgebra of A, then A is ccc i B is.
(g) Let A be a Dedekind -complete Boolean algebra. Show that it is ccc i there is no family a

)
<
1
in A such that a

whenever < <

1
.
(h) Let A be any Boolean algebra and 1 an order-closed ideal of A. Show that A/1 is ccc i there is no
uncountable disjoint family in A 1.
(i) Let A be a ccc Boolean algebra. Show that if 1 is a -ideal of A, then it is order-closed, and A/1 is
ccc.
(j) Let A be a Boolean algebra. Show that the following are equiveridical: (i) A is ccc; (ii) every -ideal
of A is order-closed; (iii) every -subalgebra of A is order-closed; (iv) every sequentially order-continuous
Boolean homomorphism from A to another Boolean algebra is order-continuous. (Hint: 313Q.)
64 Boolean algebras 316Xk
(k) Show that any principal ideal of a weakly (, )-distributive Boolean algebra is a weakly (, )-
distributive Boolean algebra.
(l) Let A
i
)
iI
be a family of Boolean algebras, with simple product A. Show that A is weakly (, )-
distributive i every A
i
is.
>>>(m) Show that if A is a weakly (, )-distributive Boolean algebra and B is a subalgebra of A which
is regularly embedded in A, then B is weakly (, )-distributive.
(n) Show that if A is a weakly (, )-distributive Boolean algebra and 1 is an order-closed ideal of A,
then A/1 is weakly (, )-distributive.
>>>(o) (i) Show that if A is a Boolean algebra and B is an order-dense subalgebra of A, then A is weakly
(, )-distributive i B is. (ii) Let X be a zero-dimensional compact Hausdor space. Show that the
regular open algebra of X is weakly (, )-distributive i the algebra of open-and-closed subsets of X is.
(p) Let A be a Boolean algebra. Show that it is weakly (, )-distributive i whenever C
n
)
nN
is a
sequence of partitions of unity in A, there is a partition C of unity such that for every c C, n N there
is a nite set I C
n
such that c supI.
>>>(q) Show that the algebra of open-and-closed subsets of 0, 1
N
, with its usual topology, is not weakly
(, )-distributive.
(r) Let A be a Boolean algebra and B an order-dense subalgebra of A. Show that A and B have the
same atoms, so that A is atomless, or purely atomic, i B is.
(s) Let A be a Boolean algebra and B a regularly embedded subalgebra of A. Show that (i) every atom
of A is included in an atom of B (ii) if A is purely atomic, so is B (iii) if B is atomless, so is A.
>>>(t) Let A be a Dedekind complete purely atomic Boolean algebra. Show that it is isomorphic to TA,
where A is the set of atoms of A.
(u) Let A be a Boolean algebra and 1 an order-closed ideal of A. Show that (i) if A is atomless, so is
A/1 (ii) if A is purely atomic, so is A/1.
(v) Let A be a Boolean algebra. Show that (i) if A is atomless, so is every principal ideal of A (ii) if A is
purely atomic, so is every principal ideal of A.
(w) Let A
i
)
iI
be a family of Boolean algebras with simple product A. Show that (i) A is purely atomic
i every A
i
is (ii) A is atomless i every A
i
is.
>>>(x) Show that any purely atomic Boolean algebra is weakly (, )-distributive.
(y) Let A be a Boolean algebra. Show that there is a one-to-one correspondence between atoms a of A
and order-continuous Boolean homomorphisms : A Z
2
, dened by saying that corresponds to a i
(a) = 1.
316Y Further exercises (a) Let I be any set. Show that 0, 1
I
, with its usual topology, is ccc.
(Hint: show that if E 0, 1
I
is a non-empty open-and-closed set, then E > 0, where is the usual
measure on 0, 1
I
.)
(b) Show that the Stone space of the regular open algebra of R is separable. More generally, show that
if a topological space X is separable so is the Stone space of its regular open algebra.
(c) Let A be a Boolean algebra and Z its Stone space. Show that A is ccc i every nowhere dense subset
of Z is included in a nowhere dense zero set.
316Ym Further topics 65
(d) Let X be a zero-dimensional topological space. Show that X is ccc i the regular open algebra of X
is ccc i the algebra of open-and-closed subsets of X is ccc.
(e) Set X = 0, 1

1
, and for <
1
set E

= x : x X, x() = 1. Let be the algebra of subsets of

X generated by E

: <
1
x : x X, and 1 the -ideal of generated by E

: < <

1
x : x X. Show that /1 is not ccc, but that there is no uncountable disjoint family in 1.
(f ) Let X be a regular topological space and G its regular open algebra. Show that G is weakly (, )-
distributive i every meager set in X is nowhere dense.
(g) Let A be a Boolean algebra. A is weakly -distributive if whenever a
mn
)
m,nN
is a double sequence
in A such that a
mn
)
nN
is non-increasing and has inmum 0 for every m N, then
infa : m Nn N, a
mn
a = 0.
(Dedekind complete weakly -distributive algebras are also called

-bounding.) A has the Egorov

property if whenever a
mn
)
m,nN
is a double sequence in A such that a
mn
)
nN
is non-increasing and has
inmum 0 for every m N, then there is a non-increasing sequence a
m
)
mN
such that inf
mN
a
m
= 0 and
for every m N there is an n N such that a
m
a
mn
. (i) Show that if A has the Egorov property it
is weakly -distributive. (ii) Show that if A is weakly (, )-distributive it is weakly -distributive. (iii)
Show that if A is ccc then it is weakly (, )-distributive i it has the Egorov property i it is weakly
-distributive. (iv) Show that T(N
N
) does not have the Egorov property, even though it is weakly (, )-
distributive. (Hint: try a
mn
= f : f(m) n.)
(h) Let A be a Boolean algebra and Z its Stone space. (i) Show that A is weakly -distributive i the
union of any sequence of nowhere dense zero sets in Z is nowhere dense. (ii) Show that A has the Egorov
property i the union of any sequence of nowhere dense zero sets in Z is included in a nowhere dense zero
set.
(i) Let A be a Dedekind -complete weakly (, )-distributive Boolean algebra, Z its Stone space, c the
algebra of open-and-closed subsets of Z, / the -ideal of meager subsets of Z, and the Baire property
algebra EM : E c, M /, as in 314M. (i) Suppose that f : Z R is a -measurable function.
Show that there is a dense open set G Z such that fG is continuous. (ii) Now suppose that A is Dedekind
complete. Show that if f : Z R is a function such that fG is continuous for some dense open set G Z,
then f is -measurable; and that if f is also bounded, there is a continuous function g : Z R such that
z : f(z) ,= g(z) is meager. (Hint: the graph of g will be the closure of the graph of fG; because Z is
extremally disconnected, this is the graph of a function.)
(j) (i) Let X be a non-empty separable Hausdor space without isolated points. Show that its regular
open algebra is not weakly (, )-distributive. (ii) Let (X, ) be a non-empty metric space without isolated
points. Show that its regular open algebra is not weakly (, )-distributive. (iii) Let I be any innite set.
Show that the algebra of open-and-closed subsets of 0, 1
I
is not weakly (, )-distributive. Show that the
regular open algebra of 0, 1
I
is not weakly (, )-distributive.
(k) For any set X, write
(
X
= I : I X is nite X I : I X is nite.
(i) Show that (
X
is an algebra of subsets of X (the nite-conite algebra). (ii) Show that a Boolean
algebra is purely atomic i it has an order-dense subalgebra isomorphic to the nite-conite algebra of
some set. (iii) Show that a Dedekind -complete Boolean algebra is purely atomic i it has an order-dense
subalgebra isomorphic to the countable-cocountable algebra of some set (211R).
(l) Let A
i
)
iI
be a family of Boolean algebras, none of them 0, with free product A. (i) Show that A
is purely atomic i every A
i
is purely atomic and i : A
i
,= 0, 1 is nite. (ii) Show that A is atomless i
either some A
i
is atomless or i : A
i
,= 0, 1 is innite.
(m) Show that a Boolean algebra is isomorphic to the algebra of open-and-closed subsets of 0, 1
N
i it
is countable, atomless and not 0.
66 Boolean algebras 316Yn
(n) Show that a Boolean algebra is isomorphic to the regular open algebra of R i it is atomless, Dedekind
complete, has a countable order-dense subalgebra and is not 0.
(o) Let G be the regular open algebra of R. Show that there is an injective Boolean homomorphism
: G TN. (Hint: the Stone space of G is separable.) Show that there is a Boolean homomorphism
: TN G such that is the identity on TN. (Hint: 314K.)
(p) Write [N]
<
for the ideal of TN consisting of the nite subsets of N. Show that TN/[N]
<
is atomless,
weakly (, )-distributive and not ccc.
(q) Let X be a Hausdor space and G its regular open algebra. (i) Show that the atoms of G are precisely
the sets x where x is an isolated point in X. (ii) Show that G is atomless i X has no isolated points.
(iii) Show that G is purely atomic i the set of isolated points in X is dense in X.
316 Notes and comments The phrase countable chain condition is perhaps unfortunate, since the
disjoint sets to which the denition 316A refers could more naturally be called antichains; but there is in
fact a connexion between countable chains and countable antichains (316Xg). While some authors speak of
the countable antichain condition or cac, the term ccc has become solidly established. In the Boolean
algebra context, it could equally well be called the countable sup property (316E).
The countable chain condition can be thought of as a restriction on the width of a Boolean algebra; it
means that the algebra cannot spread too far laterally (see 316Xc), though it may be indenitely complex
in other ways. Generally it means that in a wide variety of contexts we need look only at countable families
and monotonic sequences, rather than arbitrary families and directed sets (316E, 316F, 316Yg). Many of
the ideas of 316B-316F have already appeared in 215B; see 322G below.
I remarked in the notes to 313 that the distributive laws described in 313B have important generaliza-
tions, of which weak (, )-distributivity and its cousin weak -distributivity (316Yg) are two. They are
characteristic of the measure algebras which are the chief subject of this volume. The Egorov property
(316Yg) is an alternative formulation applicable to ccc spaces.
Of course every property of Boolean algebras has a reection in a topological property of their Stone
spaces; happily, the concepts of this section correspond to reasonably natural topological expressions (316B,
316I, 316M, 316Yh).
With four new properties (ccc, weakly (, )-distributive, atomless, purely atomic) to incorporate into
the constructions of the last few sections, a very large number of questions can be asked; most are elementary.
Any subalgebra of a ccc algebra is ccc (316Xa). All four properties are inherited by order-dense subalgebras
and principal ideals (316Xb, 316Xf, 316Xk, 316Xo, 316Xs, 316Xv); with the exception of the countable chain
condition (316Xc), they are inherited by simple products (316Xl, 316Xw); with the exception of atomlessness,
they are inherited by regularly embedded subalgebras (316Xm, 316Xs), and, in particular, by order-closed
subalgebras of Dedekind complete algebras. As for quotient algebras (equivalently, homomorphic images),
all four properties are inherited by order-continuous images (316Xi, 316Xn, 316Xu). The countable chain
condition is so important that it is worth noting that a sequentially order-continuous image of a ccc algebra
is ccc (316Xi), and that there is a useful necessary and sucient condition for a sequentially order-continuous
image of a -complete algebra to be ccc (316C, 316D, 316Xh; but see also 316Ye). To see that sequentially
order-continuous images do not inherit weak (, )-distributivity, recall that the regular open algebra of R
is isomorphic to the quotient of the Baire-property algebra

B of R by the meager ideal / (314Yd); but that

B is purely atomic (since it contains all singletons), therefore weakly (, )-distributive (316Xx). Similarly,
TN/[N]
<
is a non-ccc image of a ccc algebra (316Yp).
The denitions here provide a language in which a remarkably interesting question can be asked: is
the free product of ccc Boolean algebras always ccc? equivalently, is the product of ccc topological spaces
always ccc? What is special about this question is that it cannot be answered within the ordinary rules of
mathematics (even including the axiom of choice); it is undecidable, in the same way that the continuum
hypothesis is. I will deal with a variety of undecidable questions in Volume 5; this particular one is treated
in Jech 78 and Fremlin 84. Note that the free product of two weakly (, )-distributive algebras need
not be weakly (, )-distributive (325Yd).
I have taken the opportunity to mention three of the most important of all Boolean algebras: the algebra of
open-and-closed subsets of 0, 1
N
(316Ym), the regular open algebra of R (316K, 316Yn) and the quotient
316 Notes Further topics 67
TN/[N]
<
(316Yp). A fourth algebra which belongs in this company is the Lebesgue measure algebra,
which is atomless, ccc and weakly (, )-distributive (so that every countable subset of its Stone space Z
is nowhere dense, and Z is a non-separable ccc space); but for this I wait for the next chapter.
68 Measure algebras
Chapter 32
Measure algebras
I now come to the real work of this volume, the study of the Boolean algebras of equivalence classes of
measurable sets. In this chapter I work through the elementary theory, dening this to consist of the parts
which do not depend on Maharams theorem or the lifting theorem or non-trivial set theory.
321 gives the denition of measure algebra, and relates this idea to its origin as the quotient of a -
algebra of measurable sets by a -ideal of negligible sets, both in its elementary properties (following those of
measure spaces treated in 112) and in an appropriate version of the Stone representation. 322 deals with
the classication of measure algebras according to the scheme already worked out in 211 for measure spaces.
323 discusses the canonical topology and uniformity of a measure algebra. 324 contains results concerning
Boolean homomorphisms between measure algebras, with the relationships between topological continuity,
order-continuity and preservation of measure. 325 is devoted to the measure algebras of product measures,
and their abstract characterization. Finally, 326-327 address the properties of additive functionals on
Boolean algebras, generalizing the ideas of Chapter 23.
321 Measure algebras
I begin by dening measure algebra and relating this concept to the work of Chapter 31 and to the
elementary properties of measure spaces.
321A Denition A measure algebra is a pair (A, ), where A is a Dedekind -complete Boolean
algebra and : A [0, ] is a function such that
0 = 0;
whenever a
n
)
nN
is a disjoint sequence in A, (sup
nN
a
n
) =

n=0
a
n
;
a > 0 whenever a A and a ,= 0.
321B Elementary properties of measure algebras Corresponding to the most elementary proper-
ties of measure spaces (112C in Volume 1), we have the following basic properties of measure algebras. Let
(A, ) be a measure algebra.
(a) If a, b A and a b = 0 then (a b) = a + b. PPP Set a
0
= a, a
1
= b, a
n
= 0 for n 2; then
(a b) = (sup
nN
a
n
) =

n=0
a
n
= a + b. QQQ
(b) If a, b A and a b then a b. PPP
a a + (b \ a) = b. QQQ
(c) For any a, b A, (a b) a + b. PPP
(a b) = a + (b \ a) a + b. QQQ
(d) If a
n
)
nN
is any sequence in A, then (sup
nN
a
n
)

n=0
a
n
. PPPFor each n, set b
n
= a
n
\ sup
i<n
a
i
.
Inducing on n, we see that sup
in
a
i
= sup
in
b
i
for each n, so sup
nN
a
n
= sup
nN
b
n
and
(sup
nN
a
n
) = (sup
nN
b
n
) =

n=0
b
n

n=0
a
n
because b
n
)
nN
is disjoint. QQQ
(e) If a
n
)
nN
is a non-decreasing sequence in A, then (sup
nN
a
n
) = lim
n
a
n
. PPP Set b
0
= a
0
,
b
n
= a
n
\ a
n1
for n 1. Then
321F Measure algebras 69
(sup
nN
a
n
) = (sup
nN
b
n
) =

n=0
b
n
= lim
n
n

i=0
b
i
= lim
n
(sup
in
b
i
) = lim
n
a
n
. QQQ
(f ) If a
n
)
nN
is a non-increasing sequence in A and inf
nN
a
n
< , then (inf
nN
a
n
) = lim
n
a
n
.
PPP (Cf. 112Cf.) Set a = inf
nN
a
n
. Take k N such that a
k
< . Set b
n
= a
k
\ a
n
for n N; then b
n
)
nN
is non-decreasing and sup
nN
b
n
= a
k
\ a (313Ab). Because a
k
is nite,
a = a
k
(a
k
\ a) = a
k
lim
n
b
n
(by (e) above)
= lim
n
(a
k
\ b
n
) = lim
n
a
n
. QQQ
321C Proposition Let (A, ) be a measure algebra, and A A a non-empty upwards-directed set. If
sup
aA
a < , then sup A is dened in A and (supA) = sup
aA
a.
proof (Compare 215A.) Set = sup
aA
a, and for each n N choose a
n
A such that a
n
2
n
.
Next, choose b
n
)
nN
in A such that b
n+1
b
n
a
n
for each n, and set b = sup
nN
b
n
. Then
b = lim
n
b
n
, a
n
b for every n N,
so b = .
If a A, then for every n N there is an a
t
n
A such that a a
n
a
t
n
, so that
(a \ b) (a \ a
n
) (a
t
n
\ a
n
) = a
t
n
a
n
a
n
2
n
.
This means that (a b) = 0, so a \ b = 0 and a b. Accordingly b is an upper bound of A, and is therefore
supA; since we already know that b = , the proof is complete.
321D Corollary Let (A, ) be a measure algebra and A A a non-empty upwards-directed set. If
supA is dened in A, then (supA) = sup
aA
a.
proof If sup
aA
a = , this is trivial; otherwise it follows from 321C.
321E Corollary Let (A, ) be a measure algebra and A A a disjoint set. If supA is dened in A,
then (supA) =

aA
a.
proof If A = then supA = 0 and the result is trivial. Otherwise, set B = a
0
. . . a
n
: a
0
, . . . , a
n
A
are distinct. Then B is upwards-directed, and sup
bB
b =

aA
a because A is disjoint. Also B has the
same upper bounds as A, so sup B = supA and
(supA) = (supB) = sup
bB
b =

aA
a.
321F Corollary Let (A, ) be a measure algebra and A A a non-empty downwards-directed set. If
inf
aA
a < , then inf A is dened in A and (inf A) = inf
aA
a.
proof Take a
0
A with a
0
< , and set B = a
0
\ a : a A. Then B is upwards-directed, and
sup
bB
b a
0
< , so sup B is dened. Accordingly inf A = a
0
supB is dened (313Aa), and
(inf A) = a
0
(supB) = a
0
sup
bB
b
= inf
bB
(a
0
\ b) = inf
aA
(a
0
a) = inf
aA
a.
70 Measure algebras 321G
321G Subalgebras If (A, ) is a measure algebra, and B is a -subalgebra of A, then (B, B) is a
measure algebra. PPP As remarked in 314Eb, B is Dedekind -complete. If b
n
)
nN
is a disjoint sequence in
B, then the supremum b = sup
nN
b
n
is the same whether taken in B or A, so that we have b =

n=0
b
n
.
QQQ
321H The measure algebra of a measure space I introduce the abstract notion of measure alge-
bra because I believe that this is the right language in which to formulate the questions addressed in this
volume. However it is very directly linked with the idea of measure space, as the next two results show.
Proposition Let (X, , ) be a measure space, and ^ the ideal of -negligible sets. Let A be the Boolean
algebra quotient / ^. Then we have a functional : A [0, ] dened by setting
E

= E for every E ,
and (A, ) is a measure algebra. The canonical map E E

: A is sequentially order-continuous.
proof (a) By 314C, A is a Dedekind -complete Boolean algebra. By 313Qb, E E

is sequentially
order-continuous, because ^ is a -ideal of .
(b) If E, F and E

= F

in A, then EF ^, so
E F +(E F) = F E +(F E) = E
and E = F. Accordingly the given formula does indeed dene a function : A [0, ].
(c) Now
0 =

= = 0.
If a
n
)
nN
is a disjoint sequence in A, choose for each n N an E
n
such that E

n
= a
n
. Set F
n
=
E
n

i<n
E
i
; then
F

n
= E

n
\ sup
i<n
E

i
= a
n
\ sup
i<n
a
i
= a
n
for each n, so a
n
= F
n
for each n. Now set E =

nN
E
n
=

nN
F
n
; then E

= sup
nN
F

n
= sup
nN
a
n
.
So
(sup
nN
a
n
) = E =

n=0
F
n
=

n=0
a
n
.
Finally, if a ,= 0, then there is an E such that E

= a, and E / ^, so a = E > 0. Thus (A, ) is a

measure algebra.
321I Denition For any measure space (X, , ) I will call (A, ), as constructed above, the measure
algebra of (X, , ).
321J The Stone representation of a measure algebra Just as with Dedekind -complete Boolean
algebras (314N), every measure algebra is obtainable from the construction above.
Theorem Let (A, ) be any measure algebra. Then it is isomorphic, as measure algebra, to the measure
algebra of some measure space.
proof (a) We know from 314M that A is isomorphic, as Boolean algebra, to a quotient algebra // where
is a -algebra of subsets of the Stone space Z of A, and / is the ideal of meager subsets of Z. Let
: // A be the canonical isomorphism, and set E = E

for each E ; then : A is a

sequentially order-continuous surjective Boolean homomorphism with kernel /.
(b) For E , set
E = (E).
Then (Z, , ) is a measure space. PPP (i) We know already that is a -algebra of subsets of Z. (ii)
= () = 0 = 0.
(iii) If E
n
)
nN
is a disjoint sequence in , then (because is a Boolean homomorphism) E
n
)
nN
is a
disjoint sequence in A and (because is sequentially order-continuous) (

nN
E
n
) = sup
nN
E
n
; so
321 Notes Measure algebras 71
(

nN
E
n
) = (sup
nN
E
n
) =

n=0
(E
n
) =

n=0
E
n
. QQQ
(c) For E ,
E = 0 (E) = 0 E = 0 E /.
So the measure algebra of (Z, , ) is just //, with
E

= E = (E) = (E

)
for every E . Thus the Boolean algebra isomorphism is also an isomorphism between the measure
algebras (//, ) and (A, ), and (A, ) is represented in the required form.
321K Denition I will call the measure space (Z, , ) constructed in the proof of 321J the Stone
space of the measure algebra (A, ).
For later reference, I repeat the description of this space as developed in 311E, 311I, 314M and 321J. Z
is a compact Hausdor space, being the Stone space of A. A can be identied with the algebra of open-and-
closed sets in Z. The ideal of -negligible sets coincides with the ideal of meager subsets of Z; in particular,
is complete. The measurable sets are precisely those expressible in the form E = aM where a A,
a Z is the corresponding open-and-closed set, and M is meager; in this case E = a and a = E is the
member of A corresponding to E.
For the most important classes of measure algebras, more can be said; see 322M et seq. below.
321X Basic exercises (a) Let (A, ) be a measure algebra, and a A. Show that (A
a
, A
a
) is a
measure algebra, writing A
a
for the principal ideal of A generated by a.
(b) Let (X, , ) be a measure space, and A its measure algebra. (i) Show that if T is a -subalgebra
of , then E

: E T is a -subalgebra of A. (ii) Show that if B is a -subalgebra of A then E : E

, E

B is a -subalgebra of .
321Y Further exercises (a) Let (A, ) be a measure algebra, and I A a -ideal. For u A/I set
u = inf a : a A, a

= u. Show that (A, ) is a measure algebra.

321 Notes and comments The idea behind taking the quotient /^, where is the algebra of measurable
sets and ^ is the ideal of negligible sets, is just that if negligible sets can be ignored as is the case for a
very large proportion of the results of measure theory then two measurable sets can be counted as virtually
the same if they dier by a negligible set, that is, if they represent the same member of the measure algebra.
The denition in 321A is designed to be an exact characterization of these quotient algebras, taking into
account the measures with which they are endowed. In the course of the present chapter I will work through
many of the basic ideas dealt with in Volumes 1 and 2 to show how they can be translated into theorems
about measure algebras, as I have done in 321B-321F. It is worth checking these correspondences carefully,
because some of the ideas mutate signicantly in translation. In measure algebras, it becomes sensible to
take seriously the suprema and inma of uncountable sets (see 321C-321F).
I should perhaps remark that while the Stone representation (321J-321K) is signicant, it is not the most
important method of representing measure algebras, which is surely Maharams theorem, to be dealt with
in the next chapter. Nevertheless, the Stone representation is a canonical one, and will appear at each point
that we meet a new construction involving measure algebras, just as the ordinary Stone representation of
Boolean algebras can be expected to throw light on any aspect of Boolean algebra.
72 Measure algebras 322 intro.
322 Taxonomy of measure algebras
Before going farther with the general theory of measure algebras, I run through those parts of the
classication of measure spaces in 211 which have expressions in terms of measure algebras. The most
important concepts at this stage are those of semi-nite, localizable and -nite measure algebra (322Ac-
322Ae); these correspond exactly to the same terms applied to measure spaces (322B). I briey investigate
the Boolean-algebra properties of semi-nite and -nite measure algebras (322F, 322G), with mentions of
completions and c.l.d. versions (322D), subspace measures (322I-322J), direct sums of measure spaces (322K,
322L) and subalgebras of measure algebras (322M). It turns out that localizability of a measure algebra is
connected in striking ways to the properties of the canonical measure on its Stone space (322N). I end the
section with a description of the localization of a semi-nite measure algebra (322O-322P) and with some
further properties of Stone spaces (322Q).
322A Denitions Let (A, ) be a measure algebra.
(a) I will say that (A, ) is a probability algebra if 1 = 1.
(b) (A, ) is totally nite if 1 < .
(c) (A, ) is -nite if there is a sequence a
n
)
nN
in A such that a
n
< for every n N and
sup
nN
a
n
= 1. Note that in this case a
n
)
nN
can be taken either to be non-decreasing (consider a
t
n
=
sup
i<n
a
i
) or to be disjoint (consider a
tt
n
= a
n
\ a
t
n
).
(d) (A, ) is semi-nite if whenever a A and a = there is a non-zero b a such that b < .
(e) (A, ) is localizable or Maharam if it is semi-nite and the Boolean algebra A is Dedekind complete.
322B The rst step is to relate these concepts to the corresponding ones for measure spaces.
Theorem Let (X, , ) be a measure space, and (A, ) its measure algebra. Then
(a) (X, , ) is a probability space i (A, ) is a probability algebra;
(b) (X, , ) is totally nite i (A, ) is;
(c) (X, , ) is -nite i (A, ) is;
(d) (X, , ) is semi-nite i (A, ) is;
(e) (X, , ) is localizable i (A, ) is;
(f) if E , then E is an atom for i E

is an atom in A;
(g) (X, , ) is atomless i A is;
(h) (X, , ) is purely atomic i A is.
proof (a), (b) are trivial, since 1 = X.
(c)(i) If is -nite, let E
n
)
nN
be a sequence of sets of nite measure covering X; then E

n
< for
every n, and
sup
nN
E

n
= (

nN
E
n
)

= 1,
so (A, ) is -nite.
(ii) If (A, ) is -nite, let a
n
)
nN
be a sequence in A such that a
n
< for every n and sup
nN
a
n
= 1.
For each n, choose E
n
such that E

n
= a
n
. Set E =

nN
E
n
; then E

= sup
nN
a
n
= 1, so E is
conegligible. Now (X E, E
0
, E
1
, . . . ) is a sequence of sets of nite measure covering X, so is -nite.
(d)(i) Suppose that is semi-nite and that a A, a = . Then there is an E such that E

= a,
so that E = a = . As is semi-nite, there is an F such that F E and 0 < F < . Set
b = F

; then b a and 0 < b < .

(ii) Suppose that (A, ) is semi-nite and that E , E = . Then E

= , so there is a
b E

such that 0 < b < . Let F be such that F

= b. Then F E , F E E and
(F E)

= E

b = b, so that (F E) = b ]0, [.
322D Taxonomy of measure algebras 73
(e)(i) Note rst that if c and F , then
E F is negligible for every E c
E

= 0 for every E c
F

is an upper bound for E

: E c.
So if c and H , then H is an essential supremum of c in , in the sense of 211G, i H

is the
supremum of A = E

: E c in A. PPP Writing T for

F : F , E F is negligible for every E c,
we see that B = F

: F T is just the set of upper bounds of A, and that H is an essential supremum

of c i H T and H

is a lower bound for B; that is, i H

= sup A. QQQ
(ii) Thus A is Dedekind complete i every family in has an essential supremum in . Since we
already know that (A, ) is semi-nite i is, we see that (A, ) is localizable i is.
(f ) This is immediate from the denitions in 211I and 316L, if we remember always that b : b E

=
F

: F , F E (312Kb).
(g), (h) follow at once from (f).
322C I copy out the relevant parts of Theorem 211L in the new context.
Theorem (a) A probability algebra is totally nite.
(b) A totally nite measure algebra is -nite.
(c) A -nite measure algebra is localizable.
(d) A localizable measure algebra is semi-nite.
proof All except (c) are trivial; and (c) may be deduced from 211Lc-211Ld, 322Bc, 322Be and 321J, or
from 316Fa and 322G below.
322D Of course not all the denitions in 211 are directly relevant to measure algebras. The concepts
of complete, locally determined and strictly localizable measure space do not correspond in any direct
way to properties of the measure algebras. Indeed, completeness is just irrelevant, as the next proposition
shows.
Proposition Let (X, , ) be a measure space, with completion (X,

, ) and c.l.d. version (X,

, ) (213E).
Write (A, ), (A
1
,
1
) and (A
2
,
2
) for the measure algebras of , and respectively.
(a) The embedding

corresponds to an isomorphism between (A, ) and (A
1
,
1
).
(b)(i) The embedding

denes an order-continuous Boolean homomorphism : A A
2
. Setting
A
f
= a : a A, a < , A
f
is a measure-preserving bijection between A
f
and A
f
2
= c : c A
2
,
2
c <
.
(ii) is injective i is semi-nite, and in this case
2
(a) = a for every a A.
(iii) If is localizable, is a bijection.
proof For E , I write E

for its image in A; for F

, I write F

for its image in A

1
; and for G

, I
write F

for its image in A

2
.
(a) This is nearly trivial. The map E E

: A
1
is a Boolean homomorphism, being the composition
of the Boolean homomorphisms E E :

and F F

:

A
1
. Its kernel is E : E , E =
0 = E : E , E = 0, so it induces an injective Boolean homomorphism : A A
1
given by the
formula (E

) = E

for every E (312F, 3A2G). To see that is surjective, take any b A

1
. There is
an F

such that F

= b, and there is an E such that E F and (F E) = 0, so that

(E

) = E

= F

= b.
Thus is a Boolean algebra isomorphism. It is a measure algebra isomorphism because for any E

1
(E

) =
1
E

= E = E = E

.
74 Measure algebras 322D
(b)(i) The map E E

: A
2
is a Boolean homomorphism with kernel E : E , E = 0 E :
E , E = 0, so induces a Boolean homomorphism : A A
2
, dened by saying that E

= E

for
every E .
If a A
f
, it is expressible as E

where E < . Then E = E (213Fa), so a = E

belongs to A
f
2
,
and
2
(a) = a. If a, a
t
are distinct members of A
f
, then

2
(a .a
t
) =
2
(a .a
t
) = (a .a
t
) > 0,
so a ,= a
t
; thus A
f
is an injective map from A
f
to A
f
2
. If c A
f
2
, then c = G

where G < ; by
213Fc, there is an E such that E G, E = G and (G E) = 0, so that E

A
f
and
E

= E

= G

= c.
As c is arbitrary, [A
f
] = A
f
2
.
Finally, is order-continuous. PPP Let A A be a non-empty downwards-directed set with inmum 0,
and b A
2
a lower bound for [A]. ??? If b ,= 0, then (because (A
2
,
2
) is semi-nite) there is a b
0
A
f
2
such
that 0 ,= b
0
b. Let a
0
A be such that a
0
= b
0
. Then a
0
,= 0, so there is an a A such that a , a
0
, that
is, a a
0
,= a
0
. But now, because A
f
is injective,
b
0
= a
0
,= (a a
0
) = a a
0
= a b
0
,
and b
0
, a, which is impossible. XXX Thus b = 0, and 0 is the only lower bound of [A]. As A is arbitrary,
is order-continuous (313L(b-ii)). QQQ
(ii) () If is semi-nite, then E = E for every E (213Hc), so

2
(E

) =
2
E

= E = E = E

for every E . In particular,

a = 0 = 0 =
2
(a) = a = a = 0,
so is injective. () If is not semi-nite, there is an E such that E = but H = 0 whenever
H , H E and H < ; so that E = 0 and
E

,= 0, E

= E

= 0.
So in this case is not injective.
(iii) Now suppose that is localizable. Then for every G

there is an E such that (EG) = 0,
by 213Hb; accordingly E

= E

= G

. As G is arbitrary, is surjective; and we know from (ii) that is

injective, so it is a bijection, as claimed.
322E Proposition Let (A, ) be a measure algebra.
(a) (A, ) is semi-nite i it has a partition of unity consisting of elements of nite measure.
(b) If (A, ) is semi-nite, a = supb : b a, b < and a = sup b : b a, b < for every a A.
proof Set A
f
= b : b A, b < .
(a)(i) If (A, ) is semi-nite, then A
f
is order-dense in A, so there is a partition of unity consisting of
members of A
f
(313K).
(ii) If there is a partition of unity C A
f
, and a = , then there is a c C such that a c ,= 0, and
now a c a and 0 < (a c) < ; as a is arbitrary, (A, ) is semi-nite.
(b) Of course A
f
is upwards-directed, by 321Bc, and we are supposing that its supremum is 1. If a A,
then
B = b : b A
f
, b a = a b : b A
f

is upwards-directed and has supremum a (313Ba), so a = sup

bB
b, by 321D.
Remark Compare 213A.
322F Proposition If (A, ) is a semi-nite measure algebra, then A is a weakly (, )-distributive
Boolean algebra.
322J Taxonomy of measure algebras 75
proof Let A
n
)
nN
be a sequence of non-empty downwards-directed subsets of A, all with inmum 0. Set
A = sup
nN
a
n
: a
n
A
n
for every n N.
If c A 0, let b c be such that 0 < b < . For each n N, inf
aA
n
(b a
n
) = 0, by 321F; so we
may choose a
n
A
n
such that (b a
n
) 2
n2
b. Set a = sup
nN
a
n
A. Then
(b a)

n=0
(b a
n
) < b,
so b , a and c , a. As c is arbitrary, inf A = 0; as A
n
)
nN
is arbitrary, A is weakly (, )-distributive.
322G Corresponding to 215B, we have the following description of -nite algebras.
Proposition Let (A, ) be a semi-nite measure algebra. Then the following are equiveridical:
(i) (A, ) is -nite;
(ii) A is ccc;
(iii) either A = 0 or there is a functional : A [0, 1] such that (A, ) is a probability algebra.
proof (i) (ii) By 321J, it is enough to consider the case in which (A, ) is the measure algebra of a
measure space (X, , ), and is semi-nite, by 322Bd. We know that A is ccc i there is no uncountable
disjoint set in ^, where ^ is the ideal of negligible sets (316D). But 215B(iii) shows that this is equivalent
to being -nite, which is equivalent to (A, ) being -nite, by 322Bc.
(i)(iii) If (A, ) is -nite, and A ,= 0, let a
n
)
nN
be a disjoint sequence in A such that a
n
<
for every n and sup
nN
a
n
= 1. Then a
n
> 0 for some n, so there are
n
> 0 such that

n=0

n
a
n
= 1.
(Set
t
n
= 2
n
/(1 + a
n
),
n
=
t
n
/(

i=0

t
i
a
i
).) Set a =

n=0

n
(a a
n
) for every a A; it is easy to
check that (A, ) is a probability algebra.
(iii)(i) is a consequence of (ii)(i).
322H Principal ideals If (A, ) is a measure algebra and a A, then it is easy to see (using 314Eb)
that (A
a
, A
a
) is a measure algebra, where A
a
is the principal ideal of A generated by a.
322I Subspace measures General subspace measures give rise to complications in the measure algebra
(see 322Xg, 322Yd). But subspaces with measurable envelopes (132D, 213K) are manageable.
Proposition Let (X, , ) be a measure space, and A X a set with a measurable envelope E. Let
A
be
the subspace measure on A, and
A
its domain; let (A, ) be the measure algebra of (X, , ) and (A
A
,
A
)
the measure algebra of (A,
A
,
A
). Set a = E

and let A
a
be the principal ideal of A generated by a. Then
we have an isomorphism between (A
a
, A
a
) and (A
A
,
A
) given by the formula
F

(F A)

whenever F and F E, writing F

for the equivalence class of F in A and (F A)

for the equivalence

class of F A in A
A
.
proof Set
E
= E F : F . For F, G
E
,
F

= G

(FG) = 0
A
(A (FG)) = 0 (F A)

= (G A)

,
because E is a measurable envelope of A. Accordingly the given formula denes an injective function from
the image F

: F
E
of
E
in A to A
A
; but this image is just the principal ideal A
a
. It is easy to
check that the map is a Boolean homomorphism from A
a
to A
A
, and it is a Boolean isomorphism because

A
= F A : F
E
. Finally, it is measure-preserving because
F

= F =

(F A) =
A
(F A) =
A
(F A)

for every F
E
, again using the fact that E is a measurable envelope of A.
322J Corollary Let (X, , ) be a measure space, with measure algebra (A, ).
(a) If E , then the measure algebra of the subspace measure
E
can be identied with the principal
ideal A
E
of A.
(b) If A X is a set of full outer measure (in particular, if

A = X < ), then the measure algebra

of the subspace measure
A
can be identied with A.
76 Measure algebras 322K
322K Simple products (a) Let (A
i
,
i
))
iI
be any indexed family of measure algebras. Let A be the
simple product Boolean algebra

iI
A
i
(315A), and for a A set a =

iI

i
a(i). Then it is easy to
check (using 315D(e-ii)) that (A, ) is a measure algebra; I will call it the simple product of the family
(A
i
,
i
))
iI
. Each of the A
i
corresponds to a principal ideal A
e
i
say in A, where e
i
A corresponds to
1
A
i
A
i
(315E), and the Boolean isomorphism between A
i
and A
e
i
is a measure algebra isomorphism
between (A
i
,
i
) and (A
e
i
, A
e
i
).
(b) If (X
i
,
i
,
i
))
iI
is a family of measure spaces, with direct sum (X, , ) (214K), then the measure
algebra (A, ) of (X, , ) can be identied with the simple product of the measure algebras (A
i
,
i
) of
the (X
i
,
i
,
i
). PPP If, as in 214K, we set X = (x, i) : i I, x X
i
, and for E X, i I we set
E
i
= x : (x, i) E, then the Boolean isomorphism E E
i
)
iI
:

iI

i
induces a Boolean
isomorphism from A to

iI
A
i
, which is also a measure algebra isomorphism, because
E

= E =

iI

i
E
i
=

iI

i
E

i
for every E . QQQ
(c) A product of measure algebras is semi-nite, or localizable, or atomless, or purely atomic, i every
factor is. (Compare 214Jb.)
(d) Let (A, ) be a localizable measure algebra.
(i) If e
i
)
iI
is any partition of unity in A, then (A, ) is isomorphic to the product

iI
(A
e
i
, A
e
i
)
of the corresponding principal ideals. PPP By 315F(iii), the map a a e
i
)
iI
is a Boolean isomorphism
between A and

iI
A
i
. Because e
i
)
iI
is disjoint and a = sup
iI
a e
i
, a =

iI
(a e
i
) (321E), for
every a A. So a a e
i
)
iI
is a measure algebra isomorphism between (A, ) and

iI
(A
i
, A
e
i
). QQQ
(ii) In particular, since A has a partition of unity consisting of elements of nite measure (322Ea),
(A, ) is isomorphic to a simple product of totally nite measure algebras. Each of these is isomorphic to
the measure algebra of a totally nite measure space, so (A, ) is isomorphic to the measure algebra of a
direct sum of totally nite measure spaces, which is strictly localizable.
Thus every localizable measure algebra is isomorphic to the measure algebra of a strictly localizable
measure space. (See also 322N below.)
*322L Strictly localizable spaces The following fact is occasionally useful.
Proposition Let (X, , ) be a strictly localizable measure space with X > 0, and (A, ) its measure
algebra. If a
i
)
iI
is a partition of unity in A, there is a disjoint family X
i
)
iI
in , with union X, such
that X

i
= a
i
for every i I and
= E : E X, E X
i
i I,
E =

iI
(E X
i
) for every E ;
that is, the isomorphism between A and the simple product

iI
A
a
i
of its principal ideals (315F) corre-
sponds to an isomorphism between (X, , ) and the direct sum of the subspace measures on X
i
.
proof (a) Suppose to begin with that X < . In this case J = i : a
i
,= 0 must be countable (322G).
For each i J, choose E
i
such that E

i
= a
i
, and set F
i
= E
i

jJ,j,=i
E
j
; then F

i
= a
i
for each
i J, and F
i
)
iJ
is disjoint. Because X > 0, J is non-empty; x some j
0
J and set
X
i
= F
j
0
(X
_
jJ
F
j
) if i = j
0
,
= F
i
for i J j
0
,
= for i I J.
Then X
i
)
iI
is a disjoint family in ,

iI
X
i
= X and X

i
= a
i
for every i. Moreover, because only
countably many of the X
i
are non-empty, we certainly have
= E : E X, E X
i
i I,
322N Taxonomy of measure algebras 77
E =

iI
(E X
i
) for every E .
(b) For the general case, start by taking a decomposition Y
j
)
jJ
of X. We can suppose that no Y
j
is
negligible, because there is certainly some j
0
such that Y
j
0
> 0, and we can if necessary replace Y
j
0
by
Y
j
0

Y
j
: Y
j
= 0. For each j, we can identify the measure algebra of the subspace measure on Y
j
with
the principal ideal A
b
j
generated by b
j
= Y

j
(322I). Now a
i
b
j
)
iI
is a partition of unity in A
b
j
, so by (a)
just above we can nd a disjoint family X
ji
)
iI
in such that

iI
X
ji
= Y
j
, X

ji
= a
i
b
j
for every i and
TY
j
= E : E Y
j
, E X
ji
i I,
E =

iI
(E X
ji
) for every E TY
j
.
Set X
i
=

jI
X
ji
for every i I. Then X
i
)
iI
is disjoint and covers X. Because X
i
Y
j
= X
ji
is
measurable for every j, X
i
. Because X

i
a
i
b
j
for every j, and b
j
)
jJ
is a partition of unity in A
(322Kb), X

i
a
i
for each i; because X

i
)
iI
is disjoint and sup
iI
a
i
= 1, X

i
= a
i
for every i. If E X is
such that E X
i
for every i, then E X
ji
for all i I and j J, so E Y
j
for every j J
and E . If E , then
E =

jJ
(E Y
j
) =

jJ,iI
(E X
ji
) =

iI
sup
KJ is nite

jK
(E X
ji
)

iI
(E X
i
) = sup
KI is nite

iK
(E X
i
) E;
so E =

iI
(E X
i
). Thus X
i
)
iI
is a suitable family.
322M Subalgebras: Proposition Let (A, ) be a measure algebra, and B a -subalgebra of A. Set
= B.
(a) (B, ) is a measure algebra.
(b) If (A, ) is totally nite, or a probability algebra, so is (B, ).
(c) If (A, ) is -nite and (B, ) is semi-nite, then (B, ) is -nite.
(d) If (A, ) is localizable and B is order-closed and (B, ) is semi-nite, then (B, ) is localizable.
(e) If (B, ) is a probability algebra, or totally nite, or -nite, so is (A, ).
proof (a) By 314Eb, B is Dedekind -complete, and the identity map : B A is sequentially order-
continuous; so that = will be countably additive and (B, ) will be a measure algebra.
(b) This is trivial.
(c) Use 322G. Every disjoint subset of B is disjoint in A, therefore countable, because A is ccc; so B is
also ccc and (B, ) (being semi-nite) is -nite.
(d) By 314Ea, B is Dedekind complete; we are supposing that (B, ) is semi-nite, so it is localizable.
(e) This is elementary.
322N The Stone space of a localizable measure algebra I said above that the concepts of strictly
localizable and locally determined measure space have no equivalents in the theory of measure algebras.
But when we look at the canonical measure on the Stone space of a measure algebra, we can of course hope
that properties of the measure algebra will be reected in the properties of this measure, as happens in the
next theorem.
Theorem Let (A, ) be a measure algebra, Z the Stone space of A, and the standard measure on Z
constructed by the method of 321J-321K. Then the following are equiveridical:
(i) (A, ) is localizable;
(ii) is localizable;
(iii) is locally determined;
(iv) is strictly localizable.
proof Write for the domain of , that is,
78 Measure algebras 322N
EA : E Z is open-and-closed, A Z is meager,
and / for the ideal of meager subsets of Z, that is, the ideal of -negligible sets (314M, 321K). Then
a a

: A // is an isomorphism between (A, ) and the measure algebra of (Z, , ) (314M). Note
that because any subset of a meager set is meager, is surely complete.
(a)(i) (ii) is a consequence of 322Be.
(b)(ii)(iii) Suppose that is localizable. Of course it is semi-nite. Let V Z be a set such that
V E whenever E and E < . Because is localizable, there is a W which is an essential
supremum in of V E : E , E < , that is, W

= sup(V E)

: E < in //. I claim

that WV is nowhere dense. PPP Let G Z be a non-empty open set. Then there is a non-zero a A such
that a G. Because (A, ) is semi-nite, we may suppose that a < . Now
(W a)

= W

= sup
E<
(V E)

= sup
E<
(V E a)

= (V a)

,
so (WV ) a is negligible, therefore meager. But we know that A is weakly (, )-distributive (322F), so
that meager sets in Z are nowhere dense (316I), and there is a non-empty open set H a (WV ). Now
H G WV . As G is arbitrary, int WV = and WV is nowhere dense. QQQ
But this means that WV / and V = W(WV ) . As V is arbitrary, is locally
determined.
(c)(iii)(iv) Assume that is locally determined. Because (A, ) is semi-nite, there is a partition of
unity C A consisting of elements of nite measure (322Ea). Set ( = c : c C. This is a disjoint family
of sets of nite measure for . Now suppose that F and F > 0. Then there is an open-and-closed set
E Z such that FE is meager, and E is of the form a for some a A. Since
a = a = F > 0,
there is some c C such that a c ,= 0, and now
(F c) = (a c) > 0.
This means that satises the conditions of 213O and must be strictly localizable.
(d)(iv)(ii) This is just 211Ld.
322O Proposition Let (A, ) be a semi-nite measure algebra, and let

A be the Dedekind completion
of A (314U). Then there is a unique extension of to a functional on

A such that (

A, ) is a localizable
measure algebra. The embedding A

A identies the ideals a : a A, a < and a : a

A, a < .
proof (I write the argument out as if A were actually a subalgebra of

A.) For c

A, set
c = sup a : a A, a c.
Evidently is a function from

A to [0, ] extending , so 0 = 0. Because A is order-dense in

A, c > 0
whenever c ,= 0, because any such c includes a non-zero member of A. If c
n
)
nN
is a disjoint sequence in

A
with supremum c, then c =

n=0
c
n
. PPP Let A be the set of all members of A expressible as a = sup
nN
a
n
where a
n
A, a
n
c
n
for every n N. Now
sup
aA
a = sup

n=0
a
n
: a
n
A, a
n
c
n
for every n N
=

n=0
sup a
n
: a
n
c
n
=

n=0
c
n
.
Also, because A is order-dense in

A, c
n
= supa : a A, a c
n
for each n, and supA, taken in

A, must be
c. But this means that if a
t
A, a
t
c then a
t
= sup
aA
a
t
a in

A and therefore also in A; so that
a
t
= sup
aA
(a
t
a) sup
aA
a.
Accordingly
322Q Taxonomy of measure algebras 79
c = sup
aA
a =

n=0
c
n
. QQQ
This shows that (

A, ) is a measure algebra. It is semi-nite because (A, ) is and every non-zero element

of

A includes a non-zero element of A, which in turn includes a non-zero element of nite measure. Since

A
is Dedekind complete, (

A, ) is localizable.
If a is nite, then surely a = a is nite. If c is nite, then A : a A, a c is upwards-directed
and sup
aA
A = c is nite, so b = supA is dened in A and b = c. Because A is order-dense in

A,
b = c (313K, 313O) and c A, with c = c.
322P Denition Let (A, ) be any semi-nite measure algebra. I will call (

A, ), as constructed above,
the localization of (A, ). Of course it is unique just in so far as the Dedekind completion of A is.
322Q Further properties of Stone spaces: Proposition Let (A, ) be a semi-nite measure algebra
and (Z, , ) its Stone space.
(a) Meager sets in Z are nowhere dense; every E is uniquely expressible as GM where G Z is
open-and-closed and M is nowhere dense, and E = supH : H E is open-and-closed.
(b) The c.l.d. version of is strictly localizable, and has the same negligible sets as .
(c) If (A, ) is totally nite then E = infH : H E is open-and-closed for every E .
proof (a) I have already remarked (in the proof of 322N) that A is weakly (, )-distributive, so that
meager sets in Z are nowhere dense. But we know that every member of is expressible as GM where G
is open-and-closed and M is meager, therefore nowhere dense. Moreover, the expression is unique, because
if GM = G
t
M
t
then GG
t
M M
t
is open and nowhere dense, therefore empty, so G = G
t
and
M = M
t
.
Now let a A be such that a = G, and consider B = b : b A,

b E. Then sup B = a in A. PPP If

b B, then

b a M is nowhere dense, therefore empty; so a is an upper bound for B. ??? If a is not the
supremum of B, then there is a non-zero c a such that b a \ c for every b B. But now c cannot be
empty, so c M is non-empty, and there is a non-zero d A such that

d c M. In this case d B and
d , a \ c. XXX Thus a = sup B. QQQ
It follows that
E = G = a = sup
bB
b
= sup
bB

b supH : H E is open-and-closed E
and E = supH : H E is open-and-closed.
(b) This is the same as part (c) of the proof of 322N. We have a disjoint family ( of sets of nite measure
for such that whenever E , E > 0 there is a C ( such that (C E) > 0. Now if F is dened
and not 0, there is an E such that E F and E > 0 (213Fc), so that there is a C ( such that
(E C) > 0; since C < , we have
(F C) (E C) = (E C) > 0.
And of course C < for every C (. This means that ( witnesses that satises the conditions of 213O,
so that is strictly localizable.
Any -negligible set is surely -negligible. If M is -negligible then it is nowhere dense. PPP If G Z is
open and not empty then there is a non-empty open-and-closed set H
1
G, and now H
1
, so there is a
non-empty open-and-closed set H H
1
such that H is nite (because is semi-nite). In this case HM
is -negligible, therefore nowhere dense, and H , M. But this means that G , M; as G is arbitrary, M is
nowhere dense. QQQ Accordingly M / and is -negligible.
Thus and have the same negligible sets.
(c) Because Z < ,
80 Measure algebras 322Q
E = Z (Z E) = Z supH : H Z E is open-and-closed
= inf(Z H) : H Z E is open-and-closed
= infH : H E is open-and-closed.
322X Basic exercises >>>(a) Let (A, ) be a measure algebra. Let I

be the set of those a A which

are purely innite, that is, a = and b = for every non-zero b a. Show that I

is a -ideal of A.
Show that there is a function
sf
: A/I

[0, ] dened by setting

sf
a

= sup b : b a, b < for

every a A. Show that (A/I

,
sf
) is a semi-nite measure algebra.
(b) Let (X, , ) be a measure space and let
sf
be the semi-nite version of , as dened in 213Xc.
Let (A, ) be the measure algebra of (X, , ). Show that the measure algebra of (X, ,
sf
) is isomorphic
to the measure algebra (A/I

,
sf
) of (a) above.
(c) Let (X, , ) be a measure space and (X,

, ) its c.l.d. version. Let (A, ) and (A
2
,
2
) be the
corresponding measure algebras, and : A A
2
the canonical homomorphism, as in 322Db. Show that
the kernel of is the ideal I

, as described in 322Xa, so that A/I

is isomorphic, as Boolean algebra, to

[A] A
2
. Show that this isomorphism identies
sf
, as described in 322Xa, with
2
[A].
(d) Give a direct proof of 322G, not relying on 215B and 321J.
>>>(e) Let (A, ) be any measure algebra, A a non-empty subset of A, and c A such that c < .
Show that (i) c
0
= supa c : a A is dened in A (ii) there is a countable set B A such that
c
0
= supa c : a B.
(f ) Let (X, , ) be a measure space and an indenite-integral measure over (234). Show that the
measure algebra of can be identied, as Boolean algebra, with a principal ideal of the measure algebra of
.
(g) Let (X, , ) be a measure space and A any subset of X; let
A
be the subspace measure on A and

A
its domain. Write (A, ) for the measure algebra of (X, , ) and (A
A
,
A
) for the measure algebra
of (A,
A
,
A
). Show that the formula F

(F A)

denes a sequentially order-continuous Boolean

homomorphism : A A
A
which has kernel I = F

: F , F A = . Show that for any a A,

A
(a) = min b : b A, a \ b I.
(h) Let (A, ) be a measure algebra and B an order-closed subalgebra of A. Suppose that (B, B) is
semi-nite. Show that (A, ) is semi-nite.
(i) Let (A, ) be any measure algebra and (Z, , ) its Stone space. Show that the c.l.d. version of is
strictly localizable.
322Y Further exercises (a) Let X be a set, a -algebra of subsets of X, and 1 a -ideal of . Set
^ = N : F 1, N F. Show that ^ is a -ideal of subsets of X. Set

= EN : E , N ^.
Show that

is a -algebra of subsets of X and that

/^ is isomorphic to /1.
(b) Let (A, ) be a semi-nite measure algebra, and (Z, , ) its Stone space. Let be the c.l.d. version
of , and

its domain. Show that

is precisely the Baire property algebra GA : G Z is open, A Z
is meager, so that

// can be identied with the regular open algebra of Z (314Yd) and the measure
algebra of can be identied with the localization of A.
(c) Give an example of a localizable measure algebra (A, ) with a -subalgebra B such that (B, B)
is semi-nite and atomless, but A is not atomless.
(d) Let (X, , ) be a measure space and A X a subset; let
A
be the subspace measure on A, A
and A
A
the measure algebras of and
A
, and : A A
A
the canonical homomorphism, as described in
322Xg. (i) Show that if
A
is semi-nite, then is order-continuous. (ii) Show that if is semi-nite but

A
is not, then is not order-continuous.
322 Notes Taxonomy of measure algebras 81
(e) Show that if (A, ) is a measure algebra, with Stone space (Z, , ), then has locally determined
negligible sets in the sense of 213I.
(f ) Let (A, ) be a localizable measure algebra and (Z, , ) its Stone space. (i) Show that a function
f : Z R is -measurable i there is a conegligible set G X such that fG is continuous. (Hint: 316Yi.)
(ii) Show that f : Z [0, 1] is -measurable i there is a continuous function g : Z [0, 1] such that
f = g -a.e.
322 Notes and comments I have taken this leisurely tour through the concepts of Chapter 21 partly to
recall them (or persuade you to look them up) and partly to give you practice in the elementary manipulations
of measure algebras. The really vital result here is the correspondence between localizability in measure
spaces and measure algebras. Part of the object of this volume (particularly in Chapter 36) is to try to make
sense of the properties of localizable measure spaces, as discussed in Chapter 24 and elsewhere, in terms of
their measure algebras. I hope that 322Be has already persuaded you that the concept really belongs to
measure algebras, and that the formulation in terms of essential suprema is a dispensable expedient.
I have given proofs of 322C and 322G depending on the realization of an arbitrary measure algebra as
the measure algebra of a measure space, and the corresponding theorems for measure spaces, because this
seems the natural approach from where we presently stand; but I am sympathetic to the view that such
proofs must be inappropriate, and that it is in some sense better style to look for arguments which speak
only of measure algebras (322Xd).
For any measure algebra (A, ), the set A
f
of elements of nite measure is an ideal of A; consequently
it is order-dense i it includes a partition of unity (322E). In 322F we have something deeper: any semi-
nite measure algebra must be weakly (, )-distributive when regarded as a Boolean algebra, and this has
signicant consequences in its Stone space, which are used in the proofs of 322N and 322Q. Of course a
result of this kind must depend on the semi-niteness of the measure algebra, since any Dedekind -complete
Boolean algebra becomes a measure algebra if we give every non-zero element the measure . It is natural
to look for algebraic conditions on a Boolean algebra sucient to make it measurable, in the sense that it
should carry a semi-nite measure; this is an unresolved problem to which I will return in Chapter 39.
Subspace measures, simple products, direct sums, principal ideals and order-closed subalgebras give no
real surprises; I spell out the details in 322I-322M and 322Xg-322Xh. It is worth noting that completing a
measure space has no eect on its measure algebra (322D, 322Ya). We see also that from the point of view
of measure algebras there is no distinction to be made between localizable and strictly localizable, since
every localizable measure algebra is representable as the measure algebra of a strictly localizable measure
space (322Kd). (But strict localizability does have implications for some processes starting in the measure
algebra; see 322L.) It is nevertheless remarkable that the canonical measure on the Stone space of a semi-
nite measure algebra is localizable i it is strictly localizable (322N). This canonical measure has many
other interesting properties, which I skim over in 322Q, 322Xi, 322Yb and 322Yf. In Chapter 21 I discussed
a number of methods of improving measure spaces, notably completions (212C) and c.l.d. versions (213E).
Neither of these is applicable in any general way to measure algebras. But in fact we have a more eective
construction, at least for semi-nite measure algebras, that of localization (322O-322P); I say that it is
more eective just because localizability is more important than completeness or local determinedness, being
of vital importance in the behaviour of function spaces (241Gb, 243Gb, 245Ec, 363M, 364O, 365J, 367N,
369A, 369C). Note that the localization of a semi-nite measure algebra does in fact correspond to the c.l.d.
version of a certain measure (322Q, 322Yb). But of course A and

A do not have the same Stone spaces,
even when

A can be eectively represented as the measure algebra of a measure on the Stone space of A.
What is happening in 322Yb is that we are using all the open sets of Z to represent members of

A, not just
the open-and-closed sets, which correspond to members of A.
82 Measure algebras 323 intro.
323 The topology of a measure algebra
I take a short section to discuss one of the fundamental tools for studying totally nite measure algebras,
the natural metric that each carries. The same ideas, suitably adapted, can be applied to an arbitrary
measure algebra, where we have a topology corresponding closely to the topology of convergence in measure
on the function space L
0
. Most of the section consists of an analysis of the relations between this topology
and the order structure of the measure algebra.
323A The pseudometrics
a
(a) Let (A, ) be a measure algebra. Write A
f
= a : a A, a < .
For a A
f
and b, c A, write
a
(b, c) = (a (b .c)). Then
a
is a pseudometric on A. PPP (i) Because
a < ,
a
takes values in [0, [. (ii) If b, c, d A then b .d (b .c) (c .d), so

a
(b, d) = (a (b .d)) ((a (b .c)) (a (c .d)))
(a (b .c)) + (a (c .d)) =
a
(b, c) +
a
(c, d).
(iii) If b, c A then

a
(b, c) = (a (b .c)) = (a (c .b)) =
a
(c, b). QQQ
(b) Now the measure-algebra topology of the measure algebra (A, ) is that generated by the family
P =
a
: a A
f
of pseudometrics on A. Similarly the measure-algebra uniformity on A is that
generated by P.
(For a general discussion of topologies dened by pseudometrics, see 2A3F et seq. For the associated
uniformities see 3A4.)
(c) Note that P is upwards-directed, since
aa
max(
a
,
a
) for all a, a
t
A
f
.
(d) When (A, ) is totally nite, it is more natural to work from the measure metric =
1
, with
(a, b) = (a .b), since this by itself denes the measure-algebra topology and uniformity.
*(e) Even when (A, ) is not totally nite, we still have a metric on A
f
dened by setting (a, b) =
(a .b) for all a, b A
f
, which is sometimes useful (323Xg). Note however that the topology on A
f
dened
from is not in general the topology induced by the measure-algebra topology of A.
323B Proposition Let (A, ) be any measure algebra. Then the operations , , \ and . are all
uniformly continuous.
proof The point is that for any b, c, b
t
, c
t
A we have
(b c) .(b
t
c
t
) (b .b
t
) (c .c
t
)
for any of the operations = , etc.; so that if a A
f
then

a
(b c, b
t
c
t
)
a
(b, b
t
) +
a
(c, c
t
).
Consequently the operation must be uniformly continuous.
323C Proposition (a) Let (A, ) be a totally nite measure algebra. Then : A [0, [ is uniformly
continuous.
(b) Let (A, ) be a semi-nite measure algebra. Then : A [0, ] is lower semi-continuous.
(c) Let (A, ) be any measure algebra. If a A and a < , then b (b a) : A R is uniformly
continuous.
proof (a) For any a, b A,
[ a b[ (a .b) =
1
(a, b).
(b) Suppose that b A and b > R. Then there is an a b such that < a < (322Eb). If
c A is such that
a
(b, c) < a , then
323D The topology of a measure algebra 83
c (a c) = a (a (b \ c)) > .
Thus b : b > is open; as is arbitrary, is lower semi-continuous.
(c) [ (a .b) (a .c)[
a
(b, c) for all b, c A.
323D The following facts are basic to any understanding of the relationship between the order structure
and topology of a measure algebra.
Lemma Let (A, ) be a measure algebra.
(a) Let B A be a non-empty upwards-directed set. For b B set F
b
= c : b c B.
(i) F
b
: b B generates a Cauchy lter T(B) on A.
(ii) If sup B is dened in A, then it is a topological limit of T(B ); in particular, it belongs to the
topological closure of B.
(b) Let B A be a non-empty downwards-directed set. For b B set F
b
= c : b c B.
(i) F
b
: b B generates a Cauchy lter T(B) on A.
(ii) If inf B is dened in A, then it is a topological limit of T(B); in particular, it belongs to the
topological closure of B.
(c)(i) Closed subsets of A are order-closed in the sense of 313D.
(ii) An order-dense subalgebra of A must be dense in the topological sense.
(d) Now suppose that (A, ) is semi-nite.
(i) The sets b : b c, b : b c are closed for every c A.
(ii) If B A is non-empty and upwards-directed and e is a cluster point of T(B), then e = sup B.
(iii) If B A is non-empty and downwards-directed and e is a cluster point of T(B), then e = inf B.
proof I use the notations A
f
,
a
from 323A.
(a)(i) () If b, c B then there is a d B such that b c d, so that F
d
F
b
F
c
; consequently
T(B) = F : F A, b B, F
b
F
is a lter on A. () Let a A
f
, > 0. Then there is a b B such that (a c) (a b) +
1
2
for every
c B, and F
b
T(B). If now c, c
t
F
b
, c .c
t
(c \ b) (c
t
\ b), so

a
(c, c
t
) (a c \ b) + (a c
t
\ b) = (a c) + (a c
t
) 2 (a b) .
As a and are arbitrary, T(B) is Cauchy.
(ii) Suppose that e = supB is dened in A. Let a A
f
, > 0. By 313Ba, a e = sup
bB
a b;
but a b : b B is upwards-directed, so (a e) = sup
bB
(a b), by 321D. Let b B be such that
(a b
0
) (a e) . Then for any c F
b
, e .c e \ b, so

a
(e, c) = (a (e .c)) (a (e \ b)) = (a e) (a b) .
As a and are arbitrary, T(B) e.
Because B T(B), e surely belongs to the topological closure of B.
(b) Either repeat the arguments above, with appropriate inversions, using 321F in place of 321D, or
apply (a) to the set 1 \ b : b B.
(c)(i) This follows at once from (a) and (b) and the denition in 313D.
(ii) If B A is an order-dense subalgebra and a A, then B = b : b B, b a is upwards-directed
and has supremum a (313K); by (a-ii), a B B. As a is arbitrary, B is topologically dense.
(d)(i) Set F = b : b c. If d A F, then (because (A, ) is semi-nite) there is an a A
f
such that
= (a d \ c) > 0; now if b F,

a
(d, b) (a d \ b) ,
so that d cannot belong to the closure of F. As d is arbitrary, F is closed. Similarly, b : b c is closed.
(ii) () If b B, then e F
b
, because F
b
T(B); but c : b c A is a closed set including F
b
,
so contains e, and b e. As b is arbitrary, e is an upper bound for B. () If d is an upper bound of B,
84 Measure algebras 323D
then c : c d is a closed set belonging to T(B), so contains e. As d is arbitrary, this shows that e is the
supremum of B, as claimed.
(iii) Use the same arguments as in (ii), but inverted.
323E Corollary Let (A, ) be a measure algebra.
(a) If b
n
)
nN
is a non-decreasing sequence in A with supremum b, then b
n
)
nN
b for the measure-
algebra topology.
(b) If b
n
)
nN
is a non-increasing sequence in A with inmum b, then b
n
)
nN
b for the measure-algebra
topology.
proof I call this a corollary because it is the special case of 323Da-323Db in which B is the set of terms of
a monotonic sequence; but it is probably easier to work directly from the denition in 323A, and use 321Be
or 321Bf to see that lim
n

a
(b
n
, b) = 0 whenever a < .
323F The following is a useful calculation.
Lemma Let (A, ) be a measure algebra and c
n
)
nN
a sequence in A such that the sum

n=0
(c
n
.c
n+1
)
is nite. Set d
0
= sup
nN
inf
mn
c
m
, d
1
= inf
nN
sup
mn
c
m
. Then d
0
= d
1
and, writing d for their common
value, lim
n
(c
n
.d) = 0.
proof Write
n
= (c
n
.c
n+1
),
n
=

k=n

k
for n N; we are supposing that lim
n

n
= 0. Set
b
n
= sup
mn
c
m
.c
m+1
; then
b
n

m=n
(c
m
.c
m+1
) =
n
for each n. If m n, then
c
m
.c
n
sup
nk<m
c
k
.c
k+1
b
n
,
so
c
n
\ b
n
c
m
c
n
b
n
.
Consequently
c
n
\ b
n
inf
km
c
k
sup
km
c
k
c
n
b
n
for every m n, and
c
n
\ b
n
d
0
d
1
c
n
b
n
,
so that
c
n
.d
0
b
n
, c
n
.d
1
b
n
, d
1
\ d
0
b
n
.
As this is true for every n,
lim
n
(c
n
.d
i
) lim
n
b
n
= 0
for both i, and
(d
1
.d
0
) inf
nN
b
n
= 0,
so that d
1
= d
0
.
323G The classication of measure algebras: Theorem Let (A, ) be a measure algebra, T its
measure-algebra topology and | its measure-algebra uniformity.
(a) (A, ) is semi-nite i T is Hausdor.
(b) (A, ) is -nite i T is metrizable, and in this case | is also metrizable.
(c) (A, ) is localizable i T is Hausdor and A is complete under |.
proof I use the notations A
f
,
a
from 323A.
(a)(i) Suppose that (A, ) is semi-nite and that b, c are distinct members of A. Then there is an
a b .c such that 0 < a < , and now
a
(b, c) > 0. As b and c are arbitrary, T is Hausdor (2A3L).
323G The topology of a measure algebra 85
(ii) Suppose that T is Hausdor and that b A has b = . Then b ,= 0 so there must be an a A
f
such that (a b) =
a
(0, b) > 0; in which case a b b and 0 < (a b) < . As b is arbitrary, is
semi-nite.
(b)(i) Suppose that is -nite. Let a
n
)
nN
be a non-decreasing sequence in A
f
with supremum 1. Set
(b, c) =

n=0

a
n
(b, c)
1 + 2
n
a
n
for b, c A. Then is a metric on A, because if (b, c) = 0 then a
n
(b .c) = 0 for every n, so b .c = 0
and b = c.
If a A
f
and > 0, take n such that (a \ a
n
)
1
2
. If b, c A and (b, c) /2(1 + 2
n
a
n
), then

a
n
(b, c)
1
2
so

a
(b, c) =
a\a
n
(b, c) +
aa
n
(b, c) (a \ a
n
) +
a
n
(b, c)

1
2
+ (1 + 2
n
a
n
)(b, c) .
In the other direction, given > 0, take n N such that 2
n

1
2
; then (b, c) whenever
a
n
(b, c)
/2(n + 1).
This shows that | is the same as the metrizable uniformity dened by ; accordingly T is also dened
by .
(ii) Now suppose that T is metrizable, and let be a metric dening T. For each n N there must be
a
n0
, . . . , a
nk
n
A
f
and
n
> 0 such that

a
ni
(b, 1)
n
for every i k
n
= (b, 1) 2
n
.
Set d = sup
nN,ik
n
a
ni
. Then
a
ni
(d, 1) = 0 for every n, i, so (d, 1) 2
n
for every n and d = 1. Thus 1
is the supremum of countably many elements of nite measure and (A, ) is -nite.
(c)(i) Suppose that (A, ) is localizable. Then T is Hausdor, by (a). Let T be a Cauchy lter on A. For
each a A
f
, choose a sequence F
n
(a))
nN
in T such that
a
(b, c) 2
n
whenever b, c F
n
(a) and n N.
Choose c
an

kn
F
k
(a) for each n; then
a
(c
an
, c
a,n+1
) 2
n
for each n. Set d
a
= sup
nN
inf
kn
a c
ak
.
Then
lim
n

a
(d
a
, c
an
) = lim
n
(d
a
.(a c
an
)) = 0,
by 323F.
If a, b A
f
and a b, then d
a
= a d
b
. PPP For each n N, F
n
(a) and F
n
(b) both belong to T, so must
have a point e in common; now

a
(d
a
, d
b
)
a
(d
a
, c
an
) +
a
(c
an
, e) +
a
(e, c
bn
) +
a
(c
bn
, d
b
)

a
(d
a
, c
an
) +
a
(c
an
, e) +
b
(e, c
bn
) +
b
(c
bn
, d
b
)

a
(d
a
, c
an
) + 2
n
+ 2
n
+
b
(c
bn
, d
b
)
0 as n .
Consequently
a
(d
a
, d
b
) = 0, that is,
d
a
= a d
a
= a d
b
. QQQ
Set d = supd
b
: b A
f
; this is dened because A is Dedekind complete. Then T d. PPP If a A
f
and
> 0, then
a d = sup
bA
f a d
b
= sup
bA
f a b d
ab
= sup
bA
f a b d
a
= a d
a
.
So if we choose n N such that 2
n
+
a
(c
an
, d
a
) , then for any e F
n
(a) we shall have

a
(e, d)
a
(e, c
an
) +
a
(c
an
, d) 2
n
+
a
(c
an
, d
a
) .
Thus
86 Measure algebras 323G
e :
a
(d, e) F
n
(a) T.
As a, are arbitrary, T converges to d. QQQ As T is arbitrary, A is complete.
(ii) Now suppose that T is Hausdor and that A is complete under |. By (a), (A, ) is semi-nite.
Let B be any non-empty subset of A, and set B
t
= b
0
. . . b
n
: b
0
, . . . , b
n
B, so that B
t
is upwards-
directed and has the same upper bounds as B. By 323Da, we have a Cauchy lter T(B
t
); because A is
complete, this is convergent; and because (A, ) is semi-nite, its limit must be supB
t
= sup B, by 323Dd.
As B is arbitrary, A is Dedekind complete, so (A, ) is localizable.
323H Closed subalgebras The ideas used in the proof of (c) above have many other applications, of
which one of the most important is the following. You may nd it helpful to read the next theorem rst on
the assumption that (A, ) is a probability algebra.
Theorem Let (A, ) be a localizable measure algebra, and B a subalgebra of A. Then it is closed for the
measure-algebra topology i it is order-closed.
proof (a) If B is closed, it must be order-closed, by 323Dc.
(b) Now suppose that B is order-closed. I repeat the ideas of part (c-i) of the proof of 323G. Let e be
any member of the closure of B in A. For each a A
f
, n N choose c
an
B such that
a
(c
an
, e) 2
n
.
Then

n=0
((a c
an
) .(a c
a,n+1
)) =

n=0

a
(c
an
, c
a,n+1
)

n=0

a
(c
an
, e) +
a
(e, c
a,n+1
) < .
So if we set e
a
= sup
nN
inf
kn
c
ak
, then

a
(e
a
, c
an
) =
a
(a e
a
, a c
an
) 0
as n , by 323F, and
a
(e, e
a
) = 0, that is, a e
a
= a e. Also, because Bis order-closed, inf
kn
c
ak
B
for every n, and e
a
B.
Because A is Dedekind complete, we can set
e
t
a
= infe
b
: b A
f
, a b;
then e
t
a
B and
e
t
a
a = inf
ba
e
b
a = inf
ba
e
b
b a = inf
ba
e b a = e a.
Now e
t
a
e
t
b
whenever a b, so B = e
t
a
: a A
f
is upwards-directed, and
supB = supe
t
a
a : a A
f
= supe a : a A
f
= e
because (A, ) is semi-nite. Accordingly e B. As e is arbitrary, B is closed, as claimed.
323I Notation In the context of 323H, I will say simply that B is a closed subalgebra of A.
323J Proposition If (A, ) is a localizable measure algebra and B is a subalgebra of A, then the
topological closure B of B in A is precisely the order-closed subalgebra of A generated by B.
proof Write B

for the smallest order-closed subset of A including B. By 313Fc, B

is a subalgebra of
A, and is the order-closed subalgebra of A generated by B. Being an order-closed subalgebra of A, it is
topologically closed, by 323H, and must include B. On the other hand, B, being topologically closed, is
order-closed (323D(c-i)), so includes B

. Thus B = B

is the order-closed subalgebra of A generated by B.

323K I note some simple results for future reference.
Lemma If (A, ) is a localizable measure algebra and B is a closed subalgebra of A, then for any a A the
subalgebra C of A generated by B a is closed.
proof By 314Ja, C is order-closed.
323Y The topology of a measure algebra 87
323L Proposition Let (A
i
,
i
))
iI
be a family of measure algebras with simple product (A, ) (322K).
Then the measure-algebra topology on A =

iI
A
i
dened by is just the product of the measure-algebra
topologies of the A
i
.
proof I use the notations A
f
,
a
from 323A. Write T for the measure-algebra topology of A and S for the
product topology. For i I, d A
f
i
dene a pseudometric
di
on A by setting

di
(b, c) =
d
(b(i), c(i))
whenever b, c A; then S is dened by P =
di
: i I, a A
f
i
(3A3Ig). Now each
di
is one of the
dening pseudometrics for T, since

di
(b, c) = (

d (bc))
where

d(i) = d,

d(j) = 0 for j ,= i. So S T.
Now suppose that a A
f
and > 0. Then

iI

i
a(i) = a is nite, so there is a nite set J I such
that

iI\J

i
a(i)
1
2
. For each j J,
j
=
a(j),j
belongs to P, and

a
(b, c) =

iI

i
(a(i) (b(i) .c(i)))

jJ

j
(a(j) (b(j) .c(j))) +
1
2

jJ

j
(b(j), c(j)) +
1
2

whenever b, c are such that
j
(b(j), c(j)) /(1 +2#(J)) for every j J. By 2A3H, the identity map from
(A, S) to (A, T) is continuous, that is, T S.
Putting these together, we see that S = T, as claimed.
323X Basic exercises (a) Let (A, ) be a semi-nite measure algebra. Show that the set (a, b) : a b
is a closed set in A A.
>>>(b) Let (X, , ) be a -nite measure space and (A, ) its measure algebra. (i) Show that if T is a
-subalgebra of , then F

: F T is a closed subalgebra of A. (ii) Show that if B is a closed subalgebra

of A, then F : F , F

B is a -subalgebra of .
(c) Let (A, ) be a localizable measure algebra, and C A a set such that sup A, inf A belong to C for
all non-empty subsets A of C. Show that C is closed for the measure-algebra topology.
(d) (i) Show that if (A, ) is any measure algebra and B is a subalgebra of A, then its topological closure
B is again a subalgebra. (ii) Use this fact instead of 313Fc to prove 323J.
(e) Let (A, ) be a measure algebra, and e A; let A
e
be the principal ideal of A generated by e, and
e
its measure (322H). Show that the topology on A
e
dened by
e
is just the subspace topology induced by
the measure-algebra topology of A.
>>>(f ) Let (X, , ) be a measure space, and (A, ) its measure algebra. (i) Show that we have an injection
: A L
0
() (see 241) given by setting (E

) = (E)

for every E . (ii) Show that is a

homeomorphism between A and its image if A is given its measure-algebra topology and L
0
() is given its
topology of convergence in measure (245A).
(g) Let (A, ) be a measure algebra, and A
f
the ideal of elements of nite measure. For a, b A
f
set (a, b) = (a .b). Show that (A
f
, ) is a complete metric space and that the operations , , \
and . are uniformly continuous on A
f
, while : A
f
R is also uniformly continuous. Show that the
embedding A
f
A is continuous for the measure-algebra topology on A. In the context of 323Xf, show that
: A
f
L
0
() is an isometry between A
f
and a subset of L
1
().
323Y Further exercises (a) Let (A, ) be a -nite measure algebra. Show that a set F A is closed
for the measure-algebra topology i e F whenever there are non-empty sets B, C A such that B is
upwards-directed, C is downwards-directed, supB = inf C = e and [b, c] F ,= for every b B, c C,
writing [b, c] = d : b d c.
88 Measure algebras 323Yb
(b) Give an example to show that (a) is false for general localizable measure algebras.
(c) Give an example of a semi-nite measure algebra (A, ) with an order-closed subalgebra which is not
closed for the measure-algebra topology.
(d) Let (A, ) be a probability algebra and write B for the family of closed subalgebras of A. For B,
C B set (B, C) = sup
bB
inf
cC
(b .c) + sup
cC
inf
bB
(b .c). Show that (B, ) is a complete metric
space.
(e) Let (A, ) be the measure algebra of Lebesgue measure on R. Show that it is separable in its measure-
algebra topology. (Hint: 245Yj.)
323Z Problem Find a localizable measure algebra (A, ), a -subalgebra Bof A and a sequence b
n
)
nN
in B which converges, for the measure-algebra topology, to a member of A B.
323 Notes and comments The message of this section is that the topology of a measure algebra is
essentially dened by its order and algebraic structure; see also 324F-324H below. Of course the results are
really about semi-nite measure algebras, and indeed this whole volume, like the rest of measure theory,
has little of interest to say about others; they are included only because they arise occasionally and it is
not absolutely essential to exclude them. We therefore expect to be able to describe such things as closed
subalgebras and continuous homomorphisms in terms of the ordering, as in 323H and 324G. For -nite
algebras, indeed, there is an easy description of the topology in terms of the order (323Ya). I think the result
of this section which I shall most often wish to quote is 323I: in most contexts, there is no need to distinguish
between closed subalgebra and order-closed subalgebra. I conjecture, however, that a -subalgebra of a
localizable measure algebra need not be topologically sequentially closed (323Z).
It is also the case that the topology of a measure algebra corresponds very closely indeed to the topology
of convergence in measure. A description of this correspondence is in 323Xf. Indeed all the results of this
section have analogues in the theory of topological Riesz spaces. I will enlarge on the idea here in 367. For
the moment, however, if you look back to Chapter 24, you will see that 323B and 323G are closely paralleled
by 245D and 245E, while 323Ya is related to 245L.
It is I think natural to ask whether there are any other topological Boolean algebras with the properties
323B-323D. In fact a question in this direction, the Control Measure Problem, is one of the most important
questions outstanding in abstract measure theory. I will discuss it in 393; the particular form relevant to
the present section is what I call CM
4
(393J).
324 Homomorphisms
In the course of Volume 2, I had occasion to remark that elementary measure theory was unusual among
abstract topics in pure mathematics in not being dominated by any particular class of structure-preserving
operators. We now come to what I think is one of the reasons for the gap: the most important operators
of the theory are not between measure spaces at all, but between their measure algebras. In this section I
run through the most elementary facts about Boolean homomorphisms between measure algebras. I start
with results on the construction of such homomorphisms from functions between measure spaces (324A-
324E), then investigate continuity and order-continuity of homomorphisms (324F-324H) before turning to
measure-preserving homomorphisms (324I-324O).
324A Theorem Let (X, , ) and (Y, T, ) be measure spaces, and (A, ), (B, ) their measure algebras.
Write

for the domain of the completion of . Let D X be a set of full outer measure (denition:
132E), and let

D
be the subspace -algebra on D induced by

. Let : D Y be a function such that

1
[F]

D
for every F T and
1
[F] is -negligible whenever F = 0. Then there is a sequentially
order-continuous Boolean homomorphism : B A dened by the formula
F

= E

whenever F T, E and (E D)
1
[F] is negligible.
324E Homomorphisms 89
proof Let F T. Then there is an H

such that H D =
1
[F]; now there is an E such
that EH is negligible, so that (E D)
1
[F] is negligible. If E
1
is another member of such that
(E
1
D)
1
[F] is negligible, then (EE
1
) D is negligible, so is included in a negligible member G of
. Since (EE
1
) G belongs to and is disjoint from D, it is negligible; accordingly EE
1
is negligible
and E

= E

1
in A.
What this means is that the formula oered denes a map : B A. It is now easy to check that is
a Boolean homomorphism, because if
(E D)
1
[F], (E
t
D)
1
[F
t
]
are negligible, so are
((X E) D)
1
[Y F], ((E E
t
) D)
1
[F F
t
].
To see that is sequentially order-continuous, let b
n
)
nN
be a sequence in B. For each n we may
choose an F
n
T such that F

n
= b
n
, and E
n
such that (E
n
D)
1
[F
n
] is negligible; now, setting
F =

nN
F
n
, E =

nN
E
n
,
(E D)
1
[F]

nN
(E
n
D)
1
[F
n
]
is negligible, so
(sup
nN
b
n
) = (F

) = E

= sup
nN
E

n
= sup
nN
b
n
.
(Recall that the maps E E

, F F

are sequentially order-continuous, by 321H.) So is sequentially

order-continuous (313L(c-iii)).
324B Corollary Let (X, , ) and (Y, T, ) be measure spaces, and (A, ), (B, ) their measure algebras.
Let : X Y be a function such that
1
[F] for every F T and
1
[F] = 0 whenever F = 0.
Then there is a sequentially order-continuous Boolean homomorphism : B A dened by the formula
F

= (
1
[F])

for every F T.
324C Remarks (a) In 235 and elsewhere in Volume 2 I spent a good deal of time on functions between
measure spaces which satisfy the conditions of 324A. Indeed, I take the trouble to spell 324A out in such
generality just in order to catch these applications. Some of the results of the present chapter (322D, 322Jb)
can also be regarded as special cases of 324A.
(b) The question of which homomorphisms between the measure algebras of measure spaces (X, , ),
(Y, T, ) can be realized by functions between X and Y is important and deep; I will return to it in 343-344.
(c) In the simplied context of 324B, I have actually dened a contravariant functor; the relevant facts
are the following.
324D Proposition Let (X, , ), (Y, T, ) and (Z, , ) be measure spaces, with measure algebras
(A, ), (B, ), (C,

). Suppose that : X Y and : Y Z satisfy the conditions of 324B, that is,

1
[F] if F T,
1
[F] = 0 if F = 0,

1
[G] T if G ,
1
[G] = 0 if G = 0.
Let

: B A,

: C B be the corresponding homomorphisms. Then : X Z is another map of

the same type, and

: C A.
proof The necessary checks are all elementary.
324E Stone spaces While in the context of general measure spaces the question of realizing homomor-
phisms is dicult, in the case of the Stone representation it is relatively straightforward.
Proposition Let (A, ) and (B, ) be measure algebras, with Stone spaces Z and W; let , be the
corresponding measures on Z and W, as described in 321J-321K, and , T their domains. If : B A is
90 Measure algebras 324E
any order-continuous Boolean homomorphism, let : Z W be the corresponding continuous function, as
described in 312P. Then
1
[F] for every F T,
1
[F] = 0 whenever F = 0, and (writing E

for
the member of A corresponding to E ) F

= (
1
[F])

for every F T.
proof Recall that E

= a i Ea is meager, where a is the open-and-closed subset of Z corresponding

to a A. In particular, E = 0 i E is meager. Now the point is that
1
[F] is nowhere dense in Z
whenever F is a nowhere dense subset of W, by 313R. Consequently
1
[F] is meager whenever F is meager
in W, since F is then just a countable union of nowhere dense sets. Thus we see already that
1
[F] = 0
whenever F = 0. If F is any member of T, there is an open-and-closed set F
0
such that FF
0
is meager;
now
1
[F
0
] is open-and-closed, so
1
[F] =
1
[F
0
]
1
[FF
0
] belongs to . Moreover, if b B is such
that

b = F
0
, and a = b, then a =
1
[F
0
], so
F

= b = a = (
1
[F
0
])

= (
1
[F])

,
as required.
324F I turn now to the behaviour of order-continuous homomorphisms between measure algebras.
Theorem Let (A, ) and B, ) be measure algebras and : A B a Boolean homomorphism.
(a) is continuous i it is continuous at 0 i it is uniformly continuous.
(b) If (B, ) is semi-nite and is continuous, then it is order-continuous.
(c) If (A, ) is semi-nite and is order-continuous, then it is continuous.
proof I use the notations A
f
,
a
from 323A.
(a) Suppose that is continuous at 0; I seek to show that it is uniformly continuous. Take b B
f
and
> 0. Then there are a
0
, . . . , a
n
A
f
and > 0 such that
(b c) =
b
(c, 0) whenever max
in

a
i
(c, 0) ;
setting a = sup
in
a
i
,
(b c) whenever (a c) .
Now suppose that
a
(c, c
t
) . Then (a (c .c
t
)) , so

b
(c, c
t
) = (b (c .c
t
)) = (b (c .c
t
)) .
As b, are arbitrary, is uniformly continuous. The rest of the implications are elementary.
(b) Let A be a non-empty downwards-directed set in A with inmum 0. Then 0 A (323D(b-ii)); because
is continuous, 0 [A]. ??? If b is a non-zero lower bound for [A] in B, then (because (B, ) is semi-nite)
there is a c b with 0 < c < ; now

c
(a, 0) = (c a) = c > 0
for every a A, so 0 / [A]. XXX
Thus inf [A] = 0 in B; as A is arbitrary, is order-continuous (313L(b-ii)).
(c) By (a), it will be enough to show that is continuous at 0. Let b B
f
, > 0. ??? Suppose, if possible,
that for every a A
f
, > 0 there is a c A such that (a c) but (b c) . For each a A
f
,
n N choose c
an
such that (a c
an
) 2
n
but (b c
an
) . Set c
a
= inf
nN
sup
mn
c
am
; then
(a c
a
) inf
nN

m=n
(a c
an
) = 0,
so c
a
a = 0. On the other hand, because is order-continuous, c
a
= inf
nN
sup
mn
c
am
, so that
(b c
a
) = lim
n
(b sup
mn
c
am
) .
This shows that

b
(1 \ a, 0) = (b (1 \ a)) (b c
a
) .
But now observe that A = 1 \ a : a A
f
is a downwards-directed subset of A with inmum 0, because
(A, ) is semi-nite. So [A] is downwards-directed and has inmum 0, and 0 must be in the closure of [A],
by 323D(b-ii); while we have just seen that
b
(d, 0) for every d [A]. XXX
324K Homomorphisms 91
Thus there must be a A
f
, > 0 such that

b
(c, 0) = (b c)
whenever

a
(c, 0) = (a c) .
As b, are arbitrary, is continuous at 0 and therefore continuous.
324G Corollary If (A, ) and (B, ) are semi-nite measure algebras, a Boolean homomorphism :
A B is continuous i it is order-continuous.
324H Corollary If A is a Boolean algebra and , are two measures both rendering A a semi-nite
measure algebra, then they endow A with the same uniformity (and, of course, the same topology).
proof By 324G, the identity map from A to itself is continuous whichever of the topologies we place on A;
and by 324F it is therefore uniformly continuous.
324I Denition Let (A, ) and (B, ) be measure algebras. A Boolean homomorphism : A B is
measure-preserving if (a) = a for every a A.
324J Proposition Let (A, ), (B, ) and (C,

) be measure algebras, and : A B, : B C

measure-preserving Boolean homomorphisms. Then : A C is a measure-preserving Boolean homomor-
phism.
proof Elementary.
324K Proposition Let (A, ) and (B, ) be measure algebras, and : A B a measure-preserving
Boolean homomorphism.
(a) is injective.
(b) (A, ) is totally nite i (B, ) is, and in this case is order-continuous, therefore continuous, and
[A] is a closed subalgebra of B.
(c) If (A, ) is semi-nite and (B, ) is -nite, then (A, ) is -nite.
(d) If (A, ) is -nite and is sequentially order-continuous, then (B, ) is -nite.
(e) If (A, ) is semi-nite and is order-continuous, then (B, ) is semi-nite.
(f) If (A, ) is atomless and semi-nite, and is order-continuous, then B is atomless.
(g) If B is purely atomic and (A, ) is semi-nite, then A is purely atomic.
proof (a) If a ,= 0 in A, then a = a > 0 so a ,= 0. By 3A2Db, is injective.
(b) Because
1
B
= 1
A
= 1
A
,
(A, ) is totally nite i (B, ) is. Now suppose that A A is downwards-directed and non-empty and that
inf A = 0. Then
inf
aA
a = inf
aA
a = 0
by 321F. So b = 0 for any lower bound b of [A], and inf [A] = 0. As A is arbitrary, is order-continuous.
By 324Fc, is continuous. By 314Fa, [A] is order-closed in B, that is, closed in the sense of 323I.
(c) I appeal to 322G. If C is a disjoint family in A 0, then c)
cC
is a disjoint family in B 0, so
is countable, and C must be countable, because is injective. Thus A is ccc and (being semi-nite) (A, )
is -nite.
(d) Let a
n
)
nN
be a sequence in A such that a
n
< for every n and sup
nN
a
n
= 1. Then a
n
<
for every n and (because is sequentially order-continuous) sup
nN
a
n
= 1, so (B, ) is -nite.
(e) Setting A
f
= a : a < , supA
f
= 1; because is order-continuous, sup[A
f
] = 1 in B. So if
b = , there is an a A
f
such that a b ,= 0, and now 0 < (b a) < .
92 Measure algebras 324K
(f ) Take any non-zero b B. As in (e), there is an a A such that a < and a b ,= 0. If a b ,= b,
then surely b is not an atom. Otherwise, set
C = c : c A, c a, b c.
Then C is downwards-directed and contains a, so c
0
= inf C is dened in A (321F), and
c
0
= inf
cC
c b > 0,
so c
0
,= 0. Because A is atomless, there is a d c
0
such that neither d nor c
0
\ d is zero, so that neither
c
0
\ d nor d can belong to C. But this means that b d and b (c
0
\ d) are both non-zero, so that again b
is not an atom. As b is arbitrary, B is atomless.
(g) Take any non-zero a A. Then there is an a
t
a such that 0 < a
t
< . Because B is purely
atomic, there is an atom b of B with b a
t
. Set
C = c : c A, c a
t
, b c.
Then C is downwards-directed and contains a
t
, so c
0
= inf C is dened in A, and
c
0
= inf
cC
c b > 0,
so c
0
,= 0. If d c
0
, then b d must be either b or 0. If b d = b, then d C and d = c
0
. If b d = 0,
then c
0
\ d C and d = 0. Thus c
0
is an atom in A. As a is arbitrary, A is purely atomic.
324L Corollary Let (A, ) be a totally nite measure algebra, (B, ) a measure algebra, and : A B
a measure-preserving homomorphism. If C A and C is the closed subalgebra of A generated by C, then
[C] is the closed subalgebra of B generated by [C].
proof This is a special case of 314Gb.
324M Proposition Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras (A, ) and
(B, ). Let : X Y be inverse-measure-preserving. Then we have a sequentially order-continuous
measure-preserving Boolean homomorphism : B A dened by setting F

=
1
[F]

for every F T.
proof This is immediate from 324B.
324N Proposition Let (A, ) and (B, ) be measure algebras, with Stone spaces Z and W; let ,
be the corresponding measures on Z and W. If : B A is an order-continuous measure-preserving
Boolean homomorphism, and : Z W the corresponding continuous function, then is inverse-measure-
preserving.
proof Use 324E. In the notation there, if F T, then
F = F

= F

=
1
[F]

=
1
[F].
324O Proposition Let (A, ) and (B, ) be totally nite measure algebras, A
0
a topologically dense
subalgebra of A, and : A
0
B a Boolean homomorphism such that a = a for every a A
0
. Then
has a unique extension to a measure-preserving homomorphism from A to B.
proof Let , be the standard metrics on A, B, as in 323Ad. Then for any a, a
t
A
0
(a, a
t
) = (aa
t
) = (aa
t
) = (aa
t
) = (a, a
t
);
that is, : A
0
B is an isometry. Because A
0
is dense in the metric space (A, ), while B is complete
under (323Gc), there is a unique continuous function : A B extending (3A4G). Now the operations
(a, a
t
) (a a
t
), (a, a
t
) a a
t
: A A B,
are continuous and agree on the dense subset A
0
A
0
of A A; because the topology of B is Hausdor,
they agree on A A, that is, (a a
t
) = a a
t
for all a, a
t
A (2A3Uc). Similarly, the operations
a (1 \ a), a 1 \ a : A B
324X Homomorphisms 93
are continuous and agree on the dense subset A
0
of A, so they agree on A, that is, (1 \ a) = 1 \ a for every
a A. Thus is a Boolean homomorphism. To see that it is measure-preserving, observe that
a a = (a, 0), a ( a) = ( a, 0) : A R
are continuous and agree on A
0
, so agree on A. Finally, is the only measure-preserving Boolean homo-
morphism extending , because any such map must be continuous (324Kb), and is the only continuous
extension of .
*324P The following fact will be useful in 386, by which time it will seem perfectly elementary; for
the moment, it may be a useful exercise.
Proposition Let (A, ) and (B, ) be totally nite measure algebras such that 1 = 1. Suppose that
A A and : A B are such that (inf
in
a
i
) = (inf
in
a
i
) for all a
0
, . . . , a
n
A. Let C be
the smallest closed subalgebra of A including A. Then has a unique extension to a measure-preserving
Boolean homomorphism from C to B.
proof (a) Let be the family of all functions extending and having the same properties; that is, is
a function from a subset of A to B, and (inf
in
a
i
) = (inf
in
a
i
) for all a
0
, . . . , a
n
dom. By Zorns
Lemma, has a maximal member . Write D for the domain of .
(b)(i) If c, d D then c d D. PPP??? Otherwise, set D
t
= D c d and extend to
t
: D
t
B by
writing
t
(c d) = c d. It is easy to check that
t
, which is supposed to be impossible. XXXQQQ
Now
(c d (c d)) = (c d) = (c d) = (c d),
so (c d) = c d.
(ii) If d D then 1 \ d D. PPP??? Otherwise, set D
t
= D 1 \ d and extend to D
t
by writing

t
(1 \ d) = 1 \ d. Once again, it is easy to check that
t
, which is impossible. XXXQQQ
Consequently (since D is certainly not empty, even if C is), D is a subalgebra of A (312B(iii)).
(iii) Since
1 = 1 = 1,
1 = 1. If d D then
(1 \ d) = (1 \ d) = 1 d = 1 d = (1 \ d),
while
(d (1 \ d)) = (d (1 \ d)) = 0,
so d (1 \ d)) = 0, (1 \ d) 1 \ d and (1 \ d) must be equal to 1 \ d.
By 312H, : D B is a Boolean homomorphism.
(iv) Let D be the topological closure of D in A. Then it is an order-closed subalgebra of A (323J), so,
with D, is a totally nite measure algebra in which D is a topologically dense subalgebra. By 324O, there
is an extension of to a measure-preserving Boolean homomorphism from D to B; of course this extension
belongs to , so in fact D = D is a closed subalgebra of A.
(c) Since A D, C D and
1
= C is a suitable extension of .
To see that
1
is unique, let
2
: C B be any other measure-preserving Boolean homomorphism
extending . Set C = a :
1
a =
2
a; then C is a topologically closed subalgebra of A including A, so is
the whole of C, and
2
=
1
.
324X Basic exercises (a) Let A and B be Boolean algebras, of which A is Dedekind -complete, and
: A B a sequentially order-continuous Boolean homomorphism. Let I be an ideal of A included in the
kernel of . Show that we have a sequentially order-continuous Boolean homomorphism : A/I B given
by setting (a

) = a for every a A.
94 Measure algebras 324Xb
(b) Let (A, ) be a measure algebra, and B a -subalgebra of A. Show that provided that (B, B) is
semi-nite, then the topology of B induced by B is just the subspace topology induced by the topology
of A. (Hint: apply 324Fc to the embedding B A.)
(c) Let (X, , ) be a measure space and (X,

, ) its c.l.d. version. Let A, A
2
be the corresponding
measure algebras and : A A
2
the canonical homomorphism (see 322Db). Show that is topologically
continuous.
(d) Let (A, ) and (B, ) be measure algebras, and : A B a bijective measure-preserving Boolean
homomorphism. Show that
1
: B A is a measure-preserving homomorphism.
(e) Let be counting measure on TN. Show that (TN, ) is a -nite measure algebra. Find a measure-
preserving Boolean homomorphism from TN to itself which is not sequentially order-continuous.
324Y Further exercises (a) Let A and B be Boolean algebras, of which A is Dedekind complete, and
: A B an order-continuous Boolean homomorphism. Let I be an ideal of A included in the kernel of .
Show that we have an order-continuous Boolean homomorphism : A/I B given by setting (a

) = a
for every a A.
(b) Let A be a Dedekind -complete Boolean algebra, and Z its Stone space. Write c for the algebra of
open-and-closed subsets of Z, and Z for the family of nowhere dense zero sets of Z; let Z

be the -ideal of
subsets of Z generated by Z. Show that = EU : E c, U Z

is a -algebra of subsets of Z, and

describe a canonical isomorphism between /Z

and A.
(c) Let A and B be Dedekind -complete Boolean algebras, with Stone spaces Z and W. Construct
Z

TZ as in 324Yb, and let J

T TW be the corresponding structure dened from B. Let

: B A be a sequentially order-continuous Boolean homomorphism, and : Z W the corresponding
continuous map. Show that if E

A corresponds to E , then F

=
1
[F]

for every F T. (Hint:

313Yb.)
(d) Let A be a Boolean algebra, B a ccc Boolean algebra and : A B an injective Boolean homomor-
phism. Show that A is ccc.
(e) Let A be a Dedekind complete Boolean algebra, B a Boolean algebra, and : A B an order-
continuous Boolean homomorphism. Show that for every atom b B there is an atom a A such that
a b. Hence show that if A is atomless so is B, and that if B is purely atomic and is injective then A is
purely atomic.
(f ) Let (A, ) and (B, ) be localizable measure algebras and A
0
an order-dense subalgebra of A. Suppose
that : A
0
B is an order-continuous Boolean homomorphism such that a = a for every a A
0
.
Show that has a unique extension to a measure-preserving Boolean homomorphism from A to B.
(g) Let (A, ) be the measure algebra of Lebesgue measure on [0, 1]. (i) Show that there is an injective
order-preserving function f : A TN. (Hint: take a countable topologically dense subset D of A, and dene
f : A T(D N) by setting f(a) = (d, q) : (a d) q.) (ii) Show that there is an order-preserving
function h : TN A such that h(f(a)) = a for every a A. (Hint: set h(I) = supa : f(a) I.) Compare
316Yo.
(h) Let (A, ) and (B, ) be probability algebras, and f : A B an isometry for the measure metrics.
Show that a f(a) .f(1) is a measure-preserving Boolean homomorphism.
324 Notes and comments If you examine the arguments of this section carefully, you will see that rather
little depends on the measures named. Really this material deals with structures (X, , 1) where X is a set,
is a -ideal of subsets of X, and 1 is a -ideal of , corresponding to the family of measurable negligible
sets. In this abstract form it is natural to think in terms of sequentially order-continuous homomorphisms,
as in 324Yc. I have stated 324E in terms of order-continuous homomorphisms just for a slight gain in
325A Free products and product measures 95
simplicity. But in fact, when there is a dierence, it is likely that order-continuity, rather than sequential
order-continuity, will be the more signicant condition. Note that when the domain algebra is -nite, the
two concepts coincide, because it is ccc (316Fd, 322G).
Of course I need to refer to measures when looking at such concepts as -nite measure algebra or measure-
preserving homomorphism, but even here the real ideas involved are such notions as order-continuity and
the countable chain condition, as you will see if you work through 324K. It is instructive to look at the
translations of these facts into the context of inverse-measure-preserving functions; see 235Xe.
324H shows that we may speak of the topology and uniformity of a Dedekind -complete Boolean algebra
which carries any semi-nite measure; the topology of such an algebra is determined by its algebraic structure.
Contrast this with the theory of normed spaces: two Banach spaces (e.g.,
1
and
2
) can be isomorphic as
linear spaces, both being of algebraic dimension c, while they are not isomorphic as topological linear spaces.
When we come to the theory of ordered linear topological spaces, however, we shall again nd ourselves
with operators whose algebraic properties guarantee continuity (355C, 367P).
325 Free products and product measures
In this section I aim to describe the measure algebras of product measures as dened in Chapter 25. This
will involve the concept of free product set out in 315. It turns out that we cannot determine the measure
algebra of a product measure from the measure algebras of the factors (325B), unless the product measure is
localizable; but that there is nevertheless a general construction of localizable measure algebra free product,
applicable to any pair of semi-nite measure algebras (325D), which represents the measure algebra of the
product measure in the most important cases (325Eb). In the second part of the section (325I-325M) I deal
with measure algebra free products of probability algebras, corresponding to the products of probability
spaces treated in 254.
325A Theorem Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras (A, ) and (B, ).
Let be the c.l.d. product measure on X Y , and its domain; let (C,

) be the corresponding measure

algebra.
(a)(i) The map E E Y : induces an order-continuous Boolean homomorphism from A to C.
(ii) The map F X F : T induces an order-continuous Boolean homomorphism from B to C.
(b) The map (E, F) E F : T induces a Boolean homomorphism : A B C.
(c) [A B] is topologically dense in C.
(d) For every c C,

c = sup

(c (a b)) : a A, b B, a < , b < .

(e) If and are semi-nite, is injective and

(a b) = a b for every a A, b B.
proof (a) By 251E, E Y for every E , and
0
(E Y ) = 0 whenever E = 0, where
0
is the
primitive product measure described in 251A-251C; consequently (E Y ) = 0 whenever E = 0 (251F).
Thus E (E Y )

: C is a Boolean homomorphism with kernel including E : E = 0, so descends

to a Boolean homomorphism
1
: A C.
To see that
1
is order-continuous, let A A
1
be a non-empty downwards-directed set with inmum 0.
??? If there is a non-zero lower bound c of
1
[A], express c as W

where W . We have (W) > 0; by the

denition of (251F), there are G , H T such that G < , H < and (W (GH)) > 0. Of
course inf
aA
a G

= 0 in A, so inf
aA
(a G

) = 0, by 321F; let a A be such that (a G

) H <
(W (GH)). Express a as E

, where E . Then (W (E Y )) = 0. But this means that

(W (GH)) ((E G) H) = (E G) H = (a G

) H,
contradicting the choice of a. XXX Thus inf
1
[A] = 0 in C; as A is arbitrary,
1
is order-continuous.
Similarly
2
: B C, induced by F X F : T , is order-continuous.
(b) Now there must be a corresponding Boolean homomorphism : A B C such that (a b) =

1
a
2
b for every a A, b B, that is,
96 Measure algebras 325A
(E

) = (E Y )

(X F)

= (E F)

for every E , F T (315I).

(c) Suppose that c, e C,

e < and > 0. Express c, e as U

, W

where U, W . By 251Ie, there

are E
0
, . . . , E
n
, F
0
, . . . , F
n
T, all of nite measure, such that ((U W)

in
E
i
F
i
) . Set
c
1
= (

in
E
i
F
i
)

[A B];
then

(e (c .c
1
)) = (W (U

in
E
i
F
i
)) .
As c, e and are arbitrary, [A B] is topologically dense in C.
(d) By the denition of , we have
W = sup(W (E F)) : E , F T, E < , F <
for every W ; so all we have to do is express c as W

.
(e) Now suppose that and are semi-nite. Then (E F) = E F for any E , F T (251J),
so

(a b) = a b for every a A, b B.
To see that is injective, take any non-zero c A B; then there must be non-zero a A, b B such
that a b c (315Jb), so that

c

(a b) = a b > 0
and c ,= 0.
325B Characterizing the measure algebra of a product space A very natural question to ask is,
whether it is possible to dene a measure algebra free product of two abstract measure algebras in a way
which will correspond to one of the constructions above. I give an example to show the diculties involved.
Example There are complete locally determined localizable measure spaces (X, ), (X
t
,
t
), with isomorphic
measure algebras, and a probability space (Y, ) such that the measure algebras of the c.l.d. product measures
on X Y , X
t
Y are not isomorphic.
proof Let (X, , ) be the complete locally determined localizable not-strictly-localizable measure space
described in 216E. Recall that, for E , E = #( : C, f

E) if this is nite, otherwise

(216Eb), where C is a set of cardinal greater than c. The map E : f

E : TC is surjective
(216Ec), so descends to an isomorphism between A, the measure algebra of , and TC. Let (X
t
,
t
,
t
) be
C with counting measure, so that its measure algebra (A
t
,
t
) is isomorphic to (A, ), while
t
is of course
strictly localizable.
Let (Y, T, ) be 0, 1
C
with its usual measure. Let ,
t
be the c.l.d. product measures on XY , X
t
Y
respectively, and (C,

), (C
t
,

t
) the corresponding measure algebras. Then is not localizable (254U), so
(C,

) is not localizable (322Be). On the other hand,

t
, being the c.l.d. product of strictly localizable
measures, is strictly localizable (251N), therefore localizable, so (C
t
,

t
) is localizable, and is not isomorphic
to (C,

).
325C Thus there can be no universally applicable method of identifying the measure algebra of a
product measure from the measure algebras of the factors. However, you have no doubt observed that the
example above involves non--nite spaces, and conjectured that this is not an accident. In contexts in
which we know that all the algebras involved are localizable, there are positive results available, such as the
following.
Theorem Let (X
1
,
1
,
1
) and (X
2
,
2
,
2
) be semi-nite measure spaces, with measure algebras (A
1
,
1
)
and (A
2
,
2
). Let be the c.l.d. product measure on X
1
X
2
, and (C,

) the corresponding measure

algebra. Let (B, ) be a localizable measure algebra, and
1
: A
1
B,
2
: A
2
B order-continuous
Boolean homomorphisms such that (
1
(a
1
)
2
(a
2
)) =
1
a
1

2
a
2
for all a
1
A
1
, a
2
A
2
. Then there is a
unique order-continuous measure-preserving Boolean homomorphism : C B such that ((a
1
a
2
)) =

1
(a
1
)
2
(a
2
) for all a
1
A
1
, a
2
A
2
, writing : A
1
A
2
C for the canonical map described in 325A.
325C Free products and product measures 97
proof (a) Because is injective, it is an isomorphism between A
1
A
2
and its image in C. I trust it will
cause no confusion if I abuse notation slightly and treat A
1
A
2
as actually a subalgebra of C. Now the
Boolean homomorphisms
1
,
2
correspond to a Boolean homomorphism : A
1
A
2
B. The point is
that c =

c for every c A B. PPP By 315Jb, every member of A
1
A
2
is expressible as sup
in
a
i
a
t
i
,
where a
i
A
1
, a
t
i
A
2
and a
i
a
t
i
)
in
is disjoint. Now for each i we have
(a
i
a
t
i
) = (
1
(a
i
)
2
(a
t
i
)) =
1
a
i

2
a
t
i
=

(a
i
a
t
i
),
by 325Ad. So
(c) =

n
i=0
(a
i
a
t
i
) =

n
i=0

(a
i
a
t
i
) =

c. QQQ
(b) The following fact will underlie many of the arguments below. If e B, e < and > 0, there
are e
1
A
f
1
, e
2
A
f
2
such that (e \ (e
1
e
2
)) , writing A
f
i
= a :
i
a < . PPP Because (A
1
,
1
)
is semi-nite, A
f
1
has supremum 1 in A
1
; because
1
is order-continuous, sup
1
(a) : a A
f
1
= 1 in
B, and infe \
1
(a) : a A
f
1
= 0 (313Aa). Because A
f
1
is upwards-directed, e \
1
(a) : a A
f
1
is
downwards-directed, so inf (e \ (a) : a A
f
1
= 0 (321F). Let e
1
A
f
1
be such that (e \
1
(e
1
))
1
2
.
In the same way, there is an e
2
A
f
2
such that (e \
2
(e
2
))
1
2
. Consider e
t
= e
1
e
2
C. Then
(e \ e
t
) = (e \ (
1
(e
1
)
2
(e
2
))) (e \
1
(e
1
)) + (e \
2
(e
2
)) . QQQ
(c) The next step is to check that is uniformly continuous for the uniformities dened by ,

. PPP Take
any e B
f
and > 0. By (b), there are e
1
, e
2
such that

(e
1
e
2
) < and (e \ (e
1
e
2
))
1
2
. Set
e
t
= e
1
e
2
. Now suppose that c, c
t
A
1
A
2
and

((c .c
t
) e
t
)
1
2
. Then
(((c) .(c
t
)) e) ((c .c
t
) e
t
) + (e \ e
t
)

((c .c
t
) e
t
) +
1
2
.
By 3A4Cc, is uniformly continuous for the subspace uniformity on A
1
A
2
. QQQ
(d) Recall that A
1
A
2
is topologically dense in C (325Ab), while B is complete for its uniformity
(323Gc). So there is a uniformly continuous function : C B extending (3A4G).
(e) Because is a Boolean homomorphism, so is . PPP (i) The functions c (1 \ c), c 1 \ (c) are
continuous and the topology of B is Hausdor, so c : (1 \ c) = 1 \ (c) is closed; as it includes A
1
A
2
,
it must be the whole of C. (ii) The functions (c, c
t
) (c c
t
), (c, c
t
) (c) (c
t
) are continuous, so
(c, c
t
) : (c c
t
) = (c) (c
t
) is closed in CC; as it includes (A
1
A
2
) (A
1
A
2
), it must be the whole
of C C. QQQ
(f ) Because is measure-preserving, so is . PPP Take any e
1
A
f
1
, e
2
A
f
2
. Then the functions
c

(c (e
1
e
2
)), c (c (e
1
e
2
)) are continuous and equal on A
1
A
2
, so are equal on C. The
argument of (b) shows that for any b B,
b = sup (b e) : e B
f

= sup (b (e
1
e
2
)) : e
1
A
f
1
, e
2
A
f
2
,
so that
(c) = sup (c (e
1
e
2
)) : e
1
A
f
1
, e
2
A
f
2

= sup

(c (e
1
e
2
)) : e
1
A
f
1
, e
2
A
f
2
=

c
for every c C. QQQ
(g) To see that is order-continuous, take any non-empty downwards-directed set C C with inmum
0. ??? If [C] has a non-zero lower bound b in B, let e b be such that 0 < e < . Let e
t
C be such that

e
t
< and (e \ (e
t
)) < e, as in (b) above, so that (e (e
t
)) > 0. Now, because inf C = 0, there is a
c C such that

(c e
t
) < (e (e
t
)). But this means that
(b (e
t
)) (c e
t
) =

(c e
t
) < (e (e
t
)) (b (e
t
)),
which is absurd. XXXThus inf [C] = 0 in B. As C is arbitrary, is order-continuous.
98 Measure algebras 325C
(h) Finally, to see that is unique, observe that any order-continuous Boolean homomorphism from C
to B must be continuous (324Fc); so that if it agrees with on A
1
A
2
it must agree with on C.
325D Theorem Let (A
1
,
1
) and (A
2
,
2
) be semi-nite measure algebras.
(a) There is a localizable measure algebra (C,

), together with order-continuous Boolean homomorphisms

1
: A
1
C,
2
: A
2
C such that whenever (B, ) is a localizable measure algebra, and
1
: A
1
B,

2
: A
2
B are order-continuous Boolean homomorphisms and (
1
(a
1
)
2
(a
2
)) =
1
a
1

2
a
2
for all
a
1
A
1
, a
2
A
2
, then there is a unique order-continuous measure-preserving Boolean homomorphism
: C B such that
j
=
j
for both j.
(b) The structure (C,

,
1
,
2
) is determined up to isomorphism by this property.
(c)(i) The Boolean homomorphism : A
1
A
2
C dened from
1
and
2
is injective, and [A
1
A
2
]
is topologically dense in C.
(ii) The order-closed subalgebra of C generated by [A
1
A
2
] is the whole of C.
(d) If j 1, 2 and (A
j
,
j
) is localizable, then
j
[A
j
] is a closed subalgebra of (C,

).
proof (a)(i) We may regard (A
1
,
1
) as the measure algebra of (Z
1
,
1
,
1
) where Z
1
is the Stone space
of A
1
,
1
is the algebra of subsets of Z
1
diering from an open-and-closed set by a meager set, and
1
is
an appropriate measure (321K). Note that in this representation, each a A
1
becomes identied with a

,
where a is the open-and-closed subset of Z
1
corresponding to a. Similarly, we may think of (A
2
,
2
) as the
measure algebra of (Z
2
,
2
,
2
), where Z
2
is the Stone space of A
2
.
(ii) Let be the c.l.d. product measure on Z
1
Z
2
. The point is that is strictly localizable. PPP By
322E, both A
1
and A
2
have partitions of unity consisting of elements of nite measure; let c
i
)
iI
, d
j
)
jJ
be
such partitions. Then c
i

d
j
)
iI,jJ
is a disjoint family of sets of nite measure in Z
1
Z
2
. If W Z
1
Z
2
is such that W > 0, there must be sets E
1
, E
2
of nite measure such that (W (E
1
E
2
)) > 0. Because
E

1
= sup
iI
E

1
c
i
, we must have

1
E
1
=
1
E

1
=

iI

1
(E

1
c
i
) =

iI

1
(E
1
c
i
).
Similarly,
2
E
2
=

iJ

2
(E
2


d
j
). But this means that there must be nite I
t
I, J
t
J such that

iI

,jJ

1
(E
1
c
i
)
2
(E
2


d
j
) >
1
E
1

2
E
2
(W (E
1
E
2
)),
so that there have to be i I
t
, j J
t
such that (W (c
i


d
j
)) > 0.
Now this means that c
i


d
j
)
iI,jJ
satises the conditions of 213O. Because is surely complete and
locally determined, it is strictly localizable. QQQ
(iii) We may therefore take (C,

) to be just the measure algebra of . The maps

1
,
2
will be the
canonical maps described in 325Aa, inducing the map : A
1
B
1
C referred to in 325C; and 325C now
gives the result.
(b) This is nearly obvious. Suppose we had an alternative structure (C
t
,

t
,
t
1
,
t
2
) with the same property.
Then we must have an order-continuous measure-preserving Boolean homomorphism : C C
t
such that

j
=
t
j
for both j; and similarly we have an order-continuous measure-preserving Boolean homomorphism

t
: C
t
C such that
t

t
j
=
j
for both j. Now
t
: C C is an order-continuous measure-preserving
Boolean homomorphism such that
t

j
=
j
for both j. By the uniqueness assertion in (a), applied with
B = C,
t
must be the identity on C. In the same way,
t
is the identity on C
t
. So and
t
are the two
halves of the required isomorphism.
(c) In view of the construction for C oered in part (a) of the proof, (i) is just a consequence of 325Ac
and 325Ae. Now (ii) follows by 323J.
(d) If A
j
is Dedekind complete then
j
[A
j
] is order-closed in C because
j
is order-continuous (314F(a-i)).
325E Remarks (a) We could say that a measure algebra (C,

), together with embeddings

1
and
2
,
as described in 325D, is a localizable measure algebra free product of (A
1
,
1
) and (A
2
,
2
); and its
uniqueness up to isomorphism makes it safe, most of the time, to call it the localizable measure algebra
free product. Observe that it can equally well be regarded as the uniform space completion of the algebraic
free product; see 325Yb.
*325H Free products and product measures 99
(b) As the example in 325B shows, the localizable measure algebra free product of the measure algebras
of given measure spaces need not appear directly as the measure algebra of their product. But there is one
context in which it must so appear: if the product measure is localizable, 325C tells us at once that it has
the right measure algebra. For -nite measure algebras, of course, any corresponding measure spaces have
to be strictly localizable, so again we can use the product measure directly.
325F I ought not to proceed to the next topic without giving another pair of examples to show the
subtlety of the concept of measure algebra free product.
Example Let (A, ) be the measure algebra of Lebesgue measure on [0, 1], and (C,

) the measure algebra

of Lebesgue measure on [0, 1]
2
. Then (C,

) can be regarded as the localizable measure algebra free product

of (A, ) with itself, by 251M and 325Eb. Let : A A C be the canonical map, as described in 325A.
Then [A A] is not order-dense in C, and is not order-continuous.
proof (a) Let
n
)
nN
be a sequence in [0, 1] such that

n=0

n
= , but

n=0

2
n
< 1; for instance, we
could take
n
=
1
n+2
. Let E
n
)
nN
be a stochastically independent sequence of measurable subsets of [0, 1]
such that E
n
=
n
for each n. In A set a
n
= E

n
, and consider c
n
= sup
in
a
i
a
i
A A for each n.
(b) We have sup
nN
c
n
= 1 in AA. PPP??? Otherwise, there is a non-zero a AA such that a a
n
= 0
for every n, and now there are non-zero b, b
t
A such that b b
t
a. Set I = n : a
n
b = 0, J = n :
a
n
b
t
= 0. Then E
n
)
nI
is an independent family and (

nI
E
i
) 1 b < 1, so

nI
E
n
< , by
the Borel-Cantelli lemma (273K). Similarly

nJ
E
n
< . Because

nN
E
n
= , there must be some
n N (I J). Now a
n
b and a
n
b
t
are both non-zero, so
0 ,= (a
n
b) (a
n
b
t
) = (a
n
a
n
) (b b
t
) = 0,
which is absurd. XXXQQQ
(c) On the other hand,

n=0

(c
n
)

n=0
( a
n
)
2
=

n=0

2
n
< ,
by the choice of the
n
. So sup
nN
(c
n
) cannot be 1 in C.
Thus is not order-continuous.
(d) By 313P(a-ii) and 313O, [AA] cannot be order-dense in C; alternatively, (b) shows that there can
be no non-zero member of [AA] included in 1 \ sup
nN
(c
n
). (Both these arguments rely tacitly on the
fact that is injective, as noted in 325Ae.)
325G Since 325F shows that the free product and the localizable measure algebra free product are very
dierent constructions, I had better repeat an idea from 315 in the new context.
Example Again, let (A, ) be the measure algebra of Lebesgue measure on [0, 1], and (C,

) the measure
algebra of Lebesgue measure on [0, 1]
2
. Then there is no order-continuous Boolean homomorphism : C A
such that (a b) = a b for all a, b A. PPP Let : C A be a Boolean homomorphism such that
(ab) = a b for all a, b A. For i < 2
n
let a
ni
be the equivalence class in A of the interval [2
n
i, 2
n
(i+1)],
and set c
n
= sup
i<2
n a
ni
a
ni
. Then c
n
= 1 for every n, but

c
n
= 2
n
for each n, so inf
nN
c
n
= 0 in C;
thus cannot be order-continuous. QQQ (Compare 315P.)
*325H Products of more than two factors We can of course extend the ideas of 325A, 325C and
325D to products of any nite number of factors. No new ideas are needed, so I spell the results out without
proofs.
(a) Let (A
i
,
i
))
iI
be a nite family of semi-nite measure algebras. Then there is a localizable mea-
sure algebra (C,

), together with order-continuous Boolean homomorphisms

i
: A
i
C for i I, such
that whenever (B, ) is a localizable measure algebra, and
i
: A
i
B are order-continuous Boolean ho-
momorphisms such that (inf
iI

i
(a
i
)) =

iI

i
a
i
whenever a
i
A
i
for each i, then there is a unique
order-continuous measure-preserving Boolean homomorphism : C B such that
i
=
i
for every i.
100 Measure algebras 325Hb
(b) The structure (C,

,
i
)
iI
) is determined up to isomorphism by this property.
(c) The Boolean homomorphism :

iI
A
i
C dened from the
i
is injective, and [

iI
A
i
] is
topologically dense in C.
(d) Write

loc
iI
(A
i
,
i
) for (a particular version of) the localizable measure algebra free product described
in (a). If (A
i
,
i
))
iI
is a nite family of semi-nite measure algebras and I(k))
kK
is a partition of I,
then

loc
iI
(A
i
,
i
) is isomorphic, in a canonical way, to

loc
kK
_

loc
iI(k)
(A
i
,
i
)
_
.
(e) Let (X
i
,
i
,
i
))
iI
be a nite family of semi-nite measure spaces, and write (A
i
,
i
) for the measure
algebra of (X
i
,
i
,
i
). Let be the c.l.d. product measure on

iI
X
i
(251W), and (C,

) the corresponding
measure algebra. Then there is a canonical order-continuous measure-preserving embedding of (C,

) into
the localizable measure algebra free product of the (A
i
,
i
). If each
i
is strictly localizable, this embedding
is an isomorphism.
325I Innite products Just as in 254, we can now turn to products of innite families of probability
algebras.
Theorem Let (X
i
,
i
,
i
))
iI
be any family of probability spaces, with measure algebras (A
i
,
i
). Let
be the product measure on

iI
X
i
, and (C,

) the corresponding measure algebra. For each i I, we have

a measure-preserving homomorphism
i
: A
i
C corresponding to the inverse-measure-preserving function
x x(i) : X X
i
. Let (B, ) be a probability algebra, and
i
: A
i
B Boolean homomorphisms such
that (inf
iJ

i
(a
i
)) =

iJ

i
a
i
whenever J I is a nite set and a
i
A
i
for every i. Then there is a
unique measure-preserving Boolean homomorphism : C B such that
i
=
i
for every i I.
proof (a) As remarked in 254Fb, all the maps x x(i) are inverse-measure-preserving, so correspond to
measure-preserving homomorphisms
i
: A
i
C (324M). It will be helpful to use some notation from 254.
Write ( for the family of subsets of X expressible in the form
E = x : x X, x(i) E
i
for every i J,
where J I is nite and E
i

i
for every i J. Note that in this case
E

= inf
iJ

i
(E

i
).
Set
C = E

: E ( C,
so that C is precisely the family of elements of C expressible in the form inf
iJ

i
(a
i
) where J I is nite
and a
i
A
i
for each i.
The homomorphisms
i
: A
i
C dene a Boolean homomorphism :

iI
A
i
C (315I), which is
injective. PPP If c

iI
A
i
is non-zero, there must be a nite set J I and a family a
i
)
iJ
such that
a
i
A
i
0 for each i and c inf
iJ

i
(a
i
) (315Jb). Express each a
i
as E

i
, where E
i

i
. Then
E = x : x X, x(i) E
i
for each i J
has measure
E =

iJ
E
i
=

iJ
a
i
,= 0,
while
E

= (inf
iJ

i
(a
i
)) (c),
so (c) ,= 0. As c is arbitrary, is injective. QQQ
(b) Because is injective, it is an isomorphism between

iI
A
i
and its image in C. I trust it will cause
no confusion if I abuse notation slightly and treat

iI
A
i
as actually a subalgebra of C, so that
j
: A
j
C
becomes identied with the canonical map
j
: A
j

iI
A
i
. Now the Boolean homomorphisms
i
: A
i

B correspond to a Boolean homomorphism :

iI
A
i
B. The point is that (c) =

c for every
325L Free products and product measures 101
c

iI
A
i
. PPP Suppose to begin with that c C. Then we have c = E

, where E = x : x(i) E
i
i J
and E
i

i
for each i J. So

c = E =

iJ
E
i
=

iJ

i
E

i
= (inf
iJ
a
i
)
= (inf
iJ

i
(a
i
)) = (inf
iJ

i
(a
i
)) = (c).
Next, any c C is expressible as the supremum of a nite disjoint family c
k
)
kK
in C (315Jb), so
(c) =

kK
(c
k
) =

kK

(c
k
) =

c. QQQ
(c) It follows that is uniformly continuous for the metrics dened by ,

, since
((c) .(c
t
)) = (c .c
t
) =

(c .c
t
)
for all c, c
t

iI
A
i
.
(d) Next,

iI
A
i
is topologically dense in C. PPP Let c C, > 0. Express c as W

. Then by 254Fe
there are H
0
, . . . , H
k
( such that (W

jk
H
j
) . Now c
j
= H

j
C for each j, so
c
t
= sup
jk
c
j
= (

jk
H
j
)

iI
A
i
,
and

(c .c
t
) . QQQ
Since B is complete for its uniformity (323Gc), there is a uniformly continuous function : C B
extending (3A4G).
(e) Because is a Boolean homomorphism, so is . PPP (i) The functions c (1 \ c), 1 \ (c) are
continuous and the topology of B is Hausdor, so c : (1 \ c) = 1 \ (c) is closed; as it includes A
1
A
2
,
it must be the whole of C. (ii) The functions (c, c
t
) (c c
t
), (c, c
t
) (c) (c
t
) are continuous, so
(c, c
t
) : (c c
t
) = (c) (c
t
) is closed in CC; as it includes (A
1
A
2
) (A
1
A
2
), it must be the whole
of C C. QQQ
(f ) Because is measure-preserving, so is . PPP The functions c

c, c (c) are continuous and
equal on A
1
A
2
, so are equal on C. QQQ
(g) Finally, to see that is unique, observe that any measure-preserving Boolean homomorphism from
C to B must be continuous, so that if it agrees with on

iI
A
i
it must agree with on C.
325J Of course this leads at once to a result corresponding to 325D.
Theorem Let (A
i
,
i
))
iI
be a family of probability algebras.
(a) There is a probability algebra (C,

), together with order-continuous Boolean homomorphisms

i
:
A
i
C such that whenever (B, ) is a probability algebra, and
i
: A
i
B are Boolean homomorphisms
such that (inf
iJ

i
(a
i
)) =

iJ

i
a
i
whenever J I is nite and a
i
A
i
for each i J, then there is a
unique measure-preserving Boolean homomorphism : C B such that
j
=
j
for every j.
(b) The structure (C,

,
i
)
iI
) is determined up to isomorphism by this property.
(c) The Boolean homomorphism :

iI
A
i
C dened from the
i
is injective, and [

iI
A
i
] is
topologically dense in C.
proof For (a) and (c), all we have to do is represent each (A
i
,
i
) as the measure algebra of a probability
space, and apply 325I. The uniqueness of C and the
i
follows from the uniqueness of the homomorphisms
, as in 325D.
325K Denition As in 325Ea, we can say that (C,

,
i
)
iI
) is a, or the, probability algebra free
product of (A
i
,
i
))
iI
.
325L Independent subalgebras If (A, ) is a probability algebra, we say that a family B
i
)
iI
of
subalgebras of A is (stochastically) independent if (inf
iJ
b
i
) =

iJ
b
i
whenever J I is nite and
b
i
B
i
for each i. (Compare 272Ab.) In this case the embeddings B
i
A give rise to an embedding of the
probability algebra free product of (B
i
, B
i
))
iI
into A. (Compare 272J, 315Xn.)
102 Measure algebras 325M
325M We can now make a general trawl through Chapters 25 and 27 seeking results which can be
expressed in the language of this section. I give some in 325Xe-325Xh. Some ideas from 254 which are
thrown into sharper relief by a reformulation are in the following theorem.
Theorem Let (A
i
,
i
))
iI
be a family of probability algebras and (C,

,
i
)
iI
) their probability algebra
free product. For J I let C
J
be the closed subalgebra of C generated by

iJ

i
[A
i
].
(a) For any J I, (C
J
,

C
J
,
i
)
iJ
) is a probability algebra free product of (A
i
,
i
))
iJ
.
(b) For any c C, there is a unique smallest J I such that c C
J
, and this J is countable.
(c) For any non-empty family TI,

J
C
J
= C

.
proof (a) If (B, ,
i
)
iJ
) is any probability algebra free product of (A
i
,
i
))
iJ
, then we have a measure-
preserving homomorphism : B C such that
i
=
i
for every i J. Because the subalgebra B
0
of B
generated by

iJ

i
[A
i
] is topologically dense in B (325Jc), and is continuous (324Kb),

iJ

i
[A
i
] is
topologically dense in [B]; also [B] is closed in C (324Kb again). But this means that [B] is just the
topological closure of

iI

i
[A
i
] and must be C
J
. Thus is an isomorphism, and
(C
J
,

C
J
,
i
)
iJ
) = ([B],
1
,
i
)
iI
)
is also a probability algebra free product of (A
i
,
i
))
iJ
.
(b) As in 325J, we may suppose that each (A
i
,
i
) is the measure algebra of a probability space (X
i
,
i
,
i
),
and that C is the measure algebra of their product (X, , ). Let W be such that c = W

.
By 254Rd, there is a unique smallest K I such that WU is negligible for some U
K
, where
K
is the set of members of which are determined by coordinates in K; and K is countable. But if we look
at any J I, x : x(i) E
J
for every i J, E
i
; so U

: U
J
is a closed subalgebra of
C including
i
[A
i
] for every i J, and therefore including C
J
. On the other hand, as observed in 254Ob,
any member of
J
is approximated, in measure, by sets in the -algebra T
J
generated by sets of the form
x : x(i) E where i J, E
i
. Of course T
J

J
, so W

: W
J
= W

: W T
J
is the closed
subalgebra of C generated by

iK

i
[A
i
], which is C
J
. Thus K is also the unique smallest subset of I such
that c C
K
.
(c) Of course C
K
C
J
whenever K J I, so

J
C
J
C

. On the other hand, suppose

that c

J
C
J
; then by (b) there is some K

such that c C
K
C

. As c is arbitrary,

J
C
J
= C

.
*325N Notation In this context, I will say that an element c of C is determined by coordinates in
J if c C
J
.
325X Basic exercises (a) Let (A
1
,
1
), (A
2
,
2
) be two semi-nite measure algebras, and suppose that
for each j we are given a closed subalgebra B
j
of A
j
such that (B
j
,
j
) is also semi-nite, where
j
=
j
B
j
.
Show that the localizable measure algebra free product of (B
1
,
1
) and (B
2
,
2
) can be thought of as a closed
subalgebra of the localizable measure algebra free product of (A
1
,
1
) and (A
2
,
2
).
(b) Let (A
1
,
1
), (A
2
,
2
) be two semi-nite measure algebras, and suppose that for each j we are given
a principal ideal B
j
of A
j
. Set
j
=
j
B
j
. Show that the localizable measure algebra free product of
(B
1
,
1
) and (B
2
,
2
) can be thought of as a principal ideal of the localizable measure algebra free product
of (A
1
,
1
) and (A
2
,
2
).
>>>(c) Let (A
i
,
i
))
iI
and (B
j
,
j
))
jJ
be families of semi-nite measure algebras, with simple products
(A, ) and (B, ) (322K). Show that the localizable measure algebra free product (A, )

loc
(B, ) can be
identied with the simple product of the family (A
i
,
i
)

loc
(B
j
,
j
))
iI,jJ
.
>>>(d) Let (A
i
,
i
))
iI
be a family of probability algebras, and (C,

,
i
)
iI
) their probability algebra
free product. Suppose that for each i I we are given a measure-preserving Boolean homomorphism

i
: A
i
A
i
. Show that there is a unique measure-preserving Boolean homomorphism : C C such that

i
=
i

i
for every i I.
325 Notes Free products and product measures 103
>>>(e) Let (A, ) be a probability algebra. We say that a family a
i
)
iI
in A is (stochastically) inde-
pendent if (inf
ij
a
j
) =

iJ
a
i
for every non-empty nite J I. Show that this is so i A
i
)
iI
is
independent, where A
i
= 0, a
i
, 1 \ a
i
, 1 for each i. (Compare 272F.)
(f ) Let (A, ) be a probability algebra, and A
i
)
iI
a stochastically independent family of closed subal-
gebras of A. Let J(k))
kK
be a disjoint family of subsets of I, and for each k K let B
k
be the closed
subalgebra of A generated by

iJ(k)
A
i
. Show that B
k
)
kK
is independent. (Compare 272K.)
(g) Let (A, ) be a probability algebra, and A
i
)
iI
a stochastically independent family of closed subal-
gebras of A. For J I let B
J
be the closed subalgebra of A generated by

iJ
A
i
. Show that

B
I\J
: J
is a nite subset of I = 0, 1. (Hint: For J I, show that (b c) = b c for every b B
I\J
, c B
J
.
Compare 272O, 325M.)
(h) Let (A
i
,
i
))
iI
be a family of probability algebras with probability algebra free product (C,

,
i
)
iI
).
For c C let J
c
be the smallest subset of I such that c belongs to the closed subalgebra of C generated by

iJ
c

i
[A
i
]. Show that if c d in C, then there is an e C such that c e d and J
e
J
c
J
d
. (Hint:
254R.)
325Y Further exercises (a) Let be counting measure on X = 0,
t
the countable-cocountable
measure on X
t
=
1
, and counting measure on Y =
1
. Show that the measure algebras of the primitive
product measures on X Y , X
t
Y are not isomorphic.
(b) Let A be a Boolean algebra, and : A [0, ] a function such that 0 = 0 and (a b) = a + b
whenever a, b A and a b = 0; suppose that A
f
= a : a < is order-dense in A. For e A
f
, a, b A
set
e
(a, b) = (e (a .b)). Give A the uniformity dened by
e
: e < . (i) Show that the completion

A of A under this uniformity has a measure , extending , under which it is a localizable measure algebra.
(ii) Show that if a

A, a < and > 0, there is a b A such that (a .b) . (iii) Show that for every
a

A there is a sequence a
n
)
nN
in A such that a sup
nN
inf
mn
a
m
and a = (sup
nN
inf
mn
a
m
).
(iv) In particular, the set of inma in

A of sequences in A is order-dense in

A. (v) Explain the relevance of
this construction to the embedding A
1
A
2
C in 325D.
(c) In 325F, set W =

nN
E
n
E
n
. Show that if A, B are any non-negligible subsets of [0, 1], then
W (AB) is not negligible.
(d) Let (A, ) be the measure algebra of Lebesgue measure on [0, 1]. Show that A A is ccc but not
weakly (, )-distributive. (Hint: (i) A A is embeddable as a subalgebra of a probability algebra (ii) in
the notation of 325F, look at c
mn
= sup
min
e
i
e
i
.)
(e) Repeat 325F-325G and 325Yc-325Yd with an arbitrary atomless probability space in place of [0, 1].
(f ) Let (A, ) be a probability algebra and a
i
)
iI
a (stochastically) independent family in A. Show that
for any a A, > 0 the set i : i I, [ (a a
i
) a a
i
[ is nite, so that i : (a a
i
) ,= a a
i
is
countable. (Hint: 272Yd.)
325 Notes and comments 325B shows that the measure algebra of a product measure may be irregular
if we have factor measures which are not strictly localizable. But two facts lead the way to the local-
izable measure algebra free product in 325D-325E. The rst is that every semi-nite measure algebra is
embeddable, in a canonical way, in a localizable measure algebra (322N); and the second is that the Stone
representation of a localizable measure algebra is strictly localizable (322M). It is a happy coincidence that
we can collapse these two facts together in the construction of 325D. Another way of looking at the localiz-
able measure algebra free product of two localizable measure algebras is to express it as the simple product
of measure algebra free products of totally nite measure algebras, using 325Xc and the fact that for -nite
measure algebras there is only one reasonable measure algebra free product, being that provided by any
representation of them as measure algebras of measure spaces (325Eb).
104 Measure algebras 325 Notes
Yet a third way of approaching measure algebra free products is as the uniform space completions of
algebraic free products, using 325Yb. This gives the same result as the construction of 325D because the
algebraic free product appears as a topologically dense subalgebra of the localizable measure algebra free
product (325Dc) which is complete as uniform space (325Dc). (I have to repeat such phrases as topologically
dense because the algebraic free product is emphatically not order-dense in the measure algebra free product
(325F).) The results in 251I on approximating measurable sets for a c.l.d. product measure by combinations of
measurable rectangles correspond to general facts about completions of nitely-additive measures (325Yb(ii),
325Yb(iii)). It is worth noting that the completion process can be regarded as made up of two steps; rst
take inma of sequences of sets of nite measure, and then take arbitrary suprema (325Yb(iv)).
The idea of 325F appears in many guises, and this is only the rst time that I shall wish to call on it.
The point of the set W =

nN
E
n
E
n
is that it is a measurable subset of the square (indeed, by taking
the E
n
to be open sets we can arrange that W should be open), of measure strictly less than 1 (in fact,
as small as we wish), such that its complement does not include any non-negligible measurable rectangle
G H; indeed, W (A B) is non-negligible for any non-negligible sets A, B [0, 1] (325Yc). I believe
that the rst published example of such a set was by Erd os & Oxtoby 55; I learnt the method of 325F
from R.O.Davies.
I include 325G as a kind of guard-rail. The relationship between preservation of measure and order-
continuity is a subtle one, as I have already tried to show in 324K, and it is often worth considering the
possibility that a result involving order-continuous measure-preserving homomorphisms has a form applying
to all order-continuous homomorphisms. However, there is no simple expression of such an idea in the
present context.
In the context of innite free products of probability algebras, there is a degree of simplication, since there
is only one algebra which can plausibly be called the probability algebra free product, and this is produced
by any realization of the algebras as measure algebras of probability spaces (325I-325K). The examples
325F-325G apply equally, of course, to this context. At this point I mention the concept of (stochastically)
independent family (325L, 325Xe) because we have the machinery to translate several results from 272
into the language of measure algebras (325Xe-325Xg). I feel that I have to use the phrase stochastically
independent here because there is the much weaker alternative concept of Boolean independence (315Xn)
also present. But I leave most of this as exercises, because the language of measure algebras oers few ideas
to the probability theory already covered in Chapter 27. All it can do is formalise the ever-present principle
that negligible sets often can and should be ignored.
326 Additive functionals on Boolean algebras
I devote two sections to the general theory of additive functionals on measure algebras. As many readers
will rightly be in a hurry to get on to the next two chapters, I remark that the only signicant result needed
for 331-332 is the Hahn decomposition of a countably additive functional (326I), and that this is no more
than a translation into the language of measure algebras of a theorem already given in Chapter 23. The
concept of standard extension of a countably additive functional from a subalgebra (327F-327G) will be
used for a theorem in 333, and as preparation for Chapter 36.
I begin with notes on the space of additive functionals on an arbitrary Boolean algebra (326A-326D),
corresponding to 231A-231B, but adding a more general form of the Jordan decomposition of a bounded
additive functional into positive and negative parts (326D). The next subsection (326E-326I) deals with
countably additive functionals, corresponding to 231C-231F. In 326J-326P I develop a new idea, that of
completely additive functional, which does not match anything in the previous treatment. In 326Q I
return to additive functionals, giving a fundamental result on the construction of additive functionals on
free products.
326A Additive functionals: Denition Let A be a Boolean algebra. A functional : A R is
nitely additive, or just additive, if (a b) = a +b whenever a, b A and a b = 0.
A non-negative additive functional is sometimes called a nitely additive measure or charge.
326D Additive functionals on Boolean algebras 105
326B Elementary facts Let A be a Boolean algebra and : A R a nitely additive functional. The
following will I hope be obvious.
(a) 0 = 0 (because 0 = 0 +0).
(b) If c A, then a (a c) is additive (because (a c) (b c) = (a b) c).
(c) is an additive functional for any R. If
t
is another nitely additive functional on A, then
+
t
is additive.
(d) If
i
)
iI
is any family of nitely additive functionals such that
t
a =

iI

i
a is dened in R for for
every a A, then
t
is additive.
(e) If B is another Boolean algebra and : B A is a Boolean homomorphism, then : B R is
additive. In particular, if B is a subalgebra of A, then B : B R is additive.
(f ) is non-negative i it is order-preserving that is,
a 0 for every a A b c whenever b c
(because c = b +(c \ b) if b c).
326C The space of additive functionals Let A be any Boolean algebra. From 326Bc we see that
the set M of all nitely additive real-valued functionals on A is a linear space (a linear subspace of R
A
). We
give it the ordering induced by that of R
A
, so that
t
i a
t
a for every a A. This renders it a
partially ordered linear space (because R
A
is).
326D The Jordan decomposition (I): Proposition Let A be a Boolean algebra, and a nitely
additive real-valued functional on A. Then the following are equiveridical:
(i) is bounded;
(ii) sup
nN
[a
n
[ < for every disjoint sequence a
n
)
nN
in A;
(iii) lim
n
[a
n
[ = 0 for every disjoint sequence a
n
)
nN
in A;
(iv)

n=0
[a
n
[ < for every disjoint sequence a
n
)
nN
in A;
(v) is expressible as the dierence of two non-negative additive functionals.
proof (a)(i)(v) Assume that is bounded. For each a A, set

+
a = supb : b a.
Because is bounded,
+
is real-valued. Now
+
is additive. PPP If a, b A and a b = 0, then

+
(a b) = sup
cab
c = sup
da,eb
(d e) = sup
da,eb
d +e
(because d e a b = 0 whenever d a, e b)
= sup
da
d + sup
eb
e =
+
a +
+
b. QQQ
Consequently

=
+
is also additive (326Bc).
Since
0 = 0
+
a, a
+
a
for every a A,
+
0 and

0. Thus =
+

is the dierence of two non-negative additive

functionals.
(b)(v)(iv) If is expressible as
1

2
, where
1
and
2
are non-negative additive functionals, and
a
n
)
nN
is disjoint, then

n
i=0

j
a
i
=
j
(sup
in
a
i
)
j
1
106 Measure algebras 326D
for every n, both j, so that

i=0
[a
i
[

i=0

1
a
i
+

i=0

2
a
i

1
1 +
2
1 < .
(c)(iv)(iii)(ii) are trivial.
(d) not-(i)not-(ii) Suppose that is unbounded. Choose sequences a
n
)
nN
, b
n
)
nN
inductively, as
follows. b
0
= 1. Given that sup
ab
n
[a[ = , choose c
n
b
n
such that [c
n
[ [b
n
[ + n; then [c
n
[ n
and
[(b
n
\ c
n
)[ = [b
n
c
n
[ [c
n
[ [b
n
[ n.
We have
= sup
ab
n
[a[ = sup
ab
n
[(a c
n
) +(a \ c
n
)[
sup
ab
n
[(a c
n
)[ +[(a \ c
n
)[ sup
ab
n
c
n
[a[ + sup
ab
n
\c
n
[a[,
so at least one of sup
ab
n
c
n
[a[, sup
ab
n
\c
n
[a[ must be innite; take b
n+1
to be one of c
n
, b
n
\ c
n
such
that sup
ab
n+1
[a[ = , and set a
n
= b
n
\ b
n+1
, so that [a
n
[ n. Continue.
On completing the induction, we have a disjoint sequence a
n
)
nN
such that [a
n
[ n for every n, so
that (ii) is false.
Remark I hope that this reminds you of the decomposition of a function of bounded variation as the
dierence of monotonic functions (224D).
326E Countably additive functionals: Denition Let A be a Boolean algebra. A functional :
A R is countably additive or -additive if

n=0
a
n
is dened and equal to (sup
nN
a
n
) whenever
a
n
)
nN
is a disjoint sequence in A and sup
nN
a
n
is dened in A.
A warning is perhaps in order. It can happen that A is presented to us as a subalgebra of a larger algebra
B; for instance, A might be an algebra of sets, a subalgebra of some -algebra TX. In this case,
there may be sequences in A which have a supremum in A which is not a supremum in B (indeed, this will
happen just when the embedding is not sequentially order-continuous). So we can have a countably additive
functional : B R such that A is not countably additive in the sense used here. A similar phenomenon
will arise when we come to the Daniell integral in Volume 4 (434).
326F Elementary facts Let A be a Boolean algebra and : A R a countably additive functional.
(a) is nitely additive. (Setting a
n
= 0 for every n, we see from the denition in 326E that 0 = 0.
Now, given a b = 0, set a
0
= a, a
1
= b, a
n
= 0 for n 2 to see that (a b) = a +b.)
(b) If a
n
)
nN
is a non-decreasing sequence in A with a supremum a A, then
a = a
0
+

n=0
(a
n+1
\ a
n
) = lim
n
a
n
.
(c) If a
n
)
nN
is a non-increasing sequence in A with an inmum a A, then
a = a
0
(a
0
\ a) = a
0
lim
n
(a
0
\ a
n
) = lim
n
a
n
.
(d) If c A, then a (a c) is countably additive. (For sup
nN
a
n
c = c sup
nN
a
n
whenever the
right-hand-side is dened, by 313Ba.)
(e) is a countably additive functional for any R. If
t
is another countably additive functional on
A, then +
t
is countably additive.
(f ) If B is another Boolean algebra and : B A is a sequentially order-continuous Boolean homomor-
phism, then is a countably additive functional on B. (For if b
n
)
nN
is a disjoint sequence in B with
supremum b, then b
n
)
nN
is a disjoint sequence with supremum b.)
326I Additive functionals on Boolean algebras 107
(g) If A is Dedekind -complete and B is a -subalgebra of A, then B : B R is countably additive.
(For the identity map from B to A is sequentially order-continuous, by 314Hb.)
326G Corollary Let A be a Boolean algebra and a nitely additive real-valued functional on A.
(a) is countably additive i lim
n
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence in A with
inmum 0 in A.
(b) If
t
is an additive functional on A and [
t
a[ a for every a A, and is countably additive, then

t
is countably additive.
(c) If is non-negative, then is countably additive i it is sequentially order-continuous.
proof (a)(i) If is countably additive and a
n
)
nN
is a non-increasing sequence in A with inmum 0, then
lim
n
a
n
= 0 by 326Fc. (ii) If satises the condition, and a
n
)
nN
is a disjoint sequence in A with
supremum a, set b
n
= a \ sup
in
a
i
for each n N; then b
n
)
nN
is non-increasing and has inmum 0, so
a

n
i=0
a
i
= a (sup
in
a
i
) = b
n
0
as n , and a =

n=0
a
n
; thus is countably additive.
(b) If a
n
)
nN
is a disjoint sequence in A with supremum a, set b
n
= sup
in
a
i
for each n; then a =
lim
n
b
n
, so
lim
n
[
t
a
t
b
n
[ = lim
n
[
t
(a \ b
n
)[ lim
n
(a \ b
n
) = 0,
and

n=0

t
a
n
= lim
n

t
b
n
=
t
a.
(c) If is countably additive, then it is sequentially order-continuous by 326Fb-326Fc. If is sequentially
order-continuous, then of course it satises the condition of (a), so is countably additive.
326H The Jordan decomposition (II): Proposition Let A be a Boolean algebra and a bounded
countably additive real-valued functional on A. Then is expressible as the dierence of two non-negative
countably additive functionals.
proof Consider the functional
+
a = sup
ba
b dened in the proof of 326D. If a
n
)
nN
is a disjoint
sequence in A with supremum a, and b a, then
b =

n=0
(b a
n
)

n=0

+
a
n
.
As b is arbitrary,
+
a

n=0

+
a
n
. But of course

+
a
+
(sup
in
a
i
) =

n
i=0

+
a
i
for every n N, so
+
a =

n=0

+
a
n
. As a
n
)
nN
is arbitrary,
+
is countably additive.
Now

=
+
is also countably additive, and =
+

is the dierence of non-negative countably

additive functionals.
326I The Hahn decomposition: Theorem Let A be a Dedekind -complete Boolean algebra and
: A R a countably additive functional. Then is bounded and there is a c A such that a 0
whenever a c, while a 0 whenever a c = 0.
rst proof By 314M, there are a set X and a -algebra of subsets of X and a sequentially order-continuous
Boolean homomorphism from onto A. Set
1
= : R. Then
1
is countably additive (326Ff).
So
1
is bounded and there is a set H such that
1
F 0 whenever F and F H and
1
F 0
whenever F and F H = (231E). Set c = H A. If a c, then there is an F such that F = a;
now (F H) = a c = a, so a =
1
(F H) 0. If a c = 0, then there is an F such that F = a;
now (F H) = a \ c = a, so a =
1
(F H) 0.
second proof (a) Note rst that is bounded. PPP If a
n
)
nN
is a disjoint sequence in A, then

n=0
a
n
must exist and be equal to (sup
nN
a
n
); in particular, lim
n
a
n
= 0. By 326D, is bounded. QQQ
108 Measure algebras 326I
(b)(i) We know that = supa : a A < . Choose a sequence a
n
)
nN
in A such that a
n
2
n
for every n N. For m n N, set b
mn
= inf
min
a
i
. Then b
mn
2 2
m
+ 2
n
for every n m.
PPP Induce on n. For n = m, this is due to the choice of a
m
= b
mm
. For the inductive step, we have
b
m,n+1
= b
mn
a
n+1
, while surely (a
n+1
b
mn
), so
+b
m,n+1
(a
n+1
b
mn
) +(a
n+1
b
mn
)
= a
n+1
+b
mn
2
n1
+ 2 2
m
+ 2
n
(by the choice of a
n+1
and the inductive hypothesis)
= 2 2 2
m
+ 2
n1
.
Subtracting from both sides, b
m,n+1
2 2
m
+ 2
n1
and the induction proceeds. QQQ
(ii) Set
b
m
= inf
nm
b
mn
= inf
nm
a
n
.
Then
b
m
= lim
n
b
mn
2 2
m
,
by 326Fc. Next, b
n
)
nN
is non-decreasing, so setting c = sup
nN
b
n
we have
c = lim
n
b
n
;
since c is surely less than or equal to , c = .
If b A and b c, then
c b = (c \ b) = c,
so b 0. If b A and b c = 0 then
c +b = (c b) = c
so b 0. This completes the proof.
326J Completely additive functionals: Denition Let A be a Boolean algebra. A functional :
A R is completely additive or -additive if it is nitely additive and inf
aA
[a[ = 0 whenever A is a
non-empty downwards-directed set in A with inmum 0.
326K Basic facts Let A be a Boolean algebra and a completely additive real-valued functional on A.
(a) is countably additive. PPP If a
n
)
nN
is a non-increasing sequence in A with inmum 0, then for any
innite I N the set a
i
: i I is downwards-directed and has inmum 0, so inf
iI
[a
i
[ = 0; which means
that lim
n
a
n
must be zero. By 326Ga, is countably additive. QQQ
(b) Let A be a non-empty downwards-directed set in A with inmum 0. Then for every > 0 there is an
a A such that [b[ whenever b a. PPP??? Suppose, if possible, otherwise. Set
B = b : [b[ , a A, b a.
If a A there is a b
t
a such that [b
t
[ > . Now a
t
\ b
t
: a
t
A, a
t
a is downwards-directed and has
inmum 0, so there is an a
t
A such that a
t
a and [(a
t
\ b
t
)[ [b
t
[ . Set b = b
t
a
t
; then a
t
b and
[b[ = [b
t
+(a
t
\ b
t
)[ [b
t
[ [(a
t
\ b
t
)[ ,
so b B. But also b a. Thus every member of A includes some member of B. Since every member
of B includes a member of A, B is downwards-directed and has inmum 0; but this is impossible, since
inf
bB
[b[ . XXXQQQ
(c) If is non-negative, it is order-continuous. PPP (i) If A is a non-empty upwards-directed set with
supremum a
0
, then a
0
\ a : a A is a non-empty downwards-directed set with inmum 0, so
326M Additive functionals on Boolean algebras 109
sup
aA
a = a
0
inf
aA
(a
0
\ a) = a
0
.
(ii) If A is a non-empty downwards-directed set with inmum a
0
, then a \ a
0
: a A is a non-empty
downwards-directed set with inmum 0, so
inf
aA
a = a
0
+ inf
aA
(a \ a
0
) = a
0
. QQQ
(d) If c A, then a (a c) is completely additive. PPP If A is a non-empty downwards-directed set
with inmum 0, so is a c : a A, so inf
aA
[(a c)[ = 0. QQQ
(e) is a completely additive functional for any R. If
t
is another completely additive functional
on A, then +
t
is completely additive. PPP We know from 326Bc that +
t
is additive. Let A be a
non-empty downwards-directed set with inmum 0. For any > 0, (b) tells us that there are a, a
t
A such
that [b[ whenever b a
t
and [
t
b[ whenever b a
t
. But now, because A is downwards-directed,
there is a b A such that b a a
t
, which means that [b+
t
b[ [b[ +[
t
b[ is at most 2. As is arbitrary,
inf
aA
[( +
t
)(a)[ = 0, and +
t
is completely additive. QQQ
(f ) If B is another Boolean algebra and : B A is an order-continuous Boolean homomorphism,
then is a completely additive functional on B. PPP By 326Be, is additive. If B B is a non-empty
downwards-directed set with inmum 0 in B, then [B] is a non-empty downwards-directed set with inmum
0 in A, because is order-continuous, so inf
bB
[b[ = 0. QQQ In particular, if B is a regularly embedded
subalgebra of A, then B is completely additive.
(g) If
t
is another additive functional on A and [
t
a[ a for every a A, then
t
is completely additive.
PPP If A A is non-empty and downwards-directed and inf A = 0, then inf
aA
[
t
a[ inf
aA
a = 0. QQQ
326L I squeeze a useful fact in here.
Proposition If A is a ccc Boolean algebra, a functional : A R is countably additive i it is completely
additive.
proof If is completely additive it is countably additive, by 326Ka. If is countably additive and A is
a non-empty downwards-directed set in A with inmum 0, then there is a (non-empty) countable subset B
of A also with inmum 0 (316E). Let b
n
)
nN
be a sequence running over B, and choose a
n
)
nN
in A such
that a
0
= b
0
, a
n+1
a
n
b
n
for every n N. Then a
n
)
nN
is a non-increasing sequence with inmum 0,
so lim
n
a
n
= 0 (326Fc) and inf
aA
[a[ = 0. As A is arbitrary, is completely additive.
326M The Jordan decomposition (III): Proposition Let A be a Boolean algebra and a com-
pletely additive real-valued functional on A. Then is bounded and expressible as the dierence of two
non-negative completely additive functionals.
proof (a) I must rst check that is bounded. PPP Let a
n
)
nN
be a disjoint sequence in A. Set
A = a : a A, there is an n N such that a
i
a for every i n.
Then A is closed under , and if b is any lower bound for A then b 1 \ a
n
A, so b a
n
= 0, for every
n N; but this means that 1 \ b A, so that b 1 \ b and b = 0. Thus inf A = 0. By 326Kb, there is
an a A such that [b[ 1 whenever b a. By the denition of A, there must be an n N such that
[a
i
[ 1 for every i n. But this means that sup
nN
[a
n
[ is nite. As a
n
)
nN
is arbitrary, is bounded,
by 326D(ii). QQQ
(b) As in 326D and 326H, set
+
a = sup
ba
b for every a A. Then
+
is completely additive. PPP
We know that
+
is additive. If A is a non-empty downwards-directed subset of A with inmum 0, then for
every > 0 there is an a A such that [b[ whenever b a; in particular,
+
a . As is arbitrary,
inf
aA

+
a = 0; as A is arbitrary,
+
is completely additive. QQQ
Consequently

=
+
is completely additive (326Ke) and =
+

is the dierence of non-negative

completely additive functionals.
110 Measure algebras 326N
326N I give an alternative denition of completely additive which you may feel claries the concept.
Proposition Let A be a Boolean algebra, and : A R a function. Then the following are equiveridical:
(i) is completely additive;
(ii) 1 =

iI
a
i
whenever a
i
)
iI
is a partition of unity in A;
(iii) a =

iI
a
i
whenever a
i
)
iI
is a disjoint family in A with supremum a.
proof (For notes on sums

iI
, see 226A.)
(a)(i)(ii) If is completely additive and a
i
)
iI
is a partition of unity in A, then (inducing on #(J))
(sup
iJ
a
i
) =

iJ
a
i
for every nite J I. Consider
A = 1 \ sup
iJ
a
i
: J I is nite.
Then A is non-empty and downwards-directed and has inmum 0, so for every > 0 there is an a A such
that [b[ whenever b a. Express a as 1 \ sup
iJ
a
i
where J I is nite. If now K is another nite
subset of I including J,
[1

iK
a
i
[ = [(1 \ sup
iK
a
i
)[ .
As remarked in 226Ad, this means that 1 =

iI
a
i
, as claimed.
(b)(ii)(iii) Suppose that satises the condition (ii), and that a
i
)
iI
is a disjoint family with supre-
mum a. Take any j / I, set J = I j and a
j
= 1 \ a; then a
i
)
iJ
, (a, 1 \ a) are both partitions of unity,
so
(1 \ a) +a = 1 =

iJ
a
i
= (1 \ a) +

iI
a
i
,
and a =

iI
a
i
.
(c)(iii)(i) Suppose that satises (iii). Then is additive.
() is bounded. PPP Let a
n
)
nN
be a disjoint sequence in A. Applying Zorns Lemma to the set (
of all disjoint families C A including a
n
: n N, we nd a partition of unity C a
n
: n N. Now

cC
c is dened in R, so sup
nN
[a
n
[ sup
cC
[c[ is nite. By 326D, is bounded. QQQ
() Dene
+
from as in 326D. Then
+
satises the same condition as . PPP Let a
i
)
iI
be a disjoint
family in A with supremum a. Then for any b a, we have b = sup
iI
b a
i
, so
b =

iI
(b a
i
)

iI

+
a
i
.
Thus
+
a

iI

+
a
i
. But of course

iI

+
a
i
= sup

iJ

+
a
i
: J I is nite
= sup
+
(sup
iJ
a
i
) : J I is nite
+
a,
so
+
a =

iI

+
a
i
. QQQ
() It follows that
+
is completely additive. PPP If A is a non-empty downwards-directed set with
inmum 0, then B = b : a A, b a = 0 is order-dense in A, so there is a partition of unity b
i
)
iI
lying in B (313K). Now if J I is nite, there is an a A such that a sup
iJ
b
i
= 0 (because A is
downwards-directed), and

+
a +

iJ

+
b
i

+
1.
Since
+
1 = sup
JI is nite

iJ

+
b
i
, inf
aA

+
a = 0. As A is arbitrary,
+
is completely additive. QQQ
() Now consider

=
+
. Of course

a =
+
a a =

iI

+
a
i

iI
a
i
=

iI

a
i
whenever a
i
)
iI
is a disjoint family in A with supremum a. Because

is non-negative, the argument of

() shows that

= (

)
+
is completely additive. So =
+

is completely additive, as required.

*326Q Additive functionals on Boolean algebras 111
326O For completely additive functionals, we have a useful renement of the Hahn decomposition. I
give it in a form adapted to the applications I have in mind.
Proposition Let A be a Dedekind -complete Boolean algebra and : A R a completely additive
functional. Then there is a unique element of A, which I will denote [[ > 0]], the region where > 0, such
that a > 0 whenever 0 ,= a [[ > 0]], while a 0 whenever a [[ > 0]] = 0.
proof Set
C
1
= c : c A 0, a > 0 whenever 0 ,= a c,
C
2
= c : c A, a 0 whenever a c.
Then C
1
C
2
is order-dense in A. PPP There is a c
0
A such that a 0 for every a c, a 0 whenever
a c = 0 (326I). Given b A0, then b \ c
0
C
2
, so if b \ c
0
,= 0 we can stop. Otherwise, b c
0
. If b C
1
we can stop. Otherwise, there is a non-zero c b such that c 0; but in this case a 0, (c \ a) 0 so
a = 0 for every a c, and c C
2
. QQQ
There is therefore a partition of unity D C
1
C
2
. Now D C
1
is countable. PPP If d D C
1
, d > 0.
Also
#(d : d D, d 2
n
) 2
n
sup
aA
a
is nite for each n, so D C
1
is the union of a sequence of nite sets, and is countable. QQQ
Accordingly D C
1
has a supremum e. If 0 ,= a e then
a =

cD
(a c) =

cDC
1
(a c) 0
by 326N. Also there must be some c DC
1
such that a c ,= 0, in which case (a c) > 0, so that a > 0.
If a e = 0, then
a =

cD
(a c) =

cDC
2
(a c) 0.
Thus e has the properties demanded of [[ > 0]]. To see that e is unique, we need observe only that if
e
t
has the same properties then (e \ e
t
) 0 (because (e \ e
t
) e
t
= 0), so e \ e
t
= 0 (because e \ e
t
e).
Similarly, e
t
\ e = 0 and e = e
t
. Thus we may properly denote e by the formula [[ > 0]].
326P Corollary Let A be a Dedekind -complete Boolean algebra and , two completely additive
functionals on A. Then there is a unique element of A, which I will denote [[ > ]], the region where > ,
such that
a > a whenever 0 ,= a [[ > ]],
a a whenever a [[ > ]] = 0.
proof Apply 326O to the functional , and set [[ > ]] = [[ > 0]].
*326Q Additive functionals on free products In Volume 4, when we return to the construction of
measures on product spaces, the following fundamental fact will be useful.
Theorem Let A
i
)
iI
be a non-empty family of Boolean algebras, with free product A; write
i
: A
i
A
for the canonical maps, and
C = inf
jJ

j
(a
j
) : J I is nite, a
j
A
j
for every j J.
Suppose that : C R is such that
c = (c
i
(a)) +(c
i
(1 \ a))
whenever c C, i I and a A
i
. Then there is a unique nitely additive functional : A R extending
.
proof (a) It will help if I note at once that 0 = 0. PPP
0 = (0
i
(0)) +(0
i
(1)) = 20
112 Measure algebras *326Q
for any i I. QQQ
(b) The key is of course the following fact: if c
r
)
rm
and d
s
)
sn
are two disjoint families in C with
the same supremum in A, then

m
r=0
c
r
=

n
s=0
d
s
. PPP Let J I be a nite set and B
i
A
i
a nite
subalgebra, for each i J, such that every c
r
and every d
s
belongs to the subalgebra A
0
of A generated by

j
(b) : j J, b B
j
. Next, if j J and b B
j
, then

m
r=0
c
r
=

m
r=0
(c
r

j
(b)) +

m
r=0
(c
r
\
j
(b)).
We can therefore nd a disjoint family c
t
r
)
rm
in C A
0
such that
sup
rm
c
t
r
= sup
rm
c
r
,

m

r=0
c
t
r
=

m
r=0
c
r
,
and whenever r m
t
, j J and b B
j
then either c
t
r

j
(b) or c
t
r

j
(b)b = 0; that is, every c
t
r
is either 0
or of the form inf
jJ

j
(b
j
) where b
j
is an atom of B
j
for every j. Similarly, we can nd d
t
s
)
sn
such that
sup
sn
d
t
s
= sup
sn
d
s
,

n

s=0
d
t
s
=

n
s=0
d
s
,
and whenever s n
t
, j J and b B
j
then d
t
s
is either 0 or of the form inf
jJ

j
(b
j
) where b
j
is an atom
of B
j
for every j. But we now have sup
rm
c
t
r
= sup
sn
d
t
s
while for any r m
t
, s n
t
either c
t
r
= d
t
s
or
c
t
r
d
t
s
= 0. It follows that the non-zero terms in the nite sequence c
t
r
)
rm
are just a rearrangement of
the non-zero terms in d
t
s
)
sn
, so that

m
r=0
c
r
=

r=0
c
t
r
=

s=0
d
t
s
=

n
s=0
d
s
,
as required. QQQ
(c) By 315Jb, this means that we have a functional : A R such that (sup
rm
c
r
) =

m
r=0
c
r
whenever c
r
)
rm
is a disjoint family in C. It is now elementary to check that is additive, and it is clearly
the only additive functional on A extending .
326X Basic exercises (a) Let A be a Boolean algebra and : A R a nitely additive functional.
Show that (i) (a b) = a + b (a b) (ii) (a b c) = a + b + c (a b) (a c) (b c) +
(a b c) for all a, b, c A. Generalize these results to longer sequences in A.
(b) Let A be a Boolean algebra and : A R a nitely additive functional. Show that the following
are equiveridical: (i) is countably additive; (ii) lim
n
a
n
= a whenever a
n
)
nN
is a non-decreasing
sequence in A with supremum a.
(c) Let A be a Dedekind -complete Boolean algebra and : A R a nitely additive functional. Show
that the following are equiveridical: (i) is countably additive; (ii) lim
n
a
n
= 0 whenever a
n
)
nN
is a
sequence in A and inf
nN
sup
mn
a
m
= 0; (iii) lim
n
a
n
= a whenever a
n
)
nN
is a sequence in A and
a = inf
nN
sup
mn
a
m
= sup
nN
inf
mn
a
m
. (Hint: for (i)(iii), consider non-negative rst.)
(d) Let X be any uncountable set, and J an innite subset of X. Let A be the nite-conite algebra of X
(316Yk), and for a A set a = #(a J) if a is nite, #(J a) if a is conite. Show that is countably
additive and unbounded.
>>>(e) Let A be the algebra of subsets of [0, 1] generated by the family of (closed) intervals. Show that
there is a unique additive functional : A R such that [, ] = whenever 0 1. Show
that is countably additive but not completely additive.
(f ) (i) Let (X, , ) be any atomless probability space. Show that : R is a countably additive
functional which is not completely additive. (ii) Let X be any uncountable set and the countable-
cocountable measure on X (211R). Show that is countably additive but not completely additive.
(g) Let A be a Boolean algebra and : A R a function. (i) Show that is nitely additive i

iI
a
i
=
1 for every nite partition of unity a
i
)
iI
. (ii) Show that is countably additive i

iI
a
i
= 1 for
every countable partition of unity a
i
)
iI
.
326Yj Additive functionals on Boolean algebras 113
(h) Show that 326O can fail if is only countably additive, rather than completely additive. (Hint:
326Xf.)
(i) Let A be a Boolean algebra and a nitely additive real-valued functional on A. Let us say that
a A is a support of if () b = 0 whenever b a = 0 () for every non-zero b a there is a c b such
that c ,= 0. (i) Check that can have at most one support. (ii) Show that if a is a support for and
is bounded, then the principal ideal A
a
generated by a is ccc. (iii) Show that if A is Dedekind -complete
and is countably additive, then is completely additive i it has a support, and that in the language of
326O this is [[ > 0]] [[ > 0]]. (iv) Taking J = X in 326Xd, show that X is the support of the functional
there.
326Y Further exercises (a) Let A be a Boolean algebra and a non-negative nitely additive func-
tional on A. Show that the following are equiveridical: (i) for every > 0 there is a nite partition a
i
)
iI
of unity in A such that a
i
for every i I; (ii) whenever
t
is a non-zero nitely additive functional
such that 0
t
there is an a A such that
t
a and
t
(1 \ a) are both non-zero. (Such functionals are
called atomless.)
(b) Let A be a Boolean algebra and
1
,
2
atomless non-negative additive functionals on A. Show that

1
+
2
,
1
are atomless for every 0, and that is atomless whenever is additive and 0
1
.
(c) Let A be a Dedekind -complete Boolean algebra and an atomless non-negative nitely additive
functional on A. Show that there is a family a
t
)
t[0,1]
in A such that a
s
a
t
and a
t
= t1 whenever
s t [0, 1].
(d) Let A be a Dedekind -complete Boolean algebra and
0
, . . . ,
n
atomless non-negative nitely additive
functionals on A. Show that there is an a A such that
i
a =
1
2

i
1 for every i n. (Hint: it is enough
to consider the case
0

1
. . .
n
. For the inductive step, use the inductive hypothesis to construct
a
t
)
t[0,1]
such that a
s
a
t
,
i
a
t
= t
i
1 if i < n, 0 s t 1. Now show that t
n
(a
t+
1
2
\ a
t
) is
continuous on [0,
1
2
].)
(e) Let A be a Dedekind -complete Boolean algebra. Let : A R
r
, where r 1, be additive (in the
sense that (a b) = a +b whenever a b = 0) and atomless (in the sense that for every > 0 there is a
nite partition of unity a
i
)
iI
such that |a| whenever i I and a a
i
). Show that a : a A is
a convex set in R
r
. (This is a version of Liapounos theorem. I am grateful to K.P.S.Bhaskara Rao for
showing it to me.)
(f ) Let A be a Dedekind -complete Boolean algebra and : A [0, [ a countably additive functional.
Show that is atomless i whenever a A and a ,= 0 there is a b a such that 0 < b < a.
(g) Show that there is a nitely additive functional : TN R such that n = 1 for every n N, so
that is not bounded. (Hint: Use Zorns Lemma to construct a maximal linearly independent subset of

including n : n N, and hence to construct a linear map f :

R such that f(n) = 1 for every

n.)
(h) Let A be any innite Boolean algebra. Show that there is an unbounded nitely additive functional
: A R. (Hint: let t
n
)
nN
be a sequence of distinct points in the Stone space of A, and set a =
t
n :
t
n
a for a suitable
t
.)
(i) Let A be a Boolean algebra, and give R
A
its product topology. Show that the space of nitely additive
functionals on A is a closed subset of R
A
, but that the space of bounded nitely additive functionals is closed
only when A is nite.
(j) Let A be a Boolean algebra, and M the linear space of all bounded nitely additive real-valued
functionals on A. For ,
t
M say that
t
if a
t
a for every a A. Show that
(i)
+
, as dened in the proof of 326D, is just sup0, in M;
(ii) M is a Dedekind complete Riesz space (241E-241F, 353G);
114 Measure algebras 326Yj
(iii) for ,
t
M, [[ = (),
t
,
t
are given by the formulae
[[(a) = sup
ba
b (a \ b), (
t
)(a) = sup
ba
b +
t
(a \ b),
(
t
)(a) = inf
ba
b +
t
(a \ b);
(iv) for any non-empty A M, A is bounded above in M i
sup

n
i=0

i
a
i
:
i
A for each i n, a
i
)
in
is disjoint
is nite, and then sup A is dened by the formula
(supA)(a) = sup

n
i=0

i
a
i
:
i
A for each i n, a
i
)
in
is disjoint, sup
in
a
i
= a
for every a A;
(v) setting || = [[(1), | | is an order-continuous norm on M under which M is a Banach lattice.
(k) Let A be a Boolean algebra. A functional : A C is nitely additive if its real and imaginary
parts are. Show that the space of bounded nitely additive functionals from A to C is a Banach space under
the norm || = sup

n
i=0
[a
i
[ : a
i
)
in
is a partition of unity in A.
(l) Let A be a Boolean algebra, and give it the topology T

for which the closed sets are the sequentially

order-closed sets. Show that a nitely additive functional : A R is countably additive i it is continuous
for T

.
(m) Let A be a Boolean algebra, and M

the set of all bounded countably additive real-valued functionals

on A. Show that M

is a closed and order-closed linear subspace of the normed space M of all additive
functionals on A (326Yj), and that [[ M

whenever M

.
(n) Let A be a Boolean algebra and a non-negative nitely additive functional on A. Set

a = infsup
nN
a
n
: a
n
)
nN
is a non-decreasing sequence with supremum a
for every a A. Show that

is countably additive, and is sup

t
:
t
is countably additive.
(o) Let A be a Dedekind -complete Boolean algebra and
n
)
nN
a sequence of countably additive real-
valued functionals on A such that a = lim
n

n
a is dened in R for every a A. Show that is countably
additive. (Hint: use arguments from part (a) of the proof of 247C to see that lim
n
sup
kN
[
k
a
n
[ = 0 for
every disjoint sequence a
n
)
nN
in A, and therefore that lim
n
sup
kN
[
k
a
n
[ = 0 whenever a
n
)
nN
is a
non-increasing sequence with inmum 0.)
(p) Let A be a Boolean algebra, and M

the set of all completely additive real-valued functionals on A.

Show that M

is a closed and order-closed linear subspace of the normed space M of all additive functionals,
and that [[ M

whenever M

.
(q) Let A be a Boolean algebra and a non-negative nitely additive functional on A. Set

b = infsup
aA
a : A is a non-empty upwards-directed set with supremum b
for every b A. Show that

is completely additive, and is sup

t
:
t
is completely additive.
(r) Let A be a Boolean algebra, and give it the topology T for which the closed sets are the order-closed
sets (313Xb). Show that a nitely additive functional : A R is completely additive i it is continuous
for T.
(s) Let X be a set, any -algebra of subsets of X, and : R a functional. Show that is completely
additive i there are sequences x
n
)
nN
,
n
)
nN
such that

n=0
[
n
[ < and E =

n=0

n
E(x
n
) for
every E .
(t) Let A and B be Boolean algebras and , nitely additive functionals on A, B respectively. Show
that there is a unique nitely additive functional : A B R such that (a b) = a b for all a A,
b B.
326 Notes Additive functionals on Boolean algebras 115
(u) Let A
i
)
iI
be a family of Boolean algebras, with free product (

iI
A
i
,
i
)
iI
), and for each i I
let
i
be a nitely additive functional on A
i
such that
i
1 = 1. Show that there is a unique nitely additive
functional :

iI
A
i
R such that (inf
iJ

i
(a
i
)) =

iJ

i
a
i
whenever J I is non-empty and nite
and a
i
A
i
for each i J.
326 Notes and comments I have not mentioned the phrase measure algebra anywhere in this section, and
in principle this material could have been part of Chapter 31; but countably additive functionals are kissing
cousins of measures, and most of the ideas here surely belong to measure theory rather than to Boolean
algebra, in so far as such divisions are meaningful at all. I have given as much as possible of the theory
in a general form because the simplications which are possible when we look only at measure algebras are
seriously confusing if they are allowed too much prominence. In particular, it is important to understand
that the principal properties of completely additive functionals do not depend on Dedekind completeness
of the algebra, provided we take care over the denitions. Similarly, the denition of countably additive
functional for algebras which are not Dedekind -complete needs a moments attention to the phrase and
sup
nN
a
n
is dened in A. It can happen that a functional is countably additive mostly because there are
too few such sequences (326Xd).
The formulations I have chosen as principal denitions (326A, 326E, 326J) are those which I nd closest
to my own intuitions of the concepts, but you may feel that 326G(i), 326Xc(iii) and 326N, or 326Yl and
326Yr, provide useful alternative patterns. The point is that countable additivity corresponds to sequential
order-continuity (326Fb, 326Fc, 326Ff), while complete additivity corresponds to order-continuity (326Kc,
326Kf); the diculty is that we must consider functionals which are not order-preserving, so that the simple
denitions in 313H cannot be applied directly. It is fair to say that all the additive functionals we need
to understand are bounded, and therefore may be studied in terms of their positive and negative parts
+
,

, which are order-preserving (326Bf); but many of the most important applications of these ideas depend
precisely on using facts about to deduce facts about
+
and

.
It is in 326D that we seem to start getting more out of the theory than we have put in. The ideas here have
vast ramications. What it amounts to is that we can discover much more than we might expect by looking
at disjoint sequences. To begin with, the conditions here lead directly to 326I and 326M: every completely
additive functional is bounded, and every countably additive functional on a Dedekind -complete Boolean
algebra is bounded. (But note 326Yg-326Yh.)
Naturally enough, the theory of countably additive functionals on general Boolean algebras corresponds
closely to the special case of countably additive functionals on -algebras of sets, already treated in 231-232
for the sake of the Radon-Nikod ym theorem. This should make 326E-326I very straightforward. When we
come to completely additive functionals, however, there is room for many surprises. The natural map from
a -algebra of measurable sets to the corresponding measure algebra is sequentially order-continuous but
rarely order-continuous, so that there can be completely additive functionals on the measure algebra which
do not correspond to completely additive functionals on the -algebra. Indeed there are very few completely
additive functionals on -algebras of sets (326Ys). Of course these surprises can arise only when there is a
dierence between completely additive and countably additive functionals, that is, when the algebra involved
is not ccc (326L). But I think that neither 326M nor 326N is obvious.
I nd myself generally using the phrase countably additive in preference to completely additive in the
context of ccc algebras, where there is no dierence between them. This is an attempt at user-friendliness;
the phrase countably additive is the commoner one in ordinary use. But I must say that my personal
inclination is to the other side. The reason why so many theorems apply to countably additive functionals
in these contexts is just that they are completely additive.
I have given two proofs of 326I. I certainly assume that if you have got this far you are acquainted with the
Radon-Nikod ym theorem and the associated basic facts about countably additive functionals on -algebras
of sets; so that the rst proof should be easy and natural. On the other hand, there are purist objections
on two fronts. First, it relies on the Stone representation, which involves a much stronger form of the axiom
of choice than is actually necessary. Second, the classical Hahn decomposition in 231E is evidently a special
case of 326I, and if we need both (as we certainly do) then one expects the ideas to stand out more clearly if
they are applied directly to the general case. In fact the two versions of the argument are so nearly identical
that (as you will observe, if you have Volume 2 to hand) they can share nearly every word. You can take
the second proof, therefore, as a worked example in the translation of ideas from the context of -algebras
116 Measure algebras 326 Notes
of sets to the context of Dedekind -complete Boolean algebras. What makes it possible is the fact that the
only limit operations referred to involve countable families.
Arguments not involving limit operations can generally, of course, be applied to all Boolean algebras; I
have lifted some exercises (326Yj, 326Yn) from 231 to give you some practice in such generalizations.
Almost any non-trivial measure provides an example of a countably additive functional on a Dedekind
-complete algebra which is not completely additive (326Xf). The question of whether such a functional
can exist on a Dedekind complete algebra is the Banach-Ulam problem, to which I will return in 363S.
In this section I have looked only at questions which can be adequately treated in terms of the underlying
algebras A, without using any auxiliary structure. To go much farther we shall need to study the function
spaces S(A) and L

(A) of Chapter 36. In particular, the ideas of 326Yg, 326Yj-326Yk and 326Ym-326Yq
will make better sense when redeveloped in 362.
327 Additive functionals on measure algebras
When we turn to measure algebras, we have a simplication, relative to the general context of 326,
because the algebras are always Dedekind -complete; but there are also elaborations, because we can ask
how the additive functionals we examine are related to the measure. In 327A-327C I work through the
relationships between the concepts of absolute continuity, (true) continuity and countable additivity,
following 232A-232B, and adding complete additivity from 326. These ideas provide a new interpretation
of the Radon-Nikod ym theorem (327D). I then use this theorem to develop some machinery (the standard
extension of an additive functional from a closed subalgebra to the whole algebra, 327F-327G) which will
be used in 333.
327A I start with the following denition and theorem corresponding to 232A-232B.
Denition Let (A, ) be a measure algebra and : A R a nitely additive functional. Then is
absolutely continuous with respect to if for every > 0 there is a > 0 such that [a[ whenever
a .
327B Theorem Let (A, ) be a measure algebra, and : A R a nitely additive functional. Give A
its measure-algebra topology and uniformity (323).
(a) If is continuous, it is completely additive.
(b) If is countably additive, it is absolutely continuous with respect to .
(c) The following are equiveridical:
(i) is continuous at 0;
(ii) is countably additive and whenever a A and a ,= 0 there is a b A such that b < and
(a b) ,= 0;
(iii) is continuous everywhere on A;
(iv) is uniformly continuous.
(d) If (A, ) is semi-nite, then is continuous i it is completely additive.
(e) If (A, ) is -nite, then is continuous i it is countably additive i it is completely additive.
(f) If (A, ) is totally nite, then is continuous i it is absolutely continuous with respect to i it is
countably additive i it is completely additive.
proof (a) If is continuous, and A A is non-empty, downwards-directed and has inmum 0, then 0 A
(323D(b-ii)), so inf
aA
[a[ = 0.
(b) ??? Suppose, if possible, that is countably additive but not absolutely continuous. Then there is an
> 0 such that for every > 0 there is an a A such that a but [a[ . For each n N we may
choose a b
n
A such that b
n
2
n
and [b
n
[ . Consider b

n
= sup
kn
b
k
, b = inf
nN
b

n
. Then we have
b inf
nN
(sup
kn
b
k
) inf
nN

k=n
2
k
= 0,
so b = 0 and b = 0. On the other hand, is expressible as a dierence
+

of non-negative countably
additive functionals (326H), each of which is sequentially order-continuous (326Gc), and
327C Additive functionals on measure algebras 117
0 = lim
n
(
+
+

)b

n
inf
nN
(
+
+

)b
n
inf
nN
[b
n
[ ,
which is absurd. XXX
(c)(i)(ii) Suppose that is continuous. Then it is completely additive, by (a), therefore countably
additive. If a ,= 0, there must be an b of nite measure such that [d[ < [a[ whenever d b = , so that
[(a \ b)[ < [a[ and (a b) ,= 0. Thus the conditions are satised.
(ii)(iv) Now suppose that satises the two conditions in (ii). Because A is Dedekind -complete,
must be bounded (326I), therefore expressible as the dierence
+

of countably additive functionals.

Set
1
=
+
+

. Set
= sup
1
b : b A, b < ,
and choose a sequence b
n
)
nN
of elements of A of nite measure such that lim
n

1
b
n
= ; set b

=
sup
nN
b
n
. If d A and d b

= then d = 0. PPP If b A and b < , then

[(d b)[
1
(d b)
1
(b \ b
n
) =
1
(b b
n
)
1
b
n

1
b
n
for every n N, so (d b) = 0. As b is arbitrary, the second condition tells us that d = 0. QQQ
Setting b

n
= sup
kn
b
k
for each n, we have lim
n

1
(b

\ b

n
) = 0. Take any > 0, and (using (b) above)
let > 0 be such that [a[ whenever a . Let n be such that
1
(b

\ b

n
) . Then
[a[ [(a b

n
)[ +[(a (b

\ b

n
))[ +[(a \ b

)[
[(a b

n
)[ +
1
(b

\ b

n
) [(a b

n
)[ +
for any a A.
Now if b, c A and ((b .c) b

n
) then
[b c[ [(b \ c)[ +[(c \ b)[
[((b \ c) b

)[ +[((c \ b) b

)[ + 2 + + 2 = 4
because ((b \ c) b

n
), ((c \ b) b

n
) are both less than or equal to . As is arbitrary, is uniformly
continuous.
(iv)(iii)(i) are trivial.
(d) One implication is covered by (a). For the other, suppose that is completely additive. Then it is
countably additive. On the other hand, if a ,= 0, consider B = b : b a, b < . Then B is upwards-
directed and sup B = a, because is semi-nite (322Eb), so a \ b : b B is downwards-directed and has
inmum 0. Accordingly inf
bB
[(a \ b)[ = 0, and there must be a b B such that b ,= 0. But this means
that condition (ii) of (c) is satised, so that is continuous.
(e) Now suppose that (A, ) is -nite. In this case A is ccc (322G) so complete additivity and countable
additivity are the same (326L) and we have a special case of (d).
(f ) Finally, suppose that 1 < and that is absolutely continuous with respect to . If A A is
non-empty and downwards-directed and has inmum 0, then inf
aA
a = 0 (321F), so inf
aA
[a[ must be
0; thus is completely additive. With (b) and (e) this shows that all four conditions are equiveridical.
327C Proposition Let (X, , ) be a measure space and (A, ) its measure algebra.
(a) There is a one-to-one correspondence between nitely additive functionals on A and nitely additive
functionals on such that E = 0 whenever E = 0, given by the formula E

= E for every E .
(b) In (a), is absolutely continuous with respect to i is absolutely continuous with respect to .
(c) In (a), is countably additive i is countably additive; so that we have a one-to-one correspondence
between the countably additive functionals on A and the absolutely continuous countably additive functionals
on .
(d) In (a), is continuous for the measure-algebra topology on A i is truly continuous in the sense of
232Ab.
(e) Suppose that is semi-nite. Then, in (a), is completely additive i is truly continuous.
118 Measure algebras 327C
proof (a) This should be nearly obvious. If : A R is additive, then the formula denes a functional
: R which is additive by 326Be. Also, of course,
E = 0 = E

= 0 = E = 0.
On the other hand, if is an additive functional on which is zero on negligible sets, then, for E, F ,
E

= F

= (E F) = (F E) = 0
= (E F) = (F E) = 0
= F = E (E F) +(F E) = E,
so we have a function : A R dened by the given formula. If E, F and E

= 0, then
(E

) = (E F)

= (E F)
= (E F) +F = E

+ F

because (E F)

= E

\ F

= E

. Thus is additive, and the correspondence is complete.

(b) This is immediate from the denitions.
(c)(i) If is countably additive, and a
n
)
nN
is a disjoint sequence in A, we can express it as E
n
)
nN
where E
n
)
nN
is a sequence in . Setting F
n
= E
n

i<n
E
i
, F
n
)
nN
is a disjoint sequence in and
F

n
= a
n
\ sup
i<n
a
i
= a
n
for each n. So
(sup
nN
a
n
) = (

nN
F
n
) =

n=0
F
n
=

n=0
a
n
.
As a
n
)
nN
is arbitrary, is countably additive.
(ii) If is countably additive, then is countably additive by 326Ff.
(iii) For the last remark, note that by 232Ba a countably additive functional on is absolutely
continuous with respect to i it is zero on the -negligible sets.
(d) The denition of truly continuous functional translates directly to continuity at 0 in the measure
algebra. But by 327Bc this is the same thing as continuity.
(e) Put (d) and 327Bd together.
327D The Radon-Nikod ym theorem We are now ready for another look at this theorem.
Theorem Let (X, , ) be a semi-nite measure space, with measure algebra (A, ). Let L
1
be the space of
equivalence classes of real-valued integrable functions on X (242), and write M

for the set of completely

additive real-valued functionals on A. Then there is an ordered linear space bijection between M

and L
1
dened by saying that M

corresponds to u L
1
if
a =
_
E
f whenever a = E

in A and f

= u in L
1
.
proof (a) Given M

, we have a truly continuous : R given by setting E = E

for every
E (327Ce). Now there is an integrable function f such that E =
_
E
f for every E (232E). There
is likely to be more than one such function, but any two must be equal almost everywhere (232Hd), so the
corresponding equivalence class u

= f

is uniquely dened.
(b) Conversely, given u L
1
, we have a well-dened functional
u
on given by setting

u
E =
_
E
u =
_
E
f whenever f

= u
for every E (242Ac). By 232E,
u
is additive and truly continuous, and of course it is zero when is
zero, so corresponds to a completely additive functional
u
on A (327Ce).
(c) Clearly the maps u
u
and u

are now the two halves of a one-to-one correspondence. To see
that it is linear, we need note only that
327F Additive functionals on measure algebras 119
(
u
+
v
)E

=
u
E

+
v
E

=
_
E
u +
_
E
v =
_
E
u +v =
u+v
E

for every E , so
u
+
v
=
u+v
for all u, v L
1
; and similarly
u
=
u
for u L
1
, R. As for the
ordering, given u and v L
1
, take integrable f, g such that u = f

, v = g

; then

u

v

u
E

v
E

for every E

_
E
u
_
E
v for every E

_
E
f
_
E
g for every E
f g a.e. u v,
using 131Ha.
327E I slip in an elementary fact.
Proposition If (A, ) is a measure algebra, then the functional a
c
a = (a c) is completely additive
whenever c A and c < .
proof
c
is additive because is additive, and by 321F inf
aA

c
a = 0 whenever Ais non-empty, downwards-
directed and has inmum 0.
327F Standard extensions The machinery of 327D provides the basis of a canonical method for
extending countably additive functionals from closed subalgebras, which we shall need in 333.
Lemma Let (A, ) be a totally nite measure algebra and C A a closed subalgebra. Write M

(A), M

(C)
for the spaces of countably additive real-valued functionals on A, C respectively.
(a) There is an operator R : M

(C) M

(A) dened by saying that, for every M

(C), R is the
unique member of M

(A) such that [[R > ]] = [[ > C]] for every R.

(b)(i) R extends for every M

(C).
(ii) R is linear and order-preserving.
(iii) R( C) = .
(iv) If
n
)
nN
is a sequence of non-negative functionals in M

(C) such that

n=0

n
c = c for every
c C, then

n=0
(R
n
)(a) = a for every a A.
Remarks When saying that C is closed, I mean, indierently, topologically closed or order-closed; see
323H-323I.
For the notation [[ > ]] see 326O-326P.
proof (a)(i) By 321J-321K, we may represent (A, ) as the measure algebra of a measure space (X, , );
write for the canonical map from to A. Write T for E : E , E C. Because C is a -subalgebra
of C and is a sequentially order-continuous Boolean homomorphism, T is a -subalgebra of .
(ii) For each M

(C), : T R is countably additive and zero on F : F T, F = 0, so we

can choose a T-measurable function f

: X R such that
_
F
f

d(T) = F for every F T. Of course

we can now think of f

as a -integrable function (233B), so we get a corresponding countably additive

functional R : A R dened by setting (R)(E) =
_
E
f

for every E (327D). (In this context, of

course, countably additive functionals are completely additive, by 327Bf.)
For R, set H

= x : f

(x) > T. Then for any E ,

E H

, E > 0 =
_
E
f

> E,
E H

= =
_
E
f

E.
Translating into terms of elements of A, and setting c

= H

C, we have
0 ,= a c

= (R)(a) > a,
a c

= 0 = (R)(a) a.
120 Measure algebras 327F
So [[R > ]] = c

C. Of course we now have

c = (R)(c) > c when c C, 0 ,= c c

,
c c when c C, c c

= 0,
so that c

is also equal to [[ > C]].

Thus the functional R satises the declared formula.
(iii) To see that R is uniquely dened, observe that if M

(A) and [[ > ]] = [[R > ]] for

every , then there is a -measurable function g : X R such that
_
E
g d = E for every E ; but
in this case (just as in (ii)) [[ > ]] = G

, where G

= x : g(x) > , for each . So we must have

G

= H

, that is, (G

) = 0, for every . Accordingly

x : f

(x) ,= g(x) =

qQ
G
q
H
q
is negligible; f

= g a.e.,
_
E
f

d =
_
E
g d for every E and = R.
(b)(i) If M

(C),
(R)(F) =
_
F
f

d =
_
F
f

d(T) = F
for every F T, so R extends .
(ii) If =
1
+
2
, we must have
_
F
f

= F =
1
F +
2
F =
_
F
f

1
+
_
F
f

2
=
_
F
f

1
+f

2
for every F T, so f

= f

1
+f

2
a.e., and we can repeat the formulae
(R)(E) =
_
E
f

=
_
E
f

1
+f

2
=
_
E
f

1
+
_
E
f

2
= (R
1
)(E) + (R
2
)(E),
in a dierent order, for every E , to see that R = R
1
+ R
2
. Similarly, if M

(C) and R,
f

= f

a.e. and R() = R. If

1

2
in M

(C), then
_
F
f

1
=
1
F
2
F =
_
F
f

2
for every F T, so f

1
f

2
a.e. (131Ha), and R
1
R
2
.
Thus R is linear and order-preserving.
(iii) If = C then
_
F
f

= F = F =
_
F
1
for every F T, so f

= 1 a.e. and R = .
(iv) Now suppose that
n
)
nN
is a sequence in M

(C) such that, for every c C,

n
c 0 for every n
and

n=0

n
c = c. Set g
n
=

n
i=0
f

i
for each n; then 0 g
n
g
n+1
1 a.e. for every n, and
lim
n
_
g
n
= lim
n

n
i=0

i
1 = 1.
But this means that, setting g = lim
n
g
n
, g 1 a.e. and
_
g =
_
1, so that g = 1 a.e. and

n=0
(R
i
)(E) = lim
n
_
E
g
n
= E
for every E , so that

n=0
(R
i
)(a) = a for every a A.
327G Denition In the context of 327F, I will call R the standard extension of to A.
Remark The point of my insistence on the uniqueness of R, and on the formula in 327Fa, is that R really
is dened by the abstract structure (A, , C, ), even though I have used a proof which runs through the
representation of (A, ) as the measure algebra of a measure space (X, , ).
327X Basic exercises (a) Let (A, ) and (B,
t
) be totally nite measure algebras, and : A B a
measure-preserving Boolean homomorphism. Let C be a closed subalgebra of A, and a countably additive
functional on the closed subalgebra [C] of B. (i) Show that is a countably additive functional on C. (ii)
Show that if is the standard extension of to B, then is the standard extension of to A.
327 Notes Additive functionals on measure algebras 121
(b) Let (X, , ) be a probability space, and T a -subalgebra of . Let (A, ) be the measure algebra
of (X, , ). Show that C = F

: F T is a closed subalgebra of A. Identify the spaces M

(A), M

(C)
of countably additive functionals with L
1
(), L
1
(T), as in 327D. Show that the conditional expectation
operator P : L
1
() L
1
(T) (242Jd) corresponds to the map C : M

(A) M

(C).
(c) Let (A, ) be a totally nite measure algebra and : A R a countably additive functional. Show
that, for any a A,
a =
_

0
(a [[ > ]])d
_
0

(a \ [[ > ]])d,
the integrals being taken with respect to Lebesgue measure. (Hint: take (A, ) to be the measure algebra
of (X, , ); represent by a -integrable function f; apply Fubinis theorem to the sets (x, t) : x E, 0
t < f(x), (x, t) : x E, f(x) t 0 in X R, where a = E

.)
(d) Let (A, ) be a totally nite measure algebra, C a closed subalgebra of A and : C R a countably
additive functional with standard extension : A R. Show that, for any a A,
a =
_

0
(a [[ > C]])d
_
0

(a \ [[ > C]])d.
(e) Let (A, ) be a probability algebra, and B, C stochastically independent closed subalgebras of A
(denition: 325L). Let be a countably additive functional on C, and its standard extension to A. Show
that (b c) = b c for every b B, c C.
(f ) Let (X, , ) be a probability space, and T a -subalgebra of . Let be a probability measure with
domain T such that E = 0 whenever E T and E = 0. Show that there is a probability measure with
domain which extends .
327Y Further exercises (a) Let (A
1
,
1
) and (A
2
,
2
) be localizable measure algebras with localizable
measure algebra free product (C,

). Show that if
1
,
2
are completely additive functionals on A
1
, A
2
respectively, there is a unique completely additive functional : C R such that (a
1
a
2
) =
1
a
1

2
a
2
for every a
1
A
1
, a
2
A
2
. (Hint: 253D.)
(b) Let (A, ) be a totally nite measure algebra and C a closed subalgebra; let R : M

(C) M

(A)
be the standard extension operator (327G). Show (i) that R is order-continuous (ii) that R(
+
) = (R)
+
,
|R| = || for every M

(C), dening
+
and || as in 326Yj.
(c) Let (A, ) be a totally nite measure algebra and C a closed subalgebra of A. For a countably additive
functional on C write for its standard extension to A. Show that if ,
n
)
nN
are countably additive
functionals on C and lim
n

n
c = c for every c C, then lim
n

n
a = a for every a A. (Hint: use
ideas from 246-247, as well as from 327G and 326Yo.)
327 Notes and comments When we come to measure algebras, it is the completely additive functionals
which t most naturally into the topological theory (327Bd); they correspond to the truly continuous
functionals which I discussed in 232 (327Cc), and therefore to the Radon-Nikod ym theorem (327D). I will
return to some of these questions in Chapter 36. I myself regard the form here as the best expression of the
essence of the Radon-Nikod ym theorem, if not the one most commonly applied.
The concept of standard extension of a countably additive functional (or, as we could equally well say,
of a completely additive functional, since in the context of 327F the two coincide) is in a sense dual to
the concept of conditional expectation. If (X, , ) is a probability space and T is a -subalgebra of ,
then we have a corresponding closed subalgebra C of the measure algebra (A, ) of , and identications
between the spaces M

(A), M

(C) of countably additive functionals and the spaces L

1
(), L
1
(T). Now
we have a natural embedding S of L
1
(T) as a subspace of L
1
() (242Jb), and a natural restriction map
from M

(A) to M

(C). These give rise to corresponding operators between the opposite members of each
pair; the standard extension operator R of 327G, and the conditional expectation operator P of 242Jd. (See
327Xb.) The fundamental fact
122 Measure algebras 327 Notes
PSv = v for every v L
1
(T)
(242Jg) is matched by the fact that
RC = for every M

(C).
The further identication of R in terms of integrals
_
(a [[ > ]])d (327Xc) is relatively inessential,
but is striking, and perhaps makes it easier to believe that R is truly standard in the abstract contexts
which will arise in 333 below. It is also useful in such calculations as 327Xe.
The isomorphisms between M

spaces and L
1
spaces described here mean that any of the concepts
involving L
1
spaces discussed in Chapter 24 can be applied to M

spaces, at least in the case of measure

algebras. In fact, as I will show in Chapter 36, there is much more to be said here; the space of bounded
additive functionals on a Boolean algebra is already an L
1
space in an abstract sense, and ideas such as
uniform integrability are relevant and signicant there, as well as in the spaces of countably additive and
completely additive functionals. I hope that 326Yj, 326Ym-326Yn, 326Yp-326Yq and 327Yb will provide
some hints to be going on with for the moment.
331B Classication of homogeneous measure algebras 123
Chapter 33
Maharams theorem
We are now ready for the astonishing central fact about measure algebras: there are very few of them.
Any localizable measure algebra has a canonical expression as a simple product of measure algebras of
easily described types. This complete classication necessarily dominates all further discussion of measure
algebras; to the point that all the results of Chapter 32 have to be regarded as elementary, since however
complex their formulation they have been proved by techniques not involving, nor providing, any particular
insight into the special nature of measure algebras. The proof depends, of course, on developing methods
of dening measure-preserving homomorphisms and isomorphisms; I give a number of results, progressively
more elaborate, but all based on the same idea. These techniques are of great power, leading, for instance,
to an eective classication of closed subalgebras and their embeddings.
Maharams theorem itself, the classication of localizable measure algebras, is in 332. I devote 331
to the denition and description of homogeneous probability algebras. In 333 I turn to the problem of
describing pairs (A, C) where A is a probability algebra and C is a closed subalgebra. Finally, in 334, I give
some straightforward results on the classication of free products of probability algebras.
331 Classication of homogeneous measure algebras
I embark directly on the principal theorem of this chapter (331I), split between 331B, 331D and 331I;
331B and 331D will be the basis of many of the results in later sections of this chapter. In 331E-331H
I introduce the concepts of Maharam type and Maharam homogeneity. I discuss the measure algebras
of products 0, 1

, showing that these provide a complete set of examples of Maharam-type-homogeneous

probability algebras (331J-331L). I end the section with a brief comment on homogeneous Boolean algebras
(331M-331N).
331A Denition The following idea is almost the key to the whole chapter. Let A be a Boolean algebra
and B an order-closed subalgebra of A. A non-zero element a of A is a relative atom over B if every c a
is of the form a b for some b B; that is, a b : b B is the principal ideal generated by a. We say
that A is relatively atomless over B if there are no relative atoms in A over B.
(Im afraid the phrases relative atom, relatively atomless are bound to seem opaque at this stage. I
hope that after the structure theory of 333 they will seem more natural. For the moment, note only that
a is an atom in A i it is a relative atom over the smallest subalgebra 0, 1, and every element of A is a
relative atom over the largest subalgebra A. In a way, a is a relative atom over B if its image is an atom in
a kind of quotient A/B. But we are two volumes away from any prospect of making sense of this kind of
quotient.)
331B The rst lemma is the heart of Maharams theorem.
Lemma Let (A, ) be a totally nite measure algebra and B a closed subalgebra of A such that A is
relatively atomless over B. Let : B R be an additive functional such that 0 b b for every b B.
Then there is a c A such that b = (b c) for every b B.
Remark Recall that by 323H we need not distinguish between order-closed and topologically closed
subalgebras.
proof (a) It is worth noting straight away that is necessarily countably additive. This is easy to check
from rst principles, but if you want to trace the underlying ideas they are in 313O (the identity map from
B to A is order-continuous), 326Ff (so B : B R is countably additive) and 326Gb (therefore is
countably additive).
(b) For each a A set
a
b = (b a) for every b B; then
a
is countably additive (326Fd). The key
idea is the following fact: for every non-zero a A there is a non-zero d a such that
d

1
2

a
. PPP Because
A is relatively atomless over B, there is an e a such that e ,= a b for any b B. Consider the countably
124 Maharams theorem 331B
additive functional =
a
2
e
: B R. By 326I, there is a b
0
B such that b 0 whenever b B,
b b
0
, while b 0 whenever b B, b b
0
= 0.
If e b
0
,= 0, try d = e b
0
. Then 0 ,= d a, and for every b B

d
b =
e
(b b
0
) =
1
2
(
a
(b b
0
) (b b
0
))
1
2

a
b
(because (b b
0
) 0) so
d

1
2

a
.
If e b
0
= 0, then (by the choice of e) e ,= a (1 \ b
0
), so d = a \ (e b
0
) ,= 0, and of course d a. In this
case, for every b B,

d
b =
a
(b \ b
0
)
e
(b \ b
0
) =
1
2
((b \ b
0
) +
a
(b \ b
0
))
1
2

a
b
(because (b \ b
0
) 0), so once again
d

1
2

a
.
Thus in either case we have a suitable d. QQQ
(c) It follows at once, by induction on n, that if a is any non-zero element of A and n N then there is
a non-zero d a such that
d
2
n

a
.
(d) Now let C be the set
a : a A,
a
.
Then 0 C, so C ,= . If D C is upwards-directed and not empty, then a = sup D is dened in A, and

sup D
b = (b supD) = (sup
dD
b d) = sup
dD
(b d) = sup
dD

d
b b
using 313Ba and 321C. So a C and is an upper bound for D in C. In particular, any non-empty totally
ordered subset of C has an upper bound in C. By Zorns Lemma, C has a maximal element c say.
(e) ??? Suppose, if possible, that
c
,= . Then there is some b

B such that
c
b

,= b

; since
c
,

c
b

< b

. Let n N be such that b

>
c
b

+ 2
n
b

, and set
b = b
c
b 2
n
b
for every b B. By 326I (for the second time), there is a b
0
B such that b 0 for b B, b b
0
, while
b 0 when b B and b b
0
= 0. We have
(b
0
\ c) = b
0
(b
0
c) b
0

c
b
0
b
0
= b

+(b
0
\ b

) (b

\ b
0
) b

> 0,
so b
0
\ c ,= 0. (This is where I use the hypothesis that B.) By (c), there is a non-zero d b
0
\ c such
that

d
2
n

b
0
\c
2
n

b
0
.
Now d c = 0 so c d c. Also, for any b B,

dc
b =
d
b +
c
(b b
0
) +
c
(b \ b
0
)
2
n

b
0
\c
b +(b b
0
) 2
n
(b b
0
) (b b
0
) +(b \ b
0
)
2
n
(b b
0
\ c) +(b b
0
) 2
n
(b b
0
) +(b \ b
0
)
b.
But this means that d c C and c is not maximal in C. XXX
Thus c is the required element of A giving a representation of .
331C Corollary Let (A, ) be an atomless semi-nite measure algebra, and a A. Suppose that
0 a. Then there is a c A such that c a and c = .
proof If = a, take c = a. If < a, there is a d A such that d a and d < (322Eb). Apply
331B to the principal ideal A
d
generated by d, with B = 0, d and d = . (The point is that because A
is atomless, no non-trivial principal ideal of A
d
can be of the form c b : b B = 0, c.)
Remark Of course this is also an easy consequence of 215D.
331Fb Classication of homogeneous measure algebras 125
331D Lemma Let (A, ), (B, ) be totally nite measure algebras and C A a closed subalgebra.
Suppose that : C B is a measure-preserving Boolean homomorphism such that B is relatively atomless
over [C]. Take any a A, and let C
1
be the subalgebra of A generated by C a. Then there is a
measure-preserving homomorphism from C
1
to B extending .
proof We know that [C] is a closed subalgebra of B (324Kb), and that is a Boolean isomorphism
between C and [C]. Consequently the countably additive functional c (c a) : C R is transferred
to a countably additive functional : [C] R, writing (c) = (c a) for every c C. Of course
(c) c = (c) for every c C. So by 331B there is a b B such that (c) = (b c) for every c C.
If c C, c a then
(b c) = (c) = (a c) = c = (c),
so c b. Similarly, if a c C, then
(b c) = (a c) = (a 1) = (b 1) = b,
so b c. It follows from 312N that there is a Boolean homomorphism
1
: C
1
B, extending , such that

1
a = b.
To see that
1
is measure-preserving, take any member of C
1
. By 312M, this is expressible as e =
(c
1
a) (c
2
\ a), where c
1
, c
2
C. Now
(
1
e) = ((c
1
b) (c
2
\ b)) = (c
1
b) + (c
2
) (c
2
b)
= (c
1
a) + c
2
(c
2
a) = e.
As e is arbitrary,
1
is measure-preserving.
331E Generating sets For the sake of the next denition, we need a language a little more precise
than I have felt the need to use so far. The point is that if A is a Boolean algebra and B is a subset of A,
there is more than one subalgebra of A which can be said to be generated by B, because we can look at
any of the three algebras
B, the smallest subalgebra of A including B;
B

, the smallest -subalgebra of A including B;

B

, the smallest order-closed subalgebra of A including B.

(See 313Fb.) Now I will say henceforth, in this context, that
B is the subalgebra of A generated by B, and B generates A if A = B;
B

is the -subalgebra of A generated by B, and B -generates A if A = B

;
B

is the order-closed subalgebra of A generated by B, and B -generates or completely generates

A if A = B

.
There is a danger inherent in these phrases, because if we have B A
t
, where A
t
is a subalgebra of
A, it is possible that the smallest order-closed subalgebra of A
t
including B might not be recoverable from
the smallest order-closed subalgebra of A including B. (See 331Yb-331Yc.) This problem will not seriously
interfere with the ideas below; but for deniteness let me say that the phrases B -generates A, B -
generates A will always refer to suprema and inma taken in A itself, not in any larger algebra in which it
may be embedded.
331F Maharam types (a) With the language of 331E established, I can now dene the Maharam
type or complete generation (A) of any Boolean algebra A; it is the smallest cardinal of any subset of
A which -generates A.
(I think that this is the rst cardinal function which I have mentioned in this treatise. All you need
to know, to conrm that the denition is well-conceived, is that there is some set which -generates A;
and obviously A -generates itself. For this means that the set A = #(B) : B A -generates A is a
non-empty class of cardinals, and therefore, assuming the axiom of choice, has a least member (2A1Lf). In
331Ye-331Yf I mention a further function, the density of a topological space, which is closely related to
Maharam type.)
(b) A Boolean algebra A is Maharam-type-homogeneous if (A
a
) = (A) for every non-zero a A,
writing A
a
for the principal ideal of A generated by a.
126 Maharams theorem 331Fc
(c) Let (X, , ) be a measure space, with measure algebra (A, ). Then the Maharam type of (X, , ),
or of , is the Maharam type of A; and (X, , ), or , is Maharam-type-homogeneous if A is.
Remark I should perhaps remark that the phrases Maharam type and Maharam-type-homogeneous,
while well established in the context of probability algebras, are not in common use for general Boolean
algebras. But the cardinal (A) is important in the general context, and is such an obvious extension of
Maharams idea (Maharam 42) that I am happy to propose this extension of terminology.
331G For the sake of those who have not mixed set theory and algebra before, I had better spell out
some basic facts.
Proposition Let A be a Boolean algebra, B a subset of A. Let B be the subalgebra of A generated by B,
B

the -subalgebra of A generated by B, and B

the order-closed subalgebra of A generated by B.

(a) B B

.
(b) If B is nite, so is B, and in this case B = B

= B

.
(c) For every a B, there is a nite B
t
B such that a belongs to the subalgebra of A generated by B
t
.
Consequently #(B) max(, #(B)).
(d) For every a B

, there is a countable B
t
B such that a belongs to the -subalgebra of A generated
by B
t
.
(e) If A is ccc, then B

= B

.
proof (a) All we need to know is that B

is a subalgebra of A including B, and that B

is a -subalgebra
of A including B.
(b) Induce on #(B), using 312M for the inductive step, to see that B is nite. In this case it must be
order-closed, so is equal to B

.
(c)(i) For I B, let C
I
be the subalgebra of A generated by I. If I, J B then C
I
C
J
C
IJ
. So

C
I
: I B is nite is a subalgebra of A, and must be equal to B, as claimed.
(ii) To estimate the size of B, recall that the set [B]
<
of all nite subsets of B has cardinal at most
max(, #(B)) (3A1Cd). For each I [B]
<
, C
I
is nite, so
#(B) = #(

I[B]
<
C
I
) max(, #(I), sup
I[B]
< #(C
I
)) max(, #(B))
by 3A1Cc.
(d) For I B, let D
I
B

be the -subalgebra of A generated by I. If I, J B then D

I
D
J
D
IJ
,
so B
t

D
I
: I B is countable is a subalgebra of A. But also it is sequentially order-closed in A. PPP
Let a
n
)
nN
be a non-decreasing sequence in B
t

with supremum a in A. For each n N there is a countable

I(n) B such that a
n
C
I(n)
. Set K =

nN
I(n); then K is a countable subset of B and every a
n
belongs
to D
K
, so a D
K
B
t

. QQQ So B
t

is a -subalgebra of A including B and must be the whole of B

.
(e) By 316Fb, B

is order-closed in A, so must be equal to B

.
331H Proposition Let A be a Boolean algebra.
(a)(i) (A) = 0 i A is either 0 or 0, 1.
(ii) (A) is nite i A is nite.
(b) If B is another Boolean algebra and : A B is a surjective order-continuous Boolean homomor-
phism, then (B) (A).
(c) If a A then (A
a
) (A), where A
a
is the principal ideal of A generated by a.
(d) If A has an atom and is Maharam-type-homogeneous, then A = 0, 1.
proof (a)(i) (A) = 0 i A has no proper subalgebras. (ii) If A is nite, then (A) #(A) is nite. If
(A) is nite, then there is a nite set B A which -generates A; by 331Gb, A is nite.
(b) We know that there is a set A A, -generating A, with #(A) = (A). Now [A] -generates
[A] = B (313Mb), so
(B) #([A]) #(A) = (A).
331I Classication of homogeneous measure algebras 127
(c) Apply (b) to the map b a b : A A
a
.
(d) If a A is an atom, then (A
a
) = 0, so if A is Maharam-type-homogeneous then (A) = 0 and
A = 0, a = 0, 1.
331I We are now ready for the theorem.
Theorem Let (A, ) and (B, ) be Maharam-type-homogeneous measure algebras of the same Maharam
type, with 1 = 1 < . Then they are isomorphic as measure algebras.
proof (a) If (A) = (B) = 0, this is trivial. So let us take = (A) = (B) > 0. In this case, because
A and B are Maharam-type-homogeneous, they can have no atoms and must be innite, so is innite
(331H). Let a

)
<
, b

)
<
enumerate -generating subsets of A, B respectively.
The strategy of the proof is to dene a measure-preserving isomorphism : A B as the last of an
increasing family

of isomorphisms between closed subalgebras C

, D

of A and B. The inductive

hypothesis will be that, for some families a
t

)
<
, b
t

)
<
to be determined,
C

is the closed subalgebra of A generated by a

: < a
t

: < ,
D

is the closed subalgebra of B generated by b

: < b
t

: < ,

: C

is a measure-preserving isomorphism,

extends

whenever < .
(Formally speaking, this will be a transnite recursion, dening a function f() = (C

, D

, a
t

, b
t

)
on the ordinal + 1 by a rule which chooses f() in terms of f, as described in 2A1B. The construction
of an actual function F for which f() = F(f) will necessitate the axiom of choice.)
(b) The induction starts with C
0
= 0, 1, D
0
= 0, 1,
0
(0) = 0,
0
(1) = 1. (The hypothesis 1 = 1
is what we need to ensure that
0
is measure-preserving.)
(c) For the inductive step to a successor ordinal + 1, where < , suppose that C

, D

and

have
been dened.
(i) For any non-zero b B, the principal ideal B
b
of B generated by b has Maharam type , because
B is Maharam-type-homogeneous. On the other hand, the Maharam type of D

is at most
#(b

: b
t

: < ) #( 0, 1) < ,
because if is nite so is 0, 1, while if is innite then #( 0, 1) = #() < . Consequently B
b
cannot be an order-continuous image of D

(331Hb). Now the map c c b : D

B
b
is order-continuous,
because D

is closed, so that the embedding D

B is order-continuous. It therefore cannot be surjective,

and
b

a : a C

= b d : d D

,= B
b
.
This means that

: C

satises the conditions of 331D, and must have an extension

to
a measure-preserving homomorphism from the subalgebra C
t

of A generated by C

to B. We
know that C
t

is a closed subalgebra of A (314Ja), so it must be the closed subalgebra of A generated

by a

: a
t

: < . Also D
t

[C
t

] will be the subalgebra of B generated by D

b
t

, where
b
t

(a

), so is closed in B, and is the closed subalgebra of B generated by b

: < b
t

: .
(ii) The next step is to repeat the whole of the argument above, but applying it to
1

: D
t

, b

in place of

: C

and a

. Once again, we have (D

t

) < = (A
a
) for every a A, so we can use
Lemma 331D to nd a measure-preserving isomorphism

: D
+1
C
+1
extending
1

, where D
+1
is
the subalgebra of B generated by D
t

, and C
+1
is the subalgebra of A generated by C
t

a
t

, setting
a
t

(b

). As in (i), we nd that C
+1
is the closed subalgebra of A generated by a

: a
t

: ,
while D
+1
is the closed subalgebra of B generated by b

: b
t

: .
(iii) We can therefore take
+1
=
1

: C
+1
D
+1
, and see that
+1
is a measure-preserving
isomorphism, extending

, such that
+1
(a

) = b
t

,
+1
(a
t

) = b

. Evidently
+1
extends

for every
because it extends

and (by the inductive hypothesis)

extends

for every < .

128 Maharams theorem 331I
(d) For the inductive step to a limit ordinal , where 0 < , suppose that C

, D

, a
t

, b
t

have
been dened for < . Set C

=

<
C

. Then C

is a subalgebra of A, because it is the union of an

upwards-directed family of subalgebras; similarly, D

=

<
D

is a subalgebra of B. Next, we have a

function

: C

dened by setting

a =

a whenever < and a C

; for if , < and

a C

, then

a =
max(,)
a =

a. Clearly

[C

] =

<

[C

] = D

.
Moreover,

a = a for every a C

, since

a = a whenever < and a C

.
Now let C

be the smallest closed subalgebra of A including C

, that is, the metric closure of C

in A
(323J). Since C

is the smallest closed subalgebra of A including C

for every < , it must be the closed

subalgebra of A generated by a

: < a
t

: < . By 324O,

has an extension to a measure-

preserving homomorphism

: C

B. Set D

[C

]; by 324Kb, D

is a closed subalgebra of B.
Because

: C

B is continuous (also noted in 324Kb),

D

[C

] =

[C

]
is topologically dense in D

(3A3Eb), and D

= D

is the closed subalgebra of B -generated by b

: <
b
t

: < . Finally, if < ,

extends

because

extends

. Thus the induction continues.

(e) The induction ends with = , C

= A, D

= B and =

: A B the required measure algebra

isomorphism.
331J Lemma Let be any innite cardinal. Let be the usual measure on 0, 1

(254J) and (A, )

its measure algebra. Then if (B, ) is any non-zero totally nite measure algebra and : A B is an
order-continuous Boolean homomorphism, (B) .
proof Set X = 0, 1

and write for the domain of .

(a) Set E

= x : x X, x() = 1, a

= E

for each < . If

n
)
nN
is any sequence of distinct
elements of ,
(

nN
E

n
) = lim
n
(

in
E

n
) = lim
n
2
n1
= 0,
so that (inf
nN
a

n
) = 0 and inf
nN
a

n
= 0. Because is order-continuous, inf
nN
(a

n
) = 0 in B.
Similarly, (

nN
E

n
) = 1 and sup
nN
(a

n
) = 1.
(b) For b B, > 0 set U(b, ) = b
t
: (b
t
.b) < , the ordinary open -neighbourhood of b. If b B,
then there is a > 0 such that : < , a

U(b, ) is nite. PPP??? Suppose, if possible, otherwise. Then

there is a sequence
n
)
nN
of distinct elements of such that (b .a

n
) 2
n2
1 for every n N. Now
inf
nN
a

n
= 0, so
b = (b \ inf
nN
a

n
)

n=0
(b \ a

n
).
Similarly
(1 \ b) = (sup
nN
a

n
\ b)

n=0
(a

n
\ b).
So
1 = b + (1 \ b)

n=0
(b \ a

n
) +

n=0
(a

n
\ b)
=

n=0
(b .a

n
)

n=0
2
n2
1 < 1,
which is impossible. XXXQQQ
(c) Note that B is innite; for if b B the set : a

= b must be nite, and is supposed to be

innite. So (B) must be innite.
(d) Now take a set B B, of cardinal (B), which -generates B. By (c), B is innite. Let C be the
subalgebra of B generated by B; then #(C) = #(B) = (B), by 331Gc, and C is topologically dense in B.
331N Classication of homogeneous measure algebras 129
If b B, there are c C, k N such that b U(c, 2
k
) and : a

U(c, 2
k
) is nite. PPP By (b), there
is a > 0 such that : a

U(b, ) is nite. Take k N such that 2 2

k
, and c C U(b, 2
k
);
then U(c, 2
k
) U(b, ) can contain only nitely many a

, so these c, k serve. QQQ

Consider
| = U(c, 2
k
) : c C, k N, : a

U(c, 2
k
) is nite.
Then #(|) max(#(C), ) = (B). Also | is a cover of B. In particular, =

U/
J
U
, where J
U
= :
a

U. But this means that

= #() max(, #(|), sup
U/
#(J
U
)) = (B),
as claimed.
331K Theorem Let be any innite cardinal. Let be the usual measure on 0, 1

and (A, ) its

measure algebra. Then A is Maharam-type-homogeneous, of Maharam type .
proof Set X = 0, 1

and write for the domain of .

(a) To see that (A) , set E

= x : x X, x() = 1, a

= E

for each < . Writing c for the

algebra of subsets of X generated by E

: < , we see that every measurable cylinder in X, as dened

in 254A, belongs to c, so that every member of is approximated, in measure, by members of c (254Fe),
that is, E

: E c is topologically dense in A. But this means just that the subalgebra of A generated
by a

: < is topologically dense in A, so that a

: < -generates A, and (A) .

(b) Next, if c A 0 and A
c
is the principal ideal of A generated by c, the map a a c is an
order-continuous Boolean homomorphism from A to A
c
, so by 331J we must have (A
c
) . Thus
(A
c
) (A) .
As c is arbitrary, A is Maharam-type-homogeneous with Maharam type .
331L Theorem Let (A, ) be a Maharam-type-homogeneous probability algebra. Then there is exactly
one , either 0 or an innite cardinal, such that (A, ) is isomorphic, as measure algebra, to the measure
algebra (A

) of the usual measure on 0, 1

.
proof If (A) is nite, it is zero, and A = 0, 1 (331Ha, 331He) so that (interpreting 0, 1
0
as ) we
have the case = 0. If = (A) is innite, then by 331K we know that (A

) is also Maharam-type-
homogeneous of Maharam type , so 331I gives the required isomorphism. Of course is uniquely dened
by A.
331M Homogeneous Boolean algebras Having introduced the word homogeneous, I think I ought
not to leave you without mentioning its standard meaning in the context of Boolean algebras, which is
connected with one of the most striking and signicant consequences of Theorem 331I.
Denition A Boolean algebra A is homogeneous if A is isomorphic, as Boolean algebra, to every non-
trivial principal ideal of A.
Remark Of course a homogeneous Boolean algebra must be Maharam-type-homogeneous, since (A) =
(A
c
) whenever A is isomorphic to A
c
. In general, a Boolean algebra can be Maharam-type-homogeneous
without being homogeneous (331Xj, 331Yj). But for -nite measure algebras this doesnt happen.
331N Proposition Let (A, ) be a Maharam-type-homogeneous -nite measure algebra. Then it is
homogeneous as a Boolean algebra.
proof If A = 0 this is trivial; so suppose that A ,= 0. By 322G, there is a measure on A such that
(A, ) is a probability algebra. Now let c be any non-zero member of A, and set = c,
t
c
=
1

c
, where
c
is the restriction of to the principal ideal A
c
of A generated by c. Then (A, ) and (A
c
,
t
c
) are Maharam-
type-homogeneous probability algebras of the same Maharam type, so are isomorphic as measure algebras,
and a fortiori as Boolean algebras.
130 Maharams theorem 331X
331X Basic exercises (a) Let (X, , ) be a probability space, T a -subalgebra of such that for any
non-negligible E there in an F such that F E and (F(EH)) > 0 for every H T. Suppose
that f : X [0, 1] is a measurable function. Show that there is an F such that
_
H
f = (H F) for
every H T.
>>>(b) Write out a direct proof of 331C not relying on 331B.
(c) Let A be a nite Boolean algebra with n atoms. Show that (A) is the least k such that n 2
k
.
>>>(d) Show that the measure algebra of Lebesgue measure on R is Maharam-type-homogeneous and of
Maharam type . (Hint: show that it is -generated by ], q]

: q Q.)
(e) Show that the measure algebra of Lebesgue measure on R
r
is Maharam-type-homogeneous, of Ma-
haram type , for any r 1. (Hint: show that it is -generated by ], q]

: q Q
r
.)
(f ) Show that the measure algebra of any Radon measure on R
r
(256A) has countable Maharam type.
(Hint: show that it is -generated by ], q]

: q Q
r
.)
>>>(g) Show that TR has Maharam type . (Hint: show that it is -generated by ], q] : q Q.)
>>>(h) Show that the regular open algebra of R is Maharam-type-homogeneous, of Maharam type . (Hint:
show that it is -generated by ], q]

: q Q.)
(i) Let (A, ) be a totally nite measure algebra, and an innite cardinal. Suppose that there is a family
a

)
<
in A such that inf
I
a

= 0, sup
I
a

= 1 for every innite I . Show that (A

a
) for every
non-zero principal ideal A
a
of A.
(j) Let A be the measure algebra of Lebesgue measure on R, and G the regular open algebra of R.
Show that the simple product A G is Maharam-type-homogeneous, with Maharam type , but is not
homogeneous. (Hint: A is weakly (, )-distributive, but G is not, so they are not isomorphic.)
331Y Further exercises (a) Suppose that A is a Dedekind complete Boolean algebra, B is an order-
closed subalgebra of A and C is an order-closed subalgebra of B. Show that if a A is a relative atom in
A over C, then upr(a, B) is a relative atom in B over C. So if B is relatively atomless over C, then A is
relatively atomless over C.
(b) Give an example of a Boolean algebra A with a subalgebra A
t
and a proper subalgebra B of A
t
which
is order-closed in A
t
, but -generates A. (Hint: take A to be the measure algebra A
L
of Lebesgue measure
on R and B the subalgebra B
Q
of A generated by [a, b]

: a, b Q. Take E R such that I E, I E have

non-zero measure for every non-trivial interval I R (134Jb), and let A
t
be the subalgebra of A generated
by B E

.)
(c) Give an example of a Boolean algebra A with a subalgebra A
t
and a proper subalgebra B of A
t
which is order-closed in A, but -generates A
t
. (Hint: in the notation of 331Yb, take Z to be the Stone
space of A
L
, and set A
t
= a : a A
L
, B = a : a B
Q
; let A be the subalgebra of TZ generated by
A
t
z : z Z.)
(d) Let A be a Dedekind complete purely atomic Boolean algebra, and A the set of its atoms. Show that
(A) is the least cardinal such that #(A) 2

.
(e) Let (A, ) be a measure algebra. Write d(A) for the smallest cardinal of any subset of A which is
dense for the measure-algebra topology. Show that d(A) max(, (A)). Show that if (A, ) is localizable,
then (A) d(A).
(f ) Let (X, ) be a metric space. Write d(X) for the density of X, the smallest cardinal of any dense
subset of X. (i) Show that if ( is any family of open subsets of X, there is a family H ( such that

H =

( and #(H) max(, d(X)). (ii) Show that if > max(, d(X)) and x

)
<
is any family in X,
then there is an x X such that #( : x

G) > max(, d(X)) for every open set G containing x, and

that there is a sequence
n
)
nN
of distinct members of such that x = lim
n
x

n
.
331 Notes Classication of homogeneous measure algebras 131
(g) Show that the algebra of open-and-closed subsets of 0, 1
N
is homogeneous.
(h) Show that if A is a homogeneous Boolean algebra so is its Dedekind completion

A (314U).
(i) Show that the regular open algebra of R is a homogeneous Boolean algebra.
(j) Let (A, ) be the simple product (322K) of c copies of the measure algebra of the usual measure on
0, 1
c
. Show that A is Maharam-type-homogeneous but not homogeneous.
331 Notes and comments Maharams theorem belongs with the Radon-Nikod ym theorem, Fubinis the-
orem and the strong law of large numbers as one of the theorems which make measure theory what it is.
Once you have this theorem and its consequences in the next section properly absorbed, you will never again
look at a measure space without classifying its measure algebra in terms of the types of its homogeneous
principal ideals. As one might expect, a very large proportion of the important measure spaces of analysis
are homogeneous, and indeed a great many are homogeneous with Maharam type .
In this section I have contented myself with the basic statement of Theorem 331I on the isomorphism of
Maharam-type-homogeneous measure algebras and the identication of representative homogeneous prob-
ability algebras (331K). The same techniques lead to an enormous number of further facts, some of which
I will describe in the rest of the chapter. For the moment, it gives us a complete description of Maharam-
type-homogeneous probability algebras (331L). There is the atomic algebra 0, 1, with Maharam type 0,
and for each innite cardinal there is the measure algebra of 0, 1

, with Maharam type ; these are all

non-isomorphic, and every Maharam-type-homogeneous probability algebra is isomorphic to exactly one of
them. The isomorphisms here are not unique; indeed, it is characteristic of measure algebras that they have
very large automorphism groups (see Chapter 38 below), and there are correspondingly large numbers of
isomorphisms between any isomorphic pair. The proof of 331I already suggests this, since we have such a
vast amount of choice concerning the lists a

)
<
, b

)
<
, and even with these xed there remains a good
deal of scope in the choice of a
t

)
<
and b
t

)
<
.
The isomorphisms described in Theorem 331I are measure algebra isomorphisms, that is, measure-
preserving Boolean isomorphisms. Obvious questions arise concerning Boolean isomorphisms which are
not necessarily measure-preserving; the theorem also helps us to settle many of these (see 331M-331N). But
we can observe straight away the remarkable fact that two homogeneous probability algebras which are
isomorphic as Boolean algebras are also isomorphic as probability algebras, since they must have the same
Maharam type.
I have already mentioned certain measure space isomorphisms (254K, 255A). Of course any isomorphism
between measure spaces must induce an isomorphism between their measure algebras (see 324M), and any
isomorphism between measure algebras corresponds to an isomorphism between their Stone spaces (see
324N). But there are many important examples of isomorphisms between measure algebras which do not
correspond to isomorphisms between the measure spaces most naturally involved. (I describe one in 343J.)
Maharams theorem really is a theorem about measure algebras rather than measure spaces.
The particular method I use to show that the measure algebra of the usual measure on 0, 1

is homo-
geneous for innite (331J-331K) is chosen with a view to a question in the next section (332O). There are
other ways of doing it. But I recommend study of this particular one because of the way in which it involves
the topological, algebraic and order properties of the algebra B. I have extracted some of the elements of
the argument in 331Xi and 331Ye-331Yf. These use the concept of density of a topological space. This
does not seem the moment to go farther along this road, but I hope you can see that there are likely to be
many further cardinal functions to provide useful measures of complexity in both algebraic and topological
structures.
132 Maharams theorem 332 intro.
332 Classication of localizable measure algebras
In this section I present what I call Maharams theorem, that every localizable measure algebra is
expressible as a weighted simple product of measure algebras of spaces of the form 0, 1

(332B). Among
its many consequences is a complete description of the isomorphism classes of localizable measure algebras
(332J). This description needs the concepts of cellularity of a Boolean algebra (332D) and its renement,
the magnitude of a measure algebra (332G). I end this section with a discussion of those pairs of measure
algebras for which there is a measure-preserving homomorphism from one to the other (332P-332Q), and a
general formula for the Maharam type of a localizable measure algebra (332S).
332A Lemma Let A be any Boolean algebra. Writing A
a
for the principal ideal generated by a A,
the set a : a A, A
a
is Maharam-type-homogeneous is order-dense in A.
proof Take any a A 0. Then A = (A
b
) : 0 ,= b a has a least member; take c a such that c ,= 0
and (A
c
) = min A. If 0 ,= b c, then (A
b
) (A
c
), by 331Hc, while (A
b
) A, so (A
c
) (A
b
). Thus
(A
b
) = (A
c
) for every non-zero b c, and A
c
is Maharam-type-homogeneous.
332B Maharams Theorem Let (A, ) be any localizable measure algebra. Then it is isomorphic to
the simple product of a family (A
i
,
i
))
iI
of measure algebras, where for each i I (A
i
,
i
) is isomorphic,
up to a re-normalization of the measure, to the measure algebra of the usual measure on 0, 1

i
, where
i
is either 0 or an innite cardinal.
proof (a) For a A, let A
a
be the principal ideal of A generated by a. Then
D = a : a A, 0 < a < , A
a
is Maharam-type-homogeneous
is order-dense in A. PPP If a A 0, then (because (A, ) is semi-nite) there is a b a such that
0 < b < ; now by 332A there is a non-zero d b such that A
d
is Maharam-type-homogeneous. By 331N,
A
d
is Maharam-type-homogeneous, and d D. QQQ
(b) By 313K, there is a partition of unity e
i
)
iI
consisting of members of D; by 322Kd, (A, ) is
isomorphic, as measure algebra, to the simple product of the principal ideals A
i
= A
e
i
.
(c) For each i I, (A
i
,
i
) is a non-trivial totally nite Maharam-type-homogeneous measure algebra,
writing
i
= A
i
. Take
i
=
i
(1
A
i
) = e
i
, and set
t
i
=
1
i

i
. Then (A
i
,
t
i
) is a Maharam-type-
homogeneous probability algebra, so by 331L is isomorphic to the measure algebra (B

i
,

i
) of the usual
measure on 0, 1

i
, where
i
is either 0 or an innite cardinal. Thus (A
i
,
i
) is isomorphic, up to a scalar
multiple of the measure, to (B

i
,

i
).
Remark For the case of totally nite measure algebras, this is Theorem 2 of Maharam 42.
332C Corollary Let (A, ) be a localizable measure algebra. For any cardinal , write

for the usual

measure on 0, 1

, and

for its domain. Then we can nd families

i
)
iI
,
i
)
iI
such that every
i
is
either 0 or an innite cardinal, every
i
is a strictly positive real number, and (A, ) is isomorphic to the
measure algebra of (X, , ), where
X = (x, i) : i I, x 0, 1

i
,
= E : E X, x : (x, i) E

i
for every i I,
E =

iI

i

i
x : (x, i) E
for every E .
proof Take the family
i
)
iI
from the last theorem, take the
i
= e
i
to be the normalizing factors of the
proof there, and apply 322Kb to identify the simple product of the measure algebras of (0, 1

i
,

i
,
i

i
)
with the measure algebra of their direct sum (X, , ).
332H Classication of localizable measure algebras 133
332D The cellularity of a Boolean algebra In order to properly describe non-sigma-nite measure
algebras, we need the following concept. If A is any Boolean algebra, write
c(A) = sup#(C) : C A 0 is disjoint,
the cellularity of A. (If A = 0, take c(A) = 0.) Thus A is ccc (316A) i c(A) .
332E Proposition Let (A, ) be any semi-nite measure algebra, and C any partition of unity in A
consisting of elements of nite measure. Then max(, #(C)) = max(, c(A)).
proof Of course #(C 0) c(A), because C 0 is disjoint, so
max(, #(C)) = max(, #(C 0) max(, c(A)).
Now suppose that D is any disjoint set in A 0. For c C, d c : d D is a disjoint set in
the principal ideal A
c
generated by c. But A
c
is ccc (322G), so d c : d D must be countable, and
D
c
= d : d D, d c ,= 0 is countable. Because sup C = 1, D =

cC
D
c
, so
#(D) max(, #(C), sup
cC
#(D
c
)) = max(, #(C)).
As D is arbitrary, c(A) max(, #(C)) and max(, c(A)) = max(, #(C)).
332F Corollary Let (A, ) be any semi-nite measure algebra. Then there is a disjoint set in A 0
of cardinal c(A).
proof Start by taking any partition of unity C consisting of non-zero elements of nite measure. If
#(C) = c(A) we can stop, because C is a disjoint set in A 0. Otherwise, by 332E, we must have C
nite and c(A) = . Let A be the set of atoms in A. If A is innite, it is a disjoint set of cardinal , so
we can stop. Otherwise, since there is certainly a disjoint set D A 0 of cardinal greater than #(A),
and since each member of A can meet at most one member of D, there must be a member d of D which
does not include any atom. Accordingly we can choose inductively a sequence d
n
)
nN
such that d
0
= d,
0 ,= d
n+1
d
n
for every n. Now d
n
\ d
n+1
: n N is a disjoint set in A 0 of cardinal = c(A).
332G Denitions For the next theorem, it will be convenient to have some special terminology.
(a) The rst word I wish to introduce is a variant of the idea of cellularity, adapted to measure algebras.
If (A, ) is a semi-nite measure algebra, let us say that the magnitude of an a A is a if a is nite,
and otherwise is the cellularity of the principal ideal A
a
generated by a. (This is necessarily innite, since
any partition of a into sets of nite measure must be innite.) If we take it that any real number is less
than any innite cardinal, then the class of possible magnitudes is totally ordered.
I shall sometimes speak of the magnitude of the measure algebra (A, ) itself, meaning the magnitude
of 1
A
. Similarly, if (X, , ) is a semi-nite measure space, the magnitude of (X, , ), or of , is the
magnitude of its measure algebra.
(b) Next, for any Dedekind complete Boolean algebra A, and any cardinal , we can look at the element
e

= supa : a A 0, A
a
is Maharam-type-homogeneous of Maharam type ,
writing A
a
for the principal ideal of A generated by a, as usual. I will call this the Maharam-type-
component of A. Of course e

= 0 whenever , are distinct cardinals. PPP a b = 0 whenever A

a
, A
b
are Maharam-type-homogeneous of dierent Maharam types, since (A
ab
) cannot be equal simultaneously
to (A
a
) and (A
b
). QQQ
Also e

: is a cardinal is a partition of unity in A, because

supe

: is a cardinal = supa : A
a
is Maharam-type-homogeneous = 1
by 332A. Note that there is no claim that A
e

itself is homogeneous; but we do have a useful result in this

direction.
332H Lemma Let A be a Dedekind complete Boolean algebra and an innite cardinal. Let e be the
Maharam-type- component of A. If 0 ,= d e and the principal ideal A
d
generated by d is ccc, then it is
Maharam-type-homogeneous with Maharam type .
134 Maharams theorem 332H
proof (a) The point is that (A
d
) . PPP Set
A = a : a A 0, A
a
is Maharam-type-homogeneous of Maharam type .
Then d = supa d : a A. Because A
d
is ccc, there is a sequence a
n
)
nN
in A such that d = sup
nN
d a
n
(316E); set b
n
= d a
n
. We have (A
b
n
) (A
a
n
) = for each n; let D
n
be a subset of A
b
n
, of cardinal at
most , which -generates A
b
n
. Set
D =

nN
D
n
b
n
: n N A
d
.
If C is the order-closed subalgebra of A
d
generated by D, then C A
b
n
is an order-closed subalgebra of A
b
n
including D
n
, so is equal to A
b
n
, for every n. But a = sup
nN
a b
n
for every a A
d
, so C = A
d
. Thus
D -generates A
d
, and
(A
d
) #(D) max(, sup
nN
#(D
n
)) = . QQQ
(b) If now b is any non-zero member of A
d
, there is some a A such that b a ,= 0, so that
= (A
ba
) (A
b
) (A
d
) .
Thus we must have (A
b
) = for every non-zero b A
d
, and A
d
is Maharam-type-homogeneous of type ,
as claimed.
332I Lemma Let (A, ) be an atomless semi-nite measure algebra which is not totally nite. Then it
has a partition of unity consisting of elements of measure 1.
proof Let A be the set a : a = 1, and ( the family of disjoint subsets of A. By Zorns lemma, ( has a
maximal member C
0
(compare the proof of 313K). Set D = d : d A, d c = 0 for every c C
0
. Then
D is upwards-directed. If d D, then a ,= 1 for every a d, so d < 1, by 331C. So d
0
= sup D is dened
in A (321C); of course d
0
D, so d
0
< 1. Observe that supC
0
= 1 \ d
0
.
Because 1 = , C
0
must be innite; let a
n
)
nN
be any sequence of distinct elements of C
0
. For each
n N, use 331C again to choose an a
t
n
a
n
such that a
t
n
= d
0
. Set
b
0
= d
0
(a
0
\ a
t
0
), b
n
= a
t
n1
(a
n
\ a
t
n
)
for every n 1. Then b
n
)
nN
is a disjoint sequence of elements of measure 1 and sup
nN
b
n
= sup
nN
a
n
d
0
.
Now
(C
0
a
n
: n N) b
n
: n N
is a partition of unity consisting of elements of measure 1.
332J Now I can formulate a complete classication theorem for localizable measure algebras, rening
the expression in 332B.
Theorem Let (A, ) and (B, ) be localizable measure algebras. For each cardinal , let e

, f

be the
Maharam-type- components of A, B respectively. Then (A, ) and (B, ) are isomorphic, as measure
algebras, i (i) e

and f

have the same magnitude for every innite cardinal (ii) for every ]0, [,
(A, ) and (B, ) have the same number of atoms of measure .
proof Throughout the proof, write A
a
for the principal ideal of A generated by a, and
a
for the restriction
of to A
a
; and dene B
b
,
b
similarly for b B.
(a) If (A, ) and (B, ) are isomorphic, then of course the isomorphism matches their Maharam-type
components together and retains their magnitudes, and matches atoms of the same measure together; so
the conditions are surely satised.
(b) Now suppose that the conditions are satised. Set
K = : is an innite cardinal, e

,= 0 = : is an innite cardinal, f

,= 0.
For ]0, [, let A

be the set of atoms of measure in A, and set e

= sup A

. Write I = K ]0, [.
Then e
i
)
iI
is a partition of unity in A, so (A, ) is isomorphic to the simple product of (A
e
i
,
e
i
))
iI
,
writing A
e
i
for the principal ideal generated by e
i
and
e
i
for the restriction A
e
i
.
332L Classication of localizable measure algebras 135
In the same way, writing B

for the set of atoms of measure in B, f

for sup B

, B
f
i
for the principal
ideal generated by f
i
and
f
i
for the restriction of fo B
f
i
, we have (B, ) isomorphic to the simple product
of (B
f
i
,
f
i
))
iI
.
(c) It will therefore be enough if I can show that (A
e
i
,
e
i
)

= (B
f
i
,
f
i
) for every i I.
(i) For K, the hypothesis is that e

and f

have the same magnitude. If they are both of nite

magnitude, that is, e

= f

< , then both (A

e

,
e

) and (B
f

,
f

) are homogeneous and of Maharam

type , by 332H. So 331I tells us that they are isomorphic. If they are both of innite magnitude , then
332I tells us that both A
e

, B
f

have partitions of unity C, D consisting of sets of measure 1. So (A

e

,
e

)
is isomorphic to the simple product of (A
c
,
c
))
cC
, while (B
f

,
f

) is isomorphic to the simple product

of (B
d
,
d
))
dD
. But we know also that every (A
c
,
c
), (B
d
,
d
) is a homogeneous probability algebra of
Maharam type , by 332H again, so by Maharams theorem again they are all isomorphic. Since C, D and
are all innite,
#(C) = c(A
e

) = = c(B
f

) = #(D)
by 332E. So we are taking the same number of factors in each product and (A
e

,
e

) must be isomorphic
to (B
f

,
f

).
(ii) For ]0, [, our hypothesis is that #(A

) = #(B

). Now A

is a partition of unity in A
e

,
so (A
e

,
e

) is isomorphic to the simple product of (A

a
,
a
))
aA

. Similarly, (B
f

,
f

) is isomorphic to
the simple product of (B
b
,
b
))
bB

. Since every (A
a
,
a
), (B
b
,
b
) is just a simple atom of measure ,
these are all isomorphic; since we are taking the same number of factors in each product, (A
e

,
e

) must
be isomorphic to (B
f

,
f

).
(iii) Thus we have the full set of required isomorphisms, and (A, ) is isomorphic to (B, ).
332K Remarks (a) The partition of unity e
i
: i I of A used in the above theorem is in some sense
canonical. (You might feel it more economical to replace I by K : A

,= .) The further partition of

the atomic part into individual atoms (part (c-ii) of the proof) is also canonical. But of course the partition
of the e

of innite magnitude into elements of measure 1 requires a degree of arbitrary choice.

The value of the expressions in 332C is that the parameters
i
,
i
there are sucient to identify the
measure algebra up to isomorphism. For, amalgamating the language of 332C and 332J, we see that the
magnitude of e

in 332J is just

i
=

i
if this is nite, #(i :
i
= ) otherwise (using 332E, as usual);
while the number of atoms of measure is #(i :
i
= 0,
i
= ).
(b) The classication which Maharams theorem gives us is not merely a listing. It involves a real insight
into the nature of the algebras, enabling us to answer a very wide variety of natural questions. I give the
next couple of results as a sample of what we can expect these methods to do for us.
332L Proposition Let (A, ) be a measure algebra, and a, b A two elements of nite measure.
Suppose that : A
a
A
b
is a measure-preserving isomorphism, where A
a
, A
b
are the principal ideals
generated by a and b. Then there is a measure-preserving automorphism : A A which extends .
proof The point is that A
b\a
is isomorphic, as measure algebra, to A
a\b
. PPP Set c = a b. For each innite
cardinal , let e

be the Maharam-type- component of A

c
. Then e

a is the Maharam-type- component of

A
a
, because if d c and A
d
is Maharam homogeneous with Maharam type , then A
da
is either 0 or again
Maharam-type-homogeneous with Maharam type . Similarly, e

\ a is the Maharam-type- component of

A
c\a
= A
b\a
, e

b is the Maharam-type- component of A

b
and e

\ b is the Maharam-type- component

of A
a\b
. Now : A
a
A
b
is an isomorphism, so (e

a) must be e

b, and
(e

a) = e

(e

a) = e

(e

a)
= e

(e

b) = (e

b).
In the same way, if we write n

(d) for the number of atoms of measure in A

d
, then
n

(b \ a) = n

(c) n

(a) = n

(c) n

(b) = n

(a \ b)
136 Maharams theorem 332L
for every ]0, [. By 332J, there is a measure-preserving isomorphism
1
: A
b\a
A
a\b
. QQQ
If we now set
d = (d a)
1
(d b \ a) (d \ c)
for every d A, : A A is a measure-preserving isomorphism which agrees with on A
a
.
332M Lemma Suppose that (A, ) and (B, ) are homogeneous measure algebras, with (A) (B)
and 1 = 1 < . Then there is a measure-preserving Boolean homomorphism from A to B.
proof The case (A) = 0 is trivial. Otherwise, considering normalized versions of the measures, we are
reduced to the case 1 = 1 = 1, (A) = , (B) = , so that (A, ) is isomorphic to the
measure algebra (A

) of the usual measure

on 0, 1

; and similarly (B, ) is isomorphic to the

measure algebra of the usual measure on 0, 1

. Now (identifying the cardinals , with von Neumann

ordinals, as usual), , so we have an inverse-measure-preserving map x x : 0, 1

0, 1

(254Oa), which induces a measure-preserving Boolean homomorphism from A

to A

(324M), and hence a

measure-preserving homomorphism from A to B.
332N Lemma If (A, ) is a probability algebra and max(, (A)), then there is a measure-preserving
homomorphism from (A, ) to the measure algebra (B, ) of the usual measure on 0, 1

; that is, (A, )

is isomorphic to a closed subalgebra of (B, ).
proof Let c
i
)
iI
be a partition of unity in A such that every principal ideal A
c
i
is homogeneous and no
c
i
is zero. Then I is countable and

iI
c
i
= 1. Accordingly there is a partition of unity d
i
)
iI
in B

such that d
i
= c
i
for every i. PPP Because I is countable, we may suppose that it is either N or an initial
segment of N. In this case, choose d
i
)
iI
inductively such that d
i
1 \ sup
j<i
d
j
and d
i
= d
i
for each
i I, using 331C. QQQ
If i I, then (A
c
i
) = (B
d
i
), so there is a measure-preserving Boolean homomorphism
i
:
A
c
i
B
d
i
. Setting a = sup
iI

i
(a c
i
) for a A, we have a measure-preserving Boolean homomorphism
: A B. By 324Kb, [A] is a closed subalgebra of B, and of course ([A], [A]) is isomorphic to (A, ).
332O Lemma Let (A, ), (B, ) be localizable measure algebras. For each innite cardinal let e

, f

be their Maharam-type- components, and for ]0, [ let e

, f

be the suprema of the atoms of measure

in A, B respectively. If there is a measure-preserving Boolean homomorphism from A to B, then the
magnitude of sup

is not greater than the magnitude of sup

whenever is an innite cardinal,

while the magnitude of sup

sup

is not greater than the magnitude of sup

sup

for any ]0, [.

proof Suppose that : A B is a measure-preserving Boolean homomorphism. For innite car-
dinals , set e

= sup

, f

= sup

, while for ]0, [ set e

= sup

sup

,
f

= sup

sup

. Let c
i
)
iI
be a partition of unity in A such that all the principal ideals
A
c
i
are totally nite and homogeneous, as in 332B. Then c
i
e

whenever = (A
c
i
) is innite, and
c
i
e

if c
i
is an atom of measure . Take v to be either an innite cardinal or a strictly positive real
number. Set
J = i : i I, c
i
e

v
;
then e

v
= sup
iJ
c
i
.
Now the point is that if i J then c
i
f

v
. PPP We need to consider two cases. (i) If c
i
is an atom,
then v ]0, [ and c
i
v. So we need only observe that 1 \ f

v
is just the supremum in B of the atoms
of measure greater than v, none of which can meet c
i
, since this has measure at most v. (ii) Now suppose
that A
c
i
is atomless, with (A
c
i
) = v. If 0 ,= b c
i
, then a b a : A
c
i
B
b
is an order-continuous
Boolean homomorphism, while A
c
i
is isomorphic (as Boolean algebra) to the measure algebra of 0, 1

,
so 331J tells us that (B
b
) . This means, rst, that b cannot be an atom, so that c
i
cannot meet
sup
]0,[
f

; and also that b cannot be included in f

for any innite

t
< , so that c
i
cannot meet
sup

<
f

. Thus c
i
must be included in sup

= f

v
. QQQ
Of course c
i
)
iJ
is disjoint. So if e

v
has nite magnitude, the magnitude of f

v
is at least
332P Classication of localizable measure algebras 137

iJ
c
i
=

iJ
c
i
= e

v
,
the magnitude of e

v
. While if e

v
has innite magnitude, this is #(J), by 332E, which is not greater than
the magnitude of f

v
.
332P Proposition Let (A, ), (B, ) be atomless totally nite measure algebras. For each innite
cardinal let e

, f

be their Maharam-type- components. Then the following are equiveridical:

(i) (A, ) is isomorphic to a closed subalgebra of a principal ideal of (B, );
(ii) for every cardinal ,
(sup

) (sup

).
proof (a)(i)(ii) Suppose that : A B
d
is a measure-preserving isomorphism between A and a closed
subalgebra of a principal ideal B
d
of B. The Maharam-type- component of B
d
is just d f

, so 332O tells
us that
(sup

) (sup

d f

) (sup

)
for every .
(b)(ii)(i) Now suppose that the condition is satised.
() Let P be the set of all measure-preserving Boolean homomorphisms from principal ideals A
c

of
A to principal ideals B
d

of B such that
(sup

\ c

) (sup

\ d

)
for every cardinal . Then the trivial homomorphism from A
0
to B
0
belongs to P, so P is not empty.
Order P by saying that
t
if
t
extends , that is, if c

and
t
a = a for every a A
c

. Then P
is a partially ordered set.
() If Q P is non-empty and totally ordered, it is bounded above in P. PPP Set c

= sup
Q
c

,
d

= sup
Q
d

. For a c

set

a = sup
Q
(a c

). Because Q is totally ordered,

extends all the

functions in Q. It is also easy to check that

0 = 0,

(a a
t
) =

a
t
and

(a a
t
) =

a
t
for
all a, a
t
A
c
,

= d

and that

a = a for every a A
c
; so that

is a measure-preserving Boolean
homomorphism from A
c
to B
d
.
Now suppose that is any cardinal; then
(sup

\ c

) = inf
Q
(sup

\ c

) inf
Q
(sup

\ d

) = (sup

\ d

).
So

P and is the required upper bound of Q. QQQ

() By Zorns Lemma, P has a maximal element say. Now c

= 1. PPP??? If not, then let
0
be the
least cardinal such that e

0
\ c

,= 0. Then
0 < (sup

0
e

\ c

) (sup

0
f

\ d

),
so there is a least
1

0
such that f

1
\ d

,= 0. Set = min( (e

0
\ c

), (f

1
\ d

)) > 0. Because A and B
are atomless, there are a e

0
\ c

, b f

1
\ d

such that a = b = (331C). Now A
a
is homogeneous with
Maharam type
0
, while B
b
is homogeneous with Maharam type
1
(332H), so there is a measure-preserving
Boolean homomorphism : A
a
B
b
(332M). Set
c

= c

a, d

= d

b,
and dene

: A
c
B
d
by setting

(g) = (g c

) (g a) for every g c

. It is easy to check that

is a measure-preserving Boolean homomorphism.

If is a cardinal and
0
,
(sup

\ c

) = (sup

\ c

) (sup

\ d

) = (sup

\ d

).
If
0
<
1
,
138 Maharams theorem 332P
(sup

\ c

) = (sup

\ c

) ( sup

0
e

\ c

) (e

0
\ c

)
( sup

0
e

\ c

) ( sup

0
f

\ d

)
= ( sup

1
f

\ d

)
(by the choice of
1
)
= ( sup

1
f

\ d

) (sup

\ d

).
If >
1
,
(sup

\ c

) = (sup

\ c

) (sup

\ d

) = (sup

\ d

).
But this means that

P, and evidently it is a proper extension of , which is supposed to be impossible.

XXXQQQ
() Thus has domain A and is the required measure-preserving homomorphism fromA to the principal
ideal B
d

of B.
332Q Proposition Let (A, ) and B, ) be totally nite measure algebras, and suppose that there are
measure-preserving Boolean homomorphisms
1
: A B and
2
: B A. Then (A, ) and (B, ) are
isomorphic.
proof Writing e

, f

for their Maharam-type- components, 332O (applied to both

1
and
2
) tells us that
(sup

) = (sup

)
for every . Because all these measures are nite,
e

= (sup

) (sup
>
e

)
= (sup

(sup
>
f

) = f

for every .
Similarly, writing e

, f

for the suprema in A, B of the atoms of measure , 332O tells us that

(sup

) = (sup

)
for every ]0, [, and hence that e

= f

for every , that is, that A and B have the same number of
atoms of measure .
So (A, ) and (B, ) are isomorphic, by 332J.
332R 332J tells us that if we know the magnitudes of the Maharam-type- components of a localizable
measure algebra, we shall have specied the algebra completely, so that all its properties are determined.
The calculation of its Maharam type is straightforward and useful, so I give the details.
Lemma Let (A, ) be a semi-nite measure algebra. Then c(A) 2
(A)
.
proof Let C A 0 be a disjoint set, and B A a -generating set of size (A).
(a) If A is purely atomic, then for each c C choose an atom c
t
c, and set f(c) = b : b B, c
t
b.
If c
1
, c
2
are distinct members of C, the set
a : a A, c
t
1
a c
t
2
a
is an order-closed subalgebra of A not containing either c
t
1
or c
t
2
, so cannot include B, and f(c
1
) ,= f(c
2
).
Thus f is injective, and
#(C) #(TB) = 2
(A)
.
332T Classication of localizable measure algebras 139
(b) Now suppose that A is not purely atomic; in this case (A) is innite. For each c C choose
an element c
t
c of non-zero nite measure. Let B be the subalgebra of A generated by B. Then the
topological closure of B is A itself (323J), and #(B) = (A) (331Gc). For c C set
f(c) = b : b B, (b c
t
)
1
2
c
t
.
Then f : C TB is injective. PPP If c
1
, c
2
are distinct members of C, then (because B is topologically
dense in A) there is a b B such that
((c
t
1
c
t
2
) (c
t
1
.b))
1
3
min( c
t
1
, c
t
2
).
But in this case
(c
t
1
\ b)
1
3
c
t
1
, (c
t
2
b)
1
3
c
t
2
,
and b f(c
1
)f(c
2
), so f(c
1
) ,= f(c
2
). QQQ Accordingly #(C) 2
#(B)
= 2
(A)
in this case also.
As C is arbitrary, c(A) 2
(A)
.
332S Theorem Let (A, ) be a localizable measure algebra. Then (A) is the least cardinal such
that () c(A) 2

() (A
a
) for every Maharam-type-homogeneous principal ideal A
a
of A.
proof Fix as the least cardinal satisfying () and ().
(a) By 331Hc, (A
a
) (A) for every a A, while c(A) 2
(A)
by 332R; so (A).
(b) Let C be a partition of unity in A consisting of elements of non-zero nite measure generating
Maharam-type-homogeneous principal ideals (as in the proof of 332B); then #(C) c(A) 2

, and there
is an injective function f : C T. For each c C, let B
c
A
c
be a -generating set of cardinal (A
c
),
and f
c
: B
c
an injection. Set
b

= supc : c C, f(c),
b
t

= supb : there is some c C such that b B

c
, f
c
(b) =
for < . Set B = b

: < b
t

: < if is innite, b

: < if is nite; then #(B) . Note

that if c C and b B
c
there is a b
t
B such that b = b
t
c. PPP Since B
c
,= , (A
c
) > 0; but this means
that (A
c
) is innite (see 331H) so is innite and b
t

B, where = f
c
(b); now b = b
t

c. QQQ
Let B be the closed subalgebra of A generated by B. Then C B. PPP For c C, we surely have c b

if f(c); but also, because C is disjoint, c b

= 0 if f(c). Consequently
c

= inf
f(c)
b

inf
\f(c)
(1 \ b

)
includes c. On the other hand, if d is any other member of C, there is some f(c)f(d), so that
d

(1 \ b

) = 0.
Since sup C = 1, it follows that c = c

; but c

B, so c B. QQQ
For any c C, look at b c : b B B. This is a closed subalgebra of A
c
(314F(a-i)) including B
c
, so
must be the whole of A
c
. Thus A
c
B for every c C. But sup C = 1, so a = sup
cC
a c B for every
a A, and A = B. Consequently (A) #(B) , and (A) = .
332T Proposition Let (A, ) be a localizable measure algebra and B a closed subalgebra of A. Then
(a) there is a function : B [0, ] such that (B, ) is a localizable measure algebra;
(b) (B) (A).
proof (a) Let D be the set of those b B such that the principal ideal B
b
has Maharam type at most (A)
and is a totally nite measure algebra when endowed with an appropriate measure. Then D is order-dense
in B. PPP Take any non-zero b
0
B. Then there is an a A such that a b
0
and 0 < a < . Set
c = infb : b B, a b; then c B and a c b
0
. If 0 ,= b B
c
, then c \ b belongs to B and is properly
included in c, so cannot include a; accordingly a b ,= 0. For b B
c
, set b = (a b). Because the
140 Maharams theorem 332T
map b a b is an injective order-continuous Boolean homomorphism, is countably additive and strictly
positive, that is, (B
c
, ) is a measure algebra. It is totally nite because c = a < .
Let d B
c
0 be such that B
d
is Maharam-type-homogeneous; suppose that its Maharam type is .
The map b b a is a measure-preserving Boolean homomorphism from B
d
to A
ad
, so by 332O A
ad
must have a non-zero Maharam-type-
t
component for some
t
; but this means that
(B
d
)
t
(A
ad
) (A).
Thus d D, while 0 ,= d c b
0
. As b
0
is arbitrary, D is order-dense. QQQ
Accordingly there is a partition of unity C in B such that C D. For each c C we have a functional
c
such that (B
c
,
c
) is a totally nite measure algebra of Maharam type at most (A); dene : B [0, ] by
setting b =

cC

c
(b c) for every b B. It is easy to check that (B, ) is a measure algebra (compare
322Ka); it is localizable because B (being order-closed in a Dedekind complete partially ordered set) is
Dedekind complete.
(b) The construction above ensures that every homogeneous principal ideal of B can have Maharam type
at most (A), since it must share a principal ideal with some B
c
for c C. Moreover, any disjoint set in B
is also a disjoint set in A, so c(B) c(A). So 332S tells us that (B) (A).
Remark I think the only direct appeal I shall make to this result will be when (A, ) is a probability algebra,
in which case (a) above becomes trivial, and the proof of (b) can be shortened to some extent, though I
think we still need some of the ideas of 332S.
332X Basic exercises (a) Let A be a Dedekind complete Boolean algebra. Show that it is isomorphic
to a simple product of Maharam-type-homogeneous Boolean algebras.
(b) Let A be a Boolean algebra of nite cellularity. Show that A is purely atomic.
(c) Let A be a purely atomic Boolean algebra. Show that c(A) is the number of atoms in A.
(d) Let A be any Boolean algebra, and Z its Stone space. Show that c(A) is equal to
c(Z) = sup#(() : G is a disjoint family of non-empty open subsets of Z,
the cellularity of the topological space Z.
(e) Let X be a topological space, and G its regular open algebra. Show that c(G) = c(X) as dened in
332Xd.
(f ) Let A be a Boolean algebra, and B any subalgebra of A. Show that c(B) c(A).
(g) Let A
i
)
iI
be any family of Boolean algebras, with simple product A. Show that the cellularity of A
is at most max(, #(I), sup
iI
c(A
i
)). Devise an elegant expression of a necessary and sucient condition
for equality.
(h) Let A be any Boolean algebra, and a A; let A
a
be the principal ideal generated by a. Show that
c(A
a
) c(A).
(i) Let (A, ) be a semi-nite measure algebra. Show that it has a partition of unity of cardinal c(A).
(j) Let (A, ) and (B, ) be localizable measure algebras. For each cardinal let e

, f

be their Maharam-
type- components, and A
e

, B
f

the corresponding principal ideals. Show that A and B are isomorphic,

as Boolean algebras, i c(A
e

) = c(B
f

) for every .
(k) Let be an ordinal, and

)
<
,

)
<
two families of non-negative real numbers such that

<

<

< for every . Show that there is a family

)
<
of non-negative real
numbers such that

<

for every < ,

for every < . (If only nitely many

of the

are non-zero, this is an easy special case of the max-ow min-cut theorem; see Bollob as 79,
III.1 or Anderson 87, 12.3.1.); there is a statement of the theorem in 4A3M in the next volume.) Show
that

can be chosen in such a way that if <

t
,
t
< then at least one of

is zero.
332 Notes Classication of localizable measure algebras 141
(l) Use 332Xk and 332M to give another proof of 332P.
(m) For each cardinal , write (B

) for the measure algebra of the usual measure on 0, 1

. Let
(A, ) be the simple product of (B

n
,

n
))
nN
and (B, ) the simple product of (A, ) with (B

).
(See 3A1E for the notation
n
,

.) Show that there is a measure-preserving Boolean homomorphism from

A to B, but that no such homomorphism can be order-continuous.
(n) For each cardinal , write (B

) for the measure algebra of the usual measure on 0, 1

. Let (A, )
be the simple product of (B

n
,

n
))
nN
and (B, ) the simple product of (B

n
,

n
))
nN
, where
n
=
for even n,
n
for odd n, while
n
= for odd n,
n
for even n. Show that there are order-continuous
measure-preserving Boolean homomorphisms from A to B and from B to A, but that these two measure
algebras are not isomorphic.
(o) Let C be a Boolean algebra. Show that the following are equiveridical: (i) C is isomorphic (as Boolean
algebra) to a closed subalgebra of a localizable measure algebra; (ii) there is a such that (C, ) is itself a
localizable measure algebra; (iii) C is Dedekind complete and for every non-zero c C there is a completely
additive real-valued functional on C such that c ,= 0. (Hint for (iii)(ii): show that the set of supports
of non-negative completely additive functionals is order-dense in C, so includes a partition of unity.)
332Y Further exercises (a) Let (A, ), (B, ) be atomless localizable measure algebras. For each
innite cardinal let e

, f

be their Maharam-type- components. Show that the following are equiveridical:

(i) (A, ) is isomorphic to a closed subalgebra of a principal ideal of (B, ); (ii) for every cardinal , the
magnitude of sup

is not greater than the magnitude of sup

.
(b) Let (A, ) and (B, ) be any semi-nite measure algebras, and (

A, ), (

B, ) their localizations
(322O-322P). Let e
i
)
iI
, f
j
)
jJ
be partitions of unity in A, B respectively into elements of nite measure
generating homogeneous principal ideals A
e
i
, B
f
j
. For each innite cardinal set I

= i : (A
e
i
) = ,
J

= j : (B
f
j
) = ; for ]0, [, set I

= i : e
i
is an atom, e
i
= , J

= j : f
j
is an atom,
f
j
= . Show that (

A, ) and (

B, ) are isomorphic i for each u, either

iI
u
e
i
=

jJ
u
f
j
< or

iI
u
e
i
=

jJ
u
f
j
= and #(I
u
) = #(J
u
).
(c) Let (A, ) and (B, ) be non-zero localizable measure algebras; let e

, f

be their Maharam-type-
components. Show that the following are equiveridical: (i) A is isomorphic, as Boolean algebra, to an order-
closed subalgebra of a principal ideal of B; (ii) c(A

) c(B

) for every cardinal , where A

, B

are the
principal ideals generated by sup

and sup

respectively.
332 Notes and comments Maharams theorem tells us that all localizable measure algebras in particular,
all -nite measure algebras can be obtained from the basic algebra A = 0, a, 1 \ a, 1, with a = (1 \ a) =
1
2
, by combining the constructions of probability algebra free products, scalar multiples of measures and
simple products. But what is much more important is the fact that we get a description of our measure
algebras in terms suciently explicit to make a very wide variety of questions resolvable. The description
I oer in 332J hinges on the complementary concepts of Maharam type and magnitude. If you like, the
magnitude of a measure algebra is a measure of its width, while its Maharam type is a measure of its depth.
The latter is more important just because, for localizable algebras, we have such a simple decomposition
into algebras of nite magnitude. Of course there is a good deal of scope for further complications if we seek
to consider non-localizable semi-nite algebras. For these, the natural starting point is a proper description
of their localizations, which is not dicult (332Yb).
Observe that 332C gives a representation of a localizable measure algebra as the measure algebra of a
measure space which is completely dierent from the Stone representation in 321K. It is less canonical (since
there is a degree of choice about the partition e
i
)
iI
) but very much more informative, since the
i
,
i
carry enough information to identify the measure algebra up to isomorphism (332K).
Cellularity is the second cardinal function I have introduced in this chapter. It refers as much to
topological spaces as to Boolean algebras (see 332Xd-332Xe). There is an interesting question in this
context. If A is an arbitrary Boolean algebra, is there necessarily a disjoint set in A of cardinal c(A)? This
142 Maharams theorem 332 Notes
is believed to be undecidable from the ordinary axioms of set theory (including the axiom of choice); see
Juh asz 71, 3.1 and 6.5. But for semi-nite measure algebras we have a denite answer (332F).
Maharams classication not only describes the isomorphism classes of localizable measure algebras, but
also tells us when to expect Boolean homomorphisms between them (332P, 332Yc). I have given 332P
only for atomless totally nite measure algebras because the non-totally-nite case (332Ya, 332Yc) seems
to require a new idea, while atoms introduce acute combinatorial complications.
I oer 332T as an example of the kind of result which these methods make very simple. It fails for general
Boolean algebras; in fact, there is for any a countably -generated Dedekind complete Boolean algebra
A with cellularity (Koppelberg 89, 13.1), so that T is isomorphic to an order-closed subalgebra of A,
and if > c then (T) > (332R).
For totally nite measure algebras we have a kind of weak Schroder-Bernstein theorem: if we have two
of them, each isomorphic to a closed subalgebra of the other, they are isomorphic (332Q). This fails for
-nite algebras (332Xn). I call it a weak Schroder-Bernstein theorem because it is not clear how to build
the isomorphism from the two injections; strong Schroder-Bernstein theorems include denite recipes for
constructing the isomorphisms declared to exist (see, for instance, 344D below).
333 Closed subalgebras
Proposition 332P tells us, in eect, which totally nite measure algebras can be embedded as closed subal-
gebras of each other. Similar techniques make it possible to describe the possible forms of such embeddings.
In this section I give the fundamental theorems on extension of measure-preserving homomorphisms from
closed subalgebras (333C, 333D); these rely on the concept of relative Maharam type (333A). I go on to
describe possible canonical forms for structures (A, , C), where (A, ) is a totally nite measure algebra and
C is a closed subalgebra of A (333K, 333N). I end the section with a description of xed-point subalgebras
(333R).
333A Denitions (a) Let A be a Boolean algebra and C a subalgebra of A. The relative Maharam
type of A over C,
C
(A), is the smallest cardinal of any set A A such that A C -generates A.
(b) In this section, I will regularly use the following notation: if A is a Boolean algebra, C is a subalgebra
of A, and a A, then I will write C
a
for c a : c C. Observe that C
a
is a subalgebra of the principal
ideal A
a
(because c c a : C A
a
is a Boolean homomorphism); it is included in C i a C..
(c) Still taking A to be a Boolean algebra and C to be a subalgebra of A, I will say that an element a of
A is relatively Maharam-type-homogeneous over C if
C
b
(A
b
) =
C
a
(A
a
) for every non-zero b a.
333B Evidently this is a generalization of the ordinary concept of Maharam type as used in 331-332;
if C = 0, 1 then
C
(A) = (A). The rst step is naturally to check the results corresponding to 331H.
Lemma Let A be a Boolean algebra and C a subalgebra of A.
(a) If a b in A, then
C
a
(A
a
)
C
b
(A
b
). In particular,
C
a
(A
a
)
C
(A) for every a A.
(b) The set a : a A is relatively Maharam-type-homogeneous over C is order-dense in A.
(c) If A is Dedekind complete and C is order-closed in A, then C
a
is order-closed in A
a
.
(d) If a A is relatively Maharam-type-homogeneous over C then either A
a
= C
a
, so that
C
a
(A
a
) = 0
and a is a relative atom of A over C (denition: 331A), or
C
a
(A
a
) .
(e) If D is another subalgebra of A and D C, then
(A
a
) =
0,a]
(A
a
)
D
a
(A
a
)
C
a
(A
a
)
A
a
(A
a
) = 0
for every a A.
proof (a) Let D A
b
be a set of cardinal
C
b
(A
b
) such that DC
b
-generates A
b
. Set D
t
= d a : d D.
Then D
t
C
a
-generates A
a
. PPPApply 313Mc to the map d d a : A
b
A
a
, as in 331Hc. QQQ Consequently

C
a
(A
a
) #(D
t
) #(D) =
C
b
(A
b
),
333C Closed subalgebras 143
as claimed. Setting b = 1 we get
C
a
(A
a
)
C
(A).
(b) Just as in the proof of 332A, given b A 0, there is an a A
b
0 minimising
C
a
(A
a
), and this
a must be relatively Maharam-type-homogeneous over C.
(c) C
a
is the image of the Dedekind complete Boolean algebra C under the order-continuous Boolean
homomorphism c c a, so must be order-closed (314Fa).
(d) Suppose that
C
a
(A
a
) is nite. Let D A
a
be a nite set such that D C
a
-generates A
a
. Then
there is a non-zero b A
a
such that b d is either 0 or b for every d D. But this means that C
b
= d b :
d D C
a
, which -generates A
b
; so that
C
b
(A
b
) = 0. Since a is relatively Maharam-type-homogeneous
over C,
C
a
(A
a
) must be zero, that is, A
a
= C
a
.
(e) The middle inequality is true just because A
a
will be -generated by DC
a
whenever it is -generated
by D D
a
. The neighbouring inequalities are special cases of the middle one, and the outer equalities are
elementary.
333C Theorem Let (A, ) and (B, ) be totally nite measure algebras, and C a closed subalgebra of
A. Let : C B be a measure-preserving Boolean homomorphism.
(a) If, in the notation of 333A,
C
(A)
[C]
b
(B
b
) for every non-zero b B, there is a measure-preserving
Boolean homomorphism : A B extending .
(b) If
C
a
(A
a
) =
[C]
b
(B
b
) for every non-zero a A, b B, then there is a measure algebra isomorphism
: A B extending .
proof In both parts, the idea is to use the technique of the proof of 331I to construct as the last of an
increasing family

of measure-preserving homomorphisms from closed subalgebras C

of A, where
=
C
(A). Let a

)
<
be a family in A such that C a

: < -generates A. Write D for [C];

remember that D is a closed subalgebra of B (324L).
(a)(i) In this case, we can describe the C

immediately; C

will be the closed subalgebra of A generated

by C a

: < . The induction starts with C

0
= C,
0
= .
(ii) For the inductive step to a successor ordinal + 1, where < , suppose that C

and

have
been dened. Take any non-zero b B. We are supposing that
D
b
(B
b
) > #(), so B
b
cannot be
-generated by
D = D
b
b

: < =

[C]
b
b

: < = [C a

: < ],
writing c = b

c for c C

. As is order-continuous, [C

] is precisely the closed subalgebra of B

b
generated by D (314Gb), and is therefore not the whole of B
b
.
But this means that B
b
,= b

c : c C

. As b is arbitrary,

satises the conditions of 331D, and

has an extension to a measure-preserving Boolean homomorphism
+1
: C
+1
B, since C
+1
is just the
closed subalgebra of A generated by C a

.
(iii) For the inductive step to a non-zero limit ordinal , we can argue exactly as in part (d) of the
proof of 331I; C

will be the metric closure of C

<
C

, so we can take

: C

B to be the unique
measure-preserving homomorphism extending

<

.
Thus the induction proceeds, and evidently =

will be a measure-preserving homomorphism from A

to B extending .
(b) (This is rather closer to the proof of 331I, being indeed a direct generalization of it.) Observe that
the hypothesis (b) implies that 1
A
is relatively Maharam-type-homogeneous over C; so either = 0, in which
case A = C, B = [C] and the result is trivial, or , by 333Bd. Let us therefore take it that is innite.
We are supposing, among other things, that
D
(B) = ; let b

)
<
be a family in B such that B is
-generated by Db

: < . This time, as in 331I, we shall have to choose further families a

t

)
<
and
b
t

)
<
, and
C

will be the closed subalgebra of A generated by

C a

: < a
t

: < ,
D

will be the closed subalgebra of B generated by

144 Maharams theorem 333C
D b

: < b
t

: < ,

: C

will be a measure-preserving homomorphism.

The induction will start with C
0
= C, D
0
= D and
0
= , as in (a).
(i) For the inductive step to a successor ordinal + 1, where < , suppose that C

, D

and

have
been dened.
() Let b B 0. Because

D
b
(B
b
) = > #(b

: < b
t

: < ),
B
b
cannot be -generated by D
b
b b

: < b b
t

: < , and cannot be equal to b d : d D

.
As b is arbitrary, there is an extension of

to a measure-preserving homomorphism

from C
t

to B, where
C
t

is the closed subalgebra of A generated by Ca

: a
t

: < . Setting b
t

(a

), the image
D
t

[C

] will be the closed subalgebra of B generated by D b

: < b
t

: .
() Now, as in 331I, we must repeat the argument of (), applying it now to
1

: D

A. If
a A 0,

C
a
(A
a
) = > #(a

: a
t

: < ),
so that A
a
cannot be a c : c C
t

. As a is arbitrary,
1

has an extension to a measure-preserving

homomorphism

: D
+1
C
+1
, where D
+1
is the subalgebra of B generated by D
t

, that is,
the closed subalgebra of B generated by D b

: b
t

: < , and C
+1
is the subalgebra of A
generated by C
t

a
t

, setting a
t

(b

).
We can therefore take
+1
=
1

: C
+1
D
+1
, as in 331I.
(ii) The inductive step to a non-zero limit ordinal is exactly the same as in (a) above or in 331I;
C

is the metric closure of C

<
C

, D

is the metric closure of D

<
D

, and

is the unique
measure-preserving homomorphism from C

to D

extending every

for < .
(iii) The induction stops, as before, with =

: C

, where C

= A, D

= B.
333D Corollary Let (A, ) and (B, ) be totally nite measure algebras and C a closed subalgebra of
A. Suppose that
(C) < max(, (A)) min(B
b
) : b B 0.
Then any measure-preserving Boolean homomorphism : C B can be extended to a measure-preserving
Boolean homomorphism : A B.
proof Set = min(B
b
) : b B 0. Then for any non-zero b B,

[C]
b
(B
b
) .
PPP There is a set C C, of cardinal (C), which -generates C, so that C
t
= b c : c C -generates
[C]
b
. Now there is a set D B
b
, of cardinal
[C]
b
(B
b
), such that [C]
b
D -generates B
b
. In this case
C
t
D must -generate B
b
, so #(C
t
D). But #(C
t
) #(C) < and is innite, so we must have
#(D) , as claimed. QQQ
On the other hand,
C
(A) (A) . So we can apply 333C to give the result.
333E Theorem Let (C, ) be a totally nite measure algebra, an innite cardinal, and (B

) the
measure algebra of the usual measure on 0, 1

. Let (A,

) be the localizable measure algebra free product

of (C, ) and (B

), and : C A the corresponding homomorphism. Then for any non-zero a A,

[C]
a
(A
a
) = ,
in the notation of 333A above.
proof Recall from 325Dd that [C] is a closed subalgebra of A.
(a) Let b

)
<
be the canonical independent family in B

of elements of measure
1
2
. Let
t
: B

A
be the canonical map, and set b
t

=
t
b

for each .
333E Closed subalgebras 145
We know that b

: < -generates B

(see part (a) of the proof of 331K). Consequently [C] b

t

:
< -generates A. PPP Let A
1
be the closed subalgebra of A generated by [C] b
t

: < . Because

t
: B

A is order-continuous (325Da),
t
[B

] A
1
(313Mb). But this means that A
1
includes
[C]
t
[B

] and therefore includes the image of C B

in A; because this is topologically dense in A

(325Dc), A
1
= A, as claimed. QQQ
(b) It follows that

[C]
a
A
a

[C]
A .
(c) We need to know that if < and e belongs to the closed subalgebra E

of A generated by
[C] b
t

: ,= , then

(e b
t

) =
1
2

e. PPP Set
E = [C] b
t

: ,= , F = e
0
. . . e
n
: e
0
, . . . , e
n
E.
Then every member of F is expressible in the form
d = a inf
J
b
t

,
where a C and J is nite. Now

d = a (inf
J
b

) = 2
#(J)
a,

(b
t

d) = a (b

inf
J
b

) = 2
#(J])
a =
1
2

d.
Now consider the set
G = d : d A,

(b

d) =
1
2

d.
We have 1
A
F G, and F is closed under . Secondly, if d, d
t
G and d d
t
, then

(b

(d
t
\ d)) =

(b

d
t
)

(b

d) =
1
2

d
t

1
2

d =
1
2

(d
t
\ d),
so d
t
\ d G. Also, if H G is non-empty and upwards-directed,

(b

supH) =

(sup
dH
b

d) = sup
dH

(b

d) = sup
dH
1
2

d =
1
2

(supH),
so supH G. By the Monotone Class Theorem (313Gc), G includes the order-closed subalgebra of D
generated by F. But this is just E

. QQQ
(d) The next step is to see that
[C]
a
(A
a
) > 0. PPP By (a) and 323J, A is the metric closure of the
subalgebra A
0
generated by [C] b
t

: < , so there must be an a

0
A
0
such that

(a
0
.a)
1
4

a.
Now there is a nite J such that a
0
belongs to the subalgebra A
1
generated by [C] b
t

: J.
Take any J (this is where I use the hypothesis that is innite). If c C, then by (c) we have

((a c) .(a b
t

)) =

(a (c .b
t

))

(a
0
(c .b
t

))

(a .a
0
)
=

(a
0
b
t

) +

(a
0
c) 2

(a
0
c b
t

(a .a
0
)
=
1
2

a
0

(a .a
0
)
(because both a
0
and a
0
c belong to E

1
2

a
3
2

(a .a
0
) > 0.
Thus a b
t

is not of the form a c for any c C, and A

a
,= [C]
a
, so that
[C]
a
(A
a
) > 0. QQQ
(e) It follows that
[C]
a
(A
a
) is innite. PPP There is a non-zero d a which is relatively Maharam-type-
homogeneous over [C]. By (d), applied to d,
[C]
d
(A
d
) > 0; but now 333Bd tells us that
[C]
d
(A
d
) must
be innite, so
[C]
a
(A
a
) is innite. QQQ
146 Maharams theorem 333E
(f ) If = , we can stop here. If > , we continue, as follows. Let D A
a
be any set of cardinal
less than . Each d D a belongs to the closed subalgebra of A generated by C = [C] b
t

: < .
But because A is ccc, this is just the -subalgebra of A generated by C (331Ge). So d belongs to the
closed subalgebra of A generated by some countable subset C
d
of C, by 331Gd. Now J
d
= : b
t

C
d
is
countable. Set J =

dDa]
J
d
; then
#(J) max(, #(D a)) = max(, #(D)) < ,
so J ,= , and there is a J. Accordingly [C] D a is included in E

, as dened in (c) above,

and [C]
a
D E

. As A
a
E

is a closed subalgebra of A
a
, it includes the closed subalgebra generated
by [C]
a
D. But a b
t

surely does not belong to E

, since

(a b
t

b
t

) =

(a b
t

) =
1
2

a > 0,
and

(a b
t

b
t

) ,=
1
2

(a b
t

). Thus a b
t

cannot belong to the closed subalgebra of A

a
generated by
[C]
a
D, and [C]
a
D does not -generate A
a
. As D is arbitrary,
[C]
a
(A
a
) .
This completes the proof.
333F Corollary Let (A, ) be a totally nite measure algebra, C a closed subalgebra of A and an
innite cardinal. Let (B

) be the measure algebra of the usual measure on 0, 1

.
(i) Suppose that
C
(A). Let (C

) be the localizable measure algebra free product of (C, C)

and (B

), and : C C

the corresponding homomorphism. Then there is a measure-preserving

Boolean homomorphism : A C

extending .
(ii) Suppose further that =
C
a
(A
a
) for every non-zero a A. Then can be taken to be an isomorphism.
proof All we have to do is apply 333C with B = C

, using 333E to see that the hypothesis

[C]
b
(B
b
) = for every non-zero b B
is satised.
333G Corollary Let (C, ) be a totally nite measure algebra. Suppose that max(, (C)) is a
cardinal, and write (B

) for the measure algebra of the usual measure on 0, 1

. Let (C

) be the
localizable measure algebra free product of (C, ) and (B

). Then
(a) C

is Maharam-type-homogeneous, with Maharam type ;

(b) for every measure-preserving Boolean homomorphism : C C there is a measure-preserving auto-
morphism : C

such that (c 1) = c 1 for every c C, writing c 1 for the canonical

image in C

of any c C.
proof Write A for C

, as in 333E.
(a) If C C is a set of cardinal (C) which -generates C, and B B

a set of cardinal which

-generates B

(331K), then c b : c C, b B is a set of cardinal at most max(, (C), ) = which

-generates A (because the subalgebra it generates is topologically dense in A, by 325Dc). So (A) .
On the other hand, if a A is non-zero, then (A
a
)
[C]
a
(A
a
) , by 333E; so A is Maharam-type-
homogeneous, with Maharam type .
(b) Writing D = c 1 : c C for the canonical image of C in A, we have a measure-preserving
automorphism
1
: D D dened by setting
1
(c 1) = c 1 for every c C. Because
1
[D] D, 333Be
and 333E tell us that
= (A
a
)

1
[D]
a
(A
a
)
D
a
(A
a
) =
for every non-zero a A, so we can use 333Cb, with B = A, to see that
1
can be extended to a measure-
preserving automorphism on A.
333I Closed subalgebras 147
333H I turn now to the classication of closed subalgebras.
Theorem Let (A, ) be a localizable measure algebra and C a closed subalgebra of A. Then there are

i
)
iI
, c
i
)
iI
,
i
)
iI
such that
for each i I,
i
is a non-negative completely additive functional on C,
c
i
= [[
i
> 0]] C,

i
is 0 or an innite cardinal,
(C
c
i
,
i
C
c
i
) is a totally nite measure algebra, writing C
c
i
for the principal ideal of C
generated by c
i
,

iI

i
c = c for every c C,
there is a measure-preserving isomorphism from A to the simple product

iI
C
c
i

i
of the localizable measure algebra free products C
c
i

i
of (C
c
i
,
i
C
c
i
), (B

i
,

i
), writing
(B

) for the measure algebra of the usual measure on 0, 1

.
Moreover, may be taken such that
for every c C, c = (c c
i
) 1)
iI
, writing c 1 for the image in C
c
i

i
of c C
c
i
.
Remark Recall that [[
i
> 0]] is that element of C such that
i
c > 0 whenever c C and 0 ,= c [[
i
> 0]],

i
c 0 whenever c C and c [[
i
> 0]] = 0 (326O).
proof (a) Let A be the set of those elements of A which are relatively Maharam-type-homogeneous over C
(see 333Ac). By 333Bb, A is order-dense in A (compare part (a) of the proof of 332B), and consequently
A
t
= a : a A, a < is order-dense in A. So there is a partition of unity a
i
)
iI
in A consisting of
members of A
t
(313K). For each i I, set
i
c = (a
i
c) for every c C; then
i
is non-negative, and it is
completely additive by 327E. Because a
i
)
iI
is a partition of the identity in A,
c =

iI
(c a
i
) =

iI

i
c
for every c C. Next, (C
c
i
,
i
C
c
i
) is a totally nite measure algebra. PPP C
c
i
is a Dedekind -complete
Boolean algebra because C is.
i
C
c
i
is a non-negative countably additive functional because
i
is. If c C
c
i
and
i
c = 0, then c = 0 by the choice of c
i
. QQQ Note also that
(a
i
\ c
i
) =
i
(1 \ c
i
) = 0,
so that a
i
c
i
.
(b) By 333Bd, any nite
i
must actually be zero. The next element we need is the fact that, for
each i I, we have a measure-preserving isomorphism c c a
i
from (C
c
i
,
i
C
c
i
) to (C
a
i
, C
a
i
). PPP
Of course this is a ring homomorphism. Because a
i
c
i
, it is a surjective Boolean homomorphism. It is
measure-preserving by the denition of
i
, and therefore injective. QQQ
(c) Still focusing on a particular i I, let A
a
i
be the principal ideal of A generated by a
i
. Then
we have a measure-preserving isomorphism
i
: A
a
i
C
a
i

i
, extending the canonical homomorphism
c c 1 : C
a
i
C
a
i

i
. PPP When
i
is innite, this is just 333F(ii). But the only other case is when

i
= 0, that is, C
a
i
= A
a
i
, while B

i
= 0, 1 and C
a
i

= C
c
i
. QQQ
The isomorphism between (C
c
i
,
i
C
c
i
) and (C
a
i
, C
a
i
) induces an isomorphism between C
c
i

i
and
C
a
i

i
. So we have a measure-preserving isomorphism
i
: A
a
i
C
c
i

i
such that
i
(c a
i
) = c 1
for every c C
c
i
.
(d) By 322Kd, we have a measure-preserving isomorphism a a a
i
)
iI
: A

iI
A
a
i
.
Putting this together with the isomorphisms of (c), we have a measure-preserving isomorphism : A

iI
C
c
i

i
, setting a =
i
(a a
i
))
iI
for a A. Observe that, for c C,
c =
i
(c a
i
))
iI
= (c c
i
) 1)
iI
,
as required.
333I Remarks (a) I hope it is clear that whenever (C, ) is a Dedekind complete measure algebra,

i
)
iI
is a family of non-negative completely additive functionals on C such that

iI

i
= , and
i
)
iI
148 Maharams theorem 333I
is a family of cardinals all innite or zero, then the construction above can be applied to give a measure al-
gebra (A, ), the product of the family C
c
i

i
)
iI
, together with an order-continuous measure-preserving
homomorphism : C A; and that the partition of unity a
i
)
iI
in A corresponding to this product (315E)
has
i
c = (a
i
c) for every c C and i I, while each principal ideal A
a
i
can be identied with C
c
i

i
,
so that a
i
is relatively Maharam-type-homogeneous over [C]. Thus any structure (C, ,
i
)
iI
,
i
)
iI
) of
the type described here corresponds to an embedding of C as a closed subalgebra of a localizable measure
algebra.
(b) The obvious next step is to seek a complete classication of objects (A, , C), where (A, ) is a
localizable measure algebra and C is a closed subalgebra, corresponding to the classication of localizable
measure algebras in terms of the magnitudes of their Maharam-type- components in 332J. The general
case seems to be complex. But I can deal with the special case in which (A, ) is totally nite. In this case,
we have the following facts.
333J Lemma Let (A, ) be a totally nite measure algebra, and C a closed subalgebra. Let A be the
set of relative atoms of A over C. Then there is a unique sequence
n
)
nN
of additive functionals on C such
that (i)
n+1

n
for every n (ii) there is a disjoint sequence a
n
)
nN
in A such that sup
nN
a
n
= sup A
and
n
c = (a
n
c) for every n N, c C.
Remark I hope it is plain from my wording that it is the
n
which are unique, not the a
n
.
proof (a) For each a A set
a
(c) = (c a) for c C. Then
a
is a non-negative completely additive
real-valued functional on C (see 326Kd).
The key step is I suppose in (c) below; I approach by a two-stage argument. For each b A write A
b
for
a : a A, a b = 0.
(b) For every b A, non-zero c C there are a A
b
, c
t
C such that 0 ,= c
t
c and
a
(d)
e
(d)
whenever d C, e A
b
and d c
t
. PPP??? Otherwise, choose a
n
)
nN
and c
n
)
nN
as follows. Since 0, c wont
serve for a, c
t
, there must be an a
0
A
b
such that
a
0
(c) > 0. Let > 0 be such that
a
0
(c) > c and set
c
0
= c [[
a
0
> C]]; then c
0
C and 0 ,= c
0
c. Given that a
n
A
b
, c
n
C and 0 ,= c
n
c, then there
must be a
n+1
A
b
, d
n
C such that d
n
c
n
and
a
n+1
(d
n
) >
a
n
(d
n
). Set c
n+1
= d
n
[[
a
n+1
>
a
n
]], so
that c
n+1
C and 0 ,= c
n+1
c
n
, and continue.
There is some n N such that n 1. For any i < n, the construction ensures that
0 ,= c
n+1
c
i+1
[[
a
i+1
>
a
i
]],
so
a
i
(c
n+1
) <
a
i+1
(c
n+1
); also c
n+1
c
0
so
(a
i
c
n+1
) =
a
i
(c
n+1
)
a
0
(c
n+1
) > c
n+1
.
But this means that

n1
i=0
(a
i
c
n+1
) > c
n+1
and there must be distinct j, k < n such that a
j
a
k
c
n+1
is non-zero. Because a
j
, a
k
A there are d
t
, d
tt
C such that a
j
a
k
= a
j
d
t
= a
k
d
tt
; set d =
c
n+1
d
t
d
tt
, so that d C and
a
j
d = a
j
a
k
c
n+1
= a
k
d,
a
j
(d) = (a
j
a
k
c
n+1
) =
a
k
(d).
But as 0 ,= d [[
a
i+1
>
a
i
]] for every i < n,
a
0
(d) <
a
1
(d) < . . . <
a
n
(d), so this is impossible. XXXQQQ
(c) Now for a global, rather than local, version of the same idea. For every b A there is an a A
b
such
that and
a

e
whenever e A
b
. PPP (i) By (b), the set C of those c C such that there is an a A
b
such
that
a
C
c

e
C
c
for every e A
b
is order-dense in C. Let c
i
)
iI
be a partition of unity in C consisting
of members of C, and for each i I choose a
i
A
b
such that
a
i
C
c
i

e
C
c
i
for every e A
b
. Consider
a = sup
iI
a
i
c
i
. (ii) If a
t
A and a
t
a, then for each i I there is a d
i
C such that a
i
a
t
= a
i
d
i
.
Set d
t
= sup
iI
c
i
d
i
; then (because c
i
)
iI
is disjoint)
a d
t
= sup
iI
a
i
c
i
d
i
= sup
iI
a
i
c
i
a
t
= a a
t
= a
t
.
As a
t
is arbitrary, this shows that a A. (iii) Of course a b = 0, so a A
b
. Now take any e A
b
, d C.
Then

a
(d) =

iI

a
i
(c
i
d)

iI

e
(c
i
d) =
e
(d).
333K Closed subalgebras 149
So this a has the required property. QQQ
(d) Choose a
n
)
nN
inductively in A so that, for each n, a
n
sup
i<n
a
i
= 0 and
a
n

e
whenever
e A and e sup
i<n
a
i
= 0. Set
n
=
a
n
. Because a
n+1
sup
i<n
a
i
= 0,
n+1

n
for each n. Also
sup
nN
a
n
= supA. PPP Take any a A and set e = a \ sup
nN
a
n
. Then e A and, for any n N,
e sup
i<n
a
i
= 0, so
e

a
n
and
e =
e
(1)
a
n
(1) = a
n
.
But as a
n
)
nN
is disjoint, this means that e = 0, that is, a sup
nN
a
n
. As a is arbitrary, supA sup
nN
a
n
.
QQQ
(e) Thus we have a sequence
n
)
nN
of the required type, witnessed by a
n
)
nN
. To see that it is
unique, suppose that
t
n
)
nN
, a
t
n
)
nN
are another pair of sequences with the same properties. Note
rst that if c C and 0 ,= c [[
t
i
> 0]] there is some k N such that c a
t
i
a
k
,= 0; this is because
(a
t
i
c) =
t
i
(c) > 0, so that a
t
i
c ,= 0, while a
t
i
supA = sup
kN
a
k
. ??? Suppose, if possible, that there is
some n such that
n
,=
t
n
; since the situation is symmetric, there is no loss of generality in supposing that

t
n
,
n
, that is, that c = [[
t
n
>
n
]] ,= 0. For any i n,
t
i

t
n
so c [[
t
i
> 0]]. We may therefore choose
c
0
, . . . , c
n+1
C
c
0 and k(0), . . . , k(n) N such that c
0
= c and, for i n,
c
i
a
t
i
a
k(i)
,= 0
(choosing k(i), recalling that 0 ,= c
i
c [[
t
i
> 0]]),
c
i+1
C, c
i+1
c
i
, c
i+1
a
t
i
= c
i+1
a
k(i)
= c
i
a
t
i
a
k(i)
(choosing c
i+1
, using the fact that a
t
i
and a
k(i)
both belong to A see the penultimate sentence in part (b)
of the proof.) On reaching c
n+1
, we have 0 ,= c
n+1
c so
n
(c
n+1
) <
t
n
(c
n+1
). On the other hand, for
each i n,
c
n+1
a
t
i
a
k(i)
= c
n+1
c
i+1
a
t
i
a
k(i)
= c
n+1
a
t
i
= c
n+1
a
k(i)
,
so

n
(c
n+1
) <
t
n
(c
n+1
)
t
i
(c
n+1
) = (c
n+1
a
t
i
) = (c
n+1
a
k(i)
) =
k(i)
(c
n+1
),
and k(i) must be less than n. There are therefore distinct i, j n such that k(i) = k(j). But in this case
c
n+1
a
t
i
= c
n+1
a
k(i)
= c
n+1
a
k(j)
= c
n+1
a
t
j
,= 0
because 0 ,= c
n+1
[[
t
j
> 0]]. So a
t
i
, a
t
j
cannot be disjoint, breaking one of the rules of the construction. XXX
Thus
n
=
t
n
for every n N.
This completes the proof.
333K Theorem Let (A, ) be a totally nite measure algebra and C a closed subalgebra of A. Then
there are unique families
n
)
nN
,

)
K
such that
K is a countable set of innite cardinals,
for i N K,
i
is a non-negative countably additive functional on C, and

iNK

i
c = c
for every c C,

n+1

n
for every n N, and

,= 0 for K,
setting e
i
= [[
i
> 0]] C, and giving the principal ideal C
e
i
generated by e
i
the measure

i
C
e
i
for each i NK, and writing (B

) for the measure algebra of the usual measure on

0, 1

for K, we have a measure algebra isomorphism

: A

nN
C
e
n

K
C
e

such that
c = (c e
n
)
nN
, (c e

) 1)
K
)
for each c C, writing c 1 for the canonical image in C
e

of c C
e

.
proof (a) I aim to use the construction of 333H, but taking much more care over the choice of a
i
)
iI
in
part (a) of the proof there. We start by taking a
n
)
nN
as in 333J, and setting
n
c = (a
n
c) for every
n N, c C; then these a
n
will deal with the part in supA, as dened in the proof of 333J.
150 Maharams theorem 333K
(b) The further idea required here concerns the treatment of innite . Let b
i
)
iI
be any partition
of unity in A consisting of non-zero members of A which are relatively Maharam-type-homogeneous over
C, and
i
)
iI
the corresponding cardinals, so that
i
= 0 i b
i
A. Set I
1
= i : i I,
i
. Set
K =
i
: i I
1
, so that K is a countable set of innite cardinals, and for K set J

= i :
i
= ,
a

= sup
iJ

b
i
for K. Now every a

is relatively Maharam-type-homogeneous over C. PPP (Compare

332H.) J

must be countable, because A is ccc. If 0 ,= a a

, there is some i J

such that a b
i
,= 0; now

C
a
(A
a
)
C
ab
i
(A
ab
i
) =
i
= .
At the same time, for each i J

, there is a set D
i
A
b
i
such that #(D
i
) = and C
b
i
D
i
-generates
A
b
i
. Set D =

iJ

D
i
b
i
: i J

; then
#(D) max(, #(J

), sup
iK
#(D
i
)) = .
Let B be the closed subalgebra of A
a

generated by C
a

D. Then
C
b
i
D
i
b b
i
: b B = B A
b
i
so B A
b
i
for each i J

, and B = A
a

. Thus C
a

D -generates A
a

, and

C
a

(A
a

) min
0,=aa

C
a
(A
a
).
This shows that a

is relatively Maharam-type-homogeneousover C, with

C
a

(A
a

) = . QQQ
Since evidently J

)
K
and a

)
K
are disjoint, and sup
K
a

= sup
iI
1
b
i
, this process yields a
partition a
i
)
iNK
of unity in A. Now the arguments of 333H show that we get an isomorphism of the
kind described.
(c) To see that the families
n
)
nN
,

)
K
(and therefore the e
i
and the (C
e
i
,
i
C
e
i
), but not ) are
uniquely dened, argue as follows. Take families
n
)
nN
,

K
which correspond to an isomorphism
: A D =

nN
C
e
n

K
C
e

,
writing e
i
= [[
i
> 0]] for i N

K. In the simple product

nN
C
e
n

K
C
e

, we have a partition
of unity e

i
)
iN

K
corresponding to the product structure. Now for d e

i
, we have

[C]
d
(D
d
) = 0 if i N,
= if i =

K.
So

K must be
: , a A,
C
a
(A
a
) = = K,
and for

K,

1
e

= supa : a A,
C
a
(A
a
) = = a

,
so that

. On the other hand,

1
e

n
)
nN
must be a disjoint sequence with supremum supA, and
the corresponding functionals
n
are supposed to form a non-increasing sequence, so must be equal to the

n
by 333J.
333L Remark Thus for the classication of structures (A, , C), where (A, ) is a totally nite measure
algebra and C is a closed subalgebra, it will be enough to classify objects (C, ,
n
)
nN
,

)
K
), where
(C, ) is a totally nite measure algebra,

n
)
nN
is a non-increasing sequence of non-negative countably additive functionals on C,
K is a countable set of innite cardinals (possibly empty),

)
K
is a family of non-zero non-negative countably additive functionals on C,

n=0

n
+

= .
To do this we need the concept of standard extension of a countably additive functional on a closed
subalgebra of a measure algebra, treated in 327F-327G, together with the following idea.
333M Closed subalgebras 151
333M Lemma Let (C, ) be a totally nite measure algebra and
i
)
iI
a family of countably additive
functionals on C. For i I, R set e
i
= [[
i
> ]] (326P), and let C
0
be the closed subalgebra of C
generated by e
i
: i I, R. Write for the -algebra of subsets of R
I
generated by sets of the form
E
i
= x : x(i) > as i runs through I, runs over R. Then
(a) there is a measure , with domain , such that there is a measure-preserving isomorphism : /^

C
0
for which E

i
= e
i
for every i I, R, writing ^

for
1
[0];
(b) this formula determines both and ;
(c) for every E , i I, we have

i
E

=
_
E
x(i)(dx);
(d) for every i I,
i
is the standard extension of
i
C
0
to C;
(e) for every i I,
i
0 i x(i) 0 for -almost every x;
(f) for every i, j I,
i

j
i x(i) x(j) for -almost every x;
(g) for every i I,
i
= 0 i x(i) = 0 for -almost every x.
proof (a) Express (C, ) as the measure algebra of a measure space (Y, T, ); write : T C for the
corresponding homomorphism. For each i I let f
i
: Y R be a T-measurable, -integrable function such
that
_
H
f
i
=
i
H for every H T. Dene : Y R
I
by setting (y) = f
i
(y))
iI
; then
1
[E
i
] , and
e
i
= (
1
[E
i
]) for every i I, R. (See part (a) of the proof of 327F.) So E : E R
I
,
1
[E] T,
which is a -algebra of subsets of R
I
, contains every E
i
, and therefore includes ; that is,
1
[E] T
for every E . Accordingly we may dene by setting E =
1
[E] for every E , and will be
a measure on R
I
with domain . The Boolean homomorphism E
1
[E] : C has kernel ^

, so
descends to a homomorphism : /^

C, which is measure-preserving. To see that [/^

] = C
0
,
observe that because is the -algebra generated by E
i
: i I, R, [/^

] must be the closed

subalgebra of C generated by E

i
: i I, R = e
i
: i I, R, which is C
0
.
(b) Now suppose that
t
,
t
have the same properties. Consider
/ = E : E , E

=
t
E

,
where I write E

for the equivalence class of E in /^

, and E

for the equivalence class of E in /^

.
Then / is a -subalgebra of , because E E

, E
t
E

are both sequentially order-continuous Boolean

homomorphisms, and contains every E
i
, so must be the whole of . Consequently
E = E

=
t
E

=
t
E
for every E , and
t
= ; it follows at once that
t
= . So and are uniquely determined.
(c) If E and i I,
_
E
x(i)(dx) =
_
x(i)E(x)(dx) =
_
(y)(i)E((y))(dy)
(applying 235I to the inverse-measure-preserving function : Y R
I
)
=
_

1
[E]
f
i
(y)(dy)
(by the denition of )
=
i
(
1
[E])
(by the choice of f
i
)
=
i
E

by the denition of .
(d) Because [[
i
> ]] C
0
, it is equal to [[
i
C
0
> C
0
]], for every R. So
i
must be the standard
extension of
i
C
0
(327F).
(e)-(g) The point is that, because the standard-extension operator is order-preserving (327F(b-ii)),
152 Maharams theorem 333M

i
0
i
C
0
0

_
E
x(i)(dx) 0 for every E
x(i) 0 -a.e.,

i

j

i
C
0

j
C
0

_
E
x(i)(dx)
_
E
x(j)(dx) for every E
x(i) x(j) -a.e.,

i
= 0
i
C
0
= 0

_
E
x(i)(dx) = 0 for every E
x(i) = 0 -a.e..
333N A canonical form for closed subalgebras We now have all the elements required to describe
a canonical form for structures
(A, , C),
where (A, ) is a totally nite measure algebra and C is a closed subalgebra of A. The rst step is the
matching of such structures with structures
(C, ,
n
)
nN
,

)
K
),
where (C, ) is a totally nite measure algebra,
n
)
nN
is a non-increasing sequence of non-negative count-
ably additive functionals on C, K is a countable set of innite cardinals,

)
K
is a family of non-zero
non-negative countably additive functionals on C, and

n=0

n
+

= ; this is the burden of 333K.

Next, given any structure of this second kind, we have a corresponding closed subalgebra C
0
of C, a
measure on R
I
, where I = N K, and an isomorphism from the measure algebra C

0
of to C
0
, all
uniquely dened from the family
i
)
iI
by the process of 333M. For any E belonging to the domain of
, and i I, we have

i
E

=
_
E
x(i)(dx)
(333Mc), so that
i
C
0
is xed by and . Moreover, the functionals
i
can be recovered from their
restrictions to C
0
by the formulae of 327F (333Md). Thus from (C, ,
i
)
iI
) we are led, by a canonical and
reversible process, to the structure
(C, , C
0
, I, , ).
But the extension C of C
0
= [C

0
] can be described, up to isomorphism, by the same process as before;
that is, it corresponds to a sequence
n
)
nN
and a family

)
L
of countably additive functionals on C
0
satisfying the conditions of 333K. We can transfer these to C

0
, where they correspond to families
n
)
nN
,

)
L
of absolutely continuous countably additive functionals dened on , setting

j
E =
j
E

for E , j NL. This process too is reversible; every absolutely continuous countably additive functional
on corresponds to countably additive functionals on C

0
and C
0
. Let me repeat that the results of 327F
mean that the whole structure (C, ,
i
)
iI
) can be recovered from (C
0
, C
0
,
i
C
0
)
iI
) if we can get the
description of (C, ) right, and that the requirements
i
0,
n

n+1
,

,= 0,

iI

i
= imposed in
333K will survive the process (327F(b-iv)).
Putting all this together, a structure (A, , C) leads, in a canonical and (up to isomorphism) reversible
way, to a structure
(K, , L,

)
NL
)
such that
333P Closed subalgebras 153
K and L are countable sets of innite cardinals,
is a totally nite measure on R
I
, where I = N K, and its domain is precisely the
-algebra of subsets of R
I
dened by the coordinate functionals,
for -almost every x R
I
we have x(i) 0 for every i I, x(n) x(n + 1) for every n N
and

iI
x(i) = 1,
for K, x : x() > 0 > 0,
(these two sections corresponding to the requirements
i
0,
n

n+1
,

iI

i
= ,

,= 0 see
333M(e)-(g))
for j J = N L,
j
is a non-negative countably additive functional on ,

n

n+1
for every n N,

,= 0 for every L,

jJ

j
= .
333O Remark I do not envisage quoting the result above very often. Indeed I do not claim that its nal
form adds anything to the constituent results 333K, 327F and 333M. I have taken the trouble to spell it out,
however, because it does not seem to me obvious that the trail is going to end quite as quickly as it does.
We need to use 333K twice, but only twice. The most important use of the ideas expressed here, I suppose,
is in constructing examples to strengthen our intuition for the structures (A, , C) under consideration, and
I hope that you will experiment in this direction.
333P At the risk of trespassing on the province of Chapter 38, I turn now to a special type of closed
subalgebra, in which there is a particularly elegant alternative form for a canonical description. The rst
step is an important result concerning automorphisms of homogeneous probability algebras.
Proposition Let (B, ) be a homogeneous probability algebra. Then there is a measure-preserving auto-
morphism : B B such that
lim
n
(c
n
(b)) = c b
for all b, c B.
proof (a) The case B = 0, 1 is trivial ( is, and must be, the identity map) so we may take it that
B is the measure algebra of 0, 1

with its usual measure

, where is an innite cardinal. Because

#( Z) = max(, ) = , there must be a bijection : such that every orbit of in is
innite (take to correspond to the bijection (, n) (, n + 1) : Z Z). This induces a
bijection

: 0, 1

0, 1

through the formula

(x) = x for every x 0, 1

, and of course

is an
automorphism of the measure space (0, 1

). It therefore induces a corresponding automorphism of

B, setting E

= (

1
[E])

for every E in the domain of

.
(b) Let
0
be the family of subsets E of 0, 1

which are determined by coordinates in nite sets,

that is, are expressible in the form E = x : xJ

E for some nite set J and some

E 0, 1
J
;
equivalently, expressible as a nite union of basic cylinder sets x : xJ = y. Then
0
is a subalgebra of
, so C = E

: E
0
is a subalgebra of B.
(c) Now if b, c C, there is an n N such that (c
m
(b)) = c b for every m n. PPP Express b, c
as E

, F

where E = x : xJ

E, F = x : xK

F and J, K are nite subsets of . For K, all
the
n
() are distinct, so only nitely many of them can belong to J; as K is also nite, there is an n such
that
m
[J] K = for every m n. Fix m n. Then
m
(b) = H

where
H = x : x
m
E = x : x
m
J

E = x : xL

H,
where L =
m
[J] and

H = z
m
: z

E. So (c
m
(b)) = (F H). But L and K are disjoint, because
m n, so F and H must be independent (cf. 272K), and
(c
m
(b)) = F H = F E = c b,
as claimed. QQQ
(d) Now recall that for every E , > 0 there is an E
t

0
such that (EE
t
) (254Fe). So,
given b, c B and > 0, we can nd b
t
, c
t
C such that (b .b
t
) and (c .c
t
) , and in this case
154 Maharams theorem 333P
limsup
n
[ (c
n
(b)) c b[
limsup
n
[ (c
n
(b)) (c
t

n
(b
t
))[
+[ (c
t

n
(b
t
)) c
t
b
t
[ +[ c b c
t
b
t
[
= limsup
n
[ (c
n
(b)) (c
t

n
(b
t
))[ +[ c b c
t
b
t
[
limsup
n
(c .c
t
) + (
n
(b) .
n
(b
t
))
+ c[ b b
t
[ +[ c c
t
[ b
t
(c .c
t
) + (b .b
t
) + c (b .b
t
) + (c .c
t
) b
t
4.
As is arbitrary,
lim
n
(c
n
(b)) = c b,
as required.
Remark Automorphisms of this type are called mixing (see 372P below).
333Q Corollary Let (C,
0
) be a totally nite measure algebra and (B, ) a probability algebra which
is either homogeneous or purely atomic with nitely many atoms all of the same measure. Let (A, ) be
the localizable measure algebra free product of (C,
0
) and (B, ). Then there is a measure-preserving
automorphism : A A such that
a : a A, a = a = c 1 : c C.
Remark I am following 315M in using the notation c b for the intersection in A of the canonical images
of c C and b B. By 325Dc I need not distinguish between the free product C B and its image in A.
proof (a) Let me deal with the case of atomic B rst. In this case, if B has n atoms b
0
, . . . , b
n1
, let
: B B be the measure-preserving homomorphism cyclically permuting these atoms, so that b
0
=
b
1
, . . . , b
n1
= b
0
. Because is an automorphism of (B, ), it induces an automorphism of (A, );
any member of A is uniquely expressible as a = sup
i<n
c
i
b
i
, and now a = sup
i<n
c
i
b
i+1
, if we set
b
n
= b
0
. So a = a i c
i
= c
i+1
for i < n 1 and c
n1
= c
0
, that is, i all the c
i
are the same and
a = sup
i<n
c b
i
= c 1 for some c C.
(b) If B is homogeneous, then take a mixing measure-preserving automorphism : B B as described
in 333P. As in (a), this corresponds to an automorphism of A, dened by saying that (c b) = c (b)
for every c C, b A. Of course (c 1) = c 1 for every c C.
Now suppose that a A and a = a; I need to show that a C
1
= c 1 : c C. Take any > 0. We
know that C B is topologically dense in A (325Dc), so there is an a
t
C B such that (a .a
t
)
2
.
Express a
t
as sup
iI
c
i
b
i
, where c
i
)
iI
is a nite partition of unity in C (315Na). Then
a
t
= sup
iI
c
i
(b
i
),
n
(a
t
) = sup
iI
c
i

n
(b
i
) for every n N.
So we can get a formula for
lim
n
(a
t

n
(a
t
)) = lim
n
(sup
iI
c
i
(b
i

n
(b
i
)))
= lim
n

iI

0
c
i
(b
i

n
(b
i
)) =

iI

0
c
i
( b
i
)
2
.
It follows that

iI

0
c
i
( b
i
)
2
= lim
n
(a
t

n
(a
t
))
limsup
n
(a
n
(a)) (a .a
t
) (
n
(a) .
n
(a
t
))
= a 2 (a .a
t
) a
t
3 (a .a
t
)

iI

0
c
i
b
i
3
2
,
333R Closed subalgebras 155
that is,

iI

0
c
i
b
i
(1 b
i
) 3
2
.
But this means that, setting J = i : i I, b
i
(1 b
i
) , we must have

iJ

0
c
i
3. Set
K = i : i I, b
i
1 2, L = i : i I K, b
i
2, c = sup
iK
c
i
.
Then I (K L) J, so
(a
t
.(c 1)) =

iI\K

0
c
i
b
i
+

iK

0
c
i
(1 b
i
)

iJ

0
c
i
b
i
+

iL

0
c
i
b
i
+

iK

0
c
i
(1 b
i
)

iJ

0
c
i
+ 2

iL

0
c
i
+ 2

iK

0
c
i
3 + 2 = 5,
and
(a .(c 1))
2
+ 5.
As is arbitrary, a belongs to the topological closure of C
1
. But of course C
1
is a closed subalgebra of A
(325Dd), so must actually contain a.
As a is arbitrary, has the required property.
333R Now for the promised special type of closed subalgebra. It will be convenient to have the following
temporary notation. Write Card

for the (proper) class of all non-zero cardinals. For innite Card

, let
(B

) be the measure algebra of the usual measure on 0, 1

. For nite n Card

, let B
n
be the power
set of 0, . . . , n 1 and set
n
b =
1
n
#(b) for b B
n
.
Theorem Let (A, ) be a totally nite measure algebra and C a subset of A. Then the following are
equiveridical:
(i) there is some set G of measure-preserving automorphisms of A such that
C = c : c A, c = c for every G;
(ii) C is a closed subalgebra of A and there is a partition of unity e
i
)
iI
of C, where I is a countable
subset of Card

, such that A is isomorphic to

iI
C
e
i

B
i
, writing C
e
i
for the principal ideal of C generated
by e
i
and endowed with C
e
i
, and C
e
i

B
i
for the localizable measure algebra free product of C
e
i
and B
i
the isomorphism being one which takes any c C to (c e
i
) 1)
iI
, as in 333H and 333K;
(iii) there is a single measure-preserving automorphism of A such that
C = c : c A, c = c.
proof (a)(i)(ii)() C is a subalgebra because every G is a Boolean homomorphism, and it is order-
closed because every is order-continuous (324Kb). (Or, if you prefer, C is topologically closed because
every is continuous.)
() Because C is a closed subalgebra of A, its embedding can be described in terms of families
n
)
nN
,

)
K
as in Theorem 333K. Set I
t
= KN. Recall that each
i
is dened by setting
i
c = (c a
i
), where
a
i
)
iI
is a partition of unity in A (see the proofs of 333H and 333K). Now for K, a

is the maximal
element of A which is relatively Maharam-type-homogeneous over C with relative Maharam type (part
(b) of the proof of 333K). Consequently we must have a

= a

for any measure algebra automorphism of

(A, ) which leaves C invariant; in particular, for every G. Thus a

C for every K.
() Now consider the relatively atomic part of A. The elements a
n
, for n N, are not uniquely
dened. However, the functionals
n
and their supports e
t
n
= [[
n
> 0]] are uniquely dened from the
structure (A, , C) and therefore invariant under G. Observe also that because sup
nN
a
n
= 1 \ sup
K
a

belongs to C, and e
t
n
= infc : c C, c a
n
, while e
t
n
e
t
n+1
for every n, we must have e
t
0
= sup
nN
a
n
.
156 Maharams theorem 333R
Let G

be the set of all those automorphisms of the measure algebra (A, ) such that c = c for every
c C. Then of course G

is a group including G. Now sup

G
a
n
must be invariant under every member
of G

, so belongs to C; it includes a
n
and is included in any member of C including a
n
, so must be e
t
n
.
() I claim now that if n N then e
t
n
[[
0
>
n
]] = 0. PPP??? Otherwise, set c = [[
0
>
n
]] e
t
n
. Then

0
c > 0 so c a
0
,= 0. By the last remark in (), there is a G

such that c a
0
a
n
,= 0. Now there is
a c
t
C such that c a
0
a
n
= c
t
a
0
, and of course we may suppose that c
t
c. But this means that
(c
t
a
n
) = c
t
a
n
c
t
a
0
a
n
= c
t
a
0
,
so that

n
c
t
= (c
t
a
n
) = (c
t
a
n
) (c
t
a
0
) =
0
c
t
,
which is impossible, because 0 ,= c
t
[[
0
>
n
]]. XXXQQQ
So
0
c
n
c whenever c C and c e
t
n
. Because the
k
have been chosen to be a non-increasing
sequence, we must have
0
c =
1
c = . . . =
n
c for every c e
t
n
.
() Recalling now that

iI

i
= C, we see that
0
c
1
n+1
c for every c e
t
n
. It follows that if
e

= inf
nN
e
t
n
,
0
e

= 0; but this must mean that e

= 0. Consequently, setting I = I
t
0, e
n
= e
t
n1
\ e
t
n
for n 1, e

= a

for K, we nd that e
i
)
iI
is a partition of unity in C.
Moreover, for n 1 and c e
n
, we must have
c =

iI

i
c =

k<n

k
c = n
0
c,
so that
k
c =
1
n
c for every k < n. But this means that we have a measure-preserving homomorphism

n
: A
e
n
C
e
n

B
n
given by setting

n
(a
k
c) = c k
whenever c C
e
n
and k < n; this is well-dened because e
n
e
t
k
, so that a
k
c ,= a
k
c
t
if c, c
t
are distinct
members of C
e
n
, and it is measure-preserving because
(a
k
c) =
k
c =
1
n
c = c
n
k
for all relevant k and c. Because B
n
is nite,
n
is surjective.
() Just as in 333H, we now see that because e
i
)
iI
is a partition of unity in A as well as in C, we can
identify A with

iI
A
e
i
and therefore with

iI
C
e
i

B
i
.
(b)(ii)(iii) Let us work in D =

iI
C
e
i

B
i
, writing : A D for the canonical map. For each i I,
we have a measure-preserving automorphism
i
of C
e
i

B
i
with xed-point subalgebra c 1 : c C
e
i

(333Q). For d = d
i
)
iI
D, set
d =
i
d
i
)
iI
.
Then is a measure-preserving automorphism because every
i
is. If d = d, then for every i I there
must be a c
i
e
i
such that d
i
= c
i
1. But this means that d = c, where c = sup
iI
c
i
C. Thus
the xed-point subalgebra of is just [C]. Transferring the structure (D, [C], ) back to A, we obtain a
measure-preserving automorphism
1
of A with xed-point subalgebra C, as required.
(c)(iii)(i) is trivial.
333X Basic exercises (a) Show that, in the proof of 333H, c
i
= upr(a
i
, C) (denition: 314V) for every
i I.
(b) In the context of Lemma 333M, show that we have a one-to-one correspondence between atoms c of
C
0
and points x of non-zero mass in R
I
, given by the formula x

= c.
(c) Let (A, ) be totally nite measure algebra and G a set of measure-preserving Boolean homomorphisms
from A to itself such that G for all , G. (i) Show that a sup
G
a for every a A. (Hint: if
c c, where G and c A, then c = c; apply this to c = sup
G
a.) (ii) Set C = c : c A, c = c
for every G. Show that sup
G
a = upr(a, C) for every a A.
333 Notes Closed subalgebras 157
333Y Further exercises (a) Show that when I = N the algebra of subsets of R
I
, used in 333M, is
precisely the Borel -algebra as described in 271Ya.
(b) Let (A, ) be a totally nite measure algebra, and B, C two closed subalgebras of A with C B.
Show that
C
(B)
C
(A). (Hint: use 333K and the ideas of 332T.)
(c) Let (A, ) be a probability algebra. Show that A is homogeneous i there is a measure-preserving
automorphism of A which is mixing in the sense of 333P.
(d) Let (A, ) be a totally nite measure algebra, and G a set of measure-preserving Boolean homomor-
phisms from A to itself. Set C = c : c A, c = c for every G. Show that C is a closed subalgebra of A
of the type described in 333R. (Hint: in the language of part (a) of the proof of 333R, show that sup
nN
a
n
still belongs to C.)
333 Notes and comments I have done my best, in the rst part of this section, to follow the lines already
laid out in 331-332, using what should (once you have seen them) be natural generalizations of the former
denitions and arguments. Thus the Maharam type (A) of an algebra is just the relative Maharam type

0,1]
(A), and A is Maharam-type-homogeneous i it is relatively Maharam-type-homogeneous over 0, 1.
To help you trace through the correspondence, I list the code numbers: 331Fa333Aa, 331Fb333Ac,
331Hc333Ba, 331Hd333Bd, 332A333Bb, 331I333Cb, 331K333E, 331L333F(ii), 332B333H,
332J333K. 333D overlaps with 332P. Throughout, the principle is the same: everything can be built up
from products and free products.
Theorem 333Ca does not generalize any explicitly stated result, but overlaps with Proposition 332P. In
the proof of 333E I have used a new idea; the same method would of course have worked just as well for
331K, but I thought it worth while to give an example of an alternative technique, which displays a dierent
facet of homogeneous algebras, and a dierent way in which the algebraic, topological and metric properties
of homogeneous algebras interact. The argument of 331K-331L relies (without using the term) on the fact
that measure algebras of Maharam type have topological density at most max(, ) (see 331Ye), while
the the argument of 333E uses the rather more sophisticated concept of stochastic independence.
Corollary 333F(i) is cruder than the more complicated results which follow, but I think that it is invaluable
as a rst step in forming a picture of the possible embeddings of a given (totally nite) measure algebra C
in a larger algebra A. If we think of C as the measure algebra of a measure space (X, , ), then we can be
sure that A is representable as a closed subalgebra of the measure algebra of X 0, 1

for some , that

is, the measure algebra of T where is the product measure on X 0, 1

and T is some -subalgebra

of the domain of ; the embedding of C in A being dened by the formula E

(E 0, 1

for E
(325A, 325D). Identifying, in our imaginations, both X and 0, 1

with the unit interval, we can try to

picture everything in the unit square and these pictures, although necessarily inadequate for algebras of
uncountable Maharam type, already give a great deal of scope for invention.
I said above that everything can be constructed from simple products and free products, judiciously
combined; of course some further ideas must be mixed with these. The dierence between 332B and 333H,
for instance, is partly in the need for the functionals
i
in the latter, whereas in the former the decomposition
involves only principal ideals with the induced measures. Because the
i
are completely additive, they all
have supports c
i
(326Xi) and we get measure algebras (C
c
i
,
i
C
c
i
) to use in the products. (I note that the
c
i
can be obtained directly from the a
i
, without mentioning the functionals
i
, by the process of 333Xa.)
The fact that the c
i
can overlap means that the relatively atomic part of the larger algebra A needs a much
more careful description than before; this is the burden of 333J, and also the principal complication in the
proof of 333R. The relatively atomless part is (comparatively) straightforward, since we can use the same
kind of amalgamation as before (part (c-i) of the proof of 332J, part (b) of the proof of 333K), simplied
because I am no longer seeking to deal with algebras of innite magnitude.
Theorem 333K gives a canonical form for superalgebras of a given totally nite measure algebra (C, ),
taking the structure (C, ) itself for granted. I hope it is clear that while the
i
amd e
i
and the algebra

A =

nN
C
e
n

K
C
e

and the embedding of C in

A are uniquely dened, the rest of the isomorphism
: A

A generally is not. Even when the a

are uniquely dened the isomorphisms between A

a

and
C
e

depend on choosing generating families in the A

a

; see the proof of 333Cb.

158 Maharams theorem 333 Notes
To understand the possible structures (C,
i
)
iI
) of that theorem, we have to go rather deeper. The route
I have chosen is to pick out the subalgebra C
0
of C determined by
i
)
iI
and identify it with the measure
algebra of a particular measure on R
I
. Perhaps I should apologise for not stating explicitly in the course of
Lemma 333M that the measures here are Borel measures (see 333Ya); but I am afraid of opening a door
to an invasion of ideas which belong in Volume 4. Besides, if I were going to do anything more with these
measures than observe that they are uniquely dened by the construction proposed, I would complete them
and call them Radon measures. In order to validate this approach, I must show that the
i
can be recovered
from their restrictions to C
0
; this is 333Md, and is the motive for the discussion of standard extensions in
327. No doubt there are other ways of doing it. One temptation which I felt it right to resist was the idea
of decomposing C into its homogeneous principal ideals; this seemed merely an additional complication. Of
course the subalgebra C
0
has countable Maharam type (being -generated by the elements e
iq
, for i I and
q Q, of 333M), so that its decomposition is relatively simple, being just a matter of picking out the atoms
(333Xb).
In 333P I nd myself presenting an important fact about homogeneous measure algebras, rather out of
context; but I hope that it will help you to believe that I have by no means nished with the insights which
Maharams theorem provides. I give it here for the sake of 333R. For the moment, I invite you to think of
333R as just a demonstration of the power of the techniques I have developed in this chapter, and of the
kind of simplication (in the equivalence of conditions (i) and (iii)) which seems to arise repeatedly in the
theory of measure algebras. But you will see that the rst step to understanding any automorphism will
be a description of its xed-point subalgebra, so 333R will also be basic to the theory of automorphisms of
measure algebras. I note that the hypothesis (i) of 333R can in fact be relaxed (333Yd), but this seems to
need an extra idea.
334 Products
I devote a short section to results on the Maharam classication of the measure algebras of product
measures, or, if you prefer, of the free products of measure algebras. The complete classication, even
for probability algebras, is complex (334Xe, 334Ya), so I content myself with a handful of the most useful
results. I start with upper bounds for the Maharam type of the c.l.d. product of two measure spaces (334A)
and the localizable measure algebra free product of two semi-nite measure algebras (334B), and go on to
the corresponding results for products of probability spaces and algebras (334C-334D). Finally, I show that
any innite power of a probability space is Maharam-type-homogeneous (334E).
334A Theorem Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras A, B. Let be
the c.l.d. product measure on X Y , and (C,

) its measure algebra. Then the Maharam type (C) is at

most max(, (A), (B)).
proof Recall from 325A that we have order-continuous Boolean homomorphisms
1
: A C and
2
: B C
dened by setting
1
E

= (EY )

,
2
F

= (XF)

for E , F T. Let A A, B B be -generating

sets with #(A) = (A), #(B) = (B); set C =
1
[A]
2
[B]. Then C -generates C. PPP Let C
1
be the
order-closed subalgebra of C generated by C. Because
1
is order-continuous,
1
1
[C
1
] is an order-closed
subalgebra of A, and it includes A, so must be the whole of A; thus
1
a C
1
for every a A. Similarly,

2
b C
1
for every b B.
This means that

1
= W : W , W

C
1

contains E F for every E , F T. Also

1
is a -algebra of subsets of X Y , because C
1
is
(sequentially) order-closed in C. So
1

T (denition: 251D). But this means that if W there is

a V
1
such that V W and V = W (251Ib); that is, if c C there is a d C
1
such that d c and

d =

c. Thus C
1
is order-dense in C, and
c = supd : d C
1
, d c C
1
for every c C. So C
1
= C and C -generates C, as claimed. QQQ
334E Products 159
Consequently
(C) #(C) max(, (A), (B)).
334B Corollary Let (A
1
,
1
), (A
2
,
2
) be semi-nite measure algebras, with localizable measure algebra
free product (C,

) (325E). Then (C) max(, (A), (B)).

proof By the construction of part (a) of the proof of 325D, C can be regarded as the measure algebra of
the c.l.d. product of the Stone representations of (A
1
,
1
) and (A
2
,
2
); so the result follows at once from
334A.
334C Theorem Let (X
i
,
i
,
i
))
iI
be a family of probability spaces, with product (X, , ). Let A
i
,
C be the corresponding measure algebras. Then
(C) max(, #(I), sup
iI
(A
i
)).
proof Recall from 325I that we have order-continuous Boolean homomorphisms
i
: A
i
C corresponding
to the inverse-measure-preserving maps x
i
(x) = x(i) : X X
i
. For each i I, let A
i
A
i
be a
set of cardinal (A
i
) which -generates A
i
. Set C =

iI

i
[A
i
]. Then C -generates C. PPP Let C
1
be the
order-closed subalgebra of C generated by C. Because
i
is order-continuous,
1
i
[C
1
] is an order-closed
subalgebra of A
i
, and it includes A
i
, so must be the whole of A
i
; thus
i
a C
1
for every a A
i
, i I.
This means that

1
= W : W , W

C
1

contains
1
i
[E] for every E
i
, i I. Also
1
is a -algebra of subsets of X, because C
1
is (sequentially)
order-closed in C. So
1

iI

i
. But this means that if W there is a V
1
such that V

= W

(254Ff), that is, that C

1
= C, and C -generates C, as claimed. QQQ
Consequently
(C) #(C) max(, #(I), sup
iI
(A
i
)).
334D Corollary Let (A
i
,
i
))
iI
be a family of probability algebras, with probability algebra free
product (C,

). Then
(C) max(, #(I), sup
iI
(A
i
)).
proof See 325J-325K.
334E I come now to the question of when a probability algebra free product is homogeneous. I give
just one result in detail, leaving others to the exercises.
Theorem Let (X, , ) be a probability space, with measure algebra A, and I an innite set; let C be
the measure algebra of the product measure on X
I
. Then C is homogeneous. If (A) = 0 then (C) = 0;
otherwise (C) = max((A), #(I)).
proof (a) As usual, write for the measure of A; let be the measure on X
I
,

the measure on C. If
(A) = 0, that is, A = 0, 1, then C = 0, 1 (by 254Fe, or 325Jc, or otherwise), and in this case is surely
homogeneous, with (C) = 0. So let us suppose hencforth that (A) > 0. We have
(C) max(, #(I), (A)) = max(#(I), (A)),
by 334C.
(b) Fix on b A 0, 1. For each i I, let
i
: A C be the canonical measure-preserving homomor-
phism corresponding to the inverse-measure-preserving function x x(i) : X
I
X. For each n N, there
is a set J I of cardinal n, and now the nite subalgebra of C generated by
i
b : i J has atoms of
measure at most
n
, where = max( b, 1 b) < 1. Consequently C can have no atom of measure greater
than
n
, for any n, and is therefore atomless.
160 Maharams theorem 334E
(c) Because I is innite, there is a bijection between I and I N; that is, there is a partition J
i
)
iI
of
I into countably innite sets. Now (X
I
, ) can be identied with the product of the family (X
J
i
,
i
))
iI
,
where
i
is the product measure on X
J
i
(254N). By (b), every
i
is atomless, so there are sets E
i
X
J
i
of measure
1
2
. The sets E
t
i
= x : xJ
i
E
i
are now stochastically independent in X. Accordingly we
have an inverse-measure-preserving function f : X 0, 1
I
, endowed with its usual measure
I
, dened
by setting f(x)(i) = 1 if x E
t
i
, 0 otherwise, and therefore a measure-preserving Boolean homomorphism
: B
I
C, writing B
I
for the measure algebra of
I
.
Now if c C0 and C
c
is the corresponding ideal, b c b : B
I
C
c
is an order-continuous Boolean
homomorphism. It follows that (C
c
) #(I) (331J).
(d) Again take any non-zero c C. For each i I, set a
i
= infa :
i
a c. Writing A
a
i
for the
corresponding principal ideal of A, we have an order-continuous Boolean homomorphism
t
i
: A
a
i
C
c
,
given by the formula

t
i
a =
i
a c for every a A
a
i
.
Now
t
i
is injective, so is a Boolean isomorphism between A
a
i
and its image
t
i
[A
a
i
], which by 314F(a-i) is
a closed subalgebra of C
c
. So
(A
a
i
) = (
t
i
[A
a
i
]) (C
c
)
by 332Tb.
For any nite J I,
0 <

c

(inf
iJ

i
a
i
) =

iJ

(
i
a
i
) =

iJ
a
i
.
So for any < 1, i : a
i
must be nite, and sup
iI
a
i
= 1. In particular, sup
iI
a
i
= 1 in A. But
this means that if is any cardinal such that the Maharam-type- component e

of A is non-zero, then
e

a
i
,= 0 for some i I, so that
(A
e

a
i
) (A
a
i
) (C
c
).
As is arbitrary, (A) max(, (C
c
)) (332S).
(e) Putting (a)-(d) together, we have
max((A), #(I)) max(, (C
c
)) = (C
c
) (C) max((A), #(I))
for every non-zero c C; so C is homogeneous.
334X Basic exercises (a) Let (X, , ) and (Y, T, ) be complete locally determined measure spaces
with c.l.d. product (XY, , ). Let A, C be the measure algebras of , respectively. Show that if Y > 0
then (A) (C).
(b) Let X be a set and and two totally nite measures on X with the same domain ; then = +
is also a totally nite measure. Write (A, ), (B, ) and (C,

) for the three measure algebras. Show

that (i) there is a surjective order-continuous Boolean homomorphism from C onto A; (ii) C is isomorphic
to a closed subalgebra of the localizable measure algebra free product of A and B; (iii) (A) (C)
max(, (B), (C)).
>>>(c) Let (A
i
,
i
))
iI
be a family of probability algebras, with probability algebra free product (C,

).
Show that (A
i
) (C) for every i, and that
#(i : i I, (A
i
) > 0) (C).
(d) Let X be a set and
n
)
nN
a sequence of totally nite measures on X all with the same domain
. Let
n
)
nN
be a sequence of non-negative real numbers such that

n=0

n
X < , and set E =

n=0

n
E for E . Check that is a measure. Write (A
n
,
n
) for the measure algebra of
n
and
(C,

) for the measure algebra of . Show that (C) max(, sup

nN
(A
n
)).
(e) Let (X, , ) and (Y, T, ) be -nite measure spaces, and the product measure on X Y . Show
that is Maharam-type-homogeneous i one of , is Maharam-type-homogeneous with Maharam type at
least as great as the Maharam type of the other.
334 Notes Products 161
(f ) Show that the product of any family of Maharam-type-homogeneous probability spaces is again
Maharam-type-homogeneous.
>>>(g) Let (X, , ) be a probability space of Maharam type , and I any set of cardinal at least max(, ).
Show that the product measure on X 0, 1
I
is Maharam-type-homogeneous, with Maharam type #(I).
334Y Further exercises (a) Let (X
i
,
i
,
i
))
iI
be an innite family of probability spaces, with
product (X, , ). Let
i
be the Maharam type of
i
for each i; set = max(#(I), sup
iI

i
). Show that
either is Maharam-type-homogeneous, with Maharam type , or there are
t
< , X
t
i

i
such that

iI

i
(X
i
X
t
i
) < and the Maharam type of the subspace measure on X
t
i
is at most
t
for every i I
and either
t
= 0 or #(I) < .
334 Notes and comments The results above are all very natural ones; I have spelt them out partly
for completeness and partly for the sake of an application in 346 below. But note the second alternative
in 334Ya; it is possible, even in an innite product, for a kernel of relatively small Maharam type to be
preserved.
162 Liftings
Chapter 34
The lifting theorem
Whenever we have a surjective homomorphism : P Q, where P and Q are mathematical structures,
we can ask whether there is a right inverse of , a homomorphism : Q P such that is the identity on
Q. As a general rule, we expect a negative answer; those categories in which epimorphisms always have right
inverses (e.g., the category of linear spaces) are rather special, and elsewhere the phenomenon is relatively
rare and almost always important. So it is notable that we have a case of this at the very heart of the
theory of measure algebras: for any complete probability space (X, , ) (in fact, for any complete strictly
localizable space of non-zero measure) the canonical homomorphism from to the measure algebra of has
a right inverse (341K). This is the von Neumann-Maharam lifting theorem. Its proof, together with some
essentially elementary remarks, takes up the whole of of 341.
As a rst application of the theorem (there will be others in Volume 4) I apply it to one of the central
problems of measure theory: under what circumstances will a homomorphism between measure algebras
be representable by a function between measure spaces? Variations on this question are addressed in 343.
For a reasonably large proportion of the measure spaces arising naturally in analysis, homomorphisms are
representable (343B). New diculties arise if we ask for isomorphisms of measure algebras to be representable
by isomorphisms of measure spaces, and here we have to work rather hard for rather narrowly applicable
results; but in the case of Lebesgue measure and its closest relatives, a good deal can be done, as in 344H
and 344I.
Returning to liftings, there are many dicult questions concerning the extent to which liftings can be
required to have special properties, reecting the natural symmetries of the standard measure spaces. For
instance, Lebesgue measure is translation-invariant; if liftings were in any sense canonical, they could be
expected to be automatically translation-invariant in some sense. It seems sure that there is no canonical
lifting for Lebesgue measure all constructions of liftings involve radical use of the axiom of choice but even
so we do have many translation-invariant liftings (345). We have less luck with product spaces; here the
construction of liftings which respect the product structure is fraught with diculties. I give the currently
known results in 346.
341 The lifting theorem
I embark directly on the principal theorem of this chapter (341K, every non-trivial complete strictly
localizable measure space has a lifting), using the minimum of advance preparation. 341A-341B give the
denition of lifting; the main argument is in 341F-341K, using the concept of lower density (341C-341E)
and a theorem on martingales from 275. In 341P I describe an alternative way of thinking about liftings
in terms of the Stone space of the measure algebra.
341A Denition Let (X, , ) be a measure space, and A its measure algebra. By a lifting of A (or
of (X, , ), or of ) I shall mean
either a Boolean homomorphism : A such that (a)

= a for every a A
or a Boolean homomorphism : such that (i) E = whenever E = 0 (ii) (EE) = 0 for
every E .
341B Remarks (a) I trust that the ambiguities permitted by this terminology will not cause any
confusion. The point is that there is a natural one-to-one correspondence between liftings : A and
liftings : given by the formula
E

= E for every E .
PPP (i) Given a lifting : A , the formula denes a Boolean homomorphism : such that
= 0 = , (E .E)

= E

.(E

= 0 E ,
341E The lifting theorem 163
so that is a lifting. (ii) Given a lifting : , the kernel of includes E : E = 0, so there is a
Boolean homomorphism : A such that E

= E for every E (3A2G), and now

(E

= (E)

= E

for every E , so is a lifting. QQQ

I suppose that the word lifting applies most naturally to functions from A to ; but for applications in
measure theory the other type of lifting is used at least equally often.
(b) Note that if : is a lifting then
2
= . PPP For any E ,

2
E .E = (E .E) = . QQQ
If is associated with : A , then a = a for every a A. PPP a = ((a)

) = a. QQQ
341C Denition Let (X, , ) be a measure space, and A its measure algebra. By a lower density
of A (or of (X, , ), or of ) I shall mean
either a function : A such that (i) (a)

= a for every a A (ii) 0 = (iii) (a b) = a b for

all a, b A
or a function : such that (i) E = F whenever E, F and (EF) = 0 (ii) (EE) = 0
for every E (iii) = (iv) (E F) = E F for all E, F .
341D Remarks (a) As in 341B, there is a natural one-to-one correspondence between lower densities
: A and lower densities : given by the formula
E

= E for every E .
(For the requirement E = F whenever E

= F

in A means that every corresponds to a function ,

and the other clauses match each other directly.)
(b) As before, if : is a lower density then
2
= . If is associated with : A , then = .
(c) It will be convenient, in the course of the proofs of 341F-341H below, to have the following concept
available. If (X, , ) is a measure space with measure algebra A, a partial lower density of A is a function
: B such that (i) the domain B of is a subalgebra of A (ii) (b)

= b for every b B (iii) 0 =

(iv) (a b) = a b for all a, b B.
Similarly, if T is a subalgebra of , a function : T is a partial lower density if (i) E = F
whenever E, F T and (EF) = 0 (ii) (EE) = 0 for every E T (iii) = (iv) (EF) = EF
for all E, F T.
(d) Note that lower densities and partial lower densities are order-preserving; if a b in A, and is a
lower density of A, then
a = (a b) = a b b.
(e) Of course a Boolean homomorphism from A to , or from to itself, is a lifting i it is a lower density.
341E Example Let be Lebesgue measure on R
r
, where r 1, and its domain. For E set
E = x : x R
r
, lim
0
(EB(x,))
B(x,)
= 1.
(Here B(x, ) is the closed ball with centre x and radius .) Then is a lower density for ; we may call it
lower Lebesgue density. PPP (You may prefer at rst to suppose that r = 1, so that B(x, ) = [x, x+]
and B(x, ) = 2.) By 261Db (or 223B, for the one-dimensional case) EE is negligible for every E;
in particular, E for every E . If EF is negligible, then (E B(x, )) = (F B(x, )) for
every x and , so E = F. If E F, then (E B(x, )) (F B(x, )) for every x, , so E F;
consequently (E F) E F for all E, F . If E, F and x E F, then
164 Liftings 341E
(E F B(x, )) = (E B(x, )) +(F B(x, )) ((E F) B(x, ))
(E B(x, )) +(F B(x, )) (B(x, ))
for every , so
(EFB(x,))
B(x,)

(EB(x,))
B(x,)
+
(FB(x,))
B(x,)
1 1
as 0, and x (E F). Thus (E F) = E F for all E, F , and is a lower density. QQQ
341F The hard work of this section is in the proof of 341H below. To make it a little more digestible,
I extract two parts of the proof as separate lemmas.
Lemma Let (X, , ) be a probability space and A its measure algebra. Let B be a closed subalgebra of A
and : B a partial lower density. Then for any e A there is a partial lower density
1
, extending ,
dened on the subalgebra B
1
of A generated by B e.
proof (a) Because B is order-closed,
v = upr(e, B) = infa : a B, a e, w = upr(1 \ e, B)
are dened in B (314V). Let E be such that E

= e.
(b) We have a function
1
: B
1
dened by writing

1
((a e) (b \ e)) =
_
((a v) (b \ v)) E
_

_
((a \ w) (b w)) E
_
for a, b B. PPP By 312M, every element of B
1
is expressible as (a e) (b \ e) for some a, b B. If a, a
t
,
b, b
t
B are such that (a e) (b \ e) = (a
t
e) (b
t
\ e), then a e = a
t
e and b \ e = b
t
\ e, that is,
a .a
t
1 \ e w, b .b
t
e v.
This means that e 1 \ (a .a
t
) B and 1 \ e 1 \ (b .b
t
) B. So we also have v 1 \ (a .a
t
) and
w 1 \ (b .b
t
). Accordingly
a v = a
t
v, b w = b
t
w. a \ w = a
t
\ w, b \ v = b
t
\ v.
But this means that
_
((a v) (b \ v)) E
_

_
((a \ w) (b w)) E
_
=
_
((a
t
v) (b
t
\ v)) E
_

_
((a
t
\ w) (b
t
w)) E
_
.
Thus the formula given denes
1
uniquely. QQQ
(c) Now
1
is a lower density.
PPP(i) If a, b B,
(
1
((a e) (b \ e)))

=
__
((a v) (b \ v)) E
_

_
((a \ w) (b w)) E
__

=
_
((a v) (b \ v)) e
_

_
((a \ w) (b w)) \ e
_
= (a e) (b \ e).
So (
1
c)

= c for every c B
1
.
(ii)

1
(0) =
_
((0 v) (0 \ v)) E
_

_
((0 \ w) (0 w)) E
_
= .
(iii) If a, a
t
, b, b
t
B, then
341G The lifting theorem 165

1
(((a e) (b \ e)) ((a
t
e) (b
t
\ e)))
=
1
((a a
t
e) (b b
t
\ e))
=
_
((a a
t
v) (b b
t
\ v)) E
_

_
((a a
t
\ w) (b b
t
w)) E
_
=
_
(((a v) (b \ v)) ((a
t
v) (b
t
\ v))) E
_

_
(((a \ w) (b w)) ((a
t
\ w) (b
t
w))) E
_
=
_
((a v) (b \ v)) ((a
t
v) (b
t
\ v)) E
_

_
((a \ w) (b w)) ((a
t
\ w) (b
t
w)) E
_
=
__
((a v) (b \ v)) E
_

_
((a \ w) (b w)) E
__

__
((a
t
v) (b
t
\ v)) E
_

_
((a
t
\ w) (b
t
w)) E
__
=
1
((a e) (b \ e))
1
((a
t
e) (b
t
\ e)).
So
1
(c c
t
) =
1
(c)
1
(c
t
) for all c, c
t
B
1
. QQQ
(d) If a B, then

1
(a) =
1
((a e) (a \ e))
=
_
((a v) (a \ v)) E
_

_
((a \ w) (a w)) E
_
=
_
(a) E
_

_
(a) E
_
= a.
Thus
1
extends , as required.
341G Lemma Let (X, , ) be a probability space and (A, ) its measure algebra. Suppose we have
a sequence
n
)
nN
of partial lower densities such that, for each n, (i) the domain B
n
of
n
is a closed
subalgebra of A (ii) B
n
B
n+1
and
n+1
extends
n
. Let B be the closed subalgebra of A generated by

nN
B
n
. Then there is a partial lower density , with domain B, extending every
n
.
proof (a) For each n, set

n
= E : E , E

B
n
,
and set

= E : E , E

B.
Then (because all the B
n
, B are -subalgebras of A, and E E

is sequentially order-continuous) all

the
n
,

are -subalgebras of . We need to know that

is just the -algebra

of subsets of X
generated by

nN

n
. PPP Because

is a -algebra including

nN

n
,

. On the other hand,

B

= E

: E

is a -subalgebra of A including B
n
for every n N. Because A is ccc, B

is
(order-)closed (316Fb), so includes B. This means that if E

there must be an F

such that
E

= F

. But now (EF)

= 0 B
0
, so EF
0

, and E also belongs to

. This shows that

and the two algebras are equal. QQQ

(b) For each n N, we have the partial lower density
n
: B
n
. Since (
n
a)

= a B
n
for every
a B
n
,
n
takes all its values in
n
. For n N, let
n
:
n

n
be the lower density corresponding to

n
(341Ba), that is,
n
E =
n
E

for every E
n
.
(c) For a A, n N choose G
a
, g
an
such that G

a
= a and g
an
is a conditional expectation of G
a
on
n
; that is,
_
E
g
an
=
_
E
G
a
= (E G
a
) = (E

a)
for every E
n
. As remarked in 233Db, such a function g
an
can always be found, and moreover we may
take it to be
n
-measurable and dened everywhere on X, by 232He. Now if a B, lim
n
g
an
(x) exists
and is equal to G
a
(x) for almost every x. PPP By Levys martingale theorem (275I), lim
n
g
an
is dened
almost everywhere and is a conditional expectation of G
a
on the -algebra generated by

nN

n
. As
166 Liftings 341G
observed in (a), this is just

; and as G
a
is itself

-measurable, it is also a conditional expectation of

itself on

, and must be equal almost everywhere to lim

n
g
an
. QQQ
(d) For a B, k 1, n N set
H
kn
(a) = x : x X, g
an
(x) 1 2
k

n
,

H
kn
(a) =
n
(H
kn
(a)),
a =

k1

nN

mn

H
km
(a).
The rest of the proof is devoted to showing that : B has the required properties.
(e) G
0
is negligible, so every g
0n
is zero almost everywhere, every H
kn
(0) is negligible and every

H
kn
(0)
is empty; so 0 = .
(f ) If a b in B, then a b. PPP G
a
G
b
is negligible, g
an
g
bn
almost everywhere for every n, every
H
kn
(a) H
kn
(b) is negligible,

H
kn
(a)

H
kn
(b) for every n and k, and a b. QQQ
(g) If a, b B then (a b) = a b. PPP G
ab
G
a
+G
b
1 a.e. so g
ab,n
g
an
+g
bn
1 a.e. for
every n. Accordingly
H
k+1,n
(a) H
k+1,n
(b) H
kn
(a b)
is negligible, and (because
n
is a lower density)

H
kn
(a b)
n
(H
k+1,n
(a) H
k+1,n
(b)) =

H
k+1,n
(a)

H
k+1,n
(b)
for all k 1, n N. Now, if x a b, then, for any k 1, there are n
1
, n
2
N such that
x

mn
1

H
k+1,m
(a), x

mn
2

H
k+1,m
(b).
But this means that
x

mmax(n
1
,n
2
)

H
km
(a b).
As k is arbitrary, x (a b); as x is arbitrary, a b (a b). We know already from (f) that
(a b) a b, so (a b) = a b. QQQ
(h) If a B, then a

= a. PPP g
an
)
nN
G
a
a.e., so setting
V
a
=

k1

nN

mn
H
km
(a) = x : liminf
n
g
an
(x) 1,
V
a
G
a
is negligible, and V

a
= a; but
aV
a

k1,nN
H
kn
(a)

H
kn
(a)
is also negligible, so a

is also equal to a. QQQ Thus is a partial lower density with domain B.

(i) Finally, extends
n
for every n N. PPP If a B
n
, then G
a

m
for every m n, so g
am
= G
a
a.e. for every m n; H
km
(a)G
a
is negligible for k 1, m n;

H
km
=
m
G
a
=
m
a =
n
a
for k 1, m n (this is where I use the hypothesis that
m+1
extends
m
for every m); and
a =

k1
_
rN

mr

H
km
(a)
=

k1
_
rn

mr

H
km
(a) =

k1
_
rn

n
a =
n
a. QQQ
The proof is complete.
341H Now for the rst main theorem.
Theorem Let (X, , ) be any strictly localizable measure space. Then it has a lower density : .
If X > 0 we can take X = X.
341I The lifting theorem 167
proof : Part A I deal rst with the case of probability spaces. Let (X, , ) be a probability space, and
(A, ) its measure algebra.
(a) Set = #(A) and enumerate A as a

)
<
. For let A

be the closed subalgebra of A generated

by a

: < . I seek to dene a lower density : A as the last of a family

, where

: A

is a partial lower density for each . The inductive hypothesis will be that

extends

whenever .
To start the induction, we have A
0
= 0, 1,
0
0 = ,
0
1 = X.
(b) Inductive step to a successor ordinal Given a succesor ordinal , express it as + 1; we are
supposing that

: A

has been dened. Now A

is the subalgebra of A generated by A

(because this is a closed subalgebra, by 323K). So 341F tells us that

can be extended to a partial lower

density

with domain A

.
(c) Inductive step to a non-zero limit ordinal of countable conality In this case, there is a strictly
increasing sequence (n))
nN
with supremum . Applying 341G with B
n
= A
(n)
, we see that there is a
partial lower density

, with domain the closed subalgebra B generated by

nN
A
(n)
, extending every

(n)
. Now A
(n)
A

for every , so B A

; but also, if < , there is an n N such that < (n),

so that a

A
(n)
B; as is arbitrary, A

B and A

= B. Again, if < , there is an n such that

(n), so that
(n)
extends

and

extends

. Thus the induction continues.

(d) Inductive step to a limit ordinal of uncountable conality In this case, A

<
A

. PPP Because
A is ccc, every member a of A

must be in the closed subalgebra of A generated by some countable subset A

of a

: < (331Gd-e). Now A can be expressed as a

: I for some countable I . As I cannot

be conal with , there is a < such that < for every I, so that A A

and a A

. QQQ
But now, because

extends

whenever < , we have a function

: A

dened by writing

a =

a whenever < and a A

. Because the family A

: < is totally ordered and every

is a
partial lower density,

is a partial lower density.

Thus the induction proceeds when is a limit ordinal of uncountable conality.
(e) The induction stops when we reach

: A , which is a lower density such that

1 = X. Setting
E =

, is a lower density such that X = X.

Part B The general case of a strictly localizable measure space follows easily. First, if X = 0, then
A = 0 and we can set 0 = . Second, if is totally nite but not zero, we can replace it by , where
E = E/X for every E ; a lower density for is also a lower density for . Third, if is not totally
nite, let X
i
)
iI
be a decomposition of X (211E). There is surely some j such that X
j
> 0; replacing
X
j
by X
j

X
i
: i I, X
i
= 0, we may assume that X
i
> 0 for every i I. For each i I, let

i
:
i

i
be a lower density of
i
, where
i
= TX
i
and
i
=
i
, such that
i
X
i
= X
i
. Then it is
easy to check that we have a lower density : given by setting
E =

iI

i
(E X
i
)
for every E , and that X = X.
341I The next step is to give a method of moving from lower densities to liftings. I start with an
elementary remark on lower densities on complete measure spaces.
Lemma Let (X, , ) be a complete measure space with measure algebra A.
(a) Suppose that : A is a lower density and
1
: A TX is a function such that
1
0 = ,

1
(a b) =
1
a
1
b for all a, b A and
1
a a for all a A. Then
1
is a lower density. If
1
is a
Boolean homomorphism, it is a lifting.
(b) Suppose that : is a lower density and
1
: TX is a function such that
1
E =
1
F
whenever EF is negligible,
1
= ,
1
(E F) =
1
E
1
F for all E, F and
1
E E for all
E . Then
1
is a lower density. If
1
is a Boolean homomorphism, it is a lifting.
proof (a) All I have to check is that
1
a and (
1
a)

= a for every a A. But

a
1
a, (1 \ a)
1
(1 \ a),
1
a
1
(1 \ a) =
1
0 = .
So
168 Liftings 341I
a
1
a X (1 \ a).
Since
(a)

= a = (X (1 \ a))

,
and is complete,
1
is a lower density. If it is a Boolean homomorphism, then it is also a lifting (341De).
(b) This follows by the same argument, or by looking at the functions from A to dened by and
1
and using (a).
341J Proposition Let (X, , ) be a complete measure space such that X > 0, and A its measure
algebra.
(a) If : A is any lower density, there is a lifting : A such that a a for every a A.
(b) If : is any lower density, there is a lifting : such that E E for every E .
proof (a) For each x 1, set
I
x
= a : a A, x (1 \ a).
Then I
x
is a proper ideal of A. PPP We have
0 I
x
, because x 1,
if b a I
x
then b I
x
, because x (1 \ a) (1 \ b),
if a, b I
x
then a b I
x
, because x (1 \ a) (1 \ b) = (1 \ (a b)),
1 / I
x
because x / = 0. QQQ
For x X1, set I
x
= 0; this is also a proper ideal of A, because A ,= 0. By 311D, there is a surjective
Boolean homomorphism
x
: A 0, 1 such that
x
d = 0 for every d I
x
.
Dene : A TX by setting
a = x : x X,
x
(a) = 1
for every a A. It is easy to check that, because every
x
is a surjective Boolean homomorphism, is a
Boolean homomorphism. Now for any a A, x X,
x a = 1 \ a I
x
=
x
(1 \ a) = 0 =
x
a = 1 = x a.
Thus a a for every a A. By 341I, is a lifting, as required.
(b) Repeat the argument above, or apply it, dening by setting (E

) = E for every E , and

by setting E = (E

) for every E.
341K The Lifting Theorem Every complete strictly localizable measure space of non-zero measure
has a lifting.
proof By 341H, it has a lower density, so by 341J it has a lifting.
341L Remarks If we count 341F-341K as a single argument, it may be the longest proof, after Carlesons
theorem (286), which I have yet presented in this treatise, and perhaps it will be helpful if I suggest ways
of looking at its components.
(a) The rst point is that the theorem should be thought of as one about probability spaces. The shift to
general strictly localizable spaces (Part B of the proof of 341H) is purely a matter of technique. I would not
have presented it if I did not think that its worth doing, for a variety of reasons, but there is no signicant
idea needed, and if for instance the result were valid only for -nite spaces, it would still be one of
the great theorems of mathematics. So the rest of these remarks will be directed to the ideas needed in
probability spaces.
(b) All the proofs I know of the theorem depend in one way or another on an inductive construction. We
do not, of course, need a transnite induction written out in the way I have presented it in 341H above.
Essentially the same proof can be presented as an application of Zorns Lemma; if we take P to be the set
of partial lower densities, then the arguments of 341G and part (A-d) of the proof of 341H can be adapted
341Lf The lifting theorem 169
to prove that any totally ordered subset of P has an upper bound in P, while the argument of 341F shows
that any maximal element of P must have domain A. I think it is purely a matter of taste which form
one prefers. I suppose I have used the ordinal-indexed form largely because that seemed appropriate for
Maharams theorem in the last chapter.
(c) There are then three types of inductive step to examine, corresponding to 341F, 341G and (A-d)
in 341H. The rst and last are easier than the second. Seeking the one-step extension of : B to

1
: B
1
, the natural model to use is the one-step extension of a Boolean homomorphism presented in
312N. The situation here is rather more complicated, as
1
is not fully specied by the value of
1
e, and we
do in fact have more freedom at this point than is entirely welcome. The formula used in the proof of 341F
is derived from Graf & Weizs acker 76.
(d) At this point I must call attention to the way in which the whole proof is dominated by the choice
of closed subalgebras as the domains of our partial liftings. This is what makes the inductive step to a
limit ordinal of countable conality dicult, because A

will ordinarily be larger than

<
A

. But it is
absolutely essential in the one-step extensions.
Because we are dealing with a ccc algebra A, the requirement that the A

should be closed is not a

problem when cf() is uncountable, since in this case

<
A

is already a closed subalgebra; this is the only

idea needed in (A-d) of 341H.
(e) So we are left with the inductive step to when cf() = , which is 341G. Here we actually need
some measure theory, and a particularly striking bit. (You will see that the measure , as opposed to the
algebras and A and the homomorphism E E

and the ideal of negligible sets, is simply not mentioned

anywhere else in the whole argument.)
(i) The central idea is to use the fact that bounded martingales converge to dene a in terms of a
sequence of conditional expectations. Because I have chosen a fairly direct assault on the problem, some of
the surrounding facts are not perhaps so clearly visible as they might have been if I had used a more leisurely
route. For each a A, I start by choosing a representative G
a
; let me emphasize that this is a crude
application of the axiom of choice, and that the dierent sets G
a
are in no way coordinated. (The theorem
we are proving is that they can be coordinated, but we have not reached that point yet.) Next, I choose,
arbitrarily, a conditional expectation g
an
of G
a
on
n
. Once again, the choices are not coordinated; but
the martingale theorem assures us that g
a
= lim
n
g
an
is dened almost everywhere, and is equal almost
everywhere to G
a
if a B. Of course I could have gone to the g
an
without mentioning the G
a
; they are set
up as Radon-Nikod ym derivatives of the countably additive functionals E (E

a) :
n
R. Now the
g
an
, like the G
a
, are not uniquely dened. But they are dened up to a negligible set, so that any alternative
functions g
t
an
would have g
t
an
= g
an
a.e. This means that the sets H
kn
(a) = x : g
an
(x) 1 2
k
are
also dened up to a negligible set, and consequently the sets

H
kn
(a) =
n
(H
kn
(a)) are uniquely dened. I
point this out to show that it is not a complete miracle that we have formulae

H
kn
(a)

H
kn
(b) if a b,

H
kn
(a b)

H
k+1,n
(a)

H
k+1,n
(b) for all a, b A
which do not ask us to turn a blind eye to any negligible sets. I note in passing that I could have dened
the

H
kn
(a) without mentioning the g
an
; in fact

H
kn
(a) =
n
(supc : c B
n
, (a d) 1 2
k
d whenever d B
n
, d c).
(ii) Now, with the sets

H
kn
(a) in hand, we can look at

V
a
=

k1

nN

mn

H
kn
(a);
because g
an
G
a
a.e.,

V
a
G
a
is negligible and

V

a
= a for every a A

. The rest of the argument

amounts to checking that a

V
a
will serve for .
(f ) The arguments above apply to all probability spaces, and show that every probability space has a
lower density. The next step is to convert a lower density into a lifting. It is here that we need to assume
170 Liftings 341Lf
completeness. The point is that we can nd a Boolean homomorphism : A TX such that a a for
every a; this corresponds just to extending the ideals I
x
= a : x (1 \ a) to maximal ideals (and giving
a moments thought to x X 1). In order to ensure that a and (a)

= a, we have to observe that

a is sandwiched between a and X (1 \ a), which dier by a negligible set; so that if is complete all
will be well.
(g) The fact that completeness is needed at only one point in the argument makes it natural to wonder
whether the theorem might be true for probability spaces in general. (I will come later, in 341M, to non-
strictly-localizable spaces.) There is as yet no satisfactory answer to this. For Borel measure on R, the
question is known to be undecidable from the ordinary axioms of set theory (including the axiom of choice,
but not the continuum hypothesis, as usual); I will give some of the arguments in Volume 5. (For the
moment, I refer you to the discussion in Fremlin 89, 4, and to Burke 93.) But I conjecture that there
is an counter-example under the ordinary axioms (see 341Z below).
(h) Quite apart from whether completeness is needed in the argument, it is not absolutely clear why
measure theory is required. The general question of whether a lifting exists can be formulated for any triple
(X, , 1) where X is a set, is a -algebra of subsets of X, and 1 is a -ideal of . (See 341Ya below.)
S.Shelah has given an example of such a triple without a lifting in which two of the basic properties of the
measure-theoretic case are satised: (X, , 1) is complete in the sense that every subset of any member of
1 belongs to (and therefore to 1), and 1 is
1
-saturated in in the sense of 316C (see Shelah Sh636,
Burke n96). But many other cases are known (e.g., 341Yb) in which liftings do exist.
(i) It is of course possible to prove 341K without mentioning lower densities, and there are even some
advantages in doing so. The idea is to follow the lines of 341H, but with liftings instead of lower densities
throughout. The inductive step to a successor ordinal is actually easier, because we have a Boolean homo-
morphism in 341F to extend, and we can use 312N as it stands if we can choose the pair E, F = X E
correctly. The inductive step to an ordinal of uncountable conality remains straightforward. But in the
inductive step to an ordinal of countable conality, we nd that in 341G we get no help from assuming that
the
n
are actually liftings; we are still led to to a lower density . So at this point we have to interpolate
the argument of 341J to convert this lower density into a lifting.
I have chosen the more leisurely exposition, with the extra concept, partly in order to get as far as possible
without assuming completeness of the measure and partly because lower densities are an important tool for
further work (see 345-346).
(j) For more light on the argument of 341G see also 363Xe and 363Yg below.
341M I remarked above that the shift from probability spaces to general strictly localizable spaces
was simply a matter of technique. The question of which spaces have liftings is also primarily a matter
concerning probability spaces, as the next result shows.
Proposition Let (X, , ) be a complete locally determined space with X > 0. Then it has a lifting i it
has a lower density i it is strictly localizable.
proof If (X, , ) is strictly localizable then it has a lifting, by 341K. A lifting is already a lower density,
and if (X, , ) has a lower density it has a lifting, by 341J. So we have only to prove that if it has a lifting
then it is strictly localizable.
Let : A be a lifting, where A is the measure algebra of (X, , ). Let C be a partition of unity in
A consisting of elements of nite measure (322Ea). Set / = c : c C. Because C is disjoint, so is /.
Because supC = 1 in A, every set of positive measure meets some member of / in a set of positive measure.
So the conditions of 213O are satised, and (X, , ) is strictly localizable.
341N Extension of partial liftings The following facts are obvious from the proof of 341H, but it
will be useful to have them out in the open.
341P The lifting theorem 171
Proposition Let (X, , ) be a probability space and T a -subalgebra of .
(a) Any partial lower density
0
: T has an extension to a lower density : .
(b) Suppose now that is complete. If
0
is a Boolean homomorphism, it has an extension to a lifting
for .
proof (a) In Part A of the proof of 341H, let A

be the closed subalgebra of A generated by E

: E
T a

: < , and set

0
E

=
0
E for every E T. Proceed with the induction as before. The only
dierence is that we no longer have a guarantee that X = X.
(b) Suppose now that
0
is a Boolean homomorphism. 341J tells us that there is a lifting :
such that E E for every E . But if E T we must have E
0
E,
E
0
E = E
0
(X E) E (X E) = ,
so that E =
0
E, and extends
0
.
341O Liftings and Stone spaces The arguments of this section so far involve repeated use of the
axiom of choice, and oer no suggestion that any liftings (or lower densities) are in any sense canonical.
There is however one context in which we have a distinguished lifting. Suppose that we have the Stone
space (Z, T, ) of a measure algebra (A, ); as in 311E, I think of Z as being the set of surjective Boolean
homomorphisms from A to Z
2
, so that each a A corresponds to the open-and-closed set a = z : z(a) = 1.
Then we have a lifting : A T dened by setting a = a for each a A. (I am identifying A with the
measure algebra of , as in 321J.) The corresponding lifting : T T is dened by taking E to be that
unique open-and-closed set such that EE is negligible (or, if you prefer, meager).
Generally, liftings can be described in terms of Stone spaces, as follows.
341P Proposition Let (X, , ) be a measure space, (A, ) its measure algebra, and (Z, T, ) the Stone
space of (A, ) with its canonical measure.
(a) There is a one-to-one correspondence between liftings : A and functions f : X Z such that
f
1
[a] and (f
1
[a])

= a for every a A, dened by the formula

a = f
1
[a] for every a A.
(b) If (X, , ) is complete and locally determined, then a function f : X Z satises the conditions of
(a) i () it is inverse-measure-preserving () the homomorphism it induces between the measure algebras
of and is the canonical isomorphism dened by the construction of Z.
proof Recall that T is just the set aM : a A, M Z is meager, and that (aM) = a for all such
a, M; while the canonical isomorphism between A and the measure algebra of is dened by the formula
F

= a whenever F T, a A and Fa is meager

(341K).
(a) If : A is any Boolean homomorphism, then for every x X we have a surjective Boolean
homomorphism f

(x) : A Z
2
dened by saying that f

(x)(a) = 1 if x a, 0 otherwise. f

is a function
from X to Z. We can recover from f

by the formula
a = x : f

(x)(a) = 1 = x : f

(x) a = f
1

[a].
So f
1

[a] and, if is a lifting,

(f
1

[a])

= (a)

= a.
for every a A.
Similarly, given a function f : X Z with this property, then we can set a = f
1
[a] for every a A to
obtain a lifting : A ; and of course we now have
f(x)(a) = 1 f(x) a x a,
so f

= f.
172 Liftings 341P
(b) Assume now that (X, , ) is complete and locally determined.
(i) Let f : X Z be the function associated with a lifting , as in (a). I show rst that f is inverse-
measure-preserving. PPP If F T, express it as aM, where a A and M Z is meager. By 322F, A is
weakly (, )-distributive, so M is nowhere dense (316I). Consider f
1
[M]. If E X is measurable and
of nite measure, then E f
1
[M] has a measurable envelope H (132Ed). ??? If H > 0, then b = H

,= 0
and

b is a non-empty open set in Z. Because M is nowhere dense, there is a non-zero a A such that
a

b M. Now (f
1
[

b]H) = 0, so f
1
[a] H is negligible, and f
1
[a] H is a non-negligible measurable
set disjoint from E f
1
[M] and included in H; which is impossible. XXX Thus H and E f
1
[M] are
negligible. This is true for every measurable set E of nite measure. Because is complete and locally
determined, f
1
[M] and f
1
[M] = 0. So f
1
[F] = f
1
[a]f
1
[M] is measurable, and
f
1
[F] = f
1
[a] = a = a = a = F.
As F is arbitrary, f is inverse-measure-preserving. QQQ
It follows at once that for any F T,
f
1
[F]

= a = F

where a is that element of A such that M = Fa is meager, because in this case f

1
[a]

= a, by (a), while
f
1
[M] is negligible. So is the homomorphism induced by f.
(ii) Now suppose that f : X Z is an inverse-measure-preserving function such that f
1
[F]

= F

for every F T. Then, in particular,

f
1
[a]

= a

= a
for every a A, so that f satises the conditions of (a).
341Q Corollary Let (X, , ) be a strictly localizable measure space, (A, ) its measure algebra, and
Z the Stone space of A; suppose that X > 0. For E write E

for the open-and-closed subset of Z

corresponding to E

A. Then there is a function f : X Z such that Ef

1
[E

] is negligible for every

E . If is complete, then f is inverse-measure-preserving.
proof Let be the completion of , and

its domain. Then we can identify (A, ) with the measure
algebra of (322Da). Let : A

be a lifting, and f : X Z the corresponding function. If E then
E

= a where a = E

, so Ef
1
[E

] = EE

is negligible. If is itself complete, so that

= , then f
is inverse-measure-preserving, by 341Pb.
341X Basic exercises (a) Let (X, , ) be a measure space and : a function. Show that
is a lifting i it is a lower density and E (X E) = X for every E .
>>>(b) Let be the usual measure on X = 0, 1
N
, and its domain. For x X and n N set
U
n
(x) = y : y X, yn = xn. For E set E = x : lim
n
2
n
(E U
n
(x)) = 1. Show that is a
lower density of (X, , ).
>>>(c) Let P be the set of all lower densities of a complete measure space (X, , ), with measure algebra
A, ordered by saying that
t
if a
t
a for every a A. Show that any non-empty totally ordered
subset of P has an upper bound in P. Show that if P and a A and x X (a (1 \ a)), then

t
: A is a lower density, where
t
b = b x if either a b or there is a c A such that x c and
a c b, and
t
b = b otherwise. Hence prove 341J.
(d) Let (X, , ) and (Y, T, ) be measure spaces and suppose that there is an inverse-measure-preserving
function f : X Y such that the associated homomorphism from the measure algebra of to that of
is an isomorphism. Show that for every lifting of (Y, T, ) we have a corresponding lifting of (X, , )
dened uniquely by the formula
(f
1
[F]) = f
1
[F] for every F T.
341 Notes The lifting theorem 173
(e) Let (X, , ) be a measure space, and write L

() for the linear space of all bounded -measurable

functions from X to R. Show that for any lifting : of there is a unique linear operator
T : L

() L

() such that T(E)

= (E) for every E and Tu 0 in L

() whenever u 0 in
L

(). Show that (i) (Tu)

= u and sup
xX
[(Tu)(x)[ = |u|

for every u L

() (ii) T(uv) = TuTv

for all u, v L

().
341Y Further exercises (a) Let X be a set, an algebra of subsets of X and 1 an ideal of ; let
A be the quotient Boolean algebra /1. We say that a function : A is a lifting if it is a Boolean
homomorphism and (a)

= a for every a A, and that : A is a lower density if 0 = ,

(a b) = a b for all a, b A, and (a)

= a for every a A.
Show that if (X, , 1) is complete in the sense that F whenever F E 1, and if X / 1, and
: A is a lower density, then there is a lifting : A such that a a for every a A.
(b) Let X be a non-empty Baire space,

B the -algebra of subsets of X with the Baire property (314Yd)
and / the ideal of meager subsets of X. Show that there is a lifting from

B// to

B such that G

G
for every open G X. (Hint: in 341Ya, set (G

) = G for every regular open set G.)

(c) Let (X, , ) be a Maharam-type-homogeneous probability space with Maharam type . Let
be the Baire -algebra of Y = 0, 1

, that is, the -algebra of subsets of Y generated by the family

x : x() = 1 : < , and let be the restriction to of the usual measure on 0, 1

. Show that there

is an inverse-measure-preserving function f : X Y which induces an isomorphism between their measure
algebras.
(d) Let (X, , ) be a complete Maharam-type-homogeneous probability space with Maharam type ,
and let be the usual measure on 0, 1

. Show that there is an inverse-measure-preserving function

f : X Y which induces an isomorphism between their measure algebras.
(e) Let (X, , ) be a semi-nite measure space which is not purely atomic. Write L
1
strict
for the linear
space of integrable functions f : X R. Show that there is no operator T : L
1
() L
1
strict
such that (i)
(Tu)

= u for every u L
1
() (ii) Tu Tv whenever u v in L
1
(). (Hint: Suppose rst that is the
usual measure on 0, 1
N
. Let F be the countable set of continuous functions f : 0, 1
N
N. Show that
if T satises (i) then there is an x 0, 1
N
such that T(f

)(x) = f(x) for every f F; nd a sequence

f
n
)
nN
in F such that f

n
: n N is bounded above in L
1
() but sup
nN
f
n
(x) = . Now transfer this
argument to some atomless fragment of X.)
341Z Problems (a) Can we construct, using the ordinary axioms of mathematics (including the axiom
of choice, but not the continuum hypothesis), a probability space (X, , ) with no lifting?
(b) Set =
3
. (There is a reason for taking
3
here; see Volume 5, when it appears, or Fremlin 89.)
Let be the Baire -algebra of X = 0, 1

(as in 341Yc), and let be the restriction to of the usual

measure on 0, 1

. Does (X, , ) have a lifting?

341 Notes and comments Innumerable variations of the proof of 341K have been devised, as each author
has struggled with the technical complications. I have discussed the reasons for my own choices in 341L.
The theorem has a curious history. It was originally announced by von Neumann, but he seems never to
have written his proof down, and the rst published proof is that of Maharam 58. That argument is based
on Maharams theorem, 341Xd and 341Yd, which show that it is enough to nd liftings for every 0, 1

;
this requires most of the ideas presented above, but feels more concrete, and some of the details are slightly
simpler. The argument as I have written it owes a great deal to Ionescu Tulcea & Ionescu Tulcea 69.
The lifting theorem and Maharams theorem are the twin pillars of modern abstract measure theory. But
there remains a degree of mystery about the lifting theorem which is absent from the other. The rst point
is that there is nothing canonical about the liftings we can construct, except in the quite exceptional case
of Stone spaces (341O). Even when there is a more or less canonical lower density present (341E, 341Xb),
the conversion of this into a lifting requires arbitrary choices, as in 341J. While we can distinguish some
174 The lifting theorem 341 Notes
liftings as being somewhat more regular than others, I know of no criterion which marks out any particular
lifting of Lebesgue measure, for instance, among the rest. Perhaps associated with this arbitrariness is the
extreme diculty of deciding whether liftings of any given type exist. Neither positive nor negative results
are easily come by (I will present a few in the later sections of this chapter), and the nature of the obstacles
remains quite unclear.
342 Compact measure spaces
The next three sections amount to an extended parenthesis, showing how the Lifting Theorem can be used
to attack one of the fundamental problems of measure theory: the representation of Boolean homomorphisms
between measure algebras by functions between appropriate measure spaces. This section prepares for the
main idea by introducing the class of locally compact measures (342Ad), with the associated concepts of
compact and perfect measures (342Ac, 342K). These depend on the notions of inner regularity (342Aa,
342B) and compact class (342Ab, 342D). I list the basic permanence properties for compact and locally
compact measures (342G-342I) and mention some of the compact measures which we have already seen
(342J). Concerning perfect measures, I content myself with the proof that a locally compact measure is
perfect (342L). I end the section with two examples (342M, 342N).
342A Denitions (a) Let (X, , ) be a measure space. If / TX, I will say that is inner regular
with respect to / if
E = supK : K / , K E
for every E .
Of course is inner regular with respect to / i it is inner regular with respect to / .
(b) A family / of sets is a compact class if

/
t
,= whenever /
t
/ has the nite intersection
property.
Note that any subset of a compact class is again a compact class. (In particular, it is convenient to allow
the empty set as a compact class.)
(c) A measure space (X, , ), or a measure , is compact if is inner regular with respect to some
compact class of subsets of X.
Allowing as a compact class, and interpreting sup as 0 in (a) above, is a compact measure whenever
X = 0.
(d) A measure space (X, , ), or a measure , is locally compact if the subspace measure
E
is compact
whenever E and E < .
Remark I ought to point out that the original denitions of compact class and compact measure (Mar-
czewski 53) correspond to what I call countably compact class and countably compact measure in
Volume 4. For another variation on the concept of compact class see condition () in 343B(ii)-(iii).
For examples of compact measure spaces see 342J.
342B I prepare the ground with some straightforward lemmas.
Lemma Let (X, , ) be a measure space, and / a set such that whenever E and E > 0 there
is a K / such that K E and K > 0. Let E .
(a) There is a countable disjoint set /
1
/ such that K E for every K /
1
and (

/
1
) = E.
(b) If E < then (E

/
1
) = 0.
(c) In any case, there is for any < E a nite disjoint /
0
/ such that K E for every K /
0
and
(

/
0
) .
342F Compact measure spaces 175
proof Set /
t
= K : K /, K E, K > 0. Let /

be a maximal disjoint subfamily of /

t
. If /

is
uncountable, then there is some n N such that K : K /

, K 2
n
is innite, so that there is a
countable /
1
/

such that (

/
1
) = = E.
If /

is countable, set /
1
= /

. Then F =

/
1
is measurable, and F E. Moreover, there is no
member of /
t
disjoint from F; but this means that E F must be negligible. So F = E, and (a) is true.
Now (b) and (c) follow at once, because
(

/
1
) = sup(

/
0
) : /
0
/
1
is nite.
Remark This lemma can be thought of as two more versions of the principle of exhaustion; compare 215A.
342C Corollary Let (X, , ) be a measure space and / TX a family of sets such that () KK
t
/
whenever K, K
t
/ and K K
t
= () whenever E and E > 0, there is a K / such that
K E and K > 0. Then is inner regular with respect to /.
proof Apply 342Bc to / .
342D Lemma Let X be a set and / a family of subsets of X.
(a) The following are equiveridical:
(i) / is a compact class;
(ii) there is a topology T on X such that X is compact and every member of / is a closed set for T.
(b) If / is a compact class, so are the families /
1
= K
0
. . . K
n
: K
0
, . . . , K
n
/ and /
2
=

/
t
:
,= /
t
/.
proof (a)(i)(ii) Let T be the topology generated by X K : K /. Then of course every member of
/ is closed for T. Let T be an ultralter on X. Then /T has the nite intersection property; because /
is a compact class, it has non-empty intersection; take x X

(/ T). The family

G : G X, either G T or x / G
is easily seen to be a topology on X, and contains X K for every K / (because if X K / T then
K T and x K), so includes T; but this just means that every T-open set containing x belongs to T,
that is, that T x. As T is arbitrary, X is compact for T (2A3R).
(ii)(i) Use 3A3Da.
(b) Let T be a topology on X such that X is compact and every member of / is closed for T; then the
same is true of every member of /
1
or /
2
.
342E Corollary Suppose that (X, , ) is a measure space and that / is a compact class such that
whenever E and E > 0 there is a K / such that K E and K > 0. Then is compact.
proof Set /
1
= K
0
. . . K
n
: K
0
, . . . , K
n
/. By 342Db, /
1
is a compact class, and by 342C is
inner regular with respect to /
1
.
342F Corollary A measure space (X, , ) is compact i there is a topology on X such that X is
compact and is inner regular with respect to the closed sets.
proof (a) If is inner regular with respect to a compact class /, then there is a compact topology on X
such that every member of / is closed; now the family T of closed sets includes /, so is also inner regular
with respect to T.
(b) If there is a compact topology on X such that is inner regular with respect to the family / of
closed sets, then this is a compact class, so is a compact measure.
176 The lifting theorem 342G
342G Now I look at the standard questions concerning preservation of the properties of compactness
or local compactness under the usual manipulations.
Proposition (a) Any measurable subspace of a compact measure space is compact.
(b) The completion and c.l.d. version of a compact measure space are compact.
(c) A semi-nite measure space is compact i its completion is compact i its c.l.d. version is compact.
(d) The direct sum of a family of compact measure spaces is compact.
(e) The c.l.d. product of two compact measure spaces is compact.
(f) The product of any family of compact probability spaces is compact.
proof (a) Let (X, , ) be a compact measure space, and E . If / is a compact class such that is
inner regular with respect to /, then /
E
= /TE is a compact class (just because it is a subset of /) and
the subspace measure
E
is inner regular with respect to /
E
.
(b) Let (X, , ) be a compact measure space. Write (X,

, ) for either the completion or the c.l.d.
version of (X, , ). Let / TX be a compact class such that is inner regular with respect to /. Then
is also inner regular with respect to /. PPP If E

and < E there is an E
t
such that E
t
E
and E
t
> ; if is the c.l.d. version of , we may take E
t
to be nite. There is a K / such that
K E
t
and K . Now K = K and K E and K /

. QQQ
(c) Now suppose that (X, , ) is semi-nite; again write (X,

, ) for either its completion or its c.l.d.
version. We already know that if is compact, so is . If is compact, let / TX be a compact class
such that is inner regular with respect to /. Set /

/
t
: ,= /
t
/; then /

is a compact class
(342Db). Now is inner regular with respect to /

. PPP Take E and < E. Choose E

n
)
nN
, K
n
)
nN
as follows. Because is semi-nite, there is an E
0
E such that E
0
and < E
0
< . Given E
n

such that E
n
> , there is a K
n
/

such that K
n
E
n
and K
n
> . Now there is an E
n+1
such
that E
n+1
K
n
and E
n+1
> . Continue. On completing the induction, set K =

nN
K
n
=

nN
E
n
,
so that K /

and K E and K = lim

n
E
n
. As E, are arbitrary, is inner regular with
respect to /

. QQQ As /

is a compact class, is a compact measure.

(d) Let (X
i
,
i
,
i
))
iI
be a family of compact measure spaces, with direct sum (X, , ). We may
suppose that each X
i
is actually a subset of X, with
i
the subspace measure. For each i I let /
i
TX
i
be a compact class such that
i
is inner regular with respect to /
i
. Then / =

iI
/
i
is a compact class,
for if /
t
/ has the nite intersection property, then /
t
/
i
for some i, so has non-empty intersection.
Now if E , E > 0 there is some i I such that
i
(EX
i
) > 0, and we can nd a K /
i

i
/
such that K E X
i
and
i
K > 0, in which case K > 0. By 342E, is compact.
(e) Let (X, , ) and (Y, T, ) be two compact measure spaces, with c.l.d. product measure (XY, , ).
Let T, S be topologies on X, Y respectively such that X and Y are compact spaces and , are inner
regular with respect to the closed sets. Then the product topology on X Y is compact (3A3J).
The point is that is inner regular with respect to the family / of closed subsets of X Y . PPP Suppose
that W and W > . Then there are E , F T such that E < , F < and (W(EF)) >
(251F). Now there are sequences E
n
)
nN
, F
n
)
nN
in , T respectively such that
(E F) W

nN
E
n
F
n
,

n=0
E
n
F
n
< ((E F) W) +((E F) W) = (E F)
(251C). Set
W
t
= (E F)

nN
E
n
F
n
=

nN
((E (F F
n
)) ((E E
n
) F)).
Then W
t
W, and
((E F) W
t
) (

nN
E
n
F
n
)

n=0
E
n
F
n
< (E F) ,
so W
t
> .
Set =
1
4
(W
t
)/(1 + E + F). For each n, we can nd closed measurable sets K
n
, K
t
n
X and
L
n
, L
t
n
Y such that
K
n
E, (E K
n
) 2
n
,
342H Compact measure spaces 177
L
t
n
F F
n
, ((F F
n
) L
t
n
) 2
n
,
K
t
n
E E
n
, ((E E
n
) K
t
n
) 2
n
,
L
n
F, (F L
n
) 2
n
.
Set
V =

nN
(K
n
L
t
n
) (K
t
n
L
n
) W
t
W.
Now
W
t
V
_
nN
((E K
n
) F) (E ((F F
n
) L
t
n
))
(((E E
n
) K
t
n
) F) (E (F L
n
)),
so
(W
t
V )

n=0
(E K
n
) F +E ((F F
n
) L
t
n
)
+((E E
n
) K
t
n
) F +E (F L
n
)

n=0
2
n
(2E + 2F) W
t
,
and V . But V is a countable intersection of nite unions of products of closed measurable sets, so is
itself a closed measurable set, and belongs to / . QQQ
Accordingly the product topology on X Y witnesses that is a compact measure.
(f ) The same method works. In detail: let (X
i
,
i
,
i
))
iI
be a family of compact probability spaces,
with product (X, , ). For each i, let T
i
be a topology on X
i
such that X
i
is compact and
i
is inner regular
with respect to the closed sets. Give X the product topology; this is compact. If W and > 0, let
C
n
)
nN
be a sequence of measurable cylinder sets (in the sense of 254A) such that X W

nN
C
n
and

n=0
C
n
(XW)+. Express each C
n
as

iI
E
ni
where E
ni

i
for each i and J
n
= i : E
ni
,= X
i

is nite. For n N, i J
n
set
ni
= 2
n
/(1 +#(J
n
)). Choose closed measurable sets K
ni
X
i
E
ni
such
that
i
((X
i
E
ni
) K
ni
)
ni
for n N, i J
n
. For each n N, set
V
n
=

iJ
n
x : x X, x(i) K
ni
,
so that V
n
is a closed measurable subset of X. Observe that
X V
n
= x : x(i) X K
ni
for i J
n

includes C
n
, and that
(X (V
n
C
n
))

iJ
n
x : x(i) X
i
(K
ni
E
ni
)

iJ
n

ni
2
n
.
Now set V =

nN
V
n
; then V is again a closed measurable set, and
X V

nN
C
n
(X (C
n
V
n
))
has measure at most

n=0
C
n
+ 2
n
1 W + + 2,
so V W 3. As W and are arbitrary, is inner regular with respect to the closed sets, and is a
compact measure.
342H Proposition (a) A compact measure space is locally compact.
(b) A strictly localizable locally compact measure space is compact.
(c) Let (X, , ) be a measure space. Suppose that whenever E and E > 0 there is an F such
that F E, F > 0 and the subspace measure on F is compact. Then is locally compact.
proof (a) This is immediate from 342Ga and the denition of locally compact measure space.
178 The lifting theorem 342H
(b) Suppose that (X, , ) is a strictly localizable locally compact measure space. Let X
i
)
iI
be a
decomposition of X, and for each i I let
i
be the subspace measure on X
i
. Then
i
is compact. Now
can be identied with the direct sum of the
i
, so itself is compact, by 342Gd.
(c) Write T for the set of measurable sets F X such that the subspace measures
F
are compact.
Take E with E < . By 342Bb, there is a countable disjoint family F
i
)
iI
in T such that F
i
E
for each i, and F
t
= E

iI
F
i
is negligible; now this means that F
t
T (342Ac), so we may take it that
E =

iI
F
i
. In this case
E
is isomorphic to the direct sum of the measures
F
i
and is compact. As E is
arbitrary, is locally compact.
342I Proposition (a) Any measurable subspace of a locally compact measure space is locally compact.
(b) A measure space is locally compact i its completion is locally compact i its c.l.d. version is locally
compact.
(c) The direct sum of a family of locally compact measure spaces is locally compact.
(d) The c.l.d. product of two locally compact measure spaces is locally compact.
proof (a) Trivial: if (X, , ) is locally compact, and E , and F E is a measurable set of nite
measure for the subspace measure on E, then F and F < , so the subspace measure on F is
compact.
(b) Let (X, , ) be a measure space, and write (X,

, ) for either its completion or its c.l.d. version.
(i) Suppose that is locally compact, and that F < . Then there is an E such that E F
and E = F. Let
E
be the subspace measure on E induced by the measure ; then we are assuming that

E
is compact. Let / TE be a compact class such that
E
is inner regular with respect to /. Then, as
in the proof of 342Gb, the subspace measure
F
on F induced by is also inner regular with respect to /,
so
F
is compact; as F is arbitrary, is locally compact.
(ii) Now suppose that is locally compact, and that E < . Then the subspace measure
E
is
compact. But this is just the completion of the subspace measure
E
, so
E
is compact, by 342Gc; as E is
arbitrary, is locally compact.
(c) Put (a) and 342Hc together.
(d) Let (X, , ) and (Y, T, ) be locally compact measure spaces, with product (X Y, , ). If W
and W > 0, there are E , F T such that E < , F < and (W (E F)) > 0. Now the
subspace measure
EF
induced by on E F is just the product of the subspace measures (251P(ii-),
so is compact, and the subspace measure
W(EF)
is therefore again compact, by 342Ga. By 342Hc, this
is enough to show that is locally compact.
342J Examples It is time I listed some examples of compact measure spaces.
(a) Lebesgue measure on R
r
is compact. (Let / be the family of subsets of R
r
which are compact for
the usual topology. By 134Fb, Lebesgue measure is inner regular with respect to /.)
(b) Similarly, any Radon measure on R
r
(256A) is compact.
(c) If (A, ) is any semi-nite measure algebra, the standard measure on its Stone space Z is compact.
(By 322Qa, is inner regular with respect to the family of open-and-closed subsets of Z, which are all
compact for the standard topology of Z, so form a compact class.)
(d) The usual measure on 0, 1
I
is compact, for any set I. (It is obvious that the usual measure on
0, 1 is compact; now use 342Gf.)
Remark (a)-(c) above are special cases of the fact that all Radon measures are compact; I will return to
this in 416.
342K One of the most important properties of (locally) compact measure spaces has been studied
under the following name.
Denition Let (X, , ) be a measure space. Then (X, , ), or , is perfect if whenever f : X R is
measurable, E and E > 0, then there is a compact set K f[E] such that f
1
[K] > 0.
*342N Compact measure spaces 179
342L Theorem A semi-nite locally compact measure space is perfect.
proof Let (X, , ) be a semi-nite locally compact measure space, f : X R a measurable function,
and E a set of non-zero measure. Because is semi-nite, there is an F such that F E and
0 < F < . Now the subspace measure
F
is compact; let T be a topology on F such that F is compact
and
F
is inner regular with respect to the family / of closed sets for T.
Let
q
)
qQ
be a family of strictly positive real numbers such that

qQ

q
<
1
2
F. (For instance, you
could set
q(n)
= 2
n3
F where q(n))
nN
is an enumeration of Q.) For each q Q, set E
q
= x : x
F, f(x) q, E
t
q
= x : x F, f(x) > q, and choose K
q
, K
t
q
/ such that K
q
E
q
, K
t
q
E
t
q
,
(E
q
K
q
)
q
and (E
t
q
K
t
q
)
q
. Then K =

qQ
(K
q
K
t
q
) / , K F and
(F K)

qQ
(E
q
K
q
) +(E
t
q
K
t
q
) < F,
so K > 0.
The point is that fK is continuous. PPP For any q Q, x : x K, f(x) q = K K
q
and
x : x K, f(x) > q = K K
t
q
. If H R is open and x K f
1
[H], take q, q
t
Q such that
f(x) ]q, q
t
] H; then G = K (K
q
K
t
q
) is a relatively open subset of K containing x and included in
f
1
[H]. Thus K f
1
[H] is relatively open in K; as H is arbitrary, f is continuous. QQQ
Accordingly f[K] is a continuous image of a compact set, therefore compact; it is a subset of f[E], and
f
1
[f[K]] K > 0. As f and E are arbitrary, is perfect.
342M I ought to give examples to distinguish between the concepts introduced here, partly on general
principles, but also because it is not obvious that the concept of locally compact measure space is worth
spending time on at all. It is easy to distinguish between perfect and (locally) compact; locally compact
and compact are harder to separate.
Example Let X be an uncountable set and the countable-cocountable measure on X (211R). Then is
perfect but not compact or locally compact.
proof (a) If f : X R is measurable and E X is measurable, with measure greater than 0, set
A = : R, x : x X, f(x) r is negligible. Then A whenever A. Since
X =

nN
x : f(x) n, there is some n such that n / A, in which case A is bounded above by
n. Also there is some m N such that x : f(x) > m is non-negligible, in which case it must be
conegligible, and m A, so A is non-empty. Accordingly = supA is dened in R. Now for any k N,
x : f(x) 2
k
is negligible, so x : f(x) < is negligible. Also, for any k, x : f(x) + 2
k
is
non-negligible, so x : f(x) >
k
2
must be negligible; accordingly, x : f(x) > is negligible. But this
means that x : f(x) = is conegligible and has measure 1. Thus we have a compact set K = such
that f
1
[K] = 1, and must belong to f[E]. As f and E are arbitrary, is perfect.
(b) is not compact. PPP??? Suppose, if possible, that / TX is a compact class such that is inner
regular with respect to /. Then for every x X there is a measurable set K
x
/ such that K
x
X x
and K
x
> 0, that is, K
x
is conegligible. But this means that K
x
: x X must have the nite intersection
property; as it also has empty intersection, / cannot be a compact class. XXXQQQ
(c) Because is totally nite, it cannot be locally compact (342Hb).
*342N Example There is a complete locally determined localizable locally compact measure space
which is not compact.
proof (a) I refer to the example of 216E. In that construction, we have a set I and a family f

)
C
in
X = 0, 1
I
such that for every D C there is an i I such that D = : f

(i) = 1; moreover, #(C) > c.

The -algebra is the family of sets E X such that for every there is a countable set J I such that
x : xJ = f

J is a subset of either E or X E; and for E , E is #( : f

E) if this is nite,
otherwise. Note that any subset of X determined by a countable set of coordinates belongs to .
For each C, let i

I be such that f

(i

) = 1, f

(i

) = 0 for ,= . (In 216E I took I to be TC,

and i

would be .) Set
Y = x : x X, : C, x(i

) = 1 is nite.
180 The lifting theorem *342N
Give Y its subspace measure
Y
with domain
Y
. Then
Y
is complete, locally determined and localizable
(214Id). Note that f

Y for every C.
(b)
Y
is locally compact. PPP Suppose that F
Y
and
Y
F < . If
Y
F = 0 then surely the subspace
measure
F
is compact. Otherwise, we can express F as E Y where E and E =
Y
F. Then
D = : f

E = : f

F is nite. For D set

G
t

= x : x X, x(i

) = 1, x(i

) = 0 for every D ,
/

= K : f

K F G
t

.
Then each /

is a compact class, and members of dierent /

s are disjoint, so / =

D
/

is a compact
class.
Now suppose that H belongs to the subpsace -algebra
F
and
F
H > 0. Then there is a D such
that f

H, so that H G
t

/
F
and
F
(H G
t

) > 0. By 342E, this is enough to show that

F
is
compact. As F is arbitrary,
Y
is locally compact. QQQ
(c)
Y
is not compact. PPP??? Suppose, if possible, that
Y
is inner regular with respect to a compact class
/ TY . For each C set G

= x : x X, x(i

) = 1, so that f

and
Y
(G

Y ) = 1.
There must therefore be a K

/ such that K

Y and
Y
K

= 1 (since
Y
takes no value in ]0, 1[).
Express K

as Y E

, where E

, and let J

I be a countable set such that

E

x : x X, xJ

= f

.
At this point I call on the full strength of 2A1P. There is a set B C, of cardinal greater than c, such
that f

= f

for all , B. But this means that, for any nite set D B, we can dene
x X by setting
x(i) = f

(i) if D, i J

,
= 0 if i I
_
D
J

.
It is easy to check that : C, x(i

) = 1 = D, so that x Y ; but now

x Y

D
E

D
K

.
What this shows is that K

: B has the nite intersection property. It must therefore have

non-empty intersection; say
y

B
K

B
G

.
But now we have a member y of Y such that : y(i

) = 1 B is innite, contrary to the denition of Y .

XXXQQQ
342X Basic exercises >>>(a) Show that a measure space (X, , ) is semi-nite i is inner regular
with respect to E : E < .
(b) Find a proof of 342B based on 215A.
(c) Let (X, , ) be a locally compact semi-nite measure space in which all singleton sets are negligible.
Show that it is atomless.
(d) Let (X, , ) be a measure space, and an indenite-integral measure over (234B). Show that is
compact, or locally compact, if is. (Hint: if / satises the conditions of 342E with respect to , then it
satises them for .)
(e) Let f : R R be any non-decreasing function, and
f
the corresponding Lebesgue-Stieltjes measure.
Show that
f
is compact. (Hint: 256Xg.)
(f ) Let be Lebesgue measure on [0, 1], the countable-cocountable measure on [0, 1], and their c.l.d.
product. Show that is a compact measure. (Hint: let / be the family of sets K A where A [0, 1] is
cocountable and K A is compact.)
342 Notes Compact measure spaces 181
(g) (i) Give an example of a compact probability space (X, , ), a set Y and a function f : X Y such
that the image measure f
1
is not compact. (ii) Give an example of a compact probability space (X, , )
and a -subalgebra T of such that (X, T, T) is not compact. (Hint: 342Xf.)
(h) Let (X, , ) be a perfect measure space, and f : X R a measurable function. Show that the image
measure f
1
is inner regular with respect to the compact subsets of R, so is a compact measure.
(i) Let (X, , ) be a -nite measure space. Show that it is perfect i for every measurable f : X R
there is a Borel set H f[X] such that f
1
[H] is conegligible in X. (Hint: 342Xh for only if, 256C for
if.)
(j) Let (X, , ) be a complete totally nite perfect measure space and f : X R a measurable function.
Show that the image measure f
1
is a Radon measure, and is the only Radon measure on R for which f
is inverse-measure-preserving. (Hint: 256G.)
(k) Suppose that (X, , ) is a perfect measure space. (i) Show that if (Y, T, ) is a measure space,
and f : X Y is a function such that f
1
[F] for every F T and f
1
[F] is -negligible for every
-negligible set F, then (Y, T, ) is perfect. (ii) Show that if T is a -subalgebra of then (X, T, T) is
perfect.
(l) Let (X, , ) be a perfect measure space such that is the -algebra generated by a sequence of sets.
Show that is compact. (Hint: if is generated by E
n
: n N, set f =

n=0
3
n
E
n
and consider
f
1
[K] : K f[X] is compact.)
(m) Let (X, , ) be a semi-nite measure space. Show that is perfect i T is compact for every
countably generated -subalgebra T of .
(n) Show that (i) a measurable subspace of a perfect measure space is perfect (ii) a semi-nite measure
space is perfect i all its totally nite subspaces are perfect (iii) the direct sum of any family of perfect
measure spaces is perfect (iv) the c.l.d. product of two perfect measure spaces is perfect (hint: put 342Xm
and 342Ge together) (v) the product of any family of perfect probability spaces is perfect (vi) a measure
space is perfect i its completion is perfect (vii) the c.l.d. version of a perfect measure space is perfect (viii)
any purely atomic measure space is perfect (ix) an indenite-integral measure over a perfect measure is
perfect.
(o) Let be Lebesgue measure on R, let A be a subset of R, and let
A
be the subspace measure. Show
that
A
is compact i it is perfect i A is Lebesgue measurable. (Hint: if
A
is perfect, consider the image
measure h
1
A
on R, where h(x) = x for x A.)
342Y Further exercises (a) Show that the space (X, , ) of 216E and 342N is a compact measure
space. (Hint: use the usual topology on X = 0, 1
I
.)
(b) Give an example of a compact complete locally determined measure space which is not localizable.
(Hint: in 216D, add a point to each horizontal and vertical section of X, so that all the sections become
compact measure spaces.)
342 Notes and comments The terminology I nd myself using in this section compact, locally
compact, countably compact, perfect is not entirely satisfactory, in that it risks collision with the same
words applied to topological spaces. For the moment, this is not a serious problem; but when in Volume 4
we come to the systematic analysis of spaces which have both topologies and measures present, it will be
necessary to watch our language carefully. Of course there are cases in which a compact class of the sort
discussed here can be taken to be the family of compact sets for some familiar topology, as in 342Ja-342Jd,
but in others this is not so (see 342Xf); and even when we have a familiar compact class, the topology
constructed from it by the method of 342Da need not be one we might expect. (Consider, for instance, the
topology on R for which the closed sets are just the sets which are compact for the usual topology.)
182 The lifting theorem 342 Notes
I suppose that compact and perfect measure spaces look reasonably natural objects to study; they oer
to illuminate one of the basic properties of Radon measures, the fact that (at least for totally nite Radon
measures on Euclidean space) the image measure of a Radon measure under a measurable function is again
Radon (256G, 342Xj). Indeed this was the original impetus for the study of perfect measures (Gnedenko
& Kolmogorov 54, Sazonov 66). It is not obvious that there is any need to examine locally compact
measure spaces, but actually they are the chief purpose of this section, since the main theorem of the next
section is an alternative characterization of semi-nite locally compact measure spaces (343B). Of course you
may feel that the fact that locally compact and compact coincide for strictly localizable spaces (342Hb)
excuses you from troubling about the distinction at rst reading.
As with any new classication of measure spaces, it is worth nding out how the classes of compact and
perfect measure spaces behave with respect to the standard constructions. I run through the basic facts in
342G-342I, 342Xd, 342Xk and 342Xn. We can also look for relationships between the new properties and
those already studied. Here, in fact, there is not much to be said; 342N and 342Yb show that compactness
is largely independent of the classication in 211. However there are interactions with the concept of atom
(342Xc, 342Xn(viii)).
I give examples to show that perfect measure spaces need not be locally compact, and that locally compact
measure spaces need not be compact (342M, 342N). The standard examples of measure spaces which are
not perfect are non-measurable subspaces (342Xo); I will return to these in the next section (343L-343M).
Something which is not important to us at the moment, but is perhaps worth taking note of, is the
following observation. To determine whether a measure space (X, , ) is compact, we need only the
structure (X, , ^), where ^ is the -ideal of negligible sets, since that is all that is referred to in the
criterion of 342E. The same is true of local compactness, by 342Hc, and of perfectness, by the denition in
342K. Compare 342Xd, 342Xk, 342Xn(ix).
Much of the material of this section will be repeated in Volume 4 as part of a more systematic analysis
of inner regularity.
343 Realization of homomorphisms
We are now in a position to make progress in one of the basic questions of abstract measure theory. In
324 I have already described the way in which a function between two measure spaces can give rise to a
homomorphism between their measure algebras. In this section I discuss some conditions under which we
can be sure that a homomorphism can be represented by a function.
The principal theorem of the section is 343B. If a measure space (X, , ) is locally compact, then many
homomorphisms from the measure algebra of to other measure algebras will be representable by functions
into X; moreover, this characterizes locally compact spaces. In general, a homomorphism between measure
algebras can be represented by widely dierent functions (343I, 343J). But in some of the most important
cases (e.g., Lebesgue measure) representing functions are almost uniquely dened; I introduce the concept
of countably separated measure space to describe these (343D-343H).
343A Preliminary remarks It will be helpful to establish some vocabulary and a couple of elementary
facts.
(a) If (X, , ) and (Y, T, ) are measure spaces, with measure algebras A and B, I will say that a
function f : X Y represents a homomorphism : B A if f
1
[F] and (f
1
[F])

= (F

) for
every F T.
(Perhaps I should emphasize here that some homomorphisms are representable in this sense, and some
are not; see 343M below for examples of non-representable homomorphisms.)
(b) If (X, , ) and (Y, T, ) are measure spaces, with measure algebras A and B, f : X Y is a function,
and : B A is a sequentially order-continuous Boolean homomorphism, then
F : F T, f
1
[F] and f
1
[F]

= F

is a -subalgebra of T. (The verication is elementary.)

343B Realization of homomorphisms 183
(c) Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras A and B, and : B A a
Boolean homomorphism which is represented by a function f : X Y . Let (X,

, ), (Y,

T, ) be the
completions of (X, , ), (Y, T, ); then A and B can be identied with the measure algebras of and
(322Da). Now f still represents when regarded as a function from (X,

, ) to (Y,

T, ). PPP If G is
-negligible, there is a negligible F T such that G F; since
f
1
[F]

= F

= 0,
f
1
[F] is -negligible, so f
1
[E] is negligible, therefore belongs to

. If G is any element of

T, there is an
F T such that GF is negligible, so that
f
1
[G] = f
1
[F]f
1
[GF]

,
and
f
1
[G]

= f
1
[F]

= F

= G

. QQQ
(d) In particular, if (X, , ) and (Y, T, ) are measure spaces, and f : X Y is inverse-measure-
preserving, then f is still inverse-measure-preserving with respect to the completed measures and (apply
(c) with : B A the homomorphism induced by f). (See 235Hc.)
343B Theorem Let (X, , ) be a non-empty semi-nite measure space, and (A, ) its measure alge-
bra. Let (Z, , ) be the Stone space of (A, ); for E write E

for the open-and-closed subset of Z

corresponding to the image E

of E in A. Then the following are equiveridical:

(i) (X, , ) is locally compact in the sense of 342Ad.
(ii) There is a family / such that () whenever E and E > 0 there is a K / such that
K E and K > 0 () whenever /
t
/ is such that (

/
0
) > 0 for every non-empty nite set /
0
/
t
,
then

/
t
,= .
(iii) There is a family / such that ()
t
is inner regular with respect to / () whenever /
t
/ is
such that (

/
0
) > 0 for every non-empty nite set /
0
/
t
, then

/
t
,= .
(iv) There is a function f : Z X such that f
1
[E]E

is negligible for every E .

(v) Whenever (Y, T, ) is a complete strictly localizable measure space, with measure algebra B, and
: A B is an order-continuous Boolean homomorphism, then there is a g : Y X representing .
(vi) Whenever (Y, T, ) is a complete strictly localizable measure space, with measure algebra B, and
: A B is an order-continuous measure-preserving Boolean homomorphism, then there is a g : Y X
representing .
proof (a)(i)(ii) Because is semi-nite, there is a partition of unity a
i
)
iI
in A such that a
i
< for
each i. For each i I, let E
i
be such that E

i
= a
i
. Then the subspace measure
E
i
on E
i
is compact;
let /
i
TE
i
be a compact class such that
E
i
is inner regular with respect to /
i
. Set / =

iI
/
i
. If
/
t
/ and (

/
0
) > 0 for every non-empty nite /
0
/, then /
t
/
i
for some i, and surely has the
nite intersection property, so

/
t
,= ; thus /
t
satises () of condition (ii). And if E , E > 0 then
there must be some i I such that E

i
a
i
,= 0, that is, (E E
i
) > 0, in which case there is a K /
i
/
such that K E E
i
and K > 0; so that / satises condition ().
(b)(ii)(iii) Suppose that / witnesses that (ii) is true. If X = 0 then / already witnesses that
(iii) is true, so we need consider only the case X > 0. Set L = K
0
. . . K
n
: K
0
, . . . , K
n
/. Then
L witnesses that (iii) is true. PPP By 342Ba, is inner regular with respect to L. Let L
t
L be such that
(

L
0
) > 0 for every non-empty nite L
0
L
t
. Then
T
0
= A : A X, there is a nite L
0
L
t
such that X

L
0
A is negligible
is a lter on X, so there is an ultralter T on X including T
0
. Note that every conegligible set belongs to
T
0
, so no negligible set can belong to T. Set /
t
= / T; then

/
0
belongs to T, so is not negligible, for
every non-empty nite /
0
/
t
. Accordingly there is some x

/
t
. But any member of L
t
is of the form
L = K
0
. . . K
n
where each K
i
/; because T is an ultralter and L T, there must be some i n
such that K
i
T, in which case x K
i
L. Thus x

L
t
. As L
t
is arbitrary, L satises the condition
(). QQQ
184 The lifting theorem 343B
(c)(iii)(iv) Let / witness that (iii) is true. For any z Z, set /
z
= K : K /, z K

. If
K
0
, . . . , K
n
/
z
, then z

in
K

i
= (

in
K
i
)

, so (

in
K
i
)

,= and (

in
K
i
) > 0. By () of
condition (iii),

/
z
,= ; and even if /
z
= , X

/
z
,= because X is non-empty. So we may choose
f(z) X

/
z
. This denes a function f : Z X. Observe that, for K / and z Z,
z K

= K /
z
= f(z) K = z f
1
[K],
so that K

f
1
[K].
Now take any E . Consider
U
1
=

: K /, K E

f
1
[K] : K /, K E E

f
1
[E],
U
2
=

: K /, K X E (X E)

f
1
[X E] = Z (f
1
[E] E

),
so that f
1
[E]E

Z (U
1
U
2
). Now U
1
and U
2
are open subsets of Z, so M = Z (U
1
U
2
) is closed,
and in fact M is nowhere dense. PPP??? Otherwise, there is a non-zero a A such that the corresponding
open-and-closed set a is included in M, and an F of non-zero measure such that a = F

. At least one
of F E, F E is non-negligible and therefore includes a non-negligible member K of /. But in this case
K

is a non-empty open subset of M which is included in either U

1
or U
2
, which is impossible. XXXQQQ
By the denition of (321J-321K), M is -negligible, so f
1
[E]E

M is negligible, as required.
(d)(iv)(v) Now assume that f : Z X witnesses (iv), and let (Y, T, ) be a complete strictly localiz-
able measure space, with measure algebra B, and : A B an order-continuous Boolean homomorphism.
If Y = 0 then any function from Y to X will represent , so we may suppose that Y > 0. Write W for
the Stone space of B. Then we have a continuous function : W Z such that
1
[a] = a for every
a A (312P), and
1
[M] is nowhere dense in W for every nowhere dense M Z (313R). It follows that

1
[M] is meager for every meager M Z, that is,
1
[M] is negligible in W for every negligible M Z.
By 341Q, there is an inverse-measure-preserving function h : Y W such that h
1
[

b]

= b for every b B.
Consider g = fh : Y X.
If E , set a = E

A, so that E

= a Z, and M = f
1
[E]E

is -negligible; consequently

1
[M] is negligible in W. Because h is inverse-measure-preserving,
g
1
[E]h
1
[
1
[E

]] = h
1
[
1
[f
1
[E]]]h
1
[
1
[E

]] = h
1
[
1
[M]]
is negligible. But
1
[E

] = a, so
g
1
[E]

= h
1
[
1
[E

]]

= a.
As E is arbitrary, g induces the homomorphism .
(e)(v)(vi) is trivial.
(f )(vi)(iv) Assume (vi). Let be the c.l.d. version of , T its domain, and B its measure algebra;
then is strictly localizable (322Qb). The embedding T corresponds to an order-continuous measure-
preserving Boolean homomorphism from A to B (322Db). By (vi), there is a function f : Z X such
that f
1
[E] T and f
1
[E]

= (E

in B for every E . But as and have the same negligible sets

(322Qb), f
1
[E]E

is -negligible for every E , as required by (iv).

(g)(iv)(i)() To begin with (down to the end of () below) I suppose that is totally nite. In this
case we have a function g : X Z such that Eg
1
[E

] is negligible for every E (341Q again). We

are supposing also that there is a function f : Z X such that f
1
[E]E

is negligible for every E .

Write / for the family of sets K E such that K and there is a compact set L Z such that
f[L] K g
1
[L].
() is inner regular with respect to /. PPP Take F and < F. Choose V
n
)
nN
, F
n
)
nN
as
follows. F
0
= F. Given that F
n
> , then
(f
1
[F
n
] F

n
) = F

n
= F
n
> ,
so there is an open-and-closed set V
n
f
1
[F
n
] F

n
with V
n
> . Express V
n
as F

n+1
where F
n+1
;
since F
n
g
1
[F

n
] is negligible, and V
n
F

n
, we may take it that F
n+1
g
1
[F

n
]. Continue.
At the end of the induction, set K =

nN
F
n
and L =

nN
F

n
. Because F
n+1
F
n
g
1
[F

n
] F
n
is negligible for each n, K = lim
n
F
n
, while K F and L is surely compact. We have
343D Realization of homomorphisms 185
L

nN
V
n

nN
f
1
[F
n
] = f
1
[K],
so f[L] K. Also
K

nN
F
n+1

nN
g
1
[F

n
] = g
1
[L].
So K /. As F and are arbitrary, is inner regular with respect to /. QQQ
() Next, / is a compact class. PPP Suppose that /
t
/ has the nite intersection property. If /
t
= ,
of course

/
t
,= ; suppose that /
t
is non-empty. Let L be the family of closed sets L Z such that
g
1
[L] includes some member of /
t
. Then L has the nite intersection property, and Z is compact, so
there is some z

L; also Z L, so z Z. For any K /
t
, there is some closed set L Z such that
f[L] K g
1
[L], so that L L and z L and f(z) K. Thus f(z)

/
t
. As /
t
is arbitrary, / is a
compact class. QQQ
So / witnesses that is a compact measure.
() Now consider the general case. Take any E of nite measure. If E = then surely the subspace
measure
E
is compact. Otherwise, we can identify the measure algebra of
E
with the principal ideal A
E

of A generated by E

(322Ja), and E

Z with the Stone space of A

E
(312S). Take any x
0
E and
dene

f : E

E by setting

f(z) = f(z) if z E

f
1
[E], x
0
if z E

f
1
[E]. Then f and

f agree
almost everywhere on E

, so

f
1
[F]F

is negligible for every F

E
, that is,

f represents the canonical
isomorphism between the measure algebras of
E
and the subspace measure
E
on E

. But this means

that condition (iv) is true of
E
, so
E
is compact, by ()-() above. As E is arbitrary, is locally compact.
This completes the proof.
343C Examples (a) Let be an innite cardinal. We know that the usual measure

on 0, 1

is compact (342Jd). It follows that if (X, , ) is any complete probability space such that the measure
algebra B

of

can be embedded as a subalgebra of the measure algebra A of , there is an inverse-

measure-preserving function from X to 0, 1

. By 332P, this is so i every non-zero principal ideal of A has

Maharam type at least . Of course this does not depend in any way on the results of the present chapter.
If B

can be embedded in A, there must be a stochastically independent family E

)
<
of sets of measure
1
2
; now we get a map h : X 0, 1

by saying that h(x)() = 1 i x E

, which by 254G is inverse-

measure-preserving.
(b) In particular, if is atomless, there is an inverse-measure-preserving function from X to 0, 1
N
; since
this is isomorphic, as measure space, to [0, 1] with Lebesgue measure (254K), there is an inverse-measure-
preserving function from X to [0, 1].
(c) More generally, if (X, , ) is any complete atomless totally nite measure space, there is an inverse-
measure-preserving function from X to the interval [0, X] endowed with Lebesgue measure. (If X > 0,
apply (b) to the normalized measure (X)
1
; or argue directly from 343B, using the fact that Lebesgue
measure on [0, X] is compact; or use the idea suggested in 343Xd.)
(d) Throughout the work above in 254 as well as in 343B I have taken the measures involved to be
complete. It does occasionally happen, in this context, that this restriction is inconvenient. Typical results
not depending on completeness in the domain space X are in 343Xc-343Xd. Of course these depend not
only on the very special nature of the codomain spaces 0, 1
I
or [0, 1], but also on the measures on these
spaces being taken to be incomplete.
343D Uniqueness of realizations The results of 342E-342J, together with 343B, give a respectable
number of contexts in which homomorphisms between measure algebras can be represented by functions
between measure spaces. They say nothing about whether such functions are unique, or whether we can
distinguish, among the possible representations of a homomorphism, any canonical one. In fact the proof of
343B, using the Lifting Theorem as it does, strongly suggests that this is like looking for a canonical lifting,
and I am sure that (outside a handful of very special cases) any such search is vain. Nevertheless, we do
have a weak kind of uniqueness theorem, valid in a useful number of spaces, as follows.
186 The lifting theorem 343D
Denition A measure space (X, , ) is countably separated if there is a countable set / separating
the points of X in the sense that for any distinct x, y X there is an E / containing one but not the
other. (Of course this is a property of the structure (X, ) rather than of (X, , ).)
343E Lemma A measure space (X, , ) is countably separated i there is an injective measurable
function from X to R.
proof If (X, , ) is countably separated, let / be a countable set separating the points of X. Let
E
n
)
nN
be a sequence running over / . Set
f =

n=0
3
n
E
n
: X R.
Then f is measurable (because every E
n
is measurable) and injective (because if x ,= y in X and n = mini :
#(E
i
x, y) = 1 and x E
n
, then
f(x) 3
n
+

i<n
3
i
E
i
(x) >

i>n
3
i
+

i<n
3
i
E
i
(y) f(y).)
On the other hand, if f : X R is measurable and injective, then / = f
1
[ ], q] ] : q Q is a
countable subset of separating the points of X, so (X, , ) is countably separated.
Remark The construction of the function f from the sequence E
n
)
nN
in the proof above is a standard
trick; such f are sometimes called Marczewski functionals.
343F Proposition Let (X, , ) be a countably separated measure space and (Y, T, ) any measure
space. Let f, g : Y X be two functions such that f
1
[E] and g
1
[E] both belong to T, and f
1
[E]g
1
[E]
is -negligible, for every E . Then f = g -almost everywhere, and y : y Y, f(y) ,= g(y) is measurable
as well as negligible.
proof Let / be a countable set separating the points of X. Then
y : f(y) ,= g(y) =

E,
f
1
[E]g
1
[E]
is measurable and negligible.
343G Corollary If, in 343B, (X, , ) is countably separated, then the functions f : Y X of 343B(v)-
(vi) are almost uniquely dened in the sense that if f, g both represent the same homomorphism from A to
B then f = g a.e.
343H Examples Leading examples of countably separated spaces are
(i) R (take / = ], q] : q Q);
(ii) 0, 1
N
(take / = E
n
: n N, where E
n
= x : x(n) = 1);
(iii) subspaces (measurable or not) of countably separated spaces;
(iv) nite products of countably separated spaces;
(v) countable products of countably separated probability spaces;
(vi) completions and c.l.d. versions of countably separated spaces.
As soon as we move away from these elementary ideas, however, some interesting diculties arise.
343I Example Let be the usual measure on X = 0, 1
c
, where c = #(R), and its domain. Then
there is a function f : X X such that f(x) ,= x for every x X, but Ef
1
[E] is negligible for every
E . PPP The set c is still of cardinal c, so there is an injection h : 0, 1

c . (As usual, I am
identifying the cardinal number c with the corresponding initial ordinal. But if you prefer to argue without
the full axiom of choice, you can express all the same ideas with R in the place of c and N in the place of
.) For x X, set
f(x)() = 1 x() if = h(x),
= x() otherwise .
Evidently f(x) ,= x for every x. If E X is measurable, then we can nd a countable set J c and sets E
t
,
E
tt
, both determined by coordinates in J, such that E
t
E E
tt
and E
tt
E
t
is negligible (254Oc). Now for
343K Realization of homomorphisms 187
any particular c, x : h(x) = is negligible, being either empty or of the form x : x(n) = z(n) for
every n < for some z 0, 1

. So H = x : h(x) J is negligible. Now we see that for x X H,

f(x)J = xJ, so for x X (H (E
tt
E
t
)),
x E = x E
t
= f(x) E
t
= f(x) E,
x / E = x / E
tt
= f(x) / E
tt
= f(x) / E.
Thus Ef
1
[E] H (E
tt
E
t
) is negligible. QQQ
343J The split interval I introduce a construction which here will seem essentially elementary, but in
other contexts is of great interest, as will appear in Volume 4.
(a) Take I
|
to consist of two copies of each point of the unit interval, so that I
|
= t
+
: t [0, 1] t

:
t [0, 1]. For A I
|
write A
l
= t : t

A, A
r
= t : t
+
A. Let be the set
E : E I
|
, E
l
and E
r
are Lebesgue measurable and E
l
E
r
is Lebesgue negligible.
For E , set
E =
L
E
l
=
L
E
r
where
L
is Lebesgue measure on [0, 1]. It is easy to check that (I
|
, , ) is a complete probability space.
Also it is compact. PPP Take / to be the family of sets K I
|
such that K
l
= K
r
is a compact subset
of [0, 1], and check that / is a compact class and that is inner regular with respect to /; or use 343Xa
below. QQQ The sets t

: t [0, 1] and t
+
: t [0, 1] are non-measurable subsets of I
|
; on both of them
the subspace measures correspond exactly to
L
. We have a canonical inverse-measure-preserving function
h : I
|
[0, 1] given by setting h(t
+
) = h(t

) = t for every t [0, 1]; h induces an isomorphism between the

measure algebras of and
L
.
I
|
is called the split interval or (especially when given its standard topology, as in 343Yc below) the
double arrow space or two arrows space.
Now the relevance to the present discussion is this: we have a map f : I
|
I
|
given by setting
f(t
+
) = t

, f(t

) = t
+
for every t [0, 1]
such that f(x) ,= x for every x, but Ef
1
[E] is negligible for every E , so that f represents the
identity homomorphism on the measure algebra of . The function h : I
|
[0, 1] is canonical enough,
but is two-to-one, and the canonical map from the measure algebra of to the measure algebra of
L
is
represented equally by the functions t t

and t t
+
, which are nowhere equal.
(b) Consider the direct sum (Y, ) of (I
|
, ) and ([0, 1],
L
); for deniteness, take Y to be (I
|
0)
([0, 1] 1). Setting
h
1
(t
+
, 0) = h
1
(t

, 0) = (t, 1), h
1
(t, 1) = (t
+
, 0),
we see that h
1
: Y Y induces a measure-preserving involution of the measure algebra Bof , corresponding
to its expression as a simple product of the isomorphic measure algebras of and
L
. But h
1
is not invertible,
and indeed there is no invertible function from Y to itself which induces this involution of B. PPP??? Suppose,
if possible, that g : Y Y were such a function. Looking at the sets
E
q
= [0, q] 1, F
q
= (t
+
, 0) : t [0, q] (t

, 0) : t [0, q]
for q Q, we must have g
1
[E
q
]F
q
negligible for every q, so that we must have g(t
+
, 0) = g(t

, 0) = (t, 1)
for almost every t [0, 1], and g cannot be injective. XXXQQQ
(c) Thus even with a compact probability space, and an automorphism of its measure algebra, we
cannot be sure of representing and
1
by functions which will be inverses of each other.
343K 342L has a partial converse.
Proposition If (X, , ) is a semi-nite countably separated measure space, it is compact i it is locally
compact i it is perfect.
188 The lifting theorem 343K
proof We already know that compact measure spaces are locally compact and locally compact semi-nite
measure spaces are perfect (342Ha, 342L). So suppose that (X, , ) is a perfect semi-nite countably
separated measure space. Let f : X R be an injective measurable function (343E). Consider
/ = f
1
[L] : L f[X], L is compact in R.
The denition of perfect measure space states exactly that whenever E and E > 0 there is a K /
such that K E and K > 0. And / is a compact class. PPP If /
t
/ has the nite intersection
property, L = f[K] : K /
t
is a family of compact sets in R with the nite intersection property, and
has non-empty intersection; so that

/
t
is also non-empty, because f is injective. QQQ By 342E, (X, , ) is
compact.
343L The time has come to give examples of spaces which are not locally compact, so that we can expect
to have measure-preserving homomorphisms not representable by inverse-measure-preserving functions. The
most commonly arising ones are covered by the following result.
Proposition Let (X, , ) be a complete locally determined countably separated measure space, and A X
a set such that the subspace measure
A
is perfect. Then A is measurable.
proof ??? Otherwise, there is a set E such that E < and B = A E / . Let f : X R be an
injective measurable function (343E again). Then fB is
B
-measurable, where
B
is the domain of the
subspace measure
B
on B. Set
/ = f
1
[L] : L f[B], L is compact in R.
Just as in the proof of 343K, / is a compact class and
B
is inner regular with respect to /. By 342Bb,
there is a sequence K
n
)
nN
in / such that
B
(B

nN
K
n
) = 0. But of course / , because f is
-measurable, so

nN
K
n
. Because is complete, B

nN
K
n
and B . XXX
343M Example 343L tells us that any non-measurable set X of R
r
, or of 0, 1
N
, with their usual
measures, is not perfect, therefore not (locally) compact, when given its subspace measure.
To nd a non-representable homomorphism, we do not need to go through the whole apparatus of 343B.
Take Y to be a measurable envelope of X (132Ed). Then the identity function from X to Y induces an
isomorphism of their measure algebras. But there is no function from Y to X inducing the same isomorphism.
PPP??? Writing Z for R
r
or 0, 1
N
and for its measure, Z is countably separated; suppose E
n
)
nN
is a
sequence of measurable sets in Z separating its points. For each n, (Y E
n
)

in the measure algebra of

Y
corresponds to (X E
n
)

in the measure algebra of

X
. So if f : Y X were a function representing
the isomorphism of the measure algebras, (Y E
n
)f
1
[E
n
] would have to be negligible for each n, and
A =

nN
(Y E
n
)f
1
[E
n
] would be negligible. But for y Y A, f(y) belongs to just the same E
n
as
y does, so must be equal to y. Accordingly X Y A and X is measurable. XXXQQQ
343X Basic exercises (a) Let (X, , ) be a semi-nite measure space. (i) Suppose that there is a
set A X, of full outer measure, such that the subspace measure on A is compact. Show that is locally
compact. (Hint: show that satises (ii) or (v) of 343B.) (ii) Suppose that for every non-negligible E
there is a non-negligible set A E such that the subspace measure on A is compact. Show that is locally
compact.
(b) Let X
i
)
iI
be a family of non-empty sets, with product X; write
i
: X X
i
for the coordinate
map. Suppose we are given a -algebra
i
of subsets of X
i
for each i; let =

iI

i
be the corresponding
-algebra of subsets of X generated by
1
i
[E] : i I, E
i
. Let be a totally nite measure with
domain , and for i I let
i
be the image measure
1
i
. Check that the domain of
i
is
i
. Show that if
every (X
i
,
i
,
i
) is compact, then so is (X, , ). (Hint: either show that satises (v) of 343B or adapt
the method of 342Gf.)
(c) Let I be any set. Let T be the -algebra of subsets of 0, 1
I
generated by the sets F
i
= z : z(i) = 1
for i I, and any probability measure with domain T; let B be the measure algebra of . Let (X, , )
be a measure space with measure algebra A, and : B A an order-continuous Boolean homomorphism.
Show that there is an inverse-measure-preserving function f : X 0, 1
I
representing . (Hint: for each
i I, take E
i
such that E

i
= F

i
; set f(x)(i) = 1 if x E
i
, and use 343Ab.)
343Yc Realization of homomorphisms 189
(d) Let (X, , ) be an atomless probability space. Let
B
be the restriction of Lebesgue measure to
the -algebra of Borel subsets of [0, 1]. Show that there is a function g : X [0, 1] which is inverse-
measure-preserving for and
B
. (Hint: nd an f : X 0, 1
N
as in 343Xc, and set g = hf where
h(z) =

n=0
2
n1
g(n), as in 254K; or choose E
q
such that E
q
= q, E
q
E
q
whenever q q
t
in
[0, 1] Q, and set f(x) = infq : x E
q
for x E
1
.)
(e) Let (X, , ) be a countably separated measure space, with measure algebra A. (i) Show that x
for every x X. (ii) Show that every atom of A is of the form x

for some x X.
(f ) Let I
|
be the split interval, with its usual measure described in 343J, and h : I
|
[0, 1] the
canonical surjection. Show that the canonical isomorphism between the measure algebras of and Lebesgue
measure on [0, 1] is given by the formula E

h[E]

for every measurable E I

|
.
(g) Let (X, , ) and (Y, T, ) be measure spaces with measure algebras (A, ), (B, ). Suppose that
X Y = and that we have a measure-preserving isomorphism : A B. Set
= W : W X Y, W X , W Y T, (W X)

= (W Y )

,
and for W set W = (W X) = (W Y ). Show that (X Y, , ) is a measure space which is
locally compact, or perfect, if (X, , ) is.
(h) Let (X, , ) be a complete perfect totally nite measure space, (Y, T, ) a complete countably
separated measure space, and f : X Y an inverse-measure-preserving function. Show that T = F : F
Y, f
1
[F] , so that a function h : Y R is -integrable i hf is -integrable. (Hint: if A Y and
E = f
1
[A] , fE is inverse-measure-preserving for the subspace measures
E
,
A
; by 342Xk,
A
is
perfect, so by 343L A T. Now use 235L.)
343Y Further exercises (a) Let (X, , ) be a semi-nite measure space, and suppose that there is
a compact class / TX such that () whenever E and E > 0 there is a non-negligible K / such
that K E () whenever K
0
, . . . , K
n
/ and

in
K
i
= then there are measurable sets E
0
, . . . , E
n
such that E
i
K
i
for every i and

in
E
i
is negligible. Show that is locally compact.
(b) (i) Show that a countably separated semi-nite measure space has magnitude at most c and Maharam
type at most 2
c
. (ii) Show that the direct sum of c or fewer countably separated measure spaces is countably
separated.
(c) Let I
|
= t
+
: t [0, 1] t

: t [0, 1] be the split interval (343J). (i) Show that the rules
s

s
+
t
+
s t, s
+
t

s < t,
t

t
+
for all t [0, 1]
dene a Dedekind complete total order on I
|
with greatest and least elements. (ii) Show that the intervals
[0

, t

], [t
+
, 1
+
], interpreted for this ordering, generate a compact Hausdor topology on I
|
for which the
map h : I
|
[0, 1] of 343J is continuous. (iii) Show that a subset E of I
|
is Borel for this topology i the
sets E
r
, E
l
[0, 1], as described in 343J, are Borel and E
r
E
l
is countable. (iv) Show that if f : [0, 1] R
is of bounded variation then there is a continuous g : I
|
R such that g = fh except perhaps at countably
many points. (v) Show that the measure of 343J is inner regular with respect to the compact subsets of
I
|
. (vi) Show that we have a lower density for dened by setting
E = t

: 0 < t 1, lim
0
1

(E [(t )
+
, t

]) = 1
t
+
: 0 t < 1, lim
0
1

(E [t
+
, (t +)

]) = 1
for measurable sets E I
|
.
190 The lifting theorem 343Yd
(d) Set X = 0, 1
c
, with its usual measure . Show that there is an inverse-measure-preserving function
f : X X such that f[X] is non-measurable but f induces the identity automorphism of the measure
algebra of . (Hint: use the idea of 343I.) Show that under these conditions f[X], with its subspace
measure, must be compact. (Hint: use 343B(iv).)
(e) Let
Hr
be r-dimensional Hausdor measure on R
s
, where s 1 is an integer and r 0 (264). (i)
Show that
Hr
is countably separated. (ii) Show that the c.l.d. version of
Hr
is compact. (Hint: 264Yi.)
(f ) Give an example of a countably separated probability space (X, , ) and a function f from X to a
set Y such that the image measure f
1
is not countably separated. (Hint: use 223B to show that if E R
is Lebesgue measurable and not negligible, then E +Q is conegligible; or use the zero-one law to show that
if E TN is measurable and not negligible for the usual measure on TN, then ab : a E, b [N]
<
is
conegligible.)
343 Notes and comments The points at which the Lifting Theorem impinges on the work of this section
are in the proofs of (iv)(i) and (iv)(v) in Theorem 343B. In fact the ideas can be rearranged to give a
proof of 343B which does not rely on the Lifting Theorem; I give a hint in Volume 4 (413Yc).
I suppose the signicant new ideas of this section are in 343B and 343K. The rest is mostly a matter
of being thorough and careful. But I take this material at a slow pace because there are some potentially
confusing features, and the underlying question is of the greatest importance: when, given a Boolean
homomorphism from one measure algebra to another, can we be sure of representing it by a measurable
function between measure spaces? The concept of compact space puts the burden rmly on the measure
space corresponding to the domain of the Boolean homomorphism, which will be the codomain of the
measurable function. So the rst step is to try to understand properly which spaces are compact, and what
other properties they can be expected to have; which accounts for much of the length of 342. But having
understood that many of our favourite spaces are compact, we have to come to terms with the fact that
we still cannot count on a measure algebra isomorphism corresponding to a measure space isomorphism.
I introduce the split interval (343J, 343Xf, 343Yc) as a close approximation to Lebesgue measure on [0, 1]
which is not isomorphic to it. Of course we have already seen a more dramatic example: the Stone space
of the Lebesgue measure algebra also has the same measure algebra as Lebesgue measure, while being in
almost every other way very much more complex, as will appear in Volumes 4 and 5.
As 343C suggests, elementary cases in which 343B can be applied are often amenable to more primitive
methods, avoiding not only the concept of compact measure, but also Stone spaces and the Lifting Theorem.
For substantial examples in which we can prove that a measure space (X, ) is compact, without simulta-
neously nding direct constructions for inverse-measure-preserving functions into X (as in 343Xc-343Xd), I
think we shall have to wait until Volume 4.
The concept of countably separated measure space does not involve the measure at all, nor even the
ideal of negligible sets; it belongs to the theory of -algebras of sets. Some simple permanence properties
are in 343H and 343Yb(ii). Let us note in passing that 343Xh describes some more situations in which the
image measure catastrophe, described in 235J, cannot arise.
I include the variants 343B(ii), 343B(iii) and 343Ya of the notion of local compactness because they are
not obvious and may illuminate it.
344 Realization of automorphisms
In 343Jb, I gave an example of a good (compact, complete) probability space X with an automorphism
of its measure algebra such that both and
1
are representable by functions from X to itself, but there
is no such representation in which the two functions are inverses of each other. The present section is an
attempt to describe the further renements necessary to ensure that automorphisms of measure algebras can
be represented by automorphisms of the measure spaces. It turns out that in the most important contexts
in which this can be done, a little extra work yields a signicant generalization: the simultaneous realization
of countably many homomorphisms by a consistent family of functions.
344B Realization of automorphisms 191
I will describe three cases in which such simultaneous realizations can be achieved: Stone spaces (344A),
perfect complete countably separated spaces (344C) and suitable measures on 0, 1
I
(344E-344G). The
arguments for 344C, suitably rened, give a complete description of perfect complete countably separated
strictly localizable spaces which are not purely atomic (344I, 344Xc). At the same time we nd that Lebesgue
measure, and the usual measure on 0, 1
I
, are homogeneous in the strong sense that two measurable
subspaces (of non-zero measure) are isomorphic i they have the same measure (344J, 344L).
344A Stone spaces The rst case is immediate from the work of 312, 313 and 321, as collected
in 324E. If (Z, , ) is actually the Stone space of a measure algebra (A, ), then every order-continuous
Boolean homomorphism : A A corresponds to a unique continuous function f

: Z Z (312P)
which represents (324E). The uniqueness of f

means that we can be sure that f

= f

for all order-

continuous homomorphisms and ; and of course f

is the identity map on Z, so that f

1 will have to
be f
1

whenever is invertible. Thus in this special case we can consistently, and canonically, represent all
order-continuous Boolean homomorphisms from A to itself.
Now for two cases where we have to work for the results.
344B Theorem Let (X, , ) be a countably separated measure space with measure algebra A, and
G a countable semigroup of Boolean homomorphisms from A to itself such that every member of G can be
represented by some function from X to itself. Then a family f

)
G
of such representatives can be chosen
in such a way that f

= f

for all , G; and if the identity automorphism belongs to G, then we

may arrange that f

is the identity function on X.

proof (a) Because G satises the same conditions as G, we may suppose from the beginning that
belongs to G itself. Let / be a countable set separating the points of X. For each G take some
representing function g

: X X; take g

to be the identity function. If , G, then of course

((g

)
1
[E])

= (g
1

[g
1

[E]])

= (g
1

[E])

= E

= (g
1

[E])

for every E . By 343F, the set

H

= x : g

(x) ,= g

(x)
is negligible and belongs to .
(b) Set
H =

,G
H

;
because G is countable, H is also measurable and negligible. Try dening f

: X X by setting f

(x) =
g

(x) if x X H, f

(x) = x if x H. Because H is measurable, f

1

[E] for every E ; because H

is negligible,
(f
1

[E])

= (g
1

[E])

= E

for every E , and f

represents , for every G. Of course f

= g

is the identity function on X.

(c) If G then f
1

[H] = H. PPP (i) If x H then f

(x) = x H. (ii) If f

(x) H and f

(x) = x then
of course x H. (iii) If f

(x) = g

(x) H then there are , G such that g

(x) ,= g

(x). So
either
g

(x) ,= g

(x),
or
g

(x) ,= g

(x)
or
g

(x) ,= g

(x),
and in any case x H. QQQ
192 The lifting theorem 344B
(d) It follows that f

= f

for every , G. PPP (i) If x H then

f

(x) = x = f

(x).
(ii) If x X H then f

(x) / H, by (c), so
f

(x) = g

(x) = g

(x) = f

(x). QQQ
344C Corollary Let (X, , ) be a countably separated perfect complete strictly localizable measure
space with measure algebra A, and G a countable semigroup of order-continuous Boolean homomorphisms
from A to itself. Then we can choose simultaneously, for each G, a function f

: X X representing
, in such a way that f

= f

for all , G; and if the identity automorphism belongs to G, then

we may arrange that f

is the identity function on X. In particular, if G is invertible, and

1
G, we
shall have f

1 = f
1

; so that if moreover and

1
are measure-preserving, f

will be an automorphism
of the measure space (X, , ).
proof By 343K, (X, , ) is compact. So 343B(v) tells us that every member of G is representable, and we
can apply 344B.
Reminder: Spaces satisfying the conditions of this corollary include Lebesgue measure on R
r
, the usual
measure on 0, 1
N
, and their measurable subspaces; see also 342J, 342Xe, 343H and 343Ye.
344D The third case I wish to present requires a more elaborate argument. I start with a kind of
Schroder-Bernstein theorem for measurable spaces.
Lemma Let X and Y be sets, and TX, T TY -algebras. Suppose that there are f : X Y ,
g : Y X such that F = f[X] T, E = g[Y ] , f is an isomorphism between (X, ) and (F, T
F
) and
g is an isomorphism between (Y, T) and (E,
E
), writing
E
, T
F
for the subspace -algebras (see 121A).
Then (X, ) and (Y, T) are isomorphic, and there is an isomorphism h : X Y which is covered by f and
g in the sense that
(x, h(x)) : x X (x, f(x)) : x X (g(y), y) : y Y .
proof Set X
0
= X, Y
0
= Y , X
n+1
= g[Y
n
] and Y
n+1
= f[X
n
] for each n N; then X
n
)
nN
is a non-
increasing sequence in and Y
n
)
nN
is a non-increasing sequence in T. Set X

nN
X
n
, Y

nN
Y
n
.
Then fX
2k
X
2k+1
is an isomorphism between X
2k
X
2k+1
and Y
2k+1
Y
2k+2
, while gY
2k
Y
2k+1
is an
isomorphism between Y
2k
Y
2k+1
and X
2k+1
X
2k+2
; and gY

is an isomorphism between Y

and X

.
So the formula
h(x) = f(x) if x
_
kN
X
2k
X
2k+1
,
= g
1
(x) for other x X
gives the required isomorphism between X and Y .
Remark You will recognise the ordinary Schroder-Bernstein theorem (2A1G) as the case = TX, T = TY .
344E Theorem Let I be any set, and let be a -nite measure on X = 0, 1
I
with domain the
-algebra B generated by the sets x : x(i) = 1 as i runs over I; write A for the measure algebra of .
Let G be a countable semigroup of order-continuous Boolean homomorphisms from A to itself. Then we
can choose simultaneously, for each G, a function f

: X X representing , in such a way that

f

= f

for all , G; and if the identity automorphism belongs to G, then we may arrange that f

is the identity function on X. In particular, if G is invertible and

1
G, we shall have f

1 = f
1

;
so that if moreover is measure-preserving, f

will be an automorphism of the measure space (X, B, ).

proof (a) As in 344C, we may as well suppose from the beginning that G. The case of nite I is trivial,
so I will suppose that I is innite. For i I, set E
i
= x : x(i) = 1; for J I, let B
J
be the -subalgebra
of B generated by E
i
: I J. For i I, G choose F
i
B such that F

i
= E

i
. Let be the family
of those subsets J of I such that F
i
B
J
for every i J, G.
344E Realization of automorphisms 193
(b) For the purposes of this proof, I will say that a pair (J, g

)
G
) is consistent if J and, for
each G, g

is a function from X to itself such that

g
1

[E
i
] B
J
and (g
1

[E
i
])

= E

i
whenever i J, G,
g
1

[E
i
] = E
i
whenever i I J, G,
g

= g

whenever , G,
g

(x) = x for every x X.

Now the key to the proof is the following fact: if (J, g

)
G
) is consistent, and

J is a member of such
that

J J is countably innite, then there is a family g

)
G
such that (

J, g

)
G
) is consistent and
g
1

[E
i
] = g
1

[E
i
] whenever i J, G. The construction is as follows.
(i) Start by xing on any innite set K

J J such that (

J J) K is also innite. For z 0, 1
K
,
set V
z
= x : x X, xK = z; then V
z
B

J
. All the sets V
z
, as z runs over the uncountable set 0, 1
K
,
are disjoint, so they cannot all have non-zero measure (because is -nite), and we can choose z such that
V
z
is -negligible.
(ii) Dene h

: X X, for G, by setting
h

(x)(i) = g

(x)(i) if i J,
= x(i) if i I

J,
= x(i) if i

J J and x V
z
,
= 1 if i

J J and x F
i
V
z
,
= 0 if i

J J and x / F
i
V
z
.
Because V
z
B

J
and V
z
= 0, we see that
() h
1

[E
i
] = g
1

[E
i
] B
J
if i J,
() h
1

[E
i
] B

J
and h
1

[E
i
]F
i
is negligible if i

J J,
and consequently
() (h
1

[E
i
])

= E

i
for every i

J,
() (h
1

[E])

= E

for every E B

J
(by 343Ab); moreover,
() h
1

[E] = g
1

[E] for every E B

J
,
() h
1

[E] B

J
for every E B

J
,
() h
1

[E
i
] = E
i
if i I

J,
so that
() h
1

[E] B for every E B;

nally
() h

(x) = x for every x X.

(iii) The next step is to note that if , G then
H
,
= x : x X, h

(x) ,= h

(x)
belongs to B

J
and is negligible. PPP
H
,
=

iI
h
1

[h
1

[E
i
]]h
1

[E
i
].
Now if i J, then h
1

[E
i
] = g
1

[E
i
] B
J
, so
h
1

[h
1

[E
i
]] = h
1

[g
1

[E
i
]] = g
1

[g
1

[E
i
]] = g
1

[E
i
] = h
1

[E
i
].
Next, for i I

J,
h
1

[h
1

[E
i
]] = h
1

[E
i
] = E
i
= h
1

[E
i
].
So
194 The lifting theorem 344E
H
,
=

J\J
h
1

[h
1

[E
i
]]h
1

[E
i
].
But for any particular i

J J, E
i
and h
1

[E
i
] belong to B

J
, so
(h
1

[h
1

[E
i
]])

= (h
1

[E
i
])

= E

i
= (h
1

[E
i
])

,
and h
1

[h
1

[E
i
]]h
1

[E
i
] is a negligible set, which by (ii-) belongs to B

J
. So H
,
is a countable union
of sets of measure 0 in B

J
and is itself a negligible member of B

J
, as claimed. QQQ
(iv) Set
H =

,G
H
,

G
h
1

[V
z
].
Then H B

J
and H = 0. PPP We know that every H
,
is negligible and belongs to B

J
((iii) above), that
every h
1

[V
z
] belongs to B

J
(by (ii-), and that (h
1

[V
z
])

= V

z
= 0, so that h
1

[V
z
] is negligible, for every
G (by (ii-)). Consequently H is negligible and belongs to B

J
. QQQ Also, of course, V
z
= h
1

[V
z
] H.
Next, h

(x) / H whenever x X H and G. PPP If , G then

h

(x) = h

(x) = h

(x) = h

(x),
h

(x) = h

(x) / V
z
because
x / H
,
H
,
H
,
H
,
h
1

[V
z
];
thus h

(x) / H
,
h
1

[V
z
]; as and are arbitrary, h

(x) / H. QQQ
(v) The next fact we need is that there is a bijection q : X H such that () for E H, E B

J
i
q
1
[E] B

J
() q(x)(i) = x(i) for every i I (

J J), x X. PPP Fix any bijection r :

J J

J (J K).
Consider the maps p
1
: X H, p
2
: H X given by
p
1
(x)(i) = x(r
1
(i)) if i

J (J K),
= z(i) if i K,
= x(i) if i X (

J J),
p
2
(y) = y
for x X, y H. Then p
1
is actually an isomorphism between (X, B

J
) and (V
z
, B

J
TV
z
). So p
1
, p
2
are
isomorphisms between (X, B

J
), (H, B

J
TH) and measurable subspaces of H, X respectively. By 344D,
there is an isomorphism q between X and H such that, for every x X, either q(x) = p
1
(x) or p
2
(q(x)) = x.
Since p
1
(x)I (

J J) = xI (

J J) for every x X, and p
2
(y)I (

J J) = yI (

J J) for every y H,
q(x)I (

J J) = xI (

J J) for every x X. QQQ
(vi) An incidental fact which will be used below is the following: if i

J, G then g
1

[E
i
] belongs
to B

J
, because it belongs to B
J
if i J, and otherwise is equal to E
i
. Consequently g
1

[E] B

J
for every
E B

J
.
(vii) I am at last ready to give a formula for g

. For G set
g

(x) = h

(x) if x X H,
= qg

q
1
(x) if x H.
Now (

J, g

)
G
) is consistent. PPP
() If i

J, G,
g
1

[E
i
] = (h
1

[E
i
] H) q[g
1

[q
1
[E
i
H]]]

B
J
because H B

J
and h
1

[E], q
1
[H E], g
1

[E] and q[E] all belong to B

J
for every E B

J
. At the same
time, because g

agrees with h

on the conegligible set X H,

( g
1

[E
i
])

= (h
1

[E
i
])

= E

i
.
344E Realization of automorphisms 195
() If i I

J, G, x X then
g

(x)(i) = h

(x)(i) = q(x)(i) = x(i),

and if x H then q
1
(x)(i) is also equal to x(i); so g

(x)(i) = x(i). But this means that g

1

[E
i
] = E
i
.
() If , G and x X H, then
g

(x) = h

(x) X H
by (iv) above. So
g

(x) = h

(x) = h

(x) = g

(x)
because x / H
,
. While if x H, then
g

(x) = qg

q
1
(x) H,
so
g

(x) = qg

q
1
qg

q
1
(x) = qg

q
1
(x) = qg

q
1
(x) = g

(x).
Thus g

= g

.
() Because g

(x) = h

(x) = x for every x, g

(x) = x for every x. QQQ

(viii) Finally, if i J and G, q
1
[E
i
] = E
i
, so that q[E
i
H] = E
i
. Accordingly q(x)J = xJ for
every x X, while q
1
(x)J = xJ for x H. So g

q
1
(x)J = g

(x)J for x H, and

g

(x)(i) = h

(x)(i) = g

(x)(i) if x X H,
= qg

q
1
(x)(i) = g

q
1
(x)(i) = g

(x)(i) if x X H.
Thus (

J, g

)
G
) satises all the required conditions.
(c) The remaining idea we need is the following: there is a non-decreasing family J

in , for some
cardinal , such that J
+1
J

is countably innite for every < , J

=

<
J

for every limit ordinal

< , and J

= I. PPP Recall that I am already supposing that I is innite. If I is countable, set = 1,

J
0
= , J
1
= I. Otherwise, set = #(I) and let i

)
<
be an enumeration of I. For i I, G let K
i
I
be a countable set such that F
i
B
K
i
. Choose the J

inductively, as follows. The inductive hypothesis

must include the requirement that #(J

) max(, #()) for every . Start by setting J

0
= . Given <
and J

with #(J

) max(, #()) < , take an innite set L J

and set J
+1
= J

nN
L
n
,
where
L
0
= L i

,
L
n+1
=

iL
n
,G
K
i
for n N, so that every L
n
is countable,
F
i
B
L
n+1
whenever i L
n
, G
and J
+1
; since L J
+1
J

nN
L
n
, J
+1
J

is countably innite, and

#(J
+1
) = max(, #(J

)) max(, #()) = max(, #( + 1)).

For non-zero limit ordinals < , set J

<
J

; then
#(J

) max(, #(), sup

<
#(J

)) max(, #()).
Thus the induction proceeds. Observing that the construction puts i

into J
+1
for every , we see that J

will be the whole of I, as required. QQQ

(d) Now put (b) and (c) together, as follows. Take J

from (c). Set f

0
(x) = x for every G,
x X; then, because J
0
= , (J
0
, f
0
)
G
) is consistent in the sense of (b). Given that (J

, f

)
G
) is
consistent, where < , use the construction of (b) to nd a family f
,+1
)
G
such that (J
+1
, f
,+1
)
G
)
is consistent and f
,+1
(x)(i) = f

(x)(i) for every i J

. At a non-zero limit ordinal , set

f

(x)(i) = f

(x)(i) if x X, < , i J

,
= x(i) if i I J

.
196 The lifting theorem 344E
(The inductive hypothesis includes the requirement that f

(x)J

= f

(x)J

whenever G, x X
and < .) To see that (J

, f

)
G
) is consistent, the only non-trivial point to check is that
f
,
f
,
= f
,
for all , G. But if i J

there is some < such that i J

, and in this case

f
1
,
[E
i
] = f
1
,
[E
i
] B
J

is determined by coordinates in J

, so that (because f
,
(x)J

= f
,
(x)J

for every x)
f
1
,
[f
1
,
[E
i
]] = f
1
,
[f
1
,
[E
i
]] = f
1
,
[E
i
] = f
1
,
[E
i
];
while if i I J

then
f
1
,
[E
i
] = E
i
= f
1
,
[E
i
] = f
1
,
[E
i
] = f
1
,
[f
1
,
[E
i
]].
Thus
f
1
,
[f
1
,
[E
i
]] = f
1
,
[E
i
]
for every i, and f
,
f
,
= f
,
.
On completing the induction, set f

= f

for every G; it is easy to see that f

)
G
satises the
conditions of the theorem.
344F Corollary Let I be any set, and let be a -nite measure on X = 0, 1
I
. Suppose that is the
completion of its restriction to the -algebra B generated by the sets x : x(i) = 1 as i runs over I. Write A
for the measure algebra of . Let G be a countable semigroup of order-continuous Boolean homomorphisms
from A to itself. Then we can choose simultaneously, for each G, a function f

: X X representing
, in such a way that f

= f

for all , G; and if the identity automorphism belongs to G, then

we may arrange that f

is the identity function on X. In particular, if G is invertible and

1
G,
we shall have f

1 = f
1

; so that if moreover is measure-preserving, f

will be an automorphism of the

measure space (X, , ).
proof Apply 344E to B; of course A is canonically isomorphic to the measure algebra of B (322Da).
The functions f

provided by 344E still represent the homomorphisms when re-interpreted as functions

on the completed measure space (0, 1
I
, ), by 343Ac.
344G Corollary Let I be any set, the usual measure on 0, 1
I
, and A its measure algebra. Then any
measure-preserving automorphism of A is representable by a measure space automorphism of (0, 1
I
, ).
344H Lemma Let (X, , ) be a locally compact semi-nite measure space which is not purely atomic.
Then there is a negligible subset of X of cardinal c.
proof Let E be a set of non-zero nite measure not including any atom. Let / TE be a compact class
such that the subspace measure
E
is inner regular with respect to /. Set S =

nN
0, 1
n
, and choose
K
z
)
zS
inductively , as follows. K

is to be any non-negligible member of / included in E. Given that

K
z
E and K
z
> 0, where z 0, 1
n
, take F
z
, F
t
z
K
z
to be disjoint non-negligible measurable sets
both of measure at most 3
n
; such exist because is semi-nite and E does not include any atom. Choose
K
z

0
F
z
, K
z

1
F
t
z
to be non-negligible members of / .
For each w 0, 1
N
, K
wn
)
nN
is a decreasing sequence of members of / all of non-zero measure, so
has non-empty intersection; choose a point x
w


nN
K
wn
. Since K
z

0
K
z

1
= for every z S, all
the x
w
are distinct, and A = x
w
: w 0, 1
N
has cardinal c. Also
A

z0,1]
n
K
z
which has measure at most 2
n
3
(n1)
for every n 1, so

A = 0 and A is negligible.
Remark I see that in this proof I have slipped into a notation which is a touch more sophisticated than
what I have used so far. See 3A1H for a note on the interpretations of 0, 1
n
, 0, 1
N
which make sense
of the formulae here.
I ought to note also that the lemma is valid for all perfect spaces; see 344Yf.
344K Realization of automorphisms 197
344I Theorem Let (X, , ) and (Y, T, ) be atomless, perfect, complete, strictly localizable, countably
separated measure spaces of the same non-zero magnitude. Then they are isomorphic.
proof (a) The point is that the measure algebra (A, ) of has Maharam type . PPP Let E
n
)
nN
be a
sequence in separating the points of X. Let
0
be the -subalgebra of generated by E
n
: n N,
and A
0
the order-closed subalgebra of A generated by E

n
: n N; then E

A
0
for every E
0
, and
(X,
0
,
0
) is countably separated. Let f : X R be
0
-measurable and injective (343E). Of course f
is also -measurable. If a A 0, express a as E

where E . Because (X, , ) is perfect, there is a

compact K R such that K f[E] and f
1
[K] > 0. K is surely a Borel set, so f
1
[K]
0
and
b = f
1
[K]

A
0
0.
But because f is injective, we also have f
1
[K] E and b a. As a is arbitrary, A
0
is order-dense in A;
but A
0
is order-closed, so must be the whole of A. Thus A is -generated by the countable set E

n
: n N,
and (A) . QQQ
On the other hand, because A is atomless, and not 0, none of its principal ideals can have nite
Maharam type, and it is Maharam-type-homogeneous, with type .
(b) Writing (B, ) for the measure algebra of , we see that the argument of (a) applies equally to
(B, ), so that (A, ) and (B, ) are atomless localizable measure algebras, of Maharam type and the
same magnitude. Consequently they are isomorphic as measure algebras, by 332J. Let : A B be a
measure-preserving isomorphism.
By 343B, there are functions g : Y X and f : X Y representing and
1
, because and
are compact and complete and strictly localizable. Now fg : Y Y and gf : X X represent the
identity automorphisms on B, A, so by 343F are equal almost everywhere to the identity functions on Y ,
X respectively. Set
E = x : x X, gf(x) = x, F = y : y Y, fg(y) = y;
then both E and F are conegligible. Of course f[E] F (since fgf(x) = f(x) for every x E), and similarly
g[F] E; consequently fE, gF are the two halves of a one-to-one correspondence between E and F.
Because is measure-preserving, f
1
[H] = H, g
1
[G] = G for every G , H T; accordingly fE
is an isomorphism between the subspace measures on E and F.
(c) By 344H, applied to the subspace measure on E, there is a negligible set A E of cardinal c. Now
X and Y , being countably separated, both have cardinal at most c. (There are injective functions from X
and Y to R.) Set
B = A (X E), C = f[A] (Y F).
Then B and C are negligible subsets of X, Y respectively, and both have cardinal c precisely, so there is a
bijection h : B C. Set
f
1
(x) = f(x) if x X B = E A,
= h(x) if x B.
Then, because and are complete, f
1
is an isomorphism between the measure spaces (X, , ) and
(Y, T, ), as required.
344J Corollary Suppose that E, F are two Lebesgue measurable subsets of R
r
of the same non-zero
measure. Then the subspace measures on E and F are isomorphic.
344K Corollary (a) A measure space is isomorphic to Lebesgue measure on [0, 1] i it is an atomless
countably separated compact (or perfect) complete probability space; in this case it is also isomorphic to
the usual measure on 0, 1
N
.
(b) A measure space is isomorphic to Lebesgue measure on R i it is an atomless countably separated
compact (or perfect) -nite measure space which is not totally nite; in this case it is also isomorphic to
Lebesgue measure on any Euclidean space R
r
.
(c) Let be Lebesgue measure on R. If 0 < E < and we set F =
1
E
F for every measurable
F E, then (E, ) is isomorphic to Lebesgue measure on [0, 1].
198 The lifting theorem 344L
344L The homogeneity property of Lebesgue measure described in 344J is repeated in 0, 1
I
for any
I.
Theorem Let I be any set, and the usual measure on 0, 1
I
. If E, F 0, 1
I
are two measurable sets
of the same non-zero nite measure, the subspace measures on E and F are isomorphic.
proof Write X = 0, 1
I
.
(a) If I is nite, then X, E and F are all nite, and #(E) = #(F), so the result is trivial. If I is
countably innite, then the subspace measures are perfect and complete and countably separated, so the
result follows from 344I. So let us suppose that I is uncountable.
(b) Let (A, ) be the measure algebra of . Then A is homogeneous, and E

= F

, so there is
an automorphism : A A such that E

= F

. (Apply 333D to the nite subalgebra C generated

by E

, F

; or argue directly from 331I.) By 344F, there is a measure space isomorphism f : X X

representing , so that f
1
[E]F and f[F]E are negligible.
(c) We can nd a countably innite set J I and a measurable set E
t
such that E
t
is determined by
coordinates in J, E
t
E f[F] and E E
t
is negligible; so that E
t
is non-empty. Take any x
0
E
t
and set V = x : x X, xJ = x
0
J; then V is a negligible subset of E and #(V ) = #(X) (because
#(I J) = #(I) and x xI J : V 0, 1
I\J
is a bijection). Setting E
1
= E f[F], F
1
= F f
1
[E],
fE
1
is a bijection between E
1
and F
1
. Now
#(V (E E
1
)) = #(X) = #(f
1
[V ] (F F
1
)),
so there is a bijection h : f
1
[V ] (F F
1
) V (E E
1
). Dene g : E F by writing
g(x) = f(x) if x F
1
f
1
[V ],
= h(x) if x f
1
[V ] (F F
1
).
Then g is a bijection, and because g = f almost everywhere on F, g
1
= f
1
almost everywhere on E, g is
an isomorphism for the subspace measures.
344X Basic exercises (a) Let (X, , ) and (Y, T, ) be measure spaces, and suppose that there are
E , F T such that (X, , ) is isomorphic to the subspace (F, T
F
,
F
), while (Y, T, ) is isomorphic to
(E,
E
,
E
). Show that (X, , ) and (Y, T, ) are isomorphic.
(b) Let (X, , ) and (Y, T, ) be perfect countably separated complete strictly localizable measure spaces
with isomorphic measure algebras. Show that there are conegligible subsets X
t
X, Y
t
Y such that X
t
and Y
t
, with the subspace measures, are isomorphic.
(c) Let (X, , ) and (Y, T, ) be perfect countably separated complete strictly localizable measure spaces
with isomorphic measure algebras. Suppose that they are not purely atomic. Show that they are isomorphic.
(d) Give an example of two perfect countably separated complete probability spaces, with isomorphic
measure algebras, which are not isomorphic.
(e) Let (Z, , ) be the Stone space of a Maharam-type-homogeneous measure algebra. Show that if E,
F have the same non-zero nite measure, then the subspace measures on E and F are isomorphic.
(f ) Let (I
|
, , ) be the split interval with its usual measure (343J), and A its measure algebra. (i) Show
that every automorphism of A is represented by a measure space automorphism of I
|
. (ii) Show that if E,
F and E = F > 0 then the subspace measures on E and F are isomorphic.
(g) Let I be an innite set, and the usual measure on 0, 1
I
. Show that if E 0, 1
I
is any set of
non-zero measure, then the subspace measure on E is isomorphic to a multiple of .
344 Notes Realization of automorphisms 199
344Y Further exercises (a) Let X be a set, a -algebra of subsets of X, 1 a -ideal of , and
A the quotient /1. Suppose that there is a countable set / separating the points of X. Let G be
a countable semigroup of Boolean homomorphisms from A to itself such that every member of G can be
represented by some function from X to itself. Show that a family f

)
G
of such representatives can be
chosen in such a way that f

= f

for all , G; and if the identity automorphism belongs to G,

then we may arrange that f

is the identity function on X.

(b) Let A, B be Dedekind -complete Boolean algebras. Suppose that each is isomorphic to a principal
ideal of the other. Show that they are isomorphic.
(c) Let I be an innite set, and write B for the -algebra of subsets of X = 0, 1
I
generated by the
sets x : x(i) = 1 as i runs over I. Let and be -nite measures on X, both with domain B, and
with measure algebras (A, ), (B, ). Show that any Boolean isomorphism : A B is represented by a
bijection f : X X such that f
1
represents
1
: B A, and hence that (A, ) is isomorphic to (B, )
i (X, B, ) is isomorphic to (X, B, ).
(d) Let I be any set, and write B for the -algebra of subsets of X = 0, 1
I
generated by the sets
x : x(i) = 1 as i runs over I. Let 1 be an
1
-saturated ideal of B, and write A for the quotient Boolean
algebra B/1. Let G be a countable semigroup of order-continuous Boolean homomorphisms from A to itself.
Show that we can choose simultaneously, for each G, a function f

: X X representing , in such
a way that f

= f

for all , G; and if the identity automorphism belongs to G, then we may

arrange that f

is the identity function on X. In particular, if G is invertible and

1
G, f

will be
an automorphism of the structure (X, B, 1).
(e) Let I be any set, and write B for the -algebra of subsets of X = 0, 1
I
generated by the sets
x : x(i) = 1 as i runs over I. Let 1, be
1
-saturated ideals of B. Show that if the Boolean algebras
B/1 and B/ are isomorphic, so are the structures (X, B, 1) and (X, B, ).
(f ) Show that if (X, , ) is a perfect semi-nite measure space which is not purely atomic, there is a
negligible set of cardinal c. (Hint: reduce to the case in which is atomless and totally nite; in this case,
construct a measurable function f : X R such that the image measure = f
1
is atomless, and apply
344H to .)
344 Notes and comments In this section and the last, I have allowed myself to drift some distance from
the avowed subject of this chapter; but it seemed a suitable place for this material, which is fundamental
to abstract measure theory. We nd that the concepts of 342-343 are just what is needed to characterise
Lebesgue measure (344K), and the characterization shows that among non-negligible measurable subspaces
of R
r
the isomorphism classes are determined by a single parameter, the measure of the subspace. Of course
a very large number of other spaces indeed, most of those appearing in ordinary applications of measure
theory to other topics are perfect and countably separated (for example, those of 342Xe and 343Ye), and
therefore covered by this classication. I note that it includes, as a special case, the isomorphism between
Lebesgue measure on [0, 1] and the usual measure on 0, 1
N
already described in 254K.
In 344I, the rst part of the proof is devoted to showing that a perfect countably separated measure space
has countable Maharam type; I ought perhaps to note here that we must resist the temptation to suppose
that all countably separated measure spaces have countable Maharam type. In fact there are countably
separated probability spaces with Maharam type as high as 2
c
. The arguments are elementary but seem to
t better into Volume 5 than here.
I have oered three contexts in which automorphisms of measure algebras are represented by automor-
phisms of measure spaces (344A, 344C, 344E). In the rst case, every automorphism can be represented
simultaneously in a consistent way. In the other two cases, there is, I am sure, no such consistent family of
representations which can be constructed within ZFC; but the theorems I give oer consistent simultaneous
representations of countably many homomorphisms. The question arises, whether countably many is the
true natural limit of the arguments. In fact it is possible to extend both results to families of at most
1
automorphisms. I hope to return to this in Volume 5.
200 Liftings 344 Notes
Having successfully characterized Lebesgue measure or, what is very nearly the same thing, the usual
measure on 0, 1
N
it is natural to seek similar characterizations of the usual measures on 0, 1

for
uncountable cardinals . This seems to be hard. A variety of examples (which I hope to describe in Volume
5) show that none of the most natural conjectures can be provable in ZFC.
In fact the principal new ideas of this section do not belong specically to measure theory; rather, they
belong to the general theory of -algebras and -ideals of sets. In the case of the Schroder-Bernstein-type
theorem 344D, this is obvious from the formulation I give. (See also 344Yb.) In the case of 344B and
344E, I oer generalizations in 344Ya-344Ye. Of course the applications of 344B here, in 344C and its
corollaries, depend on Maharams theorem and the concept of compact measure space. The former has
no generalization to the wider context, and the value of the latter is based on the equivalences in Theorem
343B, which also do not have simple generalizations.
The property described in 344J and 344L a measure space (X, , ) in which any two measurable subsets
of the same non-zero measure are isomorphic seems to be a natural concept of homogeneity for measure
spaces; it seems unreasonable to ask for all sets of zero measure to be isomorphic, since nite sets of dierent
cardinalities can be expected to be of zero measure. An extra property, shared by Lebesgue measure and
the usual measure on 0, 1
I
(and by the measure on the split interval, 344Xf) but not by counting measure,
would be the requirement that measurable sets of dierent non-zero nite measures should be isomorphic up
to a scalar multiple of the measure. All these examples have the further property, that all automorphisms
of their measure algebras correspond to automorphisms of the measure spaces.
345 Translation-invariant liftings
In this section and the next I complement the work of 341 by describing some important special properties
which can, in appropriate circumstances, be engineered into our liftings. I begin with some remarks on
translation-invariance. I restrict my attention to measure spaces which we have already seen, delaying a
general discussion of translation-invariant measures on groups until Volume 4, and to results which can be
proved without special axioms, delaying the use of the continuum hypothesis, in particular, until Volume 5.
345A Translation-invariant liftings In this section I shall consider two forms of translation-invariance,
as follows.
(a) Let be Lebesgue measure on R
r
, and its domain. A lifting : is translation-invariant
if (E +x) = E +x for every E , x R
r
. (Recall from 134A that E +x = y +x : y E belongs to
for every E , x R
r
.)
Similarly, writing A for the measure algebra of , a lifting : A is translation-invariant if
(E +x)

= E

+x for every E , x R
r
.
It is easy to see that if and correspond to each other in the manner of 341B, then one is translation-
invariant if and only it the other is.
(b) Now let I be any set, and let be the usual measure on X = 0, 1
I
, with its domain and A its
measure algebra. For x, y X, dene x + y X by setting (x + y)(i) = x(i) +
2
y(i) for every i I; that
is, give X the group structure of the product group Z
I
2
. This makes X an abelian group (isomorphic to the
additive group (TI, ) of the Boolean algebra TI, if we match x X with i : x(i) = 1 I).
Recall that the measure is a product measure (254J), being the product of copies of the fair-coin
probability on the two-element set 0, 1. If x X, then for each i I the map +
2
x(i) : 0, 1 0, 1
is a measure space automorphism of 0, 1, since the two singleton sets 0 and 1 have the same measure
1
2
. It follows at once that the map y y +x : X X is a measure space automorphism.
Accordingly we can again say that a lifting : A , or : , is translation-invariant if
(E +x)

= E

+x, (E +x) = E +x
for every E , x X.
345C Translation-invariant liftings 201
345B Theorem For any r 1, there is a translation-invariant lifting of Lebesgue measure on R
r
.
proof (a) Write for Lebesgue measure on R
r
, for its domain. Let : be lower Lebesgue density
(341E). Then is translation-invariant in the sense that (E +x) = E +x for every E , x R
r
. PPP
(E +x) = y : y R
r
, lim
0
(E+x)B(y,)
(B(y,))
= 1
= y : y R
r
, lim
0
(EB(yx,))
(B(yx,))
= 1
(because is translation-invariant)
= y +x : y R
r
, lim
0
(EB(y,))
(B(y,))
= 1
= E +x. QQQ
(b) Let
0
be any lifting of such that
0
E E for every E (341J). Consider
E = y : 0
0
(E y)
for E . It is easy to check that : is a Boolean homomorphism because
0
is, so that, for
instance,
y EF 0
0
(E y)
0
(F y)
0
0
((E y)(F y)) =
0
((EF) y))
y (EF).
(c) If E = 0, then E y is negligible for every y R
r
, so
0
(E y) is always empty and E = .
(d) Next, E E for every E . PPP If y E, then
0 = y y E y = (E y)
0
(E y),
so y E. QQQ By 341Ib, is a lifting for .
(e) Finally, is translation-invariant, because if E and x, y R
r
then
y (E +x) 0
0
(E +x y) =
0
(E (y x))
y x E
y E +x.
345C Theorem For any set I, there is a translation-invariant lifting of the usual measure on 0, 1
I
.
proof I base the argument on the same programme as in 345B. This time we have to work rather harder,
as we have no simple formula for a translation-invariant lower density. However, the ideas already used in
341F-341H are in fact adequate, if we take care, to produce one.
(a) Since there is certainly a bijection between I and its cardinal = #(I), it is enough to consider the
case I = . Write for the usual measure on X = 0, 1
I
= 0, 1

and for its domain. For each <

set E

= x : x X, x() = 1, and let

be the -algebra generated by E

: < . Because x + E

is
either E

or XE

, and in either case belongs to

, for every < and x X,

is translation-invariant.
(Consider the algebra

= E : E +x

for every x X;
this must be

.) Let

be the set of partial lower densities :

which are translation-invariant in

the sense that (E +x) = E +x for any E

, x X.
(b)(i) For < ,
+1
is just the algebra of subsets of X generated by

, that is, sets of the

form (F E

) (G E

) where F, G

(312M). Moreover, the expression is unique. PPP Dene x

X by
202 Liftings 345C
setting x

() = 1, x

() = 0 if ,= . Then x

+E

= E

for every < , so x

+F = F for every F

.
If H = (F E

) (G E

) where F, G

, then
x

+H = ((x

+F) (x

+E

)) ((x

+G) (x

+E

)) = (F E

) (G E

),
so
F = (H E

) ((x

+H) E

) = F
H
,
G = (H E

) ((x

+H) E

) = G
H
are determined by H. QQQ
(ii) The functions H F
H
, H G
H
:
+1

dened above are clearly Boolean homomorphisms;

moreover, if H, H
t

+1
and HH
t
is negligible, then
(F
H
F
H
) (G
H
G
H
) (HH
t
) (x

+ (HH
t
))
is negligible. It follows at once that if < and

, we can dene
1
:
+1
by setting

1
H = (F
H
E

) (G
H
E

),
and
1
will be a lower density. If H

then F
H
= G
H
= H, so
1
H = H.
(iii) To see that
1
is translation-invariant, observe that if x X and x() = 0 then x +E

= E

, so,
for any F, G

1
(x + ((F E

) (G E

))) =
1
(((F +x) E

) ((G+x) E

))
= ((F +x) E

) ((G+x) E

)
= ((F +x) E

) ((G+x) E

)
= x + (F E

) (G E

)
= x +
1
((F E

) (G E

)).
While if x() = 1 then x +E

= X E

, so

1
(x + ((F E

) (G E

))) =
1
(((F +x) E

) ((G+x) E

))
= ((F +x) E

) ((G+x) E

)
= ((F +x) E

) ((G+x) E

)
= x + (F E

) (G E

)
= x +
1
((F E

) (G E

)).
So
1

+1
.
(iv) Thus every member of

has an extension to a member of

+1
.
(c) Now suppose that (n))
nN
is a non-decreasing sequence in with supremum < . Then

is
just the -algebra generated by

nN

(n)
. If we have a sequence
n
)
nN
such that
n

(n)
and
n+1
extends
n
for every n, then there is a

extending every
n
. PPP I repeat the ideas of 341G.
(i) For E

, n N choose g
En
such that g
En
is a conditional expectation of E on
(n)
; that is,
_
F
g
En
=
_
F
E = (F E)
for every E
(n)
. Moreover, make these choices in such a way that () every g
En
is
(n)
-measurable
and dened everywhere on X () g
En
= g
E

n
for every n if EE
t
is negligible. Now lim
n
g
En
exists and
is equal to E almost everywhere, by Levys martingale theorem (275I).
(ii) For E

, k 1, n N set
H
kn
(E) = x : x X, g
En
(x) 1 2
k

(n)
,

H
kn
(E) =
n
(H
kn
(E)),
345C Translation-invariant liftings 203
E =

k1

nN

mn

H
km
(E).
(iii) Every g
n
is zero almost everywhere, every H
kn
() is negligible and every

H
kn
() is empty;
so = . If E, E
t

and EE
t
is negligible, g
En
= g
E

n
for every n, H
nk
(E) = H
nk
(E
t
) and

H
nk
(E) =

H
nk
(E
t
) for all n, k, and E = E
t
.
(iv) If E F in

, then g
En
g
Fn
almost everywhere for every n, every H
kn
(E) H
kn
(F) is
negligible,

H
kn
(E)

H
kn
(F) for every n, k, and E F.
(v) If E, F

then (E F) E + F 1 a.e. so g
EF,n
g
En
+ g
Fn
1 a.e. for every n.
Accordingly
H
k+1,n
(E) H
k+1,n
(F) H
kn
(E F)
is negligible, and (because
n
is a lower density)

H
kn
(E F)
n
(H
k+1,n
(E) H
k+1,n
(F)) =

H
k+1,n
(E)

H
k+1,n
(F)
for all k 1, n N. Now, if x E F, then, for any k 1, there are n
1
, n
2
N such that
x

mn
1

H
k+1,m
(E), x

mn
2

H
k+1,m
(F).
But this means that
x

mmax(n
1
,n
2
)

H
km
(E F).
As k is arbitrary, x (E F); as x is arbitrary, E F (E F). We know already from (iv) that
(E F) E F, so (E F) = E F.
(vi) If E

, then g
En
E a.e., so setting
V =

k1

nN

mn
H
km
(E) = x : limsup
n
g
En
(x) 1,
V
E
E is negligible; but
EV

k1,nN
H
kn
(E)

H
kn
(E)
is also negligible, so EE is negligible. Thus is a partial lower density with domain

.
(vii) If E
(n)
, then E
(m)
for every m n, so g
Em
= E a.e. for every m n; H
km
(E)E
is negligible for k 1, m n;

H
km
(E) =
m
E =
n
E
for k 1, m n; and E =
n
E. Thus extends every
n
.
(viii) I have still to check the translation-invariance of . If E

and x X, consider g
t
n
, dened
by setting
g
t
n
(y) = g
En
(y x)
for every y X, n N; that is, g
t
n
is the composition g
En
, where (y) = y x for y X. (I am not sure
whether it is more, or less, confusing to distinguish between the operations of addition and subtraction in
X. Of course y x = y + (x) = y + x for every y.) Because is a measure space automorphism, and in
particular is inverse-measure-preserving, we have
_
F+x
g
t
n
=
_

1
[F]
g
t
n
=
_
F
g
En
= (E F)
whenever F
(n)
(235Ic). But because
(n)
is itself translation-invariant, we can apply this to F x to
get
_
F
g
t
n
= (E (F x)) = ((E +x) F)
for every F
(n)
. Moreover, for any R,
y : g
t
n
(y) = y : g
En
(y) +x
(n)
204 Liftings 345C
for every , and g
t
n
is
(n)
-measurable. So g
t
n
is a conditional expectation of (E +x) on
(n)
, and must
be equal almost everywhere to g
E+x,n
.
This means that if we set
H
t
kn
= y : g
t
n
(y) 1 2
k
= H
kn
(E) +x
for k, n N, we shall have H
t
kn

(n)
and H
t
kn
H
kn
(E +x) will be negligible, so

H
kn
(E +x) =
n
(H
kn
(E +x)) =
n
(H
t
kn
)
=
n
(H
kn
(E) +x) =
n
(H
kn
(E)) +x =

H
kn
(E) +x.
Consequently
(E +x) =

k1
_
nN

mn

H
kn
(E +x)
=

k1
_
nN

mn

H
kn
(E) +x = E +x.
As E and x are arbitrary, is translation-invariant and belongs to

. QQQ
(d) We are now ready for the proof that there is a translation-invariant lower density on X. PPP Build
inductively a family

such that ()

for each ()

extends

whenever . The
induction starts with
0
= , X,
0
= ,
0
X = X. The inductive step to a successor ordinal is dealt
with in (b), and the inductive step to a non-zero ordinal of countable conality is dealt with in (c). If
has uncountable conality, then

=

<

, so we can (and must) take

to be the unique common

extension of all the previous

.
The induction ends with

. Note that

is not in general the whole of . But for every

E there is an F

such that EF is negligible (254Ff). So we can extend

to a function dened
on the whole of by setting
E =

F whenever E , F

and (EF) = 0
(the point being that

F =

F
t
if F, F
t

and (EF) = (EF

t
) = 0). It is easy to check that
is a lower density, and it is translation-invariant because if E , x X, F

and EF is negligible,
then (E +x)(F +x) = (EF) +x is negligible, so
(E +x) =

(F +x) =

F +x = E +x. QQQ
(e) The rest of the argument is exactly that of parts (b)-(e) of the proof of 345B; you have to change R
r
into X wherever it appears, but otherwise you can use it word for word, interpreting 0 as the identity of
the group X, that is, the constant function with value 0.
345D Translation-invariant liftings are of great importance, and I will return to them in 447 with a
theorem dramatically generalizing the results above. Here I shall content myself with giving one of their
basic properties, set out for the two kinds of translation-invariant lifting we have seen.
Proposition Let (X, , ) be either Lebesgue measure on R
r
or the usual measure on 0, 1
I
for some
set I, and let : be a translation-invariant lifting. Then for any open set G X we must have
G G G, and for any closed set F we must have int F F F.
proof (a) Suppose that G X is open and that x G. Then there is an open set U such that 0 U and
x+UU = x+yz : y, z U G. PPP () If X = R
r
, take > 0 such that y : |yx| G, and set
U = y : |y x| <
1
2
. () If X = 0, 1
I
, then there is a nite set K I such that y : yK = xK G
(3A3K); set U = y : y(i) = 0 for every i K. QQQ
It follows that x G. PPP Consider H = x +U. Then H = U > 0 so H H ,= . Let y U be such
that x +y H. Then
x = (x +y) y (H y) G
345F Translation-invariant liftings 205
because
H y x +U U G. QQQ
(b) Thus G G for every open set G X. But it follows at once that if G is open and F is closed,
int F (int F) F,
G = X int(X G) X (X G) = G,
F = X (X F) X (X F) = F.
345E I remarked in 341Lg that it is undecidable in ordinary set theory whether there is a lifting for
Borel measure on R. It is however known that there can be no translation-invariant Borel lifting. The
argument depends on the following fact about measurable sets in 0, 1
N
.
Lemma Let be the usual measure on X = 0, 1
N
, and E any non-negligible measurable set. Then there
are x, x
t
E which dier at exactly one coordinate.
proof By 254Fe, there is a set F, determined by coordinates in a nite set, such that (EF)
1
4
E; we
have F
3
4
E, so (EF)
1
3
F. Suppose that F is determined by coordinates in 0, . . . , n1. Then
the map : X X, dened by setting (x)(n) = 1 x(n), (x)(i) = x(i) for i ,= n, is a measure space
automorphism, and
(
1
[EF] (EF)) 2(EF) < F.
Take any x F ((EF)
1
[EF]). Then x
t
= x diers from x at exactly one coordinate; but also
x
t
F, by the choice of n, so both x and x
t
belong to E.
345F Proposition Let be Borel measure on R, that is, the restriction of Lebesgue measure to the
algebra B of Borel sets. Then is translation-invariant, but has no translation-invariant lifting.
proof (a) To see that is translation-invariant all we have to know is that B is translation-invariant and that
Lebesgue measure is translation-invariant. I have already cited 134A for the proof that Lebesgue measure
is invariant, and B is invariant because G+x is open for every open set G and every x R.
(b) The argument below is most easily expressed in terms of the geometry of the Cantor set C. Recall
that C is dened as the intersection

nN
C
n
of a sequence of closed subsets of [0, 1]; each C
n
consists of
2
n
closed intervals of length 3
n
; C
n+1
is obtained from C
n
by deleting the middle third of each interval of
C
n
. Any point of C is uniquely expressible as f(e) =
2
3

n=0
3
n
e(n) for some e 0, 1
N
. (See 134G.) Let
be the usual measure of 0, 1
N
. Because the map e e(n) : 0, 1
N
0, 1 is measurable for each n,
f : 0, 1
N
R is measurable.
We can label the closed intervals constituting C
n
as J
z
)
z0,1]
n, taking J

to be the unit interval [0, 1]

and, for z 0, 1
n
, taking J
z

0
to be the left-hand third of J
z
and J
z

1
to be the right-hand third of J
z
.
(If the notation here seems odd to you, there is an explanation in 3A1H.)
For n N, z 0, 1
n
, let J
t
z
be the open interval with the same centre as J
z
and twice the length. Then
J
t
z
J
z
consists of two open intervals of length 3
n
/2 on either side of J
z
; call the left-hand one V
z
and the
right-hand one W
z
. Thus V
z

1
is the right-hand half of the middle third of J
z
, and W
z

0
is the left-hand
half of the middle third of J
z
.
Construct sets G, H R as follows.
G is to be the union of the intervals V
z
where z takes the value 1 an even number of times,
together with the intervals W
z
where z takes the value 0 an odd number of times;
H is to be the union of the intervals V
z
where z takes the value 1 an odd number of times,
together with the intervals W
z
where z takes the value 0 an even number of times.
G and H are open sets. The intervals V
z
, W
z
between them cover the whole of the interval

1
2
,
3
2
_
with
the exception of the set C and the countable set of midpoints of the intervals J
z
; so that

1
2
,
3
2
_
(G H)
is negligible. We have to observe that G H = . PPP For each z, J
t
z

0
and J
t
z

1
are disjoint subsets of J
t
z
.
Consequently J
t
z
J
t
w
is non-empty just when one of z, w extends the other, and we need consider only
206 Liftings 345F
the intersections of the four sets V
z
, W
z
, V
w
, W
w
when w is a proper extension of z; say w 0, 1
n
and
z = wm, where m < n. () If in the extension (w(m), . . . , w(n1)) both values 0 and 1 appear, J
t
w
will be
a subset of J
z
, and certainly the four sets will all be disjoint. () If w(i) = 0 for m i < n, then W
w
J
z
is disjoint from the rest, while V
w
V
z
; but z and w take the value 1 the same number of times, so V
w
is
assigned to G i V
z
is, and otherwise both are assigned to H. () Similarly, if w(i) = 1 for m i < n,
V
w
J
z
, W
w
W
z
and z, w take the value 0 the same number of times, so W
z
and W
w
are assigned to the
same set. QQQ
The following diagram may help you to see what is supposed to be happening:
V
0
J
1
W
V
0
J
0
W
0
V
1
J
1
W
1
V
00
J
00
W
00
V
01
J
01
W
01
V
10
J
10
W
10
V
11
J
11
W
11
G H G H G H
The assignment rule can be restated as follows:
V = V

is assigned to G, W = W

is assigned to H;
V
z

0
is assigned to the same set as V
z
, and V
z

1
to the other;
W
z

1
is assigned to the same set as W
z
, and W
z

0
to the other.
(c) Now take any n N and z 0, 1
n
. Consider the two open intervals I
0
= J
t
z

0
, I
1
= J
t
z

1
. These
are both of length = 2 3
n1
and abut at the centre of J
z
, so I
1
is just the translate I
0
+. I claim that
I
1
H = (I
0
G) +. PPP Let A be the set

m>n
w : w 0, 1
m
, w extends z

0,
and for w A let w
t
be the nite sequence obtained from w by changing w(n) = 0 into w
t
(n) = 1 but
leaving the other values of w unaltered. Then V
w
= V
w
+ and W
w
= W
w
+ for every w A. Now
I
0
G =
_
V
w
: w A, w takes the value 1 an even number of times

_
W
w
: w A, w takes the value 0 an odd number of times,
so
(I
0
G) + =
_
V
w
: w A, w takes the value 1 an even number of times

_
W
w
: w A, w takes the value 0 an odd number of times
=
_
V
w
: w A, w
t
takes the value 1 an odd number of times

_
W
w
: w A, w
t
takes the value 0 an even number of times
= I
1
H. QQQ
(d) ??? Now suppose, if possible, that : B B is a translation-invariant lifting. Note rst that U U
for every open U R. PPP The argument is exactly that of 345D as applied to R = R
1
. QQQ Consequently
J
t

1
2
,
3
2
_
J
t

.
But as J
t

(G H) is negligible,
C

1
2
,
3
2
_
G H.
345Xf Translation-invariant liftings 207
Consider the sets E = f
1
[G], F = 0, 1
N
E = f
1
[H]. Because f is measurable and G, H are
Borel sets, E and F are measurable subsets of 0, 1
N
, and at least one of them has positive measure. There
must therefore be e, e
t
0, 1
N
, diering at exactly one coordinate, such that either both belong to E or
both belong to F (345E). Let us suppose that n is such that e(n) = 0, e
t
(n) = 1 and e(i) = e
t
(i) for i ,= n.
Set z = en = e
t
n. Then f(e) belongs to the open interval I
0
= J
t
z

0
, so f(e) I
0
and f(e) G i
f(e) (I
0
G). But now
f(e
t
) = f(e) + 2 3
n1
I
1
= J
t
z

1
,
so
e E f(e) G f(e) (I
0
G)
f(e
t
) ((I
0
G) + 2 3
n1
)
(because is translation-invariant)
f(e
t
) (I
1
H)
(by (c) above)
f(e
t
) H
(because f(e
t
) I
1
I
1
)
e
t
F.
But this contradicts the choice of e. XXX
Thus there is no translation-invariant lifting of .
Remark This result is due to Johnson 80; the proof here follows Talagrand 82b. For references to
various generalizations see Burke 93, 3.
345X Basic exercises (a) In 345Ab I wrote It follows at once that the map y y +x : X X is a
measure space automorphism. Write the details out in full, using 254G or otherwise.
(b) Let S
1
be the unit circle in R
2
, and let be one-dimensional Hausdor measure on S
1
(264-265).
Show that is translation-invariant, if S
1
is given its usual group operation corresponding to complex
multiplication (255M), and that it has a translation-invariant lifting . (Hint: Identifying S
1
with ], ]
with the group operation +
2
, show that we can set E = ], ]
t
(

nZ
E + 2n), where
t
is any
translation-invariant lifting for Lebesgue measure.)
>>>(c) Show that there is no lifting of Lebesgue measure on R which is symmetric in the sense that
(E) = E for every measurable set E, writing E = x : x E. (Hint: can 0 belong to ([0, [)?)
>>>(d) Let be Lebesgue measure on X = R0. Show that there is a lifting of such that (xE) = xE
for every x X and every measurable E X, writing xE = xy : y E.
(e) Let be the usual measure on X = 0, 1
I
, for some set I, its domain, and (A, ) its measure
algebra. (i) Show that we can dene
x
(a) = a+x, for a A and x X, by the formula E

+x = (E+x)

;
and that x
x
is a group homomorphism from X to the group of measure-preserving automorphisms of
A. (ii) Dene

as in the proof of 345C, and set A

= E

: E

. Say that a partial lifting : A

is translation-invariant if (a +x) = a +x for every a A

and x X. Show that any such partial lifting

can be extended to a translation-invariant partial lifting on A
+1
. (iii) Write out a proof of 345C in the
language of 341F-341H.
>>>(f ) Let be a lower density for Lebesgue measure on R
r
which is translation-invariant in the sense that
(E + x) = E + x for every x R
r
and every measurable set E. Show that G G for every open set
G R
r
.
208 Liftings 345Xg
(g) Let be 1-dimensional Hausdor measure on S
1
, as in 345Xb. Show that there is no translation-
invariant lifting of such that E is a Borel set for every E dom.
345Y Further exercises (a) Let (X, , ) be a complete measure space, and suppose that X has a
group operation (x, y) xy (not necessarily abelian!) such that is left-translation-invariant, in the sense
that xE = xy : y E and (xE) = E whenever E and x X. Suppose that : is a
lower density which is left-translation-invariant in the sense that (xE) = x(E) for every E , x X.
Show that there is a left-translation-invariant lifting : such that E E for every E .
(b) Write for the -algebra of Lebesgue measurable subsets of R, and L
0
() for the linear space of
-measurable functions from R to itself. Show that there is a linear operator T : L
0
() L
0
() such that
() (Tu)

= u for every u L
0
() () sup
xR
[(Tu)(x)[ = |u|

for every u L

() () Tu 0 whenever
u L

() and u 0 () T is translation-invariant in the sense that T(S

x
f)

= S
x
Tf

for every x R and

f L
0
(), where (S
x
f)(y) = f(x +y) for f L
0
() and x, y R () T is reection-invariant in the sense
that T(Rf)

= RTf

for every f L
0
(), where (Rf)(x) = f(x) for f L
0
() and x R. (Hint: for
f L
0
(), set
p(f

) = inf : [0, ], lim

0
1
2
x : [x[ , [f(x)[ > = 0.
Set V = u : u L
0
(), p(u) < and show that V is a linear subspace of L
0
() and that pV is a
seminorm. Let h
0
: V R be a linear functional such that h
0
(R)

= 1 and h
0
(u) p(u) for every
u V . Extend h
0
arbitrarily to a linear functional h
1
: L
0
() R; set h(f

) =
1
2
(h
1
(f

) + h
1
(Rf)

). Set
(Tf

)(x) = h(S
x
f)

. You will need 231C.) Show that there must be a u L

1
() such that u 0 but
Tu , 0.
(c) Show that there is no translation-invariant lifting of the usual measure on 0, 1
N
such that E is
a Borel set for every measurable set E.
345 Notes and comments I have taken a great deal of care over the concept of translation-invariance. I
hope that you are already a little impatient with some of the details as I have written them out; but while it
is very easy to guess at the structure of such arguments as part (e) of the proof of 345B, or (b-iii) and (c-viii)
in the proof of 345C, I am not sure that one can always be certain of guessing correctly. A fair test of your
intuition will be how quickly you can generate the formulae appropriate to a non-abelian group operation,
as in 345Ya.
Part (b) of the proof of 345C is based on the same idea as the proof of 341F. There is a useful simplication
because the set E

in 345C, corresponding to the set E of the proof of 341F, is independent of the algebra

in a very strong sense, so that the expression of an element of

+1
in the form (F E

) (G E

) is
unique. Interpreted in the terms of 341F, we have w = v = 1, so that the formula

1
((a e) (b \ e)) =
_
((a v) (b \ v)) E
_

_
((a \ w) (b w)) E
_
used there becomes

1
((a e) (b \ e)) = (a E) (b E),
matching the formula for
1
in the proof of 345C.
The results of this section are satisfying and natural; they have obvious generalizations, many of which
are true. The most important measure spaces come equipped with a variety of automorphisms, and we can
always ask which of these can be preserved by a lifting. The answers are not always obvious; I oer 345Xc
and 346Xb as warnings, and 345Xd as an encouragement. 345Yb is striking (I have made it as striking as I
can), but slightly o the most natural target; the sting is in the last sentence (see 341Ye).
346D Consistent liftings 209
346 Consistent liftings
I turn now to a dierent type of condition which we should naturally prefer our liftings to satisfy. If we
have a product measure on a product X =

iI
X
i
of probability spaces, then we can look for liftings
which respect coordinates, that is, are compatible with the product structure in the sense that they factor
through subproducts (346A). There seem to be obstacles in the way of the natural conjecture (346Za), and I
give the partial results which are known. For Maharam-type-homogeneous spaces X
i
, there is always a lifting
which respects coordinates (346E), and indeed the translation-invariant liftings of 345 on 0, 1
I
already
have this property (346C). There is always a lower density on the product which respects coordinates, and
we can ask for a little more (346G); using the full strength of 346G, we can enlarge this lower density to a
lifting which respects single coordinates and initial segments of a well-ordered product (346H). In the case
in which all the factors are copies of each other, we can arrange for the induced liftings on the factors to
be copies also (346I, 346J, 346Yd). I end the section with an important fact about Stone spaces which is
relevant here (346K-346L).
346A Denition Let (X
i
,
i
,
i
))
iI
be a family of probability spaces, with product (X, , ). I will
say that a lifting : respects coordinates if E is determined by coordinates in J whenever E
is determined by coordinates in J I.
Remark Recall that a set E X is determined by coordinates in J if x
t
E whenever x E, x
t
X
and x
t
J = xJ; that is, if E is expressible as
1
J
[F] for some F

iJ
X
i
, where
J
(x) = xJ for every
x X; that is, if E =
1
J
[
J
[E]]. See 254M. Recall also that in this case, if E is measurable for the product
measure on X, then
J
[E] is measurable for the product measure on

iJ
X
i
(254Ob).
346B Proposition Let (X
i
,
i
,
i
))
iI
be a family of probability spaces, with product (Z, , ). For
J I let (Z
J
,
J
,
J
) be the product of (X
i
,
i
,
i
))
iJ
, and
J
: Z Z
J
the canonical map. Let
: be a lifting. If J I is such that W is determined by coordinates in J whenever W is
determined by coordinates in J, then induces a lifting
J
:
J

J
dened by the formula

1
J
[
J
E] = (
1
J
[E]) for every E
J
.
proof If E
J
, then
1
[E] and (
1
J
[E]) are determined by coordinates in J, so (
1
J
[E]) is of the
form
1
J
[F] for some F
J
; this denes
J
:
J

J
. It is now easy to see that
J
is a lifting.
Remark Of course we frequently wish to use this result with a singleton set J = j. In this case we must
remember that (Z
J
,
J
,
J
) corresponds to the completion of the probability space (X
j
,
j
,
j
).
346C Theorem Let I be any set, and the usual measure on X = 0, 1
I
. Then any translation-
invariant lifting of respects coordinates.
proof Suppose that E X is a measurable set determined by coordinates in J I; take x E and
x
t
X such that x
t
J = xJ. Set y = x
t
x; then y(i) = 0 for i J, so that E +y = y. Now
x
t
= x +y E +y = (E +y) = E
because is translation-invariant. As x, x
t
are arbitrary, E is determined by coordinates in J. As E and
J are arbitrary, respects coordinates.
346D I describe a standard method of constructing liftings from other liftings.
Lemma Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras A, B; suppose that f : X Y
is a (, T)-measurable function inducing an isomorphism F

f
1
[F]

: B A. Then if : T T is a
lifting for , there is a corresponding lifting
t
: given by the formula

t
E = f
1
[F] whenever (Ef
1
[F]) = 0.
proof If we say that : B A is the isomorphism induced by f, then
210 Liftings 346D

t
E = f
1
[(
1
E

)],
where : B T is the lifting corresponding to : T T. Since ,
1
and F f
1
[F] are all Boolean
homomorphisms, so is
t
, and it is easy to check that (
t
E)

= E

for every E and that

t
E = if
E = 0.
Remark Compare the construction in 341P.
346E Theorem Let (X
i
,
i
,
i
))
iI
be a family of Maharam-type-homogeneous probability spaces,
with product (X, , ). Then there is a lifting of which respects coordinates.
proof (a) Replacing each
i
by its completion does not change (254I), so we may suppose that all the
i
are complete. In this case there is for each i an isomorphism between the measure algebra (A
i
,
i
) of
i
and
the measure algebra (B
i
,
i
) of some 0, 1
J
i
with its usual measure
i
(331L). We may suppose that the
sets J
i
are disjoint. Each
i
is compact (342Jd), so the isomorphisms are represented by inverse-measure-
preserving functions f
i
: X
i
0, 1
J
i
(343Ca).
Set K =

iI
J
i
, and let be the usual measure on Y = 0, 1
K
, T its domain. We have a natural
bijection between

iI
0, 1
J
i
and Y , so we obtain a function f : X Y ; literally speaking,
f(x)(j) = f
i
(x(i))(j)
for i I, j J
i
, x X.
(b) Now f is inverse-measure-preserving and induces an isomorphism between the measure algebras A,
B of , .
PPP(i) If L K is nite and z 0, 1
L
, then, setting L
i
= LcapJ
i
for i I,
x : x X, f(x)L = z = (

iI
w : w X
i
, f
i
(w)L
i
= zL
i
)
=

iI

i
w : w X
i
, f
i
(w)L
i
= zL
i

iI

i
v : v 0, 1
J
i
, vL
i
= zL
i

(because every f
i
is inverse-measure-preserving)
=

iI
2
#(L
i
)
= 2
#(L)
= y : y Y, yL = z.
So f
1
[C] = C for every basic cylinder set C Y . By 254G, f is inverse-measure-preserving.
(ii) Accordingly f induces a measure-preserving homomorphism : B A. To see that is surjective,
consider

t
= E : E is -measurable, E

[B].
Because [B] is a closed subalgebra of A (324Kb),
t
is a -subalgebra of the domain of , and of course
it contains all -negligible sets. If i J and G
i
, then there is an H 0, 1
J
i
such that Gf
1
i
[H] is

i
-negligible. Now if E = x : x X, x(i) G and F = y : y Y, yJ
i
H,
Ef
1
[F] = x : x(i) Gf
1
i
[H]
is -negligible, and E
t
. But this means that
t

iI

i
, and must therefore be the whole of
(254Ff). QQQ
(c) By 345C, there is a translation-invariant lifting for ; by 346C, this respects coordinates. By 346D,
we have a corresponding lifting
t
for such that

t
f
1
[F] = f
1
[F]
for every F T. Now suppose that E is determined by coordinates in L I. Then there is an E
t
belonging to the -algebra
t
L
generated by
346G Consistent liftings 211
x : x(i) G : i L, G
i

such that (EE

t
) = 0 (254Ob). Write T
L
for the family of sets in T determined by coordinates in

iL
J
i
.
Then, just as in (b-ii), every member of
t
L
diers by a negligible set from some set of the form f
1
[F] with
F T
L
. So there is an F T
L
such that Ef
1
[F] is -negligible. Consequently

t
E =
t
f
1
[F] = f
1
[F].
But respects coordinates, so F is determined by coordinates in

iL
J
i
. It follows at once that f
1
[F]
is determined by coordinates in L; that is, that
t
E is determined by coordinates in L. As E and L are
arbitrary,
t
respects coordinates, and witnesses the truth of the theorem.
346F It seems to be unknown whether 346E is true of arbitrary probability spaces (346Za); I give some
partial results in this direction. The following general method of constructing lower densities will be useful.
Lemma Let (X, , ) and (Y, T, ) be complete probability spaces, with product (XY, , ). If :
is a lower density, then we have a lower density
1
: dened by saying that

1
E = x : x X, y : (x, y) (E Y ) is conegligible in Y
for every E .
proof For E , (E Y )(E Y ) is negligible, so that
H
x
= y : (x, y) (E Y )(E Y )
is -negligible for almost every x X (252D). Now E
1
E = x : H
x
is not negligible is negligible, so

1
E . If E, F , then
((E F) Y ) = ((E Y ) (F Y )) = (E Y ) (F Y ),
so that
y : (x, y) ((E F) Y ) = y : (x, y) (E Y ) y : (x, y) (F Y )
is conegligible i both y : (x, y) (EY ) and y : (x, y) (F Y ) are conegligible, and
1
(EF) =

1
E
1
F.
The rest is easy. Of course ( Y ) = so
1
= . If E, F and EF is negligible, then
(E Y )(F Y ) is negligible, (E Y ) = (F Y ) and
1
E =
1
F. So
1
is a lower density, as claimed.
346G Theorem Let (X
i
,
i
,
i
))
iI
be a family of probability spaces with product (X, , ). For
J I let
J
be the set of members of which are determined by coordinates in J. Then there is a lower
density : such that
(i) whenever J I and E
J
then E
J
,
(ii) whenever J, K I are disjoint, E
J
and F
K
then (E F) = E F.
proof For each i I, set Y
i
= X
N
i
, with the product measure
i
; set Y =

iI
Y
i
, with its product measure
; set Z
i
= X
i
Y
i
, with its product measure
i
, and Z =

iI
Z
i
, with its product measure . Then the
natural identication of Z =

iI
X
i
Y
i
with

iI
X
i

iI
Y
i
= X Y makes correspond to the
product of and (254N).
Each (Z
i
,
i
) can be identied with an innite power of (X
i
,
i
), and is therefore Maharam-type-homo-
geneous (334E). Consequently there is a lifting : which respects coordinates (346E). Regarding
(Z, ) as the product of (X, ) and (Y, ), we see that induces a lower density : by the formula
of 346F.
If J I and E is determined by coordinates in J, then E Y (regarded as a subset of

iI
Z
i
) is
determined by coordinates in J, so (E Y ) also is. Now suppose that x E, x
t
X and xJ = x
t
J.
Then for any y Y , (xJ, yJ) = (x
t
J, yJ), so (x, y) (E Y ) i (x
t
, y) (E Y ). Thus
y : (x
t
, y) (E Y ) = y : (x, y) (E Y )
is conegligible in Y , and x
t
E. This shows that E is determined by coordinates in J.
212 Liftings 346G
Now suppose that J and K are disjoint subsets of I, that E, F are determined by coordinates in J,
K respectively, and that x / E F. Then A = y : (x, y) / (E Y ) and B = y : (x, y) / (F Y )
are non-negligible. As noted just above, (E Y ) is determined by coordinates in J, so A is determined
by coordinates in J, and can be expressed as y : yJ A
t
, where A
t
Y
J
=

iJ
Y
i
. Because
y yJ : Y Y
J
is inverse-measure-preserving, A
t
cannot be negligible in Y
J
. Similarly, B can be
expressed as y : yK B
t
for some non-negligible B
t
Y
K
.
By 251R/251Wm, A
t
B
t
Y
I\(JK)
, regarded as a subset of Y , is non-negligible, that is,
C = y : y Y, yJ A
t
, yK B
t

is non-negligible. But
C = A B = y : (x, y) / (E Y ) (F Y ) = y : (x, y) / ((E F) Y .
So x / (E F). As x is arbitrary, (E F) E F; but of course E F (E F), because is
a lower density, so that (E F) = E F, as required.
Remark See Macheras Musial & Strauss p99 for an alternative proof.
346H Theorem Let be an ordinal, and (X

))
<
any family of probability spaces, with
product (Z, , ). For J let
J
be the set of those W which are determined by coordinates in J.
Then there is a lifting : such that W
J
whenever W
J
and J is either a singleton subset
of or an initial segment of .
proof (a) Let P be the set of all lower densities : such that, for every < , (i) whenever
E

then E

(ii) whenever E
]
then E
]
(iii) whenever E

and F
\
then
(E F) = E F. By 346G, P is not empty. Order P by saying that
t
if E
t
E for every
E ; then P is a partially ordered set. Note that if P then Z = Z (because
0
= , Z).
(b) Any non-empty totally ordered subset Q of P has an upper bound in P. PPP Dene

: TX by
setting

E =

Q
E for every E . (i)

Q
= .
(ii) If E, F and (EF) = 0 then E = F for every Q so

E =

F. (iii) If E, F and
E F then E F for every Q so

F. (iv) If E, F and x

F, then there are

1
,
2
Q such that x
1
E
2
F; now either
1

2
or
2

1
, so that
x (
1
E
1
F) (
2
E
2
F) =
1
(E F)
2
(E F)

(E F).
Accordingly

(E F) and

F =

(E F). (v) Taking any

0
Q, we have

0
E

E for every E , so (because is complete)

is a lower density, by 341Ib. (vi) Now suppose

that J I is either a singleton or an initial segment , and that E
J
. Then E is determined by
coordinates in J for every Q, so

E is determined by coordinates in J. (vii) Finally, suppose that

< and that E

, F
\
. If x

(E F) then there is a Q such that

x (E F) = E F

F.
So

(E F)

F and (using (iii) again)

(E F) =

F. Thus

belongs to P and is
an upper bound for Q in P. QQQ
By Zorns Lemma, P has a maximal element

.
(c) For any H we may dene a function
H
as follows. Set A
H
= Z (

H

(Z H)),

H
E =

E (A
H


(H E))
for E . Then
H
is a lower density. PPP (i) Because H

H and (ZH)

(ZH) are both negligible, A

H
is negligible and
H
E is measurable and (
H
E)

= (

E)

= E

for every E . (ii) Because A

H

H = ,

H
= . (iii) If E, F and (EF) = 0 then

E =

F and

(E H) =

(F H), so
H
E =
H
F.
(iv) If E, F and E F then

E

F and

(E H)

(F H), so
H
E
H
F. (v) If E, F
and x
H
E
H
F, then
346H Consistent liftings 213
() if x / A
H
,
x

E

F =

(E F)
H
(E F),
() if x A
H
,
x

(E H)

(F H) =

((E F) H)
H
(E F).
Thus
H
E
H
F
H
(E F) and
H
E
H
F =
H
(E F). QQQ
(d) It is worth noting the following.
(i) If E, H and

(E H) =

E

H then
H
E =

E. PPP We have

H
E =

E (A
H


(E H)) =

E (A
H


E) (A
H


H) =

E
because A
H


H = . QQQ
(ii) If H and
H
P then

H

(Z H) = Z. PPP By the maximality of

, we must have
H
=

.
But
A
H
=
H
(Z H)

(Z H),
so A
H
= , that is,

H

(Z H) = Z. QQQ
(iii) If E, F and

E

(Z E) = Z, then

(E F) =

E

F. PPP

(E F)

E =

(E F)

(Z E) =

((E F) (Z E)) =

(F E)

F,
so

(E F)

E

F; as the reverse inclusion is true for all E and F, we have the result. QQQ
(e) If < and H
]
, then
H
P.
PPP(i) If J I is either a singleton or an inital segment, and E
J
, then
() if J, E H and

E and

(E H) and A
H
all belong to
J
, so
H
E
J
.
() If / J,

(E H) =

E

H, because there is some such that J and ); so

H
E =

E
J
by (d-i).
(ii) If < and E

, F
\
, then
if < , E H

so

(E F H) =

(E H)

F, and

H
(E F) =

(E F) (A
H


(E F H))
=

E

F (A
H


(E H)) (A
H


F)
H
E
H
F;
if , F H
\
so

(E F H) =

(E)

(F H), and

H
(E F) =

(E F) (A
H


(E F H))
=

E

F (A
H


E)) (A
H


(F H))
H
E
H
F;
accordingly
H
(E F) =
H
E
H
F. QQQ
By (d-ii) we have

H

(Z H) = Z
whenever < and H
]
.
(f ) If and H

, then
H
P. PPP Induce on . For = 0, H
0
= , Z so

H is either or
Z and the result is trivial. For the inductive step to , we have the following.
(i) If < and E

, then
() if , E H and

E and

(E H) and A
H
all belong to

, so
H
E

.
214 Liftings 346H
() if < , then, by the inductive hypothesis,
E
P,

E = Z

(ZE) and

(EH) =

E

H,
by (d-ii) and (d-iii) above; so
H
E =

E

by (d-i).
(ii) If < and E
]
, then, by (e),

E

(Z E) = Z, so that

(E H) =

E

H, by (d-iii),
and
H
E =

E
]
, by (d-i).
(iii) If < and E

, F
\
, then
() if , then E H

and F
\
, so that

(E F H) =

(E H)

F, and

H
(E F) =

(E F) (A
H


(E F H))
=

E

F (A
H


(E H)) (A
H


F)
H
E
H
F,
as in (e-ii) above, and accordingly
H
(E F) =
H
E
H
F.
() If < then, as in (ii), using the inductive hypothesis, we have

(E F H) =

E

(F H),
and (just as in ()) we get
H
(E F) =
H
E
H
F.
Thus
H
P and the induction continues. QQQ
(g) But the case = of (f) just tells us that

H

(Z H) = Z
for every H . This means that

is actually a lifting (since it preserves intersections and complements).
And the denition of P is just what is needed to ensure that it is a lifting of the right type.
Remark This result is due to Macheras & Strauss 96b.
346I Theorem Let (X, , ) be a complete probability space. For any set I, write
I
for the product
measure on X
I
,
I
for its domain and
Ii
(x) = x(i) for x X
I
, i I. Then there is a lifting :
such that for every set I there is a lifting :
I

I
such that (
1
Ii
[E]) =
1
Ii
[E] for every E ,
i I.
proof ??? Suppose, if possible, otherwise.
Let be the set of all liftings for . We are supposing that for every there is a set I

for
which there is no lifting for
I

consistent with in the sense above. Let be a cardinal greater than

max(, #(), sup

#(I

)). Let
0
:

be a lifting satisfying the conditions of 346H. 346B tells us

that for every < we have a lifting for dened by the formula
1

[E] =
0
(
1

[E]). For set

K

= : < ,
0
(
1

[E]) =
1

[E] for every E .

Then

= , so max(, #(), sup

#(K

)) and there is some such that #(K

) >
#(I

). Take I K

such that #(I) = #(I

).
We may regard X

as X
I
X
\I
, and in this form we can use the method of 346F to obtain a lower
density :
I

I
from
0
:

. Now
(
1
I
[E]) =
1
I
[E] for every E , I.
PPP The point is that
1
I
[E] X
\I
corresponds to
1

[E] X

, while
0
(
1

[E]) =
1

[E] can be
identied with
1
I
[E] X
\I
. Now the construction of 346F obviously makes (
1
I
[E]) equal to
1
I
[E].
QQQ
By 341Jb, there is a lifting :
I

I
such that W W for every W
I
. But now we must have

1
I
[E] = (
1
I
[E]) (
1
I
[E])
= X
I
(
1
I
[X E]) X
I
(
1
I
[X E])
= X
I

1
I
[(X E)] = X
I

1
I
[X E] =
1
I
[E]
and (
1
I
[E]) =
1
I
[E] for every E , I. But since #(I) = #(I

), this must be impossible, by the

choice of I

. XXX
This contradiction proves the theorem.
346K Consistent liftings 215
346J Consistent liftings Let (X, , ) be a measure space. A lifting : is consistent if for
every n 1 there is a lifting
n
of the product measure on X
n
such that
n
(E
1
. . .E
n
) = E
1
. . .E
n
for all E
1
, . . . , E
n
. Thus 346I tells us, in part, that every complete probability space has a consistent
lifting; it follows that every non-trivial complete totally nite measure space has a consistent lifting.
I do not suppose you will be surprised to be told that not all liftings on probability spaces are consistent.
What may be surprising is the fact that one of the standard liftings already introduced is not consistent. This
depends on a general fact about Stone spaces of measure algebras which has further important applications,
so I present it as a lemma.
346K Lemma Let (Z, T, ) be the Stone space of the measure algebra of Lebesgue measure on [0, 1],
and let be the product measure on Z Z, with its domain. Then there is a set W , with W < 1,
such that

W = 1, where

W =

GH : G, H Z are open-and-closed, (GH) W is negligible.

Remark For the sake of anybody who has already become acquainted with the alternative measures which
can be put on the product of topological measure spaces, I ought to insist here that the product measure
is, as always in this volume, the ordinary completed product measure as dened in Chapter 25.
proof (a) Let E
n
)
nN
be a sequence of measurable subsets of [0, 1], stochastically independent for Lebesgue
measure on [0, 1], such that E
n
=
1
n+2
for each n. Set a
n
= E

n
in the measure of algebra of , and
E

n
= a
n
the corresponding compact open subset of Z. Set W =

nN
E

n
E

n
. Then
W

n=0
(E
n
)
2
=

n=2
1
n
2
< 1.
??? Suppose, if possible, that

W < 1. Then there are sequences G
n
)
nN
, H
n
)
nN
in T such that

nN
G
n
H
n
and (

nN
G
n
H
n
) < 1. Recall from 322Qc that
F = infG : G is compact and open, F G
for every F T. Accordingly we can nd compact open sets

G
n
,

H
n
such that G
n


G
n
, H
n


H
n
for
every n N and

n=0
(

G
n
G
n
) +

n=0
(

H
n
H
n
) < 1 (

nN
G
n
H
n
),
so that (

nN

G
n


H
n
) < 1.
Let |
0
be the family
Z E

n
: n N Z

G
n
: n N Z

H
n
: n N,
so that |
0
is a countable subset of T. Let | be the set of nite intersections U
0
U
1
. . . U
n
where
U
0
, . . . , U
n
|
0
, so that | is also a countable subset of T, and | is closed under .
(b) For U |, dene Q(U) as follows. If U = 0, then Q(U) = U. Otherwise,
Q(U) = Z

n
: n N, (E

n
U) > 0.
Then Q(U) is always 0. PPP Of course this is true if U = 0, so suppose that U > 0. Set I = n :
(E

n
U) = 0. Then we have U
t
> 0, where U
t
= U

nI
E

n
, and ZE

n
U
t
for every n I. Because
E
n
)
nN
is stochastically independent for , E

n
)
nN
is stochastically independent for , while
(

nI
E

n
) 1 U
t
< 1.
By the Borel-Cantelli lemma (273K),

nI
E

n
< . Consequently

nN\I
E

n
= , because

n=0
1
n+2
is innite, so
(Z Q(U)) = (

nN\I
E

n
) = 1,
and Q(U) = 0. QQQ
(c) Set Q
0
=

U/
Q(U); because | is countable, Q
0
is negligible. Accordingly (Z Q
0
)
2
has measure
1 and cannot be included in

nN

G
n


H
n
; take (w, z) (Z Q
0
)
2

nN

G
n


H
n
.
216 Liftings 346K
(d) We can nd sequences C
n
)
nN
, D
n
)
nN
, U
n
)
nN
and V
n
)
nN
in | such that
w U
n+1
U
n
, z V
n+1
V
n
, (U
n+1
V
n+1
) (

G
n


H
n
) = ,
C
n
> 0, D
n
> 0,
C
n
U
n
, D
n
V
n+1
,
C
n
V
n+1
W, U
n+1
D
n
W
for every n N. PPP Build the sequences inductively, as follows. Start with U
0
= V
0
= Z. Given that
w U
n
|, z V
n
|, then we know that (w, z) /

G
n


H
n
. If w /

G
n
, set U
t
n
= U
n


G
n
,
V
t
n
= V
n
; otherwise set U
t
n
= U
n
, V
t
n
= V
n


H
n
. In either case, we have w U
t
n
|, z V
t
n
| and
(U
t
n
V
t
n
) (

G
n


H
n
) = .
Because U
t
n
|, w / Q(U
t
n
). But w U
t
n
, so this must be because U
t
n
> 0. Now z / Q(U
t
n
), so
z

k
: k N, (E

k
U
t
n
) > 0. Take some k N such that z E

k
and (E

k
U
t
n
) > 0, and set
V
n+1
= V
t
n
E

k
, C
n
= E

k
U
t
n
,
so that
z V
n+1
|, C
n
U
n
, C
n
V
n+1
E

k
E

k
W, C
n
> 0.
Next, z / Q(V
n+1
) and V
n+1
> 0; also w / Q(V
n+1
), so there is an l such that w E

l
and (E

l
V
n+1
) > 0.
Set
U
n+1
= U
t
n
E

l
, D
n
= E

l
V
n+1
,
so that
w U
n+1
|, D
n
V
n+1
, U
n+1
D
n
E

l
E

l
W, D
n
> 0,
(U
n+1
V
n+1
) (

G
n


H
n
) (U
t
n
V
t
n
) (

G
n


H
n
) = ,
and continue the process. QQQ
(e) Setting C =

nN
C
n
, D =

nN
D
n
we see that C D W. PPP If m n, D
n
V
n+1
V
m+1
, so
C
m
D
n
W. If m > n, C
m
U
m
U
n+1
, so C
m
D
n
W. QQQ
Recall from 322Qa that the measurable sets of Z are precisely those of the form GM where M is
nowhere dense and negligible and G is compact and open. There must therefore be compact open sets G,
H Z such that GC and HD are negligible. Consequently
(GH) W ((G C) Z) (Z (H D))
is negligible, and
GH

W

nN

G
n


H
n
.
But because G H is compact (3A3J), and all the

G
n


H
n
are open, there must be some n such that
GH

kn

G
k


H
k
= S say. Now (U
k+1
V
k+1
) (

G
k


H
k
) = for every k, so
(C
n+2
D
n+2
) (GH) (U
n+1
V
n+1
) S = ,
and either C
n+2
G = or D
n+2
H = . Since
C
n+2
G C G, D
n+2
H D H
are both negligible, one of C
n+2
, D
n+2
is negligible. But the construction took care to ensure that all the
C
k
, D
k
were non-negligible. XXX
(f ) Thus

W = 1, as required.
346L Proposition Let (Z, T, ) be the Stone space of the measure algebra of Lebesgue measure on [0, 1].
Let : T T be the canonical lifting, dened by setting E = G whenever E T, G is open-and-closed
and EG is negligible (341O). Then is not consistent.
proof ??? Suppose, if possible, that is a lifting on Z Z such that (E F) = E F for every E,
F T. Let W Z Z be a set as in 346K, and consider W. If G, H Z are open-and-closed and
(GH) W is negligible, then
346Yd Consistent liftings 217
GH = GH = (GH) W;
that is, in the language of 346K, we must have

W W. But this means that
(W)

W = 1 > W,
which is impossible. XXX
Thus fails the rst test and cannot be consistent.
346X Basic exercises (a) Let (X, , ) be a measure space and a lower density for . Take H
and set A = X (H (Z H)),
t
E = E (A (H E)) for E . Show that
t
is a lower density.
>>>(b) Show that there is no lifting of Lebesgue measure on [0, 1]
2
which is symmetric in the sense that
(E
1
) = (E)
1
for every measurable set E, writing E
1
= (y, x) : (x, y) E.
(c) Let (X, , ) be a measure space and
n
)
nN
a sequence of lower densities for . (i) Show that
E

nN

n
E and E

nN

mn

m
E are also lower densities for . (ii) Show that if is complete
and T is any lter on N, then E

FJ

nF

n
E is a lower density for .
(d) Let (X, , ) be a strictly localizable measure space, and G a countable group of measure space
automorphisms from X to itself. Show that there is a lower density : which is G-invariant in the
sense that (g
1
[E]) = g
1
[E] for every E , g G. (Hint: set E =

gG
g[
0
(g
1
[E])].)
>>>(e) Let be lower Lebesgue density on R, and any lifting of Lebesgue measure on R such that
E E for every measurable set E. Show that is consistent. (Hint: given n 1, let
n
be lower
Lebesgue density on R
n
. For x R
n
let 1
x
be the ideal generated by
W : x
n
(R
n
W)

i<n

1
i
[E] : x(i) (R E);
show that R
n
/ 1
x
, so that we can use the method of 341J to construct a lifting of Lebesgue measure on
R
n
.)
(f ) Show that Lemma 346K is valid for any (Z, T, ) which is the Stone space of an atomless probability
space.
>>>(g) Suppose, in 341H, that (X, , ) is a product of probability spaces, and that in the proof, instead of
taking a

)
<
to run over the whole measure algebra A, we take it to run over the elements of A expressible
as E

where E is determined by a single coordinate. Show that the resulting lower density respects
coordinates in the sense that E

is determined by coordinates in J whenever E is determined by

coordinates in J. (Compare Macheras & Strauss 95, Theorem 2.)
346Y Further exercises (a) Suppose that (X, , ) and (Y, T, ) are complete probability spaces with
product (X Y, , ). Show that for any lifting
1
: there are liftings
2
: T T and :
such that (E F) =
1
E
2
F for all E , F T. (Hint: use the methods of 341. In the inductive
construction of 341H, start with
0
(E Y ) =
1
E Y for every E . Extend each lower density

to
the algebra generated by dom(

) X F

for some F

T. Make sure that

(X F) is always of the
form XF
t
, and that

((E Y ) (XF)) =

(E Y )

(XF); adapt the construction of 341G to

maintain this. Use the method of 346H to generate a lifting from the nal lower density . See Macheras
& Strauss 96a, Theorem 4.)
(b) Use 346Ya and induction on to prove 346H. (Macheras & Strauss 96b.)
(c) Let (X
1
,
1
,
1
), . . . , (X
n
,
n
,
n
) be probability spaces with product (X, , ). Show that there is a
lifting for which respects coordinates. (Burke n95.)
(d) Let (X, , ) be a complete probability space. Show that there is a lifting : such that
whenever (X
i
,
i
,
i
))
iI
is a family of probability spaces, with product measure , there is a lifting for
such that (
1
i
[E]) =
1
i
[E] whenever E and i I is such that (X
i
,
i
,
i
) = (X, , ), writing

i
(x) = x(i) for x

iI
X
i
.
218 Liftings 346Ye
(e) In 346Ya, nd a modication of the countable-conality inductive step of 341G which will ensure that
the lower density obtained in the product satises both conditions of 346G.
(f ) Let (X, , ) be a probability space, I any set, and the product measure on X
I
. Show that there
is a lower density for which is invariant under transpositions of pairs of coordinates.
346Z Problems (a) Let (X
i
,
i
,
i
))
iI
be a family of probability spaces, with product (X, , ). Is
there always a lifting for which respects coordinates in the sense of 346A?
(b) Is there a lower density for the usual measure on 0, 1
N
which is invariant under all permutations
of coordinates?
346 Notes and comments I ought to say at once that in writing this section I have been greatly assisted
by M.R.Burke.
The theorem that every complete probability space has a consistent lifting (346J) is due to Talagrand
82a; it is the inspiration for the whole of the section. Consistent liftings were devised in response to some
very interesting questions (see Talagrand 84, 6) which I do not discuss here; one will be mentioned in
465 in Volume 4. My aim here is rather to suggest further ways in which a lifting on a product space can
be consistent with the product structure. The labour is substantial and the results achieved are curiously
partial. I oer 346Za as the easiest natural question which does not appear amenable to the methods I
describe.
The arguments I use are based on the fact that the translation-invariant measures of 345C already
respect coordinates. Maharams theorem now makes it easy to show that any product of Maharam-type-
homogeneous probability spaces has a lifting which respects coordinates. A kind of projection argument
(346F) makes it possible to obtain a lower density which respects coordinates on any product of probability
spaces (346G). In fact the methods of 341, very slightly rened, automatically produce such lower densities
(346Xg). But the extra power of 346G lies in the condition (ii): if E and F are fully independent in the
sense of being determined by coordinates in disjoint sets, then (E F) = E F, that is, is making a
tentative step towards being a lifting. (Remember that the dierence between a lifting and a lower density
is mostly that a lifting preserves nite unions as well as nite intersections; see 341Xa.) This can also be
achieved by a modication of the previous method, but we have to work harder at one point in the proof
(346Ye).
The next step is to move to liftings which continue, as far as possible, to respect coordinates. Here there
seem to be quite new obstacles, and 346H is the best result I know; the lifting respects individual coordinates,
and also, for a given well-ordering of the index set, initial segments of the coordinates. The treatment of
initial segments makes essential use of the well-ordering, which is what leaves 346Za open.
Finally, if all the factors are identical, we can seek lower densities and liftings which are invariant under
permutation of coordinates. I give 345Xc and 346Xb as examples to show that we must not just assume
that a symmetry in the underlying measure space can be reected in a symmetry of a lifting. The problems
there concern liftings themselves, not lower densities, since we can frequently nd lower densities which share
symmetries (346Xd, 346Yf). (Even for lower densities there seem to be diculties if we are more ambitious
(346Zb).) However a very simple argument (346I) shows that at least we can make each individual coordinate
look more or less the same, as long as we do not investigate its relations with others.
Still on the question of whether, and when, liftings can be good, note 346L/346Xf and 346Xe. The most
natural liftings of Lebesgue measure are necessarily consistent; but the only example we have of a truly
canonical lifting is not consistent in any non-trivial context.
I have deliberately used a variety of techniques here, even though 346H (for instance) has an alternative
proof based on the ideas of 341 (346Ya-346Yb). In particular, I give some of the standard methods of
constructing liftings and lower densities (346D, 346F, 346B, 346Xa, 346Xc). In fact 346D was one of the
elements of Maharams original proof of the lifting theorem (Maharam 58).
351Bc Partially ordered linear spaces 219
Chapter 35
Riesz spaces
The next three chapters are devoted to an abstract description of the function spaces described in
Chapter 24, this time concentrating on their internal structure and their relationships with their associated
measure algebras. I nd that any convincing account of these must involve a substantial amount of general
theory concerning partially ordered linear spaces, and in particular various types of Riesz space or vector
lattice. I therefore provide an introduction to this theory, a kind of appendix built into the middle of the
volume. The relation of this chapter to the next two is very like the relation of Chapter 31 to Chapter 32.
As with Chapter 31, it is not really meant to be read for its own sake; those with a particular interest in
Riesz spaces might be better served by Luxemburg & Zaanen 71, Schaefer 74, Zaanen 83 or my own
book Fremlin 74a.
I begin with three sections in an easy gradation towards the particular class of spaces which we need
to understand: partially ordered linear spaces (351), general Riesz spaces (352) and Archimedean Riesz
spaces (353); the last includes notes on Dedekind ()-complete spaces. These sections cover the fragments
of the algebraic theory of Riesz spaces which I will use. In the second half of the chapter, I deal with normed
Riesz spaces (in particular, L- and M-spaces)(354), spaces of linear operators (355) and dual Riesz spaces
(356).
351 Partially ordered linear spaces
I begin with an account of the most basic structures which involve an order relation on a linear space,
partially ordered linear spaces. As often in this volume, I nd myself impelled to do some of the work in very
much greater generality than is strictly required, in order to show more clearly the nature of the arguments
being used. I give the denition (351A) and most elementary properties (351B-351L) of partially ordered
linear spaces; then I describe a general representation theorem for arbitrary partially ordered linear spaces
as subspaces of reduced powers of R (351M-351Q). I end with a brief note on Archimedean partially ordered
linear spaces (351R).
351A Denition I repeat a denition mentioned in 241E. A partially ordered linear space is a
linear space (U, +, ) over R together with a partial order on U such that
u v = u +w v +w,
u 0, 0 = u 0
for u, v, w U and R.
351B Elementary facts Let U be a partially ordered linear space. We have the following elementary
consequences of the denition above, corresponding to the familiar rules for manipulating inequalities among
real numbers.
(a) For u, v U,
u v = 0 = u + (u) v + (u) = v u = u = 0 +u v u +u = v,
u v = v = u + (v u) v + (v u) = u.
(b) Suppose that u, v U and u v. Then u v for every 0 and v u for every 0. PPP
(i) If 0, then (v u) 0 so v u. (ii) If 0 then ()u ()v so
v = ()v (u) = u. QQQ
(c) If u 0 and in R, then ( )u 0, so u u. If 0 u v in U and 0 in R, then
u u v.
220 Riesz spaces 351C
351C Positive cones Let U be a partially ordered linear space.
(a) I will write U
+
for the positive cone of U, the set u : u U, u 0.
(b) By 351Ba, the ordering is determined by the positive cone U
+
, in the sense that u v v u
U
+
.
(c) It is easy to characterize positive cones. If U is a real linear space, a set C U is the positive cone
for some ordering rendering U a partially ordered linear space i
u +v C, u C whenever u, v C and 0,
0 C, u C & u C = u = 0.
PPP (i) If C = U
+
for some partially ordered linear space ordering on U, then
u, v C = 0 u u +v = u +v C,
u C, 0 = u 0, i.e., u C,
0 0 so 0 C,
u, u C = u = 0 +u (u) +u = 0 u = u = 0.
(ii) On the other hand, if C satises the conditions, dene the relation by writing u v v u C;
then
u u = 0 C so u u for every u U,
if u v and v w then w u = (w v) + (v u) C so u w,
if u v and v u then u v, v u C so u v = 0 and u = v
and is a partial order; moreover,
if u v and w U then (v +w) (u +w) = v u C and u +w v +w,
if u, 0 then u C and u 0,
u 0 u C.
So makes U a partially ordered linear space in which C is the positive cone. QQQ
(d) An incidental useful fact. Let U be a partially ordered linear space, and u U. Then u 0 i
u u. PPP If u 0 then 0 u so u u. If u u then 2u 0 so u =
1
2
2u 0. QQQ
(e) I have called U
+
a positive cone without dening the term cone. I think this is something we can
pass by for the moment; but it will be useful to recognise that U
+
is always convex, for if u, v U
+
and
[0, 1] then u, (1 )v 0 and u + (1 )v U
+
.
351D Suprema and inma Let U be a partially ordered linear space.
(a) The denition of partially ordered linear space implies that u u+w is always an order-isomorphism;
on the other hand, u u is order-reversing, by 351Ba.
(b) It follows that if A U, v U then
sup
uA
v +u = v + sup A if either side is dened,
inf
uA
v +u = v + inf A if either side is dened,
sup
uA
v u = v inf A if either side is dened,
inf
uA
v u = v supA if either side is dened.
351G Partially ordered linear spaces 221
(c) Moreover, we nd that if A, B U and supA and sup B are dened, then sup(A+B) is dened and
equal to supA + supB, writing A + B = u + v : u A, v B as usual. PPP Set u
0
= sup A, v
0
= sup B.
Using (b), we have
u
0
+v
0
= sup
uA
(u +v
0
)
= sup
uA
(sup
vB
(u +v)) = sup(A+B). QQQ
Similarly, if A, B U and inf A, inf B are dened then inf(A+B) = inf A+ inf B.
(d) If > 0 then u u is an order-isomorphism, so we have sup(A) = supA if either side is dened;
similarly, inf(A) = inf A.
351E Linear subspaces If U is a partially ordered linear space, and V is any linear subspace of U,
then V , with the induced linear and order structures, is a partially ordered linear space; this is obvious from
the denition.
351F Positive linear operators Let U and V be partially ordered linear spaces, and write L(U; V ) for
the linear space of all linear operators from U to V . For S, T L(U; V ) say that S T i Su Tu for every
u U
+
. Under this ordering, L(U; V ) is a partially ordered linear space; its positive cone is T : Tu 0 for
every u U
+
. PPP This is an elementary verication. QQQ Note that, for T L(U; V ),
T 0 = Tu Tu +T(v u) = Tv whenever u v in U
= 0 = T0 Tu for every u U
+
= T 0,
so that T 0 i T is order-preserving. In this case we say that T is a positive linear operator.
Clearly ST is a positive linear operator whenever U, V and W are partially ordered linear spaces and
T : U V , S : V W are positive linear operators (cf. 313Ia).
351G Order-continuous positive linear operators: Proposition Let U and V be partially ordered
linear spaces and T : U V a positive linear operator.
(a) The following are equiveridical:
(i) T is order-continuous;
(ii) inf T[A] = 0 in V whenever A U is a non-empty downwards-directed set with inmum 0 in U;
(iii) supT[A] = Tw in V whenever A U
+
is a non-empty upwards-directed set with supremum w in
U.
(b) The following are equiveridical:
(i) T is sequentially order-continuous;
(ii) inf
nN
Tu
n
= 0 in V whenever u
n
)
nN
is a non-increasing sequence in U with inmum 0 in U;
(iii) sup
nN
Tu
n
= Tw in V whenever u
n
)
nN
is a non-decreasing sequence in U
+
with supremum w
in U.
proof (a)(i)(iii) is trivial.
(iii)(ii) Assuming (iii), and given that A is non-empty, downwards-directed and has inmum 0, take
any u
0
A and consider A
t
= u : u A, u u
0
, B = u
0
A
t
. Then A
t
is non-empty, downwards-directed
and has inmum 0, so B is non-empty, upwards-directed and has supremum u
0
(using 351Db); by (iii),
supT[B] = Tu
0
and (inverting again)
inf T[A
t
] = inf T[u
0
B] = inf Tu
0
T[B] = Tu
0
supT[B] = 0.
But (because T is positive) 0 is surely a lower bound for T[A], so it is also the inmum of T[A]. As A is
arbitrary, (ii) is true.
(ii)(i) Suppose now that (ii) is true. () If A U is non-empty, downwards-directed and has
inmum w, then Aw is non-empty, downwards-directed and has inmum 0, so
222 Riesz spaces 351G
inf T[Aw] = 0, inf T[A] = inf(T[Aw] +Tw) = Tw + inf T[Aw] = Tw.
() If A U is non-empty, upwards-directed and has supremum w, then A is non-empty, downwards-
directed and has inmum w, so
supT[A] = inf(T[A]) = inf T[A] = T(w) = Tw.
Putting these together, T is order-continuous.
(b) The arguments are identical, replacing each directed set by an appropriate sequence.
351H Riesz homomorphisms (a) For the sake of a representation theorem below (351Q), I introduce
the following denition. Let U, V be partially ordered linear spaces. A Riesz homomorphism from U to
V is a linear operator T : U V such that whenever A U is a nite non-empty set and inf A = 0 in U,
then inf T[A] = 0 in V . The following facts are now nearly obvious.
(b) Any Riesz homomorphism is a positive linear operator. (For if T is a Riesz homomorphism and u 0,
then inf0, u = 0 so inf0, Tu = 0 and Tu 0.)
(c) Let U and V be partially ordered linear spaces and T : U V a Riesz homomorphism. Then
inf T[A] exists = T(inf A), sup T[A] exists = T(supA)
whenever A U is a nite non-empty set and inf A, supA exist. (Apply the denition in (a) to
A
t
= u inf A : u A, A
tt
= supAu : u A.)
(d) If U, V and W are partially ordered linear spaces and T : U V , S : V W are Riesz homomor-
phisms then ST : U W is a Riesz homomorphism.
351I Solid sets Let U be a partially ordered linear space. I will say that a subset A of U is solid if
A = v : v U, u v u for some u A =

uU
[u, u]
in the notation of 2A1Ab. (I should perhaps remark that while this denition is well established in the case
of Riesz spaces (352), the extension to general partially ordered linear spaces is not standard. See 351Yb
for a warning.)
351J Proposition Let U be a partially ordered linear space and V a solid linear subspace of U. Then
the quotient linear space U/V has a partially ordered linear space structure dened by either of the rules
u

i there is some v V such that u v +w,

(U/V )
+
= u

: u U
+
,
and for this partial ordering on U/V the map u u

: U U/V is a Riesz homomorphism.

proof (a) I had better start by giving priority to one of the descriptions of the relation on U/V ; I choose
the rst. To see that this makes U/V a partially ordered linear space, we have to check the following.
(i) 0 V and u u + 0, so u

for every u U.
(ii) If u
1
, u
2
, u
3
U and u

1
u

2
, u

2
u

3
then there are v
1
, v
2
V such that u
1
u
2
+v
1
, u
2
u
3
+v
2
;
in which case v
1
+v
2
V and u
1
u
3
+v
1
+v
2
, so u

1
u

3
.
(iii) If u, w U and u

, w

then there are v, v

t
V such that u w + v, w u + v
t
. Now
there are v
0
, v
t
0
V such that v
0
v v
0
, v
t
0
v
t
v
t
0
, and in this case v
0
, v
t
0
0 (351Cd), so
v
0
v
t
0
v
t
u w v v
0
+v
t
0
V ,
.
Accordingly u w V and u

= w

. Thus U/V is a partially ordered set.

(iv) If u
1
, u
2
, w U and u

1
u

2
, then there is a v V such that u
1
u
2
+ v, in which case
u
1
+w u
2
+w +v and u

1
+w

2
+w

.
(v) If u U, R and u

0, 0 then there is a v V such that u + v 0; now v V and

u +v 0, so u

= (u)

0.
351N Partially ordered linear spaces 223
Thus U/V is a partially ordered linear space.
(b) Now (U/V )
+
= u

: u 0. PPP If u 0 then of course u

0 because 0 V and u +0 0. On the

other hand, if we have any element p of (U/V )
+
, there are u U, v V such that u

= p and u + v 0;
but now p = (u +v)

is of the required form. QQQ

(c) Finally, u u

is a Riesz homomorphism. PPP Suppose that A U is a non-empty nite set and that
inf A = 0 in U. Then u

0 for every u A, that is, 0 is a lower bound for u

: u A. Let p U/V be
any other lower bound for u

: u A. Express p as w

where w U. For each u A, w

so there is
a v
u
V such that w u +v
u
. Next, there is a v
t
u
V such that v
t
u
v
u
v
t
u
. Set v

uA
v
t
u
V .
Then v
u
v
t
u
v

so w u + v

for every u A, and w v

is a lower bound for A in U. Accordingly

w v

0, w 0 +v

and p = w

0. As p is arbitrary, infu

: u A = 0; as A is arbitrary, u u

is
a Riesz homomorphism. QQQ
351K Lemma Suppose that U is a partially ordered linear space, and that W, V are solid linear
subspaces of U such that W V . Then V
1
= v

: v V is a solid linear subspace of U/W.

proof (i) Because the map u u

is linear, V
1
is a linear subspace of U/W. (ii) If p V
1
, there is a v V
such that p = v

; because V is solid in U, there is a v

0
V such that v
0
v v
0
; now v

0
V
1
and
v

0
p v

0
. (iii) If p V
1
, q U/W and p q p, take v
0
V , u U such that v

0
= p, u

= q.
Because v

0
u

0
, there are w, w
t
W such that v
0
w u v
0
+ w
t
. Now v
0
w, v
0
+ w
t
both belong to V , which is solid, so both are mapped to 0 by the canonical Riesz homomorphism from U
to U/V , and u must also be, that is, u V and q = u

V
1
. (iv) Putting (ii) and (iii) together, V
1
is solid.
351L Products If U
i
)
iI
is any family of partially ordered linear spaces, we have a product linear
space U =

iI
U
i
; if we set u v in U i u(i) v(i) for every i I, U becomes a partially ordered linear
space, with positive cone u : u(i) 0 for every i I. For each i I the map u u(i) : U U
i
is an
order-continuous Riesz homomorphism (in fact, it preserves arbitrary suprema and inma).
351M Reduced powers of R (a) Let X be any set. Then R
X
is a partially ordered linear space if we
say that f g means that f(x) g(x) for every x X, as in 351L. If now T is a lter on X, we have a
corresponding set
V = f : f R
X
, x : f(x) = 0 T;
it is easy to see that V is a linear subspace of R
X
, and is solid because f V i [f[ V . By the reduced
power R
X
[T I shall mean the quotient partially ordered linear space R
X
/V .
(b) Note that for f R
X
,
f

0 in R
X
[T x : f(x) 0 T.
PPP (i) If f

0, there is a g V such that f +g 0; now

x : f(x) 0 x : g(x) = 0 T.
(ii) If x : f(x) 0 T, then x : ([f[ f)(x) = 0 T, so f

= [f[

0. QQQ
351N On the way to the next theorem, the main result (in terms of mathematical depth) of this section,
we need a string of lemmas.
Lemma Let U be a partially ordered linear space. If u, v
0
, . . . , v
n
U are such that u ,= 0 and v
0
, . . . , v
n
0
then there is a linear functional f : U R such that f(u) ,= 0 and f(v
i
) 0 for every i.
proof The point is that at most one of u, u can belong to the convex cone C generated by v
0
, . . . , v
n
,
because this is included in the convex cone set U
+
, and since u ,= 0 at most one of u, u can belong to U
+
.
Now however the Hahn-Banach theorem, in the form 3A5D, tells us that if u / C there is a linear
functional f : U R such that f(u) < 0, f(v
i
) 0 for every i; while if u / C we can get f(u) < 0 and
f(v
i
) 0 for every i. Thus in either case we have a functional of the required type.
224 Riesz spaces 351O
351O Lemma Let U be a partially ordered linear space, and u
0
a non-zero member of U. Then there is
a solid linear subspace V of U such that u
0
/ V and whenever A U is nite, non-empty and has inmum
0 then A V ,= .
proof (a) Let J be the family of all solid linear subspaces of U not containing u
0
. Then any non-empty
totally ordered 1 J has an upper bound

1 in J. By Zorns Lemma, J has a maximal element V say.

This is surely a solid linear subspace of U not containing u
0
.
(b) Now for any w U
+
V there are 0, v V
+
such that w v u
0
w +v. PPP Let V
1
be
u : u U, there are 0, v V
+
such that w v u w +v.
Then it is easy to check that V
1
is a solid linear subspace of U, including V , and containing w; because
w / V , V
1
,= V , so V
1
/ J and u V
1
, as claimed. QQQ
(c) It follows that if A U is nite and non-empty and inf A = 0 in U then AV ,= . PPP??? Otherwise, for
every w A there must be
w
0, v
w
V
+
such that
w
wv
w
u
0

w
w+v
w
. Set = 1+

wA

w
,
v =

wA
v
w
V ; then w v u
0
w + v for every w A. Accordingly
1

(u
0
v) w for every
w A and
1

(u
0
v) 0, so u
0
v. Similarly,
1

(v +u
0
) w for every w A and v u
0
. But (because
V is solid) this means that u
0
V , which is not so. XXXQQQ
Accordingly V has the required properties.
351P Lemma Let U be a partially ordered linear space and u a non-zero element of U, and suppose
that A
0
, . . . , A
n
are nite non-empty subsets of U such that inf A
j
= 0 for every j n. Then there is a
linear functional f : U R such that f(u) ,= 0 and minf[A
j
] = 0 for every j n.
proof By 351O, there is a solid linear subspace V of U such that u / V and A
j
V ,= 0 for every j n.
Give the quotient space U/V its standard partial ordering (351J), and in U/V set C = v

: v

jn
A
j
.
Then C is a nite subset of (U/V )
+
, while u

,= 0, so by 351N there is a linear functional g : U/V R such

that g(u

) ,= 0 but g(p) 0 for every p C. Set f(v) = g(v

) for v U; then f : U R is linear, f(u) ,= 0

and f(v) 0 for every v

jn
A
j
. But also, for each j n, there is a v
j
A
j
V , so that f(v
j
) = 0;
and this means that minf[A
j
] must be 0, as required.
351Q Now we are ready for the theorem.
Theorem Let U be any partially ordered linear space. Then we can nd a set X, a lter T on X and an
injective Riesz homomorphism from U to the reduced power R
X
[T described in 351M.
proof Let X be the set of all linear functionals f : U R; dene : U R
X
by setting (u)(f) = f(u)
for every f X, u U, so that is linear. Let / be the family of non-empty nite sets A U such that
inf A = 0. For A / let F
A
be the set of those f X such that min f[A] = 0. Since 0 F
A
for every
A /, the set
T = F : F X, there are A
0
, . . . , A
n
/ such that F

jn
F
A
j

is a lter on X. Set (u) = (u)

R
X
[T for u U. The : U R
X
[T is an injective Riesz
homomorphism.
PPP (i) is linear because and h h

: R
X
R
x
[T are. (ii) If A /, then F
A
T. So, rst, if
v A, then f : (v)(f) 0 T, so that (v) = (v)

0 in R
X
[T (351Mb). Next, if w R
X
[T
and w (v) for every v A, we can express w as h

where h

(v)

for every v A, that is,

H
v
= f : h(f) (v)(f) T for every v A. But now H = F
A

vA
H
v
T, and for f H we have
h(f) min
vA
f(v) = 0. This means that w = h

0. As w is arbitrary, inf [A] = 0. As A is arbitrary,

is a Riesz homomorphism. (iii) Finally, ??? suppose, if possible, that there is a non-zero u U such that
(u) = 0. Then F = f : f(u) = 0 T, and there are A
0
, . . . , A
n
/ such that F

jn
F
A
j
. By 351P,
there is an f

jn
F
A
j
such that f(u) ,= 0; which is impossible. XXX Accordingly is injective, as claimed.
QQQ
351 Notes Partially ordered linear spaces 225
351R Archimedean spaces (a) For a partially ordered linear space U, the following are equiveridical:
(i) if u, v U are such that nu v for every n N then u 0 (ii) if u 0 in U then inf
>0
u = 0. PPP
(i)(ii) If (i) is true and u 0, then of course u 0 for every > 0; on the other hand, if v u for
every > 0, then nv n
1
n
u = u for every n 1, while of course 0v = 0 u, so v 0. Thus 0 is the
greatest lower bound of u : > 0. (ii)(i) If (ii) is true and nu v for every n N, then 0 v and
u
1
n
v for every n 1. If now > 0, then there is an n 1 such that
1
n
, so that u
1
n
v v (351Bc).
Accordingly u is a lower bound for v : > 0 and u 0. QQQ
(b) I will say that partially ordered linear spaces satisfying the equiveridical conditions of (a) above are
Archimedean.
(c) Any linear subspace of an Archimedean partially ordered linear space, with the induced partially
ordered linear space structure, is Archimedean.
(d) Any product of Archimedean partially ordered linear spaces is Archimedean. PPP If U =

iI
U
i
is
a product of Archimedean spaces, and nu v in U for every n N, then for each i I we must have
nu(i) v(i) for every n, so that u(i) 0; accordingly u 0. QQQ In particular, R
X
is Archimedean for any
set X.
351X Basic exercises >>>(a) Let be any ordinal. The lexicographic ordering on R

is dened by
saying that f g i either f = g or there is a < such that f() = g() for < and f() < g(). Show
that this is a total order on R

which renders R

a partially ordered linear space.

(b) Let U be a partially ordered linear space and V a linear subspace of U. Show that the formulae of
351J dene a partially ordered linear space structure on the quotient U/V i V is order-convex, that is,
u V whenever v
1
, v
2
V and v
1
u v
2
.
(c) Let U
i
)
iI
be a family of partially ordered linear spaces with product U. Dene T
i
: U
i
U by
setting T
i
x = u where u(i) = x, u(j) = 0 for j ,= i. Show that T
i
is an injective order-continuous Riesz
homomorphism.
>>>(d) Let U be a partially ordered linear space and V
i
)
iI
a family of partially ordered linear spaces with
product V . Show that L(U; V ) can be identied, as partially ordered linear space, with

iI
L(U; V
i
).
>>>(e) Show that if U, V are partially ordered linear spaces and V is Archimedean, then L(U; V ) is
Archimedean.
351Y Further exercises (a) Give an example of two partially ordered linear spaces U and V and a
bijective Riesz homomorphism T : U V such that T
1
: V U is not a Riesz homomorphism.
(b) (i) Let U be a partially ordered linear space. Show that U is a solid subset of itself (on the denition
351I) i U = U
+
U
+
. (ii) Give an example of a partially ordered linear space U satisfying this condition
with an element u U such that the intersection of the solid sets containing u is not solid.
(c) Let U be a partially ordered linear space, and suppose that A, B U are two non-empty nite sets
such that () u v = supu, v is dened for every u A, v B ( inf A and inf B and (inf A) (inf B)
are dened. Show that infu v : u A, v B = (inf A) (inf B). (Hint: show that this is true if U = R,
if U = R
X
and if U = R
X
[T, and use 351Q.)
(d) Show that a reduced power R
X
[T, as described in 351M, is totally ordered i T is an ultralter, and
in this case has a natural structure as a totally ordered eld.
351 Notes and comments The idea of partially ordered linear space is a very natural abstraction from
the elementary examples of R
X
and its subspaces, and the only possible diculty lies in guessing the exact
boundary at which ones standard manipulations with such familiar spaces cease to be valid in the general
case. (For instance, most peoples favourite examples are Archimedean, in the sense of 351R, so it is prudent
226 Riesz spaces 351 Notes
to check your intuitions against a non-Archimedean space like that of 351Xa.) There is really no room for
any deep idea to appear in 351B-351F. When I come to what I call Riesz homomorphisms, however (351H),
there are some more interesting possibilities in the background.
I shall not discuss the applications of Theorem 351Q to general partially ordered linear spaces; it is here
for the sake of its application to Riesz spaces in the next section. But I think it is a very striking fact that
not only does any partially ordered linear space U appear as a linear subspace of some reduced power of R,
but the embedding can be taken to preserve any suprema and inma of nite sets which exist in U. This is
in a sense a result of the same kind as the Stone representation theorem for Boolean algebras; it gives us a
chance to conrm that an intuition valid for R or R
X
may in fact apply to arbitrary partially ordered linear
spaces. If you like, this provides a metamathematical foundation for such results as those in 351B. I have to
say that for partially ordered linear spaces it is generally quicker to nd a proof directly from the denition
than to trace through an argument relying on 351Q; but this is not always the case for Riesz spaces. I oer
351Yc as an example of a result where a direct proof does at least call for a moments thought, while the
argument through 351Q is straightforward.
Reduced powers are of course of great importance for other reasons; I mention 351Yd as a hint of what
can be done.
352 Riesz spaces
In this section I sketch those fragments of the theory we need which can be expressed as theorems about
general Riesz spaces or vector lattices. I begin with the denition (352A) and most elementary properties
(352C-352F). In 352G-352J I discuss Riesz homomorphisms and the associated subspaces (Riesz subspaces,
solid linear subspaces); I mention product spaces (352K, 352T) and quotient spaces (352Jb, 352U) and the
form the representation theorem 351Q takes in the present context (352L-352M). Most of the second half
of the section concerns the theory of bands in Riesz spaces, with the algebras of complemented bands
(352Q) and projection bands (352S) and a description of bands generated by upwards-directed sets (352V).
I conclude with a description of f-algebras (352W).
352A I repeat a denition from 241E.
Denition A Riesz space or vector lattice is a partially ordered linear space which is a lattice.
352B Lemma If U is a partially ordered linear space, then it is a Riesz space i sup0, u is dened
for every u U.
proof If U is a lattice, then of course supu, 0 is dened for every u. If supu, 0 is dened for every u,
and v
1
, v
2
are any two members of U, consider w = v
1
+sup0, v
2
v
1
; by 351Db, w = supv
1
, v
2
. Next,
infv
1
, v
2
= supv
1
, v
2

must also be dened in U; as v

1
and v
2
are arbitrary, U is a lattice.
352C Notation In any Riesz space U I will write
u
+
= u 0, u

= (u) 0 = (u)
+
, [u[ = u (u)
where (as in any lattice) u v = supu, v (and u v = infu, v).
I mention immediately a term which will be useful: a family u
i
)
iI
in U is disjoint if [u
i
[ [u
j
[ = 0 for
all distinct i, j I. Similarly, a set C U is disjoint if [u[ [v[ = 0 for all distinct u, v C.
352D Elementary identities Let U be a Riesz space. The translation-invariance of the order, and
its invariance under positive scalar multiplication, reversal under negative multiplication, lead directly to
the following, which are in eect special cases of 351D:
u + (v w) = (u +v) (u +w), u + (v w) = (u +v) (u +w),
352Ec Riesz spaces 227
(u v) = u v, (u v) = u v if 0,
(u v) = (u) (v).
Combining and elaborating on these facts, we get
u
+
u

= (u 0) ((u) 0) = u + (0 (u)) ((u) 0) = u,

u
+
+u

= 2u
+
u = (2u 0) u = u (u) = [u[,
u 0 u 0 u

= 0 u = u
+
u = [u[,
[ u[ = [u[, [ [u[ [ = [u[, [u[ = [[[u[
(looking at the cases 0, 0 separately),
u v +u v = u + (0 (v u)) +v + ((u v) 0)
= u + (0 (v u)) +v ((v u) 0) = u +v,
u v = u + (0 (v u)) = u + (v u)
+
,
u v = u + (0 (v u)) = u (0 (u v)) = u (u v)
+
,
u v =
1
2
(2u 2v) =
1
2
(u +v + (u v) (v u)) =
1
2
(u +v +[u v[),
u v = u +v u v =
1
2
(u +v [u v[),
u
+
u

= u (u) 0 = [u[, u
+
u

= u
+
+u

(u
+
u

) = 0,
[u +v[ = (u +v) ((u) + (v)) ([u[ +[v[) ([u[ +[v[) =[u[ +[v[,
[[u[ [v[[ = ([u[ [v[) ([v[ [u[) [u v[ +[v u[ = [u v[,
[u v[ [u[ +[v[
(because [u[ u v [u[ [v[ [u[ +[v[)
for u, v, w U and R.
352E Distributive laws Let U be a Riesz space.
(a) If A, B U have suprema a, b in U, then C = u v : u A, v B has supremum a b. PPP Of
course uv a b for all u A, v B, so a b is an upper bound for C. Now suppose that c is any upper
bound for C. If u A, v B then
u (u v)
+
= u v c, u c + (u v)
+
c + (a v)
+
(because (a v)
+
= supa v, 0 supu v, 0 = (u v)
+
). As u is arbitrary, a c + (a v)
+
and
a v c. Now turn the argument round:
v = (a v) + (v a)
+
c + (v a)
+
c + (b a)
+
,
and this is true for every v B, so b c + (b a)
+
and a b c. As c is arbitrary, a b = supC, as
claimed. QQQ
(b) Similarly, or applying (a) to A and B, infu v : u A, v B = inf A inf B whenever A,
B U and the right-hand-side is dened.
(c) In particular, U is a distributive lattice (denition: 3A1Ic).
228 Riesz spaces 352F
352F Further identities and inequalities At a slightly deeper level we have the following facts.
Proposition Let U be a Riesz space.
(a) If u, v, w 0 in U then u (v +w) (u v) + (u w).
(b) If u
0
, . . . , u
n
U and [u
i
[ [u
j
[ = 0 for i ,= j, then [

n
i=0

i
u
i
[ =

n
i=0
[
i
[[u
i
[ for any
0
, . . . ,
n
.
(c) If u, v U then
u
+
v
+
(u +v)
+
u
+
+v
+
.
(d) If u
0
, . . . , u
m
, v
0
, . . . , v
n
U
+
and

m
i=0
u
i
=

n
j=0
v
j
, then there is a family w
ij
)
im,jn
in U
+
such that

m
i=0
w
ij
= v
j
for every j n and

n
j=0
w
ij
= u
i
for every i m.
proof (a)
u (v +w) [(u +w) (v +w)] u
[(u v) +w] [(u v) +u] = (u v) + (u w).
(b)(i)() A simple induction, using (a) for the inductive step, shows that if v
0
, . . . , v
m
, w
0
, . . . , w
n
are
non-negative then

m
i=0
v
i

n
j=0
w
j

m
i=0

n
j=0
v
i
w
j
. () Next, if u v = 0 then
(u v)
+
= u (u v) = u, (u v)

= (v u)
+
= v (v u) = v,
[u v[ = (u v)
+
+ (u v)

= u +v = [u +v[,
so if [u[ [v[ = 0 then
(u
+
+v
+
) (u

+v

) (u
+
u

) + (u
+
v

) + (v
+
u

) + (v
+
v

)
0 + ([u[ [v[) + ([v[ [u[) + 0 = 0
and
[u +v[ = [(u
+
+v
+
) (u

+v

)[ = u
+
+v
+
+u

+v

= [u[ +[v[.
() Finally, if [u[ [v[ = 0 and , R,
[u[ [v[ = [[[u[ [[[v[ ([[ +[[)[u[ ([[ +[[)[v[ = ([[ +[[)([u[ [v[) = 0.
(ii) We may therefore proceed by induction. The case n = 0 is trivial. For the inductive step to n +1,
setting u
t
i
=
i
u
i
we have [u
t
i
[ [u
t
j
[ = 0 for all i ,= j, by (i-). By (i-),
[u
t
n+1
[ [

n
i=0
u
t
i
[ [u
t
n+1
[

n
i=0
[u
t
i
[

n
i=0
[u
t
n+1
[ [u
t
i
[ = 0,
so by (i-) and the inductive hypothesis
[

n+1
i=0
u
t
i
[ = [u
t
n+1
[ +[

n
i=0
u
t
i
[ =

n+1
i=0
[u
t
i
[
as required.
(c) By 352E,
u
+
v
+
= (u 0) (v 0) = (u v) 0.
Now
u v =
1
2
(u +v [u v[)
1
2
(u +v +[u +v[) = (u +v)
+
,
and of course 0 (u +v)
+
, so u
+
v
+
(u +v)
+
.
For the other inequality we need only note that u + v u
+
+ v
+
(because u u
+
, v v
+
) and
0 u
+
+v
+
.
(d) Write w for the common value of

m
i=0
u
i
and

n
j=0
v
j
.
Induce on k = #((i, j) : i m, j n, u
i
v
j
> 0). If k = 0, that is, u
i
v
j
= 0 for all i, j, then (by
(a), used repeatedly) we must have w w = 0, that is, w = 0, and we can take w
ij
= 0 for all i, j. For the
inductive step to k 1, take i

, j

such that w = u
i
v
j
> 0. Set
352Ic Riesz spaces 229
u
i
= u
i
w, u
i
= u
i
for i ,= i

,
v
j
= v
j
w, v
j
= v
j
for j ,= j

.
Then

m
i=0
u
i
=

n
j=0
v
j
= w w and u
i
v
j
u
i
v
j
for all i, j, while u
i
v
j
= 0; so that
#((i, j) : u
i
v
j
> 0) < k.
By the inductive hypothesis, there are w
ij
0, for i m and j n, such that u
i
=

n
j=0
w
ij
for each
i, v
j
=

m
i=0
w
ij
for each j. Set w
i

j
= w
i

j
+ w, w
ij
= w
ij
for (i, j) ,= (i

, j

); then u
i
=

n
j=0
w
ij
,
v
j
=

m
i=0
w
ij
so the induction proceeds.
352G Riesz homomorphisms: Proposition Let U be a Riesz space, V a partially ordered linear
space and T : U V a linear operator. Then the following are equiveridical:
(i) T is a Riesz homomorphism in the sense of 351H;
(ii) (Tu)
+
= supTu, 0 is dened and equal to T(u
+
) for every u U;
(iii) supTu, Tu is dened and equal to T[u[ for every u U;
(iv) infTu, Tv = 0 in V whenever u v = 0 in U.
proof (i)(iii) and (i)(iv) are special cases of 351Hc. For (iii)(ii) we have
supTu, 0 =
1
2
Tu + sup
1
2
Tu,
1
2
Tu =
1
2
Tu +
1
2
T[u[ = T(u
+
).
For (ii)(i), argue as follows. If (ii) is true and u, v U, then
Tu Tv = infTu, Tv = Tu + inf0, Tv Tu = Tu sup0, T(u v)
is dened and equal to
Tu T((u v)
+
) = T(u (u v)
+
) = T(u v).
Inducing on n,
inf
in
Tu
i
= T(inf
in
u
i
)
for all u
0
, . . . , u
n
U; in particular, if inf
in
u
i
= 0 then inf
in
Tu
i
= 0; which is the denition I gave of
Riesz homomorphism.
Finally, for (iv)(ii), we know from (iv) that 0 = infT(u
+
), T(u

), so T(u
+
) = inf0, Tu and
T(u
+
) = sup0, Tu.
352H Proposition If U and V are Riesz spaces and T : U V is a bijective Riesz homomorphism,
then T is a partially-ordered-linear-space isomorphism, and T
1
: V U is a Riesz homomorphism.
proof Use 352G(ii). If v V , set u = T
1
v; then T(u
+
) = v
+
so T
1
(v
+
) = u
+
= (T
1
v)
+
. Thus T
1
is
a Riesz homomorphism; in particular, it is order-preserving, so T is an isomorphism for the order structures
as well as for the linear structures.
352I Riesz subspaces (a) If U is a partially ordered linear space, a Riesz subspace of U is a linear
subspace V such that supu, v and infu, v are dened in U and belong to V for every u, v V . In this
case they are the supremum and inmum of u, v in V , so V , with the induced order and linear structure,
is a Riesz space in its own right, and the embedding map u u : V U is a Riesz homomorphism.
(b) Generally, if U is a Riesz space, V is a partially ordered linear space and T : U V is a Riesz
homomorphism, then T[U] is a Riesz subspace of V (because, by 351Hc, Tu Tu
t
= T(u u
t
), T(u u
t
) =
Tu Tu
t
belong to T[U] for all u, u
t
U).
(c) If U is a Riesz space and V is a linear subspace of U, then V is a Riesz subspace of U i [u[ V for
every u V . PPP In this case,
u v =
1
2
(u +v +[u v[), u v =
1
2
(u +v [u v[
belong to V for all u, v V . QQQ
230 Riesz spaces 352J
352J Solid subsets (a) If U is a Riesz space, a subset A of U is solid (in the sense of 351I) i v A
whenever u A and [v[ [u[. PPP () If A is solid, u V and [v[ [u[, then there is some w A such
that w u w; in this case [v[ [u[ w and w v w and v A. () Suppose that A satises
the condition. If u A, then [u[ A and [u[ u [u[. If w A and w u w then u w,
[u[ w = [w[ and u A. Thus A is solid. QQQ In particular, if A is solid, then v A i [v[ A.
For any set A U, the set
u : there is some v A such that [u[ [v[
is a solid subset of U, the smallest solid set including A; we call it the solid hull of A in U.
Any solid linear subspace of U is a Riesz subspace (use 352Fc). If V U is a Riesz subspace, then the
solid hull of V in U is
u : there is some v V such that [u[ v
and is a solid linear subspace of U.
(b) If T is a Riesz homomorphism from a Riesz space U to a partially ordered linear space V , then its
kernel W is a solid linear subspace of U. PPP If u W and [v[ [u[, then T[u[ = supTu, T(u) = 0, while
[u[ v [u[, so that 0 Tv 0 and v W. QQQ Now the quotient space U/W, as dened in 351J,
is a Riesz space, and is isomorphic, as partially ordered linear space, to the Riesz space T[U]. PPP Because
U/W is the linear space quotient of V by the kernel of the linear operator T, we have an induced linear
space isomorphism S : U/W T[U] given by setting Su

= Tu for every u U. If p 0 in U/W there is

a u U
+
such that u

= p (351J), so that Sp = Tu 0. On the other hand, if p U/W and Sp 0, take

u V such that u

= p. We have
T(u
+
) = (Tu)
+
= (Sp)
+
= Sp = Tu,
so that T(u

) = Tu
+
Tu = 0 and u

W, p = (u
+
)

0. Thus Sp 0 i p 0, and S is a
partially-ordered-linear-space isomorphism. QQQ
(c) Because a subset of a Riesz space is a solid linear subspace i it is the kernel of a Riesz homomorphism,
such subspaces are sometimes called ideals.
352K Products If U
i
)
iI
is any family of Riesz spaces, then the product partially ordered linear space
U =

iI
U
i
(351L) is a Riesz space, with
u v = u(i) v(i))
iI
, u v = u(i) v(i))
iI
, [u[ = [u(i)[)
iI
for all u, v U.
352L Theorem Let U be any Riesz space. Then there are a set X, a lter T on X and a Riesz subspace
of the Riesz space R
X
[T (351M) which is isomorphic, as Riesz space, to U.
proof By 351Q, we can nd such X and T and an injective Riesz homomorphism T : U R
X
[T. By
352K, or otherwise, R
X
is a Riesz space; by 352Ib, R
X
[T is a Riesz space (recall that it is a quotient of R
X
by a solid linear subspace, as explained in 351M); by 352I, T[U] is a Riesz subspace of R
X
[T; and by 352H
it is isomorphic to U.
352M Corollary Any identity involving the operations +, , , ,
+
,

, [ [ and scalar multiplication,
and the relation , which is valid in R, is valid in all Riesz spaces.
Remark I suppose some would say that a strict proof of this must begin with a formal description of what
the phrase any identity involving the operations. . . means. However I think it is clear in practice what is
involved. Given a proposed identity like
0

n
i=0
[
i
[[u
i
[ [

n
i=0

i
u
i
[

i,=j
([
i
[ +[
j
[)([u
i
[ [u
j
[),
(compare 352Fb), then to check that it is valid in all Riesz spaces you need only check (i) that it is true in R
(ii) that it is true in R
X
(iii) that it is true in any R
X
[T (iv) that it is true in any Riesz subspace of R
X
[T;
and you can hope that the arguments for (ii)-(iv) will be nearly trivial, since (ii) is generally nothing but a
coordinate-by-coordinate repetition of (i), and (iii) and (iv) involve only transformations of the formula by
Riesz homomorphisms which preserve its structure.
352Oc Riesz spaces 231
352N Order-density and order-continuity Let U be a Riesz space.
(a) Denition A Riesz subspace V of U is quasi-order-dense if for every u > 0 in U there is a v V
such that 0 < v u; it is order-dense if u = supv : v V, 0 v u for every u U
+
.
(b) If U is a Riesz space and V is a quasi-order-dense Riesz subspace of U, then the embedding V U
is order-continuous. PPP Let A V be a non-empty set such that inf A = 0 in V . ??? If 0 is not the inmum
of A in U, then there is a u > 0 such that u is a lower bound for A in U; now there is a v V such that
0 < v u, and v is a lower bound for A in V which is strictly greater than 0. XXX Thus 0 = inf A in U. As
A is arbitrary, the embedding is order-continuous. QQQ
(c) (i) If V U is an order-dense Riesz subspace, it is quasi-order-dense. (ii) If V is a quasi-order-dense
Riesz subspace of U and W is a quasi-order-dense Riesz subspace of V , then W is a quasi-order-dense Riesz
subspace of U. (iii) If V is an order-dense Riesz subspace of U and W is an order-dense Riesz subspace of
V , then W is an order-dense Riesz subspace of U. (Use (b).) (iv) If V is a quasi-order-dense solid linear
subspace of U and W is a quasi-order-dense Riesz subspace of U then V W is quasi-order-dense in V ,
therefore in U.
(d) I ought somewhere to remark that a Riesz homomorphism, being a lattice homomorphism, is order-
continuous i it preserves arbitrary suprema and inma; compare 313L(b-iv) and (b-v).
(e) If V is a Riesz subspace of U, we say that it is regularly embedded in U if the identity map from
V to U is order-continuous, that is, whenever A V is non-empty and has inmum 0 in V , then 0 is still its
greatest lower bound in U. Thus quasi-order-dense Riesz subspaces and solid linear subspaces are regularly
embedded.
352O Bands Let U be a Riesz space.
(a) Denition A band or normal subspace of U is an order-closed solid linear subspace.
(b) If V U is a solid linear subspace then it is a band i supA V whenever A V
+
is a non-empty,
upwards-directed subset of V with a supremum in U. PPP Of course the condition is necessary; I have to show
that it is sucient. (i) Let A V be any non-empty upwards-directed set with a supremum in V . Take any
u
0
A and set A
1
= uu
0
: u A, u u
0
. Then A
1
is a non-empty upwards-directed subset of V
+
, and
u
0
+A
1
= u : u A, u u
0
has the same upper bounds as A, so supA
1
= sup Au
0
is dened in U and
belongs to V . Now supA = u
0
+sup A
1
also belongs to V . (ii) If A V is non-empty, downwards-directed
and has an inmum in U, then A V is upwards-directed, so inf A = sup(A) belongs to V . Thus V is
order-closed. QQQ
(c) For any set A U set A

= v : v U, [u[ [v[ = 0 for every u A. Then A

is a band. PPP (i) Of

course 0 A

. (ii) If v, w A

and u A, then
[u[ [v +w[ ([u[ [v[) + ([u[ [w[) = 0,
so v +w A

. (iii) If v A

and [w[ [v[ then

0 [u[ [w[ [u[ [v[ = 0
for every u A, so w A

. (iv) If v A

then nv A

for every n, by (ii). So if R, take n N such

that [[ n; then
[v[ = [[[v[ n[v[ A

and v A

. Thus A

is a solid linear subspace of U. (v) If B (A

)
+
is non-empty and upwards-directed
and has a supremum w in U, then
[u[ [w[ = [u[ w = sup
vB
[u[ v = 0
by 352Ea, so w A

. Thus A

is a band. QQQ
232 Riesz spaces 352Od
(d) For any A U, A (A

. Also B

whenever A B. So
A

and A

= A

.
(e) If W is another Riesz space and T : U W is an order-continuous Riesz homomorphism then
its kernel is a band. (For 0 is order-closed in W and the inverse image of an order-closed set by an
order-continuous order-preserving function is order-closed.)
352P Complemented bands Let U be a Riesz space. A band V U is complemented if V

= V ,
that is, if V is of the form A

for some A U (352Od). In this case its complement is the complemented

band V

.
352Q Theorem In any Riesz space U, the set C of complemented bands forms a Dedekind complete
Boolean algebra, with
V
C
W = V W, V
C
W = (V +W)

,
1
C
= U, 0
C
= 0, 1
C
\
C
V = V

,
V
C
W V W
for V , W C.
proof To show that C is a Boolean algebra, I use the identication of Boolean algebras with complemented
distributive lattices (311L).
(a) Of course C is partially ordered by . If V , W C then
V W = V

= (V

C,
and V W must be infV, W in C. The map V V

: C C is an order-reversing bijection, so that

V W i W

and V W = supV, W will be (V

; thus C is a lattice. Note also that

V W must be the smallest complemented band including V +W, that is, it is (V +W)

.
(b) If V
1
, V
2
, W C then (V
1
V
2
) W = (V
1
W) (V
2
W). PPP Of course (V
1
V
2
) W
(V
1
W) (V
2
W). ??? Suppose, if possible, that there is a u (V
1
V
2
) W ((V
1
W) (V
2
W)). Then
u / ((V
1
W)

(V
2
W)

, so there is a v (V
1
W)

(V
2
W)

such that u
1
= [u[ [v[ > 0. Now
u
1
V
1
V
2
= (V

1
V

2
)

so u
1
/ V

1
V

2
; say u
1
/ V

j
, and there is a v
j
V
j
such that u
2
= u
1
[v
j
[ > 0.
In this case we still have u
2
(V
j
W)

, because u
2
[v[, but also u
2
V
j
and u
2
W because u
2
[u[;
but this means that u
2
= u
2
u
2
= 0, which is absurd. XXX Thus (V
1
V
2
) W (V
1
W) (V
2
W) and
the two are equal. QQQ
(c) Now if V C,
V V

= 0
is the least member of C, because if v V V

then [v[ = [v[ [v[ = 0. By 311L, C has a Boolean algebra

structure, with the Boolean relations described; by 312L, this structure is uniquely dened.
(d) Finally, if 1 C is non-empty, then

1 = (

V \
V

C
and is inf 1 in C. So C is Dedekind complete.
352R Projection bands Let U be a Riesz space.
(a) A projection band in U is a set V U such that V + V

= U. In this case V is a complemented

band. PPP If v V

then v is expressible as v
1
+ v
2
where v
1
V , v
2
V

. Now [v[ = [v
1
[ + [v
2
[ [v
2
[
(352Fb), so
352S Riesz spaces 233
[v
2
[ = [v
2
[ [v
2
[ [v
2
[ [v[ = 0
and v = v
1
V . Thus V = V

is a complemented band. QQQ Observe that U = V

+V

so V

is also a
projection band.
(b) Because V V

is always 0, we must have U = V V

for any projection band V U; accordingly

there is a corresponding band projection P
V
: U U dened by setting P(v + w) = v whenever v V ,
w V

. In this context I will say that v is the component of v +w in V . The kernel of P is V

, the set
of values is V , and P
2
= P. Moreover, P is an order-continuous Riesz homomorphism. PPP (i) P is a linear
operator because V and V

are linear subspaces. (ii) If v V , w V

then [v+w[ = [v[ +[w[, by 352Fb, so

P[v +w[ = [v[ = [P(v +w)[; consequently P is a Riesz homomorphism (352G). (iii) If A U is downwards-
directed and has inmum 0, then Pu u for every u A, so inf P[A] = 0; thus P is order-continuous.
QQQ
(c) Note that for any band projection P, and any u U, we have [Pu[ [u Pu[ = 0, so that [u[ =
[Pu[ +[u Pu[ and (in particular) [Pu[ [u[; consequently P[W] W for any solid linear subspace W of
U.
(d) A linear operator P : U U is a band projection i Pu (u Pu) = 0 for every u U
+
. PPP I
remarked in (c) that the condition is satised for any band projection. Now suppose that P has the property.
(i) For any u U
+
, Pu 0 and u Pu 0; in particular, P is a positive linear operator. (ii) If u, v U
+
then u Pu (u +v) P(u +v), so
Pv (u Pu) P(u +v) ((u +v) P(u +v)) = 0
and Pv (u Pu) = 0. (iii) If u, v U then [Pv[ P[v[, [u Pu[ [u[ P[u[ (because w w Pw is a
positive linear operator), so
[Pv[ [u Pu[ P[v[ ([u[ P[u[) = 0.
(iv) Setting V = P[U], we see that u Pu V

for every u U, so that

u = u + (u Pu) V +V

for every u, and U = V + V

; thus V is a projection band. (v) Since Pu V and u Pu V

for every
u U, P is the band projection onto V . QQQ
352S Proposition Let U be any Riesz space.
(a) The family B of projection bands in U is a subalgebra of the Boolean algebra C of complemented
bands in U.
(b) For V B let P
V
: U V be the corresponding projection. Then for any e U
+
,
P
V W
e = P
V
e P
W
e = P
V
P
W
e, P
V W
e = P
V
e P
W
e
for all V , W B. In particular, band projections commute.
(c) If V B then the algebra of projection bands of V is just the principal ideal of B generated by V .
proof (a) Of course 0
C
= 0 B. If V B then V

= 1
C
\ V belongs to B. If now W is another member
of B, then
(V W) + (V W)

(V W) +V

+W

.
But if u U then we can express u as v +v
t
, where v V and v
t
V

, and v as w+w
t
, where w W and
w
t
W

; and as [w[ [v[, we also have w V , so that

u = w +v
t
+w
t
(V W) +V

+W

.
This shows that V W B. Thus B is closed under intersection and complements and is a subalgebra of
C.
(b) If V , W B and e U
+
, we have e = e
1
+e
2
+e
3
+e
4
where
e
1
= P
W
P
V
e V W, e
2
= P
W
P
V
e V W

,
234 Riesz spaces 352S
e
3
= P
W
P
V
e V

W, e
4
= P
W
P
V
e V

,
e
1
+e
2
= P
V
e, e
1
+e
3
= P
W
e.
Now e
2
+ e
3
+ e
4
belongs to (V W)

, so e
1
must be the component of e in V W; similarly e
4
is the
component of e in V

, and e
1
+e
2
+e
3
is the component of e in V W. But as e
2
e
3
= 0, we have
P
V W
e = e
1
= (e
1
+e
2
) (e
1
+e
3
) = P
V
e P
W
e,
P
V W
e = e
1
+e
2
+e
3
= (e
1
+e
2
) (e
1
+e
3
) = P
V
e P
W
e,
as required.
It follows that
P
V
P
W
= P
V W
= P
WV
= P
W
P
V
.
(c) If V , W B and W V , then of course W is a band in the Riesz space V (because V is order-closed
in U, so that for any set A W its supremum in U will be its supremum in V ). For any v V , we have an
expression of it as w +w
t
, where w W and w
t
W

, taken in U; but as [w[ +[w

t
[ = [w +w
t
[ = [v[ V ,
w
t
belongs to V , and is in W

V
, the band in V orthogonal to W. Thus W +W

V
= V and W is a projection
band in V . Conversely, if W is a projection band in V , then W

(taken in U) includes W

V
+V

, so that
W +W

W +W

V
+V

= V +V

= U
and W B.
Thus the algebra of projection bands in V is, as a set, equal to the principal ideal B
V
; because their
orderings agree, or otherwise, their Boolean algebra structures coincide.
352T Products again (a) If U =

iI
U
i
is a product of Riesz spaces, then for any J I we have a
subspace
V
J
= u : u U, u(i) = 0 for all i I J
of U, canonically isomorphic to

iJ
U
i
. Each V
J
is a projection band, its complement being V
I\J
; the map
J V
J
is a Boolean homomorphism from TI to the algebra B of projection bands in U, and V
i]
)
iI
is a
partition of unity in B.
(b) Conversely, if U is a Riesz space and (V
0
, . . . , V
n
) is a nite partition of unity in the algebra B of
projection bands in U, then every element of U is uniquely expressible as

n
i=0
u
i
where u
i
V
i
for each i.
(Induce on n.) This decomposition corresponds to a Riesz space isomorphism between U and

in
V
i
.
352U Quotient spaces (a) If U is a Riesz space and V is a solid linear subspace, then the quotient
partially ordered linear space U/V (351J) is a Riesz space; if U and W are Riesz spaces and T : U W a
Riesz homomorphism, then the kernel V of T is a solid linear subspace of U and the Riesz subspace T[U] of
W is isomorphic to U/V (352Jb).
(b) Suppose that U is a Riesz space and V a solid linear space. Then the canonical map from U to U/V
is order-continuous i V is a band. PPP (i) If u u

is order-continuous, its kernel V is a band, by 352Oe.

(ii) If V is a band, and A U is non-empty and downwards-directed and has inmum 0, let p U/V
be any lower bound for u

: u A. Express p as w

. Then ((u w)

= (u

= 0, that is,
(w u)
+
= (u w)

V for every u A. But this means that

w
+
= sup
uA
(w u)
+
V , p
+
= (w
+
)

= 0,
that is, p 0. As p is arbitrary, inf
uA
u

= 0; as A is arbitrary, u u

is order-continuous. QQQ
352V Principal bands Let U be a Riesz space. Evidently the intersection of any family of Riesz
subspaces of U is a Riesz subspace, the intersection of any family of solid linear subspaces is a solid linear
subspace, the intersection of any family of bands is a band; we may therefore speak of the band generated
by a subset A of U, the intersection of all the bands including A. Now we have the following description of
the band generated by a single element.
352W Riesz spaces 235
Lemma Let U be a Riesz space.
(a) If A U
+
is upwards-directed and 2w A for every w A, then an element u of U belongs to the
band generated by A i [u[ = sup
wA
[u[ w.
(b) If u U and w U
+
, then u belongs to the band of U generated by w i [u[ = sup
nN
[u[ nw.
proof (a) Let W be the band generated by A and W
t
the set of elements of U satisfying the condition.
(i) If u W
t
then [u[ w W for every w A, because W is a solid linear subspace; because W is
also order-closed, [u[ and u belong to W. Thus W
t
W.
(ii) Now W
t
is a band.
PPP() If u W
t
and [v[ [u[ then
sup
wA
[v[ w = sup
wA
[v[ [u[ w = [v[ sup
wA
[u[ w = [v[ [u[ = [v[
by 352Ea, so v W
t
.
() If u, v W
t
then, for any w
1
, w
2
A there is a w A such that w w
1
w
2
. Now
w
1
+w
2
2w A, and
([u[ +[v[) 2w ([u[ w
1
) + ([v[ w
2
).
So any upper bound for ([u[+[v[)w : w A must also be an upper bound for [u[w : w A+[v[w :
w A and therefore greater than or equal to
sup([u[ w : w A +[v[ w : w A) = sup
wA
[u[ w + sup
wA
[v[ w = [u[ +[v[
(351Dc). But this means that sup
wA
([u[ +[v[) w must be [u[ +[v[, and [u[ +[v[ belongs to W
t
; it follows
from (i) that u +v belongs to W
t
.
() Just as in 352Oc, we now have
nu W
t
for every n N, u W
t
,
and therefore u W
t
for every R, u W
t
, since [u[ [nu[ if [[ n. Thus W
t
is a solid linear
subspace of U.
() Now suppose that C (W
t
)
+
has a supremum v in U. Then any upper bound of vw : w A
must also be an upper bound of u w : u C, w A and greater than or equal to u = sup
wA
u w for
every u C, therefore greater than or equal to v = sup C. Thus v = sup
wA
v w and v W
t
. As C is
arbitrary, W
t
is a band (352Ob). QQQ
(iii) Since A is obviously included in W
t
, W
t
must include W; putting this together with (i), W = W
t
,
as claimed.
(b) Apply (a) with A = nw : n N.
352W f-algebras Some of the most important Riesz spaces have multiplicative structures as well as
their order and linear structures. A particular class of these structures appears suciently often for it to be
useful to develop a little of its theory. The following denition is a common approach.
(a) Denition An f-algebra is a Riesz space U with a multiplication : U U U such that
u (v w) = (u v) w,
(u +v) w = (u w) + (v w), u (v +w) = (u v) + (u w),
(u v) = (u) v = u (v)
for all u, v, w U and R, and
u v 0 whenever u, v 0,
if u v = 0 then (u w) v = (w u) v = 0 for every w 0.
An f-algebra is commutative if u v = v u for all u, v.
236 Riesz spaces 352Wb
(b) Let U be an f-algebra.
(i) If u v = 0 in U, then u v = 0. PPP u (u v) = 0 so (u v) (u v) = 0. QQQ
(ii) u u 0 for every u U. PPP
(u
+
u

) (u
+
u

) = u
+
u
+
u
+
u

u
+
+u

= u
+
u
+
+u

0. QQQ
(iii) If u, v U then [u v[ = [u[ [v[. PPP u
+
v
+
, u
+
v

, u

v
+
and u
+
u

are disjoint, so
[u v[ = [u
+
v
+
u
+
v

v
+
+u

[
= u
+
v
+
+u
+
v

+u

v
+
+u

= [u[ [v[
by 352Fb. QQQ
(iv) If v U
+
the maps u u v, u v u : U U are Riesz homomorphisms. PPP The rst four
clauses of the denition in (a) ensure that they are linear operators. If u U, then
[u[ v = [u v[, v [u[ = [v u[
by (iii), so we have Riesz homomorphisms, by 352G(iii). QQQ
(c) Let U
i
)
iI
be a family of f-algebras, with Riesz space product U (352K). If we set u v = u(i)
v(i))
iI
for all u, v U, then U becomes an f-algebra.
352X Basic exercises >>>(a) Let U be any Riesz space. Show that [u
+
v
+
[ [u v[ for all u, v U.
>>>(b) Let U, V be a Riesz spaces and T : U V a linear operator. Show that the following are
equiveridical: (i) T is a Riesz homomorphism; (ii) T(uv) = TuTv for all u, v U; (iii) T(uv) = TuTv
for all u, v U; (iv) [Tu[ = T[u[ for every u U.
(c) Let U be a Riesz space and V a solid linear subspace; for u U write u

for the corresponding element

of U/V . Show that if A U is solid then u

: u A is solid in U/W.
(d) Let U and V be Riesz spaces and T : U V a Riesz homomorphism with kernel W. Show that if W
is a band in U and T[U] is regularly embedded in V then T is order-continuous.
(e) Give U = R
2
its lexicographic ordering (351Xa). Show that it has a band V which is not complemented.
(f ) Let U be a Riesz space, C the algebra of complemented bands in U. Show that for any V C the
algebra of complemented bands of V is just the principal ideal of C generated by V .
>>>(g) Let U = C([0, 1]) be the space of continuous functions from [0, 1] to R, with its usual linear and
order structures, so that it is a Riesz subspace of R
[0,1]
. Set V = u : u U, u(t) = 0 if t
1
2
. Show that
V is a band in U and that V

= u : u(t) = 0 if t
1
2
, so that V is complemented but is not a projection
band.
(h) Show that the Boolean homomorphism J V
J
: TI B of 352Ta is order-continuous.
(i) Let U be a Riesz space and A U
+
an upwards-directed set. Show that the band generated by A is
u : [u[ = sup
nN,wA
[u[ nw.
>>>(j) (i) Let X be any set. Setting (u v)(x) = u(x)v(x) for u, v R
X
, x X, show that R
X
is a
commutative f-algebra. (ii) With the same denition of , show that

(X) is an f-algebra. (iii) If X is

a topological space, show that C(X), C
b
(X) are f-algebras. (iv) If (X, , ) is a measure space, show that
L
0
(), L

() (241, 243) are f-algebras.

353A Archimedean and Dedekind complete Riesz spaces 237
(k) Let U R
Z
be the set of sequences u such that n : u(n) ,= 0 is bounded above in Z. For u, v U
(i) say that u v if either u = v or there is an n Z such that u(n) < v(n), u(i) = v(i) for every i > n (ii)
say that (u v)(n) =

i=
u(i)v(n i) for every n Z. Show that U is an f-algebra under this ordering
and multiplication.
(l) Let U be an f-algebra. (i) Show that any complemented band of U is an ideal in the ring (U, +, ).
(ii) Show that if P : U U is a band projection, then P(u v) = Pu Pv for every u, v U.
(m) Let U be an f-algebra with multiplicative identity e. Show that ue
1

u
2
for every u U, > 0.
(Hint: (u
+
e)
2
0.)
352 Notes and comments In this section we begin to see a striking characteristic of the theory of Riesz
spaces: repeated reections of results in Boolean algebra. Without spelling out a complete list, I mention
the distributive laws (313Bc, 352Ea) and the behaviour of order-continuous homomorphisms (313Pa, 352N,
352Oe, 352Ub, 352Xd). Riesz subspaces correspond to subalgebras, solid linear subspaces to ideals and
Riesz homomorphisms to Boolean homomorphisms. We even have a correspondence, though a weaker one,
between the representation theorems available; every Boolean algebra is isomorphic to a subalgebra of a
power of Z
2
(311D-311E), while every Riesz space is isomorphic to a Riesz subspace of a quotient of a power
of R (352L). It would be a closer parallel if every Riesz space were embeddable in some R
X
; I must emphasize
that the dierences are as important as the agreements. Subspaces of R
X
are of great importance, but are
by no means adequate for our needs. And of course the details for instance, the identities in 352D-352F,
or 352V frequently involve new techniques in the case of Riesz spaces. Elsewhere, as in 352G, I nd myself
arguing rather from the opposite side, when applying results from the theory of general partially ordered
linear spaces, which has little to do with Boolean algebra.
In the theory of bands in Riesz spaces corresponding to order-closed ideals in Boolean algebras we have
a new complication in the form of bands which are not complemented, which does not arise in the Boolean
algebra context; but it disappears again when we come to specialize to Archimedean Riesz spaces (353B).
(Similarly, order-density and quasi-order-density coincide in both Boolean algebras (313K) and Archimedean
Riesz spaces (353A).) Otherwise the algebra of complemented bands in a Riesz space looks very like the
algebra of order-closed ideals in a Boolean algebra (314Yh, 352Q). The algebra of projection bands in a
Riesz space (352S) would correspond, in a Boolean algebra, to the algebra itself.
I draw your attention to 352H. The result is nearly trivial, but it amounts to saying that the theory of
Riesz spaces will be algebraic, like the theories of groups or linear spaces, rather than analytic, like the
theories of partially ordered linear spaces or topological spaces, in which we can have bijective morphisms
which are not isomorphisms.
353 Archimedean and Dedekind complete Riesz spaces
I take a few pages over elementary properties of Archimedean and Dedekind ()-complete Riesz spaces.
353A Proposition Let U be an Archimedean Riesz space. Then every quasi-order-dense Riesz subspace
of U is order-dense.
proof Let V U be a quasi-order-dense Riesz subspace, and u 0 in U. Set A = v : v V, v u. ???
Suppose, if possible, that u is not the least upper bound of A. Then there is a u
1
< u such that v u
1
for every v A. Because 0 A, u
1
0. Because V is quasi-order-dense, there is a v > 0 in V such that
v u u
1
. Now nv u
1
for every n N. PPP Induce on n. For n = 0 this is trivial. For the inductive step,
given nv u
1
, then (n+1)v u
1
+v u, so (n+1)v A and (n+1)v u
1
. Thus the induction proceeds.
QQQ But this is impossible, because v > 0 and U is supposed to be Archimedean. XXX
So u = sup A. As u is arbitrary, V is order-dense.
238 Riesz spaces 353B
353B Proposition Let U be an Archimedean Riesz space. Then
(a) for every A U, the band generated by A is A

,
(b) every band in U is complemented.
proof (a) Let V be the band generated by A. Then V is surely included in A

, because this is a band

including A (352O). ??? Suppose, if possible, that V ,= A

. Then there is a w A

V , so that [w[ / V .
Set B = v : v V, v [w[; then B is upwards-directed and non-empty. Because V is order-closed, [w[
cannot be the supremum of A, and there is a u
0
> 0 such that [w[u
0
v for every v B. Now u
0
[w[ ,= 0,
so u
0
/ A

, and there is a u
1
A such that v = u
0
[u
1
[ > 0. In this case nv B for every n N. PPP
Induce on n. For n = 0 this is trivial. For the inductive step, given that nv B, then nv [w[ u
0
so
(n + 1)v nv + u
0
[w[; but also (n + 1)v nv + [u
1
[ V , so (n + 1)v B. QQQ But this means that
nv [w[ for every n, which is impossible, because U is Archimedean. XXX
(b) Now if V U is any band, it is surely the band generated by itself, so is equal to V

, and is
complemented.
Remark We may therefore speak of the band algebra of an Archimedean Riesz space, rather than the
complemented band algebra (352Q).
353C Corollary Let U be an Archimedean Riesz space and v U. Let V be the band in U generated
by v. If u U, then u V i there is no w such that 0 < w [u[ and w [v[ = 0.
proof By 353B, V = v

. Now, for u U,
u / V w v

, [u[ [w[ > 0 w v

, 0 < w [u[.
Turning this round, we have the condition announced.
353D Proposition Let U be an Archimedean Riesz space and V an order-dense Riesz subspace of U.
Then the map W W V is an isomorphism between the band algebras of U and V .
proof If W U is a band, then W V is surely a band in V (it is order-closed in V because it is the
inverse image of the order-closed set W under the embedding V U, which is order-continuous by 352Nc
and 352Nb). If W, W
t
are distinct bands in U, say W
t
, W, then W
t
, W

, by 353B, so W
t
W

,= 0;
because V is order-dense, V W
t
W

,= 0, and V W
t
,= V W. Thus W W V is injective.
If Q V is a band in V , then its complementary band in V is just Q

V , where Q

is taken in U. So
(because V , like U, is Archimedean, by 351Rc) Q = (Q

V )

V = W V , where W = (Q

V )

is a
band in U. Thus the map W W V is an order-preserving bijection between the two band algebras. By
312L, it is a Boolean isomorphism, as claimed.
353E Lemma Let U be an Archimedean Riesz space and V U a band such that supv : v V, 0
v u is dened for every u U
+
. Then V is a projection band.
proof Take any u U
+
and set v = supv
t
: v
t
V
+
, v
t
u, w = u v. v V because V is a band.
Also w V

. PPP??? If not, there is some v

0
V
+
such that w v
0
> 0. Now for any n N we see that
nv
0
u = nv
0
v = (n + 1)v
0
v +w = u,
so an induction on n shows that nv
0
u for every n; which is impossible, because U is supposed to be
Archimedean. XXXQQQ Accordingly u = v+w V +V

. As u is arbitrary, U
+
V +V

, and V is a projection
band (352R).
353F Lemma Let U be an Archimedean Riesz space. If A U is non-empty and bounded above and
B is the set of its upper bounds, then inf(B A) = 0.
proof ??? If not, let w > 0 be a lower bound for B A. If u A and v B, then v u w, that is,
u v w; as u is arbitrary, v w B. Take any u
0
A, v
0
B. Inducing on n, we see that v
0
nw B
for every n N, so that v
0
nw u
0
, nw v
0
u
0
for every n; but this is impossible, because U is
supposed to be Archimedean. XXX
353J Archimedean and Dedekind complete Riesz spaces 239
353G Dedekind completeness Recall from 314A that a partially ordered set P is Dedekind ()-
complete if (countable) non-empty sets with upper and lower bounds have suprema and inma in P. For a
Riesz space U, U is Dedekind complete i every non-empty upwards-directed subset of U
+
with an upper
bound has a least upper bound, and is Dedekind -complete i every non-decreasing sequence in U
+
with
an upper bound has a least upper bound. PPP (Compare 314Bc.) (i) Suppose that any non-empty upwards-
directed order-bounded subset of U
+
has an upper bound, and that A U is any non-empty set with an
upper bound. Take u
0
A and set
B = u
0
u
1
. . . u
n
u
0
: u
1
, . . . , u
n
A.
Then B is an upwards-directed subset of U
+
, and if w is an upper bound of A then w u
0
is an upper
bound of B. So supB is dened in U, and in this case u
0
+supB = sup A. As A is arbitrary, U is Dedekind
complete. (ii) Suppose that order-bounded non-decreasing sequences in U
+
have suprema, and that A U
is any countable non-empty set with an upper bound. Let u
n
)
nN
be a sequence running over A, and set
v
n
= sup
in
u
i
u
0
for each n. Then v
n
)
nN
is a non-decreasing order-bounded sequence in U
+
, and
u
0
+ sup
nN
v
n
= supA. (iii) Finally, still supposing that order-bounded non-decreasing sequences in U
+
have suprema, if A U is non-empty, countable and bounded below, inf A will be dened and equal to
sup(A). QQQ
353H Proposition Let U be a Dedekind -complete Riesz space.
(a) U is Archimedean.
(b) For any v U the band generated by v is a projection band.
(c) If u, v U, then u is uniquely expressible as u
1
+ u
2
, where u
1
belongs to the band generated by v
and [u
2
[ [v[ = 0.
proof (a) Suppose that u, v U are such that nu v for every n N. Then nu
+
v
+
for every n, and
A = nu
+
: n N is a countable non-empty upwards-directed set with an upper bound; say w = supA.
Since A+u
+
A, w +u
+
= sup(A+u
+
) w, and u u
+
0. As u, v are arbitrary, U is Archimedean.
(b) Let V be the band generated by v. Take any u U
+
and set A = v
t
: v
t
V, 0 v
t
u. Then
u n[v[ : n N is a countable set with an upper bound, so has a supremum u
1
say in U. Now u
1
is an
upper bound for A. PPP If v
t
A, then
v
t
= sup
nN
v
t
n[v[ u
1
by 352V. QQQ Since u n[v[ A V for every n, u
1
V and u
1
= supA.
As u is arbitrary, 353E tells us that V is a projection band.
(c) Again let V be the band generated by v. Then v

is a band containing v, so
v V v

, v

= v

(352Od), and V

= v

.
Now, if u U, u is uniquely expressible in the form u
1
+u
2
where u
1
V and u
2
V

, by (b). But
u
2
V

u
2
v

[u
2
[ [v[ = 0.
So we have the result.
353I Proposition In a Dedekind complete Riesz space, all bands are projection bands.
proof Use 353E, noting that the sets v : v V, 0 v u there are always non-empty, upwards-directed
and bounded above, so always have suprema.
353J Proposition (a) Let U be a Dedekind -complete Riesz space.
(i) If V is a solid linear subspace of U, then V is (in itself) Dedekind -complete.
(ii) If W is a sequentially order-closed Riesz subspace of U then W is Dedekind -complete.
(iii) If V is a sequentially order-closed solid linear subspace of U, the canonical map from U to V is
sequentially order-continuous, and the quotient Riesz space U/V is also Dedekind -complete.
(b) Let U be a Dedekind complete Riesz space.
240 Riesz spaces 353J
(i) If V is a solid linear subspace of U, then V is (in itself) Dedekind complete.
(ii) If W U is an order-closed Riesz subspace then W is Dedekind complete.
proof (a)(i) If u
n
)
nN
is a non-decreasing sequence in V
+
with an upper bound v V , then w = sup
nN
u
n
is dened in U; but as 0 w v, w V and w = sup
nN
u
n
in V . Thus V is Dedekind -complete.
(ii) If u
n
)
nN
is a non-decreasing order-bounded sequence in W, then u = sup
nN
u
n
is dened in U;
but because W is sequentially order-closed, u W and u = sup
nN
u
n
in W.
(iii) Let u
n
)
nN
be a non-decreasing sequence in U with supremum u. Then of course u

is an upper
bound for A = u

n
: n N in U/V . Now let p be any other upper bound for A. Express p as v

.
Then for each n N we have u

n
p, so that (u
n
v)
+
V . Because V is sequentially order-closed,
(u v)
+
= sup
nN
(u
n
v)
+
V and u

p. Thus u

is the least upper bound of A.

Similarly, if u
n
)
nN
is a non-increasing sequence in U with inmum u, then u

= inf
nN
u

n
in U/V .
Thus u u

is sequentially order-continuous.
Now suppose that p
n
)
nN
is a non-decreasing sequence in (U/V )
+
with an upper bound p (U/V )
+
. Let
u U
+
be such that u

= p, and for each n N let u

n
U
+
be such that u

n
= p
n
. Set v
n
= u sup
in
u
i
for each n; then v

n
= p
n
for each n, and v
n
)
nN
is a non-decreasing order-bounded sequence in U. Set
v = sup
nN
v
n
; by the last paragraph, v

= sup
nN
p
n
in U/V . As p
n
)
nN
is arbitrary, U/V is Dedekind
-complete, as claimed.
(b) The argument is the same as parts (i) and (ii) of the proof of (a).
353K Proposition Let U be a Riesz space and V a quasi-order-dense Riesz subspace of U which is (in
itself) Dedekind complete. Then V is a solid linear subspace of U.
proof Suppose that v V , u U and that [u[ [v[. Consider A = w : w V, 0 w u
+
. Then A
is a non-empty subset of V with an upper bound in V (viz., [v[). So A has a supremum v
0
in V . Because
the embedding V U is order-continuous (352N), v
0
is the supremum of A in U. But as V is order-dense
(353A), v
0
= u
+
and u
+
V . Similarly, u

V and u V . As u, v are arbitrary, V is solid.

353L Order units Let U be a Riesz space.
(a) An element e of U
+
is an order unit in U if U is the solid linear subspace of itself generated by e;
that is, if for every u U there is an n N such that [u[ ne. (For the solid linear subspace generated by
v U
+
is

nN
[nv, nv].)
(b) An element e of U
+
is a weak order unit in U if U is the principal band generated by e; that is, if
u = sup
nN
u ne for every u U
+
(352Vb).
Of course an order unit is a weak order unit.
(c) If U is Archimedean, then an element e of U
+
is a weak order unit i e

= U (353B), that is, i

e

= 0 (because
e

= 0 = e

= 0

= U = e

= e

= U

= 0,)
that is, i u e > 0 whenever u > 0.
353M Theorem Let U be an Archimedean Riesz space with order unit e. Then it can be embedded
as an order-dense and norm-dense Riesz subspace of C(X), where X is a compact Hausdor space, in such
a way that e corresponds to X; moreover, this embedding is essentially unique.
Remark Here C(X) is the space of all continuous functions from X to R; because X is compact, they are
all bounded, so that X is an order unit in C(X) = C
b
(X).
proof (a) Let X be the set of Riesz homomorphisms x from U to R such that x(e) = 1. Dene T : U R
X
by setting (Tu)(x) = x(u) for x X, u U; then it is easy to check that T is a Riesz homomorphism, just
because every member of X is a Riesz homomorphism, and of course Te = X.
353M Archimedean and Dedekind complete Riesz spaces 241
(b) The key to the proof is the fact that X separates the points of U, that is, that T is injective. I choose
the following method to show this. Suppose that w U and w > 0. Because U is Archimedean, there is a
> 0 such that (w e)
+
,= 0. Now there is an x X such that x(w) . PPP (i) By 351O, there is a solid
linear subspace V of U such that (we)
+
/ V and whenever uv = 0 in U then one of u, v belongs to V .
(ii) Because V ,= U, e / V , so no non-zero multiple of e can belong to V . Also observe that if u, v U V ,
then one of (uv)
+
, (v u)
+
must belong to V , while neither u = uv +(uv)
+
nor v = uv +(v u)
+
does; so u v / V . (iii) For each u U set A
u
= : R, (u e)
+
V . Then
A
u
= 0 (u e)
+
(u e)
+
V = A
u
.
Also A
u
is non-empty and bounded below, because if 0 is such that e u e then A
u
and
1 / A
u
(since (u ( 1)e)
+
e / V ). (iv) Set x(u) = inf A
u
for every u U; then A
u
for
every > x(u), / A
u
for every < x(u). (v) If u, v U and > x(u), > x(v) then
((u +v) ( +)e)
+
(u e)
+
+ (v e)
+
V
(352Fc), so + A
u+v
; as and are arbitrary, x(u +v) x(u) +x(v). (vi) If u, v U and < x(u),
< x(v) then
((u +v) ( +)e)
+
(u e)
+
(v e)
+
/ V ,
using (ii) of this argument and 352Fc, so + / A
u+v
. As and are arbitrary, x(u +v) x(u) +x(v).
(vii) Thus x : U R is additive. (viii) If u U, > 0 then
A
u
= (u e)
+
= (u e)
+
V = A
u
;
thus A
u
A
u
; similarly, A
u

1
A
u
so A
u
= A
u
and x(u) = x(u). (ix) Consequently x is linear,
since we know already from (vii) that x(0u) = 0.x(u), x(u) = x(u). (x) If u 0 then u + e e / V
for every > 0, that is, / A
u
for every > 0, and x(u) 0; thus x is a positive linear functional. (xi)
If u v = 0, then one of u, v belongs to V , so min(x(u), x(v)) 0 and (using (x)) min(x(u), x(v)) = 0;
thus x is a Riesz homomorphism (352G(iv)). (xii) A
e
= [1, [ so x(e) = 1. Thus x X. (xiii) / A
w
so
x(w) . QQQ
(c) Thus Tw ,= 0 whenever w > 0; consequently [Tw[ = T[w[ ,= 0 whenever w ,= 0, and T is injective. I
now have to dene the topology of X. This is just the subspace topology on X if we regard X as a subset
of R
U
with its product topology. To see that X is compact, observe that if for each u U we choose an

u
such that [u[
u
e, then X is a subspace of Q =

uU
[
u
,
u
]. Because Q is a product of compact
spaces, it is compact, by Tychonos theorem (3A3J). Now X is a closed subset of Q. PPP X is just the
intersection of the sets
x : x(u +v) = x(u) +x(v), x : x(u) = x(u),
x : x(u
+
) = max(x(u), 0), x : x(e) = 1
as u, v run over U and over R; and each of these is closed, so X is an intersection of closed sets and
therefore itself closed. QQQ Consequently X also is compact. Moreover, the coordinate functionals x x(u)
are continuous on Q, therefore on X also, that is, Tu : X R is a continuous function for every u U.
Note also that because Q is a product of Hausdor spaces, Q and X are Hausdor (3A3Id).
(d) So T is a Riesz homomorphism from U to C(X). Now T[U] is a Riesz subspace of C(X), containing
X, and such that if x, y X are distinct there is an f T[U] such that f(x) ,= f(y) (because there is
surely a u U such that x(u) ,= y(u)). By the Stone-Weierstrass theorem (281A), T[U] is | |

-dense in
C(X).
Consequently it is also order-dense. PPP If f > 0 in C(X), set =
1
3
|f|

, and let u U be such that

|f Tu|

; set v = (u e)
+
. Since
0 < (f 2X)
+
(Tu X)
+
f
+
= f,
0 < Tv f. As f is arbitrary, T[U] is quasi-order-dense, therefore order-dense (353A). QQQ
(e) I have still to show that the representation is (essentially) unique. Suppose, then, that we have
another representation of U as a norm-dense Riesz subspace of C(Z), with e this time corresponding to
Z; to simplify the notation, let us suppose that U is actually a subspace of C(Z). Then for each z Z,
242 Riesz spaces 353M
we have a functional z : U R dened by setting z(u) = u(z) for every u U; of course z is a Riesz
homomorphism such that z(e) = 1, that is, z X. Thus we have a function z z : Z X. For any
u U, the function z z(u) = u(z) is continuous, so the function z z is continuous (3A3Ib). If z
1
, z
2
are distinct members of Z, there is an f C(Z) such that f(z
1
) ,= f(z
2
) (3A3Bf); now there is a u U
such that |f u|

1
3
[f(z
1
) f(z
2
)[, so that u(z
1
) ,= u(z
2
) and z
1
,= z
2
. Thus z z is injective. Finally,
it is also surjective. PPP Suppose that x X. Set V = u : u U, x(u) = 0; then V is a solid linear
subspace of U (352Jb), not containing e. For z V
+
set G
v
= z : v(z) > 1. Because e / V , G
v
,= Z.
( = G
v
: v V
+
is an upwards-directed family of open sets in Z, not containing Z; consequently, because
Z is compact, ( cannot be an open cover of Z. Take z Z

(. Then v(z) 1 for every v V

+
; because
[v[ V
+
whenever v V , 0, we must have v(z) = 0 for every v V . Now, given any u U, consider
v = u x(u)e. Then x(v) = 0 so v V and v(z) = 0, that is,
u(z) = (v +x(u)e)(z) = v(z) +x(u)e(z) = x(u).
As u is arbitrary, z = x; as x is arbitrary, we have the result. QQQ
Thus z z is a continuous bijection from the compact Hausdor space Z to the compact Hausdor
space X; it must therefore be a homeomorphism (3A3Dd).
This argument shows that if U is embedded as a norm-dense Riesz subspace of C(Z), where Z is compact
and Hausdor, then Z must be homeomorphic to X. But it shows also that a homeomorphism is canonically
dened by the embedding; z Z corresponds to the Riesz homomorphism u u(z) in X.
353N Lemma Let U be a Riesz space, V an Archimedean Riesz space and S, T : U V Riesz
homomorphisms such that Su Tu
t
= 0 in V whenever u u
t
= 0 in U. Set W = u : Su = Tu. Then W
is a solid linear subspace of U; if S and T are order-continuous, W is a band.
proof (a) It is easy to check that, because S and T are Riesz homomorphisms, W is a Riesz subspace of U.
(b) If w W and 0 u w in U, then Su Tu. PPP??? Otherwise, set e = Sw = Tw, and let V
e
be
the solid linear subspace of V generated by e, so that V
e
is an Archimedean Riesz space with order unit,
containing both Su and Tu. By 353M (or its proof), there is a Riesz homomorphism x : V
e
R such
that x(e) = 1 and x(Su) > x(Tu). Take such that x(Su) > > x(Tu), and consider u
t
= (u w)
+
,
u
tt
= (w u)
+
. Then
x(Su
t
) = max(0, x(Su) x(Sw)) = max(0, x(Su) ) > 0,
x(Tu
tt
) = max(0, x(Tw) x(Tu)) = max(0, x(Tu)) > 0,
so
x(Su
t
Tu
tt
) = min(x(Su
t
), x(Tu
tt
)) > 0
and Su
t
Tu
tt
> 0, while u
t
u
tt
= 0. XXXQQQ
Similarly, Tu Su and u W. As u and w are arbitrary, W is a solid linear subspace.
(c) Finally, suppose that S and T are order-continuous, and that A W is a non-empty upwards-directed
set with supremum u in U. Then
Su = sup S[A] = supT[A] = Tu
and u W. As u and A are arbitrary, W is a band (352Ob).
353O f-algebras I give two results on f-algebras, intended to clarify the connexions between the
multiplicative and lattice structures of the Riesz spaces in Chapter 36.
Proposition Let U be an Archimedean f-algebra (352W). Then
(a) the multiplication is separately order-continuous in the sense that the maps u u w, u w u
are order-continuous for every w U
+
;
(b) the multiplication is commutative.
proof (a) Let A U be a non-empty set with inmum 0, and v
0
U
+
a lower bound for u w : u A.
Fix u
0
A. If u A and > 0, then v
0
(u
0

1

u)
+
u
0
w. PPP Set v = v
0
(u
0

1

u)
+
. Then
353P Archimedean and Dedekind complete Riesz spaces 243
v (u u
0
)
+
(u
0
u)
+
(u u
0
)
+
= 0,
so v (u u
0
)
+
= 0 and v ((u u
0
)
+
w) = 0. But
v v
0
u w (u u
0
)
+
w +u
0
w,
so
v ((u u
0
)
+
w) v + (u
0
w) v u
0
w,
by 352Fa. QQQ
Taking the inmum over u, and using the distributive laws (352E), we get
v
0
u
0
u
0
w.
Taking the inmum over , and using the hypothesis that U is Archimedean,
v
0
u
0
= 0.
But this means that v
0
(u
0
w) = 0, while v
0
u
0
w, so v
0
= 0. As v
0
is arbitrary, inf
uA
u w = 0;
as A is arbitrary, u u w is order-continuous. Similarly, u w u is order-continuous.
(b)(i) Fix v U
+
, and for u U set
Su = u v, Tu = v u.
Then S and T are both order-continuous Riesz homomorphisms from U to itself (352W(b-iv) and (a) above).
Also, Su Tu
t
= 0 whenever u u
t
= 0. PPP
0 = (u v) u
t
= (u v) (v u
t
). QQQ
So W = u : u v = v u is a band in U (353N). Of course v W (because Sv = Tv = v
2
). If
u W

, then v [u[ = 0 so Su = Tu = 0 (352W(b-i)), and u W; but this means that W

= 0 and
W = W

= U (353Bb). Thus v u = u v for every u U.

(ii) This is true for every v U
+
. Of course it follows that v u = u v for every u, v U, so that
multiplication is commutative.
353P Proposition Let U be an Archimedean f-algebra with multiplicative identity e.
(a) e is a weak order unit in U.
(b) If u, v U then u v = 0 i [u[ [v[ = 0.
(c) If u U has a multiplicative inverse u
1
then [u[ also has a multiplicative inverse; if u 0 then
u
1
0 and u is a weak order unit.
(d) If V is another Archimedean f-algebra with multiplicative identity e
t
, and T : U V is a positive
linear operator such that Te = e
t
, then T is a Riesz homomorphism i T(u v) = Tu Tv for all u, v U.
proof (a) e = e
2
0 by 352W(b-iii). If u U and e [u[ = 0 then [u[ = (e [u[) [u[ = 0; by 353Lc, e is
a weak order unit.
(b) If [u[ [v[ = 0 then u v = 0, by 352W(b-ii). If w = [u[ [v[ > 0, then w
2
[u[ [v[. Let n N be
such that nw , e, and set w
1
= (nw e)
+
, w
2
= (e nw)
+
. Then
0 ,= w
1
= w
1
e = w
1
w
2
+w
1
(e nw)
= w
1
(e nw) (nw)
2
n
2
[u[ [v[ = n
2
[u v[,
so u v ,= 0.
(c) u u
1
= e so [u[ [u
1
[ = [e[ = e (352W(b-iii)), and [u
1
[ = [u[
1
. (Recall that inverses in any
semigroup with identity are unique, so that we need have no inhibitions in using the formulae u
1
, [u[
1
.)
Now suppose that u 0. Then u
1
= [u
1
[ 0. If u [v[ = 0 then
e [v[ = (u u
1
) [v[ = 0,
so v = 0; accordingly u is a weak order unit.
(d)(i) If T is multiplicative, and u v = 0 in U, then Tu Tv = T(u v) = 0 and Tu Tv = 0, by (b).
So T is a Riesz homomorphism, by 352G.
244 Riesz spaces 353P
(ii) Accordingly I shall henceforth assume that T is a Riesz homomorphism and seek to show that it
is multiplicative.
If u, v U
+
, then T(u v) and Tu Tv both belong to the band generated by Tu. PPP Write W for this
band. () For any n 1 we have (v ne)
2
0, that is, 2nv v
2
+n
2
e, so
n(v ne) 2nv n
2
e v
2
.
Consequently
T(u v) nTu = T(u v) nT(u e) = T(u (v ne))
1
n
T(u v
2
)
because v
t
T(u v
t
) is a positive linear operator; as V is Archimedean, inf
nN
(T(u v) nTu)
+
= 0
and T(u v) = sup
nN
T(u v) nTu belongs to W. () If w [Tu[ = 0 then
w [Tu Tv[ = w ([Tu[ [Tv[) = 0;
so Tu Tv W

= W. QQQ
(iii) Fix v U
+
. For u U, set S
1
u = Tu Tv and S
2
u = T(u v). Then S
1
and S
2
are both
Riesz homomorphisms from U to V . If u u
t
= 0 in U, then S
1
u S
2
u
t
= 0 in V , because (by (ii) just
above) S
1
u belongs to the band generated by Tu, while S
2
u
t
belongs to the band generated by Tu
t
, and
Tu Tu
t
= T(u u
t
) = 0. By 353N, W = u : S
1
u = S
2
u is a solid linear subspace of U. Of course it
contains e, since
S
1
e = Te Tv = e
t
Tv = Tv = T(e v) = S
2
e.
In fact u W for every u U
+
. PPP As noted in (ii) just above, u ne
1
n
u
2
for every n 1. So
[S
1
u S
2
u[ = [S
1
(u ne)
+
+S
1
(u ne) S
2
(u ne)
+
S
2
(u ne)[
S
1
(u ne)
+
+S
2
(u ne)
+

1
n
(S
1
u
2
+S
2
u
2
)
for every n 1, and [S
1
u S
2
u[ = 0, that is, S
1
u = S
2
u. QQQ
So W = U, that is, Tu Tv = T(u v) for every u U. And this is true for every v U
+
. It follows at
once that it is true for every v U, so that T is multiplicative, as claimed.
353X Basic exercises >>>(a) Let U be a Riesz space in which every band is complemented. Show that
U is Archimedean.
(b) A Riesz space U has the principal projection property i the band generated by any single member
of U is a projection band. Show that any Dedekind -complete Riesz space has the principal projection
property, and that any Riesz space with the principal projection property is Archimedean.
>>>(c) Fill in the missing part (b-iii) of 353J.
(d) Let U be an Archimedean f-algebra with an order-unit which is a multiplicative identity. Show that
U can be identied, as f-algebra, with a subspace of C(X) for some compact Hausdor space X.
353Y Further exercises (a) Let U be a Riesz space in which every quasi-order-dense solid linear
subspace is order-dense. Show that U is Archimedean.
(b) Let X be a completely regular Hausdor space. Show that C(X) is Dedekind complete i C
b
(X) is
Dedekind complete i X is extremally disconnected.
(c) Let X be a compact Hausdor space. Show that C(X) is Dedekind -complete i G is open for every
cozero set G X. (Cf. 314Yf.) Show that in this case X is zero-dimensional.
(d) Let U be an Archimedean Riesz space such that u
n
: n N has a supremum in U whenever u
n
)
nN
is a sequence in U such that u
m
u
n
= 0 whenever m ,= n. Show that U has the principal projection property,
but need not be Dedekind -complete.
353 Notes Archimedean and Dedekind complete Riesz spaces 245
(e) Let U be an Archimedean Riesz space. Show that the following are equiveridical: (i) U has the
countable sup property (241Yd) (ii) for every A U there is a countable B A such that A and B have
the same upper bounds; (iii) every disjoint subset of U
+
is countable.
(f ) Let U be an Archimedean Riesz space with order unit e, and | |
e
the corresponding norm. Let Z be
the unit ball of U

. Show that for a linear functional f : U R the following are equiveridical: (i) f is
an extreme point of Z, that is, f Z and Z f is convex (ii) [f(e)[ = 1 and one of f, f is a Riesz
homomorphism.
(g) Let U be an Archimedean f-algebra. Show that an element e of U is a multiplicative identity i
e
2
= e and e is a weak order unit. (Hint: start by showing that under these conditions, e u = 0 u = 0.)
(h) Let U be an Archimedean f-algebra with a multiplicative identity. Show that if u U then u is
invertible i [u[ is invertible.
353 Notes and comments As in the last section, many of the results above have parallels in the theory
of Boolean algebras; thus 353A corresponds to 313K, 353G corresponds in part to remarks in 314Bc and
314Xa, and 353J corresponds to 314C-314E. Riesz spaces are more complicated; for instance, principal ideals
in Boolean algebras are straightforward, while in Riesz spaces we have to distinguish between the solid linear
subspace generated by an element and the band generated by the same element. Thus an order unit in a
Boolean ring would just be an identity, while in a Riesz space we must distinguish between order unit and
weak order unit. As this remark may suggest to you, (Archimedean) Riesz spaces are actually closer in
spirit to arbitrary Boolean rings than to the Boolean algebras we have been concentrating on so far; to the
point that in 361 below I will return briey to general Boolean rings.
Note that the standard denition of order-dense in Boolean algebras, as given in 313J, corresponds to
the denition of quasi-order-dense in Riesz spaces (352Na); the point here being that Boolean algebras
behave like Archimedean Riesz spaces, in which there is no need to make a distinction.
I give the representation theorem 353M more for completeness than because we need it in any formal
sense. In 351Q and 352L I have given representation theorems for general partially ordered linear spaces,
and general Riesz spaces, as quotients of spaces of functions; in 368F below I give a theorem for Archimedean
Riesz spaces corresponding rather more closely to the expressions of the L
p
spaces as quotients of spaces of
measurable functions. In 353M, by contrast, we have a theorem expressing Archimedean Riesz spaces with
order units as true spaces of functions, rather than as spaces of equivalence classes of functions. All these
theorems are important in forming an appropriate mental picture of ordered linear spaces, as in 352M.
I give a bare-handed proof of 353M, using only the Riesz space structure of C(X); if you know a little
about extreme points of dual unit balls you can approach from that direction instead, using 353Yf. The point
is that (as part (d) of the proof makes clear) the space X can be regarded as a subset of the normed space
dual U

of U with its weak* topology. In this treatise generally, and in the present chapter in particular,
I allow myself to be slightly prejudiced against normed-space methods; you can nd them in any book on
functional analysis, and I prefer here to develop techniques like those in part (b) of the proof of 353M, which
will be a useful preparation for such theorems as 368E.
There is a very close analogy between 353M and the Stone representation of Boolean algebras (311E,
311I-311K). Just as the proof of 311E looked at the set of ring homomorphisms from A to the elementary
Boolean algebra Z
2
, so the proof of 353M looks at Riesz homomorphisms from U to the elementary M-space
R. Later on, the most important M-spaces, from the point of view of this treatise, will be the L

spaces of
363, explicitly dened in terms of Stone representations (363A).
Of the two parts of 353O, it is (a) which is most important for the purposes of this book. The f-algebras
we shall encounter in Chapter 36 can be seen to be commutative for dierent, and more elementary, reasons.
The (separate) order-continuity of multiplication, however, is not always immediately obvious. Similarly,
the uniferent Riesz homomorphisms we shall encounter can generally be seen to be multiplicative without
relying on the arguments of 353Pd.
246 Riesz Spaces 354 intro.
354 Banach lattices
The next step is a brief discussion of norms on Riesz spaces. I start with the essential denitions (354A,
354D) with the principal properties of general Riesz norms (354B-354C) and order-continuous norms (354E).
I then describe two of the most important classes of Banach lattice: M-spaces (354F-354L) and L-spaces
(354M-354R), with their elementary properties. For M-spaces I give the basic representation theorem
(354K-354L), and for L-spaces I give a note on uniform integrability (354P-354R).
354A Denitions (a) If U is a Riesz space, a Riesz norm or lattice norm on U is a norm | | such
that |u| |v| whenever [u[ [v[; that is, a norm such that |[u[| = |u| for every u and |u| |v|
whenever 0 u v.
(b) A Banach lattice is a Riesz space with a Riesz norm under which it is complete.
Remark We have already seen many examples of Banach lattices; I list some in 354Xa below.
354B Lemma Let U be a Riesz space with a Riesz norm | |.
(a) U is Archimedean.
(b) The maps u [u[ and u u
+
are uniformly continuous.
(c) For any u U, the sets v : v u and v : v u are closed; in particular, the positive cone of U is
closed.
(d) Any band in U is closed.
(e) If V is a norm-dense Riesz subspace of U, then V
+
= v : v V, v 0 is norm-dense in the positive
cone U
+
of U.
proof (a) If u, v U are such that nu v for every n N, then nu
+
v
+
so n|u
+
| |v
+
| for every n,
and |u
+
| = 0, that is, u
+
= 0 and u 0. As u, v are arbitrary, U is Archimedean.
(b) For any u, v U, [[u[ [v[[ [u v[ (352D), so |[u[ [v[| |u v|; thus u [u[ is uniformly
continuous. Consequently u
1
2
(u +[u[) = u
+
is uniformly continuous.
(c) Now v : v u = v : (v u)
+
= 0 is closed because the function v (v u)
+
is continuous and
0 is closed. Similarly v : v u = v : (u v)
+
= 0 is closed.
(d) If V U is a band, then V = V

(353Bb), that is, V = v : [v[ [w[ = 0 for every w V

.
Because the function v [v[ [w[ =
1
2
([v[ + [w[ [[v[ [w[[) is continuous, all the sets v : [v[ [w[ = 0
are closed, and so is their intersection V .
(e) Observe that V
+
= v
+
: v V and U
+
= u
+
: u U; recall that u u
+
is continuous, and
apply 3A3Eb.
354C Lemma If U is a Banach lattice and u
n
)
nN
is a sequence in U such that

n=0
|u
n
| < , then
sup
nN
u
n
is dened in U, with | sup
nN
u
n
|

n=0
|u
n
|.
proof Set v
n
= sup
in
u
i
for each n. Then
0 v
n+1
v
n
(u
n+1
u
n
)
+
[u
n+1
u
n
[
for each n N, so

n=0
|v
n+1
v
n
|

n=0
|u
n+1
u
n
|

n=0
|u
n+1
| +|u
n
|
is nite, and v
n
)
nN
is Cauchy. Let u be its limit; because v
n
)
nN
is non-decreasing, and the sets v : v
v
n
are all closed, u v
n
for each n N. On the other hand, if v v
n
for every n, then
(u v)
+
= lim
n
(v
n
v)
+
= 0,
and u v. So
u = sup
nN
v
n
= sup
nN
u
n
is the required supremum.
354E Banach lattices 247
To estimate its norm, observe that [v
n
[

n
i=0
[u
i
[ for each n (induce on n, using the last item in 352D
for the inductive step), so that
|u| = lim
n
|v
n
|

i=0
|[u
i
[| =

i=0
|u
i
|.
354D I come now to the basic properties according to which we classify Riesz norms.
Denitions (a) A Fatou norm on a Riesz space U is a Riesz norm on U such that whenever A U
+
is
non-empty and upwards-directed and has a least upper bound in U, then | supA| = sup
uA
|u|. (Observe
that, once we know that | | is a Riesz norm, we can be sure that |u| | supA| for every u A, so that
all we shall need to check is that | supA| sup
uA
|u|.)
(b) A Riesz norm on a Riesz space U has the Levi property if every upwards-directed norm-bounded
set is bounded above.
(c) A Riesz norm on a Riesz space U is order-continuous if inf
uA
|u| = 0 whenever A U is a
non-empty downwards-directed set with inmum 0.
354E Proposition Let U be a Riesz space with an order-continuous Riesz norm | |.
(a) If A U is non-empty and upwards-directed and has a supremum, then supA A.
(b) | | is Fatou.
(c) If A U is non-empty and upwards-directed and bounded above, then for every > 0 there is a
u A such that |(v u)
+
| for every v A.
(d) Any non-decreasing order-bounded sequence in U is Cauchy.
(e) If U is a Banach lattice it is Dedekind complete.
(f) Every order-dense Riesz subspace of U is norm-dense.
proof (a) Suppose that A U is non-empty and upwards-directed and has a least upper bound u
0
. Then
B = u
0
u : u A is downwards-directed and has inmum 0. So inf
uA
|u
0
u| = 0, and u
0
A.
(b) If, in (a), A U
+
, then we must have
|u
0
| inf
uA
|u| +|u u
0
| sup
uA
|u|.
As A is arbitary, | | is a Fatou norm.
(c) Let B be the set of upper bounds for A. Then B is downwards-directed; because A is upwards-
directed, B A = v u : v B, u A is downwards-directed. By 353F, inf(B A) = 0. So there are
w B, u A such that |w u| . Now if v A,
(v u)
+
= (v u) u w u,
so |(v u)
+
| .
(d) If u
n
)
nN
is a non-decreasing order-bounded sequence, and > 0, then, applying (c) to u
n
: n N,
we nd that there is an m N such that |u
m
u
n
| whenever m n.
(e) Now suppose that U is a Banach lattice. Let A U be any non-empty set with an upper bound. Set
A
t
= u
0
. . . u
n
: u
0
, . . . , u
n
A, so that A
t
is upwards-directed and has the same upper bounds as A.
For each n, choose u
n
A
t
such that |(u u
n
)
+
| 2
n
for every u A
t
. Set v
n
= sup
in
u
i
for each n;
then v
n
A
t
and |v
m
v
n
| |(v
m
u
n
)
+
| 2
n
for all m n. So v
n
)
nN
is Cauchy and has a limit v
say. If u A, then |(u v)
+
| = lim
n
|(u v
n
)
+
| = 0, so u v; while if w is any upper bound for A,
then |(v w)
+
| = lim
n
|(v
n
w)
+
| = 0 and v w. Thus v = sup A and A has a supremum.
(f ) If V is an order-dense Riesz subspace of U and u U
+
, set A = v : v V, v u. Then A is
upwards-directed and has supremum u, so u A V , by (a). Thus U
+
V ; it follows at once that
U = U
+
U
+
V .
248 Riesz Spaces 354F
354F Lemma If U is an Archimedean Riesz space with an order unit e (denition: 353L), there is a
Riesz norm | |
e
dened on U by the formula
|u|
e
= min : [u[ e
for every u U.
proof This is a routine verication. Because e is an order-unit, : [u[ e is always non-empty,
so always has an inmum
0
say; now [u[
0
e e for every > 0, so (because U is Archimedean)
[u[
0
e 0 and [u[
0
e, so that the minimum is attained. In particular, |u|
e
= 0 i u = 0. The
subadditivity and homogeneity of | |
e
are immediate from the facts that [u +v[ [u[ +[v[, [u[ = [[[u[.
354G Denitions (a) If U is an Archimedean Riesz space and e an order unit in U, the norm | |
e
as
dened in 354F is the order-unit norm on U associated with e.
(b) An M-space is a Banach lattice in which the norm is an order-unit norm.
(c) If U is an M-space, its standard order unit is the order unit e such that | |
e
is the norm of U. (To
see that e is uniquely dened, observe that it is supu : u U, |u| 1.)
354H Examples (a) For any set X,

(X) is an M-space with standard order unit X. (As remarked

in 243Xl, the completeness of

(X) can be regarded as the special case of 243E in which X is given counting
measure.)
(b) For any topological space X, the space C
b
(X) of bounded continuous real-valued functions on X is
an M-space with standard order unit X. (It is a Riesz subspace of

(X) containing the order unit of

(X), therefore in its own right an Archimedean Riesz space with order unit. To see that it is complete, it
is enough to observe that it is closed in

(X) because a uniform limit of continuous functions is continuous

(3A3Nb).)
(c) For any measure space (X, , ), the space L

() is an M-space with standard order unit X

.
354I Lemma Let U be an Archimedean Riesz space with order unit e, and V a subset of U which is
dense for the order-unit norm | |
e
. Then for any u U there are sequences v
n
)
nN
, w
n
)
nN
in V such
that v
n
v
n+1
u w
n+1
w
n
and |w
n
v
n
|
e
2
n
for every n; so that u = sup
nN
v
n
= inf
nN
w
n
in
U.
If V is a Riesz subspace of U, and u 0, we may suppose that v
n
0 for every n. Consequently V is
order-dense in U.
proof For each n N, take v
n
, w
n
V such that
|u
3
2
n+3
e v
n
|
e

1
2
n+3
, |u +
3
2
n+3
e w
n
|
e

1
2
n+3
.
Then
u
1
2
n+1
e v
n
u
1
2
n+2
e u u +
1
2
n+2
e w
n
u +
1
2
n+1
e.
Accordingly v
n
)
nN
is non-decreasing, w
n
)
nN
is non-increasing and |w
n
v
n
|
e
2
n
for every n. Because
U is Archimedean, sup
nN
v
n
= inf
nN
w
n
= u.
If V is a Riesz subspace of U, then replacing v
n
by v
+
n
if necessary we may suppose that every v
n
is
non-negative; and V is order-dense by the denition in 352N.
354J Proposition Let U be an Archimedean Riesz space with an order unit e. Then | |
e
is Fatou and
has the Levi property.
proof This is elementary. If A U
+
is non-empty, upwards-directed and norm-bounded, then it is bounded
above by e, where = sup
uA
|u|
e
. This is all that is called for in the Levi property. If moreover supA
is dened, then supA e so | supA| , as required in the Fatou property.
354O Banach lattices 249
354K Theorem Let U be an Archimedean Riesz space with order unit e. Then it can be embedded as
an order-dense and norm-dense Riesz subspace of C(X), where X is a compact Hausdor space, in such a
way that e corresponds to X and | |
e
corresponds to | |

; moreover, this embedding is essentially unique.

proof This is nearly word-for-word a repetition of 353M. The only addition is the mention of the norms.
But let X and T : U C(X) be as in 353M. Then, for any u U, [u[ |u|
e
e, so that
[Tu[ = T[u[ |u|
e
Te = |u|
e
X,
and |Tu|

|u|
e
. On the other hand, if 0 < < |u|
e
then u
1
= ([u[ e)
+
> 0, so that Tu
1
=
([Tu[ X)
+
> 0 and |Tu|

; as is arbitrary, |Tu|

|u|
e
.
354L Corollary Any M-space U is isomorphic, as Banach lattice, to C(X) for some compact Hausdor
X, and the isomorphism is essentially unique. X can be identied with the set of Riesz homomorphisms
x : U R such that x(e) = 1, where e is the standard order unit of U, with the topology induced by the
product topology on R
U
.
proof By 354K, there are a compact Hausdor space X and an embedding of U as a norm-dense Riesz
subspace of C(X) matching | |
e
to | |

. Since U is complete under | |

e
, its image is closed in C(X) (3A4Ff),
and must be the whole of C(X). The expression is unique just in so far as the expression of 353M/354K is
unique. In particular, we may, if we wish, take X to be the set of normalized Riesz homomorphisms from
U to R, as in the proof of 353M.
Remark If U is an M-space, then the construction of 353M represents U as C(X), where X is the set of
uniferent Riesz homomorphisms from U to R; this is sometimes called the spectrum of U.
354M I come now to a second fundamental class of Banach lattices, in a strong sense dual to the class
of M-spaces, as will appear in 356.
Denition An L-space is a Banach lattice U such that |u +v| = |u| +|v| whenever u, v U
+
.
Example If (X, , ) is any measure space, then L
1
(), with its norm | |
1
, is an L-space (242D, 242F). In
particular, taking to be counting measure on N,
1
is an L-space (242Xa).
354N Theorem If U is an L-space, then its norm is order-continuous and has the Levi property.
proof (a) Both of these are consequences of the following fact: if A U is norm-bounded and non-empty
and upwards-directed, then sup A is dened in U and belongs to the norm-closure of A in U. PPP For u A,
set (u) = sup|v u| : v A, v u. We surely have (u) |u| + sup
vA
|v| < . Choose a sequence
u
n
)
nN
in A such that u
n+1
u
n
and |u
n+1
u
n
|
1
2
(u
n
) for each n. Then
|u
n+1
u
0
| =

n
i=0
|u
i+1
u
i
|
1
2

n
i=0
(u
i
)
for every n, using the denition of L-space. Because A is bounded,

i=0
(u
i
) < and lim
n
(u
n
) = 0.
But |u
m
u
n
| (u
n
) whenever m n, so u
n
)
nN
is Cauchy and has a limit u

in U.
For each n N, u

u
n
because u
m
u
n
for every m n (see 354Bc). If u A, n N then there
is a u
t
A such that u
t
u u
n
; now (u u

)
+
u
t
u
n
so |(u u

)
+
| |u
t
u
n
| (u
n
); as n is
arbitrary, |(u u

)
+
| = 0 and u u

. Thus u

is an upper bound for A. But if v is any upper bound for

A, then u
n
v for every n so u

v. Thus u

is the least upper bound of A; and u

A because it is the
norm limit of u
n
)
nN
. QQQ
(b) This shows immediately that the norm has the Levi property. But also it must be order-continuous.
PPP If A U is non-empty and downwards-directed and has inmum 0, take any u
0
A and consider
B = u
0
u : u A, u u
0
. Then B is upwards-directed and has supremum u
0
, so u
0
B and
inf
uA
|u| inf
vB
|u
0
v| = 0. QQQ
354O Proposition If U is an L-space and V is a norm-closed Riesz subspace of U, then V is an L-space
in its own right. In particular, any band of U is an L-space.
250 Riesz Spaces 354O
proof For any Riesz subspace V of U, we surely have |u + v| = |u[ + |v| whenever u, v V
+
; so if
V is norm-closed, therefore a Banach lattice, it must be an L-space. But in any Banach lattice, a band is
norm-closed (354Bd), so a band in an L-space is again an L-space.
354P Uniform integrability in L-spaces Some of the ideas of 246 can be readily expressed in this
abstract context.
Denition Let U be an L-space. A set A U is uniformly integrable if for every > 0 there is a w U
+
such that |([u[ w)
+
| for every u A.
354Q Since I have already used the phrase uniformly integrable based on a dierent formula, I had
better check instantly that the two denitions are consistent.
Proposition If (X, , ) is any measure space, then a subset of L
1
= L
1
() is uniformly integrable in the
sense of 354P i it is uniformly integrable in the sense of 246A.
proof (a) If A L
1
is uniformly integrable in the sense of 246A, then for any > 0 there are M 0,
E such that E < and
_
([u[ ME

)
+
for every u A; now w = ME

belongs to (L
1
)
+
and
|([u[ w)
+
| for every u A. As is arbitrary, A is uniformly integrable in the sense of 354P.
(b) Now suppose that A is uniformly integrable in the sense of 354P. Let > 0. Then there is a w (L
1
)
+
such that |([u[ w)
+
|
1
2
for every u A. There is a simple function h : X R such that |wh

|
1
2

(242M); now take E = x : h(x) ,= 0, M = sup

xX
[h(x)[ (I pass over the trivial case X = ), so that
h ME and
([u[ ME

)
+
([u[ w)
+
+ (w ME

)
+
([u[ w)
+
+ (w h

)
+
,
_
([u[ ME

)
+
|([u[ w)
+
| +|w h

|
for every u A. As is arbitrary, A is uniformly integrable in the sense of 354P.
354R I give abstract versions of the easiest results from 246.
Theorem Let U be an L-space.
(a) If A U is uniformly integrable, then
(i) A is norm-bounded;
(ii) every subset of A is uniformly integrable;
(iii) for any R, A is uniformly integrable;
(iv) there is a uniformly integrable, solid, convex, norm-closed set C A;
(v) for any other uniformly integrable set B U, A B and A+B are uniformly integrable.
(b) For any set A U, the following are equiveridical:
(i) A is uniformly integrable;
(ii) lim
n
([u
n
[ sup
i<n
[u
i
[)
+
= 0 for every sequence u
n
)
nN
in A;
(iii) either A is empty or for every > 0 there are u
0
, . . . , u
n
A such that |([u[ sup
in
[u
i
[)
+
|
for every u A;
(iv) A is norm-bounded and any disjoint sequence in the solid hull of A is norm-convergent to 0.
(c) If V U is a closed Riesz subspace then a subset of V is uniformly integrable when regarded as a
subset of V i it is uniformly integrable when regarded as a subset of U.
proof (a)(i) There must be a w U
+
such that
_
([u[ w)
+
1 for every u A; now
[u[ [u[ w +[w[ ([u[ w)
+
+[w[, |u| |([u[ w)
+
| +|w| 1 +|w|
for every u A, so A is norm-bounded.
(ii) This is immediate from the denition.
(iii) Given > 0, we can nd w U
+
such that [[|([u[ w)
+
| for every u A; now |([v[
[[w)
+
| for every v A.
(iv) If A is empty, take C = A. Otherwise, try
354R Banach lattices 251
C = v : v U, |([v[ w)
+
| sup
uA
|([u[ w)
+
| for every w U
+
.
Evidently A C, and C satises the denition 354M because A does. The functionals
v |([v[ w)
+
| : U R
are all continuous for | | (because the operators v [v[, v v w, v v
+
, v |v| are continuous), so
C is closed. If [v
t
[ [v[ and v C, then
|([v
t
[ w)
+
| |([v[ w)
+
| sup
uA
|([u[ w)
+
|
for every w, and v
t
C. If v = v
1
+v
2
where v
1
, v
2
C, [0, 1] and = 1, then [v[ [v
1
[ +[v
2
[,
so
[v[ w ([v
1
[ w) + ([v
2
[ w) ([v
1
[ w)
+
+ ([v
2
[ w)
+
and
([v[ w)
+
([v
1
[ w)
+
+([v
2
[ w)
+
for every w; accordingly
|([v[ w)
+
| |([v
1
[ w)
+
| +|([v
2
[ w)
+
|
( +) sup
uA
|([u[ w)
+
| = sup
uA
|([u[ w)
+
|
for every w, and v C.
Thus C has all the required properties.
(v) I show rst that A B is uniformly integrable. PPP Given > 0, let w
1
, w
2
U
+
be such that
|([u[ w
1
)
+
| for every u A, |([u[ w
2
)
+
| for every u B.
Set w = w
1
w
2
; then |([u[ w)
+
| for every u AB. As is arbitrary, AB is uniformly integrable.
QQQ
Now (iv) tells us that there is a convex uniformly integrable set C including A B, and in this case
A+B 2C, so A+B is also uniformly integrable, using (ii) and (iii).
(b)(i)(ii)&(iv) Suppose that A is uniformly integrable and that u
n
)
nN
is any sequence in the solid
hull of A. Set v
n
= sup
in
[u
i
[ for n N and
v
t
0
= v
0
= [u
0
[, v
t
n
= v
n
v
n1
= ([u
n
[ sup
i<n
[u
i
[)
+
for n 1. Given > 0, there is a w U
+
such that |([u[ w)
+
| for every u A, and therefore for
every u in the solid hull of A. Of course sup
nN
|v
n
w| |w| is nite, so there is an n N such that
|v
i
w| +|v
n
w| for every i N. But now, for m > n,
v
t
m
([u
m
[ v
n
)
+
([u
m
[ [u
m
[ w)
+
+ (([u
m
[ w) v
n
)
+
([u
m
[ w)
+
+ (v
m
w) (v
n
w),
so that
|v
t
m
| |([u
m
[ w)
+
| +|(v
m
w) (v
n
w)|
= |([u
m
[ w)
+
| +|v
m
w| |v
n
w| 2,
using the L-space property of the norm for the equality in the middle. As is arbitrary, lim
n
v
t
n
= 0. As
u
n
)
nN
is arbitrary, condition (ii) is satised; but so is condition (iv), because we know from (a-i) that A
is norm-bounded, and if u
n
)
nN
is disjoint then v
t
n
= [u
n
[ for every n, so that in this case lim
n
u
n
= 0.
(ii)(iii)(i) are elementary.
not-(i)not-(iv) Now suppose that A is not uniformly integrable. If it is not norm-bounded, we can
stop. Otherwise, there is some > 0 such that sup
uA
|([u[ w)
+
| > for every w U
+
. Consequently
we shall be able to choose inductively a sequence u
n
)
nN
in A such that |([u
n
[ 2
n
sup
i<n
[u
i
[)
+
| > for
every n 1. Because A is norm-bounded,

i=0
2
i
|u
i
| is nite, and we can set
252 Riesz Spaces 354R
v
n
= ([u
n
[ 2
n
sup
i<n
[u
i
[

i=n+1
2
i
[u
i
[)
+
for each n. (The sum

i=n+1
2
i
[u
i
[ is dened because

m
i=n+1
2
i
[u
i
[)
mn+1
is a Cauchy sequence. We
have v
m
[u
m
[,
v
m
v
n
([u
m
[ 2
n
[u
n
[)
+
([u
n
[ 2
n
[u
m
[)
+
(2
n
[u
m
[ [u
n
[)
+
([u
n
[ 2
n
[u
m
[)
+
= 0
whenever m < n, so v
n
)
nN
is a disjoint sequence in the solid hull of A; while
|v
n
| |([u
n
[ 2
n
sup
i<n
[u
i
[)
+
|

i=n+1
2
i
|u
i
| 2
n
sup
uA
|u|
as n , so condition (iv) is not satised.
(c) Now this follows at once, because conditions (b-ii) and (b-iv) are satised in V i they are satsied
in U.
354X Basic exercises >>>(a) Work through the proofs that the following are all Banach lattices: (i)
R
r
with () |x|
1
=

r
i=1
[
i
[ () |x|
2
=
_

r
i=1
[
i
[
2
() |x|

= max
ir
[
i
[, where x = (
1
, . . . ,
r
). (ii)

p
(X), for any set X and any p [1, ] (242Xa, 243Xl, 244Xn). (iii) L
p
(), for any measure space (X, , )
and any p [1, ] (242F, 243E, 244G). (iv) ccc
0
, the space of sequences convergent to 0, with the norm | |

inherited from

.
(b) Let U
i
)
iI
be any family of Banach lattices. Write U for their Riesz space product (352K), and in
U set
|u|
1
=

iI
|u(i)|, V
1
= u : |u|
1
< ,
|u|

= sup
iI
|u(i)|, V

= u : |u|

< .
Show that V
1
, V

are solid linear subspaces of U and are Banach lattices under their norms | |
1
, | |

.
(c) Let U be a Riesz space with a Riesz norm. Show that the maps (u, v) uv, (u, v) uv : UU U
are uniformly continuous.
>>>(d) Let U be a Riesz space with a Riesz norm. (i) Show that any order-bounded set in U is norm-
bounded. (ii) Show that in R
r
, with any of the standard Riesz norms (354Xa(i)), norm-bounded sets are
order-bounded. (iii) Show that in
1
(N) there is a sequence converging to 0 (for the norm) which is not order-
bounded. (iv) Show that in ccc
0
any sequence converging to 0 is order-bounded, but there is a norm-bounded
set which is not order-bounded.
(e) Let U be a Riesz space with a Riesz norm. Show that it is a Banach lattice i non-decreasing Cauchy
sequences are convergent. (Hint: if |u
n+1
u
n
| 2
n
for every n, show that sup
in
u
i
)
nN
is Cauchy,
and that u
n
)
nN
converges to inf
nN
sup
mn
u
m
.)
(f ) Let U be a Riesz space with a Riesz norm. Show that U is a Banach lattice i every non-decreasing
Cauchy sequence u
n
)
nN
in U
+
has a least upper bound u with |u| = lim
n
|u
n
|.
(g) Let U be a Banach lattice. Suppose that B U is solid and sup
nN
u
n
B whenever u
n
)
nN
is a
non-decreasing sequence in B with a supremum in U. Show that B is closed. (Hint: show rst that u B
whenever there is a sequence u
n
)
nN
in BU
+
such that |uu
n
| 2
n
for every n; do this by considering
v
m
= inf
nm
u
n
.)
(h) Let U be any Riesz space with a Riesz norm. Show that the Banach space completion of U (3A5Ib)
has a unique partial ordering under which it is a Banach lattice.
>>>(i) Show that the space ccc
0
of sequences convergent to 0, with | |

, is a Banach lattice with an order-

continuous norm which does not have the Levi property.
354Yc Banach lattices 253
>>>(j) Show that

, with | |

, is a Banach lattice with a Fatou norm which has the Levi property but is
not order-continuous.
(k) Let U be a Riesz space with a Fatou norm. Show that if V U is a regularly embedded Riesz
subspace (denition: 352Ne) then the induced norm on V is a Fatou norm.
(l) Let U be a Riesz space and | | a Riesz norm on U which is order-continuous in the sense of 354Dc.
Show that it is order-continuous in the sense of 313H when regarded as a function from U
+
to [0, [.
(m) Let U be a Riesz space with an order-continuous norm. Show that if V U is a regularly embedded
Riesz subspace then the induced norm on V is order-continuous.
(n) Let U be a Dedekind -complete Riesz space with a Fatou norm which has the Levi property. Show
that it is a Banach lattice. (Hint: 354Xf.)
(o) Let U
i
)
iI
be any family of Banach lattices and let V
1
, V

be the subspaces of U =

iI
U
i
as
described in 354Xb. Show that V
1
, V

have norms which are Fatou, or have the Levi property, i every U
i
has. Show that the norm of V
1
is order-continuous i the norm of every U
i
is.
(p) Let U be a Banach lattice with an order-continuous norm. Show that a Riesz subspace of U (indeed,
any sublattice of U) is norm-closed i it is order-closed in the sense of 313D, and in this case is itself a
Banach lattice with an order-continuous norm.
>>>(q) Let U be an M-space and V a norm-closed Riesz subspace of U containing the standard order unit
of U. (i) Show that V , with the induced norm, is an M-space. (ii) Deduce that the space ccc of convergent
sequences is an M-space if given the norm | |

inherited from

.
(r) Show that a Banach lattice U is an M-space i (i) its norm is a Fatou norm with the Levi property
(ii) |u v| = max(|u|, |v|) for all u, v U
+
.
>>>(s) Describe a topological space X such that the space ccc of convergent sequences (354Xq) can be
identied with C(X).
(t) Let D R be any non-empty set, and V the space of functions f : D R of bounded variation
(224). For f V set |f| = sup[f(t
0
)[ +

n
i=1
[f(t
i
) f(t
i1
)[ : t
0
t
1
. . . t
n
in D (224Yb). Let
V
+
be the set of non-negative, non-decreasing functions in V . Show that V
+
is the positive cone of V for
a Riesz space ordering under which V is an L-space.
354Y Further exercises (a) Let U be a Riesz space with a Riesz norm, and V a norm-dense Riesz
subspace of U. Suppose that the induced norm on V is Fatou, when regarded as a norm on the Riesz space
V . Show (i) that V is order-dense in U (ii) that the norm of U is Fatou. (Hint: for (i), show that if u U
+
,
v
n
V
+
and |u v
n
| 2
n2
|u| for every n, then |v
0
inf
in
v
i
|
1
4
|u| for every n, so that 0 cannot
be inf
nN
v
n
in V .)
(b) Let U be a Riesz space with a Riesz norm. Show that the following are equiveridical: (i) lim
n
u
n
=
0 whenever u
n
)
nN
is a disjoint order-bounded sequence in U
+
(ii) lim
n
u
n+1
u
n
= 0 for every order-
bounded non-decreasing sequence u
n
)
nN
in U (iii) whenever A U
+
is a non-empty downwards-directed
set in U
+
with inmum 0, inf
uA
sup
vA,vu
|u v| = 0. (Hint: for (i)(ii), show by induction that
lim
n
u
n
= 0 whenever u
n
)
nN
is an order-bounded sequence such that, for some xed k, inf
iK
u
i
= 0
for every K N of size k; now show that if u
n
)
nN
is non-decreasing and 0 u
n
u for every n, then
inf
iK
(u
i+1
u
i

1
k
u)
+
= 0 whenever K N, #(K) = k 1. For (iii)(i), set A = u : n, u u
i
i
n. See Fremlin 74a, 24H.)
(c) Show that any Riesz space with an order-continuous norm has the countable sup property (denition:
241Yd).
254 Riesz Spaces 354Yd
(d) Let U be a Banach lattice. Show that the following are equiveridical: (i) the norm on U is order-
continuous; (ii) U satises the conditions of 354Yb; (iii) every order-bounded monotonic sequence in U is
Cauchy.
(e) Let U be a Riesz space with a Fatou norm. Show that the norm on U is order-continuous i it satises
the conditions of 354Yb.
(f ) For f C([0, 1]), set |f|
1
=
_
[f(x)[dx. Show that | |
1
is a Riesz norm on C([0, 1]) satisfying the
conditions of 354Yb, but is not order-continuous.
(g) Let U be a Riesz space with a Riesz norm | |. Show that (U, | |) satises the conditions of 354Yb i
the norm of its completion is order-continuous.
(h) Let U be a Riesz space with a Riesz norm, and V U a norm-dense Riesz subspace such that the
induced norm on V is order-continuous. Show that the norm of U is order-continuous. (Hint: use 354Ya.)
(i) Let U be an Archimedean Riesz space. For any e U
+
, let U
e
be the solid linear subspace of U
generated by e, so that e is an order unit in U
e
, and let | |
e
be the corresponding order-unit norm on U
e
.
We say that U is uniformly complete if U
e
is complete under | |
e
for every e U
+
. (i) Show that any
Banach lattice is uniformly complete. (ii) Show that any Dedekind -complete Riesz space is uniformly
complete (cf. 354Xn). (iii) Show that if U is a uniformly complete Riesz space with a Riesz norm which has
the Levi property, then U is a Banach lattice. (iv) Show that if U is a Banach lattice then a set A U is
closed, for the norm topology, i AU
e
is | |
e
-closed for every e U
+
. (v) Let V be a solid linear subspace
of U. Show that the quotient Riesz space U/V (352U) is Archimedean i V U
e
is | |
e
-closed for every
e U
+
. (vi) Show that if U is uniformly complete and V U is a solid linear subspace such that U/V is
Archimedean, then U/V is uniformly complete. (vii) Show that U is Dedekind -complete i it is uniformly
complete and has the principal projection property (353Xb). (Hint: for (vii), use 353Yc.)
(j) Let U be a Banach lattice such that |u + v| = |u| + |v| whenever u v = 0. Show that U is
an L-space. (Hint: by 354Yd, the norm is order-continuous, so U is Dedekind complete. If u, v 0, set
e = u +v, and represent U
e
as C(X) where X is extremally disconnected (353Yb); now approximate u and
v by functions taking only nitely many values to show that |u +v| = |u| +|v|.)
(k) Let U be a uniformly complete Archimedean Riesz space (354Yi). Set U
C
= U U with the complex
linear structure dened by identifying (u, v) U U as u+iv U
C
, so that u = e(u+iv), v = 1m(u+iv)
and ( +i)(u +iv) = (u v) +i(v +u). (i) Show that for w U
C
we can dene [w[ U by setting
[w[ = sup
]]=1
e(w). (ii) Show that if U is a uniformly complete Riesz subspace of R
X
for some set X,
then we can identify U
C
with the linear subspace of C
X
generated by U. (iii) Show that [w+w
t
[ [w[ +[w
t
[,
[w[ = [[[w[ for all w U
C
, C. (iv) Show that if w U
C
and [w[ u
1
+ u
2
, where u
1
, u
2
U
+
, then
w is expressible as w
1
+w
2
where [w
j
[ u
j
for both j. (Hint: set e = u
1
+u
2
and represent U
e
as C(X).)
(v) Show that if U
0
is a solid linear subspace of U, then, for w U
C
, [w[ U
0
i e w, 1mw both belong
to U
0
. (vi) Show that if U has a Riesz norm then we have a norm on U
C
dened by setting |w| = |[w[|,
and that if U is a Banach lattice then U
C
is a (complex) Banach space. (vii) Show that if U = L
p
(), where
(X, , ) is a measure space and p [1, ], then U
C
can be identied with L
p
C
() as dened in 242P, 243K,
244O. (We may call U
C
the complexication of the Riesz space U.)
(l) Let (X, , ) be a measure space and V a Banach lattice. Write L
1
V
for the space of Bochner integrable
functions from conegligible subsets of X to V , and L
1
V
for the corresponding set of equivalence classes
(253Yf). (i) Show that L
1
V
is a Banach lattice under the ordering dened by saying that f

i
f(x) g(x) in V for -almost every x X. (ii) Show that when V = L
1
(), for some other measure space
(Y, T, ), then this ordering on L
1
V
agrees with the ordering of L
1
() where is the (c.l.d.) product measure
on X Y and we identify L
1
V
with L
1
(), as in 253Yi. (iii) Show that if V has an order-continuous norm,
so has L
1
V
. (Hint: 354Yd(ii).) (iv) Show that if is Lebesgue measure on [0, 1] and V =

, then L
1
V
is
not Dedekind -complete.
355A Spaces of linear operators 255
354 Notes and comments Apart from some of the exercises, the material of this section is pretty strictly
conned to ideas which will be useful later in this volume. The basic Banach lattices of measure theory
are the L
p
spaces of Chapter 24; these all have Fatou norms with the Levi property (244Ye-244Yf), and for
p < their norms are order-continuous (244Yd). In Chapter 36 I will return to these spaces in a more
abstract context. Here I am mostly concerned to establish a vocabulary in which their various properties,
and the relationships between these properties, can be expressed.
In normed Riesz spaces we have a very rich mixture of structures, and must take particular care over the
concepts of boundedness, convergence and density, which have more than one possible interpretation.
In particular, we must scrupulously distinguish between order-bounded and norm-bounded sets. I have
not yet formally introduced any of the various concepts of order-convergence (see 367), but I think that
even so it is best to get into the habit of reminding oneself, when a convergent sequence appears, that it is
convergent for the norm topology, rather than in any sense related directly to the order structure.
I should perhaps warn you that for the study of M-spaces 354L is not as helpful as it may look. The
trouble is that apart from a few special cases (as in 354Xs) the topological space used in the representation
is actually more complicated and mysterious than the M-space it is representing.
After the introduction of M-spaces, this section becomes a natural place for uniformly complete spaces
(354Yi). For the moment I leave these in the exercises. But I mention them now because they oer a
straightforward route towards a theory of complex Riesz spaces (354Yk). In large parts of functional
analysis it is natural, and in some parts it is necessary, to work with normed spaces over C rather than over
R, and for L
2
spaces in particular it is useful to have a proper grasp of the complex case. And while the
insights oered by the theory of Riesz spaces are not especially important in such areas, I think we should
always seek connexions between dierent topics. So it is worth remembering that uniformly complete Riesz
spaces have complexications.
I shall have a great deal more to say about L-spaces when I come to spaces of additive functionals (362)
and to L
1
spaces again (365) and to linear operators on them (371); and before that, there will be something
in the next section on their duals, and on L-spaces which are themselves dual spaces. For the moment I just
give some easy results, direct translations of the corresponding facts in 246, which have natural expressions
in the language of this section, holding deeper ideas over. In particular, the characterization of uniformly
integrable sets as relatively weakly compact sets (247C) is valid in general L-spaces (356Q).
For an extensive treatment of Banach lattices, going very much deeper than I have space for in this
volume, see Lindenstrauss & Tzafriri 79. For a careful exposition of a great deal of useful information,
see Schaefer 74.
355 Spaces of linear operators
We come now to a discussion of linear operators between Riesz spaces. Of course linear operators are
central to any kind of functional analysis, and a feature of the theory of Riesz spaces is the way the order
structure picks out certain classes of operators for special consideration. Here I concentrate on positive and
order-continuous operators, with a brief mention of sequential order-continuity. It turns out, in fact, that
we need to work with operators which are dierences of positive operators or of order-continuous positive
operators. I dene the basic spaces L

, L

and L

c
(355A, 355G), with their most important properties
(355B, 355E, 355H-355I) and some remarks on the special case of Banach lattices (355C, 355K). At the
same time I give an important theorem on extension of operators (355F) and a corollary (355J).
The most important case is of course that in which the codomain is R, so that our operators become
real-valued functionals; I shall come to these in the next section.
355A Denition Let U and V be Riesz spaces. A linear operator T : U V is order-bounded if
T[A] is order-bounded in V for every order-bounded A U.
I will write L

(U; V ) for the set of order-bounded linear operators from U to V .

256 Riesz spaces 355B
355B Lemma If U and V are Riesz spaces,
(a) a linear operator T : U V is order-bounded i Tu : 0 u w is bounded above in V for every
w U
+
;
(b) in particular, any positive linear operator from U to V belongs to L

= L

(U; V );
(c) L

is a linear space;
(d) if W is another Riesz space and T : U V and S : V W are order-bounded linear operators, then
ST : U W is order-bounded.
proof (a) This is elementary. If T L

and w U
+
, [0, w] is order-bounded, so its image must be order-
bounded in V , and in particular bounded above. On the other hand, if T satises the condition, and A is
order-bounded, then A [u
1
, u
2
] for some u
1
u
2
, and
T[A] T[u
1
+ [0, u
2
u
1
]] = Tu
1
+T[[0, u
2
u
1
]]
is bounded above; similarly, T[A] is bounded above, so T[A] is bounded below; as A is arbitrary, T is
order-bounded.
(b) If T is positive then Tu : 0 u w is bounded above by Tw for every w 0, so T L

.
(c) If T
1
, T
2
L

, R and A U is order-bounded, then there are v

1
, v
2
V such that T
i
[A] [v
i
, v
i
]
for both i. Setting v = (1 +[[)v
1
+v
2
, (T
1
+T
2
)[A] [v, v]; as A is arbitrary, T
1
+T
2
belongs to L

;
as , T
1
, T
2
are arbitrary, and since the zero operator surely belongs to L

, L

is a linear subspace of the

space of all linear operators from U to V .
(d) This is immediate from the denition; if A U is order-bounded, then T[A] V and (ST)[A] =
S[T[A]] W are order-bounded.
355C Theorem If U and V are Banach lattices then every order-bounded linear operator (in particular,
every positive linear operator) from U to V is continuous.
proof ??? Suppose, if possible, that T : U V is an order-bounded linear operator which is not continuous.
Then for each n N we can nd a u
n
U such that |u
n
| 2
n
but |Tu
n
| n. Now u = sup
nN
[u
n
[ is
dened in U (354C), and there is a v V such that v Tw v whenever u w u; but this means
that |v| |Tu
n
| n for every n, which is impossible. XXX
355D Lemma Let U be a Riesz space and V any linear space over R. Then a function T : U
+
V
extends to a linear operator from U to V i
T(u +u
t
) = Tu +Tu
t
, T(u) = Tu
for all u, u
t
U
+
and every > 0, and in this case the extension is unique.
proof For in this case we can, and must, set
T
1
u = Tu
1
Tu
2
whenever u
1
, u
2
U
+
and u = u
1
u
2
;
it is elementary to check that this denes T
1
u uniquely for every u U, and that T
1
is a linear operator
extending T.
355E Theorem Let U be a Riesz space and V a Dedekind complete Riesz space.
(a) The space L

of order-bounded linear operators from U to V is a Dedekind complete Riesz space; its

positive cone is the set of positive linear operators from U to V . In particular, every order-bounded linear
operator from U to V is expressible as the dierence of positive linear operators.
(b) For T L

, T
+
and [T[ are dened in the Riesz space L

by the formulae
T
+
(w) = supTu : 0 u w,
[T[(w) = supTu : [u[ w = sup

n
i=0
[Tu
i
[ :

n
i=0
[u
i
[ = w
for every w U
+
.
(c) If S, T L

then
355E Spaces of linear operators 257
(S T)(w) = sup
0uw
Su +T(w u), (S T)(w) = inf
0uw
Su +T(w u)
for every w U
+
.
(d) Suppose that A L

is non-empty and upwards-directed. Then A is bounded above in L

i
Tu : T A is bounded above in V for every u U
+
, and in this case (supA)(u) = sup
TA
Tu for every
u 0.
(e) Suppose that A (L

)
+
is non-empty and downwards-directed. Then inf A = 0 in L

i inf
TA
Tu =
0 in V for every u U
+
.
proof (a)(i) Suppose that T L

. For w U
+
set R
T
(w) = supTu : 0 u w; this is dened because
V is Dedekind complete and Tu : 0 u w is bounded above in V . Then R
T
(w
1
+w
2
) = R
T
w
1
+R
T
w
2
for all w
1
, w
2
U
+
. PPP Setting A
i
= [0, w
i
] for each i, and A = [0, w], then of course A
1
+A
2
A; but also
A A
1
+ A
2
, because if u A then u = (u w
1
) + (u w
1
)
+
, and 0 (u w
1
)
+
(w w
1
)
+
= w
2
, so
u A
1
+A
2
. Consequently
R
T
w = sup T[A] = supT[A
1
+A
2
] = sup(T[A
1
] +T[A
2
])
= sup T[A
1
] + sup T[A
2
] = R
T
w
1
+R
T
w
2
by 351Dc. QQQ Next, it is easy to see that R
T
(w) = R
T
w for w U
+
, > 0, just because u u, v v
are isomorphisms of the partially ordered linear spaces U and V . It follows from 355D that we can extend
R
T
to a linear operator from U to V .
Because R
T
u T0 = 0 for every u U
+
, R
T
is a positive linear operator. But also R
T
u Tu for every
u U
+
, so R
T
T is also positive, and T = R
T
(R
T
T) is the dierence of two positive linear operators.
(ii) This shows that every order-bounded operator is a dierence of positive operators. But of course if
T
1
and T
2
are positive, then (T
1
T
2
)u T
1
w whenever 0 u w in U, so that T
1
T
2
is order-bounded,
by the criterion in 355Ba. Thus L

is precisely the set of dierences of positive operators.

(iii) Just as in 351F, L

is a partially ordered linear space if we say that S T i Su Tu for

every u U
+
. Now it is a Riesz space. PPP Take any T L

. Then R
T
, as dened in (i), is an upper
bound for 0, T in L

. If S L

is any other upper bound for 0, T, then for any w U

+
we must
have Sw Su Tu whenever u [0, w], so that Sw R
T
w; as w is arbitrary, S R
T
; as S is arbitrary,
R
T
= sup0, T in L

. Thus sup0, T is dened in L

for every T L

; by 352B, L

is a Riesz space. QQQ

(I defer the proof that it is Dedekind complete to (d-ii) below.)
(b) As remarked in (a-iii), R
T
= T
+
for each T L

; but this is just the formula given for T

+
. Now, if
T L

and w U
+
,
[T[(w) = 2T
+
w Tw = 2 sup
u[0,w]
Tu Tw
= sup
u[0,w]
T(2u w) = sup
u[w,w]
Tu,
which is the rst formula oered for [T[. In particular, if [u[ w then Tu, Tu = T(u) are both less than
or equal to [T[(w), so that [Tu[ [T[(w). So if u
0
, . . . , u
n
are such that

n
i=0
[u
i
[ = w, then

n
i=0
[Tu
i
[

n
i=0
[T[([u
i
[) = [T[(w).
Thus B =

n
i=0
[Tu
i
[ :

n
i=0
[u
i
[ = w is bounded above by [T[(w). On the other hand, if v is an upper
bound for B and [u[ w, then
Tu [Tu[ +[T(w [u[)[ v;
as u is arbitrary, [T[(w) v; thus [T[(w) is the least upper bound for B. This completes the proof of part
(ii) of the theorem.
(c) We know that S T = T + (S T)
+
(352D), so that
(S T)(w) = Tw + (S T)
+
(w) = Tw + sup
0uw
(S T)(u)
= sup
0uw
Tw + (S T)(u) = sup
0uw
Su +T(w u)
258 Riesz spaces 355E
for every w U
+
, by the formula in (b). Also from 352D we have S T = S +T T S, so that
(S T)(w) = Sw +Tw sup
0uw
Tu +S(w u)
= inf
0uw
Sw +Tw Tu S(w u)
(351Db)
= inf
0uw
Su +T(w u)
for w U
+
.
(d)(i) Now suppose that A L

is non-empty and upwards-directed and that Tu : T A is bounded

above in V for every u U
+
. In this case, because V is Dedekind complete, we may set Ru = sup
TA
Tu
for every u U
+
. Now R(u
1
+u
2
) = Ru
1
+Ru
2
for all u
1
, u
2
U
+
. PPP Set B
i
= Tu
i
: T A for each i,
B = T(u
1
+u
2
) : T A. Then B B
1
+B
2
, so
R(u
1
+u
2
) = sup B sup(B
1
+B
2
) = sup B
1
+ sup B
2
= Ru
1
+Ru
2
.
On the other hand, if v
i
B
i
for both i, there are T
i
A such that v
i
= T
i
u
i
for each i; because A is
upwards-directed, there is a T A such that T T
i
for both i, and now
R(u
1
+u
2
) T(u
1
+u
2
) = Tu
1
+Tu
2
T
1
u
1
+t
2
u
2
= v
1
+v
2
.
As v
1
, v
2
are arbitrary,
R(u
1
+u
2
) sup(B
1
+B
2
) = supB
1
+ supB
2
= Ru
1
+Ru
2
. QQQ
It is also easy to see that R(u) = Ru for every u U
+
, > 0. So, using 355D again, R has an extension
to a linear operator from U to V .
Now if we x any T
0
A, we have T
0
u Ru for every u U
+
, so R T
0
is a positive linear operator,
and R = (R T
0
) + T
0
belongs to L

. Again, Tu Ru for every T A, u U

+
, so R is an upper bound
for A in L

; and, nally, if S is any upper bound for A in L

, then Su is an upper bound for Tu : T A,

and must be greater than or equal to Ru, for every u U
+
; so that R S and R = supA in L

.
(ii) Now L

is Dedekind complete. PPP If A L

is non-empty and bounded above by S say, then

A
t
= T
0
T
1
. . . T
n
: T
0
, . . . , T
n
A is upwards-directed and bounded above by S, so Tu : T A
t

is bounded above by Su for every u U

+
; by (i) just above, A
t
has a supremum in L

, which will also be

a supremum for A. QQQ
(e) Suppose that A (L

)
+
is non-empty and downwards-directed. Then A = T : T A is
non-empty and upwards-directed, so
inf A = 0 sup(A) = 0
sup
TA
(Tw) = 0 for every w U
+
inf
TA
Tw = 0 for every w U
+
.
355F Theorem Let U and V be Riesz spaces and U
0
U a Riesz subspace which is either order-dense
or a solid linear subspace. Suppose that T
0
: U
0
V is an order-continuous positive linear operator such
that Su = supT
0
w : w U
0
, 0 w u is dened in V for every u U
+
. Then
(i) T
0
has an extension to an order-continuous positive linear operator T : U V .
(ii) If T
0
is a Riesz homomorphism so is T.
(iii) If U
0
is order-dense then T is unique.
(iv) If U
0
is order-dense and T
0
is an injective Riesz homomorphism, then so is T.
proof (a) I check rst that
S(u +u
t
) = Su +Su
t
, S(u) = Su
355F Spaces of linear operators 259
for all u, u
t
U
+
and > 0. PPP For every u U set A(u) = v : v U
0
, 0 v u. Note that for any
u U
+
, v U
+
0
v u = sup
wA(u)
v w = supA(v u);
this is because supA(u) = u if U
0
is order-dense in U, while v u A(u) if U
0
is solid. (i) If v A(u +u
t
)
then
sup(A(v u) +A(v u)
+
) = sup A(v u) + sup A(v u)
+
= v u + (v u)
+
= v.
Now
T
0
v = supT
0
[A(v u) +A(v u)
+
]
= sup(T
0
[A(v u)] +T
0
[A(v u)
+
]) Su +Su
t
because (v u)
+
u
t
. As v is arbitrary, S(u +u
t
) Su +Su
t
. (ii) Next,
Su +Su
t
= sup T
0
[A(u)] + supT
0
[A(u
t
)] = sup T
0
[A(u) +A(u
t
)]
supT
0
[A(u +u
t
)] = S(u +u
t
)
because A(u) +A(u
t
) A(u+u
t
), so S(u+u
t
) = Su+Su
t
. (iii) Of course A(u) = A(u) so S(u) = Su.
QQQ
By 355D, S has an extension to a linear operator from U to V ; call this operator T. Of course Tu = Su 0
whenever u 0, so T is positive. If u U
+
0
then Tu = Su = T
0
u, so T extends T
0
.
(b) If B U
+
is non-empty and upwards-directed and has a supremum u
0
U, then of course Tu Tu
0
for every u B, so sup T[B] Tu
0
. On the other hand, for any v A(u
0
) we have
v = sup
uB
v u = sup
uB
sup
wA(u)
v w;
also

uB
A(u) is upwards-directed, so
T
0
v = supT
0
(v w) : w

uB
A(u) sup
uB
Tu.
As v is arbitrary, Tu
0
= Su
0
sup
uB
Tu. As B is arbitrary, T is order-continuous (351Ga).
(c) Now suppose that T
0
is a Riesz homomorphism. If u U then, in the language of (a) above,
Tu
+
Tu

= supT
0
[A(u
+
)] T
0
[A(u

)]
= supw w
t
: w T
0
[A(u
+
)], w
t
T
0
[A(u

)]
(352Ea)
= supT
0
(v v
t
) : v A(u
+
), v
t
A(u

)
(because T
0
is a Riesz homomorphism)
= 0.
So T is a Riesz homomorphism (352G(iv)).
(d) If U
0
is order-dense, any order-continuous linear operator extending T
0
agrees with S and T on U
+
,
so is equal to T.
(e) Finally, if U
0
is order-dense and T
0
is an injective Riesz homomorphism, then for any non-zero u U
there is a non-zero v U
0
such that [v[ [u[; so that
[Tu[ = T[u[ T
0
[v[ > 0
because T is a Riesz homomorphism, by (c). As u is arbitrary, T is injective.
260 Riesz spaces 355G
355G Denition Let U be a Riesz space and V a Dedekind complete Riesz space. Then L

(U; V ) will be
the set of those T L

(U; V ) expressible as the dierence of order-continuous positive linear operators, and

L

c
(U; V ) will be the set of those T L

(U; V ) expressible as the dierence of sequentially order-continuous

positive linear operators.
Because a composition of (sequentially) order-continuous functions is (sequentially) order-continuous, we
shall have
ST L

(U; W) whenever S L

(V ; W), T L

(U; V ),
ST L

c
(U; W) whenever S L

c
(V ; W), T L

c
(U; V ),
for all Riesz spaces U and all Dedekind complete Riesz spaces V , W.
355H Theorem Let U be a Riesz space and V a Dedekind complete Riesz space. Then
(i) L

= L

(U; V ) is a band in L

= L

(U; V ), therefore a Dedekind complete Riesz space in its own

right
(ii) a member T of L

belongs to L

i [T[ is order-continuous.
proof There is a fair bit to check, but each individual step is easy enough.
(a) Suppose that S, T are order-continuous positive linear operators from U to V . Then S +T is order-
continuous. PPP If A U is non-empty, downwards-directed and has inmum 0, then for any u
1
, u
2
A
there is a u A such that u u
1
, u u
2
, and now (S +T)(u) Su
1
+Tu
2
. Consequently any lower bound
for (S +T)[A] must also be a lower bound for S[A] +T[A]. But since
inf(S[A] +T[A]) = inf S[A] + inf T[A] = 0
(351Dc), inf(S +T)[A] must also be 0; as A is arbitrary, S +T is order-continuous, by 351Ga. QQQ
(b) Consequently S + T L

for all S, T L

. Since T and T belong to L

for every T L

and
0, we see that L

is a linear subspace of L

.
(c) If T : U V is an order-continuous linear operator, S : U V is linear and 0 S T, then S is
order-continuous. PPP If A U is non-empty, downwards-directed and has inmum 0, then any lower bound
of S[A] must also be a lower bound of T[A], so inf S[A] = 0; as A is arbitrary, S is order-continuous. QQQ
It follows that L

is a solid linear subspace of L

. PPP If T L

and [S[ [T[ in L

, express T as T
1
T
2
where T
1
, T
2
are order-continuous positive linear operators. Then
S
+
, S

[S[ [T[ T
1
+T
2
,
so S
+
and S

are order-continuous and S = S

+
S

. QQQ
Accordingly L

is a Dedekind complete Riesz space in its own right (353J(b-i)).

(d) The argument of (c) also shows that if T L

then [T[ is order-continuous; so that for T L

,
T L

[T[ L

[T[ is order-continuous.
(e) If C (L

)
+
is non-empty, upwards-directed and has a supremum T L

, then T is order-
continuous, so belongs to L

. PPP Suppose that A U

+
is non-empty, upwards-directed and has supremum
w. Then
Tw = sup
SC
Sw = sup
SC
sup
uA
Su = sup
uA
Tu,
putting 355Ed and 351G(a-iii) together. So (using 351Ga again) T is order-continuous. QQQ Consequently
L

is a band in L

(352Ob).
This completes the proof.
355I Theorem Let U be a Riesz space and V a Dedekind complete Riesz space. Then L

c
(U; V ) is a
band in L

(U; V ), and a member T of L

(U; V ) belongs to L

c
(U; V ) i [T[ is sequentially order-continuous.
proof Copy the arguments of 355H.
355Xe Spaces of linear operators 261
355J Proposition Let U be a Riesz space and V a Dedekind complete Riesz space. Let U
0
U be
an order-dense Riesz subspace; then T TU
0
is an embedding of L

(U; V ) as a solid linear subspace of

L

(U
0
; V ). In particular, any T
0
L

(U
0
; V ) has at most one extension in L

(U; V ).
proof (a) Because the embedding U
0
U is positive and order-continuous (352Nb), TU
0
is positive and
order-continuous whenever T is; so TU
0
L

(U
0
; V ) whenever T L

(U; V ). Because the map T TU

0
is linear, the image W of L

(U; V ) is a linear subspace of L

(U
0
; V ).
(b) If T L

(U; V ) and TU
0
0, then T 0. PPP??? Suppose, if possible, that there is a u U
+
such
that Tu , 0. Because [T[ L

(U; V ) is order-continuous and A = v : v U

0
, v u is an upwards-
directed set with supremum u, inf[T[(uv) : v A = 0 and there is a v A such that Tu+[T[(uv) , 0.
But Tv = Tu +T(v u) Tu +[T[(u v) so Tv , 0 and TU
0
, 0. XXXQQQ
This shows that the map T TU
0
is an order-isomorphism between L

(U; V ) and W, and in particular

is injective.
(c) Now suppose that S
0
W and that [S[ [S
0
[ in L

(U
0
; V ). Then S W. PPP Take T
0
L

(U; V )
such that T
0
U
0
= S
0
. Then S
1
= [T
0
[U
0
is a positive member of W such that S
0
S
1
, S
0
S
1
, so
S
+
S
1
. Consequently, for any u U
+
,
supS
+
v : v U
0
, 0 v u supS
1
v : v U
0
, 0 v u [T
0
[(u)
is dened in V (recall that we are assuming that V is Dedekind complete). But this means that S
+
has an
extension to an order-continuous positive linear operator from U to V (355F), and belongs to W. Similarly,
S

W, so S W. QQQ
This shows that W is a solid linear subspace of L

(U
0
; V ), as claimed.
355K Proposition Let U be a Banach lattice with an order-continuous norm.
(a) If V is any Archimedean Riesz space and T : U V is a positive linear operator, then T is order-
continuous.
(b) If V is a Dedekind complete Riesz space then L

(U; V ) = L

(U; V ).
proof (a) Suppose that A U
+
is non-empty and downwards-directed and has inmum 0. Then for
each n N there is a u
n
A such that |u
n
| 4
n
. By 354C, u = sup
nN
2
n
u
n
is dened in U. Now
Tu
n
2
n
Tu for every n, so any lower bound for T[A] must also be a lower bound for 2
n
Tu : n N
and therefore (because V is Archimedean) less than or equal to 0. Thus inf T[A] = 0; as A is arbitrary, T
is order-continuous.
(b) This is now immediate from 355Ea and the denition of L

.
355X Basic exercises >>>(a) Let U and V be arbitrary Riesz spaces. (i) Show that the set L(U; V ) of
all linear operators from U to V is a partially ordered linear space if we say that S T whenever Su Tu
for every u U
+
. (ii) Show that if U and V are Banach lattices then the set of positive operators is closed
in the normed space B(U; V ) of bounded linear operators from U to V .
>>>(b) If U is a Riesz space and | |, | |
t
are two norms on U both rendering it a Banach lattice, show that
they are equivalent, that is, give rise to the same topology.
(c) Let U be a Riesz space with a Riesz norm, V an Archimedean Riesz space with an order unit, and
T : U V a linear operator which is continuous for the given norm on U and the order-unit norm on V .
Show that T is order-bounded.
(d) Let U be a Riesz space, V an Archimedean Riesz space, and T : U
+
V
+
a map such that
T(u
1
+u
2
) = Tu
1
+Tu
2
for all u
1
, u
2
U
+
. Show that T has an extension to a linear operator from U to
V .
>>>(e) Show that if r, s 1 are integers then the Riesz space L

(R
r
; R
s
) can be identied with the
space of real s r matrices, saying that a matrix is positive i every coecient is positive, so that if
T =
ij
)
1is,1jr
then [T[, taken in L

(R
r
; R
s
), is [
ij
[)
1is,1jr
. Show that a matrix represents a
Riesz homomorphism i each row has at most one non-zero coecient.
262 Riesz spaces 355Xf
>>>(f ) Let U be a Riesz space and V a Dedekind complete Riesz space. Show that if T
0
, . . . , T
n
L

(U; V )
then
(T
0
. . . T
n
)(w) = sup

n
i=0
T
i
u
i
: u
i
0 i n,

n
i=0
u
i
= w
for every w U
+
.
>>>(g) Let U be a Riesz space, V a Dedekind complete Riesz space, and A L

(U; V ) a non-empty
set. Show that A is bounded above in L

(U; V ) i C
w
=

n
i=0
T
i
u
i
: T
0
, . . . , T
n
A, u
0
, . . . , u
n

U
+
,

n
i=0
u
i
= w is bounded above in V for every w U
+
, and in this case (supA)(w) = supC
w
for every
w U
+
.
355Y Further exercises (a) Let U and V be Banach lattices. For T L

= L

(U; V ), set |T|

=
sup
wU
+
,|w|1
inf|v| : [Tu[ v whenever [u[ w. Show that | |

is a norm on L

under which L

is a
Banach space, and that the set of positive linear operators is closed in L

.
(b) Give an example of a continuous linear operator from
2
to itself which is not order-bounded.
(c) Let U and V be Riesz spaces and T : U V a linear operator. (i) Show that for any w U
+
,
C
w
=

n
i=0
[Tu
i
[ : u
0
, . . . , u
n
U
+
,

n
i=0
u
i
= w is upwards-directed, and has the same upper bounds
as Tu : [u[ w. (Hint: induce on m and n to see that if u
0
, . . . , u
n
, u
t
0
, . . . , u
t
m
U
+
are such
that

n
i=0
u
i
=

m
j=0
u
t
j
, there is a family u
ij
)
in,jm
in U
+
such that

m
j=0
u
ij
= u
i
for every i n,

n
i=0
u
ij
= u
t
j
for every j m.) (ii) Show that if supC
w
is dened for every w U
+
, then S = T (T)
is dened in the partially ordered linear space L

(U; V ) and Sw = sup C

w
for every w U
+
.
(d) Let U, V and W be Riesz spaces, of which V and W are Dedekind complete. (i) Show that for
any S L

(V ; W), the map T ST : L

(U; V ) L

(U; W) belongs to L

(L

(U; V ); L

(U; W)),
and is a Riesz homomorphism if S is. (Hint: 355Yc.) (ii) Show that for any T L

(U; V ), the map

S ST : L

(V ; W) L

(U; W) belongs to L

(L

(V ; W); L

(U; W)).
(e) Let be the usual measure on 0, 1
N
and ccc the Banach lattice of convergent sequences. Find a linear
operator T : L
2
() ccc which is norm-continuous, therefore order-bounded, such that 0 and T have no
common upper bound in the partially ordered linear space of all linear operators from L
2
() to ccc.
(f ) Let U and V be Banach lattices. Let L
reg
be the linear space of operators from U to V expressible as
the dierence of positive operators. For T L
reg
let |T|
reg
be
inf|T
1
+T
2
| : T
1
, T
2
: U V are positive, T = T
1
T
2
.
Show that | |
reg
is a norm under which L
reg
is complete.
(g) Let U and V be Riesz spaces. For this exercise only, say that L

(U; V ) is to be the set of linear

operators T : U V such that whenever A U is non-empty, downwards-directed and has inmum 0 then
v : v V
+
, w A, [Tu[ v whenever [u[ w has inmum 0 in V . (i) Show that L

(U; V ) is a linear
space. (ii) Show that if U is Archimedean then L

(U; V ) L

(U; V ). (iii) Show that if U is Archimedean

and V is Dedekind complete then this denition agrees with that of 355G. (iv) Show that for any Riesz
spaces U, V and W, ST L

(U; W) for every S L

(V ; W), T L

(U; V ). (v) Show that if U and V

are Banach lattices, then L

(U; V ) is closed in L

(U; V ) for the norm | |

of 355Ya. (vi) Show that if V is

Archimedean and U is a Banach lattice with an order-continuous norm, then L

(U; V ) = L

(U; V ).
(h) Let U be a Riesz space and V a Dedekind complete Riesz space. Show that the band projection
P : L

(U; V ) L

(U; V ) is given by the formula

(PT)(w) = infsup
uA
Tu : A U
+
is non-empty, upwards-directed
and has supremum w
for every w U
+
, T (L

)
+
. (Cf. 362Bd.)
355 Notes Spaces of linear operators 263
(i) Show that if U is a Riesz space with the countable sup property (241Yd), then L

c
(U; V ) = L

(U; V )
for every Dedekind complete Riesz space V .
(j) Let U and V be Riesz spaces, of which V is Dedekind complete, and U
0
a solid linear subspace of
U. Show that the map T TU
0
is an order-continuous Riesz homomorphism from L

(U; V ) onto a solid

linear subspace of L

(U
0
; V ).
(k) Let U be a uniformly complete Riesz space (354Yi) and V a Dedekind complete Riesz space. Let U
C
,
V
C
be their complexications (354Yk). Show that the complexication of L

(U; V ) can be identied with

the complex linear space of linear operators T : U
C
V
C
such that B
T
(w) = [Tu[ : [u[ w is bounded
above in V for every w U
+
, and that now [T[(w) = supB
T
(w) for every T L

(U; V )
C
, w U
+
. (Hint:
if u, v U and [u+iv[ = w, then u and v can be simultaneously approximated for the order-unit norm | |
w
on the solid linear subspace generated by w by nite sums

n
j=0
(cos
j
)w
j
,

n
j=0
(sin
j
)w
j
where w
j
U
+
,

n
j=0
w
j
= w. Consequently [T(u +iv)[ [T[(w) for every T L

C
.)
355 Notes and comments I have had to make some choices in the basic denitions of this chapter
(355A, 355G). For Dedekind complete codomains V , there is no doubt what L

(U; V ) should be, since

the order-bounded operators (in the sense of 355A) are just the dierences of positive operators (335Ea).
(These are sometimes called regular operators.) When V is not Dedekind complete, we have to choose
between the two notions, as not every order-bounded operator need be regular (355Ye). In my previous book
(Fremlin 74a) I chose the regular operators; I have still not encountered any really persuasive reason to
settle denitively on either class. In 355G the technical complications in dealing with any natural equivalent
of the larger space (see 355Yg) are such that I have settled for the narrower class, but explicitly restricting
the denition to the case in which V is Dedekind complete. In the applications in this book, the codomains
are nearly always Dedekind complete, so we can pass these questions by.
The elementary extension technique in 355D may recall the denition of the Lebesgue integral (122L-
122M). In the same way, 351G may remind you of the theorem that a linear operator between normed
spaces is continuous everywhere if it is continuous anywhere, or of the corresponding results about Boolean
homomorphisms and additive functionals on Boolean algebras (313L, 326Ga, 326N).
Of course 355Ea is the central fact about the space L

(U; V ) for Dedekind complete V ; because we get a

new Riesz space from old ones, the prospect of indenite recursion immediately presents itself. For Banach
lattices, L

(U; V ) is a linear subspace of the space B(U; V ) of bounded linear operators (355C); the question
of when the two are equal will be of great importance to us. I give only the vaguest hints on how to show
that they can be dierent (355Yb, 355Ye), but these should be enough to make it plain that equality is
the exception rather than the rule. It is also very useful that we have eective formulae to describe the
Riesz space operations on L

(U; V ) (355E, 355Xf-355Xg, 355Yc). You may wish to compare these with the
corresponding formulae for additive functionals on Boolean algebras in 326Yj and 362B.
If we think of L

as somehow corresponding to the space of bounded additive functionals on a Boolean

algebra, the bands L

c
and L

correspond to the spaces of countably additive and completely additive

functionals. In fact (as will appear in 362) this correspondence is very close indeed. For the moment, all
I have sought to establish is that L

c
and L

are indeed bands. Of course any case in which L

(U; V ) =
L

c
(U; V ) or L

c
(U; V ) = L

(U; V ) is of interest (355Kb, 355Yi).

Between Banach lattices, positive linear operators are continuous (355C); it follows at once that the Riesz
space structure determines the topology (355Xb), so that it is not to be wondered at that there are further
connexions between the norm and the spaces L

and L

, as in 355K.
355F will be a basic tool in the theory of representations of Riesz spaces; if we can represent an order-
dense Riesz subspace of U as a subspace of a Dedekind complete space V , we have at least some chance of
expressing U also as a subspace of V . Of course it has other applications, starting with analysis of the dual
spaces.
264 Riesz spaces 356 intro.
356 Dual spaces
As always in functional analysis, large parts of the theory of Riesz spaces are based on the study of linear
functionals. Following the scheme of the last section, I dene spaces U

, U

c
and U

, the order-bounded,
sequentially order-continuous and order-continuous duals of a Riesz space U (356A). These are Dedekind
complete Riesz spaces (356B). If U carries a Riesz norm they are closely connected with the normed space
dual U

, which is itself a Banach lattice (356D). For each of them, we have a canonical Riesz homomorphism
from U to the corresponding bidual. The map from U to U

is particularly important (356I); when this

map is an isomorphism we call U perfect (356J). The last third of the section deals with L- and M-spaces
and the duality between them (356N, 356P), with two important theorems on uniform integrability (356O,
356Q).
356A Denition Let U be a Riesz space.
(a) I write U

for the space L

(U; R) of order-bounded real-valued linear functionals on U, the order-

bounded dual of U.
(b) U

c
will be the space L

c
(U; R) of dierences of sequentially order-continuous positive real-valued
linear functionals on U, the sequentially order-continuous dual of U.
(c) U

will be the space L

(U; R) of dierences of order-continuous positive real-valued linear functionals

on U, the order-continuous dual of U.
Remark It is easy to check that the three spaces U

, U

c
and U

are in general dierent (356Xa-356Xc).

But the examples there leave open the question: can we nd a Riesz space U, for which U

c
,= U

, and
which is actually Dedekind complete, rather than just Dedekind -complete, as in 356Xc? This leads to
unexpectedly deep water; it is yet another form of the Banach-Ulam problem. Really this is a question for
Volume 5, but in 363S below I collect the relevant ideas which are within the scope of the present volume.
356B Theorem For any Riesz space U, its order-bounded dual U

is a Dedekind complete Riesz space

in which U

c
and U

are bands, therefore Dedekind complete Riesz spaces in their own right. For f U

,
f
+
and [f[ U

are dened by the formulae

f
+
(w) = supf(u) : 0 u w, [f[(w) = supf(u) : [u[ w
for every w U
+
. A non-empty upwards-directed set A U

is bounded above i sup

fA
f(u) is nite for
every u U, and in this case (supA)(u) = sup
fA
f(u) for every u U
+
.
proof 355E, 355H, 355I.
356C Proposition Let U be any Riesz space and P a band projection on U. Then its adjoint P
t
:
U

, dened by setting P
t
(f) = fP for every f U

, is a band projection on U

.
proof (a) Because P : U U is a positive linear operator, P
t
f U

for every f U

(355Bd), and P
t
is
a positive linear operator from U

to itself. Set Q = I P, the complementary band projection on U; then

Q
t
is another positive linear operator on U

, and P
t
f +Q
t
f = f for every f. Now P
t
f Q
t
f = 0 for every
f 0. PPP For any w U
+
,
(P
t
f Q
t
f)
+
(w) = sup
0uw
(P
t
f Q
t
f)(u) = sup
0uw
f(Pu Qu)
= f(Pw) = (P
t
f)(w),
so (P
t
f Q
t
f)
+
= P
t
f, that is, P
t
f Q
t
f = 0. QQQ By 352Rd, P
t
is a band projection.
356D Proposition Let U be a Riesz space with a Riesz norm.
(a) The normed space dual U

of U is a solid linear subspace of U

, and in itself is a Banach lattice with

a Fatou norm and has the Levi property.
356E Dual spaces 265
(b) The norm of U is order-continuous i U

.
(c) If U is a Banach lattice, then U

= U

, so that U

, U

and U

c
are all Banach lattices.
(d) If U is a Banach lattice with order-continuous norm then U

= U

= U

.
proof (a)(i) If f U

then
sup
]u]w
f(u) sup
]u]w
|f||u| = |f||w| <
for every w U
+
, so f U

(355Ba). Thus U

.
(ii) If f, g U

and [f[ [g[, then for any w U

[f(w)[ [f[([w[) [g[([w[) = sup
]u]]w]
g(u) sup
]u]]w]
|g||u| |g||w|.
As w is arbitrary, f U

and |f| |g|; as f and g are arbitrary, U

is a solid linear subspace of U

and
the norm of U

is a Riesz norm. Because U

is a Banach space it is also a Banach lattice.

(iii) If A (U

)
+
is non-empty, upwards-directed and M = sup
fA
|f| is nite, then sup
fA
f(u)
M|u| is nite for every u U
+
, so g = supA is dened in U

(355Ed). Now g(u) = sup

fA
f(u) for every
u U
+
, as also noted in 355Ed, so
[g(u)[ g([u[) M|[u[| = M|u|
for every u U, and |g| M. But as A is arbitrary, this simultaneously proves that the norm of U

is
Fatou and has the Levi property.
(b)(i) Suppose that the norm is order-continuous. If f U

and A U is a non-empty downwards-

directed set with inmum 0, then
inf
uA
[f[(u) inf
uA
|f||u| = 0,
so [f[ U

and f U

. Thus U

.
(ii) Now suppose that the norm is not order-continuous. Then there is a non-empty downwards-directed
set A U, with inmum 0, such that inf
uA
|u| = > 0. Set
B = v : v u for some u A.
Then B is convex. PPP If v
1
, v
2
B and [0, 1], there are u
1
, u
2
A such that v
i
u
i
for both i; now
there is a u A such that u u
1
u
2
, so that
u = u + (1 )u v
1
+ (1 )v
2
,
and v
1
+ (1 )v
2
B. QQQ Also inf
vB
|v| = > 0. By the Hahn-Banach theorem (3A5Cb), there is an
f U

such that inf

vB
f(v) > 0. But now
inf
uA
[f[(u) inf
uA
f(u) > 0
and [f[ is not order-continuous; so U

, U

.
(c) By 355C, U

, so U

= U

. Now U

and U

c
, being bands, are closed linear subspaces (354Bd),
so are Banach lattices in their own right.
(d) Put (b) and (c) together.
356E Biduals If you have studied any functional analysis at all, it will come as no surprise that
duals-of-duals are important in the theory of Riesz spaces. I start with a simple lemma.
Lemma Let U be a Riesz space and f : U R a positive linear functional. Then for any u U
+
there is
a positive linear functional g : U R such that 0 g f, g(u) = f(u) and g(v) = 0 whenever u v = 0.
proof Set g(v) = sup
0
f(v u) for every v U
+
. Then it is easy to see that g(v) = g(v) for every
v U
+
, [0, [. If v, w U
+
then
(v u) + (w u) (v +w) 2u (v 2u) + (w 2u)
for every 0 (352Fa), so g(v + w) = g(v) + g(w). Accordingly g has an extension to a linear functional
from U to R (355D). Of course 0 g(v) f(v) for v 0, so 0 g f in U

. We have g(u) = f(u), while

if u v = 0 then u v = 0 for every 0, so g(v) = 0.
266 Riesz spaces 356F
356F Theorem Let U be a Riesz space and V a solid linear subspace of U

. For u U dene u : V R
by setting u(f) = f(u) for every f V . Then u u is a Riesz homomorphism from U to V

.
proof (a) By the denition of addition and scalar multiplication in V , u is linear for every u; also u = u,
(u
1
+ u
2
)= u
1
+ u
2
for all u, u
1
, u
2
U and R. If u 0 then u(f) = f(u) 0 for every f V
+
, so
u 0; accordingly every u is the dierence of two positive functionals, and u u is a linear operator from
U to V

.
(b) If B V is a non-empty downwards-directed set with inmum 0, then inf
fB
f(u) = 0 for every
u U
+
, by 355Ee. But this means that u is order-continuous for every u U
+
, so that u V

for every
u U.
(c) If uv = 0 in U, then for any f V
+
there is a g [0, f] such that g(u) = f(u), g(v) = 0 (356D). So
( u v)(f) u(f g) + v(g) = f(u) g(u) +g(v) = 0.
As f is arbitrary, u v = 0. As u and v are arbitrary, u u is a Riesz homomorphism (352G).
356G Lemma Suppose that U is a Riesz space such that U

separates the points of U. Then U is

Archimedean.
proof ??? Otherwise, there are u, v U such that v > 0 and nv u for every n N. Now there is an
f U

such that f(v) ,= 0; but f(v) [f[(v)

1
n
[f[(u) for every n, so this is impossible. XXX
356H Lemma Let U be an Archimedean Riesz space and f > 0 in U

. Then there is a u U such

that (i) u > 0 (ii) f(v) > 0 whenever 0 < v u (iii) g(u) = 0 whenever g f = 0 in U

. Moreover, if
u
0
U
+
is such that f(u
0
) > 0, we can arrange that u u
0
.
proof (Because f > 0 there certainly is some u
0
U such that f(u
0
) > 0.)
(a) Set A = v : 0 v u
0
, f(v) = 0. Then (v
1
+ v
2
) u
0
A for all v
1
, v
2
A, so A is upwards-
directed. Because f(u
0
) > 0 = supf[A] and f is order-continuous, u
0
cannot be the least upper bound of
A, and there is another upper bound u
1
of A strictly less than u
0
.
Set u = u
0
u
1
> 0. If 0 v u and f(v) = 0, then
w A = w u
1
= w +v u
0
= w +v A;
consequently nv A and nv u
0
for every n N, so v = 0. Thus u has properties (i) and (ii).
(b) Now suppose that g f = 0 in U

. Let > 0. Then for each n N there is a v

n
[0, u] such that
f(v
n
) + g(u v
n
) 2
n
(355Ec). If v v
n
for every n N then f(v) = 0 so v = 0; thus inf
nN
v
n
= 0.
Set w
n
= inf
in
v
i
for each n N; then w
n
)
nN
is non-increasing and has inmum 0 so (because g is
order-continuous) inf
nN
g(w
n
) = 0. But
u w
n
= sup
in
u v
i

n
i=0
u v
i
,
so
g(u w
n
)

n
i=0
g(u v
i
) 2
for every n, and
g(u) 2 + inf
nN
g(w
n
) = 2.
As is arbitrary, g(u) = 0; as g is arbitrary, u has the third required property.
356I Theorem Let U be any Archimedean Riesz space. Then the canonical map from U to U

(356F)
is an order-continuous Riesz homomorphism from U onto an order-dense Riesz subspace of U

. If U is
Dedekind complete, its image in U

is solid.
proof (a) By 356F, u u : U U

is a Riesz homomorphism.
To see that it is order-continuous, take any non-empty downwards-directed set A U with inmum 0.
Then C = u : u A is downwards-directed, and for any f (U

)
+
356K Dual spaces 267
inf
C
(f) = inf
uA
f(u) = 0
because f is order-continuous. As f is arbitrary, inf C = 0 (355Ee); as A is arbitrary, u u is order-
continuous (351Ga).
(b) Now suppose that > 0 in U

. By 356H, there is an f > 0 in U

such that (f) > 0 and (g) = 0

whenever g f = 0. Next, there is a u > 0 in U such that f(u) > 0. Since U 0, u 0; since u(f) > 0,
u > 0.
Because U

(being Dedekind complete) is Archimedean, inf

>0
u = 0, and there is an > 0 such that
= ( u u)
+
> 0.
Let g (U

)
+
be such that (g) > 0 and (g) = 0 whenever = 0 in U

. Let v U
+
be such that
g(v) > 0 and h(v) = 0 whenever h g = 0 in U

.
Because v(g) = g(v) > 0, v > 0. As u, v u > 0 and v u > 0. Set w = vu; then w = v u,
by 356F, so w > 0.
??? Suppose, if possible, that w , . Then = ( w)
+
> 0, so there is an h (U

)
+
such that (h) > 0
and (h
t
) > 0 whenever 0 < h
t
h (356H, for the fourth and last time). Now examine
(h g) ( u u)
+
(g)
(because w u, u , h g g)
= 0
because ( u u)
+
= 0. So h g = 0 and h(v) = 0. But this means that
(h) w(h) v(h) = 0,
which is impossible. XXX
Thus 0 < w . As is arbitrary, the image

U of U is quasi-order-dense in U

, therefore order-dense
(353A).
(c) Now suppose that U is Dedekind complete and that 0

U. Express as u where u U,
and set A = v : v U, v u
+
, v . If v U and 0 v , then w = v
+
u
+
A and w = v; thus
= sup v : v A = v
0
, where v
0
= supA. So

U. As and are arbitrary,

U is solid in U

.
356J Denition A Riesz space U is perfect if the canonical map from U to U

is an isomorphism.
356K Proposition A Riesz space U is perfect i (i) it is Dedekind complete (ii) U

separates the
points of U (iii) whenever A U is non-empty and upwards-directed and f(u) : u A is bounded for
every f U

, then A is bounded above in U.

proof (a) Suppose that U is perfect. Because it is isomorphic to U

, which is surely Dedekind complete,

U also is Dedekind complete. Because the map u u : U U

is injective, U

separates the points of

U. If A U is non-empty and upwards-directed ad f(u) : u A is bounded above for every f U

, then
B = u : u A is non-empty and upwards-directed and sup
B
(f) < for every f U

, so supB is
dened in U

(355Ed); but U

is a band in U

, so supB U

and is of the form w for some w U.

Because u u is a Riesz space isomorphism, w = sup A in U. Thus U satises the three conditions.
(b) Suppose that U satises the three conditions. We know that u u is a Riesz homomorphism onto
an order-dense Riesz subspace of U

(356I). It is injective because U

separates the points of U. If 0

in U

, set A = u : u U
+
, u . Then A is non-empty and upwards-directed and for any f U

sup
uA
f(u) sup
uA
[f[(u) sup
uA
u([f[) ([f[) < ,
so by condition (iii) A has an upper bound in U. Since U is Dedekind complete, w = sup A is dened in U.
Now
w = sup
uA
u = .
As is arbitrary, the image of U includes (U

)
+
, therefore is the whole of U

, and u u is a bijective
Riesz homomorphism, that is, a Riesz space isomorphism.
268 Riesz spaces 356L
356L Proposition (a) Any band in a perfect Riesz space is a perfect Riesz space in its own right.
(b) For any Riesz space U, U

is perfect; consequently U

c
and U

are perfect.
proof (a) I use the criterion of 356K. Let U be a perfect Riesz space and V a band in U. Then V is
Dedekind complete because U is (353Jb). If v V 0 there is an f U

such that f(v) ,= 0; but the

embedding V U is order-continuous (352N), so g = fV belongs to V

, and g(v) ,= 0. Thus V

separates
the points of V . If A V is non-empty and upwards-directed and sup
vA
g(v) is nite for every g V

,
then sup
vA
f(v) < for every f U

(again because fV V

), so A has an upper bound in U; because

U is Dedekind complete, supA is dened in U; because V is a band, supA V and is an upper bound for
A in V . Thus V satises the conditions of 356K and is perfect.
(b) U

is Dedekind complete, by 355Ea. If f U

0, there is a u U such that f(u) ,= 0; now

u(f) ,= 0, where u U

(356F). Thus U

separates the points of U

. If A U

is non-empty and
upwards-directed and sup
fA
(f) is nite for every U

, then, in particular,
sup
fA
f(u) = sup
fA
u(f) <
for every u U, so A is bounded above in U

, by 355Ed. Thus U

satises the conditions of 356K and is

perfect.
By (a), it follows at once that U

and U

c
are perfect.
356M Proposition If U is a Banach lattice in which the norm is order-continuous and has the Levi
property, then U is perfect.
proof By 356Db, U

= U

; since U

surely separates the points of U, so does U

. By 354Ee, U is
Dedekind complete. If A U is non-empty and upwards-directed and f[A] is bounded for every f U

,
then A is norm-bounded, by the Uniform Boundedness Theorem (3A5Hb). Because the norm is supposed to
have the Levi property, A is bounded above in U. Thus U satises all the conditions of 356K and is perfect.
356N L- and M-spaces I come now to the duality between L-spaces and M-spaces which I hinted at
in 354.
Proposition Let U be an Archimedean Riesz space with an order-unit norm.
(a) U

= U

is an L-space.
(b) If e is the standard order unit of U, then |f| = [f[(e) for every f U

.
(c) A linear functional f : U R is positive i it belongs to U

and |f| = f(e).

(d) If e ,= 0 there is a positive linear functional f on U such that f(e) = 1.
proof (a)-(b) We know already that U

is a Banach lattice (356Da). If f U

then
sup[f(u)[ : |u| 1 = sup[f(u)[ : [u[ e = [f[(e),
so f U

and |f| = [f[(e); thus U

= U

. If f, g 0 in U

, then
|f +g| = (f +g)(e) = f(e) +g(e) = |f| +|g|;
thus U

is an L-space.
(c) As already remarked, if f is positive then f U

and |f| = f(e). On the other hand, if f U

and
|f| = f(e), take any u 0. Set v = (1 +|u|)
1
u. Then 0 v e and |e v| 1 and
f(e v) [f(e v)[ |f| = f(e).
But this means that f(v) 0 so f(u) 0. As u is arbitrary, f 0.
(d) By the Hahn-Banach theorem (3A5Ac), there is an f U

such that f(e) = |f| = 1; by (c), f is

positive.
356O Theorem Let U be an Archimedean Riesz space with order-unit norm. Then a set A U

= U

is uniformly integrable i it is norm-bounded and lim

n
sup
fA
[f(u
n
)[ = 0 for every order-bounded
disjoint sequence u
n
)
nN
in U
+
.
356O Dual spaces 269
proof (a) Suppose that A is uniformly integrable. Then it is surely norm-bounded (354Ra). If u
n
)
nN
is
a disjoint sequence in U
+
bounded above by w, then for any > 0 we can nd an h 0 in U

such that
|([f[ h)
+
| for every f A. Now

n
i=0
h(u
i
) h(w) for every n, and lim
n0
h(u
n
) = 0; since at the
same time
[f(u
n
)[ [f[(u
n
) h(u
n
) + ([f[ h)
+
(u
n
) h(u
n
) +|u
n
| h(u
n
) +|w|
for every f A, n N, limsup
n
sup
fA
[f[(u
n
) |w|. As is arbitrary,
lim
n
sup
fA
[f[(u
n
) = 0,
and the conditions are satised.
(b)(i) Now suppose that A is norm-bounded but not uniformly integrable. Write B for the solid hull of
A, M for sup
fA
|f| = sup
fB
|f|; then there is a disjoint sequence g
n
)
nN
in B (U

)
+
which is not
norm-convergent to 0 (354R(b-iv)), that is,
=
1
2
limsup
n
g
n
(e) =
1
2
limsup
n
|g
n
| > 0,
where e is the standard order unit of U.
(ii) Set
C = v : 0 v e, sup
gB
g(v) ,
D = w : 0 w e, limsup
n
g
n
(w) > .
Then for any u D we can nd v C, w D such that v w = 0. PPP Set
t
= limsup
n
g
n
(u),
= (
t
)/(3 +M) > 0; take k N so large that k M.
Because g
n
(u)
t
for innitely many n, we can nd a set K N, of size k, such that g
i
(u)
t

for every i K. Now we know that, for each i K, g

i
k

jK,j,=i
g
j
= 0, so there is a v
i
u such that
g
i
(u v
i
) +k

jK,j,=i
g
j
(v
i
) (355Ec). Now
g
i
(v
i
) g
i
(u)
t
2, g
i
(v
j
)

k
for i, j K, i ,= j.
Set v
t
i
= (v
i

jK,j,=i
v
j
)
+
for each i K; then
g
i
(v
t
i
) g
i
(v
i
)

jK,j,=i
g
i
(v
j
)
t
3
for every i K, while v
t
j
v
t
i
= 0 for distinct i, j K.
For each n N,

iK
g
n
(u
1

v
t
i
) g
n
(u) |g
n
| k,
so there is some i(n) K such that
g
n
(u
1

v
t
i(n)
) , g
n
(u
1

v
t
i(n)
)
+
g
n
(u) .
Since n : g
n
(u) + 2 is innite, there is some m K such that J = n : g
n
(u) + 2, i(n) = m is
innite. Try
v = (v
t
m
u)
+
, w = (u
1

v
t
m
)
+
.
Then v, w [0, u] and v w = 0. Next,
g
m
(v) g
m
(v
t
m
) M
t
3 M = ,
so v C, while for any n J
g
n
(w) = g
n
(u
1

v
t
i(n)
)
+
g
n
(u) +;
since J is innite,
limsup
n
g
n
(w) + >
270 Riesz spaces 356O
and w D. QQQ
(iii) Since e D, we can choose inductively sequences w
n
)
nN
in D, v
n
)
nN
in C such that w
0
= e,
v
n
w
n+1
= 0, v
n
w
n+1
w
n
for every n N. But in this case v
n
)
nN
is a disjoint order-bounded
sequence in [0, u], while for each n N, we can nd f
n
A such that [f
n
[(v
n
) >
2
3
. Now there is a
u
n
[0, v
n
] such that [f
n
(u
n
)[
1
3
. PPP Set = sup
0vv
n
[f
n
(v)[. Then f
+
n
(v
n
), f

n
(v
n
) are both less than
or equal to , so [f
n
[(v
n
) 2 and >
1
3
; so there is a u
n
[0, v
n
] such that [f
n
(u
n
)[
1
3
. QQQ
Accordingly we have a disjoint sequence u
n
)
nN
in [0, e] such that sup
fA
[f(u
n
)[
1
3
for every n N.
(iv) All this is on the assumption that A is norm-bounded and not uniformly integrable. So, turning it
round, we see that if A is norm-bounded and lim
n
sup
fA
[f(u
n
)[ = 0 for every order-bounded disjoint
sequence u
n
)
nN
, A must be uniformly integrable.
This completes the proof.
356P Proposition Let U be an L-space.
(a) U is perfect.
(b) U

= U

= U

is an M-space; its standard order unit is the functional

_
dened by setting
_
u = |u
+
| |u

| for every u U.
(c) If A U is non-empty and upwards-directed and sup
uA
_
u is nite, then supA is dened in U and
_
supA = sup
uA
_
u.
proof (a) By 354N we know that the norm on U is order-continuous and has the Levi property, so 356M
tells us that U is perfect.
(b) 356Dd tells us that U

= U

= U

.
The L-space property tells us that the functional u |u| : U
+
R is additive; of course it is also
homogeneous, so by 355D it has an extension to a linear functional
_
: U R satisfying the given formula.
Because
_
u = |u| 0 for u 0,
_
(U

)
+
. For f U

,
[f[
_
[f[(u)
_
u for every u U
+
[f(v)[ |u| whenever [v[ u U
[f(v)[ |v| for every v U
|f| 1,
so the norm on U

= U

is the order-unit norm dened from

_
, and U

is an M-space, as claimed.
(c) Fix u
0
A, and set B = u
+
: u A, u u
0
. Then B U
+
is upwards-directed, and
sup
vB
|v| = sup
uA,uu
0
_
u
+
= sup
uA,uu
0
_
u +
_
u

sup
uA,uu
0
_
u +
_
u

0
< .
Because | | has the Levi property, B is bounded above. But (because A is upwards-directed) every member
of A is dominated by some member of B, so A also is bounded above. Because U is Dedekind complete,
supA is dened in U. Finally,
_
supA = sup
uA
_
u because
_
, being a positive member of U

, is order-
continuous.
356Q Theorem Let U be any L-space. Then a subset of U is uniformly integrable i it is relatively
weakly compact.
proof (a) Let A U be a uniformly integrable set.
(i) Suppose that T is an ultralter on X containing A. Then A ,= . Because A is norm-bounded,
sup
uA
[f(u)[ < and (f) = lim
uJ
f(u) is dened in R for every f U

(2A3Se).
If f, g U

then
356X Dual spaces 271
(f +g) = lim
uJ
f(u) +g(u) = lim
uJ
f(u) + lim
uJ
g(u) = (f) +(g)
(2A3Sf). Similarly,
(f) = lim
uJ
f(u) = (f)
for every f U

, R. Thus : U

R is linear. Also
[(f)[ sup
uA
[f(u)[ |f| sup
uA
|u|,
so U

= U

.
(ii) Now the point of this argument is that U

. PPP Suppose that B U

is non-empty and
downwards-directed and has inmum 0. Fix f
0
B. Let > 0. Then there is a w U
+
such that
|([u[ w)
+
| for every u A, which means that
[f(u)[ [f[([u[) [f[(w) +[f[([u[ w)
+
[f[(w) +|f|
for every f U

and every u A. Accordingly [(f)[ [f[(w)+|f| for every f U

. Now inf
fB
f(w) = 0
(using 355Ee, as usual), so there is an f
1
B such that f
1
f
0
and f
1
(w) . In this case
[[(f
1
) = sup
]f]f
1
[(f)[ sup
]f]f
1
[f[(w) +|f| f
1
(w) +|f
1
| (1 +|f
0
|).
As is arbitrary, inf
fB
[[(f) = 0; as B is arbitrary, [[ is order-continuous and U

. QQQ
(iii) At this point, we recall that U

= U

and that the canonical map from U to U

is surjective
(356P). So there is a u
0
U such that u
0
= . But now we see that
f(u
0
) = (f) = lim
uJ
f(u)
for every f U

; which is just what is meant by saying that T u

0
for the weak topology on U (2A3Sd).
Accordingly every ultralter on U containing A has a limit in U. But because the weak topology on U
is regular (3A3Be), it follows that the closure of A for the weak topology is compact (3A3De), so that A is
relatively weakly compact.
(b) For the converse I use the criterion of 354R(b-iv). Suppose that A U is relatively weakly compact.
Then A is norm-bounded, by the Uniform Boundedness Theorem. Now let u
n
)
nN
be any disjoint sequence
in the solid hull of A. For each n, let U
n
be the band of U generated by u
n
. Let P
n
be the band projection
from U onto U
n
(353H). Let v
n
A be such that [u
n
[ [v
n
[; then
[u
n
[ = P
n
[u
n
[ P
n
[v
n
[ = [P
n
v
n
[,
so |u
n
| |P
n
v
n
| for each n. Let g
n
U

be such that |g
n
| = 1 and g
n
(P
n
v
n
) = |P
n
v
n
|.
Dene T : U R
N
by setting Tu = g
n
(P
n
u))
nN
for each u U. Then T is a continuous linear operator
from U to
1
. PPP For m ,= n, U
m
U
n
= 0, because [u
m
[ [u
n
[ = 0. So, for any u U, P
n
u)
nN
is a
disjoint sequence in U, and

n
i=0
|P
i
u| = |

n
i=0
[P
i
u[| = | sup
in
[P
i
u[| |u|
for every n; accordingly
|Tu|
1
=

i=0
[g
i
P
i
u[

i=0
|P
i
u| |u|.
Since T is certainly a linear operator (because every coordinate functional g
i
P
i
is linear), we have the result.
QQQ
Consequently T[A] is relatively weakly compact in
1
, because T is continuous for the weak topologies
(2A5If). But
1
can be identied with L
1
(), where is counting measure on N. So T[A] is uniformly
integrable in
1
, by 247C, and in particular lim
n
sup
wT[A]
[w(n)[ = 0. But this means that
lim
n
|u
n
| lim
n
[g
n
(P
n
v
n
)[ = lim
n
[(Tv
n
)(n)[ = 0.
As u
n
)
nN
is arbitrary, A satises the conditions of 354R(b-iv) and is uniformly integrable.
356X Basic exercises (a) Show that if U =

then U

= U

c
can be identied with
1
, and is
properly included in U

. (Hint: show that if f U

c
then f(u) =

n=0
u(n)f(e
n
), where e
n
(n) = 1,
e
n
(i) = 0 for i ,= n.)
272 Riesz spaces 356Xb
(b) Show that if U = C([0, 1]) then U

= U

c
= 0. (Hint: show that if f (U

c
)
+
and q
n
)
nN
enumerates Q [0, 1], then for each n N there is a u
n
U
+
such that u
n
(q
n
) = 1 and f(u
n
) 2
n
.)
(c) Let X be an uncountable set and the countable-cocountable measure on X, its domain (211R).
Let U be the space of bounded -measurable real-valued functions on X. Show that U is a Dedekind
-complete Banach lattice if given the supremum norm | |

. Show that U

can be identied with

1
(X)
(cf. 356Xa), and that u
_
ud belongs to U

c
U

.
(d) Let U be a Dedekind -complete Riesz space and f U

c
. Let u
n
)
nN
be an order-bounded sequence
in U which is order-convergent to u U in the sense that u = inf
nN
sup
mn
u
m
= sup
nN
inf
mn
u
m
. Show
that lim
n
f(u
n
) exists and is equal to f(u).
(e) Let U be any Riesz space. Show that the band projection P : U

is dened by the formula

(Pf)(u) = infsup
vA
f(v) : A U is non-empty, upwards-directed
and has supremumu
for every f (U

)
+
, u U
+
. (Hint: show that the formula for Pf always denes an order-continuous
linear functional. Compare 355Yh, 356Yb and 362Bd.)
(f ) Let U be any Riesz space. Show that the band projection P : U

c
is dened by the formula
(Pf)(u) = infsup
nN
f(v
n
) : v
n
)
nN
is a non-decreasing sequence with supremum u
for every f (U

)
+
, u U
+
.
(g) Let U be a Riesz space with a Riesz norm. Show that U

is perfect.
(h) Let U be a Riesz space with a Riesz norm. Show that the canonical map from U to U

is a Riesz
homomorphism.
(i) Let V be a perfect Riesz space and U any Riesz space. Show that L

(U; V ) is perfect. (Hint: show

that if u U, g V

then T g(Tu) belongs to L

(U; V )

.)
(j) Let U be an M-space. Show that it is perfect i it is Dedekind complete and U

separates the points

of U.
(k) Let U be a Banach lattice which, as a Riesz space, is perfect. Show that its norm has the Levi
property.
(l) Write out a proof from rst principles that if u
n
)
nN
is a sequence in
1
such that [u
n
(n)[ > 0
for every n N, then u
n
: n N is not relatively weakly compact.
(m) Let U be an L-space and A U a non-empty set. Show that the following are equiveridical: (i) A
is uniformly integrable (ii) inf
fB
sup
uA
[f(u)[ for every non-empty downwards-directed set B U

with
inmum 0 (iii) inf
nN
sup
uA
[f
n
(u)[ = 0 for every non-increasing sequence f
n
)
nN
in U

with inmum 0
(iv) A is norm-bounded and lim
n
sup
uA
[f
n
(u)[ = 0 for every disjoint order-bounded sequence f
n
)
nN
in U

.
356Y Further exercises (a) Let U be a Riesz space with the countable sup property. Show that
U

= U

c
.
(b) Let U be a Riesz space, and / a family of non-empty downwards-directed subsets of U
+
all with
inmum 0. (i) Show that U

,
= f : f U

, inf
uA
[f[(u) = 0 for every A / is a band in U

. (ii) Set
/

= A
0
+ . . . + A
n
: A
0
, . . . , A
n
/. Sbow that U

,
= U

,
. (iii) Take any f (U

)
+
, and let g, h be
the components of f in U

,
, (U

,
)

respectively. Show that

g(u) = inf
A,
sup
vA
f(u v)
+
, h(u) = sup
A,
inf
vA
f(u v)
for every u U
+
. (Cf. 362Xi.)
356 Notes Dual spaces 273
(c) Let U be a Riesz space. For any band V U write V

for f : f U

, f(v) = 0 for every

v V . Show that V (V

is a surjective order-continuous Boolean homomorphism from the algebra of

complemented bands of U onto the band algebra of U

, and that it is injective i U

separates the points

of U.
(d) Let U be a Riesz space such that U

separates the points of U. For any band V U

write V

for
x : x U, f(x) = 0 for every f V . Show that V (V

is a surjective Boolean homomorphism from

the algebra of bands of U

onto the band algebra of U, and that it is injective i U

= U

.
(e) Let U be a Dedekind complete Riesz space such that U

separates the points of U and U is the solid

linear subspace of itself generated by a countable set. Show that U is perfect.
(f ) Let U be an L-space and u
n
)
nN
a sequence in U such that f(u
n
))
nN
is Cauchy for every f U

.
Show that u
n
)
nN
is convergent for the weak topology of U. (Hint: use 356Xm(iv) to show that u
n
: n N
is relatively weakly compact.)
(g) Let U be a perfect Banach lattice with order-continuous norm and u
n
)
nN
a sequence in U such
that f(u
n
))
nN
is Cauchy for every f U

. Show that u
n
)
nN
is convergent for the weak topology of U.
(Hint: set (f) = lim
n
f
n
(u). For any g (U

)
+
let V
g
be the solid linear subspace of U

generated
by g, W
g
= u : g([u[) = 0

, |u|
g
= g([u[) for u W
g
. Show that the completion of W
g
under | |
g
is an
L-space with dual isomorphic to V
g
, and hence (using 356Yf) that V
g
belongs to V

g
; as g is arbitrary,
V

and may be identied with an element of U.)

(h) Let U be a uniformly complete Archimedean Riesz space with complexication V (354Yk). (i) Show
that the complexication of U

can be identied with the space of linear functionals f : V C such that

sup
]v]u
[f(v)[ is nite for every u U
+
. (ii) Show that if U is a Banach lattice, then the complexication
of U

= U

can be identied (as normed space) with V

. (See 355Yk.)
356 Notes and comments The section starts easily enough, with special cases of results in 355 (356B).
When U has a Riesz norm, the identication of U

as a subspace of U

, and the characterization of order-

continuous norms (356D) are pleasingly comprehensive and straightforward. Coming to biduals, we need
to think a little (356F), but there is still no real diculty at rst. In 356H-356I, however, something more
substantial is happening. I have written these arguments out in what seems to be the shortest route to the
main theorem, at the cost perhaps of neglecting any intuitive foundation. What I think we are really doing
is matching bands in U, U

and U

, as in 356Yc.
From now on, almost the rst thing we shall ask of any new Riesz space will be whether it is perfect,
and if not, which of the three conditions of 356K it fails to satisfy. For reasons which will I hope appear
in the next chapter, perfect Riesz spaces are particularly important in measure theory; in particular, all L
p
spaces for p [1, [ are perfect (366D), as are the L

spaces of localizable measure spaces (365K). Further

examples will be discussed in 369 and 374. Of course we have to remember that there are also important
Riesz spaces which are not perfect, of which C([0, 1]) and ccc
0
are two of the simplest examples.
The duality between L- and M-spaces (356N, 356P) is natural and satisfying. We are now in a position
to make a determined attempt to tidy up the notion of uniform integrability. I give two major theorems.
The rst is yet another disjoint-sequence characterization of uniformly integrable sets, to go with 246G
and 354R. The essential dierence here is that we are looking at disjoint sequences in a predual; in a sense,
this means that the result is a sharper one, because the M-space U need not be Dedekind complete (for
instance, it could be C([0, 1]) this indeed is the archetype for applications of the theorem) and therefore
need not have as many disjoint sequences as its dual. (For instance, in the dual of C([0, 1]) we have all
the point masses
t
, where
t
(u) = u(t); these form a disjoint family in C([0, 1])

not corresponding to any

disjoint family in C([0, 1]).) The essence of the proof is a device to extract a disjoint sequence in U to match
approximately a subsequence of a given disjoint sequence in U

. In the example just suggested, this would

correspond, given a sequence t
n
)
nN
of distinct points in [0, 1], to nding a subsequence t
n(i)
)
iN
which is
discrete, so that we can nd disjoint u
i
C([0, 1]) with u
i
(t
n(i)
) = 1 for each i.
The second theorem, 356Q, is a new version of a result already given in 247: in any L-space, uniform
integrability is the same as relative weak compactness. I hope you are not exasperated by having been
274 Riesz spaces 356 Notes
asked, in Volume 2, to master a complex argument (one of the more dicult sections of that volume) which
was going to be superseded. Actually it is worse than that. A theorem of Kakutani (369E) tells us that
every L-space is isomorphic to an L
1
space. So 356Q is itself a consequence of 247C. I do at least owe you
an explanation for writing out two proofs. The rst point is that the result is suciently important for
it to be well worth while spending time in its neighbourhood, and the contrasts and similarities between
the two arguments are instructive. The second is that the proof I have just given was not really accessible
at the level of Volume 2. It does not rely on every single page of this chapter, but the key idea (that U
is isomorphic to U

, so it will be enough if we can show that A is relatively compact in U

) depends
essentially on 356I, which lies pretty deep in the abstract theory of Riesz spaces. The third is an aesthetic
one: a theorem about L-spaces ought to be proved in the category of normed Riesz spaces, without calling
on a large body of theory outside. Of course this is a book on measure theory, so I did the measure theory
rst, but if you look at everything that went into it, the proof in 247 is I believe longer, in the formal sense,
than the one here, even setting aside the labour of proving Kakutanis theorem.
Let us examine the ideas in the two proofs. First, concerning the proof that uniformly integrable sets
are relatively compact, the method here is very smooth and natural; the denition I chose of uniform
integrability is exactly adapted to showing that uniformly integrable sets are relatively compact in the
order-continuous bidual; all the eort goes into the proof that L-spaces are perfect. The previous argument
depended on identifying the dual of L
1
as L

and was disagreeably complicated by the fact that the

identication is not always valid, so that I needed to reduce the problem to the -nite case (part (b-ii) of
the proof of 247C). After that, the Radon-Nikod ym theorem did the trick. Actually Kakutanis theorem
shows that the side-step to -nite spaces is irrelevant. It directly represents an abstract L-space as L
1
()
for a localizable measure , in which case (L
1
)

= L

exactly.
In the other direction, both arguments depend on a disjoint-sequence criterion for uniform integrability
(246G(iii) or 354R(b-iv)). These criteria belong to the easy side of the topic; straightforward Riesz space
arguments do the job, whether written out in that language or not. (Of course the new one in this section,
356O, lies a little deeper.) I go a bit faster this time because I feel that you ought by now to be happy
with the Hahn-Banach theorem and the Uniform Boundedness Theorem, which I was avoiding in Volume
2. And then of course I quote the result for
1
. This looks like cheating. But
1
really is easier, as you
will nd if you just write out part (a) of the proof of 247C for this case. It is not exactly that you can
dispense with any particular element of the argument; rather it is that the formulae become much more
direct when you can write u(i) in place of
_
F
i
u, and cluster points for the weak topology become pointwise
limits of subsequences, so that the key step (the sliding hump, in which u
k(j)
(n(k(j))) is the only signicant
coordinate of u
k(j)
), is easier to nd.
We now have a wide enough variety of conditions equivalent to uniform integrability for it to be easy to
nd others; I give a couple in 356Xm, corresponding in a way to those in 246G. You may have noticed, in
the proof of 247C, that in fact the full strength of the hypothesis relatively weakly compact is never used;
all that is demanded is that a couple of sequences should have cluster points for the weak topology. So we
see that a set A is uniformly integrable i every sequence in A has a weak cluster point. But this extra
renement is nothing to do with L-spaces; it is generally true, in any normed space U, that a set A U is
relatively weakly compact i every sequence in A has a cluster point in U for the weak topology (Eberleins
theorem; see 462D in Volume 4, K othe 69, 24.2.1, or Dunford & Schwartz 57, V.6.1).
There is a very rich theory concerning weak compactness in perfect Riesz spaces, based on the ideas here;
some of it is explored in Fremlin 74a. As a sample, I give one of the basic properties of perfect Banach
lattices with order-continuous norms: they are weakly sequentially complete (356Yg).
Chap. 36 intro. Introduction 275
Chapter 36
Function Spaces
Chapter 24 of Volume 2 was devoted to the elementary theory of the function spaces L
0
, L
1
, L
2
and
L

associated with a given measure space. In this chapter I return to these spaces to show how they can
be related to the more abstract themes of the present volume. In particular, I develop constructions to
demonstrate, as clearly as I can, the way in which all the function spaces associated with a measure space in
fact depend only on its measure algebra; and how many of their features can (in my view) best be understood
in terms of constructions involving measure algebras.
The chapter is very long, not because there are many essentially new ideas, but because the intuitions
I seek to develop depend, for their logical foundations, on technically complex arguments. This is perhaps
best exemplied by 364. If two measure spaces (X, , ) and (Y, T, ) have isomorphic measure algebras
(A, ), (B, ) then the spaces L
0
(), L
0
() are isomorphic as topological f-algebras; and more: for any
isomorphism between (A, ) and (B, ) there is a unique corresponding isomorphism between the L
0
spaces.
The intuition involved is in a way very simple. If f, g are measurable real-valued functions on X and Y
respectively, then f

L
0
() will correspond to g

L
0
() if and only if [[f

> ]] = x : f(x) >

A
corresponds to [[g

> ]] = y : g(y) >

B for every . But the check that this formula is consistent,

and denes an isomorphism of the required kind, involves a good deal of detailed work. It turns out, in
fact, that the measures and do not enter this part of the argument at all, except through their ideals
of negligible sets (used in the construction of A and B). This is already evident, if you look for it, in
the theory of L
0
(); in 241, as written out, you will nd that the measure of an individual set is not
once mentioned, except in the exercises. Consequently there is an invitation to develop the theory with
algebras A which are not necessarily measure algebras. Here is another reason for the length of the chapter;
substantial parts of the work are being done in greater generality than the corresponding sections of Chapter
24, necessitating a degree of repetition. Of course this is not measure theory in the strict sense; but for
thirty years now measure theory has been coloured by the existence of these generalizations, and I think it is
useful to understand which parts of the theory apply only to measure algebras, and which can be extended
to other -complete Boolean algebras, like the algebraic theory of L
0
, or even to all Boolean algebras, like
the theory of L

.
Here, then, are two of the objectives of this chapter: rst, to express the ideas of Chapter 24 in ways
making explicit their independence of particular measure spaces, by setting up constructions based exclu-
sively on the measure algebras involved; second, to set out some natural generalizations to other algebras.
But to justify the eort needed I ought to point to some mathematically signicant idea which demands
these constructions for its expression, and here I mention the categorical nature of the constructions. Be-
tween Boolean algebras we have a variety of natural and important classes of morphism; for instance,
the Boolean homomorphisms and the order-continuous Boolean homomorphisms; while between measure
algebras we have in addition the measure-preserving Boolean homomorphisms. Now it turns out that if
we construct the L
p
spaces in the natural ways then morphisms between the underlying algebras give rise
to morphisms between their L
p
spaces. For instance, any Boolean homomorphism from A to B produces
a multiplicative norm-contractive Riesz homomorphism from L

(A) to L

(B); if A and B are Dedekind

-complete, then any sequentially order-continuous Boolean homomorphism from A to B produces a se-
quentially order-continuous multiplicative Riesz homomorphism from L
0
(A) to L
0
(B); and if (A, ) and
(B, ) are measure algebras, then any measure-preserving Boolean homomorphism from A to B produces
norm-preserving Riesz homomorphisms from L
p
(A, ) to L
p
(B, ) for every p [1, ]. All of these are
functors, that is, a composition of homomorphisms between algebras gives rise to a composition of the
corresponding operators between their function spaces, and are covariant, that is, a homomorphism from
A to B leads to an operator from L
p
(A) to L
p
(B). But the same constructions lead us to a functor which
is contravariant: starting from an order-continuous Boolean homomorphism from a semi-nite measure
algebra (A, ) to a measure algebra (B, ), we have an operator from L
1
(B, ) to L
1
(A, ). This last is in
fact a kind of conditional expectation operator. In my view it is not possible to make sense of the theory of
measure-preserving transformations without at least an intuitive grasp of these ideas.
Another theme is the characterization of each construction in terms of universal mapping theorems: for
instance, each L
p
space, for 1 p , can be characterized as Banach lattice in terms of factorizations of
functions of an appropriate class from the underlying algebra to Banach lattices.
276 Function spaces Chap. 36 intro.
Now let me try to sketch a route-map for the journey ahead. I begin with two sections on the space
S(A); this construction applies to any Boolean algebra (indeed, any Boolean ring), and corresponds to the
space of simple functions on a measure space. Just because it is especially close to the algebra (or ring) A,
there is a particularly large number of universal mapping theorems corresponding to dierent aspects of its
structure (361). In 362 I seek to relate ideas on additive functionals on Boolean algebras from Chapter
23 and 326-327 to the theory of Riesz space duals in 356. I then turn to the systematic discussion of the
function spaces of Chapter 24: L

(363), L
0
(364), L
1
(365) and other L
p
(366), followed by an account
of convergence in measure (367). While all these sections are dominated by the objectives sketched in the
paragraphs above, I do include a few major theorems not covered by the ideas of Volume 2, such as the
Kelley-Nachbin characterization of the Banach spaces L

(A) for Dedekind complete A (363R). In the last

two sections of the chapter I turn to the use of L
0
spaces in the representation of Archimedean Riesz spaces
(368) and of Banach lattices separated by their order-continuous duals (369).
361 S
This is the fundamental Riesz space associated with a Boolean ring A. When A is a ring of sets, S(A) can
be regarded as the linear space of simple functions generated by the characteristic functions of members of
A (361L). Its most important property is the universal mapping theorem 361F, which establishes a one-to-
one correspondence between (nitely) additive functions on A (361B-361C) and linear operators on S(A).
Simple universal mapping theorems of this type can be interesting, but do not by themselves lead to new
insights; what makes this one important is the fact that S(A) has a canonical Riesz space structure, norm
and multiplication (361E). From this we can deduce universal mapping theorems for many other classes of
function (361G, 361H, 361I, 361Xb). (Particularly important are countably additive and completely additive
real-valued functionals, which will be dealt with in the next section.) While the exact construction of S(A)
(and the associated map from A to S(A)) can be varied (361D, 361L, 361M, 361Ya), its structure is uniquely
dened, so homomorphisms between Boolean rings correspond to maps between their S( )-spaces (361J),
and (when A is an algebra) A can be recovered from the Riesz space S(A) as the algebra of its projection
bands (361K).
361A Boolean rings In this section I speak of Boolean rings rather than algebras; there are ideas
in 365 below which are more naturally expressed in terms of the ring of elements of nite measure in a
measure algebra than in terms of the whole algebra. I should perhaps therefore recall some of the ideas of
311, which is the last time when Boolean rings without identity were mentioned, and set out some simple
facts.
(a) Any Boolean ring A can be represented as the ring of compact open subsets of a zero-dimensional
locally compact Hausdor space X (311I); X is just the set of surjective ring homomorphisms from A onto
Z
2
(311E).
(b) If A and B are Boolean rings and : A B is a function, then the following are equiveridical:
(i) is a ring homomorphism; (ii) (a \ b) = a \ b for all a, b A; (iii) 0 = 0 and (a b) = a b,
(a b) = a b for all a, b A. PPP See 312H. To prove (ii)(iii), observe that if a, b A then
(a b) = a \ (a \ b) = a b. QQQ
(c) If A and B are Boolean rings and : A B is a ring homomorphism, then is order-continuous
i inf [A] = 0 whenever A A is non-empty and downwards-directed and inf A = 0 in A; while is
sequentially order-continuous i inf
nN
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence in A with
inmum 0. (See 313L.)
(d) The following will be a particularly important type of Boolean ring for us. If (A, ) is a measure
algebra, then the ideal A
f
= a : a A, a < is a Boolean ring in its own right. Now suppose that
(B, ) is another measure algebra and B
f
B the corresponding ring of elements of nite measure. We
361D S 277
can say that a ring homomorphism : A
f
B
f
is measure-preserving if a = a for every a A
f
.
Now in this case is order-continuous. PPP If A A
f
is non-empty, downwards-directed and has inmum 0,
then inf
aA
a = 0, by 321F; but this means that inf
aA
a = 0, and inf [A] = 0 in B
f
. QQQ
361B Denition Let A be a Boolean ring and U a linear space. A function : A U is nitely
additive, or just additive, if (a b) = a +b whenever a, b A and a b = 0.
361C Elementary facts We have the following immediate consequences of this denition, correspond-
ing to 326B and 313L. Let A be a Boolean ring, U a linear space and : A U an additive function.
(a) 0 = 0 (because 0 = 0 +0).
(b) If a
0
, . . . , a
m
are disjoint in A, then (sup
jm
a
j
) =

m
j=0
a
j
.
(c) If B is another Boolean ring and : B A is a ring homomorphism, then : B U is additive.
In particular, if B is a subring of A, then B : B U is additive.
(d) If V is another linear space and T : U V is a linear operator, then T : A V is additive.
(e) If U is a partially ordered linear space, then is order-preserving i it is non-negative, that is, a 0
for every a A. PPP () If is order-preserving, then of course 0 = 0 a for every a A. () If is
non-negative, and a b in A, then
a a +(b \ a) = b. QQQ
(f ) If U is a partially ordered linear space and is non-negative, then (i) is order-continuous i
inf [A] = 0 whenever A A is a non-empty downwards-directed set with inmum 0 (ii) is sequentially
order-continuous i inf
nN
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence in A with inmum 0. PPP
(i) If is order-continuous, then of course inf [A] = 0 = 0 whenever A A is a non-empty downwards-
directed set with inmum 0. If satises the condition, and A A is a non-empty upwards-directed set
with supremum c, then c \ a : a A is downwards-directed with inmum 0 (313Aa), so that
sup
aA
a = sup
aA
c (c \ a) = c inf
aA
(c \ a)
(by 351Db)
= c.
Similarly, if A A is a non-empty downwards-directed set with inmum c, then
inf
aA
a = inf
aA
c +(a \ c) = c + inf
aA
(a \ c) = c.
Putting these together, is order-continuous. (ii) If is sequentially order-continuous, then of course
inf
nN
a
n
= 0 = 0 whenever a
n
)
nN
is a non-increasing sequence in A with inmum 0. If satises the
condition, and a
n
)
nN
is a non-decreasing sequence in A with supremum c, then c \ a
n
)
nN
is non-increasing
and has inmum 0, so that
sup
nN
a
n
= sup
nN
c (c \ a
n
) = c inf
nN
(c \ a
n
) = c.
Similarly, if a
n
)
nN
is a non-increasing sequence in A with inmum c, then a
n
\ c)
nN
is non-increasing
and has inmum 0, so that
inf
nN
a
n
= inf
nN
c +(c \ a
n
) = c + inf
nN
(c \ a
n
) = c.
Thus is sequentially order-continuous. QQQ
361D Construction Let A be a Boolean ring, and Z its Stone space. For a A write a for the
characteristic function of the open-and-compact subset a of Z corresponding to a. Let S(A) be the linear
subspace of R
Z
generated by a : a A. Because a is a bounded function for every a, S(A) is a subspace
of the space

(Z) of all bounded real-valued functions on Z (354Ha), and | |

is a norm on S(A). Because

a b = (a b) for all a, b A (writing for pointwise multiplication of functions, as in 281B), S(A) is
closed under .
278 Function spaces 361E
361E I give a portmanteau proposition running through the elementary, mostly algebraic, properties
of S(A).
Proposition Let A be a Boolean ring, with Stone space Z. Write S for S(A).
(a) If a
0
, . . . , a
n
A, there are disjoint b
0
, . . . , b
m
such that each a
i
is expressible as the supremum of
some of the b
j
.
(b) If u S, it is expressible in the form

m
j=0

j
b
j
where b
0
, . . . , b
m
are disjoint members of A and

j
R for each j. If all the b
j
are non-zero then |u|

= sup
jm
[
j
[.
(c) If u S is non-negative, it is expressible in the form

m
j=0

j
b
j
where b
0
, . . . , b
m
are disjoint members
of A and
j
0 for each j, and simultaneously in the form

m
j=0

j
c
j
where c
0
c
1
. . . c
m
and
j
0
for every j.
(d) If u =

m
j=0

j
b
j
where b
0
, . . . , b
m
are disjoint members of A and
j
R for each j, then [u[ =

m
j=0
[
j
[b
j
S.
(e) S is a Riesz subspace of R
Z
; in its own right, it is an Archimedean Riesz space. If A is a Boolean
algebra, then S has an order unit 1 and |u|

= min : 0, [u[ 1 for every u S.

(f) The map : A S is injective, additive, non-negative, a lattice homomorphism and order-continuous.
(g) Suppose that u 0 in S and 0 in R. Then
[[u > ]] = supa : a A, ( +)a u for some > 0
is dened in A, and
[[u > ]] u [[u > 0]] |u|

[[u > ]].

In particular, u |u|

[[u > 0]] and there is an > 0 such that [[u > 0]] u. If u, v 0 in S then
u v = 0 i [[u > 0]] [[v > 0]] = 0.
(h) Under , S is an f-algebra (352W) and a commutative normed algebra (2A4J).
(i) For any u S, u 0 i u = v v for some v S.
proof Write a for the open-and-compact subset of Z corresponding to a A.
(a) Induce on n. If n = 0 take m = 0, b
0
= a
0
. For the inductive step to n 1, take disjoint
b
0
, . . . , b
m
such that a
i
is the supremum of some of the b
j
for each i < n; now replace b
0
, . . . , b
m
with
b
0
a
n
, . . . , b
m
a
n
, b
0
\ a
n
, . . . , b
m
\ a
n
, a
n
\ sup
jm
b
j
to obtain a suitable string for a
0
, . . . , a
n
.
(b) If u = 0 set m = 0, b
0
= 0,
0
= 0. Otherwise, express u as

n
i=0

i
a
i
where a
i
A and
0
, . . . ,
n
are real numbers. Let b
0
, . . . , b
m
be disjoint and such that every a
i
is expressible as the supremum of some
of the b
j
. Set
ij
= 1 if b
j
a
i
, 0 otherwise, so that, because the b
j
are disjoint, a
i
=

k
j=0

ij
b
j
for
each i. Then
u =

n
i=0

i
a
i
=

n
i=0

m
j=0

ij
b
j
=

m
j=0

j
b
j
,
setting
j
=

n
i=0

ij
for each j k.
The expression for |u|

is now obvious.
(c)(i) If u 0 in (b), we must have
j
= u(z) 0 whenever z

b
j
, so that
j
0 whenever b
j
,= 0;
consequently u =

m
j=0
[
j
[b
j
is in the required form.
(ii) If we suppose that every
j
is non-negative, and rearrange the terms of the sum so that
0
. . .

m
, then we may set
0
=
0
,
j
=
j

j1
for 1 j m, c
j
= sup
ij
b
i
to get

m
j=0

j
c
j
=

m
j=0

m
i=j

j
b
i
=

m
i=0

i
j=0

j
b
i
=

m
i=0

i
b
i
= u.
(d) is trivial, because

b
0
, . . . ,

b
n
are disjoint.
(e) By (d), [u[ S for every u S, so S is a Riesz subspace of R
Z
, and in itself is an Archimedean Riesz
space. If A is a Boolean algebra, then 1, the constant function with value 1, belongs to S, and is an order
unit of S; while
|u|

= min : 0, [u(z)[ z Z = min : 0, [u[ 1

for every u S.
361F S 279
(f ) is injective because a ,=

b whenever a ,= b. is additive because a

b = whenever a b = 0. Of
course it is non-negative. It is a lattice homomorphism because a a : A TZ and E E : TZ R
Z
are. To see that is order-continuous, take a non-empty downwards-directed A A with inmum 0. ???
Suppose, if possible, that a : a A does not have inmum 0 in S. Then there is a u > 0 in S such that
u a for every a A. Now u can be expressed as

m
j=0

j
b
j
where b
0
, . . . , b
m
are disjoint. There must
be some z
0
Z such that u(z
0
) > 0; take j such that z
0

b
j
, so that b
j
,= 0 and
j
= u(z
0
) > 0. But now,
for any z

b
j
, a A,
(a)(z) u(z) =
j
> 0
and z a. As z is arbitrary,

b
j
a and b
j
a; as a is arbitrary, b
j
is a non-zero lower bound for A in A.
XXX So inf [A] = 0 in S. As A is arbitrary, is order-continuous, by the criterion of 361C(f-i).
(g) Express u as

m
j=0

j
b
j
where b
0
, . . . , b
m
are disjoint and every
j
0. Then given 0, > 0,
a A we have ( + )a u i a supb
j
: j m,
j
+ . So [[u > ]] = supb
j
: j m,
j
> .
Writing c = [[u > ]], d = [[u > 0]] = supb
j
:
j
> 0, we have
u(z) |u|

if z c,
if z

d c,
= 0 if z /

d.
So
c u |u|

c d,
as claimed. Taking = 0 we get u |u|

d. Set
= min(1
j
: j m,
j
> 0);
then > 0 and d u.
If u, v S
+
take ,
t
> 0 such that
[[u > 0]] u,
t
[[v > 0]] v.
Then
min(,
t
)([[u > 0]] [[v > 0]]) u v max(|u|

, |v|

)([[u > 0]] [[v > 0]]).

So
u v = 0 = [[u > 0]] [[v > 0]] = 0 = u v = 0.
(h) S is a commutative f-algebra and normed algebra just because it is a Riesz subspace of the f-algebra
and commutative normed algebra

(Z) and is closed under multiplication.

(i) If u =

m
j=0

j
b
j
where b
0
, . . . , b
m
are disjoint and
j
0 for every j, then u = v v where
v =

m
j=0
_

j
b
j
.
361F I now turn to the universal mapping theorems which really dene the construction.
Theorem Let A be a Boolean ring, and U any linear space. Then there is a one-to-one correspondence
between additive functions : A U and linear operators T : S(A) U, given by the formula = T.
proof (a) The core of the proof is the following observation. Let : A U be additive. If a
0
, . . . , a
n
A
and
0
, . . . ,
n
R are such that

n
i=0

i
a
i
= 0 in S = S(A), then

n
i=0

i
a
i
= 0 in U. PPP As in 361E,
we can nd disjoint b
0
, . . . , b
m
such that each a
i
is the supremum of some of the b
j
; set
ij
= 1 if b
j
a
i
, 0
otherwise, so that a
i
=

m
j=0

ij
b
j
, a
i
=

m
j=0

ij
b
j
for each i. Set
j
=

n
i=0

ij
, so that
0 =

n
i=0

i
a
i
=

m
j=0

j
b
j
.
Now
j
b
j
= 0 in U for each j, because either b
j
= 0 and b
j
= 0, or there is some z

b
j
so that
j
must
be 0. Accordingly
280 Function spaces 361F
0 =

m
j=0

j
b
j
=

m
j=0

n
i=0

ij
b
i
=

n
i=0

i
a
i
. QQQ
(b) It follows that if u S is expressible simultaneously as

n
i=0

i
a
i
=

m
j=0

j
b
j
, then

n
i=0

i
a
i
+

m
j=0
(
j
)b
j
= 0 in S,
so that

n
i=0

i
a
i
+

m
j=0
(
j
)b
j
= 0 in U,
and

n
i=0

i
a
i
=

m
j=0

j
b
j
.
We can therefore dene T : S U by setting
T(

n
i=0

i
a
i
) =

n
i=0

i
a
i
whenever a
0
, . . . , a
n
A and
0
, . . . ,
n
R.
(c) It is now elementary to check that T is linear, and that Ta = a for every a A. Of course this
last condition uniquely denes T, because a : a A spans the linear space S.
361G Theorem Let A be a Boolean ring, and U a partially ordered linear space. Let : A U be an
additive function, and T : S(A) U the corresponding linear operator.
(a) is non-negative i T is positive.
(b) In this case,
(i) if T is order-continuous or sequentially order-continuous, so is ;
(ii) if U is Archimedean and is order-continuous or sequentially order-continuous, so is T.
(c) If U is a Riesz space, then the following are equiveridical:
(i) T is a Riesz homomorphism;
(ii) a b = 0 in U whenever a b = 0 in A;
(iii) is a lattice homomorphism.
proof Write S for S(A).
(a) If T is positive, then surely a = Ta 0 for every a A, so = T is non-negative. If is
non-negative, and u 0 in S, then u is expressible as

m
j=0

j
b
j
where b
0
, . . . , b
m
A and
j
0 for
every j (361Ec), so that
Tu =

m
j=0

j
b
j
0.
Thus T is positive.
(b)(i) If T is order-continuous (resp. sequentially order-continuous) then = T is the composition of
two order-continuous (resp. sequentially order-continuous) functions (361Ef), so must be order-continuous
(resp. sequentially order-continuous).
(ii) Assume now that U is Archimedean.
() Suppose that is order-continuous and that A S is non-empty, downwards-directed and has
inmum 0. Fix u
0
A, set = |u|

and a
0
= [[u > 0]] (in the language of 361Eg). If = 0 then of course
inf
uA
Tu = Tu
0
= 0. Otherwise, take any w U such that w , 0. Then there is some > 0 such that
w , a
0
, because U is Archimedean. Set A
t
= u : u A, u u
0
; because A is downwards-directed,
A
t
has the same lower bounds as A, and inf A
t
= 0, while A
t
is still downwards-directed. For u A
t
set
c
u
= [[u > ]], so that
c
u
u c
u
+[[u > 0]] c
u
+a
0
(361Eg). If u, v A
t
and u v, then c
u
c
v
, so C = c
u
: u A
t
is downwards-directed; but if c is any
lower bound for C in A, c is a lower bound for A
t
in S, so is zero, and c = 0 in A. Thus inf C = 0 in A,
and inf
uA
c
u
= 0 in U. But this means, in particular, that
1

(w a
0
) is not a lower bound for [C],
and there is some u A
t
such that
1

(wa
0
) , c
u
, that is, wa
0
, c
u
, that is, w , a
0
+c
u
.
As u c
u
+a
0
,
361H S 281
Tu T(c
u
+a
0
) = c
u
+a
0
,
and w , Tu. Since w is arbitrary, this means that 0 = inf T[A]; as A is arbitrary, T is order-continuous.
() The argument for sequential order-continuity is essentially the same. Suppose that is se-
quentially order-continuous and that u
n
)
nN
is a non-increasing sequence in S with inmum 0. Again set
= |u
0
|, a
0
= [[u
0
> 0]]; again we may suppose that > 0; again take any w U such that w , 0. As
before, there is some > 0 such that w , a
0
. For n N set c
n
= [[u
n
> ]], so that
c
n
u
n
c
n
+a
0
.
The sequence c
n
)
nN
is non-increasing because u
n
)
nN
is, and if c c
n
for every n, then c u
n
for
every n, so is zero, and c = 0 in A. Thus inf
nN
c
n
= 0 in A, and inf
nN
c
n
= 0 in U, because is
sequentially order-continuous. Replacing A
t
, C in the argument above by u
n
: n N, c
n
: n N we
nd an n such that w , Tu
n
. Since w is arbitrary, this means that 0 = inf
nN
Tu
n
; as u
n
)
nN
is arbitrary,
T is sequentially order-continuous.
(c)(i)(iii) If T : S(A) U is a Riesz homomorphism, and = T, then surely is a lattice homo-
morphism because T and are.
(iii)(ii) is trivial.
(ii)(i) If a b = 0 whenever a b = 0, then for any u S(A) we have an expression of u as

m
j=0

j
b
j
, where b
0
, . . . , b
m
A are disjoint. Now
[Tu[ = [

m
j=0

j
b
j
[ =

m
j=0
[
j
[b
j
= T(

m
j=0
[
j
[b
j
) = T([u[)
by 352Fb and 361Ed. As u is arbitrary, T is a Riesz homomorphism (352G).
361H Theorem Let A be a Boolean ring and U a Dedekind complete Riesz space. Suppose that
: A U is an additive function and T : S = S(A) U the corresponding linear operator. Then
T L

= L

(S; U) i b : b a is order-bounded in U for every a A, and in this case [T[ L

corresponds to [[ : A U, dened by setting

[[(a) = sup
n

j=0
[a
i
[ : a
0
, . . . , a
n
a are disjoint
= supb (a \ b) : b a
for every a A.
proof (a) Suppose that T L

and a A. Then for any b a, we have b a so

[b[ = [Tb[ [T[(a).
Accordingly b : b a is order-bounded in U.
(b) Now suppose that b : b a is order-bounded in U for every a A. Then for any a A we can
dene w
a
= sup[b[ : b a; in this case, b (a \ b) 2w
a
whenever b a, so a = sup
ba
b (a \ b)
is also dened in U. Considering b = a, b = 0 we see that a [a[. Next, : A U is additive. PPP Take
a
1
, a
2
A such that a
1
a
2
= 0; set a
0
= a
1
a
2
. For each j 2 set
A
j
= (a
j
b) (a
j
\ b) : b A U.
Then A
0
A
1
+A
2
, because
(a
0
b) (a
0
\ b) = (a
1
b) (a
1
\ b) +(a
2
b) (a
2
\ b)
for every b A. But also A
1
+A
2
A
0
, because if b
1
, b
2
A then
(a
1
b
1
) (a
1
\ b
1
) +(a
2
b
2
) (a
2
\ b
2
) = (a
0
b) (a
0
\ b)
where b = (a
1
b
1
) (a
2
b
2
). So A
0
= A
1
+A
2
, and
a
0
= sup A
0
= sup A
1
+ sup A
2
= a
1
+a
2
(351Dc). QQQ
282 Function spaces 361H
We therefore have a corresponding positive operator T
1
: S U such that = T
1
. But we also see that
a = sup

n
i=0
[a
i
[ : a
0
, . . . , a
n
a are disjoint for every a A. PPP If a
0
, . . . , a
n
are disjoint and included
in a, then

n
i=0
[a
i
[

n
i=0
a
i
= (sup
in
a
i
) a.
On the other hand,
a sup
ba
[b[ +[(a \ b)[ sup

n
i=0
[a
i
[ : a
0
, . . . , a
n
a are disjoint. QQQ
It follows that T L

. PPP Take any u 0 in S. Set a = [[u > 0]] (361Eg) and = |u|

. If 0 < [v[ u,
then v is expressible as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint and no
i
nor a
i
is zero. Since [v[ a,
we must have [
i
[ , a
i
a for each i. So
[Tv[ = [

n
i=0

i
a
i
[

n
i=0
[
i
[[a
i
[

n
i=0
[a
i
[ = a.
Thus [Tv[ : [v[ u is bounded above by a. As u is arbitrary, T L

. QQQ
(c) Thus T L

i is order-bounded on the sets b : b a, and in this case the two formulae oered
for [[ are consistent and make [[ = . Finally, = [T[. PPP Take a A. If a
0
, . . . , a
n
a are disjoint,
then

n
i=0
[a
i
[ =

n
i=0
[Ta
i
[

n
i=0
[T[(a
i
) [T[(a);
so a [T[(a). On the other hand, the argument at the end of (b) above shows that [T[(a) a for
every a. Thus [T[(a) = a for every a A, as required. QQQ
361I Theorem Let A be a Boolean ring, U a normed space and : A U an additive function. Give
S = S(A) its norm | |

, and let T : S U be the linear operator corresponding to . Then T is a bounded

linear operator i a : a A is bounded, and in this case |T| = sup
a,bA
|a b|.
proof (a) If T is bounded, then
|a b| = |T(a b)| |T||a b|

|T|
for every a A, so is bounded and sup
a,bA
|a b| |T|.
(b)(i) For the converse, we need a renement of an idea in 361Ec. If u S and u 0 and |u|

1,
then u is expressible as

m
i=0

i
c
i
where
i
0 and

m
i=0

i
= 1. PPP If u = 0, take n = 0, c
0
= 0,
0
= 1.
Otherwise, start from an expression u =

n
j=0

j
c
j
where c
0
. . . c
n
and every
j
is non-negative, as in
361Ec. We may suppose that c
n
,= 0, in which case

n
j=0

j
= u(z) 1
for every z c
n
Z, the Stone space of A. Set m = n+1, c
m
= 0 and
m
= 1

n
j=0

j
to get the required
form. QQQ
(ii) The next fact we need is an elementary property of real numbers: if
0
, . . . ,
m
,
t
0
, . . . ,
t
n
0
and

m
i=0

i
=

n
j=0

t
j
, then there are
ij
0 such that
i
=

m
j=0

ij
for every i m and
t
j
=

n
i=0

ij
for every j n. PPP This is just the case U = R of 352Fd. QQQ
(iii) Now suppose that is bounded; set
0
= sup
aA
|a| < . Then
= sup
a,bA
|a b| 2
0
is also nite. If u S and |u|

1, then we can express u as u

+
u

where u
+
, u

are non-negative and

also of norm at most 1. By (i), we can express these as
u
+
=

m
i=0

i
c
i
, u

n
j=0

t
j
c
t
j
where all the
i
,
t
j
are non-negative and

m
i=0

i
=

n
j=0

t
j
= 1. Take
ij
)
im,jn
from (ii). Set c
ij
= c
i
,
c
t
ij
= c
t
j
for all i, j, so that
u
+
=

m
i=0

n
j=0

ij
c
ij
, u

m
i=0

n
j=0

ij
c
t
ij
,
u =

m
i=0

n
j=0

ij
(c
ij
c
t
ij
),
361J S 283
Tu =

m
i=0

n
j=0

ij
(c
ij
c
t
ij
),
|Tu|

m
i=0

n
j=0

ij
|c
ij
c
t
ij
|

m
i=0

n
j=0

ij
= .
As u is arbitrary, T is a bounded linear operator and |T| , as required.
361J The last few paragraphs describe the properties of S(A) in terms of universal mapping theorems.
The next theorem looks at the construction as a functor which converts Boolean algebras into Riesz spaces
and ring homomorphisms into Riesz homomorphisms.
Theorem Let A and B be Boolean rings and : A B a ring homomorphism.
(a) We have a Riesz homomorphism T

: S(A) S(B) given by the formula

T

(a) = (a) for every a A.

For any u S(A), |T

u|

= min|u
t
|

: u
t
S(A), T

u
t
= T

u; in particular, |T

u|

|u|

.
Moreover, T

(u u
t
) = T

u T

u
t
for all u, u
t
S(A).
(b) T

is surjective i is surjective, and in this case |v|

= min|u|

: u S(A), T

u = v for every
v S(B).
(c) The kernel of T

is just the set of those u S(A) such that [[[u[ > 0]] = 0, dening [[. . . > . . . ]] as in
361Eg.
(d) T

is injective i is injective, and in this case |T

u|

= |u|

for every u S(A).

(e) T

is order-continuous i is order-continuous.
(f) T

is sequentially order-continuous i is sequentially order-continuous.

(g) If C is another Boolean ring and : B C is another ring homomorphism, then T

= T

:
S(A) S(C).
proof (a) The map : A S(B) is additive (361Cc), so corresponds to a linear operator T = T

: S(A)
S(B), by 361F. and are both lattice homomorphisms, so also is, and T is a Riesz homomorphism
(361Gc). If u =

n
i=0

i
a
i
, where a
0
, . . . , a
n
are disjoint, then look at I = i : i n, a
i
,= 0. We have
Tu =

n
i=0

i
(a
i
) =

iI

i
(a
i
)
and a
0
, . . . , a
n
are disjoint, so that
|Tu|

= sup
iI
[
i
[ = |u
t
|

sup
a
i
,=0
[
i
[ |u|

,
where u
t
=

iI

i
a
i
, so that Tu
t
= Tu. If a, a
t
A, then
T(a a
t
) = T(a a
t
) = (a a
t
) = a a
t
= Ta Ta
t
,
so T is multiplicative.
(b) If is surjective, then T[S(A)] must be the linear span of
T(a) : a A = (a) : a A = b : b B,
so is the whole of S(B). If T is surjective, and b B, then there must be a u A such that Tu = b. We
can express u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint; now
b = Tu =

n
i=0

i
(a
i
),
and a
0
, . . . , a
n
are disjoint in B, so we must have
b = sup
iI
a
i
= (sup
iI
a
i
) [A],
where I = i :
i
= 1. As b is arbitrary, is surjective. Of course the formula for |v|

is a consequence
of the formula for |Tu|

in (a).
(c)(i) If [[[u[ > 0]] = 0 then [u[ a, where = |u|

, a = [[[u[ > 0]], so

[Tu[ = T[u[ T(a) = (a) = 0,
and Tu = 0. (ii) If u S(A) and Tu = 0, express u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint and every

i
is non-zero (361Eb). In this case
0 = [Tu[ = T[u[ =

n
i=0
[
i
[(a
i
),
284 Function spaces 361J
so a
i
= 0 for every i, and
[[[u[ > 0]] = (sup
in
a
i
) = sup
in
a
i
= 0.
(d) If T is injective and a A 0, then (a) = T(a) ,= 0, so a ,= 0; as a is arbitrary, is injective.
If is injective then [[[u[ > 0]] ,= 0 for every non-zero u S(A), so T is injective, by (c). Now the formula
in (a) shows that T is norm-preserving.
(e)(i) If T is order-continuous and A A is a non-empty downwards-directed set with inmum 0 in A,
let b be any lower bound for [A] in B. Then
b (a) = T(a)
for any a A. But T is order-continuous, by 361Ef, so inf
aA
T(a) = 0, and b must be 0. As b is arbitrary,
inf
aA
a = 0; as A is arbitrary, is order-continuous. (ii) If is order-continuous, so is : A S(B),
using 361Ef again; but now by 361G(b-ii) T must be order-continuous.
(f )(i) If T is sequentially order-continuous, and a
n
)
nN
is a non-increasing sequence in A with inmum
0, let b be any lower bound for a
n
: n N in B. Then
b (a
n
) = T(a
n
)
for any a A. But T is sequentially order-continuous so inf
nN
T(a
n
) = 0, and b must be 0. As b is
arbitrary, inf
nN
a
n
= 0; as A is arbitrary, is sequentially order-continuous. (ii) If is sequentially
order-continuous, so is : A S(B); but now T must be sequentially order-continuous.
(g) We need only check that
T

(a) = ((a)) = T

((a)) = T

T(a)
for every a A.
361K Proposition Let A be a Boolean algebra. For a A write V
a
for the solid linear subspace of
S(A) generated by a. Then a V
a
is a Boolean isomorphism between A and the algebra of projection
bands in S(A).
proof Write S for S(A).
(a) The point is that, for any a A,
(i) [u[ [v[ = 0 whenever u V
a
, v V
1\a
,
(ii) V
a
+V
1\a
= S.
PPP (i) is just because a (1 \ a) = 0. As for (ii), if w S then
w = (w a) + (w (1 \ a)) V
a
+V
1\a
. QQQ
(b) Accordingly V
a
+ V

a
V
a
+ V
1\a
= U and V
a
is a projection band (352R). Next, any projection
band U S is of the form V
a
. PPP We know that 1 = u + v where u U, v U

. Since [u[ [v[ = 0,

u and v must be the characteristic functions of complementary subsets of Z, the Stone space of A. But
z : u(z) ,= 0 = z : u(z) 1 must be of the form a, where a = [[u > 0]], in which case u = a and
v = (1 \ a). Accordingly U V
a
and U

V
1\a
. But this means that U must be V
a
precisely. QQQ
(c) Thus a V
a
is a surjective function from A onto the algebra of projection bands in S. Now
a b a V
b
V
a
V
b
,
so a V
a
is order-preserving and bijective. By 312L it is a Boolean isomorphism.
361L Proposition Let X be a set, and a ring of subsets of X, that is, a subring of the Boolean ring
TX. Then S() can be identied, as ordered linear space, with the linear subspace of

(X) generated
by the characteristic functions of members of , which is a Riesz subspace of

(X). The norm of S()

corresponds to the uniform norm on

(X), and its multiplication to pointwise multiplication of functions.

proof Let Z be the Stone space of , and for E write E for the characteristic function of E as a
subset of X, E for the characteristic function of the open-and-compact subset of Z corresponding to E.
361Xb S 285
Of course :

(X) is additive, so by 361F there is a linear operator T : S

(X), writing S for

S(), such that T( E) = E for every E .
If u S, Tu 0 i u 0. PPP Express u as

m
j=0

j
E
j
where E
0
, . . . , E
m
are disjoint. Then
Tu =

m
j=0

j
E
j
, so
u 0
j
0 whenever E
j
,= Tu 0. QQQ
But this means () that
Tu = 0 Tu 0 & T(u) 0 u 0 & u 0 u = 0,
so that T is injective and is a linear space isomorphism between S and its image S, which must be the linear
space spanned by E : E () that T is an order-isomorphism between S and S.
Because E F = 0 whenever E, F and E F = , T is a Riesz homomorphism and S is a Riesz
subspace of

(X) (361Gc). Now

|u|

= inf : [u[ X = inf : [Tu[ X = |Tu|

for every u S. Finally,

T( E F) = T( (E F)) = (E F) = T( E) T( F)
for all E, F , so S is closed under pointwise multiplication and the multiplications of S, S are identied
under T.
361M Proposition Let X be a set, a ring of subsets of X, and 1 an ideal of ; write A for the
quotient ring /1. Let S be the linear span of E : E in R
X
, and write
V = f : f S, x : f(x) ,= 0 1.
Then V is a solid linear subspace of S. Now S(A) becomes identied with the quotient Riesz space S/V , if
for every E we identify (E

) S(A) with (E)

S/V . If we give S its uniform norm inherited from

(X), V is a closed linear subspace of S, and the quotient norm on S/V corresponds to the norm of S(A):
|f

| = min : x : [f(x)[ > 1.

If we write for pointwise multiplication on S, then V is an ideal of the ring (S, +, ), and the multiplication
induced on S/V corresponds to the multiplication of S(A).
proof Use 361J and 361L. We can identify S with S(). Now the canonical ring homomorphism E E

corresponds to a surjective Riesz homomorphism T from S() to S(A) which takes E to (E

). For f S,
[[[f[ > 0]] is just x : f(x) ,= 0, so the kernel of T is just the set of those f S such that x : f(x) ,= 0 1,
which is V . So
S(A) = T[S]

= S/V .
As noted in 361Ja, T(f g) = Tf Tg for all f, g S, so the multiplications of S/V and S(A) match.
As for the norms, the norm of S(A) corresponds to the norm of S/V by the formulae in 361Ja or 361Jb.
(To see that V is closed in S, we need note only that if f V then
|Tf|

= inf
gV
|f +g|

= inf
gV
|f g|

= 0,
so that Tf = 0 and f V .) To check the formula for |f

|, take any f S. Express it as

n
i=0

i
E
i
where
E
0
, . . . , E
n
are disjoint. Set I = i : E
i
/ 1; then
|Tf|

= max
iI
[
i
[ = min : x : [f(x)[ > 1.
361X Basic exercises (a) Let A be a Boolean ring and U a linear space. Show that a function
: A U is additive i 0 = 0 and (a b) +(a b) = a +b for all a, b A.
>>>(b) Let U be an algebra over R, that is, a real linear space endowed with a multiplication such that
(U, +, ) is a ring and (w z) = (w) z = w (z) for all w, z U and all R. Let A be a Boolean
ring, : A U an additive function and T : S(A) U the corresponding linear operator. Show that T is
multiplicative i (a b) = a b for all a, b A.
286 Function spaces 361Xc
>>>(c) Let A be a Boolean ring, and U a Dedekind complete Riesz space. Suppose that : A U is an
additive function such that the corresponding linear operator T : S(A) U belongs to L

= L

(S(A); U).
Show that T
+
L

corresponds to
+
: A U, where
+
a = sup
ba
b for every a A.
(d) Let A and B be Boolean algebras. Show that there is a natural one-to-one correspondence between
Boolean homomorphisms : A B and Riesz homomorphisms T : S(A) S(B) such that T(1
A
) = 1
B
,
given by setting T(a) = (a) for every a A.
(e) Let A, B be Boolean rings and T : S(A) S(B) a linear operator such that T(u v) = Tu Tv for
all u, v S(A). Show that there is a ring homomorphism : A B such that T(a) = (a) for every
a A.
(f ) Let A and B be Boolean rings. Show that any isomorphism of the algebras S(A) and S(B) (using
the word algebra in the sense of 361Xb) must be a Riesz space isomorphism, and therefore corresponds to
an isomorphism between A and B.
(g) Let A, B be Boolean algebras and T : S(A) S(B) a Riesz homomorphism. Show that there are a
ring homomorphism : A B and a non-negative v S(B) such that T(a) = v (a) for every a A.
(h) Let A be a Boolean ring. Show that for any u S(A) the solid linear subspace of S(A) generated by
u is a projection band in S(A). Show that the set of such bands is an ideal in the algebra of all projection
bands, and is isomorphic to A.
>>>(i) Let X be a set, a -algebra of subsets of X. Show that the linear span S in R
X
of E : E
is just the set of -measurable functions f : X R which take only nitely many values.
(j) For any Boolean ring A, we may dene its complex S-space S
C
(A) as the linear span in C
X
of the
characteristic functions of open-and-compact subsets of the Stone space Z of A. State and prove results
corresponding to 361Ea-361Ed, 361Eh, 361F, 361L and 361M.
361Y Further exercises (a) Let A be a Boolean ring. Let V be the linear space of all formal sums of
the form

n
i=0

i
a
i
where
0
, . . . ,
n
R and a
0
, . . . , a
n
A. Let W V be the linear subspace spanned
by members of V of the form (a b) a b where a, b A are disjoint. Dene
t
: A V/W by taking
t
a
to be the image in V/W of a V . Show, without using the axiom of choice, that the pair (V/W,
t
) has the
universal mapping property of (S(A), ) as described in 361F and that V/W has a Riesz space structure, a
norm and a multiplicative structure as described in 361D-361E. Prove results corresponding to 361E-361M.
(b) Let A be a Boolean ring and U a Dedekind complete Riesz space. Let A L

= L

(S(A); U) be a
non-empty set. Suppose that

T = sup A is dened in L

, and that =

T. Show that for any a A,
a = sup

n
i=0
T
i
(a
i
) : T
0
, . . . , T
n
A, a
0
, . . . , a
n
a are disjoint, sup
in
a
i
= a.
(c) Let A be a Boolean algebra. Show that the algebra of all bands of S(A) can be identied with the
Dedekind completion of A (314U).
(d) Let A be a Boolean ring, and U a complex normed space. Let : A U be an additive function
and T : S
C
(A) U the corresponding linear operator (cf. 361Xj). Show that (giving S
C
(A) its usual norm
| |

)
|T| = sup|

n
j=0

j
a
j
| : a
0
, . . . , a
n
A are disjoint, [
j
[ = 1 for every j
if either is nite.
(e) Let U be a Riesz space. Show that it is isomorphic to S(A), for some Boolean algebra A, i it has an
order unit and every solid linear subspace of U is a projection band.
(f ) Let A
i
)
iI
be a non-empty family of Boolean algebras, with free product A; write
i
: A
i
A for
the canonical maps, and
362 intro. S

287
C = inf
jJ

j
(a
j
) : J I is nite, a
j
A
j
for every j J.
Suppose that U is a linear space and : C U is such that
c = (c
i
(a)) +(c
i
(1 \ a))
whenever c C, i I and a A
i
. Show that there is a unique additive function : A U extending .
(Hint: 326Q.)
361 Notes and comments The space S(A) corresponds of course to the idea of simple function which
belongs to the very beginnings of the theory of integration (122A). All that 361D is trying to do is to set up
a logically sound description of this obvious concept which can be derived from the Boolean ring A itself. To
my eye, there is a defect in the construction there. It relies on the axiom of choice, since it uses the Stone
space; but none of the elementary properties of S(A) have anything to do with the axiom of choice. In 361Ya
I oer an alternative construction which is in a formal sense more elementary. If you work through the
suggestion there you will nd, however, that the technical details become signicantly more complicated, and
would be intolerable were it not for the intuition provided by the Stone space construction. Of course this
intuition is chiey valuable in the nitistic arguments used in 361E, 361F and 361I; and for these arguments
we really need the Stone representation only for nite Boolean rings, which does not depend on the axiom
of choice.
It is quite true that in most of this volume (and in most of this chapter) I use the axiom of choice without
scruple and without comment. I mention it here only because I nd myself using arguments dependent on
choice to prove theorems of a type to which the axiom cannot be relevant.
The linear space structure of S(A), together with the map , are uniquely determined by the rst universal
mapping theorem here, 361F. This result says nothing about the order structure, which needs the further
renement in 361Ga. What is striking is that the partial order dened by 361Ga is actually a lattice ordering,
so that we can have a universal mapping theorem for functions to Riesz spaces, as in 361Gc and 361Ja.
Moreover, the same ordering provides a happy abundance of results concerning order-continuous functions
(361Gb, 361Je-361Jf). When the codomain is a Dedekind complete Riesz space, so that we have a Riesz
space L

(S; U), and a corresponding modulus function T [T[ for linear operators, there are reasonably
natural formulae for [T[ in terms of T (361H); see also 361Xc and 361Yb. The multiplicative structure
of S(A) is dened by 361Xb, and the norm by 361I.
The Boolean ring A cannot be recovered from the linear space structure of S(A) alone (since this tells us
only the cardinality of A), but if we add either the ordering or the multiplication of S(A) then A is easy to
identify (361K, 361Xf).
The most important Boolean algebras of measure theory arise either as algebras of sets or as their
quotients, so it is a welcome fact that in such cases the spaces S(A) have straightforward representations in
terms of the construction of A (361L-361M).
In Chapter 24 I oered a paragraph in each section to sketch a version of the theory based on the eld of
complex numbers rather than the eld of real numbers. This was because so many of the most important
applications of these ideas involve complex numbers, even though (in my view) the ideas themselves are
most clearly and characteristically expressed in terms of real numbers. In the present chapter we are one
step farther away from these applications, and I therefore relegate complex numbers to the exercises, as in
361Xj and 361Yd.
362 S

The next stage in our journey is the systematic investigation of linear functionals on spaces S = S(A).
We already know that these correspond to additive real-valued functionals on the algebra A (361F). My
purpose here is to show how the structure of the Riesz space dual S

and its bands is related to the classes

of additive functionals introduced in 326-327. The rst step is just to check the identication of the linear
and order structures of S

and the space M of bounded nitely additive functionals (362A); all the ideas
needed for this have already been set out, and the basic properties of S

are covered by the general results in

356. Next, we need to be able to describe the operations on M corresponding to the Riesz space operations
288 Function spaces 362 intro.
[ [, , on S

, and the band projections from S

onto S

c
and S

; these are dealt with in 362B, with

a supplementary remark in 362D. In the case of measure algebras, we have some further important bands
which present themselves in M, rather than in S

, and which are treated in 362C. Since all these spaces

are L-spaces, it is worth taking a moment to identify their uniformly integrable subsets; I do this in 362F.
While some of the ideas here have interesting extensions to the case in which A is a Boolean ring without
identity, these can I think be left to one side; the work of this section will be done on the assumption that
every A is a Boolean algebra.
362A Theorem Let A be a Boolean algebra. Write S for S(A).
(a) The partially ordered linear space of all nitely additive real-valued functionals on A may be identied
with the partially ordered linear space of all real-valued linear functionals on S.
(b) The linear space of bounded nitely additive real-valued functionals on A may be identied with the
L-space S

of order-bounded linear functionals on S. If f S

corresponds to : A R, then f
+
S

corresponds to
+
, where

+
a = sup
ba
b
for every a A, and
|f| = sup
aA
a (1 \ a).
(c) The linear space of bounded countably additive real-valued functionals on A may be identied with
the L-space S

c
.
(d) The linear space of completely additive real-valued functionals on A may be identied with the L-space
S

.
proof By 361F, we have a canonical one-to-one correspondence between linear functionals f : S R and
additive functionals
f
: A R, given by setting
f
= f.
(a) Now it is clear that
f+g
=
f
+
g
,
f
=
f
for all f, g and , so this one-to-one correspondence
is a linear space isomorphism. To see that it is also an order-isomorphism, we need note only that
f
is
non-negative i f is, by 361Ga.
(b) Recall from 356N that, because S is a Riesz space with order unit (361Ee), S

has a corresponding
norm under which it is an L-space.
(i) If f S

, then
sup
bA
[
f
b[ = sup
bA
[f(b)[ sup[f(u)[ : u S, [u[ 1
is nite, and
f
is bounded.
(ii) Now suppose that
f
is bounded and that v S
+
. Then there is an 0 such that v 1
(361Ee). If u S and [u[ v, then we can express u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint (361Eb);
now [
i
[ whenever a
i
,= 0, so
[f(u)[ = [

n
i=0

f
a
i
[

n
i=0
[
f
a
i
[ = (
f
c
1

f
c
2
) 2sup
bA
[
f
b[,
setting c
1
= supa
i
: i n,
f
a
i
0, c
2
= supa
i
: i n,
f
a
i
< 0. This shows that f(u) : [u[ v is
bounded. As v is arbitrary, f S

(356Aa).
(iii) To check the correspondence between f
+
and
+
f
, rene the arguments of (i) and (ii) as follows.
Take any f S

. If a A,

+
f
a = sup
ba

f
b = sup
ba
f(b) supf(u) : u S, 0 u a = f
+
(a).
On the other hand, if u S and 0 u a, then we can express u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are
disjoint; now 0
i
1 whenever a
i
,= 0, so
f(u) =

n
i=0

f
a
i

f
c
+
f
a,
where c = supa
i
: i n,
f
a
i
0. As u is arbitrary, f
+
(a)
+
f
a. This shows that
+
f
= f
+
is nitely
additive, and that
+
f
=
f
+, as claimed.
362B S

289
(iv) Now, for any f S

,
|f| = [f[(1)
(356N)
= (2f
+
f)(1)
(352D)
= 2
+
f
1
f
1
(by (iii) just above)
= sup
aA
2
f
a
f
1 = sup
aA

f
a
f
(1 \ a).
(c) If f 0 in S

, then f is sequentially order-continuous i

f
is sequentially order-continuous (361Gb),
that is, i
f
is countably additive (326Gc). Generally, an order-bounded linear functional belongs to S

c
i it is expressible as the dierence of two sequentially order-continuous positive linear functionals (356Ab),
while a bounded nitely additive functional is countably additive i it is expressible as the dierence of two
non-negative countably additive functionals (326H); so in the present context f S

c
i
f
is bounded and
countably additive.
(d) If f 0 in S

, then f is order-continuous i
f
is order-continuous (361Gb), that is, i
f
is com-
pletely additive (326Kc). Generally, an order-bounded linear functional belongs to S

i it is expressible
as the dierence of two order-continuous positive linear functionals (356Ac), while a nitely additive func-
tional is completely additive i it is expressible as the dierence of two non-negative completely additive
functionals (326M); so in the present context f S

i
f
is completely additive.
362B Spaces of nitely additive functionals The identications in the last theorem mean that we
can relate the Riesz space structure of S(A)

to constructions involving nitely additive functionals. I have

already set out the most useful facts as exercises (326Yj, 326Ym, 326Yn, 326Yp, 326Yq); it is now time to
repeat them more formally.
Theorem Let A be a Boolean algebra. Write S = S(A), and let M be the Riesz space of bounded nitely
additive real-valued functionals on A, M

M the space of bounded countably additive functionals, and

M

the space of completely additive functionals.

(a) For any , M, , and [[ are dened by the formulae
( )(a) = sup
ba
b +(a \ b),
( )(a) = inf
ba
b +(a \ b),
[[(a) = sup
ba
b (a \ b) = sup
b,ca
b c
for every a A. Setting
|| = [[(1) = sup
aA
a (1 \ a),
M becomes an L-space.
(b) M

and M

are projection bands in M, therefore L-spaces in their own right. In particular, [[ M

for every M

, and [[ M

for every M

.
(c) The band projection P

: M M

is dened by the formula

(P

)(c) = infsup
nN
a
n
: a
n
)
nN
is a non-decreasing sequence with supremum c
whenever c A and 0 in M.
(d) The band projection P

: M M

is dened by the formula

(P

)(c) = infsup
aA
a : A is a non-empty upwards-directed set with supremum c
whenever c A and 0 in M.
290 Function spaces 362B
(e) If A M is upwards-directed, then A is bounded above in M i 1 : A is bounded above in R,
and in this case (if A ,= ) supA is dened by the formula
(supA)(a) = sup
A
a for every a A.
(f) Suppose that , M.
(i) The following are equiveridical:
() belongs to the band in M generated by ;
() for every > 0 there is a > 0 such that [a[ whenever [[a ;
() lim
n
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence in A such that lim
n
[[(a
n
) =
0.
(ii) Now suppose that , 0, and let
1
,
2
be the components of in the band generated by and
its complement. Then

1
c = sup
>0
inf
a
(c \ a),
2
c = inf
>0
sup
ac,a
a
for every c A.
proof (a) Of course = + ( )
+
, = ( )
+
, [[ = () (352D), so the formula of
362Ab gives
( )(a) = a + sup
ba
b b = sup
ba
b +(a \ b),
( )(a) = a sup
ba
b b = inf
ba
b +(a \ b),
[[(a) = sup
ba
b (a \ b) sup
b,ca
b c = sup
b,ca
(b \ c) (c \ b)
sup
b,ca
[[(b \ c) +[[(c \ b) = sup
b,ca
[[(b .c) [[(a).
The formula oered for || corresponds exactly to the formula in 362Ab for the norm of the associated
member of S(A)

; because S(A)

is an L-space under its norm, so is M.

(b) By 362Ac-362Ad, M

and M

may be identied with S(A)

c
and S(A)

, which are projection bands

in S(A)

(356B); so that M

and M

are projection bands in M, and are L-spaces in their own right (354O).
(c) Take any 0 in M. Set

c = infsup
nN
a
n
: a
n
)
nN
is a non-decreasing sequence with supremum c
for every c A. Then of course 0

c c for every c. The point is that

is countably additive. PPP Let

c
i
)
iN
be a disjoint sequence in A, with supremum c. Then for any > 0 we have non-decreasing sequences
a
n
)
nN
, a
in
)
nN
, for i N, such that
sup
nN
a
n
= c, sup
nN
a
in
= c
i
for i N,
sup
nN
a
n

c +,
sup
nN
a
in

c
i
+ 2
i
for every i N.
Set b
n
= sup
in
a
in
for each n; then b
n
)
nN
is non-decreasing, and
sup
nN
b
n
= sup
i,nN
a
in
= sup
iN
c
i
= c,
so

c sup
nN
b
n
= sup
nN
n

i=0
a
in
=

i=0
sup
nN
a
in

i=0

c
i
+ 2
i
=

i=0

c
i
+ 2.
On the other hand, a
n
c
i
)
nN
is a non-decreasing sequence with supremum c c
i
= c
i
for each i, so

c
i
sup
nN
(a
n
c
i
), and
362B S

291

i=0

c
i

i=0
sup
nN
(a
n
c
i
) = sup
nN

i=0
(a
n
c
i
)
(because a
n
)
nN
is non-decreasing)
sup
nN
a
n
(because c
i
)
iN
is disjoint)

c +.
As is arbitrary,

c =

i=0

c
i
; as c
i
)
iN
is arbitrary,

is countably additive. QQQ

Thus

. On the other hand, if

t
M

and 0
t
, then whenever c A and a
n
)
nN
is a
non-decreasing sequence with supremum c,

t
c = sup
nN

t
a
n
sup
nN
a
n
.
So we must have
t
c

c. This means that

= sup
t
:
t
M

,
t
= P

,
as claimed.
(d) The same ideas, with essentially elementary modications, deal with the completely additive part.
Take any 0 in M. Set

c = infsup
aA
a : A is a non-empty upwards-directed set with supremum c
for every c A. Then of course 0

c c for every c. The point is that

is completely additive.
PPP Note rst that if c A, > 0 there is a non-empty upwards-directed A, with supremum c, such that
sup
aA
a

c +c; for if c = 0 we can take A = c. Now let c

i
)
iI
be a partition of unity in A. Then
for any > 0 we have non-empty upwards-directed sets A, A
i
, for i I, such that
supA = 1, supA
i
= c
i
for i I, sup
aA
a

1 +1,
sup
aA
i
a

c
i
+c
i
for every i I.
Set
B = sup
iJ
a
i
: J I is nite, a
i
A
i
for every i J;
then B is non-empty and upwards-directed, and
supB = sup(

iI
A
i
) = 1,
so

1 sup
bB
b = sup

iJ
a
i
: J I is nite, a
i
A
i
i J

iI

c
i
+c
i
1 +

iI

c
i
.
On the other hand, A
t
i
= a c
i
: a A is a non-empty upwards-directed set with supremum c
i
for each i,
so

c
i
sup
aA

i
a, and

iI

c
i

iI
sup
aA
(a c
i
) = sup
aA

iI
(a c
i
)
sup
aA
a

1 +1.
As is arbitrary,

c =

iI

c
i
; as c
i
)
iI
is arbitrary,

is completely additive, by 326N. QQQ

Thus

. On the other hand, if

t
M

and 0
t
, then whenever c A and A is a non-empty
upwards-directed set with supremum c,
292 Function spaces 362B

t
c = sup
aA

t
a sup
aA
a
(using 326Kc). So we must have
t
c

c. This means that

= sup
t
:
t
M

,
t
= P

,
as claimed.
(e) If A is empty, of course it is bounded above in M, and 1 : A = is bounded above in R; so
let us suppose that A is not empty. In this case, if
0
M is an upper bound for A, then
0
1 is an upper
bound for 1 : A. On the other hand, if sup
A
1 = is nite,

= supa : A, a A is
nite. PPP Fix
0
A. Set
1
= sup
aA
[
0
a[ < . Then for any A, a A there is a
t
A such that

0

t
, so that
a
t
a =
t
1
t
(1 \ a)
0
(1 \ a) +
1
.
So

+
1
< . QQQ
Set a = sup
A
a for every a A. Then : A R is additive. PPP If a, b A are disjoint, then
(a b) = sup
A
(a b) = sup
A
a +b = sup
A
a + sup
A
b
(because A is upwards-directed)
= a +b. QQQ
Also a

for every a, so
[a[ = max(a, a) = max(a, (1 \ a) 1)

+[1[
for every a A, and is bounded.
This shows that M, so that A is bounded above in M. Of course must be actually the least upper
bound of A in M.
(f )(i)()() Suppose that belongs to the band in M generated by , that is, [[ = sup
nN
[[ n[[
(352Vb). Let > 0. Then there is an n N such that [[(1)
1
2
+ ([[ n[[)(1) ((e) above). Set
=
1
2n+1
> 0. If [[(a) , then
[a[ [[(a) = ([[ n[[)(a) + ([[ [[ n[[)(a)
n[[(a) + ([[ [[ n[[)(1) n +
1
2
.
So () is satised.
()() Suppose that does not belong to the band in M generated by [[. Then there is a
1
> 0
such that
1
[[ and
1
[[ = 0 (353C). For any > 0, there is an a A such that
1
(1 \ a) + [[(a)
min(,
1
2

1
1) ((a) above); now [[(a) but
[[(a)
1
a =
1
1
1
(1 \ a)
1
1
1
2

1
1 =
1
2

1
1.
Thus , do not satisfy () (with =
1
2

1
1).
()() is trivial.
()() Observe rst that if c
k
)
kN
is a non-increasing sequence in A such that lim
k
c
k
= 0,
then lim
k

+
c
k
= 0. PPP Let > 0. Because
+

= 0, there is a b A such that

+
b +

(1 \ b) ,
by part (a). Now c
k
\ b)
kN
is non-increasing and lim
k
(c
k
\ b) = 0, so lim
k
(c
k
\ b) = 0 and
limsup
k

+
c
k
= limsup
k

+
(c
k
b) +(c
k
\ b) +

(c
k
\ b)

+
b +

(1 \ b) .
As is arbitrary, lim
k

+
c
k
= 0. QQQ
362C S

293
??? Now suppose, if possible, that
+
does not belong to the band generated by . Then there is a
1
> 0
such that
1

+
and
1
[[ = 0. Set =
1
4

1
1 > 0. For each n N, we can choose a
n
A such that
[[a
n
+
1
(1 \ a
n
) 2
n
, by part (a) again. For n k, set b
kn
= sup
kin
a
i
; then
[[b
kn

n
i=k
[[a
i
2
k+1
,
and b
kn
)
nk
is non-decreasing. Set
k
= sup
nk

1
b
kn
and choose m(k) k such that
1
b
k,m(k)

k
2
k
.
Setting b
k
= b
k,m(k)
, we see that b
k
b
k+1
= b
kn
where n = max(m(k), m(k + 1)), so that

1
(b
k
b
k+1
)
k

1
b
k
+ 2
k

and
1
(b
k+1
\ b
k
) 2
k
. Set c
k
= inf
ik
b
i
for each k; then

1
(b
k+1
\ c
k+1
) =
1
(b
k+1
\ c
k
)
1
(b
k+1
\ b
k
) +
1
(b
k
\ c
k
) 2
k
+
1
(b
k
\ c
k
)
for each k; inducing on k, we see that

1
(b
k
\ c
k
)

k1
i=0
2
i
2
for every k. This means that

+
c
k

1
c
k

1
b
k
2
1
a
k
2 =
1
1
1
(1 \ a
k
) 2 4 2 =
for every k N. On the other hand, c
k
)
kN
is a non-decreasing sequence and
[[c
k
[[b
k
2
k+1

for every k, which contradicts the paragraph just above. XXX

This means that
+
must belong to the band generated by . Similarly

= ()
+
belongs to the band
generated by and =
+
+

also does.
(ii) Take c A. Set

1
= sup
>0
inf
a
(c \ a),
2
= inf
>0
sup
ac,a
a.
Then

1
= sup
>0
inf
ac,a
(c \ a) = c
2
.
Take any > 0. Because
1
belongs to the band generated by , part (i) tells us that there is a > 0 such
that
1
a whenever a . In this case, if a ,
(c \ a) = c (c a) c
1
c ;
thus

1
inf
a
(c \ a)
1
c .
As is arbitrary,
1

1
c. On the other hand, given , > 0, there is an a c such that a +
2
(c \ a)
min(, ), because
2
= 0 (using (a) again). In this case, of course, a , while
a
2
a =
2
c
2
(c \ a)
2
c .
Thus sup
ac,a
a
2
c . As is arbitrary,
2

2
c . As is arbitrary,
2

2
c; but as

1
+
2
= c =
1
c +
2
c,

i
=
i
c for both i, as claimed.
362C The formula in 362B(f-i) has, I hope, already reminded you of the concept of absolutely contin-
uous additive functional from the Radon-Nikod ym theorem (Chapter 23, 327). The expressions in 362Bf
are limited by the assumption that , like , is nite-valued. If we relax this we get an alternative version
of some of the same ideas.
Theorem Let (A, ) be a measure algebra. Write S = S(A), and let M be the Riesz space of bounded
nitely additive real-valued functionals on A. Write
M
ac
= : M is absolutely continuous with respect to
(see 327A),
294 Function spaces 362C
M
tc
= : M is continuous with respect to the measure-algebra topology on A,
M
t
= : M, [[1 = sup
a<
[[a.
Then M
ac
, M
tc
and M
t
are bands in M.
proof (a)(i) It is easy to check that M
ac
is a linear subspace of M.
(ii) If M
ac
,
t
M and [
t
[ [[ then
t
M
ac
. PPP Given > 0 there is a > 0 such that
[a[
1
2
whenever a . Now
[
t
a[ [
t
[(a) [[(a) 2 sup
ca
[c[
(using the formula for [[ in 362Ba) whenever a . As is arbitrary,
t
is absolutely continuous. QQQ
(iii) If A M
ac
is non-empty and upwards-directed and = supA in M, then M
ac
. PPP Let > 0.
Then there is a
t
A such that 1
t
1 +
1
2
(362Be). Now there is a > 0 such that [a[
1
2
whenever
a . If now a ,
[a[ [
t
a[ + (
t
)(a)
1
2
+ (
t
)(1) .
As is arbitrary, is absolutely continuous with respect to . QQQ
Putting these together, we see that M
ac
is a band.
(b)(i) We know that M
tc
consists just of those M which are continuous at 0 (327Bc). Of course this
is a linear subspace of M.
(ii) If M
tc
,
t
M and [
t
[ [[ then [[ M
tc
. PPP Write A
f
= d : d A, d < . Given > 0
there are d A
f
, > 0 such that [a[
1
2
whenever (a d) . Now
[
t
a[ [
t
[(a) [[(a) 2 sup
ca
[c[
whenever (a d) . As is arbitrary,
t
is continuous at 0 and belongs to M
tc
. QQQ
(iii) If A M
tc
is non-empty and upwards-directed and = sup A in M, then M
tc
. PPP Let > 0.
Then there is a
t
A such that 1
t
1 +
1
2
(362Be). There are d A
f
, > 0 such that [a[
1
2

whenever (a d) . If now (a d) ,
[a[ [
t
a[ + (
t
)(a)
1
2
+ (
t
)(1) .
As is arbitrary, is continuous at 0, therefore belongs to M
tc
. QQQ
Putting these together, we see that M
tc
is a band.
(c)(i) M
t
is a linear subspace of M. PPP Suppose that
1
,
2
M
t
and R. Given > 0, there are a
1
,
a
2
A
f
such that [
1
[(1 \ a
1
)

1+]]
, [
2
[(1 \ a
2
) . Set a = a
1
a
2
; then a < and
[
1
+
2
[(1 \ a) 2, [
1
[(1 \ a) .
As is arbitrary,
1
+
2
and
1
belong to M
t
; as
1
,
2
and are arbitrary, M
t
is a linear subspace of M.
QQQ
(ii) If M
t
,
t
M and [
t
[ [[ then
inf
a<
[
t
[(1 \ a) inf
a<
[[(1 \ a) = 0,
so
t
M
t
. Thus M
t
is a solid linear subspace of M.
(iii) If A M
+
t
is non-empty and upwards-directed and = supA is dened in M, then M
t
. PPP
[[1 = 1 = sup

t
1 = sup

A, a<

t
a = sup
a<
a.
As A is arbitrary, M
t
. QQQ Thus M
t
is a band in M.
362Xb S

295
362D For semi-nite measure algebras, among others, the formula of 362Bd takes a special form.
Proposition Let A be a weakly (, )-distributive Boolean algebra. Let M be the space of bounded nitely
additive functionals on A, M

M the space of completely additive functionals, and P

: M M

the
band projection, as in 362B. Then for any M
+
and c A there is a non-empty upwards-directed set
A A with supremum c such that (P

)(c) = sup
aA
a; that is, the inf in 362Bd can be read as min.
proof By 362Bd, we can nd for each n a non-empty upwards-directed A
n
, with supremum c, such that
sup
aA
n
a (P

)(c) + 2
n
. Set B
n
= c \ a : a A
n
for each n, so that B
n
is downwards-directed and
has inmum 0, and
B = b : for every n N there is a b
t
B
n
such that b b
t

is also a downwards-directed set with inmum 0 (316H). Consequently A = c \ b : b B is upwards-

directed and has supremum c. Moreover, for any n N and a A, there is an a
t
A
n
such that a a
t
; so,
using 362Bd again and referring to the choice of A
n
,
(P

)(c) sup
aA
a sup
a

A
n
a
t
(P

)(c) + 2
n
.
As n is arbitrary, A has the required property.
362E Uniformly integrable sets The spaces S

, S

c
and S

of 362A, or, if you prefer, the spaces

M, M

, M

, M
ac
, M
tc
, M
t
of 362B-362C, are all L-spaces, and any serious study of them must involve a
discussion of their uniformly integrable ( = relatively weakly compact) subsets. The basic work has been
done in 356O; I spell out its application in this context.
Theorem Let A be a Boolean algebra and M the L-space of bounded nitely additive functionals on A.
Then a norm-bounded set C M is uniformly integrable i lim
n
sup
C
[a
n
[ = 0 for every disjoint
sequence a
n
)
nN
in A.
proof Write

C for the set f : f S(A)

, f C. Because the map f f is a normed Riesz space

isomorphism between S

and M,

C is uniformly integrable in M i C is uniformly integrable in S

.
(a) Suppose that C is uniformly integrable and that a
n
)
nN
is a disjoint sequence in A. Then a
n
)
nN
is a disjoint order-bounded sequence in S

, while

C is uniformly integrable, so lim
n
sup
f

C
[f(a
n
)[ = 0,
by 356O; but this means that lim
n
sup
C
[a
n
[ = 0. Thus the condition is satised.
(b) Now suppose that C is not uniformly integrable. By 356O, in the other direction, there is a disjoint
sequence u
n
)
nN
in S such that 0 u
n
1 for each n and limsup
n
sup
f

C
[f(u
n
)[ > 0. For each n,
take c
n
= [[u
n
> 0]] (361Eg); then 0 u
n
c
n
and c
n
)
nN
is disjoint. Now
limsup
n
sup
C
[[(c
n
) = limsup
n
sup
f

C
[f[(c
n
)
limsup
n
sup
f

C
[f(u
n
)[ > 0.
So if we choose
n
C such that [
n
[(c
n
)
1
2
sup
C
[[(c
n
), we shall have limsup
n
[
n
[(c
n
) > 0. Next,
for each n, we can nd a
n
c
n
such that [
n
a
n
[
1
2
[
n
[(c
n
), so that
limsup
nN
sup
C
[a
n
[ limsup
n
[
n
a
n
[ > 0.
Since a
n
)
nN
, like c
n
)
nN
, is disjoint, the condition is not satised. This completes the proof.
362X Basic exercises (a) Let A be a Dedekind -complete Boolean algebra and
1
,
2
two countably
additive functionals on A. Show that [
1
[[
2
[ = 0 in the Riesz space of bounded nitely additive functionals
on A i there is a c A such that
1
a =
1
(a c) and
2
a =
2
(a \ c) for every a A.
(b) Let (A, ) be a measure algebra. Write M, M
ac
as in 362C. Show that for any non-negative M,
the component
ac
of in M
ac
is given by the formula

ac
c = sup
>0
inf
a
(c \ a).
296 Function spaces 362Xc
(c) Let (A, ) be a measure algebra. Write M, M
t
as in 362C. (i) Show that M
t
is just the set of those
M such that a = lim
bJ
(a b) for every a A, where T is the lter on A generated by the sets
b : b A
f
, b b
0
as b
0
runs through the elements of A of nite measure. (ii) Show that the complementary
band M

t
of M
t
in M is just the set of M such that a = 0 whenever a < . (iii) Show that for any
M, its component
t
in M
t
is given by the formula
t
a = lim
bJ
(a b) for every a A.
(d) Let (A, ) be a measure algebra. Write M, M

, M

, M
ac
, M
tc
and M
t
as in 362B-362C. Show that
(i) M

M
ac
(ii) M
ac
M
t
= M
tc
M

(iii) if (A, ) is -nite, then M

= M
t
M
ac
.
(e) Let A be a Boolean algebra. Let us say that a non-zero nitely additive functional : A R is an
atom if whenever a, b A and a b = 0 then at least one of a, b is zero. Show that for a non-zero nitely
additive functional the following are equiveridical: (i) is an atom; (ii) is bounded and [[ is an atom;
(iii) is bounded and the corresponding linear functional f
]]
= [f

[ S(A)

is a Riesz homomorphism;
(iv) there are a multiplicative linear functional f : S(A) R and an R such that a = f(a) for every
a A. Show that a completely additive functional : A R is an atom i there are a A, R 0
such that a is an atom in A and b = when a b, 0 when a b = 0.
(f ) Let A be a Boolean algebra. Let us say that a bounded nitely additive functional : A R is
atomless if for every > 0 there is a nite partition C of unity in A such that [[c for every c C
(cf. 326Ya). (i) Show that the atomless functionals form a band M
c
in the Riesz space M of all bounded
nitely additive functionals on A. (ii) Show that the complementary band M

c
consists of just those M
expressible as a sum

iI

i
of countably many atoms
i
M. (iii) Show that if A is purely atomic then
an atomless completely additive functional on A must be 0.
(g) Let X be a set and an algebra of subsets of X. Let M be the Riesz space of bounded nitely
additive functionals on , M

the space of completely additive functionals and M

p
the space of functionals
expressible in the form E =

xE

x
for some absolutely summable family
x
)
xX
of real numbers. (i)
Show that M
p
is a band in M. (ii) Show that if all singleton subsets of X belong to then M
p
= M

.
(iii) Show that if is a -algebra then every member of M
p
is countably additive. (iv) Show that if X is
a compact zero-dimensional Hausdor space and is the algebra of open-and-closed subsets of X then the
complementary band M

p
of M
p
in M is the band M
c
of atomless functionals described in 362Xf.
(h) Let (X, , ) be a measure space. Let M be the Riesz space of bounded nitely additive functionals
on and M

the space of bounded countably additive functionals. Let M

tc
, M
ac
be the spaces of truly
continuous and bounded absolutely continuous additive functionals as dened in 232A. Show that M
tc
and
M
ac
are bands in M and that M
tc
M

M
ac
. Show that if is -nite then M
tc
= M

M
ac
.
(i) Let A be a Boolean algebra and M the Riesz space of bounded nitely additive functionals on A.
(i) For any non-empty downwards-directed set A A set N
A
= : M, inf
aA
[[a = 0. Show
that N
A
is a band in M. (ii) For any non-empty set / of non-empty downwards-directed sets in A set
M
,
= : M, inf
aA
[[a = 0 A /. Show that M
,
is a band in M. (iii) Explain how to represent
as such M
,
the bands M

, M

, M
t
, M
ac
, M
tc
described above, and also any band generated by a single
element of M. (iv) Suppose, in (ii), that / has the property that for any A, A
t
/ there is a B / such
that for every b B there are a A, a
t
A
t
such that a a
t
b. Show that for any non-negative M,
the component
1
of in M
,
is given by the formula
1
c = inf
A,
sup
aA
(c \ a), so that the component

2
of in M

,
is given by the formula
2
c = sup
A,
inf
aA
(c a). (Cf. 356Yb.)
362Y Further exercises (a) Let A be a Boolean algebra. Let C be the band algebra of the Riesz space
M of bounded nitely additive functionals on A (353B). Show that the bands M

, M

, M
c
(362B, 362Xe,
362Xf) generate a subalgebra C
0
of C with at most six atoms. Give an example in which C
0
has six atoms.
How many atoms can it have if (i) A is atomless (ii) A is purely atomic (iii) A is Dedekind -complete?
(b) Let (A, ) be a measure algebra. Let C be the band algebra of the Riesz space M of bounded nitely
additive functionals on A. Show that the bands M

, M

, M
c
, M
ac
, M
tc
, M
t
(362B, 362C, 362Xe, 362Xf)
generate a subalgebra C
0
of C with at most twelve atoms. Give an example in which C
0
has twelve atoms.
How many atoms can it have if (i) A is atomless (ii) A is purely atomic (iii) (A, ) is semi-nite (iv) (A, )
is localizable (v) (A, ) is -nite (vi) (A, ) is totally nite?
362 Notes S

297
(c) Give an example of a set X, a -algebra of subsets of X, and a functional in M
p
(as dened in
362Xg) which is not completely additive.
(d) Let U be a Riesz space and f, g U

. Show that the following are equiveridical: () g is in the

band of U

generated by f; () for every u U

+
, > 0 there is a > 0 such that [g(v)[ whenever
0 v u and [f[(v) ; () lim
n
g(u
n
) = 0 whenever u
n
)
nN
is a non-increasing sequence in U
+
and
lim
n
f(u
n
) = 0. (Hint: 362B(f-i).)
(e) Let A be a weakly -distributive Boolean algebra (316Yg). Show that the inf in the formula for P

in 362Bc can be replaced by min.

(f ) Let A be any Boolean algebra and M the space of bounded nitely additive functionals on A. Let
C M be such that sup
C
[a[ < for every a A. (i) Suppose that sup
nN
sup
C
[a
n
[ is nite for every
disjoint sequence a
n
)
nN
in A. Show that C is norm-bounded. (ii) Suppose that lim
n
sup
C
[a
n
[ = 0
for every disjoint sequence a
n
)
nN
in A. Show that C is uniformly integrable.
(g) Let A be a Boolean algebra and M

the space of completely additive functionals on A. Let C M

be such that sup

C
[a[ < for every atom a A. (i) Suppose that sup
nN
sup
C
[a
n
[ is nite for every
disjoint sequence a
n
)
nN
in A. Show that C is norm-bounded. (ii) Suppose that lim
n
sup
C
[a
n
[ = 0
for every disjoint sequence a
n
)
nN
in A. Show that C is uniformly integrable.
(h) Let A be a Dedekind -complete Boolean algebra and
n
)
nN
a sequence of countably additive real-
valued functionals on A. Suppose that a = lim
n

n
a is dened in R for every a A. Show that is
countably additive and that
n
: n N is uniformly integrable. (Hint: 246Yg.) Show that if every
n
is
completely additive, so is .
(i) Let A be a Boolean algebra, M the Riesz space of bounded nitely additive functionals on A, and
M
c
M the space of atomless functionals (362Xf). Show that for a non-negative M the component
c
of in M
c
is given by the formula

c
a = inf
>0
sup

n
i=0
a
i
: a
0
, . . . , a
n
a are disjoint, a
i
for every i
for each a A.
(j) Let A be a Boolean algebra and M the L-space of bounded additive real-valued functionals on A. Show
that the complexication of M, as dened in 354Yk, can be identied with the Banach space of bounded
additive functionals : A C, writing
|| = sup

n
i=0
[a
i
[ : a
0
, . . . , a
n
are disjoint elements of A
for such .
362 Notes and comments The Boolean algebras most immediately important in measure theory are of
course -algebras of measurable sets and their quotient measure algebras. It is therefore natural to begin any
investigation by concentrating on Dedekind -complete algebras. Nevertheless, in this section and the last
(and in 326), I have gone to some trouble not to specialize to -complete algebras except when necessary.
Partly this is just force of habit, but partly it is because I wish to lay a foundation for a further step forward:
the investigation of the ways in which additive functionals on general Boolean algebras reect the concepts
of measure theory, and indeed can generate them. Some of the results in this direction can be surprising. I
do not think it obvious that the condition () in 362B(f-i), for instance, is sucient in the absence of any
hypothesis of Dedekind -completeness or countable additivity.
Given a Boolean algebra A with the associated Riesz space M

= S(A)

of bounded additive functionals

on A, we now have a substantial list of bands in M: M

, M

, M
c
(362Xf), and for a measure algebra the
further bands M
ac
, M
tc
and M
t
; for an algebra of sets we also have M
p
(362Xg). These bands can be used
to generate nite subalgebras of the band algebra of M (362Ya-362Yb), and for any such nite subalgebra
we have a corresponding decomposition of M as a direct sum of the bands which are the atoms of the
subalgebra (352Tb). This decomposition of M can be regarded as a recipe for decomposing its members
into nite sums of functionals with special properties. What I called the Lebesgue decomposition in 232I
298 Function spaces 362 Notes
is just such a recipe. In that context I had a measure space (X, , ) and was looking at the countably
additive functionals from to R, that is, at M

in the language of this section, and the bands involved

in the decomposition were M
p
, M
ac
and M
tc
. But I hope that it will be plain that these ideas can be
rened indenitely, as we rene the classication of additive functionals. At each stage, of course, the exact
enumeration of the subalgebra of bands generated by the classication (as in 362Ya-362Yb) is a necessary
check that we have understood the relationships between the classes we have described.
These decompositions are of such importance that it is worth examining the corresponding band pro-
jections. I give formulae for the action of band projections on (non-negative) functionals in 362Bc, 362Bd,
362B(f-ii), 362Xb, 362Xc(iii), 362Xi(iv) and 362Yi. Of course these are readily adapted to give formulae for
the projections onto the complementary bands, as in 362Bf and 362Xi.
If we have an algebra of sets, the completely additive functionals are (usually) of relatively minor impor-
tance; in the standard examples, they correspond to functionals dened as weighted sums of point masses
(362Xg(ii)). The point is that measure algebras A appear as quotients of -algebras of sets by -ideals 1;
consequently the countably additive functionals on A correspond exactly to the countably additive function-
als on which are zero on 1; but the canonical homomorphism from to A is hardly ever order-continuous,
so completely additive functionals on A rarely correspond to completely additive functionals on . On the
other hand, when we are looking at countably additive functionals on , we have to consider the possibility
that they are singular in the sense that they are carried on some member of 1; in the measure algebra
context this possibility disappears, and we can often be sure that every countably additive functional is
absolutely continuous, as in 327Bb.
For any Boolean algebra A, we can regard it as the algebra of open-and-closed subsets of its Stone space
Z; the points of Z correspond to Boolean homomorphisms from A to 0, 1, which are the fundamental
atoms in the space of additive functionals on A (362Xe, 362Xg(iv)). It is the case that all non-negative
additive functionals on a Boolean algebra A can be represented by appropriate measures on its Stone space
(see 416P in Volume 4), but I prefer to hold this result back until it can take its place among other theorems
on representing functionals by measures and integrals.
It is one of the leitmotivs of this chapter, that Boolean algebras and Riesz spaces are Siamese twins; again
and again, matching results are proved by the application of identical ideas. A typical example is the pair
362B(f-i) and 362Yd. Many of us have been tempted to try to describe something which would provide a
common generalization of Boolean algebras and Riesz spaces (and lattice-ordered groups). I have not yet
seen any such structure which was worth the trouble. Most of the time, in this chapter, I shall be using
ideas from the general theory of Riesz spaces to suggest and illuminate questions in measure theory; but if
you pursue this subject you will surely nd that intuitions often come to you rst in the context of Boolean
algebras, and the applications to Riesz spaces are secondary.
In 362E I give a condition for uniform integrability in terms of disjoint sequences, following the pattern
established in 246G and repeated in 354R and 356O. The condition of 362E assumes that the set is norm-
bounded; but if you have 246G to hand, you will see that it can be done with weaker assumptions involving
atoms, as in 362Yf-362Yg.
I mention once again the Banach-Ulam problem: if A is Dedekind complete, can S(A)

c
be dierent from
S(A)

? This is obviously equivalent to the form given in the notes to 326 above. See 363S below.
363 L

In this section I set out to describe an abstract construction for L

spaces on arbitrary Boolean algebras,

corresponding to the L

() spaces of 243. I begin with the denition of L

(A) (363A) and elementary

facts concerning its own structure and the embedding S(A) L

(A) (363B-363D). I give the basic universal

mapping theorems which dene the Banach lattice structure of L

(363E) and a description of the action

of Boolean homomorphisms on L

spaces (363F-363G) before discussing the representation of L

() and
L

(/1) for -algebras and ideals 1 of sets (363H). This leads at once to the identication of L

(), as
dened in Volume 2, with L

(A), where A is the measure algebra of (363I). Like S(A), L

(A) determines
the algebra A (363J). I briey discuss the dual spaces of L

; they correspond exactly to the duals of S

described in 362 (363K). Linear functionals on L

can for some purposes be treated as integrals (363L).

363E L

299
In the second half of the section I present some of the theory of Dedekind complete and -complete
algebras. First, L

(A) is Dedekind (-)complete i A is (363M). The spaces L

(A), for Dedekind -

complete A, are precisely the Dedekind -complete Riesz spaces with order unit (363N-363P). The spaces
L

(A), for Dedekind complete A, are precisely the normed spaces which may be put in place of R in
the Hahn-Banach theorem (363R). Finally, I mention some equivalent forms of the Banach-Ulam problem
(363S).
363A Denition Let A be a Boolean algebra, with Stone space Z. I will write L

(A) for the space

C(Z) = C
b
(Z) of continuous real-valued functions from Z to R, endowed with the linear structure, order
structure, norm and multiplication of C(Z) = C
b
(Z). (Recall that because Z is compact (311I), u(z) : z
Z is bounded for every u L

(A) = C(Z) (2A3N(b-iii)), that is, C(Z) = C

b
(Z). Of course if A = 0, so
that Z = , then C(Z) has just one member, the empty function.)
363B Theorem Let A be any Boolean algebra; write L

for L

(A).
(a) L

is an M-space; its standard order unit is the constant function taking the value 1 at each point;
in particular, it is a Banach lattice with a Fatou norm and the Levi property.
(b) L

is a commutative Banach algebra and an f-algebra.

(c) If u L

then u 0 i there is a v L

such that u = v v.
proof (a) See 354Hb and 354J.
(b)-(c) are obvious from the denitions of Banach algebra (2A4J) and f-algebra (352W) and the ordering
of L

= C(Z).
363C Proposition Let A be any Boolean algebra. Then S(A) is a norm-dense, order-dense Riesz
subspace of L

(A), closed under multiplication.

proof Let Z be the Stone space of A. Using the denition of S = S(A) set out in 361D, it is obvious that
S is a linear subspace of L

= L

(A) closed under multiplication. Because S, like L

, is a Riesz subspace
of R
X
(361Ee), S is a Riesz subspace of L

. By the Stone-Weierstrass theorem (in either of the forms given

in 281A and 281E), S is norm-dense in L

. Consequently it is order-dense (354I).

363D Proposition Let A be a Boolean algebra. If we regard a S(A) (361D) as a member of
L

(A) for each a A, then : A L

(A) is additive, order-preserving, order-continuous and a lattice

homomorphism.
proof Because the embedding S = S(A) L

(A) = L

is a Riesz homomorphism, : A L

is additive
and a lattice homomorphism (361F-361G). Because S is order-dense in L

(363C), the embedding S L

is order-continuous (352Nb), so : A L

is order-continuous (361Gb).
363E Theorem Let A be a Boolean algebra, and U a Banach space. Let : A U be a bounded
additive function.
(a) There is a unique bounded linear operator T : L

(A) U such that T = ; in this case |T| =

sup
a,bA
|a b|.
(b) If U is a Banach lattice then T is positive i is non-negative; and in this case T is order-continuous
i is order-continuous, and sequentially order-continuous i is sequentially order-continuous.
(c) If U is a Banach lattice then T is a Riesz homomorphism i is a lattice homomorphism i ab = 0
whenever a b = 0.
proof Write S = S(A), L

= L

(A).
(a) By 361I there is a unique bounded linear operator T
0
: S U such that T
0
= , and |T
0
| =
sup|a b| : a, b A. But because U is a Banach space and S is dense in L

, T
0
has a unique
extension to a bounded linear operator T : L

U with the same norm (2A4I).

(b)(i) If T is positive then T
0
is positive so is non-negative, by 361Ga.
300 Function spaces 363E
(ii) If is non-negative then T
0
is positive, by 361Ga in the other direction. But if u L
+
and > 0,
then by 354I there is a v S
+
such that |u v|

; now |Tu Tv| |T|. But Tv = T

0
v belongs to
the positive cone U
+
of U. As is arbitrary, Tu belongs to the closure of U
+
, which is U
+
(354Bc). As u
is arbitrary, T is positive.
(iii) Now suppose that is order-continuous as well as non-negative, and that A L

is a non-empty
downwards-directed set with inmum 0. Set
B = v : v S, there is some u A such that v u.
Then B is downwards-directed (indeed, v
1
v
2
B for every v
1
, v
2
B), and u = infv : v B, u v
for every u A (354I again), so B has the same lower bounds as A and inf B = 0 in L

and in S. But we
know from 361Gb that T
0
is order-continuous, while any lower bound for Tu : u A in U must also be a
lower bound for Tv : v B = T
0
v : v B, so inf
uA
Tu = inf
vB
T
0
v = 0 in U. As A is arbitrary, T is
order-continuous (351Ga).
(iv) Suppose next that is only sequentially order-continuous, and that u
n
)
nN
is a non-increasing
sequence in L

with inmum 0. For each n, k choose w

nk
S such that u
n
w
nk
and |w
nk
u
n
|

2
k
(354I once more), and set w
t
n
= inf
j,kn
w
jk
for each n. Then w
t
n
)
nN
is a non-increasing sequence in S,
and any lower bound of w
t
n
: n N is also a lower bound of u
n
: n N, so 0 = inf
nN
w
t
n
in S and L

.
Since T
0
: S U is sequentially order-continuous (361Gb),
inf
nN
Tu
n
inf
nN
Tw
t
n
= inf
nN
T
0
w
t
n
= 0
in U. As u
n
)
nN
is arbitrary, T is sequentially order-continuous.
(v) On the other hand, if T is order-continuous or sequentially order-continuous, so is = T, because
is order-continuous (363D).
(c) We know that T
0
: S U is a Riesz homomorphism i is a lattice homomorphism i a b = 0
whenever a b = 0, by 361Gc. But T
0
is a Riesz homomorphism i T is. PPP If T is a Riesz homomorphism
so is T
0
, because the embedding S L

is a Riesz homomorphism. On the other hand, if T

0
is a Riesz
homomorphism, then the functions u u
+
T(u
+
), u Tu (Tu)
+
are continuous (by 354Bb) and
agree on S, so agree on L

, and T is a Riesz homomorphism, by 352G. QQQ

363F Theorem Let A and B be Boolean algebras, and : A B a Boolean homomorphism.
(a) There is an associated multiplicative Riesz homomorphism T

: L

(A) L

(B), of norm at most

1, dened by saying that T

(a) = (a) for every a A.

(b) For any u L

(A), there is a u
t
L

(A) such that T

u = T

u
t
and |u
t
|

= |T

u|

|u|

.
(c)(i) The kernel of T

is the closed linear subspace of L

(A) generated by a : a A, a = 0.
(ii) The set of values of T

is the closed linear subspace of L

(B) generated by (a) : a A.

(d) T

is surjective i is surjective, and in this case |v|

= min|u|

: T

u = v for every v L

(B).
(e) T

is injective i is injective, and in this case |T

u|

= |u|

for every u L

(A).
(f) T

is order-continuous, or sequentially order-continuous, i is.

(g) If C is another Boolean algebra and : B C is another Boolean homomorphism, then T

= T

:
L

(A) L

(C).
proof Let Z and W be the Stone spaces of A and B. By 312P there is a continuous function : W Z
such that a =
1
[a] for every a A, where a is the open-and-closed subset of Z corresponding to a A.
Write T for T

.
(a) For u L

(A) = C(Z), set Tu = u : W R. Then Tu C(W) = L

(B). It is obvious, or at
any rate very easy to check, that T : L

(A) L

(B) is linear, multiplicative, a Riesz homomorphism and

of norm 1 unless B = 0, W = . If a A, then
T(a) = (a) = (a) = (
1
[a]) = (a),
identifying a L

(A) with the characteristic function a : Z 0, 1 of the set a. Of course T

= T is
the only continuous linear operator with these properties, by 363Ea.
(b) Set = |Tu|

, u
t
(z) = max(, min(u(z), )) for z Z; that is, u
t
= (e) (u e) in L

(A),
where e is the standard order unit of L

(A). Then Te is the standard order unit of L

(B), so
363F L

301
Tu
t
= (Te) (Tu Te) = Tu,
while
|u
t
|

= |Tu|

= |Tu
t
|

|u
t
|

|u|

.
(c)(i) Let U be the closed linear subspace of L

(A) generated by a : a = 0, and U

0
the kernel of T.
Because T is continuous and linear, U
0
is a closed linear subspace, and T(a) = 0 = 0 whenever a = 0;
so U U
0
. Now take any u U
0
and > 0. Then T(u
+
) = (Tu)
+
= 0, so u
+
U
0
. By 354I there is a
u
t
S(A) such that 0 u
t
u
+
and |u
+
u
t
|

. Now 0 Tu
t
Tu
+
= 0, so Tu
t
= 0. Express
u
t
as

n
i=0

i
a
i
where
i
0 for each i. For each i,
i
(a
i
) = T(
i
a
i
) = 0, so a
i
= 0 or
i
= 0;
in either case
i
a
i
U. Consequently u
t
U. As is arbitrary and U is closed, u
+
U. Similarly,
u

= (u)
+
U and u = u
+
u

U. As u is arbitrary, U
0
U and U
0
= U.
(ii) Let V be the closed linear subspace of L

(B) generated by (a) : a A, and V

0
= T[L

(A)].
Then T[S(A)] V , so
V
0
= T[S(A)] T[S(A)] V = V .
On the other hand, V
0
is a closed linear subspace in L

(B). PPP It is a linear subspace because T is a linear

operator. To see that it is closed, take any v V
0
. Then there is a sequence v
n
)
nN
in V
0
such that
|v v
n
|

2
n
for every n N. Choose u
n
L

(A) such that Tu

0
= v
0
, while Tu
n
= v
n
v
n1
and
|u
n
|

= |v
n
v
n1
|

for n 1 (using (b) above). Then

n=1
|u
n
|

n=1
|v v
n
|

+|v v
n1
|

is nite, so u = lim
n

n
i=0
u
i
is dened in the Banach space L

(A), and
Tu = lim
n

n
i=0
Tu
i
= lim
n
v
n
= v.
As v is arbitrary, V
0
is closed. QQQ Since (a) = T(a) V
0
for every a A, V V
0
and V = V
0
, as
required.
(d) If is surjective, then T is surjective, by (c-ii). If T is surjective and b B, then there is a u L

(A)
such that Tu = b. Now there is a u
t
S(A) such that |u u
t
|

1
3
, so that |Tu
t
b|

1
3
. Taking
a A such that z : u
t
(z)
1
2
= a, we must have a = b, since

b = w : (Tu
t
)(w)
1
2
=
1
[a] = a.
As b is arbitrary, is surjective.
Now (b) tells us that in this case |v|

= min|u|

: Tu = v for every v L

(B).
(e) By (c-i), T is injective i is injective. In this case, for any u L

(A),
|Tu|

= |T[u[|

(because T is a Riesz homomorphism)

sup|Tu
t
|

: u
t
S(A), u
t
[u[
= sup|u
t
|

: u
t
S(A), u
t
[u[
(by 361Jd)
= |u|

(by 354I)
|Tu|

,
and |Tu|

= |u|

.
(f ) If T is (sequentially) order-continuous then = T is (sequentially) order-continuous, by 363D.
If is (sequentially) order-continuous then : A L

(B) is (sequentially) order-continuous, so T is

(sequentially) order-continuous, by 363Eb.
(g) This is elementary, in view of the uniqueness of T

.
302 Function spaces 363G
363G Corollary Let A be a Boolean algebra.
(a) If C is a subalgebra of A, then L

(C) can be identied, as Banach lattice and as Banach algebra,

with the closed linear subspace of L

(A) generated by c : c C.
(b) If 1 is an ideal of A, then L

(A/1) can be identied, as Banach lattice and as Banach algebra, with

the quotient space L

(A)/V , where V is the closed linear subspace of L

(A) generated by a : a 1.
proof Apply 363Fc to the identity map from C to A and the canonical map from A onto A/1.
363H Representations of L

(A) Much of the importance of the concept of L

(A) arises from the

way it is naturally represented in the contexts in which the most familiar Boolean algebras appear.
Proposition Let X be a set and a -algebra of subsets of X.
(a) L

() can be identied, as Banach algebra and Banach lattice, with the space L

of bounded -
measurable real-valued functions on X, with the norm |f|

= sup
xX
[f(x)[ for f L

; this identication
matches E L

() with the characteristic function of E as a subset of X, for every E . In particular,

for any set X, L

(TX) can be identied with

(X).
(b) If A is a Dedekind -complete Boolean algebra and : A is a surjective sequentially order-
continuous Boolean homomorphism with kernel 1, then L

(A) can be identied, as Banach algebra and

Banach lattice, with L

/W, where W= f : f L

, x : f(x) ,= 0 1 is a closed ideal and solid linear

subspace of L

. For f L

,
|f

= min : 0, x : [f(x)[ > 1.

(c) In particular, if 1 is any -ideal of and E E

is the canonical homomorphism from onto

A = /1, then we have an identication of L

(A) with a quotient of L

, and for any E we can

identify (E

) L

(A) with the equivalence class (E)

/W of the characteristic function E.

proof (a) For the elementary properties of the space of -measurable functions, see 121. In particular, it
is easy to check that L

is a Riesz space, with a Riesz norm, and a normed algebra. To check that it is a
Banach space, observe that if f
n
)
nN
is a Cauchy sequence in L

, then [f
m
(x) f
n
(x)[ |f
m
f
n
|

0
as m, n , so f(x) = lim
n
f
n
(x) is dened for every x X; now f is -measurable, by 121Fa. Of
course |f|

sup
nN
|f
n
|

< , so f L

, while
|f f
n
|

sup
mn
|f
m
f
n
|

0
as n , so f = lim
n
f
n
in L

. As f
n
)
nN
is arbitrary, L

is complete.
By 361L we can identify S(), as Riesz space and normed algebra, with the linear span S of E : E ,
which is a subspace of L

. Now the point is that it is dense. PPP If f L

and > 0, then for each n Z

set E
n
= x : n f(x) < (n + 1) ; then E
n
= if [n[ > 1 +
1

|f|

, so g =

nZ
nE
n
belongs to
S, and of course |f g|

. QQQ
Consequently the canonical normed space isomorphism between S() and S extends (uniquely) to a
normed space isomorphism between L

() and L

(use 2A4I). Because the operations and are con-

tinuous in both L

() and L

, and their actions on S() and S are identied by our isomorphism, the
isomorphism between L

() and L

identies their lattice and algebra structures.

(b)(i) By 363F, we have a multiplicative Riesz homomorphism T = T

from L

() to L

(A) which is
surjective (363Fd) and has kernel the closed linear subspace W of L

() generated by E : E 1. Now
under the isomorphism described in (a), W corresponds to W. PPP () W is a linear subspace of L

because
x : (f +g)(x) ,= 0 x : f(x) ,= 0 x : g(x) ,= 0 1,
x : (f)(x) ,= 0 x : f(x) ,= 0 1
whenever f, g W and R. () If f
n
)
nN
is a sequence in W converging to f L

, then
x : f(x) ,= 0

nN
x : f
n
(x) ,= 0 1,
so f W. Thus W is a closed linear subspace of L

. () If E 1, then E, taken in S() or L

(),
corresponds to the function E : X 0, 1, which belongs to W; so that W must correspond to the closed
linear span in L

of such characteristic functions, which is a subspace of W. () On the other hand, if

f W and > 0, set
363K L

303
E
n
= x : n f(x) (n + 1), E
t
n
= x : (n + 1) f(x) n
for n 1; all these belong to 1, so g =

n=1
n(E
n
E
t
n
) W corresponds to a member of W, while
|f g|

. As W is closed, f also must correspond to some member of W. As f is arbitrary, W and W

match exactly. QQQ
(ii) Because T is a multiplicative Riesz homomorphism, L

(A)

= L

()/W is matched canonically,

in its linear, order and multiplicative structures, with L

/W. We know also that

|v|

= inf|u|

: u L

(), Tu = v
for every v L

(A) (363Fd), that is, that the norm of L

(A) corresponds to the quotient norm on

L

()/W.
As for the given formula for the norm, take any f L

. There is a g L

such that Tf = Tg and

|Tf|

= |g|

. (Here I am treating T as an operator from L

onto L

(A).) In this case

x : [f(x)[ > |Tf|

x : f(x) ,= g(x) 1.
On the other hand, if x : [f(x)[ > 1, and we set h = 1 (f 1), then Th = Tf, so |Tf|

|h|

.
(c) Put (a) and (b) together.
363I Corollary Let (X, , ) be a measure space, with measure algebra A. Then L

() can be
identied, as Banach lattice and Banach algebra, with L

(A); the identication matches (E)

()
with (E

) L

(A), for every E .

Remark The space I called L

() in Chapter 24 is not strictly speaking the space L

= L

() of 363H;
I took L

() L
0
() to be the set of essentially bounded, virtually measurable functions dened almost
everywhere on X, and in general this is larger. But, as remarked in the notes to 243, L

() can equally
well be regarded as a quotient of what I there called L

strict
, which is the L

above, because every function

in L

() is equal almost everywhere to some member of L

strict
.
363J Recovering the algebra A: Proposition Let A be a Boolean algebra. For a A write V
a
for
the solid linear subspace of L

(A) generated by a. Then a V

a
is a Boolean isomorphism between A
and the algebra of projection bands in L

(A).
proof The proof is nearly identical to that of 361K. If a A, u V
a
and v V
1\a
, then [u[ [v[ = 0
because a (1 \ a) = 0; and if w L

(A) then
w = (w a) + (w (1 \ a)) V
a
+V
1\a
because [wa[ |w|

a and [w(1 \ a)[ |w|

(1 \ a). So V
a
and V
1\a
are complementary projection
bands in L

= L

(A). Next, if U L

is a projection band, then 1 is expressible as u +v where u U,

v U

; thinking of L

as the space of continuous real-valued functions on the Stone space Z of A, u and v

must be the characteristic functions of complementary subsets E, F of Z, which must be open-and-closed,
so that E = a, F =

1 \ a. In this case V
a
U and V
1\a
U

, so U must be V
a
precisely. Thus a V
a
is
surjective. Finally, just as in 361K, a b V
a
V
b
, so we have a Boolean isomorphism.
363K Dual spaces of L

The questions treated in 362 yield nothing new in the present context. I
spell out the details.
Proposition Let A be a Boolean algebra. Let M, M

and M

be the L-spaces of bounded nitely additive

functionals, bounded countably additive functionals and completely additive functionals on A. Then the
embedding S(A) L

(A) induces Riesz space isomorphisms between S(A)

= M and L

(A)

= L

(A)

,
S(A)

= M

and L

(A)

c
, and S(A)

= M

and L

(A)

.
proof Write S = S(A), L

= L

(A).
(a) For the identications S

= M, S

= M

and S

= M

see 362A.
304 Function spaces 363K
(b) L

= L

either because L

is a Banach lattice (356Dc) or because L

has an order-unit norm,

so that a linear functional on L

is order-bounded i it is bounded on the unit ball.

(c) If f is a positive linear functional on L

, then fS is a positive linear functional. Because S is order-

dense in L

(363C), the embedding is order-continuous (352Nb); so if f is (sequentially) order-continuous,

so is fS. Accordingly the restriction operator f fS gives maps from L

to S

, (L

c
to S

c
and
L

to S

. If f L

and fS 0, then f(u

+
) 0 for every u S and therefore for every u L

, and
f 0; so all these restriction maps are injective positive linear operators.
(d) I need to show that they are surjective.
(i) If g S

, then g is bounded on the unit ball u : u S, |u|

1, so has an extension to a
continuous linear f : L

R (2A4I); thus S

= fS : f L

. This means that f fS is actually

a Riesz space isomorphism between L

and S

. In particular, [f[S = [fS[ for any f L

.
(ii) If f : L

R is a positive linear operator and fS S

c
, let u
n
)
nN
be a non-increasing sequence
in L

with inmum 0. For each n, k N there is a v

nk
S such that u
n
v
nk
u
n
+2
k
e, where e is the
standard order unit of L

(354I, as usual); set w

n
= inf
i,kn
v
ik
; then w
n
)
nN
is a non-increasing sequence
in S with inmum 0, so
0 inf
nN
f(u
n
) inf
nN
f(w
n
) = 0.
As u
n
)
nN
is arbitrary, f (L

c
. Consequently, for general f L

,
f (L

c
[f[ (L

c
[fS[ S

c
fS S

c
,
and the map f fS : (L

c
S

c
is a Riesz space isomorphism.
(iii) Similarly, if f L

is non-negative and fS S

, then whenever A L

is non-empty,
downwards-directed and has inmum 0, B = w : w S, u A, w u has inmum 0, so inf
uA
f(u)
inf
wB
f(w) 0 and f L

. As in (ii), it follows that f fS is a surjection from L

onto S

.
*363L Integration with respect to a nitely additive functional (a) If A is a Boolean algebra
and : A R is a bounded additive functional, then by 363K we have a corresponding functional f

(A)

dened by saying that f

(a) = a for every a A. There are contexts in which it is convenient,

and even helpful, to use the formula
_
ud in place of f

(u) for u L

= L

(A). When doing so, we must

of course remember that we may have lost some of the standard properties of integration. But enough of
our intuitions (including, for instance, the idea of stochastic independence) remain valid to make the formula
a guide to interesting ideas.
(b) Let M be the L-space of bounded nitely additive functionals on A (362B). Then we have a function
(u, )
_
ud : L

M R. Now this map is bilinear. PPP For , M, u, v L

and R,
_
u +v d =
_
ud +
_
v d,
_
ud =
_
ud
just because f

is linear. On the other side, we have

(f

+f

)(a) = f

(a) +f

(a) = a +a = ( +)(a) = f
+
(a)
for every a A, so that f

+ f

and f
+
must agree on S(A) and therefore on L

. But this means that

_
ud( +) =
_
ud +
_
ud. Similarly,
_
ud() =
_
ud. QQQ
(c) If is non-negative, we have
_
ud 0 whenever u 0, as in part (c) of the proof of 363K.
Consequently, for any M and u L

,
[
_
ud[ = [
_
u
+
d
+

_
u

d
+

_
u
+
d

+
_
u

_
u
+
d
+
+
_
u

d
+
+
_
u
+
d

+
_
u

=
_
[u[d[[
_
|u|

1 d[[ = |u|

[[(1) = |u|

||.
363Lf L

305
So (u, )
_
ud has norm (as dened in 253Ab) at most 1. If A ,= 0, the norm is exactly 1. (For this we
need to know that there is a M
+
such that 1 = 1. Take any z in the Stone space of A and set a = 1
if z a, 0 otherwise.)
(d) We do not have any result corresponding to B.Levis theorem in this language, because (even if is
non-negative and countably additive) there is no reason to suppose that sup
nN
u
n
is dened in L

just
because sup
nN
_
u
n
d is nite. But if is countably additive and A is Dedekind -complete, we have
something corresponding to Lebesgues Dominated Convergence Theorem (363Yh).
(e) One formula which we can imitate in the present context is that of 252O, where the ordinary integral
is represented in the form
_
fd =
_

0
x : f(x) tdt.
In the context of general Boolean algebras, we cannot directly represent the set [[f t]] = x : f(x) t
(though in the next section I will show that in Dedekind -complete Boolean algebras there is an eective
expression of this idea, and I will use it in the principal denition of 365). But what we can say is the
following. If A is any Boolean algebra, and : A [0, [ is a non-negative additive functional, and
u L

(A)
+
, then
_
ud =
_

0
supa : ta udt,
where the right-hand integral is taken with respect to Lebesgue measure. PPP (i) For t 0 set h(t) =
supa : ta u. Then h is non-increasing and zero for t > |u|

, so
_

0
h(t)dt is dened in R. If
we set h
n
(t) = h(2
n
(k + 1)) whenever k, n N and 2
n
k t < 2
n
(k + 1), then h
n
(t))
nN
is a non-
decreasing sequence which converges to h(t) whenever h is continuous at t, which is almost everywhere
(222A, or otherwise); so
_

0
h(t)dt = lim
n
_

0
h
n
(t)dt. Next, given n N and > 0, we can choose for
each k k

= 2
n
|u|

| an a
k
such that 2
n
(k + 1)a
k
u and a
k
h(2
n
(k + 1)) . In this case

k=0
2
n
a
k
u, so
_

0
h
n
(t)dt = 2
n
k

k=0
h(2
n
(k + 1)) |u|

+ 2
n
k

k=0
a
k
= |u|

+
_
k

k=0
2
n
a
k
d |u|

+
_
ud.
As n and are arbitrary,
_

0
h(t)dt
_
ud. (ii) In the other direction, there is for any > 0 a v S(A)
such that v u v + 1. If we express v as

m
j=0

j
c
j
where c
0
. . . c
m
and
j
0 for every j
(361Ec), then we shall have h(t) c
k
whenever t

k
j=0

j
, so
_

0
h(t)dt

m
k=0

k
c
k
=
_
v d
_
ud 1.
As is arbitrary,
_

0
h(t)dt
_
ud and the two integrals are equal. QQQ
(f ) The formula
_
d is especially natural when A is an algebra of sets, so that L

can be directly
interpreted as a space of functions (363Yf); better still, when A is actually a -algebra of subsets of a set
X, L

can be identied with the space of bounded A-measurable functions on X, as in 363Ha. So in such
contexts I may write
_
g d when g : X R is bounded and A-measurable, and : A R is an additive
functional. But I will try to take care to signal any such deviation from the normal principle that the
symbol
_
refers to the sequentially order-continuous integral dened in 122 with the minor modications
introduced in 133 and 135. My purpose in this paragraph has been only to indicate something of what
can be done with nite additivity alone.
306 Function spaces 363M
363M Now I come to a fundamental fact underlying a number of theorems in both this volume and
the last.
Theorem Let A be a Boolean algebra.
(a) A is Dedekind -complete i L

(A) is Dedekind -complete.

(b) A is Dedekind complete i L

(A) is Dedekind complete.

proof (a)(i) Suppose that A is Dedekind -complete. By 314M, we may identify A with a quotient //,
where / is the ideal of meager subsets of the Stone space Z of A, and = EA : E c, A /,
writing c = a : a A for the algebra of open-and-closed subsets of Z. By 363H, L

= L

(A) can be
identied with L

/V, where L

is the space of bounded -measurable functions from Z to R, and V is the

space of functions zero except on a member of 1.
Now suppose that u
n
)
nN
is a sequence in L

with an upper bound u L

. Express u
n
, u as f

n
, f

where f
n
, f L

. Set g(z) = sup

nN
min(f
n
(z), f(z)) for every z Z; then g L

(121F), so we have a
corresponding member v = g

of L

. For each n N, u u
n
so (f
n
f)
+
V,
z : f
n
(z) > g(z) z : f
n
(z) > f(z) /
and v u
n
. If w L

and w u
n
for every n, then express w as h

where h L

; we have (f
n
h)
+
V
for every n, so
z : g(z) > h(z)

nN
z : f
n
(z) > h(z) /
because / is a -ideal, and (g h)
+
V, w v. Thus v = sup
nN
u
n
in L

. As u
n
)
nN
is arbitrary, L

is Dedekind -complete (using 353G).

(ii) Now suppose that L

is Dedekind -complete, and that A is a countable non-empty set in A.

In this case a : a A has a least upper bound u in L

. Take a v S(A) such that 0 v u and

|u v|

1
3
; set b = [[v >
1
3
]], as dened in 361Eg. If a A, then |(a v)
+
|

|u v|

1
3
, so
2
3
a v and a b. If c A is any upper bound for A, then v u c so b c. Thus b = supA in A. As
A is arbitrary, A is Dedekind -complete.
(b)(i) For the second half of this theorem I use an argument which depends on joining the representation
described in (a-i) above with the original denition of L

in 363A. The point is that C(Z) L

, and for
any f C(Z) = L

(A) its equivalence class f

in L

/V corresponds to f itself. PPP Perhaps it will help

to give a name T to the canonical isomorphism from L

/V to L

. Then V = f : Tf

= f is a closed
linear subspace of C(Z), because f f

and T are continuous linear operators. But if a A, then (a)

, the
equivalence class of a in //, corresponds to a (see the proof of 314M), so (a)

/V corresponds
to a; that is, T(a)

= a, if we identify a L

with a : Z 0, 1. So V contains a for every a A;

because V is a linear subspace, S(A) V ; because V is closed, L

V . QQQ
For a general f L

, g = Tf

must be the unique member of C(Z) such that g

= f

, that is, such

that z : g(z) ,= f(z) is meager.
(ii) Suppose now that A is actually Dedekind complete. In this case Z is extremally disconnected
(314S). Consequently every open set belongs to . PPP If G is open, then G is open-and-closed; but A = GG
is a closed set with empty interior, so is meager, and G = GA . QQQ
Let A L

= C(Z) be any non-empty set with an upper bound in C(Z). For each z Z set
g(z) = sup
uA
u(z). Then
G

= z : g(z) > =

uA
z : u(z) >
is open for every R (that is, g is lower semi-continuous). Thus G

for every , so g L

, and
v = Tg

is dened in C(Z). For any u A, g u in L

, so
v = Tg

Tu

= u
in L

; thus v is an upper bound for A in L

. On the other hand, if w is any upper bound for A in

L

= C(Z), then surely w(z) u(z) for every z Z, u A, so w g and

w = Tw

Tg

= v.
This means that v is the least upper bound of A. As A is arbitrary, L

is Dedekind complete.
363Q L

307
(iii) Finally, if L

is Dedekind complete, then the argument of (b-ii), applied to arbitrary non-empty

subsets A of A, shows that A is also Dedekind complete.
363N Much of the importance of L

spaces in the theory of Riesz spaces arises from the next result.
Proposition Let U be a Dedekind -complete Riesz space with an order unit. Then U is isomorphic, as
Riesz space, to L

(A), where A is the algebra of projection bands in U.

proof (a) By 353M, U is isomorphic to a norm-dense Riesz subspace of C(X) for some compact Hausdor
space X; for the rest of this argument, therefore, we may suppose that U actually is such a subspace.
(b) Now U = C(X). PPP If g C(X) then by 354I there are sequences f
n
)
nN
, f
t
n
)
nN
in U such that
f
n
g g
n
and |g
n
f
n
|

2
n
for every n. Now f
n
: n N has a least upper bound f in U; since
we must have f
n
f g
n
for every n, f = g and g U. QQQ
(c) Next, X is zero-dimensional. PPP Suppose that G X is open and x G. Then there is an open set
G
1
such that x G
1
G
1
G (3A3Bc). There is an f C(X) such that 0 f G
1
and f(x) > 0 (also
by 3A3Bc); write H for y : f(y) > 0. Set g = sup
nN
(nf X), the supremum being taken in U = C(X).
For each y H, we must have g(y) min(1, nf(y)) for every n, so that g(y) = 1. On the other hand, if
y X H, there is an h C(X) such that h(y) > 0 and 0 h (X H); now h f = 0 so h g = 0 and
g(y) = 0. Thus H g H. The set y : g(y) 0, 1 is closed and includes H (X H) so must be
the whole of X; thus G
2
= y : g(y) >
1
2
= y : g(y)
1
2
is open-and-closed, and we have
x H G
2
H G
1
G.
As x, G are arbitrary, the set of open-and-closed subsets of X is a base for the topology of X, and X is
zero-dimensional. QQQ
(d) We can therefore identify X with the Stone space of its algebra c of open-and-closed sets (311J).
But in this case 363A immediately identies U = C(X) with L

(c). By 363J, c is isomorphic to A, so

U

= L

(A).
Remark Note that in part (c) of the argument above, we have to take great care over the interpretation
of sup. In the space of all real-valued functions on X, the supremum of nf X : n N is just H.
But g is supposed to be the least continuous function greater than or equal to nf X for every n, and is
therefore likely to be strictly greater than H, even though sandwiched between H and H.
363O Corollary Let U be a Dedekind -complete M-space. Then U is isomorphic, as Banach lattice,
to L

(A), where A is the algebra of projection bands of U.

proof This is merely the special case of 363N in which U is known from the start to be complete under an
order-unit norm.
363P Corollary Let U be any Dedekind -complete Riesz space and e U
+
. Then the solid linear
subspace U
e
of U generated by e is isomorphic, as Riesz space, to L

(A) for some Dedekind -complete

Boolean algebra A; and if U is Dedekind complete, so is A.
proof Because U is Dedekind -complete, so is U
e
(353J(a-i)). Apply 363N to U
e
to see that U
e

= L

(A)
for some A. Because U
e
is Dedekind -complete, so is A, by 363Ma; while if U is Dedekind complete, so are
U
e
and A, by 353J(b-i) and 363Mb.
363Q The next theorem will be a striking characterization of the Dedekind complete L

spaces as
normed spaces. As a warming-up exercise I give a much simpler result concerning their nature as Banach
lattices.
Proposition Let A be a Dedekind complete Boolean algebra. Then for any Banach lattice U, a linear
operator T : U L

= L

(A) is continuous i it is order-bounded, and in this case |T| = |[T[|, where

the modulus [T[ is taken in L

(U; L

).
308 Function spaces 363Q
proof It is generally true that order-bounded operators between Banach lattices are continuous (355C). If
T : U L

is continuous, then for any w U

+
[u[ w = |u| |w| = |Tu|

|T||w| = [Tu[ |T||w|1.

So T is order-bounded. As L

is Dedekind complete (363Mb), [T[ is dened in L

(U; L

) (355Ea). For
any w U,
[T[[w[ = sup[Tu[ : [u[ [w[ |T||w|1,
so |[T[(w)| |T||w|; accordingly |[T[| |T|. On the other hand, of course,
[Tw[ [T[[w[ |[T[||w|1
for every w U, so |T| |[T[| and the two norms are equal.
Remark Of course what is happening here is that the spaces L

(A), for Dedekind complete A, are just

the Dedekind complete M-spaces; this is an elementary consequence of 363N and 363M.
363R Now for something much deeper.
Theorem Let U be a normed space over R. Then the following are equiveridical:
(i) there is a Dedekind complete Boolean algebra A such that U is isomorphic, as normed space, to
L

(A);
(ii) whenever V is a normed space, V
0
a linear subspace of V , and T
0
: V
0
U is a bounded linear
operator, there is an extension of T
0
to a bounded linear operator T : V U with |T| = |T
0
|.
proof For the purposes of the argument below, let us say that a normed space U satisfying the condition
(ii) has the Hahn-Banach property.
Part A: (i)(ii) I have to show that L

(A) has the Hahn-Banach property for every Dedekind complete

Boolean algebra A. Let V be a normed space, V
0
a linear subspace of V , and T
0
: V
0
L

= L

(A) a
bounded linear operator. Set = |T
0
|.
Let P be the set of all functions T such that domT is a linear subspace of V including V
0
and T :
domT U is a bounded linear operator extending T
0
and with norm at most . Order P by saying that
T
1
T
2
if T
2
extends T
1
. Then any non-empty totally ordered subset Q of P has an upper bound in P. PPP
Set domT =

domT
1
: T
1
Q, Tv = T
1
v whenever T
1
Q and v domT
1
; it is elementary to check
that T P, so that T is an upper bound for Q in P. QQQ
By Zorns Lemma, P has a maximal element

T. Now dom

T = V . PPP??? Suppose, if possible, otherwise.
Write

V = dom

T and take any v V

V ; let V
1
be the linear span of

V v, that is, v+ v : v

V , R.
If v
1
, v
2


V then, writing e for the standard order unit of L

Tv
1
+

Tv
2
=

T(v
1
+v
2
) |

T(v
1
+v
2
)|

e
|v
1
+v
2
|e |v
1
v|e +|v
2
+ v|e,
so

Tv
1
|v
1
v|e |v
2
+ v|e

Tv
2
.
Because L

is Dedekind complete (363Mb),

u = sup
v
1

Tv
1
|v
1
v|e
is dened in L

and u |v
2
+ v|e Tv
2
for every v
2


V . Putting these together, we have

Tv u |v v|e,

Tv + u |v + v|e
for all v

V . Consequently, if v

V , then for > 0

Tv + u = (

T(
1

v) + u) |
1

v + v|e = |v + v|e,
while for < 0

Tv + u = [[(

T(
1

v) u) [[|
1

v v|e = |v + v|e,
and of course
363R L

309

Tv |

Tv|

e |v|e.
So we have

Tv + u |v + v|e
for every v

V , R.
Dene T
1
: V
1
L

by setting T
1
(v + v) =

Tv + u for every v

V , R. (This is well-dened
because v /

V , so any member of V
1
is uniquely expressible as v + v where v

V and R.) Then T
1
is a linear operator, extending T
0
, from a linear subspace of V to L

. But from the calculations above we

know that T
1
v |v|e for every v V
1
; since we also have
T
1
v = T
1
(v) | v|e = |v|e,
|T
1
v|

|v| for every v V

1
, and T
1
P. But now T
1
is a member of P properly extending

T, which
is supposed to be impossible. XXXQQQ
Accordingly

T : V L

is an extension of T
0
to the whole of V , with the same norm as T
0
. As V and
T
0
are arbitrary, L

has the Hahn-Banach property.

Part B: (ii)(i) Now let U be a normed space with the Hahn-Banach property. If U = 0 then of
course it is isomorphic to L

(A), where A = 0, so henceforth I will take it for granted that U ,= 0.

(a) Let Z be the unit ball of the dual U

of U, with the weak* topology. Then Z is a compact Hausdor

space (3A5F). For u U set Z
u
= f : f Z, [f(u)[ = |u|; then Z
u
is a closed subset of Z (because
f f(u) is continuous), and is non-empty, by the Hahn-Banach theorem (3A5Ab, or Part A above!) Now
let P be the set of those closed sets X Z such that X Z
u
,= for every u U. If Q P is non-empty
and totally ordered, then

Q P, because for any u U

X Z
u
: X Q
is a downwards-directed family of non-empty compact sets, so must have non-empty intersection. By Zorns
Lemma, upside down, P has a minimal element X; with its relative topology, X is a compact Hausdor
space.
(b) We have a linear operator R : U C(X) given by setting (Ru)(x) = x(u) for every u U, x X;
because X Z, |Ru|

|u|, and because X P, |Ru|

= |u|, for every u U. Moreover, if G X is

a non-empty open set (in the relative topology of X) then XG cannot belong to P, because X is minimal,
so there is a (non-zero) u U such that [x(u)[ < |u| for every x X G. Replacing u by |u|
1
u if need
be, we may suppose that |u| = 1.
What this means is that W = R[U] is a linear subspace of C(X) which is isomorphic, as normed space,
to U, and has the property that whenever G X is a non-empty relatively open set there is an f W such
that |f|

= 1 and [f(x)[ < 1 for every x X G. Observe that, because X G is compact, there is now
some < 1 such that [f(u)[ for every f X G.
Because W is isomorphic to U, it has the Hahn-Banach property.
(c) Now consider V =

(X), V
0
= W, T
0
: V
0
W the identity map. Because W has the Hahn-Banach
property, there is a linear operator T :

(X) W, extending T
0
, and of norm |T
0
| = 1.
(d) If h

(X) and x
0
X x : h(x) ,= 0, then (Th)(x
0
) = 0. PPP??? Otherwise, set G = y : y
X x : h(x) ,= 0, (Th)(y) ,= 0. This is a non-empty open set in X, so there are f W, < 1 such that
|f|

= 1, [f(x)[ for every x X G.

Because |f|

= 1, there must be some x

1
X such that [f(x
1
)[ = 1, and of course x
1
G, so
that (Th)(x
1
) ,= 0. But let > 0 be such that |h|

1 . Then, because h(x) = 0 for x G,

[f(x)[ +[h(x)[ 1 for every x X, and |f +h|

, |f h|

are both less than or equal to 1. As Tf = f

and |T| = 1, this means that
|f Th|

1, |f +Th|

1;
consequently
[f(x
1
)[ +[(Th)(x
1
)[ = max([(f +Th)(x
1
)[, [(f Th)(x
1
)[) 1.
But [f(x
1
)[ = 1 and (Th)(x
1
) ,= 0, so this is impossible. XXXQQQ
310 Function spaces 363R
(e) It follows that Th = h for every h C(X). PPP??? Suppose, if possible, otherwise. Then there is a > 0
such that G = x : [(Th)(x) h(x)[ > is not empty. Let f W be such that |f| = 1 but [f(x)[ < 1
for every x X G. Then there is an x
0
X such that [f(x
0
)[ = 1; of course x
0
must belong to G. Set
f
1
=
h(x
0
)
f(x
0
)
f, so that f
1
W and f
1
(x
0
) = h(x
0
). Set
h
1
(x) = max(h(x) , min(h(x) +, f
1
(x)))
for x X. Then h
1
C(X). Setting
H = x : [h(x) h(x
0
)[ +[f
1
(x) f
1
(x
0
)[ < ,
H is an open set containing x
0
and
[f
1
(x) h(x)[ [f
1
(x
0
) h(x
0
)[ + = , h
1
(x) = f
1
(x)
for every x H. Consequently x
0
/ x : (f
1
h
1
)(x) ,= 0, and T(f
1
h
1
)(x
0
) = 0, by (d). But this means
that
(Th
1
)(x
0
) = (Tf
1
)(x
0
) = f
1
(x
0
) = h(x
0
),
so that
[h(x
0
) (Th)(x
0
)[ = [T(h
1
h)(x
0
)[ |T(h
1
h)|

|h
1
h|

,
which is impossible, because x
0
G. XXXQQQ
(f ) This tells us at once that W = C(X). But (d) also tells us that X is extremally disconnected. PPP Let
G X be any open set. Then X = G+(X G), so
X = T(X) = h
1
+h
2
,
where h
1
= T(G), h
2
= T((X G)). Now from (d) we see that h
1
must be zero on X G while h
2
must
be zero on G. Thus we have h
1
(x) = 1 for x G; as h
1
is continuous, h
1
(x) = 1 for x G, and h
1
= G.
Of course it follows that G is open. As G is arbitrary, X is extremally disconnected. QQQ
(g) Being also compact and Hausdor, therefore regular (3A3Bc), X is zero-dimensional (3A3Bd). We
may therefore identify X with the Stone space of its regular open algebra A (314S), and W = C(X) with
L

(A). Thus R : U C(X) is a Banach space isomorphism between U and C(X)

= L

(A); so U is of the
type declared.
363S The Banach-Ulam problem At a couple of points already (232Hc, the notes to 326) I have
remarked on a problem which was early recognised as a fundamental question in abstract measure theory.
I now set out some formulations of the problem which arise naturally from the work done so far. I will do
this by writing down a list of statements which are equiveridical in the sense that if one of them is true, so
are all the others; the Banach-Ulam problem asks whether they are indeed true.
I should remark that this is not generally counted as an open problem. It is in fact believed by most of
us that these statements are independent of the usual axioms of Zermelo-Fraenkel set theory, including the
axiom of choice and even the continuum hypothesis. As such, this problem belongs to Volume 5 rather than
anywhere earlier, but its manifestations will become steadily more obtrusive as we continue through this
volume and the next, and I think it will be helpful to begin collecting them now. The ideas needed to show
that the statements here imply each other are already accessible; in particular, they involve no set theory
beyond Zorns Lemma. These implications constitute the following theorem, derived from Luxemburg 67a.
Theorem The following statements are equiveridical.
(i) There are a set X and a probability measure , with domain TX, such that x = 0 for every x X.
(ii) There are a localizable measure space (X, , ) and an absolutely continuous countably additive
functional : R which is not truly continuous, so has no Radon-Nikod ym derivative (denitions:
232Ab, 232Hf).
(iii) There are a Dedekind complete Boolean algebra A and a countably additive functional : A R
which is not completely additive.
(iv) There is a Dedekind complete Riesz space U such that U

c
,= U

.
363Xd L

311
proof (a)(i)(ii) Let X be a set with a probability measure , dened on TX, such that x = 0 for
every x X. Let be counting measure on X. Then (X, TX, ) is strictly localizable, and : TX R
is countably additive; also E = 0 whenever E is nite, so is absolutely continuous with respect to .
But if E < then E is nite and (X E) = 1, so is not truly continuous, and has no Radon-Nikod ym
derivative (232D).
(b)(ii)(iii) Let (X, , ) be a localizable measure space and : R an absolutely continuous
countably additive functional which is not truly continuous. Let (A, ) be the measure algebra of ; then
we have an absolutely continuous countably additive functional : A R dened by setting E

= E
for every E (327C). Since is not truly continuous, is not completely additive (327Ce). Also A is
Dedekind complete, because is localizable, so A and witness (iii).
(c)(iii)(i) Let A be a Dedekind complete Boolean algebra and : A R a countably additive
functional which is not completely additive. Because is bounded (326I), therefore expressible as the
dierence of non-negative countably additive functionals (326H), there must be a non-negative countably
additive functional
t
on A which is not completely additive.
By 326N, there is a partition of unity a
i
)
iI
in A such that

iI

t
a
i
<
t
1. Set K = i : i I,
t
a
i
> 0;
then K must be countable, so

t
(sup
iI\K
a
i
) =
t
1
t
(sup
iK
a
i
) =
t
1

iK

t
a
i
> 0.
For J I set J =
t
(sup
iJ\K
a
i
); the supremum is always dened because A is Dedekind complete.
Because
t
is countably additive and non-negative, so is ; because
t
a
i
= 0 for i J K, i = 0 for
every i I. Multiplying by a suitable scalar, if need be, (I, TI, ) witnesses that (i) is true.
(d)(iii)(iv) If A is a Dedekind complete Boolean algebra with a countably additive functional which
is not completely additive, then U = L

(A) is a Dedekind complete Riesz space (363Mb) and U

c
,= U

,
by 363K (recalling, as in (c) above, that the functional must be bounded).
(e)(iv)(iii) Let U be a Dedekind complete Riesz space such that U

,= U

c
. Take f U

c
U

;
replacing f by [f[ if need be, we may suppose that f 0 is sequentially order-continuous but not order-
continuous (355H, 355I). Let A be a non-empty downwards-directed set in U, with inmum 0, such that
inf
uA
f(u) > 0 (351Ga). Take e A, and consider the solid linear subspace U
e
of U generated by e; write
g for the restriction of f to U
e
. Because the embedding of U
e
in U is order-continuous, g (U
e
)

c
; because
A U
e
is downwards-directed and has inmum 0, and
inf
uAU
e
g(u) = inf
uA
f(u) > 0,
g / U

e
. But U
e
is a Riesz space with order unit e, and is Dedekind complete because U is; so it can be
identied with L

(A) for some Boolean algebra A (363N), and A is Dedekind complete, by 363M.
Accordingly we have a Dedekind complete Boolean algebra A such that L

(A)

c
,= L

(A)

. By 363K,
there is a (bounded) countably additive functional on A which is not completely additive, and (iii) is true.
363X Basic exercises (a) Let A be a Boolean algebra and U a Banach algebra. Let : A U be a
bounded additive function and T : L

(A) U the corresponding bounded linear operator. Show that T is

multiplicative i (a b) = a b for all a, b A.
>>>(b) Let A, B be Boolean algebras and T : L

(A) L

(A) a linear operator. Show that the following

are equiveridical: (i) there is a Boolean homomorphism : A A such that T = T

(ii) T(uv) = TuTv

for all u, v L

(A) (iii) T is a Riesz homomorphism and Te

A
= e
B
, where e
A
is the standard order unit of
L

(A).
(c) Let A, B be Boolean algebras and T : L

(A) L

(B) a Riesz homomorphism. Show that there are

a Boolean homomorphism : A B and a v 0 in L

(B) such that Tu = v T

u for every u L

(A),
where T

is the operator associated with (363F).

(d) Let A be a Boolean algebra and C a subalgebra of A. Show that L

(C), regarded as a subspace of

L

(A) (363Ga), is order-dense in L

(A) i C is order-dense in A.
312 Function spaces 363Xe
>>>(e) Let (X, , ) be a measure space with measure algebra A, and L

the space of bounded -

measurable real-valued functions on X. A linear lifting of is a positive linear operator T : L

(A) L

such that T(1

A
) = X and (Tu)

= u for every u L

(A), writing f f

for the canonical map from

L

to L

(A) (363H-363I). (i) Show that if : A is a lifting in the sense of 341A then T

, as dened in
363F, is a linear lifting. (ii) Show that if T : L

(A) L

is a linear lifting, then there is a corresponding

lower density : A dened by setting a = x : T(a)(x) = 1 for each a A. (iii) Show that , as
dened in (ii), is a lifting i T is a Riesz homomorphism.
(f ) Let U be any commutative ring with multiplicative identity 1. Show that the set A of idempotents
in U (that is, elements a U such that a
2
= a) is a Boolean algebra with identity 1, writing a b = ab,
1 \ a = 1 a for a, b A.
(g) Let A be a Boolean algebra. Show that A is isomorphic to the Boolean algebras of idempotents of
S(A) and L

(A).
(h) Let A be a Dedekind -complete Boolean algebra. (i) Show that for any u L

(A), R there
are elements [[u ]], [[u > ]] A, where [[u ]] is the largest a A such that u a a, and
[[u > ]] = sup
>
[[u ]]. (ii) Show that in the context of 363H, if u corresponds to f

for f L

, then
[[u ]] = x : f(x)

, [[u > ]] = x : f(x) >

. (iii) Show that if A L

is non-empty and v L

,
then v = supA i [[v > ]] = sup
uA
[[u > ]] for every R; in particular, v = u i [[v > ]] = [[u > ]] for
every . (iv) Show that a function : R A is of the form () = [[u > ]] i () () = sup
>
() for
every R () there is an M such that (M) = 0, (M) = 1. (v) Put (iii) and (iv) together to give a
proof that L

is Dedekind -complete if A is.

(i) Let A be a Dedekind -complete Boolean algebra and U L

(A) a (sequentially) order-closed Riesz

subspace containing 1. Show that U can be identied with L

(B) for some (sequentially) order-closed

subalgebra B A. (Hint: set B = b : b U and use 363N.)
(j) For a Boolean algebra A, with Stone space Z, write L

C
(A) for the Banach algebra C(Z; C) of continu-
ous complex-valued functions on Z. Prove results corresponding to 363C, 363Ea, 363F-363I for the complex
case.
363Y Further exercises (a) Let A be a Boolean algebra. Given the linear structure, ordering, multi-
plication and norm of S(A) as described in 361, show that a norm completion of S(A) will serve for L

(A)
in the sense that all the results of 363B-363Q can be proved with no use of the axiom of choice except an
occasional appeal to countably many choices in sequential forms of the theorems.
(b) Let A be a Boolean algebra. Show that A is ccc i L

(A) has the countable sup property (241Yd,

353Ye).
(c) Let X be an extremally disconnected topological space, and G its regular open algebra. Show that
there is a natural isomorphism between L

(G) and C
b
(X).
(d) Let X be a compact Hausdor space. Let us say that a linear subspace U of C(X) is

-comple-
mented in C(X) if there is a linear subspace V such that C(X) = U V and |u+v|

= max(|u|

, |v|

)
for all u U, v V . Show that there is a one-to-one correspondence between such subspaces U and open-
and-closed subsets E of X, given by setting U = f : f C(X), f(x) = 0 x X E. Hence show that
if A is any Boolean algebra, there is a canonical isomorphism between A and the partially ordered set of

-complemented subspaces of L

(A).
(e) Let A be a Boolean algebra. (i) If u L

= L

(A), show that [u[ = e, the standard order unit of

L

, i max(|u +v|

, |u v|

) > 1 whenever v L

0. (ii) Show that if u, v L

then [u[ [v[ = 0

i |u + v + w|

max(|u + w|

, |v + w|

) whenever = 1 and w L

. (iii) Show that if

T : L

is a normed space automorphism then there are a Boolean automorphism : A A and a

w L

such that [w[ = e and Tu = w T

u for every u L

.
363 Notes L

313
(f ) Let X be a set, an algebra of subsets of X, and 1 an ideal in . (i) Show that L

() can be
identied, as Banach lattice and Banach algebra, with the space L

of bounded functions f : X R such

that whenever < in R there is an E such that x : f(x) E x : f(x) . (ii) Show that
L

= g : g C(Z), where Z is the Stone space of and : X Z is a function (to be described). (iii)
Show that L

(/1) can be identied, as Banach lattice and Banach algebra, with L

/V, where V is the

set of those functions f L

such that for every > 0 there is a member of 1 including x : [f(x)[ .

(g) Let (X, , ) be a complete probability space with measure algebra A. Let B
n
)
nN
be a non-
decreasing sequence of closed subalgebras of A such that A is the closed subalgebra of itself generated by

nN
B
n
, and set
n
= F : F

B
n
for each n. Let P
n
: L
1
() L
1
(
n
) be the conditional
expectation operator for each n, so that P
n
L

() is a positive linear operator from L

()

= L

(A) to
L

(
n
)

= L

(B
n
). Suppose that we are given for each n a lifting
n
: B
n

n
and that
n+1
b =
n
b
whenever n N, b B
n
. Let T
n
: L

(B
n
) L

be the corresponding linear liftings (363Xe), and T any

non-principal ultralter on N. (i) Show that for any u L

(A), T
n
P
n
u)
nN
converges almost everywhere.
(ii) For u L

(A) set (Tu)(x) = lim

nJ
(T
n
P
n
u)(x) for x X, u L

(A). Show that T is a linear lifting

of . (iii) Use 363Xe(ii) and 341J to show that there is a lifting of extending every
n
. (iv) Use this as
the countable-conality inductive step in a proof of the Lifting Theorem (using partial liftings rather than
partial lower densities, as suggested in 341Li).
(h) Let A be a Dedekind -complete Boolean algebra and : A R a countably additive func-
tional. Suppose that u
n
)
nN
is an order-bounded sequence in L

(A) such that inf

nN
sup
mn
u
m
and
sup
nN
inf
mn
u
m
are equal to u say. Show that
_
ud = lim
n
_
u
n
d.
(i) Let be the family of those sets E [0, 1] such that (int E) = E, where is Lebesgue measure.
(i) Show that is an algebra of subsets of [0, 1] and that every member of is Lebesgue measurable. (ii)
Show that if we identify L

() with a set of real-valued functions on [0, 1], as in 363Yf, then we get just
the space of Riemann integrable functions. (iii) Show that if we write for , then
_
d, as dened in
363L, is just the Riemann integral.
(j) Show that a normed space over C has the Hahn-Banach property of 363R for complex spaces i it is
isomorphic to L

C
(A), as described in 363Xj, for some Dedekind complete Boolean algebra A.
363 Notes and comments As with S(A), I have chosen a denition of L

(A) in terms of the Stone space

of A; but as with S(A), this is optional (363Ya). By and large the basic properties of L

are derived very

naturally from those of S. The spaces L

(A), for general Boolean algebras A, are not in fact particularly

important; they have too few properties not shared by all the spaces C(X) for compact Hausdor X. The
point at which it becomes helpful to interpret C(X) as L

(A) is when C(X) is Dedekind -complete. The

spaces X for which this is true are dicult to picture, and alternative representations of L

along the lines

of 363H-363I can be easier on the imagination.
For Dedekind -complete A, there is an alternative description of members of L

(A) in terms of objects

[[u > ]] (363Xh); I will return to this idea in the next section. For the moment I remark only that it gives
an alternative approach to 363M not necessarily depending on the representation of L

as a quotient L

/V
nor on an analysis of a Stone space. I used a version of such an argument in the proof of 363M which I gave
in Fremlin 74a, 43D.
I spend so much time on 363M not only because Dedekind completeness is one of the basic properties of
any lattice, but because it oers an abstract expression of one of the central results of Chapter 24. In 243H
I showed that L

() is always Dedekind -complete, and that it is Dedekind complete if is localizable.

We can now relate this to the results of 321H and 322Be: the measure algebra of any measure is Dedekind
-complete, and the measure algebra of a localizable measure is Dedekind complete.
The ideas of the proof of 363M can of course be rearranged in various ways. One uses 353Yb: for
completely regular spaces X, C(X) is Dedekind complete i X is extremally disconnected; while for compact
Hausdor spaces, X is extremally disconnected i it is the Stone space of a Dedekind complete algebra.
With the right modication of the concept extremally disconnected (314Yf), the same approach works for
Dedekind -completeness.
314 Function spaces 363 Notes
363R is the Nachbin-Kelley theorem; it is commonly phrased a normed space U has the Hahn-Banach
extension property i it is isomorphic, as normed space, to C(X) for some compact extremally disconnected
Hausdor space X, but the expression in terms of L

spaces seems natural in the present context. The

implication in one direction (Part A of the proof) calls for nothing but a check through one of the standard
proofs of the Hahn-Banach theorem to make sure that the argument applies in the generalized form. Part
B of the proof has ideas in it; I have tried to set it out in a way suggesting that if you can remember the
construction of the set X the rest is just a matter of a little ingenuity.
One way of trying to understand the multiple structures of L

spaces is by looking at the corresponding

automorphisms. We observe, for instance, that an operator T from L

(A) to itself is a Banach algebra

automorphism i it is a Banach lattice automorphism preserving the standard order unit i it corresponds
to an automorphism of the algebra A (363Xb). Of course there are Banach space automorphisms of L

which do not respect the order or multiplicative structure; but they have to be closely related to algebra
isomorphisms (363Ye).
I devote a couple of exercises (363Xe, 363Yg) to indications of how the ideas here are relevant to the
Lifting Theorem. If you found the formulae of the proof of 341G obscure it may help to work through the
parallel argument.
A lecture by W.A.J.Luxemburg on the equivalence between (i) and (iv) in 363S was one of the turning
points in my mathematical apprenticeship. I introduce it here, even though the real importance of the
Banach-Ulam problem lies in the metamathematical ideas it has nourished, because these formulations pro-
vide a focus for questions which arise naturally in this volume and which otherwise might prove distracting.
The next group of signicant ideas in this context will appear in 438.
364 L
0
My next objective is to develop an abstract construction corresponding to the L
0
() spaces of 241. These
generalized L
0
spaces will form the basis of the work of the rest of this chapter and also the next; partly
because their own properties are remarkable, but even more because they form a framework for the study
of Archimedean Riesz spaces in general (see 368). There seem to be signicant new diculties, and I take
the space to describe an approach which can be made essentially independent of the route through Stone
spaces used in the last three sections. I embark directly on a denition in the new language (364A), and
relate it to the constructions of 241 (364C-364E, 364J) and 361-363 (364K). The ideas of Chapter 27
can also be expressed in this language; I make a start on developing the machinery for this in 364G, with
the formula [[u E]], the region in which u belongs to E, and some exercises (364Xd-364Xf). Following
through the questions addressed in 363, I discuss Dedekind completeness in L
0
(364M-364O), properties of
its multiplication (364P), the expression of the original algebra in terms of L
0
(364Q), the action of Boolean
homomorphisms on L
0
(364R) and product spaces (364S). In 364T-364W I describe representations of the
L
0
space of a regular open algebra.
364A Denition Let A be a Dedekind -complete Boolean algebra. I will write L
0
(A) for the set of
all functions [[u > ]] : R A such that
() [[u > ]] = sup
>
[[u > ]] in A for every R,
() inf
R
[[u > ]] = 0,
() sup
R
[[u > ]] = 1.
364B Remarks (a) My reasons for using the notation [[u > ]] rather than u() will I hope become
clear in the next few paragraphs. For the moment, if you think of A as a -algebra of sets and of L
0
(A)
as the family of A-measurable real-valued functions, then [[u > ]] corresponds to the set x : u(x) >
(364Ja).
(b) Some readers will recognise the formula [[. . . ]] as belonging to the language of forcing, so that [[u > ]]
could be read as the Boolean value of the proposition u > . But a vocalisation closer to my intention
might be the region where u > .
364C L
0
315
(c) Note that condition () of 364A automatically ensures that [[u > ]] [[u >
t
]] whenever
t
in
R.
(d) In fact it will sometimes be convenient to note that the conditions of 364A can be replaced by
(
t
) [[u > ]] = sup
qQ,q>
[[u > q]] in A for every R,
(
t
) inf
nN
[[u > n]] = 0,
(
t
) sup
nN
[[u > n]] = 1;
the point being that we need look only at suprema and inma of countable subsets of A.
(e) In order to make sense of this construction we need to match it with an alternative route to the same
object, based on -algebras and -ideals of sets, as follows.
364C Proposition Let X be a set, a -algebra of subsets of X, and 1 a -ideal of .
(a) Write L
0
for the space of all -measurable functions from X to R. Then L
0
, with its linear structure,
ordering and multiplication inherited from R
X
, is a Dedekind -complete f-algebra with multiplicative
identity.
(b) Set
W= f : f L
0
, x : f(x) ,= 0 1.
Then
(i) W is a sequentially order-closed solid linear subspace and ideal of L
0
;
(ii) the quotient space L
0
/W, with its inherited linear, order and multiplicative structures, is a Dedekind
-complete Riesz space and an f-algebra with a multiplicative identity;
(iii) for f, g L
0
, f

in L
0
/W i x : f(x) > g(x) 1, and f

= g

in L
0
/W i x : f(x) ,=
g(x) 1.
proof (Compare 241A-241H.)
(a) The point is just that L
0
is a sequentially order-closed Riesz subspace and subalgebra of R
X
. The
facts we need to know that constant functions belong to L
0
, that f +g, f, f g, sup
nN
f
n
belong to L
0
whenever f, g, f
n
do and f
n
: n N is bounded above are all covered by 121E-121F. Its multiplicative
identity is of course the constant function X.
(b)(i) The necessary verications are all elementary.
(ii) Because W is a solid linear subspace of the Riesz space L
0
, the quotient inherits a Riesz space
structure (352Jb); because W is an ideal of the ring (L
0
, +, ), L
0
/W inherits a multiplication; it is a
commutative algebra because L
0
is; and has a multiplicative identity e = X

because X is the identity

of L
0
.
To check that L
0
/W is an f-algebra it is enough to observe that, for any non-negative f, g , h L
0
,
f

= (f g)

0,
and if f

= 0 then x : f(x) > 0 x : g(x) > 0 1, so that x : f(x)h(x) > 0 x : g(x) > 0 1
and
(f

) g

= (h

) g

= 0.
Finally, L
0
/W is Dedekind -complete, by 353J(a-iii).
(iii) For f, g L
0
,
f

(f g)
+
W x : f(x) > g(x) = x : (f g)
+
(x) ,= 0 1
(using the fact that the canonical map from L
0
to L
0
/W is a Riesz homomorphism, so that ((f g)
+
)

=
(f

)
+
). Similarly
f

= g

f g W x : f(x) ,= g(x) = x : (f g)(x) ,= 0 1.

316 Function spaces 364D
364D Theorem Let X be a set and a -algebra of subsets of X. Let A be a Dedekind -complete
Boolean algebra and : A a surjective Boolean homomorphism, with kernel a -ideal 1; dene L
0
and
W as in 364C, so that U = L
0
/W is a Dedekind -complete f-algebra with multiplicative identity.
(a) We have a canonical bijection T : U L
0
= L
0
(A) dened by the formula
[[Tf

> ]] = x : f(x) >

for every f L
0
, R.
(b)(i) For any u, v U,
[[T(u +v) > ]] = sup
qQ
[[Tu > q]] [[Tv > q]]
for every R.
(ii) For any u U, > 0,
[[T(u) > ]] = [[Tu >

]]
for every R.
(iii) For any u, v U,
u v [[Tu > ]] [[Tv > ]] for every R.
(iv) For any u, v U
+
,
[[T(u v) > ]] = sup
qQ,q>0
[[Tu > q]] [[Tv >

q
]]
for every 0.
(v) Writing e = (X)

for the multiplicative identity of U, we have

[[Te > ]] = 1 if < 1, 0 if 1.
proof (a)(i) Given f L
0
, set
f
() = x : f(x) > for R. Then it is easy to see that
f
satises
the conditions ()
t
-()
t
of 364Bd, because is sequentially order-continuous (313Qb). Moreover, if f

= g

in U, then

f
() .
g
() = (x : f(x) > x : g(x) > ) = 0
for every R, because
x : f(x) > x : g(x) > x : f(x) ,= g(x) 1,
and
f
=
g
. So we have a well-dened member Tu of L
0
dened by the given formula, for any u U.
(ii) Next, given w L
0
, there is a u L
0
/W such that Tu = w. PPP For each q Q, choose F
q

such that F
q
= [[w > q]] in A. Note that if q
t
q then
(F
q
F
q
) = [[u > q
t
]] \ [[u > q]] = 0,
so F
q
F
q
1. Set
H =

qQ
F
q

nN

qQ,qn
F
q
,
and for x X set
f(x) = supq : q Q, x F
q
if x H,
= 0 otherwise.
(H is chosen just to make the formula here give a nite value for every x.) We have
H = sup
qQ
[[w > q]] \ inf
nN
sup
qQ,qn
[[w > q]]
= 1
A
\ inf
nN
[[w > n]] = 1
A
\ 0
A
= 1
A
,
so X H 1. Now, for any R,
364D L
0
317
x : f(x) > =
_
qQ,q>
F
q
(X H) if < 0
=
_
qQ,q>
F
q
(X H) if 0,
and in either case belongs to ; so that f L
0
and f

is dened in L
0
. Next, for any R,
[[Tf

> ]] = x : f(x) > = (

_
qQ,q>
F
q
)
= sup
qQ,q>
[[w > q]] = [[w > ]],
and Tf

= w. QQQ
(iii) Thus T is surjective. To see that it is injective, observe that if f, g L
0
, then
Tf

= Tg

= [[Tf

> ]] = [[Tg

> ]] for every R

= x : f(x) > = x : g(x) > for every R
= x : f(x) > x : g(x) > 1 for every R
= x : f(x) ,= g(x) =
_
qQ
(x : f(x) > qx : g(x) > q) 1
= f

= g

.
So we have the claimed bijection.
(b)(i) Let f, g L
0
be such that u = f

, v = g

, so that u +v = (f +g)

. For any R,
[[T(u +v) > ]] = x : f(x) +g(x) >
= (
_
qQ
x : f(x) > q x : g(x) > q)
= sup
qQ
x : f(x) > q x : g(x) > q
(because is a sequentially order-continuous Boolean homomorphism)
= sup
qQ
[[Tu > q]] [[Tv > q]].
(ii) Let f L
0
be such that f

= u, so that (f)

= u. For any R,
[[T(u) > ]] = x : f(x) > = x : f(x) >

= [[Tu >

]].
(iii) Let f, g L
0
be such that f

= u and g

= v. Then
u v x : f(x) > g(x) 1
(see 364C(b-iii))

_
qQ
x : f(x) > q g(x) 1
x : f(x) > x : g(x) > 1 for every R
x : f(x) > \ x : g(x) > = 0 for every
[[Tu > ]] [[Tv > ]] for every .
318 Function spaces 364D
(iv) Now suppose that u, v 0, so that they can be expressed as f

, g

where f, g 0 in L
0
(351J),
and u v = (f g)

. If 0, then
[[T(u v) > ]] = (
_
qQ,q>0
x : f(x) > q x : g(x) >

q
)
= sup
qQ,q>0
x : f(x) > q x : g(x) >

q

= sup
qQ,q>0
[[Tu > q]] [[Tv >

q
]].
(v) This is trivial, because
[[T(X)

> ]] = x : (X)(x) >

= X = 1 if < 1,
= = 0 if 1.
364E Theorem Let A be a Dedekind -complete Boolean algebra. Then L
0
= L
0
(A) has the structure
of a Dedekind -complete f-algebra with multiplicative identity e, dened by saying
[[u +v > ]] = sup
qQ
[[u > q]] [[v > q]],
whenever u, v L
0
and R,
[[u > ]] = [[u >

]]
whenever u L
0
, ]0, [ and R,
u v [[u > ]] [[v > ]] for every R,
[[u v > ]] = sup
qQ,q>0
[[u > q]] [[v >

q
]]
whenever u, v 0 in L
0
and 0,
[[e > ]] = 1 if < 1, 0 if 1.
proof (a) By the Loomis-Sikorski theorem (314M), we can nd a set Z (the Stone space of A), a -algebra
of subsets of Z (the algebra generated by the open-and-closed sets and the ideal / of meager sets)
and a surjective sequentially order-continuous Boolean homomorphism : A (corresponding to the
identication between A and the quotient //). Consequently, dening L
0
and W as in 364C, we have
a bijection between the Dedekind -complete f-algebra L
0
/W and L
0
(364Da). Of course this endows L
0
itself with the structure of a Dedekind -complete f-algebra; and 364Db tells us that the description of the
algebraic operations above is consistent with this structure.
(b) In fact the f-algebra structure is completely dened by the description oered. For while scalar
multiplication is not described for 0, the assertion that L
0
is a Riesz space implies that 0u = 0 and that
u = ()(u) for < 0; so if we have formulae to describe u + v and u for > 0, this suces to dene
the linear structure of L
0
. Note that we have an element 0 in L
0
dened by setting
[[0 > ]] = 0 if 0, 1 if < 0,
and the formula for u +v shows us that
[[0 +u > ]] = sup
qQ
[[0 > q]] [[u > q]] = sup
qQ,q<0
[[u > q]] = [[u > ]]
for every , so that 0 is the zero of L
0
. As for multiplication, if L
0
is to be an f-algebra we must have
[[u v > ]] [[0 > ]] = 1
whenever u, v (L
0
)
+
and < 0, because u v 0. So the formula oered is sucient to determine u v
for non-negative u and v; and for others we know that
364I L
0
319
u v = (u
+
v
+
) (u
+
v

) (u

v
+
) + (u

),
so the whole of the multiplication of L
0
is dened.
364F The rest of this section will be devoted to understanding the structure just established. I start
with a pair of elementary facts.
Lemma Let A be a Dedekind -complete Boolean algebra.
(a) If u, v L
0
= L
0
(A) and , R,
[[u +v > +]] [[u > ]] [[v > ]].
(b) If u, v 0 in L
0
and , 0 in R,
[[u v > ]] [[u > ]] [[v > ]].
proof (a) For any q Q, either q and [[u > q]] [[u > ]], or q and [[v > + q]] [[v > ]]; thus
in all cases
[[u > q]] [[v > + q]] [[u > ]] [[v > ]];
taking the supremum over q, we have the result.
(b) The same idea works, replacing + q by /q for q > 0.
364G Yet another description of L
0
is sometimes appropriate, and leads naturally to an important
construction (364I).
Proposition Let A be a Dedekind -complete Boolean algebra. Then there is a bijection between L
0
=
L
0
(A) and the set of sequentially order-continuous Boolean homomorphisms from the algebra B of Borel
subsets of R to A, dened by saying that u L
0
corresponds to i [[u > ]] = (], [) for every
R.
proof (a) If , then the map (], [) satises the conditions of 364Bd, so corresponds to an
element u

of L
0
.
(b) If , and u

= u

, then = . PPP Set / = E : E B, (E) = (E). Then / is a

-subalgebra of B, because and are both sequentially order-continuous Boolean homomorphisms, and
contains ], [ for every R. Now / contains ], ] for every , and therefore includes B (121J). But
this means that = . QQQ
(c) Thus u

is injective. But it is also surjective. PPP As in 364E, take a set Z, a -algebra

of subsets of Z and a surjective sequentially order-continuous Boolean homomorphism : A; let
T : L
0
/W L
0
be the bijection described in 364D. If u L
0
, there is an f L
0
such that Tf

= u.
Now consider E = f
1
[E] for E B. f
1
[E] always belongs to (121Ef), so E is always well-dened;
E f
1
[E] and are sequentially order-continuous, so also is; and
(], [) = z : f(z) > = [[u > ]]
for every , so u = u

. QQQ
Thus we have the declared bijection.
364H Denition In the context of 364G, I will write [[u E]], the region where u takes values in E,
for (E), where : B A is the homomorphism corresponding to u L
0
. Thus [[u > ]] = [[u ], [ ]].
In the same spirit I write [[u ]] for [[u [, [ ]] = inf
<
[[u > ]], [[u ,= 0]] = [[[u[ > 0]] = [[u > 0]] [[u < 0]]
and so on, so that (for instance) [[u = ]] = [[u ]] = [[u ]] \ [[u > ]] for u L
0
, R.
364I Proposition Let A be a Dedekind -complete Boolean algebra, E R a Borel set, and h : E R
a Borel measurable function. Then whenever u L
0
= L
0
(A) is such that [[u E]] = 1, there is an element

h(u) of L
0
dened by saying that [[

h(u) F]] = [[u h

1
[F]]] for every Borel set F R.
proof All we have to observe is that F [[u h
1
[F]]] is a sequentially order-continuous Boolean homo-
morphism. (The condition [[u E]] = 1 ensures that [[u h
1
[R]]] = 1.)
320 Function spaces 364J
364J Examples Perhaps I should spell out the most important contexts in which we apply these ideas,
even though they have in eect already been mentioned.
(a) Let X be a set and a -algebra of subsets of X. Then we may identify L
0
() with the space L
0
of -measurable real-valued functions on X. (This is the case A = of 364D.) For f L
0
, [[f E]] (364H)
is just f
1
[E], for any Borel set E R; and if h is a Borel measurable function,

h(f) (364I) is just the
composition hf, for any f L
0
.
(b) Now suppose that 1 is a -ideal of and that A = /1. Then, as in 364D, we identify L
0
(A) with
a quotient L
0
/W. For f L
0
, [[f

E]] = f
1
[E]

, and

h(f

) = (hf)

, for any Borel set E and any Borel

measurable function h : R R.
(c) In particular, if (X, , ) is a measure space with measure algebra A, then L
0
(A) becomes identied
with L
0
() as dened in 241.
The same remarks as in 363I apply here; the space L
0
() of 241 is larger than the space L
0
of the present
section. But for every f L
0
() there is a g L
0
such that g = f a.e. (241Bk), so that L
0
() can be
identied with L
0
/N, where N is the set of functions in L
0
which are zero almost everywhere (241Yh).
364K Embedding S and L

in L
0
: Proposition Let A be a Dedekind -complete Boolean algebra.
(a) We have a canonical embedding of L

= L

(A) as an order-dense solid linear subspace of L

0
= L
0
(A);
it is the solid linear subspace generated by the multiplicative identity e of L
0
. Consequently S = S(A) is
also embedded as an order-dense Riesz subspace and subalgebra of L
0
.
(b) This embedding respects the linear, lattice and multiplicative structures of L

and S.
(c) For a A, a, when regarded as a member of L
0
, can be described by the formula
[[a > ]] = 1 if < 0,
= a if 0 < 1,
= 0 if 1 .
The function : A L
0
is additive, injective, order-continuous and a lattice homomorphism.
(d) For every u (L
0
)
+
there is a non-decreasing sequence u
n
)
nN
in S such that u
0
0 and sup
nN
u
n
=
u.
proof Let Z, , L
0
, Wand be as in the proof of 364E. I dened L

to be the space C(Z) of continuous real-

valued functions on Z (363A); but because A is Dedekind -complete, there is an alternative representation
as L

/WL

, where L

is the space of bounded -measurable functions from Z to R (363Hb). Put like

this, we clearly have an embedding of L

= L

/W L

in L
0
= L
0
/W; and this embedding represents
L

as a Riesz subspace and subalgebra of L

0
, because L

is a Riesz subspace and subalgebra of L

0
. L

becomes the solid linear subspace of L

0
generated by (Z)

= e, because L

is the solid linear subspace of L

0
generated by Z. To see that L

is order-dense in L
0
, we have only to note that f = sup
nN
f nZ in L
0
for every f L
0
, and therefore (because the map f f

is sequentially order-continuous) u = sup

nN
une
in L
0
for every u L
0
.
To identify a, we have the formula (E) = (E)

, as in 363Hc; but this means that, if a = E,

[[a > ]] = z : E(z) > = Z = 1 if < 0,
= E = a if 0 < 1,
= = 0 if 1,
using the formula in 364Da. Evidently is injective.
Because S is an order-dense Riesz subspace and subalgebra of L

(363C), the same embedding represents

it as an order-dense Riesz subspace and subalgebra of L
0
. (For order-dense, use 352Nc.)
Because : A L

is additive, order-continuous and a lattice homomorphism (363D), and the embed-

ding map L

L
0
also is, : A L
0
has the same properties.
Finally, if u 0 in L
0
, we can represent it as f

where f 0 in L
0
. For n N set
364M L
0
321
f
n
(z) = 2
n
k if 2
n
k f(z) < 2
n
(k + 1) where 0 k < 4
n
,
= 0 if f(z) 2
n
;
then f

n
)
nN
is a non-decreasing sequence in S
+
with supremum u.
364L Corollary Let (A, ) be a measure algebra. Then S(A
f
) can be embedded as a Riesz subspace
of L
0
(A), which is order-dense i (A, ) is semi-nite.
proof (Recall that A
f
is the ring a : a < .) The embedding A
f
A is an injective ring homomorphism,
so induces an embedding of S(A
f
) as a Riesz subspace of S(A), by 361J. Now S(A
f
) is order-dense in S(A)
i (A, ) is semi-nite. PPP (i) If (A, ) is semi-nite and v > 0 in S(A), then v is expressible as

n
j=0

j
b
j
where
j
0 for each j and some
j
b
j
is non-zero; now there is a non-zero a A
f
such that a b
j
, so
that 0 <
j
a S(A
f
) and
j
a v. As v is arbitrary, S(A
f
) is quasi-order-dense, therefor order-dense
(353A). (ii) If S(A
f
) is order-dense in S(A) and b A 0, there is a u > 0 in S(A
f
) such that u b;
now there are > 0, a A
f
0 such that a u, in which case a b. QQQ
Now because S(A
f
) S(A) and S(A) is order-dense in L
0
(A), we must have
S(A
f
) is order-dense in L
0
(A) S(A
f
) is order-dense in S(A)
(A, ) is semi-nite.
364M Suprema and inma in L
0
We know that any L
0
(A) is a Dedekind -complete partially or-
dered set. There is a useful description of suprema for this ordering, as follows.
Proposition Let A be a Dedekind -complete Boolean algebra and A a subset of L
0
= L
0
(A).
(a) A is bounded above in L
0
i there is a sequence c
n
)
nN
in A, with inmum 0, such that [[u > n]] c
n
for every u A.
(b) If A is non-empty, then A has a supremum in L
0
i c

= sup
uA
[[u > ]] is dened in A for every
R and inf
nN
c
n
= 0; and in this case c

= [[supA > ]] for every .

(c) If A is non-empty and bounded above, then A has a supremum in L
0
i sup
uA
[[u > ]] is dened in
A for every R.
proof (a)(i) If A has an upper bound u
0
, set c
n
= [[u
0
> n]] for each n; then c
n
)
nN
satises the conditions.
(ii) If c
n
)
nN
satises the conditions, set
() = 1 if < 0,
= inf
in
c
i
if n N, [n, n + 1[ .
Then it is easy to check that satises the conditions of 364A, since inf
nN
c
n
= 0. So there is a u
0
L
0
such that () = [[u
0
> ]] for each . Now, given u A and R,
[[u > ]] 1 = [[u
0
> ]] if < 0,
inf
in
[[u > i]] inf
in
c
i
= [[u
0
> ]] if n N, [n, n + 1[ .
Thus u
0
is an upper bound for A in L
0
.
(b)(i) Suppose that c

= sup
uA
[[u > ]] is dened in A for every , and that inf
nN
c
n
= 0. Then, for
any ,
sup
qQ,q>
c
q
= sup
uA,qQ,q>
[[u > q]] = sup
uA
[[u > q]] = c

.
Also, we are supposing that A contains some u
0
, so that
sup
nN
c
n
sup
nN
[[u
0
> n]] = 1.
Accordingly there is a u

L
0
such that [[u

> ]] = c

for every R. But now, for v L

0
,
322 Function spaces 364M
v is an upper bound for A [[u > ]] [[v > ]] for every u A, R
[[u

> ]] [[v > ]] for every

u

v,
so that u

= sup A in L
0
.
(ii) Now suppose that u

= sup A is dened in L
0
. Of course [[u

> ]] must be an upper bound for

[[u > ]] : u A for every . ??? Suppose we have an for which it is not the least upper bound, that is,
there is a c [[u

> ]] which is an upper bound for [[u > ]] : u A. Dene : R A by setting

() = c [[u

> ]] if ,
= [[u

> ]] if < .
It is easy to see that satises the conditions of 364A (we need the distributive law 313Ba to check that
() = sup
>
() if ), so corresponds to a member v of L
0
. But we now nd that v is an upper
bound for A (because if u A, then
[[u > ]] [[u > ]] [[u

> ]] c [[u

> ]] = [[v > ]],)

that v u

and that v ,= u

(because [[v > ]] = c ,= [[u

> ]]); but this is impossible, because u

is
supposed to be the supremum of A. XXX Thus if u

= sup A is dened in L
0
, then sup
uA
[[u > ]] is dened
in A for every R. Also, of course,
inf
nN
sup
uA
[[u > n]] = inf
nN
[[u

> n]] = 0.
(c) This is now easy. If A has a supremum, then surely it satises the condition, by (b). If A satises
the condition, then we have a family c

)
R
as required in (b); but also, by (a) or otherwise, there is a
sequence c
t
n
)
nN
such that c
n
c
t
n
for every n and inf
nN
c
t
n
= 0, so inf
nN
c
n
is also 0, and both conditions
in (b) are satised, so A has a supremum.
364N We do not have such a simple formula for general inma (though see 364Xl), but the following
facts are useful.
Proposition Let A be a Dedekind -complete Boolean algebra.
(a) If u, v L
0
= L
0
(A), then [[u v > ]] = [[u > ]] [[v > ]] for every R.
(b) If A is a non-empty subset of (L
0
)
+
, then inf A = 0 in L
0
i inf
uA
[[u > ]] = 0 in A for every > 0.
proof (a) Take Z, L
0
and as in the proof of 364E. Express u as f

, v as g

where f, g L
0
, so that
u v = (f g)

, because the canonical map from L

0
to L
0
is a Riesz homomorphism (351J). Then
[[u v > ]] = z : min(f(z), g(z)) > = (z : f(z) > z : g(z) > )
= z : f(z) > z : g(z) > = [[u > ]] [[v > ]]
for every .
(b)(i) If inf
uA
[[u > ]] = 0 for every > 0, and v is any lower bound for A, then [[v > ]] must be 0 for
every > 0, so that [[v > 0]] = 0; now since [[0 > ]] = 0 for 0, 1 for < 0, v 0. As v is arbitrary,
inf A = 0.
(ii) If > 0 is such that inf
uA
[[u > ]] is undened, or not equal to 0, let c A be such that
0 ,= c [[u > ]] for every u A, and consider v = c. Then [[v > ]] = [[c >

]] is 1 if < 0, c if
0 < and 0 if . If u A then [[u > ]] is 1 if < 0 (since u 0), at least [[u > ]] c if 0 < ,
and always includes 0; so that v u. As u is arbitrary, inf A is either undened in L
0
or not 0.
364O Now we have a reward for our labour, in that the following basic theorem is easy.
Theorem For a Dedekind -complete Boolean algebra A, L
0
= L
0
(A) is Dedekind complete i A is.
proof The description of suprema in 364Mc makes it obvious that if A is Dedekind complete, so that
sup
uA
[[u > ]] is always dened, then L
0
must be Dedekind complete. On the other hand, if L
0
is Dedekind
complete, then so is L

(A) (by 364K and 353J(b-i)), so that A is also Dedekind complete, by 363Mb.
364R L
0
323
364P The multiplication of L
0
I have already observed that L
0
is always an f-algebra with identity;
in particular (because L
0
is surely Archimedean) the map u u v is order-continuous for every v 0
(353Oa), and multiplication is commutative (353Ob, or otherwise). The multiplicative identity is 1 (364E,
364Kc). By 353Pb, or otherwise, u v = 0 i [u[ [v[ = 0. There is one special feature of multiplication in
L
0
which I can mention here.
Proposition Let A be a Dedekind -complete Boolean algebra. Then an element u of L
0
has a multiplicative
inverse in L
0
i [u[ is a weak order unit in L
0
i [[[u[ > 0]] = 1.
proof If u is invertible, then [u[ is a weak order unit, by 353Pc or otherwise. In this case, setting c =
1 \ [[[u[ > 0]], we see that
[[[u[ c > 0]] = [[[u[ > 0]] c = 0
(364Na), so that [u[ c 0 and c = 0, that is, c = 0; so [[[u[ > 0]] must be 1. To complete the circuit,
suppose that [[[u[ > 0]] = 1. Let Z, , L
0
, , / be as in the proof of 364E, and S : L
0
L
0
the canonical
map, so that [[Sh > ]] = z : h(z) > for every h L
0
, R. Express u as Sf where f L
0
. Then
z : [f(z)[ > 0 = [[Sf > 0]] = 1, so z : f(z) = 0 /. Set
g(z) =
1
f(z)
if f(z) ,= 0, g(z) = 0 if f(z) = 0.
Then z : f(z)g(z) ,= 1 / so
u Sg = S(f g) = S(Z) = 1
and u is invertible.
Remark The repeated phrase by 353x or otherwise reects the fact that the abstract methods there can
all be replaced in this case by simple direct arguments based on the construction in 364C-364E.
364Q Recovering the algebra: Proposition Let A be a Dedekind -complete Boolean algebra. For
a A write V
a
for the band of L
0
= L
0
(A) generated by a. Then a V
a
is a Boolean isomorphism
between A and the algebra of projection bands of L
0
(A).
proof I copy from the argument for 363J, itself based on 361K. If a A, w L
0
then w a V
a
. PPP If
v V

a
then [a[ [v[ = 0, so a v = 0, so (w a) v = 0, so [w a[ [v[ = 0; thus w a V

a
,
which is equal to V
a
because L
0
is Archimedean (353B). QQQ Now, if a A, u V
a
and v V
1\a
, then
[u[ [v[ = 0 because a (1 \ a) = 0; and if w L
0
(A) then
w = (w a) + (w (1 \ a)) V
1
+V
1\a
.
So V
a
and V
1\a
are complementary projection bands in L
0
. Next, if U L
0
is a projection band, then 1
is expressible as u + v = u v where u U, v U

. Setting a = [[u >

1
2
]], a
t
= [[v >
1
2
]] we must have
a a
t
= 1, a a
t
= 0 (using 364M and 364Na), so that a
t
= 1 \ a; also
1
2
a u, so that a U, and
similarly (1 \ a) U

. In this case V
a
U and V
1\a
U

, so U must be V
a
precisely. Thus a V
a
is
surjective. Finally, just as in 361K, a b V
a
V
b
, so we have a Boolean isomorphism.
364R I come at last to the result corresponding to 361J and 363F.
Theorem Let A and B be Boolean algebras, and : A B a sequentially order-continuous Boolean
homomorphism.
(a) We have a multiplicative sequentially order-continuous Riesz homomorphism T

: L
0
(A) L
0
(B)
dened by the formula
[[T

u > ]] = [[u > ]]

for every R, u L
0
(A).
(b) Dening a L
0
(A) as in 364K, T

(a) = (a) in L
0
(B) for every a A. If we regard L

(A) and
L

(B) as embedded in L
0
(A) and L
0
(B) respectively, then T

, as dened here, agrees on L

(A) with T

as dened in 363F.
(c) T

is order-continuous i is order-continuous, injective i is injective, surjective i is surjective.

324 Function spaces 364R
(d) [[T

u E]] = [[u E]] for every u L

0
(A) and every Borel set E R; consequently

hT

= T

h for
every Borel measurable h : R R, writing

h indierently for the associated maps from L
0
(A) to itself and
from L
0
(B) to itself (364I).
(e) If C is another Dedekind -complete Boolean algebra and : B C another sequentially order-
continuous Boolean homomorphism then T

= T

: L
0
(A) L
0
(C).
proof I write T for T

.
(a)(i) To see that Tu is well-dened in L
0
(B) for every u L
0
(A), all we need to do is to check that
the map [[u > ]] : R B satises the conditions of 364Bd, and this is easy, because preserves all
countable suprema and inma.
(ii) To see that T is linear and order-preserving and multiplicative, we can use the formulae of 364E.
For instance, if u, v L
0
(A), then
[[Tu +Tv > ]] = sup
qQ
[[Tu > q]] [[Tv > q]] = sup
qQ
[[u > q]] [[v > q]]
= (sup
qQ
[[u > q]] [[v > q]]) = [[u +v > ]] = [[T(u +v) > ]]
for every R, so that Tu +Tv = T(u +v). In the same way,
T(u) = Tu whenever > 0,
Tu Tv whenever u v,
Tu Tv = T(u v) whenever u, v 0,
so that, using the distributive laws, T is linear and multiplicative.
To see that T is a sequentially order-continuous Riesz homomorphism, suppose that A L
0
(A) is a
countable non-empty set with a supremum u

L
0
(A); then T[A] is a non-empty subset of L
0
(B) with an
upper bound Tu

, and
sup
uA
[[Tu > ]] = sup
uA
[[u > ]] = (sup
uA
[[u > ]]) = [[u

> ]]
(using 364M)
= [[Tu

> ]]
for every R. So using 364M again, Tu

= sup
uA
Tu. Now this is true, in particular, for doubleton
sets A, so that T is a Riesz homomorphism; and also for non-decreasing sequences, so that T is sequentially
order-continuous.
(b) The identication of T(a) with (a) is another almost trivial verication. It follows that T agrees
with the map of 363F on S(A), if we think of S(A) as a subspace of L
0
(A). Next, if u L

(A) L
0
(A),
and = |u|

, then [u[ 1
A
, so that [Tu[ 1
B
, and Tu L

(B), with |Tu|

. Thus TL

(A)
has norm at most 1. As it agrees with the map of 363F on S(A), which is | |

-dense in L

(A) (363C), and

both are continuous, they must agree on the whole of L

(A).
(c)(i)() Suppose that is order-continuous, and that A L
0
(A) is a non-empty set with a supremum
u

L
0
(A). Then for any R,
[[Tu

> ]] = [[u

> ]] = (sup
uA
[[u > ]])
(by 364M)
= sup
uA
[[u > ]]
(because is order-continuous)
364S L
0
325
= sup
uA
[[Tu > ]].
As is arbitrary, Tu

= sup T[A], by 364M again. As A is arbitrary, T is order-continuous (351Ga).

() Now suppose that T is order-continuous and that A A is a non-empty set with supremum c
in A. Then c = sup
aA
a (364Kc) so
(c) = T(c) = sup
aA
T(a) = sup
aA
(a).
But now
c = [[(c) > 0]] = sup
aA
[[(a) > 0]] = sup
aA
a.
As A is arbitrary, is order-continuous.
(ii)() If is injective and u, v are distinct elements of L
0
(A), then there must be some such that
[[u > ]] ,= [[v > ]], in which case [[Tu > ]] ,= [[Tv > ]] and Tu ,= Tv.
() Now suppose that T is injective. It is easy to see that : A L
0
(A) is injective, so that
T : A L
0
(B) is injective; but this is the same as (by (b)), so must also be injective.
(iii)() Suppose that is surjective. Let be a -algebra of sets such that there is a sequentially
order-continuous Boolean surjection : A. Then : B is surjective. So given w L
0
(B), there
is an f L
0
() such that [[w > ]] = x : f(x) > for every R (364D). But, also by 364D, there
is a u L
0
(A) such that [[u > ]] = x : f(x) > for every . And now of course Tu = w. As w is
arbitrary, T is surjective.
() If T is surjective, and b B, there must be some u L
0
(A) such that Tu = b. Now set
a = [[u > 0]] and see that a = [[b > 0]] = b. As b is arbitrary, is surjective.
(d) The map E [[u E]] is a sequentially order-continuous Boolean homomorphism, equal to [[Tu E]]
when E is of the form ], [; so by 364G the two are equal for all Borel sets E.
If h : R R is a Borel measurable function, u L
0
(A) and E R is a Borel set, then
[[

h(Tu) E]] = [[Tu h

1
[E]]] = [[u h
1
[E]]]
= [[

h(u) E]] = [[T(

h(u)) E]].
As E and u are arbitrary, T

h =

hT.
(e) This is immediate from the denitions.
364S Products: Proposition Let A
i
)
iI
be a family of Dedekind -complete Boolean algebras,
with simple product A. If
i
: A A
i
is the coordinate map for each i, and T
i
: L
0
(A) L
0
(A
i
) the
corresponding homomorphism, then u Tu = T
i
u)
iI
: L
0
(A)

iI
L
0
(A
i
) is a multiplicative Riesz
space isomorphism, so L
0
(A) may be identied with the f-algebra product

iI
L
0
(A
i
) (352Wc).
proof Because each
i
is a surjective order-continuous Boolean homomorphism, 364R assures us that there
are corresponding surjective multiplicative Riesz homomorphisms T
i
. So all we need to check is that the
multiplicative Riesz homomorphism T : L
0
(A)

iI
L
0
(A
i
) is a bijection.
If u, v L
0
(A) are distinct, there must be some R such that [[Tu > ]] ,= [[Tv > ]]. In this case
there must be an i I such that
i
[[Tu > ]] ,=
i
[[Tv > ]], that is, [[T
i
u > ]] ,= [[T
i
v > ]]. So T
i
u ,= T
i
v
and Tu ,= Tv. As u, v are arbitrary, T is injective.
If w = w
i
)
iI
is any member of

iI
L
0
(A
i
), then for R set
() = [[w
i
> ]])
iI
A.
It is easy to check that satises the conditions of 364A, because, for instance,
sup
>

i
() = sup
>
[[w
i
> ]] = [[w
i
> ]] =
i
()
for every i, so that sup
>
() = (), for every R; and the other two conditions are also satised
because they are satised coordinate-by-coordinate. So there is a u L
0
(A) such that () = [[u > ]] for
every , that is,
i
[[u > ]] = [[w
i
> ]] for all , i, that is, T
i
u = w
i
for every i, that is, Tu = w. As w is
arbitrary, T is surjective and we are done.
326 Function spaces *364T
*364T Regular open algebras I noted in 314P that for every topological space X there is a corre-
sponding Dedekind complete Boolean algebra G of regular open sets. We have an identication of L
0
(G) as
a space of equivalence classes of functions, dierent in kind from the representations above, as follows. This
is hard work (especially if we do it in full generality), but instructive. I start with a temporary denition.
Denition Let (X, T) be a topological space, f : X R a function. For x X write
(f, x) = inf
GT,xG
sup
y,zG
[f(y) f(z)[
(allowing ).
*364U Theorem Let X be any topological space, and G its regular open algebra. Let U be the set
of functions f : X R such that x : (f, x) < is dense in X for every > 0. Then U is a Riesz
subspace of R
X
, closed under multiplication, and we have a surjective multiplicative Riesz homomorphism
T : U L
0
(G) dened by writing
[[Tf > ]] = sup
>
int x : f(x) > ,
the supremum being taken in G, for every R, f U. The kernel of T is the set W of functions f : X R
such that intx : [f(x)[ is dense for every > 0, so L
0
(G) can be identied, as f-algebra, with the
quotient space U/W.
proof (a)(i)() The rst thing to observe is that for any f R
X
, > 0 the set
x : (f, x) < =
_
G : G X is open and non-empty
and sup
y,zG
[f(y) f(z)[ <
is open.
() Next, it is easy to see that
(f +g, x) (f, x) +(g, x),
(f, x) = [[(f, x),
([f[, x) (f, x),
for all f, g R
X
and R.
() Thirdly, it is useful to know that if f U and G X is a non-empty open set, then there is a
non-empty open set G
t
G on which f is bounded. PPP Take any x
0
G such that (f, x
0
) < 1; then there
is a non-empty open set G
t
containing x
0
such that [f(y) f(z)[ < 1 for all y, z G
t
, and we may suppose
that G
t
G. But now [f(x)[ 1 +[f(x
0
)[ for every x G
t
. QQQ
(ii) So if f, g U and R then
x : (f +g, x) < x : (f, x) <
1
2
x : (g, x) <
1
2

is the intersection of two dense open sets and is therefore dense, while
x : (f, x) < x : (f, x) <

1+[[
,
x : ([f[, x) < x : (f, x) <
are also dense. As is arbitrary, f +g, f and [f[ all belong to U; as f, g and are arbitrary, U is a Riesz
subspace of R
X
.
(iii) If f, g U then f g U. PPP Take > 0 and let G
0
be a non-empty open subset of X. By the
last remark in (i) above, there is a non-empty open set G
1
G
0
such that [f[ [g[ is bounded on G
1
; say
max([f(x)[, [g(x)[) for every x G
1
.
*364U L
0
327
Set = min(1,

2(1+)
) > 0. Then there is an x G
1
such that (f, x) < and (g, x) < . Let H, H
t
be open sets containing x such that [f(y) f(z)[ < for all y, z H and [g(y) g(z)[ < for all y, z H
t
.
Consider G = G
1
H H
t
. This is an open set containing x, and if y, z G then
[f(y)g(y) f(z)g(z)[ [f(y) f(z)[[g(y) g(z)[
+[f(y) f(z)[[g(z)[ +[f(z)[[g(y) g(z)[

2
+ +.
Accordingly
(f g, x) (1 + 2) < ,
while x G
0
. As G
0
is arbitrary, x : (f g, x) < is dense; as is arbitrary, f g U. QQQ
Thus U is a subalgebra of R
X
.
(b) Now, for f U, consider the map
f
: R G dened by setting

f
() = sup
>
int x : f(x) >
for every R. Then
f
satises the conditions of 364A. PPP (See 314P for the calculation of suprema and
inma in G.) (i) If R then

f
() = sup
>
int x : f(x) > = sup
>>
int x : f(x) >
= sup
>
sup
>
int x : f(x) > = sup
>

f
().
(ii) If G
0
X is a non-empty open set, then there is a non-empty open set G
1
G
0
such that f is
bounded on G
1
; say [f(x)[ < for every x G
1
. If > then G
1
does not meet x : f(x) > , so
G
1
int x : f(x) > = ; as is arbitrary, G
1

f
() = and G
0
, inf
R

f
(). On the other hand,
G
1
x : f(x) > , so
G
1
int x : f(x) >
f
()
and G
0
sup
R

f
() ,= . As G
0
is arbitrary, inf
R

f
() = and sup
R

f
() = X. QQQ
(c) Thus we have a map T : U L
0
= L
0
(G) dened by setting [[Tf > ]] =
f
() for every R,
f U.
It is worth noting that
x : f(x) > +(f, x) [[Tf > ]] x : f(x) +(f, x)
for every f U, R. PPP (i) If f(x) > + (f, x), set =
1
2
(f(x) (f, x)) > 0. Then there is an
open set G containing x such that [f(y) f(z)[ < (f, x) + for every y, z G, so that f(y) > + for
every y G, and
x inty : f(y) > + [[Tf > ]].
(ii) If f(x) + (f, x) < , set =
1
2
( f(x) (f, x)) > 0; then there is an open neighbourhood G of x
such that [f(y) f(z)[ < (f, x) + for every y, z G, so that f(y) < for every y G. Accordingly G
does not meet y : f(y) > nor y : f(y) > for any > , G [[Tf > ]] = and x / [[Tf > ]]. QQQ
(d) T is additive. PPP Let f, g U and < R. Set =
1
5
( ) > 0, H = x : (f, x) < , (g, x) <
; then H is the intersection of two dense open sets, so is itself dense and open.
(i) If x H [[T(f +g) > ]], then (f + g)(x) + (f + g, x) ; but (f + g, x) 2 (see (a-i-)
above), so f(x) +g(x) 2 > + 2 and there is a q Q such that
f(x) > q + q +(f, x), g(x) > q + q +(g, x).
Accordingly
x [[Tf > q]] [[Tg > q]] [[Tf +Tg > ]].
328 Function spaces *364U
Thus H [[T(f +g) > ]] [[Tf +Tg > ]]. Because H is dense, [[T(f +g) > ]] [[Tf +Tg > ]].
(ii) If x H, then
x
_
qQ
([[Tf > q]] [[Tg > q]])
= q Q, f(x) +(f, x) q, g(x) +(g, x) q
= f(x) +g(x) + 2
= (f +g)(x) + 3 > +(f +g, x)
= x [[T(f +g) > ]].
Thus
H

qQ
([[Tf > q]] [[Tg > q]]) [[T(f +g) > ]].
Because H is dense and

qQ
([[Tf > q]] [[Tg > q]]) is open,
[[Tf +Tg > ]] = int
_
qQ
[[Tf > q]] [[Tg > q]]
int [[T(f +g) > ]] = [[T(f +g) > ]].
(iii) Now let ; we have
[[T(f +g) > ]] = sup
>
[[T(f +g) > ]] [[Tf +Tg > ]]
= sup
>
[[Tf +Tg > ]] [[T(f +g) > ]],
so [[T(f +g) > ]] = [[Tf +Tg > ]]. As is arbitrary, T(f +g) = Tf +Tg; as f and g are arbitrary, T is
additive. QQQ
(e) It is now easy to see that T is linear. PPP If > 0, f U, R then
[[T(f) > ]] = sup
>
int x : f(x) > = sup
>
int x : f(x) >

= sup
>/
int x : f(x) > = [[Tf >

]] = [[Tf > ]].

As is arbitrary, T(f) = Tf; because we already know that T is additive, this is enough to show that T
is linear. QQQ
(f ) In fact T is a Riesz homomorphism. PPP If f U, 0 then
[[T(f
+
) > ]] = sup
>
int x : f
+
(x) > = sup
>
int x : f(x) >
= [[Tf > ]] = [[(Tf)
+
> ]].
If < 0 then
[[T(f
+
) > ]] = sup
>
int x : f
+
(x) >
= X = [[(Tf)
+
> ]]. QQQ
(g) Of course the constant function X belongs to U, and is its multiplicative identity; and T(X) is
the multiplicative identity of L
0
, because
*364U L
0
329
[[T(X) > ]] = sup
>
int x : (X)(x) >
= X if < 1, if 1.
By 353Pd, or otherwise, T is multiplicative.
(h) The kernel of T is W. PPP (i) For f U,
Tf = 0 = [[T[f[ > 0]] = [[[Tf[ > 0]] =
= x : [f(x)[ > ([f[, x) =
= intx : [f(x)[ x : ([f[, x) <
is dense, for every > 0
= f W.
(ii) If f W, then, rst,
x : (f, x) < intx : [f(x)[
1
3

is dense for every > 0, so f U; and next, for any > 0, x : [f(x)[ > does not meet the dense open
set intx : [f(x)[ , so
[[[Tf[ > 0]] = [[T[f[ > 0]] = sup
>0
int x : [f(x)[ > =
and Tf = 0. QQQ
(i) Finally, T is surjective. PPP Take any u L
0
. Dene

f : X [, ] by setting

f(x) = sup : x
[[u > ]] for each x, counting inf as . Then
x :

f(x) > =

>
[[u > ]]
is open, for every R. The set
x :

f(x) = =

R
[[u > ]]
is nowhere dense, because inf
R
[[u > ]] = in G; while
x :

f(x) = = X

R
[[u > ]]
is also nowhere dense, because sup
R
[[u > ]] = X in G. Accordingly E = intx :

f(x) R is dense. Set
f(x) =

f(x) for x E, 0 for x X E.
Let > 0. If G X is a non-empty open set, there is an R such that G , [[u > ]], so G
1
=
G [[u > ]] ,= , and

f(x) for every x G
1
. Set

t
= sup
xG
1

f(x) < .
Because E meets G
1
,
t
> . Then G
2
= G
1
[[u >
t

1
2
]] is a non-empty open subset of G and

t

1
2


f(x)
t
for every x G
2
. Accordingly [f(y) f(z)[
1
2
for all y, z G
2
, and (f, x) < for
all x G
2
. As G is arbitrary, x : (f, x) < is dense; as is arbitrary, f U.
Take < in R, and set =
1
2
( ). Then H = E x : (f, x) < is a dense open set, and
H [[Tf > ]] H x : f(x) +(f, x) E x : f(x) >
x :

f(x) > [[u > ]].
As H is dense, [[Tf > ]] [[u > ]]. In the other direction
H [[u > ]] H x :

f(x) = H x : f(x)
x : f(x) > +(f, x) [[Tf > ]],
so [[u > ]] [[Tf > ]]. Just as in (d) above, this is enough to show that Tf = u. As u is arbitrary, T is
surjective. QQQ
This completes the proof.
330 Function spaces *364V
*364V Compact spaces Suppose now that X is a compact Hausdor topological space. In this case
the space U of 364U is just the space of functions f : X R such that x : f is continuous at x is dense
in X. PPP It is easy to see that
x : f is continuous at x = x : (f, x) = 0 =

nN
H
n
where H
n
= x : (f, x) < 2
n
for each n. Each H
n
is an open set (see part (a-i-) of the proof of 364U),
so by Baires theorem (3A3G)

nN
H
n
is dense i every H
n
is dense, that is, i f U. QQQ
Now W becomes f : f U, x : f(x) = 0 is dense. PPP (i) If f W, then T[f[ = 0, so (by the formula
in (c) of the proof of 364U) [f(x)[ ([f[, x) for every x. But x : (f, x) = 0 is dense, because f U, so
x : f(x) = 0 is also dense. (ii) If f U and x : f(x) = 0 is dense, then
(f, x) inf
xG is open
sup
yG
[f(y) f(x)[ [f(x)[
for every x X. So for any > 0, intx : [f(x)[ x : (f, x) < is dense, and f W. QQQ
In the case of extremally disconnected spaces, we can go farther.
*364W Theorem Let X be a compact Hausdor extremally disconnected space, and G its regular
open algebra. Write C

= C

(X) for the space of continuous functions g : X [, ] such that

x : g(x) = is nowhere dense. Then we have a bijection S : C

L
0
= L
0
(G) dened by saying that
[[Sg > ]] = x : g(x) >
for every R. Addition and multiplication in L
0
correspond to the operations

+,

on C

dened by
saying that g

+h, g

h are the unique elements of C

agreeing with g +h, g h on x : g(x), h(x) are both

nite. Scalar multiplication in L
0
corresponds to the operation
(g)(x) = g(x) for x X, g C

, R
on C

(counting 0 as 0), while the ordering of L

0
corresponds to the relation
g h g(x) h(x) for every x X.
proof (a) For g C

, set H
g
= x : g(x) R, so that H
g
is a dense open set, and dene Rg : X R by
setting (Rg)(x) = g(x) if x H
g
, 0 if x X H
g
. Then Rg is continuous at every point of H
g
, so belongs
to the space U of 364U-364V. Set Sg = T(Rg), where T : U L
0
is the map of 364U. Then
[[Sg > ]] = x : g(x) >
for every R. PPP (i) (g, x) = 0 for every x H
g
, so, if > ,
H
g
[[Sg > ]] x : x H
g
, (Rg)(x) x : g(x)
by the formula in part (c) of the proof of 364U. As [[Sg > ]] is open and H
g
is dense,
[[Sg > ]] H
g
[[Sg > ]] x : g(x) x : g(x) > .
Now
[[Sg > ]] = sup
>
[[Sg > ]] = int

>
[[Sg > ]] x : g(x) > .
(ii) In the other direction, H
g
x : g(x) > [[Sg > ]], by the other half of the formula in the proof of
364U. Again because x : g(x) > is open and H
g
is dense,
x : g(x) > [[Sg > ]] = [[Sg > ]]
because X is extremally disconnected (see 314S). QQQ
(b) Thus we have a function S dened by the formula oered. Now if g, h C

and g h, we surely
have x : g(x) > x : h(x) > for every , so [[Sg > ]] [[Sh > ]] for every and Sg Sh.
On the other hand, if g , h then Sg , Sh. PPP Take x
0
such that g(x
0
) > h(x
0
), and R such that
g(x
0
) > > h(x
0
); set H = x : g(x) > > h(x); this is a non-empty open set and H [[Sg > ]]. On
the other hand, H x : h(x) > = so H [[Sh > ]] = . Thus [[Sg > ]] , [[Sh > ]] and Sg , Sh. QQQ
In particular, S is injective.
(c) S is surjective. PPP If u L
0
, set
364Xe L
0
331
g(x) = sup : x [[u > ]] [, ]
for every x X, taking sup = . Then, for any R, x : g(x) > =

>
[[u > ]] is open. On the
other hand,
x : g(x) < =

<
x : x / [[u > ]]
is also open, because all the sets [[u > ]] are open-and-closed. So g : X [, ] is continuous. Also
x : g(x) > =

R
[[u > ]],
x : g(x) < =

R
X [[u > ]]
are dense, so g C

. Now, for any R,

[[Sg > ]] = x : g(x) > =
_
>
[[u > ]]
= int
_
>
[[u > ]] = sup
>
[[u > ]] = [[u > ]].
So Sg = u. As u is arbitrary, S is surjective. QQQ
(d) Accordingly S is a bijection. I have already checked (in part (b)) that it is an isomorphism of the
order structures. For the algebraic operations, observe that if g, h C

then there are f

1
, f
2
C

such
that Sg +Sh = Sf
1
and Sg Sh = Sf
2
, that is,
T(Rg +Rh) = TRg +TRh = TRf
1
, T(Rg Rh) = TRg TRh = TRf
2
.
But this means that
T(Rg +Rh Rf
1
) = T((Rg Rh) Rf
2
) = 0,
so that Rg +RhRf
1
, (Rg Rh) Rf
2
belong to W and are zero on dense sets (364V). Since we know also
that the set G = x : g(x), h(x) are both nite is a dense open set, while g, h, f
1
and f
2
are all continuous,
we must have f
1
(x) = g(x) +h(x), f
2
(x) = g(x)h(x) for every x G. And of course this uniquely species
f
1
and f
2
as members of C

.
Thus we do have operations

+,

as described, rendering S additive and multiplicative. As for scalar
multiplication, it is easy to check that R(g) = Rg (at least, unless = 0, which is trivial), so that
S(g) = Sg for every g C

.
364X Basic exercises (a) Let A be a Dedekind -complete Boolean algebra. For u, v L
0
= L
0
(A)
set [[u < v]] = [[v > u]] = [[v u > 0]], [[u v]] = [[v u]] = 1 \ [[v < u]], [[u = v]] = [[u v]] [[v u]]. (i) Show
that ([[u < v]], [[u = v]], [[u > v]]) is always a partition of unity in A. (ii) Show that for any u, u
t
, v, v
t
L
0
,
[[u u
t
]] [[v v
t
]] [[u +v u
t
+v
t
]] and [[u = u
t
]] [[v = v
t
]] [[u v = u
t
v
t
]].
(b) Let A be a Dedekind -complete Boolean algebra. (i) Show that if u, v L
0
= L
0
(A) and ,
R then [[u +v +]] [[u ]] [[v ]]. (ii) Show that if u, v (L
0
)
+
and , 0 then
[[u v ]] [[u ]] [[v ]].
(c) Let A be a Dedekind -complete Boolean algebra and u L
0
(A). Show that [[u E]] : E R is
Borel is the -subalgebra of A generated by [[u > ]] : R.
>>>(d) Let (A, ) be a probability algebra. Show that for any u L
0
(A) there is a unique Radon probability
measure on R (the distribution of u) such that E = [[u E]] for every Borel set E R. (Hint: 271B.)
(e) Let (A, ) be a probability algebra, and u
i
)
iI
any family in L
0
(A); for each i I let B
i
be the
closed subalgebra of A generated by [[u
i
> ]] : R. Show that the following are equiveridical: (i)
(inf
iJ
[[u
i
>
i
]]) =

iJ
[[u
i
>
i
]] whenever J I is nite and
i
R for each i J (ii) B
i
)
iI
is
independent in the sense of 325L. (In this case we may call u
i
)
iI
(stochastically) independent.)
332 Function spaces 364Xf
(f ) Let (A, ) be a probability algebra and u, v two stochastically independent members of L
0
(A). Show
that the distribution of their sum is the convolution of their distributions. (Hint: 272S.)
>>>(g) Let A be a Dedekind -complete Boolean algebra and g, h : R R Borel measurable functions. (i)
Show that g

h =

gh, where g,

h : L
0
L
0
are dened as in 364I. (ii) Show that g +h(u) = g(u) +

h(u),
g h(u) = g(u)

h(u) for every u L
0
= L
0
(A). (iii) Show that if h
n
)
nN
is a sequence of Borel
measurable functions on R and sup
nN
h
n
= h, then sup
nN

h
n
(u) =

h(u) for every u L
0
. (iv) Show that
if h is non-decreasing and continuous on the left, then

h(supA) = sup

h[A] whenever A L
0
is a non-empty
set with a supremum in L
0
.
(h) Let A be a Dedekind -complete Boolean algebra. (i) Show that S(A) can be identied () with the
set of those u L
0
= L
0
(A) such that [[u > ]] : R is nite () with the set of those u L
0
such that
[[u I]] = 1 for some nite I R. (ii) Show that L

(A) can be identied with the set of those u L

0
such
that [[u [, ]]] = 1 for some 0, and that |u|

is the smallest such .

(i) Show that if A is a Dedekind -complete Boolean algebra, and u L
0
(A), then for any R
[[u > ]] = inf
>
supa : a A, u a a
(compare 363Xh).
(j) Let A be a Dedekind -complete Boolean algebra and u 0 in L
0
= L
0
(A). Show that u =
sup
qQ
q[[u > q]] in L
0
.
(k) (i) Let A be a Dedekind -complete Boolean algebra and A L
0
(A) a non-empty countable set with
supremum w. Show that [[w E]] sup
uA
[[u E]] for every Borel set E R. (ii) Let (A, ) be a localizable
measure algebra and A L
0
(A) a non-empty set with supremum w. Show that [[w E]] sup
uA
[[u E]]
for every Borel set E R.
(l) Let A be a Dedekind -complete Boolean algebra and A L
0
= L
0
(A) a non-empty set which is
bounded below in L
0
. Suppose that
0
() = inf
uA
[[u > ]] is dened in A for every R. Show that
v = inf A is dened in L
0
, and that [[v > ]] = sup
>

0
() for every R.
(m) Let (X, , ) be a measure space and f : X [0, [ a function; set A = g

: g L
0
(), g f
a.e.. (i) Show that if (X, , ) either is localizable or has the measurable envelope property (213Xl), then
supA is dened in L
0
(). (ii) Show that if (X, , ) is complete and locally determined and w = sup A is
dened in L
0
(), then w A.
(n) Let A be a Dedekind -complete Boolean algebra. Show that if u, v L
0
= L
0
(A) then the following
are equiveridical: () [[[v[ > 0]] [[[u[ > 0]] () v belongs to the band of L
0
generated by u () there is a
w L
0
such that u w = v.
(o) Let A be a Dedekind -complete Boolean algebra and a A; let A
a
be the principal ideal of A
generated by a. Show that L
0
(A
a
) can be identied, as f-algebra, with the band of L
0
(A) generated by a.
(p) Let A and B be Dedekind -complete Boolean algebras, and : A B a sequentially order-
continuous Boolean homomorphism. Let T : L
0
(A) L
0
(B) be the corresponding Riesz homomorphism
(364R). Show that (i) the kernel of T is the sequentially order-closed solid linear subspace of L
0
(A) generated
by a : a A, a = 0 (ii) the set of values of T is the sequentially order-closed Riesz subspace of L
0
(B)
generated by (a) : a A.
(q) Let A and B be Dedekind -complete Boolean algebras, and : A B a sequentially order-
continuous Boolean homomorphism, with T : L
0
(A) L
0
(B) the associated operator. Suppose that h is a
Borel measurable real-valued function dened on a Borel subset of R. Show that

h(Tu) = T

h(u) whenever
u L
0
(A) and

h(u) is dened in the sense of 364I.
>>>(r) Let X and Y be sets, , T -algebras of subsets of X, Y respectively, and 1, -ideals of , T.
Set A = /1, B = T/. Suppose that : X Y is a function such that
1
[F] for every F T,
364Yk L
0
333

1
[F] 1 for every F . Show that there is a sequentially order-continuous Boolean homomorphism
: B A dened by saying that F

=
1
[F]

for every F T. Let T : L

0
(B) L
0
(A) be the Riesz
homomorphism corresponding to . Show that if we identify L
0
(B) with L
0
T
/W

and L
0
(A) with L
0

/W
1
in the manner of 364D, then T(g

) = (g)

for every g L
0
T
.
(s) Use the ideas of part (d) of the proof of 364U to show that the operator T there is multiplicative,
without appealing to 353P.
364Y Further exercises (a) Show directly, without using the Stone representation, that if A is any
Dedekind -complete Boolean algebra then the formulae of 364E dene a group operation + on L
0
(A), and
generally an f-algebra structure.
(b) Let A be a Dedekind -complete Boolean algebra. Show that A is ccc i L
0
(A) has the countable sup
property.
(c) Let A be a Dedekind -complete Boolean algebra and u
1
, . . . , u
n
members of L
0
(A). Write B
n
for
the algebra of Borel sets in R
n
. (i) Show that there is a unique sequentially order-continuous Boolean
homomorphism E [[(u
1
, . . . , u
n
) E]] : B
n
A such that [[(u
1
, . . . , u
n
) E]] = inf
in
[[u
i
>
i
]] when
E =

in
]
i
, [. (ii) Show that for every sequentially order-continuous Boolean homomorphism : B
n

A there are u
1
, . . . , u
n
L
0
(A) such that E = [[(u
1
, . . . , u
n
) E]] for every E B
n
.
(d) Let A be a Dedekind -complete Boolean algebra, n 1 and h : R
n
R a Borel measurable
function. Show that we have a corresponding function

h : L
0
(A)
n
L
0
(A) dened by saying that
[[

h(u
1
, . . . , u
n
) E]] = [[(u
1
, . . . , u
n
) h
1
[E]]] for every Borel set E R.
(e) Suppose that h
1
(x, y) = x +y, h
2
(x, y) = xy, h
3
(x, y) = max(x, y) for all x, y R. Show that, in the
language of 364Yd,

h
1
(u, v) = u +v,

h
2
(u, v) = u v,

h
3
(u, v) = u v for all u, v L
0
.
(f ) Let A and B be Dedekind -complete Boolean algebras, and T : L
0
(A) L
0
(B) a Riesz homomor-
phism such that Te = e
t
, where e, e
t
are the multiplicative identities of L
0
(A), L
0
(B) respectively. Show
that there is a unique sequentially order-continuous Boolean homomorphism : A B such that T = T

in the sense of 364R. (Hint: use 353Pd. Compare 375A below.)

(g) Suppose, in 364U, that X = Q. (i) Show that there is an f W such that f(q) > 0 for every q Q.
(ii) Show that there is a u L
0
such that no f U representing u can be continuous at any point of Q.
(h) Let X and Y be topological spaces and : X Y a continuous function such that
1
[M] is nowhere
dense in X for every nowhere dense subset M of Y . (Cf. 313R.) (i) Show that we have an order-continuous
Boolean homomorphism from the regular open algebra G
Y
of Y to the regular open algebra G
X
of X
dened by the formula G = int
1
[G] for every G G
Y
. (ii) Show that if U
X
, U
Y
are the function spaces
of 364U then g U
X
for every g U
Y
. (iii) Show that if T
X
: U
X
L
0
(G
X
), T
Y
: U
Y
L
0
(G
Y
)
are the canonical surjections, and T : L
0
(G
Y
) L
0
(G
X
) is the homomorphism corresponding to , then
T(T
Y
g) = T
X
(g) for every g U
Y
. (iv) Rewrite these ideas for the special case in which X is a dense
subset of Y and is the identity map, showing that in this case and T are isomorphisms.
(i) Let X be a Baire space, G its algebra of regular open sets, / its ideal of meager sets, and

B the Baire
property -algebra GA : G X is open, A /, so that G can be identied with

B// (314Yd). (i)
Repeat the arguments of 364V in this context. (ii) Show that the space U of 364U-364V is a subspace of
L
0
(

B), and that W = U W where W = f : f R

X
, x : f(x) ,= 0 /, so that the representations of
L
0
(G) as U/W, L
0
/W are consistent.
(j) Work through the arguments of 364U and 364Yi for the case of compact Hausdor X, seeking simpli-
cations based on 364V.
(k) Let X be an extremally disconnected compact Hausdor space with regular open algebra G. Let U
0
be the space of real-valued functions f : X R such that intx : f is continuous at x is dense. Show that
U
0
is a Riesz subspace of the space U of 364U, and that every member of L
0
(G) is represented by a member
of U
0
.
334 Function spaces 364Yl
(l) Let X be a Baire space. Let Q be the set of all continuous real-valued functions dened on subsets
of X, and Q

the set of all members of Q which are maximal in the sense that there is no member of Q
properly extending them. (i) Show that the domain of any member of Q

is a dense G

set. (ii) Show that

we can dene addition and multiplication and scalar multiplication on Q

by saying that f

+g, f

g, .f
are to be the unique members of Q

extending the partially-dened functions f + g, f g, f, and that

these denitions render Q

an f-algebra if we say that f g i f(x) g(x) for every x domf domg.

(iii) Show that every member of Q

has an extension to a member of U, as dened in 364U, and that these

extensions dene an isomorphism between Q

and L
0
(G), where G is the regular open algebra of X. (iv)
Show that if X is compact, Hausdor and extremally disconnected, then every member of Q

has a unique
extension to a member of C

(X), as dened in 364W.

(m) Let X be an extremally disconnected Hausdor space, and Z any compact Hausdor space. Show
that if D X is dense and f : D Z is continuous, there is a continuous g : X Z extending f.
(n) Let A be a Dedekind -complete Boolean algebra. Write L
0
C
= u+iv : u, v L
0
for the complexi-
cation of L
0
= L
0
(A) as dened in 354Yk. (i) Writing B(C) for the Borel -algebra of C, show that there is
a one-to-one correspondence between sequentially order-continuous Boolean homomorphisms : B(C) A
and members w = u + iv of L
0
C
dened by saying that [[u > ]] [[v > ]] = z : e z > , 1mz > .
(ii) Show that if is a -algebra of subsets of a set X and : A is a surjective sequentially order-
continuous Boolean homomorphism with kernel 1, then we can identify L
0
C
with L
0
C
/W, where L
0
C
is the set
of -measurable functions from X to C, and W is the set of those f L
0
C
such that x : f(x) ,= 0 1.
364 Notes and comments This has been a long section, and so far all we have is a supposedly thorough
grasp of the construction of L
0
spaces; discussion of their properties still lies ahead. The diculties seem
to stem from a variety of causes. First, L
0
spaces have a rich structure, being linear ordered spaces with
multiplications; consequently all the main theorems have to check rather a lot of dierent aspects. Second,
unlike L

spaces, they are not accessible by means of the theory of normed spaces, so I must expect to do
more of the work here rather than in an appendix. But this is in fact a crucial dierence, because it aects
the proof of the central theorem 364E. The point is that a given algebra A will be expressible in the form
/1 for a variety of algebras of sets. Consequently any denition of L
0
(A) as a quotient L
0
()/W must
include a check that the structure produced is independent of the particular pair , 1 chosen.
The same question arises with S(A) and L

(A). But in the case of S, I was able to use a general theory

of additive functions on A (see the proof of 361L), while in the case of L

I could quote the result for S

and a little theory of normed spaces (see the proof of 363H). The theorems of 368 will show, among other
things, that a similar approach (describing L
0
as a special kind of extension of S or L

) can be made to
work in the present situation. I have chosen, however, an alternative route using a novel technique. The
price is the time required to develop skill in the technique, and to relate it to the earlier approach (364D,
364E, 364K). The reward is a construction which is based directly on the algebra A, independent of any
representation (364A), and methods of dealing with it which are complementary to those of the previous
three sections. In particular, they can be used in the absence of the full axiom of choice (364Ya).
I have deliberately chosen the notation [[u > ]] from the theory of Forcing. I do not propose to try to
explain myself here, but I remark that much of the labour of this section is a necessary basis for understanding
real analysis in Boolean-valued models of set theory. The idea is that just as a function f : X R can
be described in terms of the sets x : f(x) > , so can an element u of L
0
(A) be described in terms of
the elements [[u > ]] of A where in some sense u is greater than . This description is well adapted to
discussion of the order struction of L
0
(A) (see 364M-364O), but rather ill-adapted to discussion of its linear
and multiplicative structures, which leads to a large part of the length of the work above. Once we have
succeeded in describing the algebraic operations on L
0
in terms of the values of [[u > ]], however, as in
364E, the fundamental result on the action of Boolean homomorphisms (364R) is elegant and reasonably
straightforward.
The concept [[u > ]] can be dramatically generalized to the concept [[(u
1
, . . . , u
n
) E]], where E is
a Borel subset of R
n
and u
1
, . . . , u
n
L
0
(A) (364H, 364Yc). This is supposed to recall the notation
Pr(X E), already used in Chapter 27. If, as sometimes seems reasonable, we wish to regard a random
variable as a member of L
0
() rather than of L
0
(), then [[u E]] is the appropriate translation of X
1
[E].
365B L
1
335
The reasons why we can reach all Borel sets E here, but then have to stop, seem to lie fairly deep. We see
that we have here another potential denition of L
0
(A), as the set of sequentially order-continuous Boolean
homomorphisms from the Borel -algebra of R to A. This is suitably independent of realizations of A, but
makes the f-algebra structure of L
0
dicult to elucidate, unless we move to a further level of abstraction
in the denitions, as in 364Ye.
I take the space to describe the L
0
spaces of general regular open algebras in detail (364U) partly to oer
a demonstration of an appropriate technique, and partly to show that we are not limited to -algebras of sets
and their quotients. This really is a new representation; for instance, it does not meld in any straightforward
way with the constructions of 364G-364I. Of course the most important examples are compact Hausdor
spaces, for which alternative methods are available (364V-364W, 364Yk, 364Yi, 364Yl); from the point of
view of applications, indeed, it is worth sorting out compact Hausdor spaces in general (364Yj). The
version in 364W is derived from Vulikh 67. But I have starred everything from 364T on, because I shall
not rely on this work later for anything essential.
365 L
1
Continuing my programme of developing the ideas of Chapter 24 at a deeper level of abstraction, I arrive
at last at L
1
. As usual, the rst step is to establish a denition which can be matched both with the
constructions of the previous sections and with the denition of L
1
() in 242 (365A-365C, 365F). Next,
I give what I regard as the most characteristic internal properties of L
1
spaces, including versions of the
Radon-Nikod ym theorem (365E), before turning to abstract versions of theorems in 235 (365H, 365T)
and the duality between L
1
and L

(365I-365K). As in 361 and 363, the construction is associated with

universal mapping theorems (365L-365N) which dene the Banach lattice structure of L
1
. As in 361, 363
and 364, homomorphisms between measure algebras correspond to operators between their L
1
spaces; but
now the duality theory gives us two types of operators (365O-365Q), of which one class can be thought of
as abstract conditional expectations (365R). For localizable measure algebras, the underlying algebra can
be recovered from its L
1
space (365S), but the measure cannot.
365A Denition Let (A, ) be a measure algebra. For u L
0
(A), write
|u|
1
=
_

0
[[[u[ > ]] d,
the integral being with respect to Lebesgue measure on R, and allowing as a value of the integral. (Because
the integrand is monotonic, it is certainly measurable.) Set L
1

= L
1
(A, ) = u : u L
0
(A), |u|
1
< .
It is convenient to note at once that if u L
1
(A, ), then [[[u[ > ]] must be nite for almost every
> 0, and therefore for every > 0, since it is a non-increasing function of ; so that [[u > ]] also belongs
to the Boolean ring A
f
= a : a < for every > 0.
365B Theorem Let (X, , ) be a measure space with measure algebra (A, ). Then the canonical
isomorphism between L
0
() and L
0
(A), dened in 364D and 364J, matches L
1
() L
0
(), dened in 242,
with L
1

L
0
(A), and the standard norm of L
1
() with | |
1
: L
1

[0, [, as dened in 365A.
proof Take any f L
0
= L
0
() (364C); write f

for its equivalence class in L

0
(), and u for the
corresponding element of L
0
(A), so that [[[u[ > ]] = x : [f(x)[ >

in A, for every R, and

|u|
1
=
_

0
x : [f(x)[ > d.
(a) If f is a non-negative simple function, it is expressible as

n
i=0

i
E
i
where E
0
, . . . , E
n
are disjoint
measurable sets of nite measure and
i
0 for each i; re-enumerating if necessary, we may suppose that

0

1
. . .
n
. In this case, setting
n+1
= 0,
336 Function spaces 365B
|u|
1
=
_
x : f(x) > d =
n

j=0
(
j

i=0
E
i
)(
j

j+1
)
=
n

i=0
n

j=i
E
i
(
j

j+1
) =
n

i=0

i
E
i
=
_
fd.
(b) If f is integrable, then there is a non-decreasing sequence f
n
)
nN
of non-negative simple functions
such that [f(x)[ = sup
nN
f
n
(x) a.e.; now
x : [f(x)[ > = sup
nN
x : f
n
(x) >
for every R, so
_
[f[ = sup
nN
_
f
n
= sup
nN
_

0
x : f
n
(x) > d
=
_

0
x : [f(x)[ > d = |u|
1
.
Thus in this case u L
1
(A, ) and |u|
1
= |f

|
1
.
It follows that every member of L
1
() corresponds to a member of L
1
(A, ), and that the norm of L
1
()
corresponds to | |
1
on L
1
(A, ).
(c) On the other hand, if u L
1
(A, ), then x : [f(x)[ > = [[[u[ > ]] is nite for every > 0
(365A). Also, if g is any simple function with 0 g [f[,
_
g d =
_

0
x : g(x) > d
_

0
x : f(x) > d = |u|
1
< .
So f is integrable (122J). This shows that every member of L
1
(A, ) corresponds to a member of L
1
().
365C Accordingly we can apply everything we know about L
1
() spaces to L
1

spaces. For instance:
Theorem For any measure algebra (A, ), L
1

= L
1
(A, ) is a solid linear subspace of L
0
(A), and | |
1
is
a norm on L
1

under which L
1

is an L-space. Consequently L
1

is a perfect Riesz space with an order-
continuous norm which has the Levi property, and if u
n
)
nN
is a non-decreasing norm-bounded sequence
in L
1

then it converges for | |
1
to sup
nN
u
n
.
proof (A, ) is isomorphic to the measure algebra of some measure space (X, , ) (321J). L
1
() is a solid
linear subspace of L
0
() (242Cb), so L
1

is a solid linear subspace of L
0
(A). L
1
() is an L-space (354M), so
L
1

also is. The rest of the properties claimed are general features of L-spaces (354N, 354E, 356P).
365D Integration Let (A, ) be any measure algebra.
(a) If u L
1
= L
1
(A, ), then u
+
, u

belong to L
1
, and we may set
_
u = |u
+
|
1
|u

|
1
=
_

0
[[u > ]] d
_

0
[[u > ]] d.
Now
_
: L
1
R is an order-continuous positive linear functional (356P), and under the translation of 365B
matches the integral on L
1
() as dened in 242Ab.
(b) Of course |u|
1
=
_
[u[ [
_
u[ for every u L
1
.
(c) If u L
1
, a A we may set
_
a
u =
_
u a. (Compare 242Ac.) If > 0 and 0 ,= a [[u > ]] then
there is a > such that a
t
= a [[u > ]] ,= 0, so that
_
a
u =
_

0
(a [[u > ]])d
_

0
a d +
_

a
t
> a.
In particular, setting a = [[u > ]], [[u > ]] must be nite.
365E L
1
337
(d) If u L
1
then u 0 i
_
a
u 0 for every a A
f
, writing A
f
= a : a < , as usual. PPP If u 0
then u a 0,
_
a
u 0 for every a A. If u , 0, then [[u

> 0]] ,= 0 and there is an > 0 such that

a = [[u

> ]] ,= 0. But now a is nite ((c) above) and

_
u a =
_
u

a =
_
(a [[u

]])d a < 0,
so
_
a
u < 0. QQQ If u, v L
1
and
_
a
u =
_
a
v for every a A
f
then u = v (cf. 242Ce). If u 0 in L
1
then
_
u = sup
_
a
u : a A
f
. PPP Of course u a u so
_
a
u u for every a A. On the other hand, setting
a
n
= [[u > 2
n
]], ua
n
)
nN
is a non-decreasing sequence with supremum u, so
_
u = lim
n
_
a
n
u, while
a
n
is nite for every n. QQQ
(e) If u L
1
, u 0 and
_
u = 0 then u = 0 (122Rc). If u L
1
, u 0 and
_
a
u = 0 then u a = 0,
that is, a [[u > 0]] = 0.
(f ) If C L
1
is non-empty and upwards-directed and sup
vC
_
v is nite, then sup C is dened in L
1
and
_
supC = sup
vC
_
v (356P).
(g) It will occasionally be convenient to adapt the conventions of 133 to the new context; so that I may
write
_
u = if u L
0
(A), u

L
1
, u
+
/ L
1
, while
_
u = if u
+
L
1
and u

/ L
1
.
(h) On this convention, we can restate (f) as follows: if C (L
0
)
+
is non-empty and upwards-directed
and has a supremum u in L
0
, then
_
u = sup
vC
_
v in [0, ]. PPP For if sup
vC
_
v is innite, then surely
_
u = ; while otherwise we can apply (f). QQQ
365E The Radon-Nikod ym theorem again (a) Let (A, ) be a semi-nite measure algebra and
: A R an additive functional. Then the following are equiveridical:
(i) there is a u L
1
= L
1
(A, ) such that a =
_
a
u for every a A;
(ii) is additive and continuous for the measure-algebra topology on A;
(iii) is completely additive.
(b) Let (A, ) be any measure algebra, and : A
f
R a function. Then the following are equiveridical:
(i) is additive and bounded and inf
aA
[a[ = 0 whenever A A
f
is downwards-directed and has
inmum 0;
(ii) there is a u L
1
such that a =
_
a
u for every a A
f
.
proof (a) The equivalence of (ii) and (iii) is 327Bd. The equivalence of (i) and (iii) is just a translation of
327D into the new context.
(b)(i)(ii)() Set M = sup
aA
f [a[ < .
Let D A
f
be a maximal disjoint set. For each d D, write A
d
for the principal ideal of A generated by
d, and
d
for the restriction of to A
d
, so that (A
d
,
d
) is a totally nite measure algebra. Set
d
= A
d
;
then
d
: A
d
R is completely additive. By (a), there is a u
d
L
1
(A
d
,
d
) such that
_
a
u
d
=
d
a = a for
every a d.
Now u
+
d
L
0
(A
d
) corresponds to a member u
+
d
of L
0
(A)
+
dened by saying
[[ u
+
d
> ]] = [[u
+
d
> ]] = [[u
d
> ]] if 0,
= 1 if 0.
Set d
+
= [[u
d
> 0]]. If a A, then
_
a
u
+
d
=
_

0
(a [[ u
+
d
> ]])d =
_

0
(a [[u
+
d
> ]])d =
_
ad
u
+
d
;
taking a = 1, we see that | u
+
d
|
1
= d
+
is nite, so that u
+
d
L
1

.
() For any nite I D, set v
I
=

dI
u
+
d
. Then
_
v
I
= (sup
dI
d
+
) M;
consequently the upwards-directed set A = v
I
: I D is nite is bounded above in L
1

, and we can set
v = sup A in L
1

. If a A, then
_
a
v
I
=

dI
_
ad
u
+
d
for each nite I D, so
_
a
v =

dD
_
ad
u
+
d
.
338 Function spaces 365E
Applying the same arguments to , there is a w L
1

such that
_
a
w =

dD
_
ad
u

d
for every a A. Try u = v w; then
_
a
u =

dD
_
ad
u
+
d

_
ad
u

d
=

dD
_
ad
u
d
=

dD
(a d)
for every a A.
() Now take any a A
f
. For J D set a
J
= sup
dJ
a d. Let > 0. Then there is a nite I D
such that
[
_
a
u a
J
[ = [

dD
(a d)

dJ
(a d)[
whenever I J D and J is nite. But now consider
A = a \ a
J
: I J D, J is nite.
Then inf A = 0, so there is a nite J such that I J D and
[a a
J
[ = [(a \ a
J
)[ .
Consequently
[a
_
a
u[ [a a
J
[ +[
_
a
u a
J
[ 2.
As is arbitrary, a =
_
a
u. As a is arbitrary, (ii) is proved.
(ii)(i) From where we now are, this is nearly trivial. Thinking of a as
_
ua, is surely additive
and bounded. Also [a[
_
[u[ a. If A A
f
is non-empty, downwards-directed and has inmum 0, the
same is true of [u[ a : a A, because a [u[ a is order-continuous, so
inf
aA
[a[ inf
aA
_
[u[ a = inf
aA
|[u[ a|
1
= 0.
365F It will be useful later to have spelt out the following elementary facts.
Lemma Let (A, ) be a measure algebra. Write S
f
for the intersection S(A)L
1
(A, ). Then S
f
is a norm-
dense and order-dense Riesz subspace of L
1
= L
1
(A, ), and can be identied with S(A
f
). The function
: A
f
S
f
L
1
is an injective order-continuous additive lattice homomorphism. If u 0 in L
1
, there is
a non-decreasing sequence u
n
)
nN
in (S
f
)
+
such that u = sup
nN
u
n
= lim
n
u
n
.
proof (a) As in 364L, we can think of S(A
f
) as a Riesz subspace of S = S(A), embedded in L
0
(A). If u S,
it is expressible as

n
i=0

i
a
i
where a
0
, . . . , a
n
A are disjoint and no
i
is zero. Now [u[ =

n
i=0
[
i
[a
i
,
so u L
1
i a
i
< for every i, that is, i u S(A
f
); thus S
f
= S(A
f
).
Now suppose that u 0 in L
1
. By 364Kd, there is a non-decreasing sequence u
n
)
nN
in S(A)
+
such
that u
0
0 and u = sup
nN
u
n
in L
0
. Because L
1
is a solid linear subspace of L
0
, every u
n
belongs to L
1
and therefore to S
f
. By 365C, u
n
)
nN
is norm-convergent to u. This shows also that S
f
is order-dense in
L
1
.
The map : A
f
S
f
is an injective order-continuous additive lattice homomorphism; because S
f
is
regularly embedded in L
1
, has the same properties when regarded as a map into L
1
.
For general u L
1
, there are sequences in S
f
converging to u
+
and to u

, so that their dierence is a

sequence in S
f
converging to u, and u belongs to the closure of S
f
; thus S
f
is norm-dense in L
1
.
Remark Of course S
f
here corresponds to the space of (equivalence classes of) simple functions.
365G Semi-nite algebras: Lemma Let (A, ) be a measure algebra.
(a) (A, ) is semi-nite i L
1
= L
1
(A, ) is order-dense in L
0
= L
0
(A).
(b) In this case, writing S
f
= S(A) L
1
(as in 365F),
_
u = sup
_
v : v S
f
, 0 v u in [0, ] for
every u (L
0
)
+
.
proof (a) If (A, ) is semi-nite then S
f
is order-dense in L
0
(364L), so L
1
must also be. If L
1
is order-dense
in L
0
, then so is S
f
, by 365F and 352Nc, so (A, ) is semi-nite, by the other half of 364L.
365J L
1
339
(b) Set C = v : v S
f
, 0 v u. Then C is an upwards-directed set with supremum u, because S
f
is order-dense in L
0
. So
_
u = sup
vC
_
v by 365Dh.
365H Measurable transformations We have a generalization of the ideas of 235 in this abstract
context.
Theorem Let (A, ) and (B, ) be measure algebras, and : A B a sequentially order-continuous
Boolean homomorphism. Let T : L
0
(A) L
0
(B) be the sequentially order-continuous Riesz homomorphism
associated with (364R).
(a) Suppose that w 0 in L
0
(B) is such that
_
a
wd = a whenever a A and a < . Then for any
u L
1
(A, ) and a A,
_
a
Tu wd is dened and equal to
_
ud .
(b) Suppose that w
t
0 in L
0
(A) is such that
_
a
w
t
d = (a) for every a A. Then
_
Tud =
_
uw
t
d
whenever u L
0
(A) and either integral is dened in [, ].
Remark I am using the convention of 365Dg concerning as the value of an integral, and the notation

_
ud is supposed to indicate that I am considering the integral in L
1
(A, ).
proof (a) If u S
f
= L
1
(A, ) S(A) then u is expressible as

n
i=0

i
a
i
where a
0
, . . . , a
n
have nite
measure, so that Tu =

n
i=0

i
(a
i
) and
_
Tu wd =

n
i=0

i
_
a
i
w =

n
i=0

i
a
i
=
_
ud .
If u 0 in L
1
(A, ) there is a non-decreasing sequence u
n
)
nN
in S
f
with supremum u, so that Tu =
sup
nN
Tu
n
and w Tu = sup
nN
w Tu
n
in L
0
(B), and
_
Tu w = sup
nN
_
Tu
n
w = sup
nN
_
u
n
=
_
u.
(365C tells us that in this context Tu w L
1
(B, ).) Finally, for general u L
1
(A, ),
_
Tu w =
_
Tu
+
w
_
Tu

w =
_
u
+

_
u

=
_
u.
(b) The argument follows the same lines: start with u = a for a A, then with u S(A), then with
u L
0
(A)
+
and conclude with general u L
0
(A). The point is that T is a Riesz homomorphism, so that
at the last step
_
Tu =
_
(Tu)
+

_
(Tu)

=
_
T(u
+
)
_
T(u

)
=
_
u
+
w
t

_
u

w
t
=
_
(u w
t
)
+

_
(u w
t
)

=
_
u w
t
whenever either side is dened in [, ].
365I The duality between L
1
and L

Let (A, ) be a measure algebra. If we identify L

= L

(A)
with the solid linear subspace of L
0
= L
0
(A) generated by e = 1
A
(364K), then we have a bilinear map
(u, v) u v : L
1
L

L
1
, because [u v[ |v|

[u[ and L
1
= L
1
(A, ) is a solid linear subspace of
L
0
. Note that |u v|
1
|u|
1
|v|

, so that the bilinear map (u, v) u v has norm at most 1 (253A).

Consequently we have a bilinear functional (u, v)
_
uv : L
1
L

R, which also has norm at most 1,

corresponding to linear operators S : L
1
(L

and T : L

(L
1
)

, both of norm at most 1. Because

L
1
and L

are both Banach lattices, we have (L

1
)

= (L
1
)

, (L

= (L

(356Dc). Because the norm

of L
1
is order-continuous, (L
1
)

= (L
1
)

(356Dd).
365J Theorem Let (A, ) be a measure algebra, and set L
1
= L
1
(A, ), L

= L

(A). Let S : L
1

(L

= (L

, T : L

(L
1
)

= (L
1
)

= (L
1
)

be the canonical maps dened by the duality between

L
1
and L

, as in 365I. Then
(a) S and T are order-continuous Riesz homomorphisms, S[L
1
] (L

, and S is norm-preserving;
(b) (A, ) is semi-nite i T is injective, and in this case T is norm-preserving, while S is a normed Riesz
space isomorphism between L
1
and (L

;
340 Function spaces 365J
(c) (A, ) is localizable i T is bijective, and in this case T is a normed Riesz space isomorphism between
L

and (L
1
)

= (L
1
)

.
proof Take a measure space (X, , ) such that (A, ) is isomorphic to its measure algebra. Then we can
identify L
1
with L
1
(), as in 365B, and L

with L

(), as in 363I. Moreover, these identications are based

on the canonical embeddings of L
1
and L

in L
0
(), so that the duality described in 365I corresponds to
the familiar duality (u, v)
_
u v already used in 243F.
(a)(i) If u 0 in L
1
and v 0 in L

then u v 0 and
(Tv)(u) =
_
u v 0.
As u is arbitrary, Tv 0 in (L
1
)

; as v is arbitrary, T is a positive linear operator.

If v L

, set a = [[v > 0]] A. (Remember that we are identifying L

0
(), as dened in 241, with
L
0
(A), as dened in 364.) Then v
+
= v a, so for any u 0 in L
1
(Tv
+
)(u) =
_
u v a = (Tv)(u a) (Tv)
+
(u).
As u is arbitrary, Tv
+
(Tv)
+
. On the other hand, because T is a positive linear operator, Tv
+
Tv and
Tv
+
0, so Tv
+
(Tv)
+
. Thus Tv
+
= (Tv)
+
. As v is arbitrary, T is a Riesz homomorphism (352G).
(ii) Exactly the same arguments show that S is a Riesz homomorphism.
(iii) Given u L
1
, set a = [[u > 0]]; then
|Su| (Su)(a (1 \ a)) =
_
a
u
_
1\a
u =
_
[u[ = |u|
1
|Su|.
So S is norm-preserving.
(iv) By 355Ka, S is order-continuous.
(v) If A L

is a non-empty downwards-directed set with inmum 0, and u (L

1
)
+
, then inf
vA
u
v = 0 for every u (L
1
)
+
, because v u v : L
0
L
0
is order-continuous. So
inf
vA
(Tv)(u) = inf
vA
_
u v = inf
vA
|u v|
1
= 0
and the only possible non-negative lower bound for T[A] in (L
1
)

is 0. As A is arbitrary, T is order-
continuous.
(vi) The ideas of (v) show also that S[L
1
] (L

. PPP If u (L
1
)
+
and A L

is non-empty,
downwards-directed and has inmum 0, then
inf
vA
(Su)(v) = inf
vA
_
u v = 0.
As A is arbitrary, Su is order-continuous. For general u L
1
, Su = Su
+
Su

belongs to (L

. QQQ
(b)(i) If (A, ) is not semi-nite, let a A be such that a = and b = whenever 0 ,= b a. If
u L
1
, then [[[u[ >
1
n
]] has nite measure for every n 1, so must be disjoint from a; accordingly
a [[[u[ > 0]] = sup
n1
a [[[u[ >
1
n
]] = 0.
This means that
_
u a = 0 for every u L
1
. Accordingly T(a) = 0 and T is not injective.
(ii) If (A, ) is semi-nite, take any v L

. Then if 0 < |v|

, a = [[[v[ > ]] ,= 0. Let b a be

such that 0 < b < . Then b L
1
, and
|Tv| = |[Tv[| = |T[v[| (T[v[)(b)/|b|
1

because [v[ b b, so
(T[v[)(b) b = |b|
1
.
As is arbitrary, |Tv| |v|

. But we already know that |Tv| |v|

, so the two are equal. As v is

arbitrary, T is norm-preserving (and, in particular, is injective).
(iii) Still supposing that (A, ) is semi-nite, S[L
1
] = (L

. PPP Take any h (L

. For a A, set
a = h(a

). By 363K, : A R is completely additive. By 365Ea, there is a u L

1
such that
365M L
1
341
(Su)(a) =
_
u a =
_
a
u = a = h(a)
for every a A. Because Su and h are both linear functionals on L

, they must agree on S(A); because

they are continuous and S(A) is dense in L

(363C), Su = h. As h is arbitrary, S is surjective. QQQ

(c) Using (b), we know that if either T is bijective or (A, ) is localizable, then (A, ) is semi-nite,
so that (X, , ) is also semi-nite (322Bd). Given this, 243G tells us that T is bijective i (X, , ) is
localizable, and in this case T is norm-preserving; but of course (X, , ) is localizable i (A, ) is (322Be).
365K Corollary If (A, ) is a localizable measure algebra, L

(A) is perfect.
proof By 365J(b)-(c), we can identify L

with (L
1
)

= (L

.
365L Theorem Let (A, ) be a measure algebra and U a Banach space. Let : A
f
U be a function.
Then the following are equiveridical:
(i) there is a continuous linear operator T : L
1
U, where L
1
= L
1
(A, ), such that a = T(a) for
every a A
f
;
(ii)() is additive () there is an M 0 such that |a| M a for every a A
f
.
Moreover, in this case, T is unique and |T| is the smallest number M satisfying the condition in (ii-).
proof (a)(i)(ii) If T : L
1
U is a continuous linear operator, then a L
1
for every a A
f
, so = T
is a function from A
f
to U. If a, b A
f
and a b = 0, then (a b) = a +b in L
0
= L
0
(A) and therefore
in L
1
, so
(a b) = T(a b) = T(a +b) = T(a) +T(b) = a +b.
If a A
f
then |a|
1
= a (using the formula in 365A, or otherwise), so
|a| = |T(a)| |T||a|
1
= |T| a.
(b)(ii)(i) Now suppose that : A
f
U is additive and that |a| M a for every a A
f
. Let
S
f
= L
1
S(A), as in 365F. Then there is a linear operator T
0
: S
f
U such that T
0
(a) = a for every
a A
f
(361F). Next, |T
0
u| M|u|
1
for every u S
f
. PPP If u S
f
= S(A
f
), then u is expressible as

m
j=0

j
b
j
where b
0
, . . . , b
m
A
f
are disjoint (361Eb). So
|T
0
u| = |

m
j=0

j
b
j
| M

m
j=0
[
j
[ b
j
= M|u|
1
. QQQ
There is therefore a continuous linear operator T : L
1
U, extending T
0
, and with |T| |T
0
| M
(2A4I). Of course we still have = T.
(c) The argument in (b) shows that T
0
= TS
f
and T are uniquely dened from . We have also seen
that if T, correspond to each other then
|a| |T| a for every a A
f
,
|T| M whenever |a| M a for every a A
f
,
so that
|T| = minM : M 0, |a| M a for every a A
f
.
365M Corollary Let (X, , ) be a measure space and U any Banach space. Set
f
= E : E
, E < . Let :
f
U be a function. Then the following are equiveridical:
(i) there is a continuous linear operator T : L
1
U, where L
1
= L
1
(), such that E = T(E)

for
every E
f
;
(ii)() (E F) = E + F whenever E, F
f
and E F = 0 () there is an M 0 auch that
|E| ME for every E
f
.
Moreover, in this case, T is unique and |T| is the smallest number M satisfying the condition in (ii-).
proof This is a direct translation of 365L. The only point to note is that if satises the conditions of
(ii), and E, F
f
are such that E

= F

in the measure algebra (A, ) of (X, , ), then (E F) =

(F E) = 0, so that (E F) = (F E) = 0 (using condition (ii-)) and
342 Function spaces 365M
E = (E F) +(E F) = (E F) +(F E) = F.
This means that we have a function : A
f
U, where
A
f
= a : a A, a < = E

: E
f
,
dened by setting E

= E for every E
f
. Of course we now have (a b) = a + b whenever a,
b A
f
and a b = 0 (since we can express them as a = E

, b = F

with E F = ), and | a| M a
for every a A
f
. Thus we have a one-to-one correspondence between functions :
f
U satisfying
the conditions (ii) here, and functions : A
f
U satisfying the conditions (ii) of 365L. The rest of the
argument is covered by the identication between L
1
() and L
1
(A, ) in 365B.
365N Theorem Let (A, ) be a measure algebra, U a Banach lattice, and T : L
1
U a bounded linear
operator, where L
1
= L
1
(A, ). Let : A
f
U be the corresponding additive function, as in 365L.
(a) T is a positive linear operator i a 0 in U for every a A
f
; in this case, T is order-continuous.
(b) If U is Dedekind complete and T L

(L
1
; U), then [T[ : L
1
U corresponds to [[ : A
f
U, where
[[(a) = sup

n
i=0
[a
i
[ : a
0
, . . . , a
n
a are disjoint
for every a A
f
.
(c) T is a Riesz homomorphism i is a lattice homomorphism.
proof As in 365F, let S
f
be L
1
S(A), identied with S(A
f
).
(a)(i) If T is a positive linear operator and a A
f
, then a 0 in L
1
, so a = T(a) 0 in U.
(ii) Now suppose that a 0 in U for every a A
f
, and let u 0 in L
1
, > 0 in R. Then there is a
v S
f
such that 0 v u and |u v|
1
(365F). Express v as

n
i=0

i
a
i
where a
i
A
f
,
i
0 for
each i. Now
|Tu Tv| |T||u v|
1
|T|.
On the other hand,
Tv =

n
i=0

i
a
i
U
+
.
As U
+
is norm-closed in U, and is arbitrary, Tu U
+
. As u is arbitrary, T is a positive linear operator.
(iii) By 355Ka, T is order-continuous.
(b) The point is that [TS
f
[ = [T[S
f
. PPP (i) Because the embedding S
f
L
1
is positive, the map
P PS
f
is a positive linear operator from L

(L
1
; U) to L

(S
f
; U) (see 355Bd). So [TS
f
[ [T[S
f
. (ii)
There is a positive linear operator T
1
: L
1
U extending [TS
f
[, by 365M and (a) above, and now T
1
S
f
dominates both TS
f
and TS
f
; since (S
f
)
+
is dense in (L
1
)
+
, T
1
T and T
1
T, so that T
1
[T[
and
[TS
f
[ = T
1
S
f
[T[S
f
. QQQ
Now 361H tells us that
[T[(a) = [TS
f
[(a) = [[a
for every a A
f
.
(c)(i) If T is a lattice homomorphism, then so is = T, because : A
f
S
f
is a lattice homomorphism.
(ii) Now suppose that is a lattice homomorphism. In this case TS
f
is a Riesz homomorphism
(361Gc), that is, [Tv[ = T[v[ for every v S
f
. Because S
f
is dense in L
1
and the map u [u[ is continuous
both in L
1
and in U (354Bb), [Tu[ = T[u[ for every u L
1
, and T is a Riesz homomorphism.
365O Theorem Let (A, ) and (B, ) be measure algebras. Let : A
f
B
f
be a measure-preserving
ring homomorphism.
(a) There is a unique order-continuous norm-preserving Riesz homomorphism T

: L
1

L
1

such that
T

(a) = (a) whenever a A

f
. We have T

(u a) = T

u (a) whenever a A
f
and u L
1

.
(b)
_
T

u =
_
u,
_
a
T

u =
_
a
u for every u L
1

, a A
f
.
365O L
1
343
(c) [[T

u > ]] = [[u > ]] for every u L

1
(A, ), > 0.
(d) T

is surjective i is.
(e) If (C,

) is another measure algebra and : B

f
C another measure-preserving Boolean homomor-
phism, then T

= T

: L
1

L
1
(C,

).
proof Throughout the proof I will write T for T

and S
f
for S(A) L
1

= S(A
f
) (see 365F).
(a)(i) We have a map : A
f
L
1

dened by writing a = (a) for a A
f
. Because
(a b) = (a b) = a +b, |(a)|
1
= (a) = a
whenever a, b A
f
and a b = 0, we get a (unique) corresponding bounded linear operator T : L
1

L
1

such that T = on A
f
(365L). Because : A
f
B
f
and : B
f
L
1

are lattice homomorphisms, so
is , and T is a Riesz homomorphism (365Nc).
(ii) If u S
f
, express u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint in A
f
. Then Tu =

n
i=0

i
(a
i
)
and a
0
, . . . , a
n
are disjoint in B
f
, so
|Tu|
1
=

n
i=0
[
i
[ (a
i
) =

n
i=0
[
i
[ a
i
= |u|
1
.
Because S
f
is dense in L
1

and u |u|
1
is continuous (in both L
1

and L
1

), |Tu|
1
= |u|
1
for every u L
1

,
that is, T is norm-preserving. As noted in 365Na, T is order-continuous.
(iii) If a, b A
f
then
T(a b) = T((a b)) = (a b) = (a b) = a b = a T(b).
Because T is linear and is bilinear, T(a u) = a Tu for every u S
f
. Because the maps
u u a : L
1

L
1

, T : L
1

L
1

and v v a : L
1

L
1

are all continuous, Tu a = T(u a)
for every u L
1

.
(iv) T is unique because the formula T(a) = a denes T on the norm-dense and order-dense
subspace S
f
.
(b) Because T is positive,
_
Tu = |Tu
+
|
1
|Tu

|
1
= |u
+
|
1
|u

|
1
=
_
u.
For a A
f
,
_
a
Tu =
_
Tu a =
_
T(u a) =
_
u a =
_
a
u.
(c) If u S
f
, express it as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint; then
[[u > ]] = (sup
iI
a
i
) = sup
iI
a
i
= [[Tu > ]]
where I = i : i n,
i
> . For u (L
1
)
+
, take a sequence u
n
)
nN
in S
f
with supremum u; then
sup
nN
Tu
n
= Tu, so
[[u > ]] = (sup
nN
[[u
n
> ]])
(364Mb; [[u > ]] A
f
by 365A)
= sup
nN
[[u
n
> ]]
(because is order-continuous, see 361Ad)
= sup
nN
[[Tu
n
> ]] = [[Tu > ]]
because T is order-continuous. For general u L
1
,
[[u > ]] = [[u
+
> ]] = [[T(u
+
) > ]] = [[(Tu)
+
> ]] = [[Tu > ]]
because T is a Riesz homomorphism.
344 Function spaces 365O
(d)(i) Suppose that T is surjective and that b B
f
. Then there is a u L
1

such that Tu = b. Now
b = [[Tu >
1
2
]] = [[u >
1
2
]] [A
f
];
as b is arbitrary, is surjective.
(ii) Suppose now that is surjective. Then T[L
1

] is a linear subspace of L
1

containing b for every
b B
f
, so includes S(B
f
). If v (L
1

)
+
there is a sequence v
n
)
nN
in S(B
f
)
+
with supremum v. For each
n, choose u
n
such that Tu
n
= v
n
. Setting u
t
n
= sup
in
u
i
, we get a non-decreasing sequence u
t
n
)
nN
such
that v
n
Tu
t
n
v for every n N. So
sup
nN
|u
t
n
|
1
= sup
nN
|Tu
t
n
|
1
|v|
1
<
and u = sup
nN
u
t
n
is dened in L
1

, with
Tu = sup
nN
Tu
t
n
= v.
Thus (L
1

)
+
T[L
1

]; consequently L
1

T[L
1

] and T is surjective.
(e) This is an immediate consequence of the uniqueness assertion in (i), because for any a A
f
T

(a) = T

(a) = (a),
so that T

: L
1

L
1

is a bounded linear operator taking the right values at elements a, and must
therefore be equal to T

.
365P Theorem Let (A, ) and (B, ) be measure algebras, and : A
f
B an order-continuous ring
homomorphism.
(a) There is a unique positive linear operator P

: L
1

L
1

such that
_
a
P

v =
_
a
v for every v L
1

,
a A
f
.
(b) P

is order-continuous and norm-continuous, and |P

| 1.
(c) If a A
f
, v L
1

then P

(v a) = P

v a.
(d) If [A
f
] is order-dense in B then P

is a norm-preserving Riesz homomorphism; in particular, P

is
injective.
(e) If (B, ) is semi-nite and is injective, then P

is surjective, and there is for every u L

1

a v L
1

such that P

v = u and |v|
1
= |u|
1
.
(f) Suppose again that (B, ) is semi-nite. If (C,

) is another measure algebra and : B C an order-

continuous Boolean homomorphism, then P

= P

: L
1
(C,

) L
1

, where I write
t
for the restriction
of to B
f
.
proof I write P for P

.
(a)-(b) For v L
1

, a A
f
set
v
(a) =
_
a
v. Then
v
: A
f
R is additive, bounded (by |v|
1
) and if
A A
f
is non-empty, downwards-directed and has inmum 0, then
inf
aA
[
v
(a)[ inf
aA
_
[v[ a = 0
because a
_
[v[ a is a composition of order-continuous functions, therefore order-continuous. So
365Eb tells us that there is a Pv L
1

such that
_
a
Pv =
v
(a) =
_
a
v for every a A
f
. By 365Dd, this
formula denes Pv uniquely. Consequently P must be linear (since Pv
1
+ Pv
2
, Pv will always have the
properties dening P(v
1
+v
2
), P(v)).
If v 0 in L
1

, then
_
a
Pv =
_
a
v 0 for every a A
f
, so Pv 0 (365Dd); thus P is positive. It must
therefore be norm-continuous and order-continuous (355C, 355Ka).
Again supposing that v 0, we have
|Pv|
1
=
_
Pv = sup
aA
f
_
a
Pv = sup
aA
f
_
a
v |v|
1
(using 365Dd). For general v L
1

,
|Pv|
1
= |[Pv[|
1
|P[v[|
1
|v|
1
.
(c) For any c A
f
,
365P L
1
345
_
c
Pv a =
_
ca
Pv =
_
(ca)
v =
_
c
v a =
_
c
P(v a).
(d) Now suppose that [A
f
] is order-dense. Take any v, v
t
L
1

such that v v
t
= 0. ??? Suppose, if
possible, that u = Pv Pv
t
> 0. Take > 0 such that a = [[u > ]] is non-zero. Since
_
a
v =
_
a
Pv
_
a
u > 0,
b = a [[v > 0]] ,= 0. Let c A
f
be such that 0 ,= c b; then (a c) = c ,= 0, so a c ,= 0, and
0 <
_
ac
u
_
ac
Pv
t

_
c
v
t
.
But c [[v > 0]] and v v
t
= 0 so
_
c
v
t
= 0. XXX
So Pv Pv
t
= 0. As v, v
t
are arbitrary, P is a Riesz homomorphism (352G).
Next, if v 0 in L
1

,
_
Pv = sup
aA
f
_
a
Pv = sup
aA
f
_
a
v =
_
v
because [A
f
] is upwards-directed and has supremum 1 in B. So, for general v L
1

,
|Pv|
1
=
_
[Pv[ =
_
P[v[ =
_
[v[ = |v|
1
,
and P is norm-preserving.
(e) Next suppose that (B, ) is semi-nite and that is injective.
(i) If u > 0 in L
1

, there is a v > 0 in L
1

such that Pv u and
_
Pv
_
v. PPP Let > 0 be such that
a = [[u > ]] ,= 0. Then a ,= 0. Because (B, ) is semi-nite, there is a non-zero b B
f
such that b a.
Set u
1
= P(b). Then u
1
0, and also
_
a
u
1
= b > 0,
_
1\a
u
1
= sup
cA
f
_
c\a
u
1
= sup
cA
f
_
c\a
b = 0.
So u
1
(1 \ a) = 0 and 0 ,= [[u
1
> 0]] a. Let > 0 be such that [[u
1
> ]] ,= [[u
1
> 0]], and set a
1
=
a \ [[u
1
> ]], v =
1
(b a
1
). Then
Pv =
1
P(b (a
1
)) =
1
P(b) a
1
=
1
u
1
a
1
a u
because
[[u
1
a
1
> ]] [[u
1
> ]] a
1
= 0
so
u
1
a
1
[[u
1
> 0]] a.
Also Pv > 0 because a
1
[[u
1
> 0]] ,= 0, so v ,= 0; and
_
Pv
_
a
1
Pv =
_
a
1
v =
_
v. QQQ
(ii) Now take any u 0 in L
1

, and set B = v : v L
1

, v 0, Pv u,
_
v
_
Pv. Let C B be a
maximal upwards-directed set (applying Zorns Lemma to the family P of all upwards-directed subsets of
B). We have
sup
vC
_
v sup
vC
_
Pv |u|
1
,
so v
0
= sup C is dened in L
1

(365Df). Because P is order-continuous, Pv
0
= sup P[C] u, and
_
Pv
0
= sup
vC
_
Pv sup
vC
_
v =
_
v
0
.
??? Suppose, if possible, that Pv
0
,= u. In this case, by (), there is a v
1
> 0 such that Pv
1
u Pv
0
,
_
v
1

_
Pv
1
. In this case, v
0
+ v
1
B, so C
t
= C v
0
+ v
1
is an upwards-directed subset of B strictly
larger than C, which is impossible. XXX Thus Pv
0
= u; also
|v
0
|
1
=
_
v
0

_
Pv
0
= |Pv
0
|
1
.
(iii) Now take any u L
1

. By (ii), there are non-negative v
1
, v
2
L
1

such that Pv
1
= u
+
, Pv
2
= u

,
|v
1
|
1
|u
+
|
1
, |v
2
|
1
|u

|
1
. Setting v = v
1
v
2
, we have Pv = u. Also we must have
346 Function spaces 365P
|v|
1
|v
1
|
1
+|v
2
|
1
|u
+
|
1
+|u

|
1
= |u|
1
|v|
1
,
so |v|
1
= |u|
1
, as required.
(f ) As usual, this is a consequence of the uniqueness of P. However (because I do not assume that
[A
f
] B
f
) there is an extra renement: we need to know that
_
b
P

w =
_
b
w for every b B, w L
1

.
PPP Because is order-continuous and (B, ) is semi-nite, b = supb
t
: b
t
B
f
, b
t
b, so if w 0 then
_
b
w = sup
b

B
f
,b

b
_
b

w = sup
b

B
f
,b

b
_
b

w =
_
b
P

w.
Expressing w as w
+
w

, we see that the same is true for every w L

1

. QQQ
Now we can say that PP

is a positive linear operator from L

1

to L
1

such that
_
a
PP

w =
_
a
P

w =
_
a
w =
_
a
P

w
whenever a A
f
, w L
1

.
365Q Proposition Let (A, ) and (B, ) be measure algebras and : A
f
B
f
a measure-preserving
ring homomorphism.
(a) In the language of 365O-365P above, P

is the identity operator on L

1
(A, ).
(b) If is surjective (so that it is an isomorphism between A
f
and B
f
) then P

= T
1

= T

1,
T

= P
1

= P

1.
proof (a) If u L
1

, a A
f
then
_
a
P

u =
_
a
T

u =
_
a
u.
So u = P

u, by 365Dd.
(b) From 365Od, we know that T

is surjective, while P

is the identity, so that P

= T
1

, T

= P
1

.
As for T

1, 365Oe tells us that T

1 = T
1

; so
P

1 = T
1

1
= T

.
365R Conditional expectations It is a nearly universal rule that any investigation of L
1
spaces must
include a look at conditional expectations. In the present context, they take the following form.
(a) Let (A, ) be a probability algebra and B a closed subalgebra; write for the restriction B. The
identity map from B to A induces operators T : L
1

L
1

and P : L
1

L
1

. If we take L
0
(A) to be dened
as the set of functions from R to A described in 364A, then L
0
(B) becomes a subset of L
0
(A) in the literal
sense, and T is actually the identity operator associated with the subset L
1

L
1

; L
1

is a norm-closed
and order-closed Riesz subspace of L
1

. P is a positive linear operator, while PT is the identity, so P is a
projection from L
1

onto L
1

. P is dened by the familiar formula
_
b
Pu =
_
b
u for every u L
1

, b B,
so is the conditional expectation operator in the sense of 242J.
(b) Just as in 233I-233J and 242K, we have a fundamental inequality concerning convex functions: if
h : R R is a convex function and u L
1

, then h(
_
u)
_

h(u); and if

h(u) L
1

(364I), then

h(Pu) P(

h(u)). PPP I repeat the proof of 233I-233J. For each q Q, take

q
R such that h(t) h
q
(t) =
h(q) +
q
(t q) for every t R, so that h(t) = sup
qQ
h
q
(t) for every t R, and

h(u) = sup
qQ

h
q
(u) for
every u L
0
= L
0
(A). (This is because
[[

h(u) > ]] = [[u h

1
[ ], [ ]]] = [[u

qQ
h
1
q
[ ], [ ]]]
= sup
qQ
[[u h
1
q
[ ], [ ]]] = sup
qQ
[[

h
q
(u) > ]]
for every R.) But setting e = 1, we see that

h
q
(u) = h(q)e +
q
(u qe) for every u L
0
, so that
_
h
q
(u) = h(q) +
q
(
_
u q) = h
q
(
_
u),
365T L
1
347
P(

h
q
(u)) = h(q)e +
q
(Pu qe) =

h
q
(Pu)
because
_
e = 1 and Pe = e. Taking the supremum over q, we get
h(
_
u) = sup
qQ
h
q
(
_
u) = sup
qQ
_

h
q
(u)
_

h(u),
and if

h(u) L
1

then

h(Pu) = sup
qQ

h
q
(Pu) = sup
qQ
P(

h
q
(u)) P(

h(u)). QQQ
Of course the result in this form can also be deduced from 233I-233J if we represent (A, ) as the measure
algebra of a probability space (X, , ) and set T = E : E , E

B.
(c) I note here a fact which is occasionally useful. If u L
1

is non-negative, then [[Pu > 0]] =
upr([[u > 0]], C), the upper envelope of [[u > 0]] in C as dened in 314V. PPP We have only to observe that, for
c C,
c [[Pu > 0]] = 0 c Pu = 0
_
c
Pu = 0
_
c
u = 0 c [[u > 0]] = 0.
Taking complements, c [[Pu > 0]] i c [[u > 0]]. QQQ
365S Recovering the algebra: Proposition (a) Let (A, ) be a localizable measure algebra. Then
A is isomorphic to the band algebra of L
1
= L
1
(A, ).
(b) Let A be a Dedekind -complete Boolean algebra, and , two measures on A such that (A, ) and
(A, ) are both semi-nite measure algebras. Then L
1
(A, ) is isomorphic, as Banach lattice, to L
1
(A, ).
proof (a) Because (A, ) is semi-nite, L
1
is order-dense in L
0
= L
0
(A) (365G). Consequently, L
1
and L
0
have isomorphic band algebras (353D). But the band algebra of L
0
is just its algebra of projection bands
(because A and therefore L
0
are Dedekind complete, see 364O and 353I), which is isomorphic to A (364Q).
(b) Let : A A be the identity map. Regarding as an order-continuous Boolean homomorphism
from A
f

= a : a < to (A, ), we have an associated positive linear operator P = P

: L
1

L
1

;
similarly, we have Q = P

1 : L
1

L
1

, and both P and Q have norm at most 1 (365Pb). Now 365Pf
assures us that QP is the identity operator on L
1

and PQ is the identity operator on L
1

. So P and Q are
the two halves of a Banach lattice isomorphism between L
1

and L
1

.
365T Having opened the question of varying measures on a single Boolean algebra, this seems an
appropriate moment for a general description of how they interact.
Proposition Let A be a Dedekind complete Boolean algebra, and : A [0, ], : A [0, ] two
functions such that (A, ) and (A, ) are both semi-nite (therefore localizable) measure algebras.
(a) There is a unique u L
0
= L
0
(A) such that (if we allow as a value of the integral)
_
a
ud = a
for every a A.
(b) For every v L
1

,
_
v d =
_
u v d .
(c) u is strictly positive (i.e., [[u > 0]] = 1) and, writing
1
u
for the multiplicative inverse of u,
_
a
1
u
d = a
for every a A.
proof (a) Because (A, ) is semi-nite, there is a partition of unity D A such that d < for every
d D. For each d D, the functional a (a d) : A R is completely additive, so there is a
u
d
L
1

such that
_
a
u
d
d = (a d) for every a A. Because
_
a
u
d
d 0 for every a, u
d
0. Because
_
1\d
u
d
= 0, [[u
d
> 0]] d. Now u = sup
dD
u
d
is dened in L
0
. PPP (See 368K below.) For n N,
set c
n
= sup
dD
[[u
d
> n]]. If d, d
t
D are distinct, then d [[u
d
> n]] = 0, so d c
n
= [[u
d
> n]]. Set
c = inf
nN
c
n
. If d D, then
d c = inf
nN
d c
n
= inf
nN
[[u
d
> n]] = 0.
But c c
0
supD, so c = 0. By 364Ma, u
d
: d D is bounded above in L
0
, so has a supremum, because
L
0
is Dedekind complete, by 364O. QQQ
For nite I D set u
I
=

dI
u
d
= sup
dI
u
d
(because u
d
u
c
= 0 for distinct c, d D). Then
u = sup u
I
: I D, I is nite. So, for any a A,
348 Function spaces 365T
_
a
ud = sup
ID is nite
_
a
u
I
d
(365Dh)
= sup
ID is nite

dI
_
a
u
d
d = sup
ID is nite

dI
(a d) = a.
Note that if a A is non-zero, then a > 0, so a [[u > 0]] ,= 0; consequently [[u > 0]] = 1.
To see that u is unique, observe that if u
t
has the same property then for any d D
_
a
(u d)d = (a d) =
_
a
(u
t
d)
for every a A, so that u d = u
t
d; because sup D = 1 in A, u must be equal to u
t
.
(b) Use 365Hb, with and T the identity maps.
(c) In the same way, there is a w L
0
such that
_
a
wd = a for every a A. To relate u and w,
observe that applying 365Hb we get
_
w a ud =
_
w a d
for every a A, that is,
_
a
w ud = a for every a. But from this we see that w u b = b at least
when b < , so that w u = 1 is the multiplicative identity of L
0
, and w =
1
u
.
365X Basic exercises (a) Let (A, ) be a measure algebra, and u L
1

. Show that
_
u =
_

0
[[u > ]] d
_
0

(1 \ [[u > ]]) d.

>>>(b) Let (A, ) be any measure algebra, and u L
1

. (i) Show that |u|
1
2 sup
aA
f [
_
a
u[. (Hint:
246F.) (ii) Show that for any > 0 there is an a A
f
such that [
_
u
_
b
u[ whenever a b A.
>>>(c) Let U be an L-space. If u
n
)
nN
is any norm-bounded sequence in U
+
, show that liminf
n
u
n
=
sup
nN
inf
mn
u
m
is dened in U, and that
_
liminf
n
u
n
liminf
n
_
u
n
.
(d) Let U be an L-space. Let T be a lter on U
+
such that u : u 0, |u| k belongs to T for some
k N. Show that u
0
= sup
FJ
inf F is dened in U, and that
_
u
0
sup
FJ
inf
uF
_
u.
(e) Let (A, ) be a measure algebra and A L
1

a non-empty set. Show that A is bounded above in L
1

i
sup

n
i=0
_
a
i
u
i
: a
0
, . . . , a
n
is a partition of unity in A, u
0
, . . . , u
n
A
is nite, and that in this case the supremum is
_
supA. (Hint: given u
0
, . . . , u
n
A, set b
ij
= [[u
i
u
j
]],
b
i
= sup
j,=i
b
ij
, a
i
= b
i
\ sup
j<i
b
j
, and show that
_
sup
in
u
i
=

n
i=0
_
a
i
u
i
.)
(f ) Let (A, ) be any measure algebra and : A
f
R a bounded additive function. Show that the
following are equiveridical: (i) there is a u L
1

such that a =
_
a
u for every a A
f
; (ii) for every > 0
there is a > 0 such that [a[ whenever a ; (iii) for every > 0, c A
f
there is a > 0 such that
a whenever a c and a ; (iv) for every > 0 there are c A
f
, > 0 such that [a[ whenever
a A
f
and (a c) ; (v) lim
n
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence in A
f
with
inmum 0.
(g) Let (A, ) and (B, ) be measure algebras, and : A B a sequentially order-continuous Boolean
homomorphism. Let T : L
0
(A) L
0
(B) be the Riesz homomorphism associated with (364R). Suppose
that w 0 in L
0
(B) is such that
_
a
wd = a whenever a A. Show that for any u L
0
(A, ),
_
Tu wd =
_
ud whenever either is dened in [, ].
365 Notes L
1
349
>>>(h) Let (A, ) be a measure algebra and a A; write A
a
for the principal ideal it generates. Show that
if is the identity embedding of A
f
A
a
into A
f
, then T

, as dened in 365O, identies L

1
(A
a
, A
a
) with
a band in L
1
(A, ).
>>>(i) Let (X, , ) and (Y, T, ) be measure spaces, with measure algebras (A, ) and (B, ). Let :
X Y be an inverse-measure-preserving function and : B A the corresponding measure-preserving
homomorphism (324M). Show that T

: L
1

L
1

(365O) corresponds to the map g

(g)

: L
1
()
L
1
() of 242Xf.
(j) Let (A, ) and (B, ) be measure algebras. Let : A
f
B
f
be a ring homomorphism such that,
for some > 0, (a) a for every a A
f
. (i) Show that there is a unique order-continuous Riesz
homomorphism T : L
1

L
1

such that T(a) = (a) whenever a A
f
, and that |T| . (ii) Show that
[[Tu > ]] = [[u > ]] for every u L
1
(A, ), > 0. (iii) Show that T is surjective i is, injective i is.
(iv) Show that T is norm-preserving i (a) = a for every a A
f
.
(k) Let (A, ) and (B, ) be measure algebras, and : A B a measure-preserving Boolean homo-
morphism. Let T : L
1

L
1

and P : L
1

L
1

be the operators corresponding to A
f
, as described
in 365O-365P, and

T : L

(A) L

(B) the operator corresponding to , as described in 363F. (i) Show

that T(u v) = Tu

Tv for every u L
1

, v L

(A). (ii) Show that if is order-continuous, then

_
Pv u =
_
v

Tu for every u L

(A), v L
1

.
>>>(l) Let (X, , ) be a probability space, with measure algebra (A, ), and let T be a -subalgebra of .
Set = T, B = F

: F T A, = B, so that (B, ) is a measure algebra. Let : B A be

the identity homomorphism. Show that T

: L
1

L
1

(365O) corresponds to the canonical embedding of
L
1
() in L
1
() described in 242Jb, while P

: L
1

L
1

(365P) corresponds to the conditional expectation
operator described in 242Jd.
(m) Let (A, ) be a semi-nite measure algebra, and (

A, ) its localization (322P). Show that the natural

embedding of A in

A induces a Banach lattice isomorphism between L
1

and L
1
(

A, ), so that the band

algebra of L
1

can be identied with the Dedekind completion

A of A.
(n) Let A be a Dedekind -complete Boolean algebra and , two functions such that (A, ), (A, ) are
measure algebras. Show that L
1

L
1

(as subsets of L
0
(A)) i there is a > 0 such that a a for
every a A. (Hint: show that the identity operator from L
1

to L
1

is bounded.)
(o) Let (A, ) be a measure algebra, I

the ideal of purely innite elements of A,

sf
the measure on
B = A/I

(322Xa). Let : A B be the canonical map. Show that T

, as dened in 365O, is a Banach

lattice isomorphism between L
1
(A, ) and L
1
(B,
sf
).
(p) Let (X, , ) be a a semi-nite measure space. Show that L
1
() is separable i is -nite and has
countable Maharam type.
365Y Further exercises (a) Let (A, ) be a semi-nite measure algebra, not 0. Show that the
topological density of L
1
(A, ) (331Yf) is max(, (A), c(A)), where (A), c(A) are the Maharam type and
cellularity of A.
365 Notes and comments You should not suppose that L
1
spaces appear in the second half of this chapter
because they are of secondary importance. Indeed I regard them as the most important of all function spaces.
I have delayed the discussion of them for so long because it is here that for the rst time we need measure
algebras in an essential way.
The actual denition of L
1

which I give is designed for speed rather than illumination; I seek only a
formula, visibly independent of any particular representation of (A, ) as the measure algebra of a measure
space, for which I can prove 365B. I use a relatively primitive argument in 365B, not appealing to Fubinis
theorem; of course the formula
_

0
x : [f(x)[ > d =
_
[f(x)[dx can also be regarded as a reversal of
350 Function spaces 365 Notes
order of integration in a double integral (252N, 252O). 365C-365D and 365Ea are now elementary. In 365Eb
I take a page to describe a form of the Radon-Nikod ym theorem which is applicable to arbitrary measure
algebras, at the cost of dealing with functionals on the ring A
f
rather than on the whole algebra A. This is
less for the sake of applications than to emphasize one of the central properties of L
1
: it depends only on
A
f
and A
f
. For alternative versions of the condition 365Eb(i) see 365Xf.
The convergence theorems (B.Levis theorem, Fatous lemma and Lebesgues dominated convergence
theorem) are so central to the theory of integrable functions that it is natural to look for versions in the
language here. Corresponding to B.Levis theorem is the Levi property of a norm in an L-space; note how
the abstract formulation makes it natural to speak of general upwards-directed families rather than of non-
decreasing sequences, though the sequential form is so often used that I have spelt it out (365C). In the
same way, the integral becomes order-continuous rather than just sequentially order-continuous (365Da).
Corresponding to Fatous lemma we have 365Xc-365Xd. For abstract versions of Lebesgues theorem I will
wait until 367.
In 365H I have deliberately followed the hypotheses of 235A and 235T. Of course 365H can be deduced
from these if we use the Stone representations of (A, ) and (B, ), so that can be represented by a
function between the Stone spaces (312P). But 365H is essentially simpler, because the technical problems
concerning measurability which took up so much of 235 have been swept under the carpet. In the same
way, 365Xg corresponds to 235E. Here we have a fair example of the way in which the abstract expression
in terms of measure algebras can be tidier than the expression in terms of measure spaces. But in my view
this is because here, at least, some of the mathematics has been left out.
365J is a re-run of 243G, but with the additional renement that I examine the action of L
1
on L

(the
operator S) as well as the action of L

on L
1
(the operator T). Note that (b-iii) and (c) of the proof of 365J
depend on the Radon-Nikod ym theorem; these parts lie a step deeper than the rest. 365L-365N correspond
closely to 361F-361H and 363E.
Theorems 365O-365Q lie at the centre of my picture of L
1
spaces, and are supposed to show their dual
nature. Starting from a semi-nite measure algebra (A, ) we have two essentially dierent routes to the
L
1
-space: we can either build it up from characteristic functions of elements of nite measure, so that it is
naturally embedded in L
0
(A), or we can think of it as the order-continuous dual of L

(A). The rst is a

covariant construction (signalled by the formula T

= T

in 365Oe) and the second is contravariant

(so that P

= P

in 365Pf). The rst construction is the natural one if we are seeking to copy the ideas
of 242, but the second arises inevitably if we follow the ordinary paths of functional analysis and study
dual spaces whenever they appear. The link between them is the Radon-Nikod ym theorem.
I have deliberately written out 365O and 365P with dierent hypotheses on the homomorphism in the
hope of showing that the two routes to L
1
really are dierent, and can be expected to tell us dierent things
about it. I use the letter P in 365P in order to echo the language of 242J; in the most important context,
in which A is actually a subalgebra of B and is the identity map, P is a kind of conditional expectation
operator (365R). I note that in the proof of 365Pe I have returned to rst principles, using some of the
ideas of the Radon-Nikod ym theorem (232E), but a dierent approach to the exhaustion step (converting
for every u > 0 there is a v > 0 such that Pv u into P is surjective). I chose the somewhat cruder
method in 232E (part (c) of the proof) in order to use the weakest possible form of the axiom of choice. In
the present context such scruples seem absurd.
I used the words covariant and contravariant above; of course this distinction depends on the side of the
mirror on which we are standing; if our measure-preserving homomorphism is derived (contravariantly) from
an inverse-measure-preserving transformation, then the Ts become contravariant (365Xi). An important
component of this work, for me, is the fact that not all measure-preserving homomorphisms between measure
algebras can be represented by inverse-measure-preserving functions (343Jb, 343M).
I have already remarked (in the notes to 244) that the properties of L
1
() are not much aected by
peculiarities in a measure space (X, , ), because (unlike L
0
or L

) they really depend only on A

f
, the
ring of elements of nite measure in the measure algebra. (See 365O-365Q and 365Xm-365Xn.) Note that
while the algebra A is uniquely determined (given that (A, ) is localizable, 365Sa), the measure is not;
if A is any algebra carrying two non-isomorphic semi-nite measures, the corresponding L
1
spaces are still
isomorphic (365Sb). For instance, the L
1
-spaces of Lebesgue measure on R, and the subspace measure

[0,1]
on [0, 1], are isomorphic, though their measure algebras are not.
I make no attempt here to add to the results in 246, 247, 354 and 356 concerning uniform integrability
366D L
p
351
and weak compactness. Once we have left measure spaces behind, these ideas belong to the theory of Banach
lattices, and there is little to relate them to the questions dealt with in this section. But see 373Xj and
373Xn below.
366 L
p
In this section I apply the methods of this chapter to L
p
spaces, where 1 < p < . The constructions
proceed without surprises up to 366E, translating the ideas of 244 by the methods used in 365. Turning
to the action of Boolean homomorphisms on L
p
spaces, I introduce a space M
0
, which can be regarded
as the part of L
0
that can be determined from the ring A
f
of elements of A of nite measure (366F), and
which includes L
p
whenever 1 p < . Now a measure-preserving ring homomorphism from A
f
to B
f
acts on the M
0
spaces in a way which includes injective Riesz homomorphisms from L
p
(A, ) to L
p
(B, )
and surjective positive linear operators from L
p
(B, ) to L
p
(A, ) (366H). The latter may be regarded as
conditional expectation operators (366J). The case p = 2 (366K-366L) is of course by far the most important.
366A Denition Let (A, ) be a measure algebra and suppose that 1 < p < . For u L
0
(A), dene
[u[
p
L
0
(A) by setting
[[[u[
p
> ]] = [[[u[ >
1/p
]] if 0
= 1 if < 0.
(In the language of 364I, [u[
p
=

h(u), where h(t) = [t[
p
for t R.) Set
L
p

= L
p
(A, ) = u : u L
0
(A), [u[
p
L
1
(A, ),
and for u L
p

set
|u|
p
= (
_
[u[
p
)
1/p
= |[u[
p
|
1/p
1
.
366B Theorem Let (X, , ) be a measure space, and (A, ) its measure algebra. Then the canonical
isomorphism between L
0
() and L
0
(A) (364Jc) makes L
p
(), as dened in 244, correspond to L
p
(A, ).
proof What we really have to check is that if w L
0
() corresponds to u L
0
(A), then [w[
p
, as dened
in 244A, corresponds to [u[
p
as dened in 366A. But this was noted in 364Jb.
Now, because the isomorphism between L
0
() and L
0
(A) matches L
1
() with L
1
(A, ) (365B), we can
be sure that [w[
p
L
1
() i [u[
p
L
1
(A, ), and that in this case
|w|
p
=
_
_
[w[
p
_
1/p
=
__
[u[
p
_
1/p
= |u|
p
,
as required.
366C Corollary For any measure algebra (A, ) and p ]1, [, L
p
= L
p
(A, ) is a solid linear subspace
of L
0
(A). It is a Dedekind complete Banach lattice under its norm | |
p
. Setting q = p/(p 1), (L
p
)

is
identied with L
q
(A, ) by the duality (u, v)
_
uv. Writing A
f
for the ring a : a A, a < , S(A
f
)
is norm-dense in L
p
.
proof Because we can nd a measure space (X, , ) such that (A, ) is isomorphic to the measure algebra
of (321J), this is just a digest of the results in 244B, 244E, 244F, 244G, 244H and 244K. (Of course S(A
f
)
corresponds to the space S of equivalence classes of simple functions in 244Ha, just as in 365F.)
366D I can add a little more, corresponding to 365C and 365J.
Theorem Let (A, ) be a measure algebra, and p ]1, [.
(a) The norm | |
p
on L
p
= L
p
(A, ) is order-continuous.
(b) L
p
has the Levi property.
352 Function spaces 366D
(c) Setting q = p/(p 1), the canonical identication of L
q
= L
q
(A, ) with (L
p
)

is a Riesz space
isomorphism between L
q
and (L
p
)

= (L
p
)

.
(d) L
p
is a perfect Riesz space.
proof (a) Suppose that A L
p
is non-empty, downwards-directed and has inmum 0. For u, v 0 in L
p
,
u v u
p
v
p
(by the denition in 366A, or otherwise), so B = u
p
: u A is downwards-directed. If
v
0
= inf B in L
1
= L
1
(A, ), then v
1/p
0
(dened by the formula in 366A, or otherwise) is less than or equal to
every member of A, so must be 0, and v
0
= 0. Accordingly inf B = 0 in L
1
. Because | |
1
is order-continuous
(365C),
inf
uA
|u|
p
= inf
uA
|u
p
|
1/p
1
= (inf
vB
|v|
1
)
1/p
= 0.
As A is arbitrary, | |
p
is order-continuous.
(b) Now suppose that A (L
p
)
+
is non-empty, upwards-directed and norm-bounded. Then B = u
p
:
u A is non-empty, upwards-directed and norm-bounded in L
1
. So v
0
= sup B is dened in L
1
, and v
1/p
0
is an upper bound for A in L
p
.
(c) By 356Dd, (L
p
)

= (L
p
)

= (L
p
)

. The extra information we need is that the identication of

L
q
with (L
p
)

is an order-isomorphism. PPP () If w (L
q
)
+
and u (L
p
)
+
then u w 0 in L
1
, so
(Tw)(u) =
_
uw 0, writing T : L
q
(L
p
)

for the canonical bijection. As u is arbitrary, Tw 0. As w

is arbitrary, T is a positive linear operator. () If w L
q
and Tw 0, consider u = (w

)
q/p
. Then u 0
in L
p
and w
+
u = 0 (because [[w
+
> 0]] [[u > 0]] = [[w
+
> 0]] [[w

> 0]] = 0), so

0 (Tw)(u) =
_
w u =
_
w

u =
_
(w

)
p
0,
and
_
(w

)
p
= 0. But as (w

)
p
0 in L
1
, this means that (w

)
p
and w

must be 0, that is, w 0. As w

is arbitrary, T
1
is positive and T is an order-isomorphism. QQQ
(d) This is an immediate consequence of (c), since p = q/(q 1), so that L
p
can be identied with
(L
q
)

= (L
q
)

. From 356M we see that it is also a consequence of (a) and (b).

366E Proposition Let (A, ) be a semi-nite measure algebra, and p [1, ]. Set q = p/(p 1) if
1 < p < , and q = if p = 1, q = 1 if p = . Then
L
q
(A, ) = u : u L
0
(A), u v L
1
(A, ) for every v L
p
(A, ).
proof (a) We already know that if u L
p
and v L
q
then u v L
1
; this is elementary if p 1, and
otherwise is covered by 366C.
(b) So suppose that u L
0
L
p
. If p = 1 then of course 1 L

= L
q
and u 1 / L
1
. If p > 1 set
A = w : w S(A
f
), 0 w [u[.
Because is semi-nite, S(A
f
) is order-dense in L
0
(364L), and [u[ = supA. Because the norm on L
p
has
the Levi property (366Db, 363Ba) and A is not bounded above in L
p
, sup
wA
|w|
p
= .
For each n N choose w
n
A with |w
n
|
p
> 4
n
. Then there is a v
n
L
q
such that |v
n
|
q
= 1 and
_
w
n
v
n
4
n
. PPP () If p < this is covered by 366C, since |w
n
|
p
= sup
_
w
n
v : |v|
q
1. () If
p = then [[w
n
> 4
n
]] ,= 0; because is semi-nite, there is a b [[w
n
> 4
n
]] such that 0 < b < , and
|
1
b
b|
1
= 1, while
_
w
n

1
b
b 4
n
. QQQ
Because L
q
is complete (363Ba, 366C), v =

n=0
2
n
[v
n
[ is dened in L
q
. But now
_
[u[ v 2
n
_
w
n
v
n
2
n
for every n, so u v / L
1
.
Remark This result is characteristic of perfect subspaces of L
0
; see 369C and 369J.
366F The next step is to look at the action of Boolean homomorphisms, as in 365O. It will be convenient
to be able to deal with all L
p
spaces at once by introducing names for a pair of spaces which include all of
them.
366H L
p
353
Denition Let (A, ) be a measure algebra. Write
M
0

= M
0
(A, ) = u : u L
0
(A), [[[u[ > ]] < for every > 0,
M
1,0

= M
1,0
(A, ) = u : u M
0

, u a L
1
(A, ) whenever a < .
366G Lemma Let (A, ) be any measure algebra. Write M
0
= M
0
(A, ), etc.
(a) M
0
and M
1,0
are Dedekind complete solid linear subspaces of L
0
which include L
p
for every p [1, [;
moreover, M
0
is closed under multiplication.
(b) If u M
0
and u 0, there is a non-decreasing sequence u
n
)
nN
in S(A
f
) such that u = sup
nN
u
n
.
(c) M
1,0
= u : u L
0
, ([u[ 1)
+
L
1
for every > 0 = L
1
+ (L

M
0
).
(d) If u, v M
1,0
and
_
a
u
_
a
v whenever a < , then u v; so if
_
a
u =
_
a
v whenever a < ,
u = v.
proof (a) If u, v M
0
and R, then for any > 0
[[[u +v[ > ]] [[[u[ >
1
2
]] [[[v[ >
1
2
]],
[[[u[ > ]] [[[u[ >

1+]]
]],
[[[u v[ > ]] [[[u[ >

]] [[[v[ >

]]
(364F) are of nite measure. So u + v, u and u v belong to M
0
. Thus M
0
is a linear subspace of L
0
closed under multiplication. If u M
0
, [v[ [u[ and > 0, then [[[v[ > ]] [[[u[ > ]] is of nite measure;
thus v M
0
and M
0
is a solid linear subspace of L
0
. It follows that M
1,0
also is. If u L
p
= L
p
(A, ),
where p < , and > 0, then [[[u[ > ]] = [[[u[
p
>
p
]] is of nite measure, so u M
0
; moreover, if a < ,
then a L
q
, where q = p/(p 1), so u a L
1
; thus u M
1,0
.
To see that M
0
is Dedekind complete, observe that if A (M
0
)
+
is non-empty and bounded above by
u
0
M
0
, and > 0, then [[u > ]] : u A is bounded above by [[u
0
> ]] A
f
, so has a supremum in A
(321C). Accordingly supA is dened in L
0
(364Mc) and belongs to M
0
. Finally, M
1,0
, being a solid linear
subspace of M
0
, must also be Dedekind complete.
(b) If u 0 in M
0
, then there is a non-decreasing sequence u
n
)
nN
in S = S(A) such that u = sup
nN
u
n
and u
0
0 (364Kd). But now every u
n
belongs to S M
0
= S(A
f
), just as in 365F.
(c)(i) If u M
1,0
and > 0, then a = [[[u[ > ]] A
f
, so u a L
1
= L
1
(A, ); but ([u[ 1)
+

[u[ a, so ([u[ 1)
+
L
1
.
(ii) Suppose that u L
0
and ([u[ 1)
+
L
1
for every > 0. Then, given > 0, v = ([u[
1
2
1)
+

L
1
, and [[v >
1
2
]] < ; but [[[u[ > ]] [[v >
1
2
]], so also has nite measure. Thus u M
0
. Next, if a A
f
,
then [u a[ a + ([u[ 1)
+
L
1
, so u M
1,0
.
(iii) Of course L
1
and L

M
0
are included in M
1,0
, so their linear sum also is. On the other hand,
if u M
1,0
, then
u = (u
+
1)
+
(u

1)
+
+ (u
+
1) (u

1) L
1
+ (L

M
0
).
(d) Take > 0 and set a = [[u u
t
> ]]. Because both u and u
t
belong to M
1,0

, a < and
_
a
u
_
a
u
t
,
that is,
_
a
u u
t
0; so a must be 0 (365Dc). As is arbitrary, u u
t
0 and u u
t
. If
_
a
u =
_
a
u
t
for
every a A
f
, then u
t
u so u = u
t
.
366H Theorem Let (A, ) and (B, ) be measure algebras. For p [1, ], write L
p

, L
p

for L
p
(A, ),
L
p
(B, ); similarly, write M
0

for M
0
(A, ), etc. Let : A
f
B
f
be a measure-preserving ring homomor-
phism.
(a)(i) We have a unique order-continuous Riesz homomorphism T = T

: M
0

M
0

such that T(a) =
(a) for every a A
f
.
(ii) [[Tu > ]] = [[u > ]] for every u M
0

and > 0.
(iii) T is injective and multiplicative.
354 Function spaces 366H
(iv) For p [1, ] and u M
0

, Tu L
p

i u L
p

, and in this case |Tu|
p
= |u|
p
. In particular,
_
Tu =
_
u whenever u L
1

.
(v) For u M
0

, Tu M
1,0

i u M
1,0

.
(b)(i) We have a unique order-continuous positive linear operator P = P

: M
1,0

M
1,0

such that
_
a
Pv =
_
a
v whenever v M
1,0

and a A
f
.
(ii) If u M
0

, v M
1,0

and v Tu M
1,0

, then P(v Tu) = u Pv.
(iii) If q [1, [ and v L
q

, then Pv L
q

and |Pv|
q
|v|
q
; if v L

M
0

, then Pv L

and
|Pv|

|v|

.
(iv) PTu = u for every u M
1,0

; in particular, P[L
p

] = L
p

for every p [1, [.
(c) If (C, ) is another measure algebra and : B
f
C
f
another measure-preserving ring homomorphism,
then T

= T

: M
0

M
0

and P

= P

: M
1,0

M
1,0

.
(d) Now suppose that [A
f
] = B
f
, so that is a measure-preserving isomorphism between the rings A
f
and B
f
. Then
(i) T is a Riesz space isomorphism between M
0

and M
0

, and its inverse is T

1.
(ii) P is a Riesz space isomorphism between M
1,0

and M
1,0

, and its inverse is P

1.
(iii) The restriction of T to M
1,0

is P
1
= P

1; the restriction of T
1
= T

1 to M
1,0

is P.
(iv) For any p [1, [, TL
p

= P

1L
p

and PL
p

= T

1L
p

are the two halves of a Banach lattice
isomorphism between L
p

and L
p

.
proof (a)(i) By 361J, induces a multiplicative Riesz homomorphism T
0
: S(A
f
) S(B
f
) which is order-
continuous because is (361Ad, 361Je). If u S(A
f
) and > 0, then [[T
0
u > ]] = [[u > ]]. PPP Express
u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint in A
f
; then T
0
u =

n
i=0

i
(a
i
), so
[[T
0
u > ]] = supa
i
: i n,
i
> = (supa
i
: i n,
i
> ) = [[u > ]]. QQQ
Now if u
0
0 in M
0

, supT
0
u : u S(A
f
), 0 u u
0
is dened in M
0

. PPP Set A = u : u S(A
f
), 0
u u
0
. Because u
0
= sup A (366Gb),
sup
uA
[[Tu > ]] = sup
uA
[[u > ]] = (sup
uA
[[u > ]]) = [[u
0
> ]]
is dened and belongs to B
f
for any > 0. Also
inf
n1
sup
uA
[[Tu > n]] = (inf
n1
[[u
0
> n]]) = 0.
By 364Mb, v
0
= supT
0
[A] is dened in L
0
(B), and [[v
0
> ]] = [[u
0
> ]] B
f
for every > 0, so v
0
M
0

,
as required. QQQ
Consequently T
0
has a unique extension to an order-continuous Riesz homomorphism T : M
0

M
0

(355F).
(ii) If u
0
M
0

and > 0, then
[[Tu
0
> ]] = [[Tu
+
0
> ]]
(because T is a Riesz homomorphism)
= sup
uS(A
f
),0uu
+
0
[[Tu > ]]
(because T is order-continuous and S(A
f
) is order-dense in M
0

)
= [[u
0
> ]]
by the argument used in (i).
(iii) I have already remarked, at the beginning of the proof of (i), that T(u u
t
) = Tu Tu
t
for u,
u
t
S(A
f
). Because both T and are order-continuous and S(A
f
) is order-dense in M
0

,
T(u
0
u
1
) = supT(u u
t
) : u, u
t
S(A
f
), 0 u u
0
, 0 u
t
u
1

= sup
u,u

Tu Tu
t
= Tu
0
Tu
1
366H L
p
355
whenever u
0
, u
1
0 in M
0

. Because T is linear and is bilinear, it follows that T is multiplicative on M
0

.
To see that it is injective, observe that if u ,= 0 in M
0

then there is some > 0 such that a = [[[u[ > ]] ,= 0,
so that 0 < a T[u[ = [Tu[ and Tu ,= 0.
(iv)() Suppose that p [1, [ and that u L
p

. Then for any > 0,
[[[Tu[
p
> ]] = [[T[u[ >
1/p
]] = [[[u[ >
1/p
]] = [[[u[
p
> ]].
So
|[Tu[
p
|
1
=
_

0
[[[Tu[
p
> ]] d =
_

0
[[[u[
p
> ]] d = |[u[
p
|
1
<
and Tu L
p

, with |Tu|
p
= |u|
p
.
() As for the case p = , if u L

and = |u|

> 0 then [[[u[ > ]] = 0, so [[[Tu[ > ]] =

[[[u[ > ]] = 0. This shows that |Tu|

. On the other hand, if 0 < < then a = [[[u[ > ]] ,= 0,

and a [u[ so (a) [Tu[; as a ,= 0 (because (a) = a > 0), |Tu|

> . This shows that

|Tu|

= |u|

, at least when u ,= 0; but the case u = 0 is trivial.

() Now take any p [1, ], and suppose that u M
0

and that Tu L
p

. Let u
n
)
nN
be a
non-decreasing sequence in S(A
f
) with supremum [u[ and u
0
0 (366Gb). Then Tu
n
Tu so |u
n
|
p
=
|Tu
n
|
p
|Tu|
p
. But this means that u
n
)
nN
is bounded above in L
p

(366Db), so that [u[ and u belong
to L
p

.
() If u L
1

, then
_
Tu = |(Tu)
+
|
1
|(Tu)

|
1
= |Tu
+
|
1
|Tu

|
1
= |u
+
|
1
|u

|
1
=
_
u.
(v) If u M
1,0

and > 0, then T([u[ 1
A
) = [Tu[ 1
B
. PPP Set a = [[[u[ > ]] A
f
. Then
[u[ 1
A
= a +[u[ [u[ a and [[[Tu[ > ]] = a. So
T([u[ 1
A
) = T(a) +T[u[ T([u[ a)
= (a) +[Tu[ [Tu[ (a) = [Tu[ 1
B
. QQQ
Consequently
T([u[ 1
A
)
+
= T([u[ [u[ 1
A
) = ([Tu[ 1
B
)
+
.
But this means that ([u[ 1
A
)
+
L
1

i ([Tu[ 1
B
)
+
L
1

. Since this is true for every > 0, 366Gc
tells us that u M
1,0

i Tu M
1,0

.
(b)(i)() By 365Pa, we have an order-continuous positive linear operator P
0
: L
1

L
1

such that
_
a
P
0
v =
_
a
v for every v L
1

and a A
f
.
() We now nd that if v
0
0 in M
1,0

and B = v : v L
1

, 0 v v
0
, then P
0
[B] has a
supremum in L
0
(A) which belongs to M
1,0

. PPP Because B is upwards-directed and P
0
is order-preserving,
P
0
[B] is upwards-directed. If > 0 and v B and a = [[P
0
v > ]], then
v (v
0


2
1
B
)
+
+

2
1
B
,
so
a
_
a
P
0
v =
_
a
v
_
(v
0


2
1
B
)
+
+

2
(a)
=
_
(v
0


2
1
B
)
+
+

2
a
and
[[P
0
v > ]]
2

_
(v
0


2
1
B
)
+
.
Thus [[P
0
v > ]] : v B is an upwards-directed set in A
f
with measures bounded above in R, and
c

= sup
vB
[[P
0
v > ]] is dened in A
f
. Also
356 Function spaces 366H
inf
n1
c
n
inf
n1
2
n
_
(v
0

n
2
1
B
)
+
= 0.
So, by 364Mb, P
0
[B] has a supremum u
0
L
0
(A), and [[u
0
> ]] = c

A
f
for every > 0, so u
0
M
0

. If
c A
f
, then
_
c
u
0
= sup
vB
_
c
P
0
v = sup
vB
_
c
v
_
c
v
0
< ,
so u
0
M
1,0

. QQQ
() Now 355F tells us that P
0
has a unique extension to an order-continuous positive linear operator
P : M
1,0

M
1,0

. If v
0
0 in M
1,0

and a A
f
, then, as remarked above,
_
a
Pv
0
= sup
_
a
P
0
v : v L
1

, 0 v v
0
= sup
_
a
v : v L
1

, 0 v v
0
=
_
a
v
0
;
because P is linear,
_
a
Pv =
_
a
v for every v M
1,0

, a A
f
.
() By 366Gd, P is uniquely dened by the formula
_
a
Pv =
_
a
v whenever v M
1,0

, a A
f
.
(ii) Because M
0

is closed under multiplication, u Pv certainly belongs to M
0

.
() Suppose that u, v 0. Fix c A
f
for the moment. Suppose that u
t
S(A
f
). Then we can
express u
t
as

n
i=0

i
a
i
where a
i
A
f
for every i n. Accordingly
_
c
u
t
Pv =

n
i=0

i
_
ca
i
Pv =

n
i=0

i
_
v (a
i
) (c) =
_
c
v Tu
t
.
Next, we can nd a non-decreasing sequence u
n
)
nN
in S(A
f
)
+
with supremum u, and
sup
nN
_
c
u
n
Pv = sup
nN
_
c
v Tu
n
=
_
c
sup
nN
v Tu
n
=
_
c
v sup
nN
Tu
n
=
_
c
v Tu,
using the order-continuity of T,
_
and . But this means that u Pv = sup
nN
u
n
Pv is integrable over
c and that
_
c
u Pv =
_
c
v Tu. As c is arbitrary, u Pv = P(v Tu) M
1,0

.
() For general u, v,
v
+
Tu
+
+ v
+
Tu

+ v

Tu
+
+ v

Tu

= [v[ T[u[ = [v Tu[ M

1,0

(because T is a Riesz homomorphism), so we may apply () to each of the four products; combining them,
we get P(v Tu) = u Pv, as required.
(iii) Because P is a positive operator, we surely have [Pv[ P[v[, so it will be enough to show that
|Pv|
q
|v|
q
for v 0 in L
q

.
() I take the case q = 1 rst. In this case, for any a A
f
, we have
_
a
Pv =
_
a
v |v|
1
. In
particular, setting a
n
= [[Pv > 2
n
]],
_
a
n
Pv |v|
1
. But Pv = sup
nN
Pv a
n
, so
|Pv|
1
= sup
nN
_
a
n
Pv |v|
1
.
() Next, suppose that q = , so that v (L

)
+
; say |v|

= . ??? If > 0 and a = [[Pv > ]] ,= 0,

then
a <
_
a
Pv =
_
a
v (a) = a. XXX
So [[Pv > ]] = 0 and Pv L

, with |Pv|

|v|

, at least when |v|

> 0; but the case |v|

= 0 is
trivial.
() I come at last to the general case q ]1, [, v L
q

. In this case set p = q/(q 1). If u L
p

then Tu L
p

so Tu v L
1

and
366J L
p
357
[
_
u Pv[ |u Pv|
1
= |P(Tu v)|
1
(by (ii))
|Tu v|
1
(by () just above)
=
_
[Tu[ [v[ |Tu|
p
|v|
q
= |u|
p
|v|
q
by (a-iii) of this theorem. But this means that u
_
u Pv is a bounded linear functional on L
p

, and
is therefore represented by some w L
q

with |w|
q
|v|
q
. If a A
f
then a L
p

, so
_
a
w =
_
a
Pv;
accordingly Pv is actually equal to w (by 366Gd) and |Pv|
q
= |w|
q
|v|
q
, as claimed.
(iv) If u M
1,0

and a A
f
, we must have
_
a
PTu =
_
a
Tu =
_
T(a) Tu =
_
T(a u) =
_
a u =
_
a
u,
using (a-iv) to see that
_
a u is dened and equal to
_
T(a u). As a is arbitrary, u M
1,0

and
PTu = u.
(c) As usual, in view of the uniqueness of T

and P

, all we have to check is that

T

T(a) = T

(a) = (a) = T

(a),
_
a
PP

w =
_
a
P

w =
_
a
w =
_
a
P

w
whenever a A
f
, w M
1,0

.
(d)(i) By (c), T

1T = T

must be the identity operator on M

0

; similarly, TT

1 is the identity
operator on M
0

. Because T and T

1 are Riesz homomorphisms, they must be the two halves of a Riesz

space isomorphism.
(ii) In the same way, P and P

1 must be the two halves of an ordered linear space isomorphism

between M
1,0

and M
1,0

, and are therefore both Riesz homomorphisms.
(iii) By (b-iv), PTu = u for every u M
1,0

, so TM
1,0

must be P
1
. Similarly P = P
1

1
is the
restriction of T
1
= T

1 to M
1,0

.
(iv) Because T
1
[L
p

] = L
p

(by (a-iv)), and T is a bijection between M
0

and M
0

, TL
p

must be a
Riesz space isomorphism between L
p

and L
p

; (a-iv) also tells us that it is norm-preserving. Now its inverse
is PL
p

, by (iii) here.
366I Corollary Let (A, ) be a measure algebra, and B a -subalgebra of A. Then, for any p [1, [,
L
p
(B, B) can be identied, as Banach lattice, with the closed linear subspace of L
p
(A, ) generated by
b : b B, b < .
proof The identity map b b : B A induces an injective Riesz homomorphism T : L
0
(B) L
0
(A)
(364R) such that Tu L
p
A
= L
p
(A, ) and |Tu|
p
= |u|
p
whenever p [1, [ and u L
p
B
= L
p
(B, B)
(366H(a-iv)). Because S(B
f
), the linear span of b : b B, b < , is dense in L
p
B
(366C), the image of
L
p
B
in L
p
A
must be the closure of the image of S(B
f
) in L
p
A
, that is, the closed linear span of b : b B
f

interpreted as a subset of L
p
A
.
366J Corollary If (A, ) is a probability algebra, B is a closed subalgebra of A, and P : L
1
(A, )
L
1
(B, B) is the conditional expectation operator (365R), then |Pu|
p
|u|
p
whenever p [1, ] and
u L
p
(A, ).
proof Because (A, ) is totally nite, M
0
(A, ) = L
1
(A, ), so that the operator P of 366Hb can be
identied with the conditional expectation operator of 365R. Now 366H(b-iii) gives the result.
Remark Of course this is also covered by 244M.
358 Function spaces 366K
366K Corollary Let (A, ) and (B, ) be measure algebras, and : A
f
B
f
a measure-preserving
ring homomorphism. Let T : L
2

L
2

and P : L
2

L
2

be the corresponding operators, as in 366H. Then
TP : L
2

L
2

is an orthogonal projection, its range TP[L
2

] being isomorphic, as Banach lattice, to L
2

.
The kernel of TP is just
v : v L
2

,
_
a
v = 0 for every a A
f
.
proof Most of this is simply because T is a norm-preserving Riesz homomorphism (so that T[L
2

] is
isomorphic to L
2

), PT is the identity on L
2

(so that (TP)
2
= TP) and |P| 1 (so that |TP| 1). These
are enough to ensure that TP is a projection of norm at most 1, that is, an orthogonal projection. Also
TPv = 0 Pv = 0
_
a
Pv = 0 for every a A
f

_
a
v = 0 for every a A
f
.
366L Corollary Let (A, ) be a measure algebra, and : A
f
A
f
a measure-preserving ring automor-
phism. Then there is a corresponding Banach lattice isomorphism T of L
2
= L
2
(A, ) dened by writing
T(a) = (a) for every a A
f
. Its inverse is dened by the formula
_
a
T
1
u =
_
a
u for every u L
2
, a A
f
.
proof In the language of 366H, T = T

and T
1
= P.
366X Basic exercises (a) Let (A, ) be a measure algebra and p ]1, [. Show that, for u L
0
(A),
u L
p
(A, ) i = p
_

0

p1
[[[u[ > ]] d is nite, and that in this case |u|
p
=
1/p
. (Cf. 263Xa.)
>>>(b) Let (A, ) be a localizable measure algebra and p [1, ]. Show that the band algebra of L
p
(A, )
is isomorphic to A. (Cf. 365S.)
(c) Let (A, ) be a measure algebra and p ]1, [. Show that L
p
(A, ) is separable i L
1
(A, ) is.
(d) Let (A, ) be a measure algebra. (i) Show that L

(A)M
0
(A, ) and L

(A)M
1,0
(A, ), as dened
in 366G, are equal. (ii) Call this intersection M
,0
(A, ). Show that it is a norm-closed solid linear subspace
of L

(A), therefore a Banach lattice in its own right.

(e) Let (A, ) be a semi-nite measure algebra and (

A, ) its localization (322P). Show that the natural

embedding of A in

A induces a Banach lattice isomorphism between L
p
(A, ) and L
p
(

A, ) for every p
[1, [, so that the band algebra of L
p
(A, ) can be identied with

A.
(f ) Let (A, ) be a semi-nite measure algebra which is not localizable (cf. 211Ya, 216D), and (

A, ) its
localization. Let : A

A be the identity embedding, so that is an order-continuous measure-preserving
Boolean homomorphism. Show that if we set v = b where b

A A, then there is no u L

(A) such that

_
a
u =
_
a
v whenever a < .
(g) In 366H, show that [[Tu E]] = [[u E]] whenever u M
0

and E R is a Borel set such that 0 / E.
>>>(h) Let (A, ) be a measure algebra and let G be the group of all measure-preserving ring automorphisms
of A
f
. Let H be the group of all Banach lattice automorphisms of L
2
(A, ). Show that the map T of
366L is an injective group homomorphism from G to H, so that G is represented as a subgroup of H.
(i) Let (A
i
,
i
))
iI
be any family of measure algebras, with simple product (A, ) (322K). Show that for
any p [1, [, L
p
(A, ) can be identied, as normed Riesz space, with the solid linear subspace
u : |u| =
_
iI
|u(i)|
p
_
1/p
<
of

iI
L
p
(A
i
,
i
).
366 Notes L
p
359
(j) Let A be a Dedekind -complete Boolean algebra and , two functionals rendering A a semi-nite
measure algebra. Show that for any p [1, [, L
p
(A, ) and L
p
(A, ) are isomorphic as normed Riesz
spaces. (Hint: use 366Xe to reduce to the case in which A is Dedekind complete. Take w L
0
(A) such that
_
a
wd = a for every a A (365T). Set Tu = w
1/p
u for u L
p
(A, ).)
(k) Let (A, ) and (B, ) be semi-nite measure algebras, and p [1, [. Show that the following are
equiveridical: (i) L
p

and L
p

are isomorphic as Banach lattices; (ii) L
p

and L
p

are isomorphic as Riesz
spaces; (iii) A and B have isomorphic Dedekind completions.
366Y Further exercises (a) Let (A, ) be a measure algebra and suppose that 0 < p < 1. Write
L
p
= L
p
(A, ) for u : u L
0
(A), [u[
p
L
1
(A, ), and for u L
p
set (u) =
_
[u[
p
. (i) Show that denes
a Hausdor linear space topology on L
p
(see 2A5B). (ii) Show that if A L
p
is non-empty, downwards-
directed and has inmum 0 then inf
uA
(u) = 0. (iii) Show that if A L
p
is non-empty, upwards-directed
and bounded in the linear topological space sense (see 245Yf) then A is bounded above. (iv) Show that
(L
p
)

= (L
p
)

is just the set of continuous linear functionals from L

p
to R, and is 0 i A is atomless.
(b) Let (A, ) be a measure algebra. Show that M
0
(A, ) has the countable sup property.
(c) Let (A, ) be a measure algebra and dene M
,0
(A, ) as in 366Xd. Show that (M
,0
(A, ))

can
be identied with L
1
(A, ).
(d) In 366H, show that if

T : M
0
(A, ) M
0
(B, ) is any positive linear operator such that

T(a) =
(a) for every a A
f
, then

T is order-continuous, so that it is equal to T

.
(e) Let (A, ) be a measure algebra. (i) Show that there is a natural one-to-one correspondence between
M
1,0
(A, ) and the set of additive functionals : A
f
R such that << in the double sense that for
every > 0 there are , M > 0 such that [a[ whenever a and [a[ a whenever a M. (ii)
Use this description of M
1,0
to prove 366H(b-i).
(f ) In 366H, show that the following are equiveridical: () [A
f
] = B
f
; () T = T

is surjective; ()
P = P

is injective; () P is a Riesz homomorphism; () there is some q [1, ] such that |Pv|

q
= |v|
q
for every v L
q

; () TPv = v for every v M
1,0

.
(g) Let (A, ) and (B, ) be measure algebras, and suppose that : A
f
B
f
is a measure-preserving
ring homomorphism, as in 366H; let T : M
0
(A, ) M
0
(B, ) be the associated linear operator. Show that
if 0 < p < 1 (as in 366Ya) then L
p
(A, ) M
0
(A, ) and T
1
[L
p
(B, )] = L
p
(A, ).
(h) Let (A, ) be a totally nite measure algebra. (i) For each Boolean automorphism : A A, let
T

: L
0
(A) L
0
(A) be the associated Riesz space isomorphism (364R), and let w

L
1
(A)
+
be such that
_
a
w

= (
1
a) for every a A (365Ea). Set Q

u = T

for u L
0
(A). Show that |Q

u|
2
= |u|
2
for every u L
2
(A). (ii) Show that if , : A A are Boolean automorphisms then Q

= Q

.
366 Notes and comments The L
p
spaces, for 1 p , constitute the most important family of
leading examples for the theory of Banach lattices, and it is not to be wondered at that their properties
reect a wide variety of general results. Thus 366Dd and 366E can both be regarded as special cases
of theorems about perfect Riesz spaces (356M and 369D). In a dierent direction, the concept of Orlicz
space (369Xd below) generalizes the L
p
spaces if they are regarded as normed subspaces of L
0
invariant
under measure-preserving automorphisms of the underlying algebra. Yet another generalization looks at the
(non-locally-convex) spaces L
p
for 0 < p < 1 (366Ya).
In 366H and its associated results I try to emphasize the way in which measure-preserving homomorphisms
of the underlying algebras induce both direct and dual operators on L
p
spaces. We have already seen
the phenomenon in 365P. I express this in a slightly dierent form in 366H, noting that we really do need
the homomorphisms to be measure-preserving, for the dual operators as well as the direct operators, so
we no longer have the shift in the hypotheses which appears between 365O and 365P. Of course all these
renements in the hypotheses are irrelevant to the principal applications of the results, and they make
360 Function spaces 366 Notes
substantial demands on the reader; but I believe that the demands are actually demands to expand ones
imagination, to encompass the dierent ways in which the spaces depend on the underlying measure algebras.
In the context of 366H, L

is set apart from the other L

p
spaces, because L

(A) is not in general

determined by the ideal A
f
, and the hypotheses of 366H do not look outside A
f
. 366H(a-iv) and 366H(b-iii)
reach only the space M
,0
as dened in 366Xd. To deal with L

we need slightly stronger hypotheses.

If we are given a measure-preserving Boolean homomorphism from A to B, rather than from A
f
to B
f
,
then of course the direct operator T has a natural version acting on L

(A) and indeed on M

1,
(A, ), as
in 363F and 369Xm. If we know that (A, ) is localizable, then A can be recovered from A
f
, and the dual
operator P acts on L

(B), as in 369Xm. But in general we cant expect this to work (366Xf).

Of course 366H can be applied to many other spaces; for reasons which will appear in 371 and 374, the
archetypes are not really L
p
spaces at all, but the spaces M
1,0
(366F) and M
1,
(369N).
I include 366L and 366Yh as pointers to one of the important applications of these ideas: the investigation
of properties of a measure-preserving homomorphism in terms of its action on L
p
spaces. The case p = 2
is the most useful because the group of unitary operators (that is, the normed space automorphisms) of L
2
has been studied intensively.
367 Convergence in measure
Continuing through the ideas of Chapter 24, I come to convergence in measure. The basic results of
245 all translate easily into the new language (367M-367N, 367Q). The associated concept of (sequential)
order-convergence can also be expressed in abstract terms (367A), and I take the trouble to do this in the
context of general lattices (367A-367B), since the concept can be applied in many ways (367C-367F, 367L,
367Xa-367Xm). In the particular case of L
0
spaces, which are the rst aim of this section, the idea is most
naturally expressed by 367G. It enables us to express some of the fundamental theorems from Volumes 1
and 2 in the language of this chapter (367J-367K).
In 367O and 367P I give two of the most characteristic properties of the topology of convergence in
measure on L
0
; it is one of the fundamental types of topological Riesz space. Another striking fact is the
way it is determined by the Riesz space structure (367T). In 367U I set out a theorem which is the basis
of many remarkable applications of the concept; for the sake of a result in 369 I give one such application
(367V).
367A Order*-convergence As I have remarked before, the function spaces of measure theory have
three interdependent structures: they are linear spaces, they have a variety of interesting topologies, and
they are ordered spaces. Ordinary elementary functional analysis studies interactions between topologies
and linear structures, in the theory of normed spaces and, more generally, of linear topological spaces.
Chapter 35 in this volume looked at interactions between linear and order structures. It is natural to seek
to complete the triangle with a theory of topological ordered spaces. The relative obscurity of any such
theory is in part due to the diculty of nding convincing denitions; that is, isolating concepts which lead
to in elegant and useful general theorems. Among the many rival ideas, however, I believe it is possible to
identify one which is particularly important in the context of measure theory.
In its natural home in the theory of L
0
spaces, this notion of order*-convergence has a very straightfor-
ward expression (367G). But, suitably interpreted, the same idea can be applied in other contexts, some of
which will be very useful to us, and I therefore begin with a denition which is applicable in any lattice.
Denition Let P be a lattice and p
n
)
nN
a sequence in P, p an element of P. I will say that p
n
)
nN
order*-converges to p, or that p is the order*-limit of p
n
)
nN
, if
p = infq : n N, q (p
t
p
i
) p
tt
i n
= supq : n N, q p
t
(p
i
p
tt
) i n
whenever p
t
p p
tt
in P.
367B Convergence in measure 361
367B Lemma Let P be a lattice.
(a) A sequence in P can order*-converge to at most one point.
(b) A constant sequence order*-converges to its constant value.
(c) Any subsequence of an order*-convergent sequence is order*-convergent, with the same limit.
(d) If p
n
)
nN
and p
t
n
)
nN
both order*-converge to p, and p
n
q
n
p
t
n
for every n, then q
n
)
nN
order*-converges to p.
(e) If p
n
)
nN
is an order-bounded sequence in P, then it order*-converges to p P i
p = infq : n N, q p
i
i n
= supq : n N, q p
i
i n.
(f) If P is a Dedekind -complete lattice (314A) and p
n
)
nN
is an order-bounded sequence in P, then it
order*-converges to p P i
p = sup
nN
inf
in
p
i
= inf
nN
sup
in
p
i
.
proof (a) Suppose that p
n
)
nN
is order*-convergent to both p and p. Set p
t
= p p, p
tt
= p p; then
p = infq : n N, q (p
t
p
i
) p
tt
i n = p.
(b) is trivial.
(c) Suppose that p
n
)
nN
is order*-convergent to p, and that p
t
n
)
nN
is a subsequence of p
n
)
nN
. Take
p
t
, p
tt
such that p
t
p p
tt
, and set
B = q : n N, q p
t
(p
i
p
tt
) i n,
B
t
= q : n N, q p
t
(p
t
i
p
tt
) i n,
C = q : n N, q (p
t
p
i
) p
tt
i n,
C
t
= q : n N, q (p
t
p
t
i
) p
tt
i n.
If q B
t
and q
t
C, then for all suciently large i
q p
t
(p
t
i
p
tt
) (p
t
p
t
i
) p
tt
q
t
.
As p = inf C, we must have q p; thus p is an upper bound for B
t
. On the other hand, p
t
i
: i n p
i
:
i n for every n, so B B
t
and p must be the least upper bound of B
t
, since p = sup B.
Similarly, p = inf C
t
. As p
t
and p
tt
are arbitrary, p
t
n
)
nN
order*-converges to p.
(d) Take p
t
, p
tt
such that p
t
p p
tt
, and set
B = q : n N, q p
t
(p
i
p
tt
) i n,
B
t
= q : n N, q p
t
(q
i
p
tt
) i n,
C = q : n N, q (p
t
p
t
i
) p
tt
i n,
C
t
= q : n N, q (p
t
q
i
) p
tt
i n.
If q B
t
and q
t
C, then for all suciently large i
q p
t
(q
i
p
tt
) (p
t
p
t
i
) p
tt
q
t
.
As p = inf C, we must have q p; thus p is an upper bound for B
t
. On the other hand, p
t
(p
i
p
tt
)
p
t
(q
i
p
tt
) for every i, so B B
t
and p = supB
t
. Similarly, p = inf C
t
. As p
t
and p
tt
are arbitrary, q
n
)
nN
order*-converges to p.
(e) Set
B = q : n N, q p
i
i n,
C = q : n N, q p
i
i n.
362 Function spaces 367B
(i) Suppose that p
n
)
nN
order*-converges to p. Let p
t
, p
tt
be such that p
t
p
n
p
tt
for every n N
and p
t
p p
tt
. Then
B = q : n N, q p
t
(p
i
p
tt
) i n,
so sup B = p. Similarly, inf C = p, so the condition is satised.
(ii) Suppose that supB = inf C = p. Take any p
t
, p
tt
such that p
t
p p
tt
and set
B
t
= q : n N, q p
t
(p
i
p
tt
) i n,
C
t
= q : n N, q (p
t
p
i
) p
tt
i n.
If q B
t
and q
t
C, then for all large enough i
q p
t
(p
i
p
tt
) p
t
q
t
= q
t
because p q
t
. As inf C = p, p is an upper bound for B
t
. On the other hand, if q B, then q p, so
q p
t
(p
i
p
tt
) whenever q p
i
, which is so for all suciently large i, and q B
t
. Thus B
t
B and p
must be the supremum of B
t
. Similarly, p = inf C
t
; as p
t
and p
tt
are arbitrary, p
n
)
nN
order*-converges to
p.
(f ) This follows at once from (e). Setting
B = q : n N, q p
i
i n,
B
t
= inf
in
p
i
: i N,
then B
t
B and for every q B there is a q
t
B
t
such that q q
t
; so supB = sup B
t
if either is dened.
Similarly,
infq : n N, q p
i
i n = inf
nN
sup
in
p
i
if either is dened.
367C Proposition Let U be a Riesz space and u
n
)
nN
, v
n
)
nN
two sequences in U order*-converging
to u, v respectively.
(a) If w U, u
n
+w)
nN
order*-converges to u +w, and u
n
order*-converges to u for every R.
(b) u
n
v
n
)
nN
order*-converges to u v.
(c) If w
n
)
nN
is any sequence in U, then it order*-converges to w U i [w
n
w[)
nN
order*-converges
to 0.
(d) u
n
+v
n
)
nN
order*-converges to u +v.
(e) If U is Archimedean, and
n
)
nN
is a sequence in R converging to R, then
n
u
n
)
nN
order*-
converges to u.
(f) Again suppose that U is Archimedean. Then a sequence w
n
)
nN
in U
+
is not order*-convergent to
0 i there is a w > 0 such that w = sup
in
w w
i
for every n N.
proof (a)(i) u
n
+w)
nN
order*-converges to u +w because the ordering of U is translation-invariant; the
map w
t
w
t
+w is an order-isomorphism.
(ii)() If > 0, then the map w
t
w
t
is an order-isomorphism, so u
n
)
nN
order*-converges to
u.
() If = 0 then u
n
)
nN
order*-converges to u = 0 by 367Bb.
() If w
t
u w
tt
then w
tt
u w
t
so
u = infw : n N, w ((w
tt
) u
i
) (w
t
) i n
= supw : n N, w (w
tt
) (u
i
(w
t
)) i n.
Turning these formulae upside down,
u = supw : n N, w (w
tt
(u
i
)) w
t
i n
= infw : n N, w w
tt
((u
i
) w
t
) i n.
367C Convergence in measure 363
As w
t
and w
tt
are arbitrary, u
n
)
nN
order*-converges to u.
() Putting () and () together, u
n
)
nN
order*-converges to u for every < 0.
(b) Suppose that w
t
u v w
tt
. Set
B = w : n N, w w
t
((u
i
v
i
) w
tt
) i n,
C = w : n N, w (w
t
(u
i
v
i
)) w
tt
i n,
B
1
= w : n N, w (w
t
u) (u
i
w
tt
) i n,
B
2
= w : n N, w (w
t
v) (v
i
w
tt
) i n,
C
1
= w : n N, w ((w
t
u) u
i
) w
tt
i n,
C
2
= w : n N, w ((w
t
v) v
i
) w
tt
i n,
If w
1
B
1
and w
2
B
2
then w
1
w
2
B. PPP There is an n N such that w
1
(w
t
u) (u
i
w
tt
) for
every i n, while w
2
(w
t
v) (v
i
w
tt
) for every i n. So
w
1
w
2
(w
t
u) (w
t
v) (u
i
w
tt
) (v
i
w
tt
)
= (w
t
(u v)) ((u
i
v
i
) w
tt
)
(by the distributive law 352Ec)
= w
t
((u
i
v
i
) w
tt
)
for every i n, and w
1
w
2
B. QQQ
Similarly, if w
1
C
1
and w
2
C
2
then w
1
w
2
C. PPPThere is an n N such that w
1
((w
t
u)u
i
)w
tt
,
w
2
((w
t
v) v
i
) w
tt
for every i n. So
w
1
w
2
(((w
t
u) u
i
) w
tt
) (((w
t
v) v
i
) w
tt
)
= ((w
t
u) u
i
(w
t
v) v
i
) w
tt
= ((w
t
(u v)) (u
i
v
i
)) w
tt
= (w
t
(u
i
v
i
)) w
tt
for every i n, so w
1
w
2
C. QQQ
At the same time, of course, w w whenever w B, w C, since there is some i N such that
w w
t
((u
i
v
i
) w
tt
) (w
t
(u
i
v
i
)) w
tt
w.
Since
supw
1
w
2
: w
1
B
1
, w
2
B
2
= (supB
1
) (supB
2
) = u v,
infw
1
w
2
: w
1
C
1
, w
2
C
2
= (inf C
1
) (inf C
2
) = u v
(using the generalized distributive laws in 352E), we must have supB = inf C = u v. As w
t
and w
tt
are
arbitrary, u
n
v
n
)
nN
is order*-convergent to u v.
(c) The hard parts are over. (i) If w
n
)
nN
order*-converges to w, then w
n
w)
nN
, w w
n
)
nN
and
[w
n
w[)
nN
= (w
n
w) (w w
n
))
nN
all order*-converge to 0, putting (a) and (b) together. (ii) If
[w
n
w[)
nN
order*-converges to 0, then so do [w
n
w[)
nN
and w
n
w)
nN
, by (a) and 367Bd; so
w
n
)
nN
order*-converges to 0, by (a) again.
(d) [u
n
u[)
nN
and [v
n
v[)
nN
order*-converge to 0, by (c), so 2([u
n
u[ [v
n
v[))
nN
also
order*-converges to 0, by (b) and (a). But
0 [(u
n
+v
n
) (u +v)[ [u
n
u[ +[v
n
v[ 2([u
n
u[ [v
n
v[)
for every n, so [(u
n
+ v
n
) (u + v)[)
nN
order*-converges to 0, by 367Bb and 367Bd, and u
n
+ v
n
)
nN
order*-converges to u +v.
364 Function spaces 367C
(e) Set
n
= sup
in
[
i
[ for each n. Then
n
)
nN
0, so inf
nN

n
[u[ = 0, because U is Archimedean.
Consequently
n
[u[)
nN
order*-converges to 0, by 367Be. But we also have
0
[u
n
u[ order*-converging to
0, by (c) and (a), so
0
[u
n
u[+
n
[u[)
nN
order*-converges to 0, by (d). As [
n
u
n
u[
0
[u
n
u[+
n
[u[
for every n,
n
u
n
)
nN
order*-converges to u, as required.
(f )(i) Suppose that w
n
)
nN
is not order*-convergent to 0. Then there are w
t
, w
tt
such that w
t
0 w
tt
and either
B = w : n N, w w
t
(w
i
w
tt
) i n
does not have supremum 0, or
C = w : n N, w (w
t
w
i
) w
tt
i n
does not have inmum 0. Now 0 B, because every w
i
0, and every member of B is a lower bound for
C; so 0 cannot be the greatest lower bound of C. Let w > 0 be a lower bound for C.
Let n N, and set
C
n
= w : w (w
t
w
i
) w
tt
i n = w : w w
i
w
tt
i n.
Because U is Archimedean, we know that inf(C
n
A
n
) = 0, where A
n
= w
i
w
tt
: i n (353F). Now w
is a lower bound for C
n
, so
inf
in
( w w
i
)
+
inf(w w
i
)
+
: w C, i n
inf(w (w
i
w
tt
))
+
: w C, i n
= infw (w
i
w
tt
) : w C, i n = inf(C
n
A
n
) = 0.
As this is true for every n N, w has the property declared.
(ii) If w > 0 is such that w = sup
in
w w
i
for every n N, then
w : n N, w (0 w
i
) w i n
cannot have inmum 0, and w
n
)
nN
is not order*-convergent to 0.
367D As an example of the use of this concept in a relatively abstract setting, I oer the following.
Proposition Let U be a Banach lattice and u
n
)
nN
a sequence in U which is norm-convergent to u U.
Then u
n
)
nN
has a subsequence which is order-bounded and order*-convergent to u. So if u
n
)
nN
itself is
order*-convergent, its order*-limit is u.
proof Let u
t
n
)
nN
be a subsequence of u
n
)
nN
such that |u
t
n
u| 2
n
for each n N. Then
v
n
= sup
in
[u
t
i
u[ is dened in U, and |v
n
| 2
n+1
, for each n (354C). Because inf
nN
|v
n
| = 0,
inf
nN
v
n
must be 0, while u v
n
u
t
i
u + v
n
whenever i n; so u
t
n
)
nN
order*-converges to u, by
367Be.
Now if u
n
)
nN
has an order*-limit, this must be u, by 367Ba and 367Bc.
367E Proposition Let U be a Riesz space with an order-continuous norm. Then any order-bounded
order*-convergent sequence is norm-convergent.
proof Suppose that u
n
)
nN
is order*-convergent to u. Then [u
n
u[)
nN
is order*-convergent to 0
(367Cc), so
C = v : n N, v [u
i
u[ i n
has inmum 0 (367Be). Because U is a lattice, C is downwards-directed, so inf
vC
|v| = 0. But
inf
vC
|v| inf
nN
sup
in
|u
i
u|,
so lim
n
|u
n
u| = 0, that is, u
n
)
nN
is norm-convergent to u.
367H Convergence in measure 365
367F One of the fundamental obstacles to the development of any satisfying general theory of ordered
topological spaces is the erratic nature of the relations between subspace topologies of order topologies and
order topologies on subspaces. The particular virtue of order*-convergence in the context of function spaces
is that it is relatively robust when transferred to the subspaces we are interested in.
Proposition Let U be an Archimedean Riesz space and V a regularly embedded Riesz subspace. (For
instance, V might be either solid or order-dense.) If v
n
)
nN
is a sequence in V and v V , then v
n
)
nN
order*-converges to v when regarded as a sequence in V , i it order*-converges to v when regarded as a
sequence in U.
proof (a) Since, in either V or U, v
n
)
nN
order*-converges to v i [v
n
v[)
nN
order*-converges to 0
(367Cc), it is enough to consider the case v
n
0, v = 0.
(b) If v
n
)
nN
is not order*-convergent to 0 in U, then, by 367Cf, there is a u > 0 in U such that
u = sup
in
u v
i
for every n N (the supremum being taken in U, of course). In particular, there is a
k N such that u v
k
> 0. Now consider the set
C = w : w V, n N, w (0 v
i
) v
k
i n.
Then for any w C,
u v
k
= sup
in
u v
i
v
k
w,
using the generalized distributive law in U, so 0 is not the greatest lower bound of C in U. But as the
embedding of V in U is order-continuous, 0 is not the greatest lower bound of C in V , and v
n
)
nN
cannot
be order*-convergent to 0 in V .
(c) Now suppose that v
n
)
nN
is not order*-convergent to 0 in V . Because V also is Archimedean (351Rc),
there is a w > 0 in V such that w = sup
in
w v
i
for every n N, the suprema being taken in V . Again
because V is regularly embedded in U, we have the same suprema in U, so, by 367Cf in the other direction,
v
n
)
nN
is not order*-convergent to 0 in U.
367G I now spell out the connexion between the denition above and the concepts introduced in 245C.
Proposition Let X be a set, a -algebra of subsets of X, A a Boolean algebra and : A a
sequentially order-continuous surjective Boolean homomorphism; let 1 be its kernel. Write L
0
for the
space of -measurable real-valued functions on X, and let T : L
0
L
0
= L
0
(A) be the canonical Riesz
homomorphism (364D, 364R). Then for any f
n
)
nN
, f in L
0
, Tf
n
)
nN
order*-converges to Tf in L
0
i
X x : f(x) = lim
n
f
n
(x) 1.
proof Set H = x : lim
n
f
n
(x) exists = f(x); of course H . Set g
n
(x) = [f
n
(x) f(x)[ for n N,
x X.
(a) If X H 1, set h
n
(x) = sup
in
g
i
(x) for x H and h
n
(x) = 0 for x X H. Then h
n
)
nN
is a non-increasing sequence with inmum 0 in L
0
, so inf
nN
Th
n
= 0 in L
0
, because T is sequentially
order-continuous (364Ra). But as H 1, Th
n
Tg
i
= [Tf
i
Tf[ whenever i n, so [Tf
n
Tf[)
nN
order*-converges to 0, by 367Be or 367Bf, and Tf
n
)
nN
order*-converges to Tf, by 367Cc.
(b) Now suppose that Tf
n
)
nN
order*-converges to Tf. Set g
t
n
(x) = min(1, g
n
(x)) for n N, x X; then
Tg
t
n
)
nN
= e [Tf
n
Tf[)
nN
order*-converges to 0, where e = T(X). By 367Bf, inf
nN
sup
in
Tg
t
i
= 0
in L
0
. But T is a sequentially order-continuous Riesz homomorphism, so T(inf
nN
sup
in
g
t
i
) = 0, that is,
X H = x : inf
nN
sup
in
g
t
i
> 0
belongs to 1.
367H Corollary Let A be a Dedekind -complete Boolean algebra.
(a) Any order*-convergent sequence in L
0
= L
0
(A) is order-bounded.
(b) If u
n
)
nN
is a sequence in L
0
, then it is order*-convergent to u L
0
i
u = inf
nN
sup
in
u
i
= sup
nN
inf
in
u
i
.
366 Function spaces 367H
proof (a) We can express A as a quotient /1 of a -algebra of sets, in which case L
0
can be identied
with the canonical image of L
0
= L
0
() (364D). If u
n
)
nN
is an order*-convergent sequence in L
0
, then
it is expressible as Tf
n
)
nN
, where T : L
0
L
0
is the canonical map, and 367G tells us that f
n
(x))
nN
converges for every x H, where X H 1. If we set h(x) = sup
nN
[f
n
(x)[ for x H, 0 for x X H,
then we see that [u
n
[ Th for every n N, so that u
n
)
nN
is order-bounded in L
0
.
(b) This now follows from 367Be, because L
0
is Dedekind -complete.
367I Proposition Suppose that E R is a Borel set and h : E R is a continuous function. Let A
be a Dedekind -complete Boolean algebra and set Q
E
= u : u L
0
, [[u E]] = 1, where L
0
= L
0
(A).
Let

h : Q
E
L
0
be the function dened by h (364I). Then

h(u
n
))
nN
order*-converges to

h(u) whenever
u
n
)
nN
is a sequence in Q
E
order*-converging to u Q
E
.
proof This is an easy consequence of 367G. We can represent A as /1 where is a -algebra of subsets
of some set X and 1 is a -ideal of (314M); let T : L
0
L
0
(A) be the corresponding homomorphism
(364D, 367G). Now we can nd -measurable functions f
n
)
nN
, f such that Tf
n
= u
n
, Tf = u, as in
367G; and the hypothesis [[u
n
E]] = 1, [[u E]] = 1 means just that, adjusting f
n
and f on a member of
1 if necessary, we can suppose that f
n
(x), f(x) E for every x X. (I am passing over the trivial case
E = , X 1, A = 0.) Accordingly

h(u
n
) = T(hf
n
),

h(u) = T(hf), and (because h is continuous)
x : h(f(x)) ,= lim
n
h(f
n
(x)) x : f(x) ,= lim
n
f
n
(x) 1,
so

h(u
n
))
nN
order*-converges to

h(u).
367J Dominated convergence We now have a suitable language in which to express an abstract
version of Lebesgues Dominated Convergence Theorem.
Theorem Let (A, ) be a measure algebra. If u
n
)
nN
is a sequence in L
1
= L
1
(A, ) which is order-bounded
and order*-convergent in L
1
, then u
n
)
nN
is norm-convergent to u in L
1
; in particular,
_
u = lim
n
_
u
n
.
proof The norm of L
1
is order-continuous (365C), so u
n
)
nN
is norm-convergent to u, by 367E. As
_
is
norm-continuous,
_
u = lim
n
_
u
n
.
367K The Martingale Theorem In the same way, we can re-write theorems from 275 in this lan-
guage.
Theorem Let (A, ) be a probability algebra, and B
n
)
nN
a non-decreasing sequence of closed subalgebras
of A. For each n N let P
n
: L
1
= L
1

L
1
L
0
(B
n
) be the conditional expectation operator (365R); let
B be the closed subalgebra of A generated by

nN
B
n
, and P the conditional expectation operator onto
L
1
L
0
(B).
(a) If u
n
)
nN
is a norm-bounded sequence in L
1
such that P
n
(u
n+1
) = u
n
for every n N, then u
n
)
nN
is order*-convergent in L
1
.
(b) If u L
1
then P
n
u)
nN
is order*-convergent to Pu.
proof If we represent (A, ) as the measure algebra of a probability space, these become mere translations
of 275G and 275I. (Note that this argument relies on the description of order*-convergence in L
0
in terms of
a.e. convergence of functions, as in 367G; so that we need to know that order*-convergence in L
1
matches
order*-convergence in L
0
, which is what 367F is for.)
367L Some of the most important applications of these ideas concern spaces of continuous functions.
I do not think that this is the time to go very far along this road, but one particular fact will be useful in
376.
Proposition Let X be a locally compact Hausdor space, and u
n
)
nN
a sequence in C(X), the space
of continuous real-valued functions on X. Then u
n
)
nN
order*-converges to 0 in C(X) i x : x
X, limsup
n
[u
n
(x)[ > 0 is meager. In particular, u
n
)
nN
order*-converges to 0 if lim
n
u
n
(x) = 0
for every x.
367N Convergence in measure 367
proof (a) The following elementary fact is worth noting: if A C(X)
+
is non-empty and inf A = 0 in
C(X), then G =

uA
x : u(x) < is dense for every > 0. PPP??? If not, take x
0
X G. Because X is
completely regular (3A3Bb), there is a continuous function w : X [0, 1] such that w(x
0
) = 1 and w(x) = 0
for every x G. But in this case 0 < w u for every u A, which is impossible. XXXQQQ
(b) Suppose that u
n
)
nN
order*-converges to 0. Set v
n
= [u
n
[ X, so that v
n
)
nN
order*-converges
to 0 (using 367C, as usual). Set
B = v : v C(X), n N, v
i
v i n,
so that inf B = 0 in C(X) (367Be). For each k N, set G
k
=

vB
x : v(x) < 2
k
; then G
k
is dense, by
(a), and of course is open. So H =

kN
X G
k
is a countable union of nowhere dense sets and is meager.
But this means that
x : limsup
n
[u
n
(x)[ > 0 = x : limsup
n
v
n
(x) > 0
x : inf
vB
v(x) > 0 H
is meager.
(c) Now suppose that u
n
)
nN
does not order*-converge to 0. By 367Cf, there is a w > 0 in C(X) such
that w = sup
in
w [u
i
[ for every n N; that is, inf
in
(w [u
i
[)
+
= 0 for every n. Set
G
n
= x : inf
in
(w [u
i
[)
+
(x) < 2
n
= x : sup
in
[u
i
(x)[ > w(x) 2
n

for each n. Then

H =

nN
G
n
= x : limsup
n
u
n
(x) w(x)
is the intersection of a sequence of dense open sets, and its complement is meager.
Let G be the non-empty open set x : w(x) > 0. Then G is not meager, by Baires theorem (3A3Ha);
so G H cannot be meager. But x : limsup
n
[u
n
(x)[ > 0 includes G H, so is also not meager.
Remark Unless the topology of X is discrete, C(X) is not regularly embedded in R
X
, and we expect to nd
sequences in C(X) which order*-converge to 0 in C(X) but not in R
X
. But the proposition tells us that if
we have a sequence in C(X) which order*-converges in R
X
to a member of C(X), then it order*-converges
in C(X).
367M Everything above concerns a particular notion of sequential convergence. There is inevitably a
suggestion that there ought to be a topological interpretation of this convergence (see 367Yc, 367Yl), but
I have taken care to avoid spelling one out. I come now to something which really is a topology, and is as
closely involved with order-convergence as any.
Convergence in measure Let (A, ) be a measure algebra. For a A
f
= a : a < and u L
0
=
L
0
(A) set
a
(u) =
_
[u[ a,
a
(u) = (a [[[u[ > ]]). Then the topology of convergence in measure
on L
0
is dened either as the topology generated by the pseudometrics (u, v)
a
(u v) or by saying that
G L
0
is open i for every u G there are a A
f
, > 0 such that v G whenever
a
(u v) .
Remark The sentences above include a number of assertions which need proving. But at this point, rather
than write out any of the relevant arguments, I refer you to 245. Since we know that L
0
(A) can be identied
with L
0
() for a suitable measure space (X, , ) (321J, 364Jc), everything we know about general spaces
L
0
() can be applied directly to L
0
(A) for measure algebras (A, ); and that is what I will do for the next
few paragraphs. So far, all I have done is to write
a
in place of the
F
of 245A, and call on the remarks in
245Bb and 245F.
367N Theorem (a) For any measure algebra (A, ), the topology T of convergence in measure on
L
0
= L
0
(A) is a linear space topology, and any order*-convergent sequence in L
0
is T-convergent to the
same limit.
(b) (A, ) is semi-nite i T is Hausdor.
(c) (A, ) is localizable i T is Hausdor and L
0
is complete under the uniformity corresponding to T.
(d) (A, ) is -nite i T is metrizable.
368 Function spaces 367N
proof 245D, 245Cb, 245E. Of course we need 322B to assure us that the phrases semi-nite, localizable,
-nite here correspond to the same phrases used in 245, and 367G to identify order*-convergence in L
0
with the order-convergence studied in 245.
367O Proposition Let (A, ) be a measure algebra and give L
0
= L
0
(A) its topology of convergence in
measure. If A L
0
is a non-empty, downwards-directed set with inmum 0, then for every neighbourhood
G of 0 in L
0
there is a u A such that v G whenever [v[ u.
proof Let a A
f
, > 0 be such that u G whenever
_
[u[ a (see 245Bb). Since ua : u A is a
downwards-directed set in L
1
= L
1
(A, ) with inmum 0 in L
1
, there must be a u A such that
_
ua
(365Da). But now [u, u] G, as required.
367P Theorem Let U be a Banach lattice and (A, ) a measure algebra. Give L
0
= L
0
(A) its topology
of convergence in measure. If T : U L
0
= L
0
(A) is a positive linear operator, then it is continuous.
proof Take any open set G L
0
. ??? Suppose, if possible, that T
1
[G] is not open. Then we can nd u,
u
n
)
nN
U such that Tu G and |u
n
u| 2
n
, Tu
n
/ G for every n. Set H = GTu; then H is an
open set containing 0 but not T(u
n
u), for any n N. Since

n=0
|u
n
u| < , v =

n=0
n[u
n
u[ is
dened in U, and [T(u
n
u)[
1
n
Tv for every n 1. But by 367N (or otherwise) we know that there is
some n such that w H whenever [w[
1
n
Tv, so that T(u
n
u) H for some n, which is impossible. XXX
367Q Proposition Let (A, ) be a -nite measure algebra.
(a) A sequence u
n
)
nN
in L
0
= L
0
(A) converges in measure to u L
0
i every subsequence of u
n
)
nN
has a sub-subsequence which order*-converges to u.
(b) A set F L
0
is closed for the topology of convergence in measure i u F whenever there is a
sequence u
n
)
nN
in F order*-converging to u L
0
.
proof 245K, 245L.
367R It will be useful later to be able to quote the following straightforward facts.
Proposition Let (A, ) be a measure algebra. Give A its measure-algebra topology (323A) and L
0
= L
0
(A)
the topology of convergence in measure. Then the map : A L
0
is a homeomorphism between A and its
image in L
0
.
proof Of course is injective (364Kc). The measure-algebra topology of A is dened by the pseudometrics

a
(b, c) = (a (bc)), while the topology of L
0
is dened by the pseudometrics
a
(u, v) =
_
[u v[ a,
in both cases taking a to run over elements of A of nite measure; as
a
(b, c) is always equal to
a
(b, c),
we have the result.
367S Proposition Let E R be a Borel set, and h : E R a continuous function. Let (A, ) be a
measure algebra, and

h : Q
E
L
0
= L
0
(A) the associated function, where Q
E
= u : u L
0
, [[u E]] = 1
(364I). Then

h is continuous for the topology of convergence in measure.
proof (Compare 245Dd.) Express (A, ) as the measure algebra of a measure space (X, , ), and write
f

for the element of L

0
corresponding to any f L
0
. Take any u Q
E
, any a A such that a < , and
any > 0. Express u as f

where f : X R is a measurable function, and a as F

where F . Then
f(x) E a.e.(x). For each n N, write E
n
for
t : t E, [h(s) h(t)[
1
3
whenever s E and [s t[ 2
n
.
Then E
n
)
nN
is a non-decreasing sequence of Borel sets with union E, so there is an n such that x : x F,
f(x) / E
n

1
3
.
Now suppose that v Q
E
is such that
_
[v u[ a
1
3
/2
n
. Express v as g

where g : X R is a
measurable function. Then g(x) E for almost every x, and
_
F
min(1, [g(x) f(x)[)(dx)
1
3
/2
n
,
367T Convergence in measure 369
so x : x F, [f(x) g(x)[ > 2
n

1
3
, and
x : x F, [h(g(x)) h(f(x))[ >
1
3

x : x F, f(x) / E
n
x : g(x) / E
x : x F, [f(x) g(x)[ > 2
n

has measure at most

2
3
. But this means that
_
[

h(v)

h(u)[ a =
_
F
min(1, [hg(x) hf(x)[)(dx) .
As u, a and are arbitrary,

h is continuous.
367T Intrinsic description of convergence in measure It is a remarkable fact that the topology
of convergence in measure, not only on L
0
but on its order-dense Riesz subspaces, can be described in
terms of the Riesz space structure alone, without referring at all to the underlying measure algebra or
to integration. (Compare 324H.) There is more than one way of doing this. As far as I know, none is
outstandingly convincing; I present a formulation which seems to me to exhibit some, at least, of the essence
of the phenomenon.
Proposition Let (A, ) be a semi-nite measure algebra, and U an order-dense Riesz subspace of L
0
=
L
0
(A). Suppose that A U and u

U. Then u

belongs to the closure of A for the topology of convergence

in measure i
there is an order-dense Riesz subspace V of U such that
for every v V
+
there is a non-empty downwards-directed B U, with inmum 0, such
that
for every w B there is a u A such that
[u u

[ v w.
proof (a) Suppose rst that u

A. Take V to be U L
1
(A, ); then V is an order-dense Riesz subspace of
L
0
, by 352Nc, and is therefore order-dense in U. (This is where I use the hypothesis that (A, ) is semi-nite,
so that L
1
is order-dense in L
0
, by 365Ga.)
Take any v V
+
. For each n N, set a
n
= [[v > 2
n
]] A
f
. Because u

A, there is a u
n
A such
that b
n
2
n
, where
b
n
= a
n
[[[u
n
u

[ > 2
n
]] = [[[u
n
u

[ v > 2
n
]].
Set c
n
= sup
in
b
i
; then c
n
2
n+1
for each n, so inf
nN
c
n
= 0 and inf
nN
w
n
= 0 in L
0
, where
w
n
= v c
n
+ 2
n
1. Also [u
n
u

[ v w
n
for each n.
The w
n
need not belong to U, so we cannot set B = w
n
: n N. But if instead we write
B = w : w U, w v w
n
for some n N,
then B is non-empty and downwards-directed (because w
n
)
nN
is non-increasing); and
inf B = v supv w : w B
= v supw : w U, w (v w
n
)
+
for some n N
= v sup
nN
(v w
n
)
+
(because U is order-dense in L
0
)
= 0.
Since for every w B there is an n such that v [u
n
u

[ v w
n
w, B witnesses that the condition is
satised.
(b) Now suppose that the condition is satised. Fix a A
f
, > 0. Because V is order-dense in U and
therefore in L
0
, there is a v V such that 0 v a and
_
v a . Let B be a downwards-directed
370 Function spaces 367T
set, with inmum 0, such that for every w B there is a u A with v [u u

[ w. Then there is a
w B such that
_
w v . Now there is a u A such that [u u

[ v w, so that
_
[u u

[ a +
_
[u u

[ v +
_
w v 2.
As a and are arbitrary, u

A.
*367U Theorem Let (A, ) be a semi-nite measure algebra; write L
1
for L
1
(A, ). Let P : (L
1
)

L
1
be the linear operator corresponding to the band projection from (L
1
)

= (L
1
)

onto (L
1
)

and the
canonical isomorphism between L
1
and (L
1
)

. For A L
1
write A

for the weak* closure of the image of

A in (L
1
)

. Then for every A L

1
P[A

] (A),
where (A) is the convex hull of A and (A) is the closure of (A) in L
0
= L
0
(A) for the topology of
convergence in measure.
proof (a) The statement of the theorem includes a number of assertions: that (L
1
)

= (L
1
)

; that (L
1
)

=
((L
1
)

; that the natural embedding of L

1
into (L
1
)

= (L
1
)

identies L
1
with (L
1
)

; and that (L
1
)

is a band in (L
1
)

. For proofs of these see 365C, 356B, 356D, 356N and 356P.
Now for the new argument. First, observe that the statement of the theorem involves the measure algebra
(A, ) and the space L
0
only in the denition of convergence in measure; everything else depends only on
the Banach lattice structure of L
1
. And since we are concerned only with the question of whether members
of P[A

], which is surely a subset of L

1
, belong to (A), 367T shows that this also can be answered in terms
of the Riesz space structure of L
1
. What this means is that we can suppose that (A, ) is localizable. PPP
Let (

A, ) be the localization of (A, ) (322P). The natural expression of A as an order-dense subalgebra of

A identies A
f
= a : a A, a < with

A
f
(322O), so that L
1
(A, ) becomes identied with L
1
(

A, ),
by 365Od. Thus we can think of L
1
as L
1
(

A, ), and (

A, ) is localizable. QQQ
(b) Take A

and set u
0
= P; I have to show that u
0
(A). Write R for the canonical map from
L
1
to (L
1
)

, so that belongs to the weak* closure of R[A].

Consider rst the case u
0
= 0. Take any c A
f
and > 0. We know that (L
1
)

= (L
1
)

= (L
1
)

can
be identied with L

= L

(A) (365Jc), so that (L

= (L

must be in the band orthogonal to

(L

. Now we can identify (L

with the Riesz space M of bounded additive functionals on A, and

if we do so then (L

corresponds to the space M

of completely additive functionals (363K). Writing

P

: M M

for the band projection, we must have P

() = 0, where M is dened by setting

a = (a) for each a A; consequently P

([[) = 0 and there is an upwards-directed family C A, with

supremum 1, such that [[(a) = 0 for every a C (362D). Since c = sup
aC
(a c), there is an a C
such that (c \ a) .
Consider the map Q : L
1
L
1
dened by setting Qw = w a for every w L
1
. Then its adjoint
Q
t
: L

(3A5Ed) can be dened by the same formula: Q

t
v = v a for every v L

. Since
[[ (L

corresponds to [[ M, we have
[(Q
t
v)[ |v|

[[(a) = |v|

[[(a) = 0
for every v L

, and Q
tt
= 0, where Q
tt
: (L

(L

is the adjoint of Q
t
. Since Q
tt
is continuous for
the weak* topology on (L

, 0 Q
tt
[R[A]], where Q
tt
R[A] is the closure for the weak* topology of (L

.
But of course Q
tt
R = RQ, while the weak* topology of (L

corresponds, on the image R[L

1
] of L
1
, to
the weak topology of L
1
; so that 0 belongs to the closure of Q[A] for the weak topology of L
1
.
Because Q is linear, Q[(A)] is convex. Since 0 belongs to the closure of Q[(A)] for the weak topology
of L
1
, it belongs to the closure of Q[(A)] for the norm topology (3A5Ee). So 0 belongs to the closure of
Q[(A)] for the norm topology, and there is a w (A) such that |w a|
1

2
. But this means that
(a [[[w[ ]]) and (c [[[w[ ]]) 2. Since c and are arbitrary, 0 (A).
(c) This deals with the case u
0
= 0. Now the general case follows at once if we set B = A u
0
and
observe that Ru
0
B

, so
0 = P( Ru
0
) (B) = (A) u
0
= (A) u
0
.
367Xf Convergence in measure 371
Remark This is a version of a theorem from Bukhvalov 95.
*367V Corollary Let (A, ) be a localizable measure algebra. Let ( be a family of convex subsets of
L
0
= L
0
(A), all closed for the topology of convergence in measure, with the nite intersection property, and
suppose that for every non-zero a A there are a non-zero b a and a C ( such that sup
uC
_
b
[u[ < .
Then

( , = .
proof Because ( has the nite intersection property, there is an ultralter T on L
0
including (. Set
I = a : a A, inf
FJ
sup
uF
_
a
[u[ < ;
because T is a lter, I is an ideal in A, and the condition on ( tells us that I is order-dense. For each a I,
dene Q
a
: L
0
L
0
by setting Q
a
u = u a. Then there is an F T such that Q
a
[F] is a norm-bounded
set in L
1
, so
a
= lim
uJ
RQ
a
u is dened in (L

for the weak* topology on (L

, writing R for the

canonical map from L
1
to (L

)

= (L
1
)

. If P : (L

L
1
is the map corresponding to the band
projection

P from (L

onto (L

, as in 367U, and C (, then 367U tells us that P(

a
) must belong
to the closure of the convex set Q
a
[C] for the topology of convergence in measure. Moreover, if a b I,
so that Q
a
= Q
a
Q
b
, then P(
a
) = Q
a
P(
b
). PPP Q
a
L
1
is a band projection on L
1
, so its adjoint Q
t
a
is a
band projection on L

= (L
1
)

(356C) and Q
tt
a
is a band projection on (L

)

= (L

. This means that

Q
tt
a
will commute with

P (352Sb). But also Q
tt
a
is continuous for the weak* topology of (L

, so
Q
tt
a
(
b
) = lim
uJ
Q
tt
a
RQ
b
u = lim
uJ
RQ
a
Q
b
u =
a
,
and
P(
a
) = R
1

P(
a
) = R
1

PQ
tt
a
(
b
) = R
1
Q
tt
a

P(
b
) = Q
a
R
1

P(
b
) = Q
a
P(
b
). QQQ
What this means is that if we take a partition D of unity included in I (313K), so that L
0
=

dD
L
0
(A
d
)
(315F(iii), 364S), and dene w L
0
by saying that w d = P(
d
) for every d D, then we shall have
w a = P(
a
) for every a I. But now, given a A
f
and > 0 and C (, there is a b I
such that (a \ b) ; w b Q
b
[C], so there is a u C such that (b [[[w u[ ]]) ; and
(a [[[w u[ ]]) 2. As a, are arbitrary and C is closed, w C; as C is arbitrary, w

( and

( ,= .
367X Basic exercises >>>(a) Let P be a lattice. (i) Show that if p P and p
n
)
nN
is a non-decreasing
sequence in P, then p
n
)
nN
is order*-convergent to p i p = sup
nN
p
n
. (ii) Suppose that p
n
)
nN
a sequence
in P order*-converging to p P. Show that p = sup
nN
pp
n
= inf
nN
pp
n
. (iii) Let p
n
)
nN
, q
n
)
nN
be
two sequences in P which are order*-convergent to p, q respectively. Show that if p
n
q
n
for every n then
p q. (iv) Let p
n
)
nN
be a sequence in P. Show that p
n
)
nN
order*-converges to p P i p
n
p)
nN
and p
n
p)
nN
order*-converge to p.
(b) Let P and Q be lattices, and f : P Q an order-preserving function. Suppose that p
n
)
nN
is an
order-bounded sequence which order*-converges to p in P. Show that f(p
n
))
nN
order*-converges to f(p)
in Q if either f is order-continuous or P is Dedekind -complete and f is sequentially order-continuous.
(c) Let P be either a Boolean algebra or a Riesz space. Suppose that p
n
)
nN
is a sequence in P such
that p
2n
)
nN
and p
2n+1
)
nN
are both order*-convergent to p P. Show that p
n
)
nN
is order*-convergent
to p. (Hint: 313B, 352E.)
>>>(d) Let A be a Boolean algebra and a
n
)
nN
, b
n
)
nN
two sequences in A order*-converging to a,
b respectively. Show that a
n
b
n
)
nN
, a
n
b
n
)
nN
, a
n
\ b
n
)
nN
, a
n
.b
n
)
nN
order*-converge to a b,
a b, a \ b and a .b respectively.
(e) Let A be a Boolean algebra and a
n
)
nN
a sequence in A. Show that a
n
)
nN
does not order*-converge
to 0 i there is a non-zero a A such that a = sup
in
a a
i
for every n N.
>>>(f ) (i) Let U be a Riesz space and u
n
)
nN
an order*-convergent sequence in U
+
with limit u. Show that
h(u) liminf
n
h(u
n
) for every h (U

)
+
. (ii) Let U be a Riesz space and u
n
)
nN
an order-bounded
order*-convergent sequence in U with limit u. Show that h(u) = lim
n
h(u
n
) for every h U

. (Compare
356Xd.)
372 Function spaces 367Xg
>>>(g) Let U be a Riesz space with a Fatou norm | |. (i) Show that if u
n
)
nN
is an order*-convergent
sequence in U with limit u, then |u| liminf
n
|u
n
|. (Hint: [u
n
[ [u[)
nN
is order*-convergent to
[u[.) (ii) Show that if u
n
)
nN
is a norm-convergent sequence in U it has an order*-convergent subsequence.
(Hint: if

n=0
|u
n
| < then u
n
)
nN
order*-converges to 0.)
(h) Let U and V be Archimedean Riesz spaces and T : U V an order-continuous Riesz homomorphism.
Show that if u
n
)
nN
is a sequence in U which order*-converges to u U, then Tu
n
)
nN
order*-converges
to Tu in V .
(i) Let A be a Boolean algebra and B an order-closed subalgebra. Show that if b
n
)
nN
is a sequence in
B and b B, then b
n
)
nN
order*-converges to b in B i it order*-converges to b in A.
(j) Let A be a Dedekind -complete Boolean algebra and u
n
)
nN
, v
n
)
nN
two sequences in L
0
(A) which
are order*-convergent to u, v respectively. Show that u
n
v
n
)
nN
order*-converges to u v. Show that if
u, u
n
all have multiplicative inverses u
1
, u
1
n
then u
1
n
)
nN
order*-converges to u
1
.
(k) Let A be a Dedekind -complete Boolean algebra and 1 a -ideal of A. Show that for any a
n
)
nN
,
a A, a

n
)
nN
order*-converges to a

in A/1 i inf
nN
sup
mn
a
m
a 1.
>>>(l) Let A be a Dedekind -complete Boolean algebra, and h
n
)
nN
a sequence of Borel measurable
functions from R to itself such that h(t) = lim
n
h
n
(t) is dened for every t R. Show that

h
n
(u))
nN
order*-converges to

h(u) for every u L
0
= L
0
(A), where

h
n
,

h : L
0
L
0
are dened as in 364I.
(m) Let (A, ) be a measure algebra, and u
n
)
nN
a sequence in L
1
= L
1
(A, ) which is order*-convergent
to u L
1
. Show that u
n
)
nN
is norm-convergent to u i u
n
: n N is uniformly integrable i |u|
1
=
lim
n
|u
n
|
1
. (Hint: 245H, 246J.)
(n) Let U be an L-space and u
n
)
nN
a norm-bounded sequence in U. Show that there are a v U and
a subsequence v
n
)
nN
of u
n
)
nN
such that
1
n+1

n
i=0
w
i
)
nN
order*-converges to v for every subsequence
w
n
)
nN
of v
n
)
nN
. (Hint: 276H.)
>>>(o) Let (A, ) be a measure algebra and give L
0
= L
0
(A) its topology of convergence in measure. Show
that u [u[, (u, v) u v, (u, v) u v are continuous.
(p) Let (A, ) be a measure algebra and p [1, [. For v (L
p
)
+
= L
p
(A, )
+
dene
v
: L
0
L
0
[0, [
by setting
v
(u
1
, u
2
) = |[u
1
u
2
[ v|
p
for all u
1
, u
2
U. Show that each
v
is a pseudometric and that
the topology on L
0
= L
0
(A) dened by
v
: v (L
p
)
+
is the topology of convergence in measure.
(q) Let (A, ) be a -nite measure algebra. Suppose we have a double sequence u
ij
)
(i,j)NN
in L
0
=
L
0
(A) such that u
ij
)
jN
order*-converges to u
i
in L
0
for each i, while u
i
)
iN
order*-converges to u. Show
that there is a strictly increasing sequence n(i))
iN
such that u
i,n(i)
)
iN
order*-converges to u.
(r) Let (X, , ) be a semi-nite measure space. Show that L
0
() is separable for the topology of
convergence in measure i is -nite and has countable Maharam type. (Cf. 365Xp.)
(s) Let (A, ) be a measure algebra. (i) Show that if a
n
)
nN
is order*-convergent to a A, then
a
n
)
nN
a for the measure-algebra topology. (ii) Show that if (A, ) is -nite, then () a sequence
converges to a for the topology of A i every subsequence has a sub-subsequence which is order*-convergent
to to a () a set F A is closed for the topology of A i a F whenever there is a sequence a
n
)
nN
in F
which is order*-convergent to a A.
(t) Let (A, ) be a measure algebra which is not -nite. Show that there is a set A L
0
(A) such
that the limit of any order*-convergent sequence in A belongs to A, but A is not closed for the topology of
convergence in measure.
(u) Let U be a Banach lattice with an order-continuous norm. (i) Show that a sequence u
n
)
nN
is
norm-convergent to u U i every subsequence has a sub-subsequence which is order-bounded and order*-
convergent to u. (ii) Show that a set F U is closed for the norm topology i u F whenever there is an
order-bounded sequence u
n
)
nN
in F order*-converging to u U.
367Yj Convergence in measure 373
(v) Let (A, ) be a probability algebra. For u L
0
= L
0
(A) let
u
be the distribution of u (364Xd).
Show that u
u
is continuous when L
0
is given the topology of convergence in measure and the space of
probability distributions on R is given the vague topology (274Ld).
(w) Let (A, ) be a probability algebra and u
n
)
nN
a stochastically independent sequence in L
0
, all
with the Cauchy distribution
C,1
with centre 0 and scale parameter 1 (285Xm). For each n let C
n
be the
convex hull of u
i
: i n, and C
n
its closure for the topology of convergence in measure. Show that every
u C
0
has distribution
C,1
. (Hint: consider rst u C
0
.) Show that C
0
is bounded for the topology of
convergence in measure. Show that

nN
C
n
= .
(x) If U is a linear space and C U is a convex set, a function f : C R is convex if f(x+(1)y)
f(x) + (1 )f(y) whenever x, y C and [0, 1]. Let (A, ) be a localizable measure algebra and
C L
1
(A, ) a non-empty convex norm-bounded set which is closed in L
0
(A) for the topology of convergence
in measure. Show that any convex function f : C R which is lower semi-continuous for the topology of
convergence in measure is bounded below and attains its inmum.
367Y Further exercises (a) Give an example of an Archimedean Riesz space U and an order-bounded
sequence u
n
)
nN
in U which is order*-convergent to 0, but such that there is no non-increasing sequence
v
n
)
nN
, with inmum 0, such that u
n
v
n
for every n N.
(b) Let P be a distributive lattice. Show that if p
n
)
nN
is a sequence in P order*-converging to p P,
then p
n
q)
nN
, p
n
q)
nN
order*-converge to p q, p q respectively for any q P.
(c) Let P be any lattice. (i) Show that there is a topology on P for which a set A P is closed i
p A whenever there is a sequence in A which is order*-convergent to p. Show that any closed set for this
topology is sequentially order-closed. (ii) Now let Q be another lattice, with the topology dened in the
same way, and f : P Q an order-preserving function. Show that if f is topologically continuous it is
sequentially order-continuous.
(d) Let us say that a lattice P is (2, )-distributive if () whenever A, B P are non-empty sets with
inma p, q respectively, then infa b : a A, b B = p q () whenever A, B P are non-empty sets
with suprema p, q respectively, then supa b : a A, b B = p q. Show that, in this case, if p
n
)
nN
order*-converges to p and q
n
)
nN
order*-converges to q, p
n
q
n
)
nN
order*-converges to p q.
(e) Give an example of a lattice P with two sequences p
n
)
nN
, q
n
)
nN
, both order*-convergent to p,
such that p
n
q
n
)
nN
is not order*-convergent to p.
(f ) (i) Give an example of a Riesz space U with an order-dense Riesz subspace V of U and a sequence
v
n
)
nN
in V such that v
n
)
nN
order*-converges to 0 in V but does not order*-converge in U. (ii) Give
an example of a Riesz space U with an order-dense Riesz subspace V of U and a sequence v
n
)
nN
in V ,
order-bounded in V , such that v
n
)
nN
order*-converges to 0 in U but does not order*-converge in V .
(g) Let U be an Archimedean f-algebra. Show that if u
n
)
nN
, v
n
)
nN
are sequences in U order*-
converging to u, v respectively, then u
n
v
n
)
nN
order*-converges to u v.
(h) Let A be a Dedekind -complete Boolean algebra and r 1. Let E R
r
be a Borel set and write
Q
E
= (u
1
, . . . , u
r
) : [[(u
1
, . . . , u
r
) E]] = 1 L
0
(A)
r
(364Yc). Let h : E R be a continuous function
and

h : Q
E
L
0
the corresponding map (364Yd). Show that if u
n
)
nN
is a sequence in Q
E
which is
order*-convergent to u Q
E
(in the lattice (L
0
)
r
), then

h(u
n
))
nN
is order*-convergent to

h(u).
(i) Let X be a completely regular Baire space (denition: 314Yd), and u
n
)
nN
a sequence in C(X).
Show that u
n
)
nN
order*-converges to 0 in C(X) i x : limsup
n
[u
n
(x)[ > 0 is meager in X.
(j) (i) Give an example of a sequence u
n
)
nN
in C([0, 1]) such that lim
n
u
n
(x) = 0 for every x [0, 1],
but u
n
: n N is not order-bounded in C([0, 1]). (ii) Give an example of an order-bounded sequence
u
n
)
nN
in C(Q) such that lim
n
u
n
(q) = 0 for every q Q, but sup
in
u
i
= Q in C(Q) for every n N.
(iii) Give an example of a sequence u
n
)
nN
in C([0, 1]) such that u
n
)
nN
order*-converges to 0 in C([0, 1]),
but lim
n
u
n
(q) > 0 for every q Q [0, 1].
374 Function spaces 367Yk
(k) Write out an alternative proof of 367K/367Yi based on the fact that, for a Baire space X, C(X) can
be identied with an order-dense Riesz subspace of a quotient of the space of -measurable functions, where
is the algebra of subsets of X with the Baire property, as in 364Yi.
(l) Let A be a ccc weakly (, )-distributive Boolean algebra. Show that there is a topology on A such
that the closure of any A A is precisely the set of limits of order*-convergent sequences in A.
(m) Give an example of a set X and a double sequence u
mn
)
m,nN
in R
X
such that lim
n
u
mn
(x) =
u
m
(x) exists for every m N and x X, lim
m
u
m
(x) = 0 for every x X, but there is no sequence
v
k
)
kN
in u
mn
: m, n N such that lim
k
v
k
(x) = 0 for every x.
(n) Let U be any Banach lattice with an order-continuous norm. For v U
+
dene
v
: U U [0, [
by setting
v
(u
1
, u
2
) = |[u
1
u
2
[ v| for all u
1
, u
2
U. Show that every
v
is a pseudometric on U, and
that
v
: v U
+
denes a Hausdor linear space topology on U.
(o) Let U be any Riesz space. For h (U

c
)
+
(356Ab), v U
+
dene
vh
: U U [0, [ by setting

vh
(u
1
, u
2
) = h([u
1
u
2
[ v) for all u
1
, u
2
U. Show that each
vh
is a pseudometric on U, and that

vh
: h (U

c
)
+
, v U
+
denes a linear space topology on U.
(p) Let (A, ) be a -nite measure algebra. Show that the function (, u) [[u > ]] : R L
0
A
is Borel measurable when L
0
= L
0
(A) is given the topology of convergence in measure and A is given its
measure-algebra topology. (Hint: if a A, 0 then (, u) : (a [[u > ]]) > is open.)
(q) Let G be the regular open algebra of R. Show that there is no Hausdor topology T on L
0
(G) such
that u
n
)
nN
is T-convergent to u whenever u
n
)
nN
is order*-convergent to u. (Hint: Let H be any T-open
set containing 0. Enumerate Q as q
n
)
nN
. Find inductively a non-decreasing sequence G
n
)
nN
in G such
that G
n
H, q
n
G
n
for every n. Conclude that R H.)
(r) Give an example of a Banach lattice with a norm which is not order-continuous, but in which every
order-bounded order*-convergent sequence is norm-convergent.
(s) Let A be a Dedekind -complete Boolean algebra and r 1. Let E R
r
be a Borel set and write
Q
E
= (u
1
, . . . , u
r
) : [[(u
1
, . . . , u
r
) E]] = 1 L
0
(A)
r
(364Yc). Let h : E R be a continuous function
and

h : Q
E
L
0
the corresponding map (364Yd). Show that if

h is continuous if L
0
is given its topology
of convergence in measure and (L
0
)
r
the product topology.
(t) Show that 367U is true for all measure algebras, whether semi-nite or not.
367 Notes and comments I have given a very general denition of order*-convergence. The general
theory of convergence structures on ordered spaces is complex and full of traps for the unwary. I have
tried to lay out a safe path to the results which are important in the context of this book. But the
propositions here are necessarily full of little conditions (e.g., the requirement that U should be Archimedean,
in 367F) whose signicance may not be immediately obvious. In particular, the denition is very much better
adapted to distributive lattices than to others (367Yb, 367Yd, 367Ye). It is useful in the study of Riesz
spaces and Boolean algebras largely because these satisfy strong distributive laws (313B, 352E). The special
feature which distinguishes the denition here from other denitions of order-convergence is the fact that
it can be applied to sequences which are not order-bounded. For order-bounded sequences there are useful
simplications (367Be-f), but the Martingale Theorem (357J), for instance, if we want to express it in terms
of its natural home in the Riesz space L
1
, refers to sequences which are hardly ever order-bounded.
The * in the phrase order*-convergent is supposed to be a warning that it may not represent exactly the
concept you expect. I think nearly any author using the phrase order-convergent would accept sequences
fullling the conditions of 367Bf; but beyond this no standard denitions have taken root.
The fact that order*-convergent sequences in an L
0
space are order-bounded (367H) is actually one of
the characteristic properties of L
0
. Related ideas will be important in the next section (368A, 368M).
It is one of the outstanding characteristics of measure algebras in this context that they provide non-
trivial linear space topologies on their L
0
spaces, related in striking ways to the order structure. Not all L
0
368B Embedding Riesz spaces in L
0
375
spaces have such topologies (367Yq). It is not known whether a topology corresponding to convergence in
measure can be dened on L
0
(A) for any A which is not a measure algebra; this is the control measure
problem, which I will discuss in 393 (see 393L).
367T shows that the topology of convergence in measure on L
0
(A) is (at least for semi-nite measure
algebras) determined by the Riesz space structure of L
0
; and that indeed the same is true of its order-dense
Riesz subspaces. This fact is important for a full understanding of the representation theorems in 369
below. If a Riesz space U can be embedded as an order-dense subspace of any such L
0
, then there is already
a topology of convergence in measure on U, independent of the embedding. It is therefore not surprising
that there should be alternative descriptions of the topology of convergence in measure on the important
subspaces of L
0
(367Xp, 367Yn).
For -nite measure algebras, the topology of convergence in measure is easily described in terms of
order-convergence (367Q). For other measure algebras, the formula fails (367Xt). 367Yq shows that trying
to apply the same ideas to Riesz spaces in general gives rise to some very curious phenomena.
367V enables us to prove results which would ordinarily be associated with some form of compactness. Of
course compactness is indeed involved, as the proof through 367U makes clear; but it is weak* compactness
in (L
1
)

, rather than in the space immediately to hand.

I hardly mention uniform integrability in this section, not because it is uninteresting, but because I have
nothing to add at this point to 246J and the exercises in 246. But I do include translations of Lebesgues
Dominated Convergence Theorem (367I) and the Martingale Theorem (367J) to show how these can be
expressed in the language of this chapter.
368 Embedding Riesz spaces in L
0
In this section I turn to the representation of Archimedean Riesz spaces as function spaces. Any Archi-
medean Riesz space U can be represented as an order-dense subspace of L
0
(A), where A is its band algebra
(368E). Consequently we get representations of Archimedean Riesz spaces as quotients of subspaces of R
X
(368F) and as subspaces of C

(X) (368G), and a notion of Dedekind completion (368I-368J). Closely

associated with these is the fact that we have a very general extension theorem for order-continuous Riesz
homomorphisms into L
0
spaces (368B). I give a characterization of L
0
spaces in terms of lateral completeness
(368M, 368Yd), and I discuss weakly (, )-distributive Riesz spaces (368N-368S).
368A Lemma Let A be a Dedekind -complete Boolean algebra, and A (L
0
)
+
a set with no upper
bound in L
0
, where L
0
= L
0
(A). If either A is countable or A is Dedekind complete, there is a v > 0 in L
0
such that nv = sup
uA
u nv for every n N.
proof The hypothesis A is countable or A is Dedekind complete ensures that c

= sup
uA
[[u > ]] is
dened for each . By 364Ma, c = inf
nN
c
n
= inf
R
c

is non-zero. Now for any n 1, R

[[sup
uA
(u nc) > ]] = sup
uA
[[u > ]] [[c >

n
]] = [[c >

n
]],
because if 0 k N then
sup
uA
[[u > ]] c
k
c [[c >

n
]],
while if < 0 then (because A is a non-empty subset of (L
0
)
+
)
sup
uA
[[u > ]] = 1 = [[c >

n
]].
So sup
uA
u nc = nc for every n 1, and we can take v = c. (The case n = 0 is of course trivial.)
368B Theorem Let A be a Dedekind complete Boolean algebra, U an Archimedean Riesz space, V an
order-dense Riesz subspace of U and T : V L
0
= L
0
(A) an order-continuous Riesz homomorphism. Then
T has a unique extension to an order-continuous Riesz homomorphism

T : U L
0
.
proof (a) The key to the proof is the following: if u 0 in U, then Tv : v V, 0 v u is
bounded above in L
0
. PPP??? Suppose, if possible, otherwise. Then by 368A there is a w > 0 in L
0
such that
376 Function spaces 368B
nw = sup
vA
nw Tv for every n N, where A = v : v V, 0 v u. In particular, there is a v
0
A
such that w
0
= wTv
0
> 0. Because U is Archimedean, inf
k1
1
k
u = 0, so v
0
= sup
k1
(v
0

1
k
u)
+
. Because
V is order-dense in U, v
0
= sup B where
B = v : v V, 0 v (v
0

1
k
u)
+
for some k 1.
Because T is order-continuous, Tv
0
= sup T[B] in L
0
, and there is a v
1
B such that w
1
= w
0
Tv
1
> 0.
Let k 1 be such that v
1
(v
0

1
k
u)
+
. Then for any m N,
mv
1
u (mv
1
kv
0
) + (mv
1
(u kv
0
)
+
)
(352Fa)
kv
0
+ (m+k)(v
1
(
1
k
u v
0
)
+
) = kv
0
.
So for any v A, m N,
mw
1
Tv = mw
1
mTv
1
Tv T(mv
1
v) T(mv
1
u) T(kv
0
) = kTv
0
.
But this means that, for m N,
mw
1
= mw
1
mw = sup
vA
mw
1
(mw Tv) = sup
vA
mw
1
Tv kTv
0
,
which is impossible because L
0
is Archimedean and w
1
> 0. XXXQQQ
(b) Because L
0
is Dedekind complete, supTv : v V, 0 v u is dened in L
0
for every u U. By
355F, T has a unique extension to an order-continuous Riesz homomorphism from U to L
0
.
368C Corollary Let A and B be Dedekind complete Boolean algebras and U, V order-dense Riesz
subspaces of L
0
(A), L
0
(B) respectively. Then any Riesz space isomorphism between U and V extends
uniquely to a Riesz space isomorphism between L
0
(A) and L
0
(B); and in this case A and B must be
isomorphic as Boolean algebras.
proof If T : U V is a Riesz space isomorphism, then 368B tells us that we have (unique) order-continuous
Riesz homomorphisms S : L
0
(A) L
0
(B) and S
t
: L
0
(B) L
0
(A) extending T, T
1
respectively. Now
S
t
S : L
0
(A) L
0
(A) is an order-continuous Riesz homomorphism agreeing with the identity on U, so must
be the identity on L
0
(A); similarly SS
t
is the identity on L
0
(B), and S is a Riesz space isomorphism. To
see that A and B are isomorphic, recall that by 364Q they can be identied with the algebras of projection
bands of L
0
(A) and L
0
(B), which must be isomorphic.
368D Corollary Suppose that A is a Dedekind -complete Boolean algebra, and that U is an order-
dense Riesz subspace of L
0
(A) which is isomorphic, as Riesz space, to L
0
(B) for some Dedekind complete
Boolean algebra B. Then U = L
0
(A) and A is isomorphic to B (so, in particular, is Dedekind complete).
proof The identity mapping U U is surely an order-continuous Riesz homomorphism, so by 368B extends
to an order-continuous Riesz homomorphism

T : L
0
(A) U. Now

T must be injective, because if u ,= 0 in
L
0
(A) there is a u
t
U such that 0 < u
t
[u[, so that 0 < u
t
[

Tu[. So we must have U = L

0
(A) and

T
the identity map. By 368C, or otherwise, A

= B.
368E Theorem Let U be any Archimedean Riesz space, and A its band algebra (353B). Then U can
be embedded as an order-dense Riesz subspace of L
0
(A).
proof (a) If U = 0 then A = 0, L
0
= L
0
(A) = 0 and the result is trivial; I shall therefore suppose
henceforth that U is non-trivial. Note that by 352Q A is Dedekind complete.
Let C U
+
0 be a maximal disjoint set (in the sense of 352C); to obtain such a set apply Zorns
lemma to the family of all disjoint subsets of U
+
0. Now I can write down the formula for the embedding
T : U L
0
immediately, though there will be a good deal of work to do in justication: for u U and
R, [[Tu > ]] will be the band in U generated by
368E Embedding Riesz spaces in L
0
377
e (u e)
+
: e C.
(For once, I allow myself to use the formula [[. . . ]] without checking immediately that it represents a member
of L
0
; all I claim for the moment is that [[Tu > ]] is a member of A determined by u and .)
(b) Before getting down to the main argument, I make some remarks which will be useful later.
(i) If u > 0 in U, then there is some e C such that ue > 0, since otherwise we ought to have added
u to C. Thus C

= 0.
(ii) If u U and e C and R, then v = e (e u)
+
belongs to [[Tu > ]]

. PPP If e
t
C, then
either e
t
,= e so
v e
t
(u e
t
)
+
e e
t
= 0,
or e
t
= e and
v e
t
(u e
t
)
+
(e u)
+
(u e)
+
= 0.
Accordingly [[Tu > ]] is included in the band v

and v [[Tu > ]]

. QQQ
(c) Now I must conrm that the formula given for [[Tu > ]] is consistent with the conditions laid down
in 364A. PPP Take u U.
(i) If then
0 e (u e)
+
e (u e)
+
[[Tu > ]]
so e (u e)
+
[[Tu > ]] for every e C and [[Tu > ]] [[Tu > ]].
(ii) Given R, set W = sup
>
[[Tu > ]] in A, that is, the band in U generated by e (u e)
+
:
e C, > . Then for each e C,
sup
>
e (u e)
+
= e (u inf
>
e)
+
= e (u e)
+
using the general distributive laws in U (352E), the translation-invariance of the order (351D) and the fact
that U is Archimedean (to see that e = inf
>
e). So e (ue)
+
W; as e is arbitrary, [[Tu > ]] W
and [[Tu > ]] = W.
(iii) Now set W = inf
nN
[[Tu > n]]. For any e C, n N we have
e (ne u)
+
[[Tu > n]]

,
so that
e (e
1
n
u
+
)
+
e (e
1
n
u)
+
W

for every n 1 and

e = sup
n1
e (e
1
n
u
+
)
+
W

.
Thus C W

and W C

= 0. So we have inf
nN
[[Tu > n]] = 0.
(iv) Finally, set W = sup
nN
[[Tu > n]]. Then
e (e
1
n
u

)
+
e (e +
1
n
u)
+
e (u +ne)
+
W
for every n 1 and e C, so
e = sup
n1
e (e
1
n
u

)
+
W
for every e C and W

= 0, W = U. Thus all three conditions of 364A are satised. QQQ

(d) Thus we have a well-dened map T : U L
0
. I show next that T(u+v) = Tu+Tv for all u, v U.
PPP I rely on the formulae in 364E and 364Fa, and on partitions of unity in A, constructed as follows. Fix
n 1 for the moment. Then we know that
sup
iZ
[[Tu >
i
n
]] = 1, inf
iZ
[[Tu >
i
n
]] = 0.
So setting
378 Function spaces 368E
V
i
= [[Tu >
i
n
]] \ [[Tu >
i+1
n
]] = [[Tu >
i
n
]] [[Tu >
i+1
n
]]

,
V
i
)
iZ
is a partition of unity in A. Similarly, W
i
)
iZ
is a partition of unity, where
W
i
= [[Tv >
i
n
]] [[Tv >
i+1
n
]]

.
Now, for any i, j, k Z such that i +j k,
V
i
W
j
[[Tu >
i
n
]] [[Tv >
j
n
]] [[Tu +Tv >
i+j
n
]] [[Tu +Tv >
k
n
]];
thus
[[Tu +Tv >
k
n
]] sup
i+jk
V
i
W
j
.
On the other hand, if q Q and k Z, there is an i Z such that
i
n
q <
i+1
n
, so that
[[Tu > q]] [[Tv >
k+1
n
q]] [[Tu >
i
n
]] [[Tv >
ki
n
]] sup
i+jk
V
i
W
j
;
thus for any k Z
[[Tu +Tv >
k+1
n
]] sup
i+jk
V
i
W
j
[[Tu +Tv >
k
n
]].
Also, if 0 < w V
i
W
j
, e C then
w e (u
i+1
n
e)
+
= w e (v
j+1
n
e)
+
= 0,
so that
w e (u +v
i+j+2
n
e)
+
= 0
because
(u +v
i+j+2
n
e)
+
(u
i+1
n
e)
+
+ (v
j+1
n
e)
+
by 352Fc. But this means that V
i
W
j
[[T(u +v) >
i+j+2
n
]] = 0. Turning this round,
[[T(u +v) >
k+1
n
]] sup
i+jk1
V
i
W
j
= 0,
and because sup
i,jZ
V
i
W
j
= U in A,
[[T(u +v) >
k+1
n
]] sup
i+jk
V
i
W
j
.
Finally, if i +j k and 0 < w V
i
V
j
, then there is an e C such that w
1
= we (u
i
n
e)
+
> 0; there
is an e
t
C such that w
2
= w
1
e
t
(v
j
n
e
t
)
+
> 0; of course e = e
t
, and
0 < w
2
e (u
i
n
e)
+
(v
j
n
)
+
e (u +v
i+j
n
e)
+
[[T(u +v) >
i+j
n
]] [[T(u +v) >
k
n
]]
using 352Fc. This shows that w , [[T(u +v) >
k
n
]]

; as w is arbitrary, V
i
W
j
[[T(u +v) >
k
n
]]; so we get
sup
i+jk
V
i
W
j
[[T(u +v) >
k
n
]].
Putting all these four together, we see that
[[T(u +v) >
k+1
n
]] sup
i+jk
V
i
W
j
[[Tu +Tv >
k
n
]],
[[Tu +Tv >
k+1
n
]] sup
i+jk
V
i
W
j
[[T(u +v) >
k
n
]]
for all n 1, k Z. But this means that we must have
[[Tu +Tv > ]] [[T(u +v) > ]], [[T(u +v) > ]] [[Tu +Tv > ]]
whenever < . Consequently
368F Embedding Riesz spaces in L
0
379
[[Tu +Tv > ]] = sup
>
[[Tu +Tv > ]] [[T(u +v) > ]]
= sup
>
[[T(u +v) > ]] [[Tu +Tv > ]]
and [[Tu +Tv > ]] = [[T(u +v) > ]] for every , that is, T(u +v) = Tu +Tv. QQQ
(e) The hardest part is over. If u U, > 0 and R, then for any e C
min(1,
1

)(e (u e)
+
) e (u

e)
+
max(1,
1

)(e (u e)
+
),
so
[[T(u) > ]] = [[Tu >

]] = [[Tu > ]];

as is arbitrary, Tu = T(u); as and u are arbitrary, T is linear. (We need only check linearity for > 0
because we know from the additivity of T that T(u) = Tu for every u.)
(f ) To see that T is a Riesz homomorphism, take any u U and R and consider the band
[[Tu > ]] [[Tu > ]] = [[[Tu[ > ]] (by 364Mb). This is the band generated by e (u e)
+
: e
C e (u e)
+
: e C. But this must also be the band generated by
(e (u e)
+
) (e (u e)
+
) : e C = e ([u[ e)
+
: e C,
which is [[T[u[ > ]]. Thus [[[Tu[ > ]] = [[T[u[ > ]] for every and [Tu[ = T[u[. As u is arbitrary, T is a
Riesz homomorphism.
(g) To see that T is injective, take any non-zero u U. Then there must be some e C such that
[u[ e ,= 0, and some > 0 such that [u[ e , e, so that e ([u[ e)
+
,= 0 and [[T[u[ > ]] ,= 0 and
T[u[ ,= 0 and Tu ,= 0.
Thus T embeds U as a Riesz subspace of L
0
.
(h) Finally, I must check that T[U] is order-dense in L
0
. PPP Let p > 0 in L
0
. Then there is some > 0
such that V = [[p > ]] ,= 0. Take u > 0 in V . Let e C be such that u e > 0. Then v = u e > 0. Now
e (v e)
+
= 0; but also e
t
v = 0 for every e
t
C distinct from e, so that [[Tv > ]] = 0. Also v V ,
so e
t
(v e
t
)
+
V whenever e
t
C and 0, and [[Tv > ]] V for every 0. Accordingly we have
[[Tv > ]] = 0 [[p > ]] if ,
V [[p > ]] if 0 < ,
= U = [[p > ]] if < 0,
and Tv p. Also Tv > 0, by (g). As p is arbitrary, T[U] is order-dense in L
0
. QQQ
368F Corollary A Riesz space U is Archimedean i it is isomorphic to a Riesz subspace of some
reduced power R
X
[T, where X is a set and T is a lter on X such that

nN
F
n
T whenever F
n
)
nN
is
a sequence in T.
proof (a) If U is an Archimedean Riesz space, then by 368E there is a space of the form L
0
= L
0
(A) such
that U can be embedded into L
0
. As in the proof of 364E, L
0
is isomorphic to some space of the form
L
0
()/W, where is a -algebra of subsets of some set X and W = f : f L
0
, x : f(x) ,= 0 1, 1
being a -ideal of . But now T = A : AE = X for some E 1 is a lter on X such that

nN
F
n
T
for every sequence F
n
)
nN
in T. (I am passing over the trivial case X 1, since then U must be 0.)
And L
0
/W is (isomorphic to) the image of L
0
in R
X
[T, since W= f : f L
0
, x : f(x) = 0 T. Thus
U is isomorphic to a Riesz subspace of R
X
[T.
(b) On the other hand, if T is a lter on X closed under countable intersections, then W = f : f
R
X
, x : f(x) = 0 T is a sequentially order-closed solid linear subspace of the Dedekind -complete
Riesz space R
X
, so that R
X
[T = R
X
/W is Dedekind -complete (353J(a-iii)) and all its Riesz subspaces
must be Archimedean (353H, 351Rc).
380 Function spaces 368G
368G Corollary Every Archimedean Riesz space U is isomorphic to an order-dense Riesz subspace of
some space C

(X), where X is an extremally disconnected compact Hausdor space.

proof Let Z be the Stone space of the band algebra A of U. Because A is Dedekind complete (352Q), Z
is extremally disconnected and A can be identied with the regular open algebra G of Z (314S). By 364W,
L
0
(G) can be identied with C

(Z). So the embedding of U as an order-dense Riesz subspace of L

0
(A)
(368E) can be regarded as an embedding of U as an order-dense Riesz subspace of C

(Z).
368H Corollary Any Dedekind complete Riesz space U is isomorphic to an order-dense solid linear
subspace of L
0
(A) for some Dedekind complete Boolean algebra A.
proof Embed U in L
0
= L
0
(A) as in 368E; because U is order-dense in L
0
and (in itself) Dedekind complete,
it is solid (353K).
368I Corollary Let U be an Archimedean Riesz space. Then U can be embedded as an order-dense
Riesz subspace of a Dedekind complete Riesz space V in such a way that the solid linear subspace of V
generated by U is V itself, and this can be done in essentially only one way. If W is any other Dedekind
complete Riesz space and T : U W is an order-continuous positive linear operator, there is a unique
positive linear operator

T : V W extending T.
proof By 368E, we may suppose that U is actually an order-dense Riesz subspace of L
0
(A), where A is a
Dedekind complete Boolean algebra. In this case, we can take V to be the solid linear subspace generated
by U, that is, v : [v[ u for some u U; being a solid linear subspace of the Dedekind complete Riesz
space L
0
(A), V is Dedekind complete, and of course U is order-dense in V .
If W is any other Dedekind complete Riesz space and T : U W is an order-continuous positive linear
operator, then for any v V
+
there is a u
0
U such that v u
0
, so that Tu
0
is an upper bound for
Tu : u U, 0 u v; as W is Dedekind complete, sup
uU,0uv
Tu is dened in W. By 355F, T has a
unique extension to an order-continuous positive linear operator from V to W.
In particular, if V
1
is another Dedekind complete Riesz space in which U can be embedded as an order-
dense Riesz subspace, this embedding of U extends to an embedding of V ; since V is Dedekind complete,
its copy in V
1
must be a solid linear subspace, so if V
1
is the solid linear subspace of itself generated by U,
we get an identication between V and V
1
, uniquely determined by the embeddings of U in V and V
1
.
368J Denition If U is an Archimedean Riesz space, a Dedekind completion of U is a Dedekind
complete Riesz space V together with an embedding of U in V as an order-dense Riesz subspace of V such
that the solid linear subspace of V generated by U is V itself. 368I tells us that every Archimedean Riesz
space U has an essentially unique Dedekind completion, so that we may speak of the Dedekind completion
of U.
368K This is a convenient point at which to give a characterization of the Riesz spaces L
0
(A).
Lemma Let A be a Dedekind -complete Boolean algebra. Suppose that A L
0
(A)
+
is disjoint. If either
A is countable or A is Dedekind complete, A is bounded above in L
0
(A).
proof If A = , this is trivial; suppose that A is not empty. For n N, set a
n
= sup
uA
[[u > n]]; this
is always dened; set a = inf
nN
a
n
. Now a = 0. PPP??? Otherwise, there must be a u A such that
a
t
= a [[u > 0]] ,= 0, since a a
0
. But now, for any n, and any v A u,
a
t
[[v > n]] [[u > 0]] [[v > 0]] = 0,
so that a
t
[[u > n]]. As n is arbitrary, inf
nN
[[u > n]] ,= 0, which is impossible. XXXQQQ
By 364Ma, A is bounded above.
368L Denition A Riesz space U is called laterally complete or universally complete if A is
bounded above whenever A U
+
is disjoint.
368N Embedding Riesz spaces in L
0
381
368M Theorem Let U be an Archimedean Riesz space. Then the following are equiveridical:
(i) there is a Dedekind complete Boolean algebra A such that U is isomorphic to L
0
(A);
(ii) U is Dedekind -complete and laterally complete;
(iii) whenever V is an Archimedean Riesz space, V
0
is an order-dense Riesz subspace of V and T
0
: V
0
U
is an order-continuous Riesz homomorphism, there is a positive linear operator T : V U extending T
0
.
proof (a)(i)(ii) and (i)(iii) are covered by 368K and 368B.
(b)(ii)(i) Assume (ii). By 368E, we may suppose that U is actually an order-dense Riesz subspace of
L
0
= L
0
(A) for a Dedekind complete Boolean algebra A.
() If u U
+
and a A then u a U. PPP Set A = v : v U, 0 v a, and let C A be
a maximal disjoint set; then w = supC is dened in U, and is also the supremum in L
0
. Set b = [[w > 0]].
As w a, b a. ??? If b ,= a, then (a \ b) > 0, and there is a v
t
U such that 0 < v
t
(a \ b); but
now v
t
A and v
t
w = 0, so v
t
v = 0 for every v C, and we ought to have added v
t
to C. XXX Thus
[[w > 0]] = a.
Now consider u
t
= sup
nN
u nw; as U is Dedekind -complete, u
t
U. Since [[u
t
> 0]] a, u
t
u a.
On the other hand,
u [[w >
1
n
]] [[u n]] u n
2
w u
t
for every n 1, so, taking the supremum over n, u a u
t
. Accordingly
u a = u
t
U,
as required. QQQ
() If w 0 in L
0
, there is a u U such that
1
2
w u w. PPP Set
A = u : u U, 0 u w,
C = a : a A, a [[u
1
2
w 0]] for some u A.
Then sup A = w, so C is order-dense in A. (If a A 0, either a [[w > 0]] = 0 and a [[0
1
2
w 0]], so
a C, or there is a u U such that 0 < u w a. In the latter case there is some n such that 2
n
u w
and 2
n+1
u , w, and now c = a [[2
n
u
1
2
w 0]] is a non-zero member of C included in a.) Let D C be
a partition of unity and for each d D choose u
d
A such that d [[u
d

1
2
w 0]]. By (), u
d
d U
for every d D, so u = sup
dD
u
d
d U. Now u w, but also [[u
1
2
w 0]] d for every d D, so is
equal to 1, and u
1
2
w, as required. QQQ
() Given w 0 in L
0
, we can therefore choose u
n
)
nN
, v
n
)
nN
inductively such that v
0
= 0 and
u
n
U,
1
2
(w v
n
) u
n
w v
n
, v
n+1
= v
n
+u
n
for every n N. Now v
n
)
nN
is a non-decreasing sequence in U and w v
n
2
n
w for every n, so
w = sup
nN
v
n
U.
As w is arbitrary, (L
0
)
+
U and U = L
0
is of the right form.
(c)(iii)(i) As in (b), we may suppose that U is an order-dense Riesz subspace of L
0
. But now apply
condition (iii) with V = L
0
, V
0
= U and T
0
the identity operator. There is an extension T : L
0
U. If
v 0 in L
0
, Tv T
0
= u whenever u U and u v, so Tv v, since v = supu : u U, 0 u v in
L
0
. Similarly, T(Tv v) Tv v. But as Tv U, T(Tv) = Tv and T(Tv v) = 0, so v = Tv U. As v
is arbitrary, U = L
0
.
368N Weakly (, )-distributive Riesz spaces We are now ready to look at the class of Riesz spaces
corresponding to the weakly (, )-distributive Boolean algebras of 316.
Denition Let U be a Riesz space. Then U is weakly (, )-distributive if whenever A
n
)
nN
is a
sequence of non-empty downwards-directed subsets of U
+
, each with inmum 0, and

nN
A
n
has an upper
bound in U, then
u : u U, for every n N there is a v A
n
such that v u
382 Function spaces 368N
has inmum 0 in U.
Remark Because the denition looks only at sequences A
n
)
nN
such that

nN
A
n
is order-bounded, we
can invert it, as follows: a Riesz space U is weakly (, )-distributive i whenever A
n
)
nN
is a sequence of
non-empty upwards-directed subsets of U
+
, all with supremum u
0
, then
u : u U
+
, for every n N there is a v A
n
such that u v
also has supremum u
0
.
368O Lemma Let U be an Archimedean Riesz space. Then the following are equiveridical:
(i) U is not weakly (, )-distributive;
(ii) there are a u > 0 in U and a sequence A
n
)
nN
of non-empty downwards-directed sets, all with
inmum 0, such that sup
nN
u
n
= u whenever u
n
A
n
for every n N.
proof (ii)(i) is immediate from the denition of weakly (, )-distributive. For (i)(ii), suppose that
U is not weakly (, )-distributive. Then there is a sequence A
n
)
nN
of non-empty downwards-directed
sets, all with inmum 0, such that

nN
A
n
is bounded above, but
A = w : w U, for every n N there is a v A
n
such that v w
does not have inmum 0. Let u > 0 be a lower bound for A, and set A
t
n
= u v : v A
n
for each n N.
Then each A
t
n
is a non-empty downwards-directed set with inmum 0. Let u
n
)
nN
be a sequence such that
u
n
A
t
n
for every n. Express each u
n
as u v
n
where v
n
A
n
. Let B be the set of upper bounds of
v
n
: n N. Then inf
wB,nN
w v
n
= 0, because U is Archimedean (353F), while B A, so u w for
every w B. If u
t
is any upper bound for u
n
: n N, then
u u
t
u u v
n
= (u v
n
)
+
(w v
n
)
+
= w v
n
for every n N, w B. So u
t
u. Thus u = sup
nN
u
n
. As u
n
)
nN
is arbitrary, u and A
t
n
)
nN
witness
that (ii) is true.
368P Proposition (a) A regularly embedded Riesz subspace of an Archimedean weakly (, )-dis-
tributive Riesz space is weakly (, )-distributive.
(b) An Archimedean Riesz space with a weakly (, )-distributive order-dense Riesz subspace is weakly
(, )-distributive.
(c) If U is a Riesz space such that U

separates the points of U, then U is weakly (, )-distributive; in

particular, U

and U

are weakly (, )-distributive for every Riesz space U.

proof (a) Suppose that U is an Archimedean Riesz space and that V U is a regularly embedded Riesz
subspace which is not weakly (, )-distributive. Then 368O tells us that there are a v > 0 in V and
a sequence A
n
)
nN
of non-empty downwards-directed subsets of V , all with inmum 0 in V , such that
sup
nN
v
n
= v in V whenever v
n
A
n
for every n N. Because V is regularly embedded in U, inf A
n
= 0
in U for every n and sup
nN
v
n
= v in U for every sequence v
n
)
nN

nN
A
n
, so U is not weakly (, )-
distributive. Turning this round, we have (a).
(b) Let U be an Archimedean Riesz space which is not weakly (, )-distributive, and V an order-
dense Riesz subspace of U. By 368O again, there are a u > 0 in U and a sequence A
n
)
nN
of non-empty
downwards-directed sets in U, all with inmum 0, such that sup
nN
u
n
= u whenever u
n
A
n
for every n.
Let v V be such that 0 < v u. Set
B
n
= w : w V , there is some u A
n
such that v u w v
for each n N. Because A
n
is downwards-directed, ww
t
B
n
for all w, w
t
B
n
; v B
n
, so B
n
,= ; and
inf B
n
= 0 in V . PPP Setting
C = w : w V
+
, there is some u A
n
such that w (v u)
+
,
then (because V is order-dense) any upper bound for C in U is also an upper bound of (v u)
+
: u A
n
.
But
sup
uA
n
(v u)
+
= (v inf A
n
)
+
= v,
368Q Embedding Riesz spaces in L
0
383
so v = supC in U and inf B
n
= infv w : w C = 0 in U and in V . QQQ
Now if v
n
B
n
for every n N, we can choose u
n
A
n
such that v u
n
v
n
v for every n, so that
v = v u = v sup
nN
u
n
= sup
nN
v u
n
sup
nN
v
n
v,
and v = sup
nN
v
n
. Thus B
n
)
nN
witnesses that V is not weakly (, )-distributive.
(c) Now suppose that U

separates the points of U. In this case U is surely Archimedean (356G). ??? If U

is not weakly (, )-distributive, there are a u > 0 in U and a sequence A
n
)
nN
of non-empty downwards-
directed sets, all with inmum 0, such that sup
nN
u
n
= u whenever u
n
A
n
for each n. Take f U

such that f(u) ,= 0; replacing f by [f[ if necessary, we may suppose that f > 0. Set = f(u) > 0. For each
n N, there is a u
n
A
n
such that f(u
n
) 2
n2
. But in this case sup
i<n
u
i
)
nN
is a non-decreasing
sequence with supremum u, so
f(u) = lim
n
f(sup
in
u
i
)

i=0
f(u
i
)
1
2
< f(u),
which is absurd. XXX Thus U is weakly (, )-distributive.
For any Riesz space U, U acts on U

as a subspace of U

(356F); as U surely separates the points of

U

, so does U

. So U

is weakly (, )-distributive. Now U

is a band in U

(356B), so is regularly
embedded, and must also be weakly (, )-distributive, by (a) above.
368Q Theorem (a) For any Boolean algebra A, A is weakly (, )-distributive i S(A) is weakly
(, )-distributive i L

(A) is weakly (, )-distributive.

(b) For a Dedekind -complete Boolean algebra A, L
0
(A) is weakly (, )-distributive i A is weakly
(, )-distributive.
proof (a)(i) ??? Suppose, if possible, that A is weakly (, )-distributive but S = S(A) is not. By 368O, as
usual, we have a u > 0 in S and a sequence A
n
)
nN
of non-empty downwards-directed sets in S, all with
inmum 0, such that u = sup
nN
u
n
whenever u
n
A
n
for every n. Let > 0 be such that c = [[u > ]] ,= 0
(361Eg), and consider
B
n
= [[v > ]] : v A
n
A
for each n N. Then each B
n
is downwards-directed (because A
n
is), and inf B
n
= 0 in A (because if b is
a lower bound of B
n
, b v for every v A
n
). Because A is weakly (, )-distributive, there must be
some a A such that a , c but there is, for every n N, a b
n
B
n
such that a b
n
. Take v
n
A
n
such
that b
n
= [[v
n
> ]], so that
v
n
1 |v
n
|

b
n
1 |u|

a.
Since u = sup
nN
v
n
, u 1 |u|

a. But in this case

c = [[u > ]] a,
contradicting the choice of a. XXX
Thus S must be weakly (, )-distributive if A is.
(ii) Now suppose that S is weakly (, )-distributive, and let B
n
)
nN
be a sequence of non-empty
downwards-directed subsets of A, all with inmum 0. Set A
n
= b : b B
n
for each n; then A
n
S is
non-empty, downwards-directed and has inmum 0 in S, because : A S is order-continuous (361Ef).
Set
A = v : v S, for every n N there is a u A
n
such that u v,
B = b : b A, for every n N there is an a B
n
such that a b.
??? If 0 is not the greatest lower bound of B, take a non-zero lower bound c. Because S is weakly (, )-
distributive, inf A = 0, and there is a v A such that c , v. Express v as

n
i=0

i
a
i
, where a
i
)
in
is
disjoint, and set a = supa
i
: i n,
i
1; then a v, so c , a. For each n there is a b
n
B
n
such that
b
n
v. But in this case b
n
a for each n N, so that a B; which means that c is not a lower bound
for B. XXX
Thus inf B = 0 in A. As B
n
)
nN
is arbitrary, A is weakly (, )-distributive.
384 Function spaces 368Q
(iii) Thus S is weakly (, )-distributive i A is. But S is order-dense in L

= L

(A) (363C),
therefore regularly embedded (352Ne), so 368Pa-b tell us that S is weakly (, )-distributive i L

is.
(b) In the same way, because S can be regarded as an order-dense Riesz subspace of L
0
= L
0
(A) (364K),
L
0
is weakly (, )-distributive i S is, that is, i A is.
368R Corollary An Archimedean Riesz space is weakly (, )-distributive i its band algebra is weakly
(, )-distributive.
proof Let U be an Archimedean Riesz space and A its band algebra. By 368E, U is isomorphic to an
order-dense Riesz subspace of L
0
= L
0
(A). By 368P, U is weakly (, )-distributive i L
0
is; and by 368Qb
L
0
is weakly (, )-distributive i A is.
368S Corollary If (A, ) is a semi-nite measure algebra, any regularly embedded Riesz subspace
(in particular, any solid linear subspace and any order-dense Riesz subspace) of L
0
(A) is weakly (, )-
distributive.
proof By 322F, A is weakly (, )-distributive; by 368Qb, L
0
(A) is weakly (, )-distributive; by 368Pa,
any regularly embedded Riesz subspace is weakly (, )-distributive.
368X Basic exercises (a) Let X be an uncountable set and the countable-cocountable -algebra
of subsets of X. Show that there is a family A L
0
= L
0
() such that u v = 0 for all distinct u, v A
but A has no upper bound in L
0
. Show moreover that if w > 0 in L
0
then there is an n N such that
nw ,= sup
uA
u nw.
(b) Let A be any Boolean algebra, and

A its Dedekind completion (314U). Show that L

A) can be
identied with the Dedekind completions of S(A) and L

(A).
(c) Explain how to prove 368K from 368A.
(d) Show that any product of weakly (, )-distributive Riesz spaces is weakly (, )-distributive.
(e) Let A be a Dedekind complete weakly (, )-distributive Boolean algebra. Show that a set A L
0
=
L
0
(A) is order-bounded i 2
n
u
n
)
nN
order*-converges to 0 in L
0
whenever u
n
)
nN
is a sequence in A.
(Hint: use 368A. If v > 0 and v = sup
uA
v 2
n
u for every n, we can nd a w > 0 and a sequence u
n
)
nN
in A such that w 2
n
u
n
for every n.)
(f ) Give a direct proof of 368S, using the ideas of 322F, but not relying on it or on 368Q.
368Y Further exercises (a) (i) Use 364U-364V to show that if X is any compact Hausdor space
then C(X) can be regarded as an order-dense Riesz subspace of L
0
(G), where G is the regular open algebra
of X. (ii) Use 353M to show that any Archimedean Riesz space with order unit can be embedded as an
order-dense Riesz subspace of some L
0
(G). (iii) Let U be an Archimedean Riesz space and C U
+
a
maximal disjoint set, as in part (a) of the proof of 368E. For e C let U
e
be the solid linear subspace of U
generated by e, and let V be the solid linear subspace of U generated by C. Show that V can be embedded
as an order-dense Riesz subspace of

eC
U
e
and therefore in

eC
L
0
(G
e
)

= L
0
(

eC
G
e
) for a suitable
family of regular open algebras G
e
. (iv) Now use 368B to complete a proof of 368E.
(b) Let U be any Archimedean Riesz space. Let 1 be the family of pairs (A, B) of non-empty subsets of
U such that B is the set of upper bounds of A and A is the set of lower bounds of B. Show that 1 can be
given the structure of a Dedekind complete Riesz space dened by the formulae
(A
1
, B
1
) + (A
2
, B
2
) = (A, B) i A
1
+A
2
A, B
1
+B
2
B,
(A, B) = (A, B) if > 0,
(A
1
, B
1
) (A
2
, B
2
) i A
1
A
2
.
Show that u (], u] , [u, [) denes an embedding of U as an order-dense Riesz subspace of 1, so that
1 may be identied with the Dedekind completion of U.
368 Notes Embedding Riesz spaces in L
0
385
(c) Work through the proof of 364U when X is compact, Hausdor and extremally disconnected, and
show that it is easier than the general case. Hence show that 368Yb can be used to shorten the proof of
368E sketched in 368Ya.
(d) Let U be a Riesz space. Show that the following are equiveridical: (i) U is isomorphic, as Riesz space,
to L
0
(A) for some Dedekind -complete Boolean algebra A (ii) U is Dedekind -complete and has a weak
order unit and whenever A U
+
is countable and disjoint then A is bounded above in U.
(e) Let U be a weakly (, )-distributive Riesz space and V a Riesz subspace of U which is either solid
or order-dense. Show that V is weakly (, )-distributive.
(f ) Show that C([0, 1]) is not weakly (, )-distributive. (Compare 316K.)
(g) Let A be a ccc weakly (, )-distributive Boolean algebra. Suppose we have a double sequence
a
ij
)
(i,j)NN
in A such that a
ij
)
jN
order*-converges to a
i
in A for each i, while a
i
)
iN
order*-converges
to a. Show that there is a strictly increasing sequence n(i))
iN
such that a
i,n(i)
)
iN
order*-converges to a.
(h) Let U be a weakly (, )-distributive Riesz space with the countable sup property. Suppose we have
an order-bounded double sequence u
ij
)
(i,j)NN
in U such that u
ij
)
jN
order*-converges to u
i
in U for
each i, while u
i
)
iN
order*-converges to u. Show that there is a strictly increasing sequence n(i))
iN
such
that u
i,n(i)
)
iN
order*-converges to u.
(i) Let A be a ccc weakly (, )-distributive Dedekind complete Boolean algebra. Show that there is a
topology on L
0
= L
0
(A) such that the closure of any A L
0
is precisely the set of order*-limits of sequences
in A.
(j) Let U be a weakly (, )-distributive Riesz space and f : U R a positive linear functional; write f

for the component of f in U

. (i) Show that for any u U

+
there is an upwards-directed A [0, u], with
supremum u, such that f

(u) = sup
vA
f(v). (See 356Xe, 362D.) (ii) Show that if f is strictly positive, so
is f

. (Compare 391D.)
368 Notes and comments 368A-368B are manifestations of a principle which will reappear in 375:
Dedekind complete L
0
spaces are in some sense maximal. If we have an order-dense subspace U of such an
L
0
, then any Archimedean Riesz space including U as an order-dense subspace can itself be embedded in L
0
(368B). In fact this property characterizes Dedekind complete L
0
spaces (368M). Moreover, any Archimedean
Riesz space U can be embedded in this way (368E); by 368C, the L
0
space (though not the embedding) is
unique up to isomorphism. If U and V are Archimedean Riesz spaces, each embedded as an order-dense
Riesz subspace of a Dedekind complete L
0
space, then any order-continuous Riesz homomorphism from
U to V extends uniquely to the L
0
spaces (368B). If one Dedekind complete L
0
space is embedded as an
order-dense Riesz subspace of another, they must in fact be the same (368D). Thus we can say that every
Archimedean Riesz space U can be extended to a Dedekind complete L
0
space, in a way which respects
order-continuous Riesz homomorphisms, and that this extension is maximal, in that U cannot be order-dense
in any larger space.
The proof of 368E which I give is long because I am using a bare-hands approach. Alternative methods
shift the burdens. For instance, if we take the trouble to develop a direct construction of the Dedekind
completion of a Riesz space (368Yb), then we need prove the theorem only for Dedekind complete Riesz
spaces. A more substantial aid is the representation theorem for Archimedean Riesz spaces with order unit
(353M); I sketch an argument in 368Ya. The drawback to this approach is the proof of Theorem 364U,
which seems to be quite as long as the direct proof of 368E which I give here. Of course we need 364U only
for compact Hausdor spaces, which are usefully easier than the general case (364V, 368Yc).
368G is a version of Ogasawaras representation theorem for Archimedean Riesz spaces. Both this and
368F can be regarded as expressions of the principle that an Archimedean Riesz space is nearly a space of
functions.
I have remarked before on the parallels between the theories of Boolean algebras and Archimedean Riesz
spaces. The notion of weak (, )-distributivity is one of the more striking correspondences. (Compare, for
386 Function spaces 368 Notes
instance, 316Xm with 368Pa.) What is really important to us, of course, is the fact that the function spaces
of measure theory are mostly weakly (, )-distributive, by 368S. Of course this is easy to prove directly
(368Xf), but I think that the argument through 368Q gives a better idea of what is really happening here.
Some of the features of order*-convergence, as dened in 367, are related to weak (, )-distributivity; in
368Yi I describe a topology which can be thought of as an abstract version of the topology of convergence
in measure on the L
0
space of a -nite measure algebra (367N).
369 Banach function spaces
In this section I continue the work of 368 with results which involve measure algebras. The rst step is
a modication of the basic representation theorem for Archimedean Riesz spaces. If U is any Archimedean
Riesz space, it can be represented as a subspace of L
0
= L
0
(A), where A is its band algebra (368E); now if
U

separates the points of U, there is a measure rendering A a localizable measure algebra (369A, 369Xa).
Moreover, we get a simultaneous representation of U

as a subspace of L
0
(369C-369D), the duality between
them corresponding exactly to the familiar duality between L
p
and L
q
.
Still drawing inspiration from the classical L
p
spaces, we have a general theory of associated Fatou
norms (369F-369M, 369R). I include notes on the spaces M
1,
, M
,1
and M
1,0
(369N-369Q), which will
be particularly useful in the next chapter.
369A Theorem Let U be a Riesz space such that U

separates the points of U. Then U can be

embedded as an order-dense Riesz subspace of L
0
(A) for some localizable measure algebra (A, ).
proof (a) Consider the canonical map S : U U

. We know that this is a Riesz homomorphism onto

an order-dense Riesz subspace of U

(356I). Because U

separates the points of U, S is injective. Let

A be the band algebra of U

and T : U

L
0
= L
0
(A) an injective Riesz homomorphism onto an
order-dense Riesz subspace V of L
0
, as in 368E. The composition TS : U L
0
is now an injective Riesz
homomorphism, so embeds U as a Riesz subspace of L
0
, which is order-dense because V is order-dense in
L
0
and TS[U] is order-dense in V (352Nc). Thus all that we need to nd is a measure on A rendering it
a localizable measure algebra.
(b) Note that V is isomorphic, as Riesz space, to U

, which is Dedekind complete (356B), so V must

be solid in L
0
(353K). Also V

must separate the points of V (356L).

Let D be the set of those d A such that the principal ideal A
d
is measurable in the sense that there
is some for which (A
d
, ) is a totally nite measure algebra. Then D is order-dense in A. PPP Take any
non-zero a A. Because V is order-dense, there is a non-zero v V such that v a. Take h 0 in V

such that h(v) > 0. Then there is a v

t
such that 0 < v
t
v and h(w) > 0 whenever 0 < w v
t
in V (356H).
Let > 0 be such that d = [[v
t
> ]] ,= 0. Then b
1

v
t
V whenever b A
d
. Set b = h(b) [0, [ for
b A
d
. Because the map b b : A L
0
is additive and order-continuous, the map b b : A
d
V also
is, and = h must be additive and order-continuous; in particular, (sup
nN
b
n
) =

n=0
b
n
whenever
b
n
)
nN
is a disjoint sequence in A
d
. Moreover, if b A
d
is non-zero, then 0 < b v
t
, so b = h(b) > 0.
Thus (A
d
, ) is a totally nite measure algebra, and d D, while 0 ,= d a. As a is arbitrary, D is
order-dense. QQQ
(c) By 313K, there is a partition of unity C D. For each c C, let
c
: A
c
[0, [ be a functional
such that (A
c
,
c
) is a totally nite measure algebra. Dene : A [0, ] by setting a =

cC

c
(a c)
for every a A. Then (A, ) is a localizable measure algebra. PPP (i) 0 =

cC
0 = 0. (ii) If a
n
)
nN
is a
disjoint sequence in A with supremum a, then
a =

cC

c
(a c) =

cC,nN

c
(a
n
c) =

n=0
a
n
.
(iii) If a A 0, then there is a c C such that a c ,= 0, so that a
c
(a c) > 0. Thus (A, ) is
a measure algebra. (iv) Moreover, in (iii), (a c) =
c
(a c) is nite. So (A, ) is semi-nite. (v) A is
Dedekind complete, being a band algebra (352Q), so (A, ) is localizable. QQQ
369C Banach function spaces 387
369B Corollary Let U be a Banach lattice with order-continuous norm. Then U can be embedded as
an order-dense solid linear subspace of L
0
(A) for some localizable measure algebra (A, ).
proof By 356Dd, U

= U

, which separates the points of U, by the Hahn-Banach theorem (3A5Ae). So

369A tells us that U can be embedded as an order-dense Riesz subspace of an appropriate L
0
(A). But also
U is Dedekind complete (354Ee), so its copy in L
0
(A) must be solid, as in 368H.
369C The representation in 369A is complemented by the following result, which is a kind of general-
ization of 365J and 366Dc.
Theorem Let (A, ) be a semi-nite measure algebra, and U L
0
= L
0
(A) an order-dense Riesz subspace.
Set
V = v : v L
0
, v u L
1
for every u U,
writing L
1
for L
1
(A, ) L
0
. Then V is a solid linear subspace of L
0
, and we have an order-continuous
injective Riesz homomorphism T : V U

dened by setting
(Tv)(u) =
_
u v for all u U, v V .
The image of V is order-dense in U

. If (A, ) is localizable, then T is surjective, so is a Riesz space

isomorphism between V and U

.
proof (a)(i) Because : L
0
L
0
L
0
is bilinear and L
1
is a linear subspace of L
0
, V is a linear subspace
of L
0
. If u U, v V , w L
0
and [w[ [v[, then
[w u[ = [w[ [u[ [v[ [u[ = [v u[ L
1
;
as L
1
is solid, w u L
1
; as u is arbitrary, w V ; this shows that V is solid.
(ii) By the denition of V , (Tv)(u) is dened in R for all u U, v V . Because is bilinear and
_
is linear, Tv : U R is linear for every v V , and T is a linear functional from V to the space of linear
operators from U to R.
(iii) If u 0 in U and v 0 in V , then u v 0 in L
1
and (Tv)(u) =
_
u v 0. This shows that
T is a positive linear operator from V to U

.
(iv) If v 0 in V and A U is a non-empty downwards-directed set with inmum 0 in U, then
inf A = 0 in L
0
, because U is order-dense (352Nb). Consequently inf
uA
u v = 0 in L
0
and in L
1
(364P),
and
inf
uA
(Tv)(u) = inf
uA
_
u v = 0
(because
_
is order-continuous). As A is arbitrary, Tv is order-continuous. As v is arbitrary, T[V ] U

.
(v) If v V and u
0
0 in U, set a = [[v > 0]]. Then v
+
= va. Set A = u : u U, 0 u u
0
a.
Because U is order-dense in L
0
, u
0
a = supA in L
0
. Because and
_
are order-continuous,
(Tv)
+
(u
0
) sup
uA
(Tv)(u) = sup
uA
_
v u
=
_
v u
0
a =
_
v
+
u
0
= (Tv
+
)(u
0
).
As u
0
is arbitrary, (Tv)
+
Tv
+
. But because T is a positive linear operator, we must have Tv
+
(Tv)
+
,
so that Tv
+
= (Tv)
+
. As v is arbitrary, T is a Riesz homomorphism.
(vi) Now T is injective. PPP If v ,= 0 in V , there is a u > 0 in U such that u [v[, because U is
order-dense. In this case u [v[ > 0 so
_
u [v[ > 0. Accordingly [Tv[ = T[v[ ,= 0 and Tv ,= 0. QQQ
(b) Putting (a-i) to (a-vi) together, we see that T is an injective Riesz homomorphism from V to U

.
All this is easy. The point of the theorem is the fact that T[V ] is order-dense in U

.
PPP Take h > 0 in U

. Let U
1
be the solid linear subspace of L
0
generated by U. Then U is an order-dense
Riesz subspace of U
1
, h : U R is an order-continuous positive linear functional, and suph(u) : u U, 0
u v is dened in R for every v 0 in U
1
; so we have an extension

h of h to U
1
such that

h U

1
(355F).
388 Function spaces 369C
Set S
1
= S(A) U
1
; then S
1
is an order-dense Riesz subspace of U
1
, because S(A) is order-dense in L
0
and U
1
is solid in L
0
. Note that S
1
is the linear span of c : c I, where I = c : c A, c U
1
, and
that I is an ideal in A.
Because h ,= 0,

h ,= 0; there must therefore be a u
0
S
1
such that

h(u
0
) > 0, and a d I such that

h(d) > 0. For a A, set a =

h(d a). Because , and

h are all order-continuous, so is , and
: A R is a non-negative completely additive functional.
By 365Ea, there is a v L
1
(A, ) such that
_
a
v = a =

h(d a)
for every a A; of course v 0. We have
_
uv

h(u) whenever u = a for a I, and therefore for every
u (S
1
)
+
. If u U
+
, then A = u
t
: u
t
S
1
, 0 u
t
u is upwards-directed, supA = u and
sup
u

A
_
v u
t
sup
u

A

h(u
t
) =

h(u) = h(u)
is nite, so v u = sup
uA
v u
t
belongs to L
1
(365Df) and
_
v u h(u). As u is arbitrary, v V and
Tv h. At the same time,
_
d
v =

h(d) > 0, so Tv > 0. As h is arbitrary, T[V ] is order-dense. QQQ
It follows that T is order-continuous (352Nb), as can also be easily proved by the argument of (a-iv)
above.
(c) Now suppose that (A, ) is localizable, that is, that A is Dedekind complete. T
1
: T[V ] V is
a Riesz space isomorphism, so certainly an order-continuous Riesz homomorphism; because V is a solid
linear subspace of L
0
, T
1
is still an injective order-continuous Riesz homomorphism when regarded as a
map from T[V ] to L
0
. Since T[V ] is order-dense in U

, T
1
has an extension to an order-continuous Riesz
homomorphism Q : U

L
0
(368B). But Q[U

] V . PPP Take h 0 in U

and u 0 in U. Then
B = g : g T[V ], 0 g h is upwards-directed and has supremum h. For g B, we know that
u T
1
g L
1
and
_
u T
1
g = g(u), by the denition of T. But this means that
sup
gB
_
u T
1
g = sup
gB
g(u) = h(u) < .
Since u T
1
g : g B is upwards-directed, it follows that
u Qh = sup
gB
u Qg = sup
gB
u T
1
g L
1
by 365Df again. As u is arbitrary, Qh V . As h is arbitrary (and Q is linear), Q[U

] V . QQQ
Also Q is injective. PPP If h U

is non-zero, there is a v V such that 0 < Tv [h[, so that

[Qh[ = Q[h[ QTv = v > 0
and Qh ,= 0. QQQ Since QT is the identity on V , Q and T must be the two halves of a Riesz space isomorphism
between V and U

.
369D Corollary Let U be any Riesz space such that U

separates the points of U. Then there is a

localizable measure algebra (A, ) such that the pair (U, U

) can be represented by a pair (V, W) of order-

dense Riesz subspaces of L
0
= L
0
(A) such that W = w : w L
0
, v w L
1
for every v V , writing L
1
for L
1
(A, ). In this case, U

becomes represented by

V = v : v L
0
, v w L
1
for every w W V .
proof Put 369A and 369C together. The construction of 369A nds (A, ) and an order-dense V which is
isomorphic to U, and 369C identies W with V

. To check that W is order-dense, take any u > 0 in L

0
.
There is a v V such that 0 < v u. There is an h (V

)
+
such that h(v) > 0, so there is a w W
+
such that w v ,= 0, that is, w v ,= 0. But now w v W, because W is solid, and 0 < w v u.
Remark Thus the canonical embedding of U in U

(356I) is represented by the embedding V

V ; U, or
V , is perfect i V =

V .
369E Kakutanis Theorem (Kakutani 41) If U is any L-space, there is a localizable measure algebra
(A, ) such that U is isomorphic, as Banach lattice, to L
1
= L
1
(A, ).
proof U is a perfect Riesz space, and U

= U

has an order unit

_
dened by saying that
_
u = |u| for
u 0 (356P). By 369D, we can nd a localizable measure algebra (A, ) and an identication of the pair
(U, U

), as dual Riesz spaces, with a pair (V, W) of subspaces of L

0
= L
0
(A); and V will be v : vw L
1
for
369H Banach function spaces 389
every w W. But W, like U

, must have an order unit; call it e. Because W is order-dense, [[e > 0]] must
be 1 and e must have a multiplicative inverse
1
e
in L
0
(364P). This means that V must be v : v e L
1
,
so that v v e is a Riesz space isomorphism between V and L
1
, which gives a Riesz space isomorphism
between U and L
1
. Moreover, if we write | |
t
for the norm on V corresponding to the norm of U, we have
|u| =
_
[u[ for u U, |v|
t
=
_
[v[ e =
_
[v e[ for v V .
Thus the Riesz space isomorphism between U and L
1
is norm-preserving, and U and L
1
are isomorphic as
Banach lattices.
369F The L
p
spaces are leading examples for a general theory of normed subspaces of L
0
, which I
proceed to sketch in the rest of the section.
Denition Let A be a Dedekind -complete Boolean algebra. An extended Fatou norm on L
0
= L
0
(A)
is a function : L
0
[0, ] such that
(i) (u +v) (u) +(v) for all u, v L
0
;
(ii) (u) = [[(u) for all u L
0
, R (counting 0 as 0, as usual);
(iii) (u) (v) whenever [u[ [v[ in L
0
;
(iv) sup
uA
(u) = (v) whenever A (L
0
)
+
is a non-empty upwards-directed set with supremum v in
L
0
;
(v) (u) > 0 for every non-zero u L
0
;
(vi) whenever u > 0 in L
0
there is a v L
0
such that 0 < v u and (v) < .
369G Proposition Let (A, ) be a Dedekind -complete Boolean algebra and an extended Fatou
norm on L
0
= L
0
(A). Then L

= u : u L
0
, (u) < is an order-dense solid linear subspace of L
0
,
and , restricted to L

, is a Fatou norm under which L

is a Banach lattice. If u
n
)
nN
is a non-decreasing
norm-bounded sequence in (L

)
+
, then it has a supremum in L

; if A is Dedekind complete, then L

has
the Levi property.
proof (a) By (i), (ii) and (iii) of 369F, L

is a solid linear subspace of L

0
; by (vi), it is order-dense.
Hypotheses (i), (ii), (iii) and (v) show that is a Riesz norm on L

, while (iv) shows that it is a Fatou

norm.
(b)(i) Suppose that u
n
)
nN
is a non-decreasing norm-bounded sequence in (L

)
+
. Then u = sup
nN
u
n
is dened in L
0
. PPP??? Otherwise, there is a v > 0 in L
0
such that kv = sup
nN
kvu
n
for every k N (368A).
By (v)-(vi) of 369F, there is a v
t
such that 0 < v
t
v and 0 < (v
t
) < . Now kv
t
= sup
nN
kv
t
u
n
for
every k, so
k(v
t
) = (kv
t
) = sup
nN
(kv
t
u
n
) sup
nN
(u
n
)
for every k, using 369F(iv), and sup
nN
(u
n
) = , contrary to hypothesis. XXXQQQ By 369F(iv) again,
(u) = sup
nN
(u
n
) < , so that u L

and u = sup
nN
u
n
in L

.
(ii) It follows that L

is complete under . PPP Let u

n
)
nN
be a sequence in L

such that (u
n+1
u
n
)
2
n
for every n N. Set v
mn
=

n
i=m
[u
i+1
u
i
[ for m n; then (v
mn
) 2
m+1
for every n, so by (i)
just above v
m
= sup
nN
v
mn
is dened in L

, and (v
m
) 2
m+1
. Now v
m
= [u
m+1
u
m
[ +v
m+1
for each
m, so u
m
v
m
)
mN
is non-decreasing and u
m
+v
m
)
mN
is non-increasing, while u
m
v
m
u
m
u
m
+v
m
for every m. Accordingly u = sup
mN
u
m
v
m
is dened in L

and [u u
m
[ v
m
for every m. But this
means that lim
m
(u u
m
) lim
m
(v
m
) = 0 and u = lim
m
u
m
in L

. As u
n
)
nN
is arbitrary,
L

is complete. QQQ
(c) Now suppose that A is Dedekind complete and A (L

)
+
is a non-empty upwards-directed norm-
bounded set in L

. By the argument of (b-i) above, using the other half of 368A, supA is dened in L
0
and
belongs to L

. As A is arbitrary, L

has the Levi property.

369H Associate norms: Denition Let (A, ) be a semi-nite measure algebra, and an extended
Fatou norm on L
0
= L
0
(A). Dene
t
: L
0
[0, ] by setting

t
(u) = sup|u v|
1
: v L
0
, (v) 1
390 Function spaces 369H
for every u L
0
; then
t
is the associate of . (The word suggests a symmetric relationship; it is justied
by the next theorem.)
369I Theorem Let (A, ) be a semi-nite measure algebra, and an extended Fatou norm on L
0
=
L
0
(A). Then
(i) its associate
t
is also an extended Fatou norm on L
0
;
(ii) is the associate of
t
;
(iii) |u v|
1
(u)
t
(v) for all u, v L
0
.
proof (a) Before embarking on the proof that
t
is an extended Fatou seminorm on L
0
, I give the greater
part of the argument needed to show that =
tt
, where

tt
(u) = sup|u w|
1
: w L
0
,
t
(w) 1
for every u L
0
.
(i) Set
B = u : u L
1
, (u) 1.
Then B is a convex set in L
1
= L
1
(A, ) and is closed for the norm topology of L
1
. PPP Suppose that u
belongs to the closure of B in L
1
. Then for each n N we can choose u
n
B such that |u u
n
|
1
2
n
.
Set v
mn
= inf
min
[u
i
[ for m n, and
v
m
= inf
nm
v
mn
= inf
nm
[u
n
[ [u[
for m N. The sequence v
m
)
mN
is non-decreasing, (v
m
) (u
m
) 1 for every m, and
|[u[ v
m
|
1
sup
nm
|[u[ v
mn
|
1

i=m
|[u[ [u
i
[|
1

i=m
|u u
i
|
1
0
as m . So [u[ = sup
mN
v
m
in L
0
,
(u) = ([u[) = sup
mN
(v
m
) 1
and u B. QQQ
(ii) Now take any u
0
L
0
such that (u
0
) > 1. Then, writing A
f
for a : a < ,
A = u : u S(A
f
), 0 u u
0

is an upwards-directed set with supremum u

0
(this is where I use the hypothesis that (A, ) is semi-nite,
so that S(A
f
) is order-dense in L
0
), and sup
uA
(u) = (u
0
) > 1. Take u
1
A such that (u
1
) > 1, that
is, u
1
/ B. By the Hahn-Banach theorem (3A5Cc), there is a continuous linear functional f : L
1
R such
that f(u
1
) > 1 but f(u) 1 for every u B. Because (L
1
)

= (L
1
)

(356Dc), [f[ is dened in (L

1
)

, and
of course
[f[(u
1
) f(u
1
) > 1, [f[(u) = supf(v) : [v[ u 1
whenever u B and u 0. Set c = [[u
1
> 0]], so that c < , and dene
a = [f[((a c))
for every a A. Then is a completely additive real-valued functional on A, so there is a w L
1
such that
a =
_
a
w for every a A (365Ea). Because a 0 for every a, w 0. Now
_
a
w = [f[(a c)
for every a A, so
_
w u = [f[(u c) [f[(u) 1
for every u S(A)
+
B. But if (v) 1, then
A
v
= u : u S(A)
+
B, u [v[
is an upwards-directed set with supremum [v[, so that
|w v|
1
= sup
uA
v
_
w u 1.
369K Banach function spaces 391
Thus
t
(w) 1. On the other hand,
|w u
0
|
1

_
w u
0

_
w u
1
= [f[(u
1
) > 1,
so
tt
(u
0
) > 1.
(iii) This shows that, for u L
0
,

tt
(u) 1 = (u) 1.
(c) Now I return to the proof that
t
is an extended Fatou norm. It is easy to check that it satises
conditions (i)-(iv) of 369F; in eect, these depend only on the fact that | |
1
is an extended Fatou norm. For
(v)-(vi), take v > 0 in L
0
. Then there is a u such that 0 u v and 0 < (u) < ; set = 1/(u). Then
(2u) > 1, so that
tt
(2u) > 1 and there is a w L
0
such that
t
(w) 1, |2u w|
1
> 1. But now set
v
1
= v [w[; then
v v
1
u [w[ > 0,
while
t
(v
1
) < . Also v u ,= 0 so

t
(v) |v
1
u|
1
> 0.
As v is arbitrary,
t
satises 369F(v)-(vi).
(d) Accordingly
tt
is also an extended Fatou norm. Now in (a) I showed that

tt
(u) 1 = (u) 1.
It follows easily that (u)
tt
(u) for every u (since otherwise there would be some such that

tt
(u) =
tt
(u) < 1 < (u) = (u).)
On the other hand, we surely have
(u) 1 = |u v|
1
1 whenever
t
(v) 1 =
tt
(u) 1,
so we must also have
tt
(u) (u) for every u. Thus
tt
= , as claimed.
(e) Of course we have |u v|
1
1 whenever (u) 1 and
t
(v) 1. It follows easily that |u v|
1

(u)
t
(v) whenever u, v L
0
and both (u),
t
(v) are non-zero. But if one of them is zero, then u v = 0,
because both and
t
satisfy (v) of 369F, so the result is trivial.
369J Theorem Let (A, ) be a semi-nite measure algebra, and an extended Fatou norm on L
0
=
L
0
(A), with associate . Then
L

= v : v L
0
, u v L
1
for every u L

.
proof (a) If v L

and u L

, then |u v|
1
is nite, by 369I(iii), so u v L
1
.
(b) If v / L

then for every n N there is a u

n
such that (u
n
) 1 and |u
n
v|
1
2
n
. Set
w
n
=

n
i=0
2
i
[u
i
[ for each n. Then w
n
)
nN
is a non-decreasing sequence and (w
n
) 2 for each n, so
w = sup
nN
w
n
is dened in L

, by 369G; now
_
w [v[ n + 1 for every n, so w v / L
1
.
369K Corollary Let (A, ) be a localizable measure algebra, and an extended Fatou norm on
L
0
= L
0
(A), with associate . Then L

may be identied, as normed Riesz space, with (L

(L

, and
L

is a perfect Riesz space.

proof Putting 369J and 369C together, we have an identication between L

and (L

. Now 369I tells

us that is the associate of , so that we can identify L

with (L

, and L

is perfect, as in 369D.
By the denition of , we have, for any v L

,
(v) = sup
(u)1
|u v|
1
= sup
(u)1,|w|

1
_
u v w = sup
(u)1
_
u v,
which is the norm of the linear functional on L

corresponding to v.
392 Function spaces 369L
369L L
p
I remarked above that the L
p
spaces are leading examples for this theory; perhaps I should
spell out the details. Let (A, ) be a semi-nite measure algebra and p [1, ]. Extend the functional | |
p
,
as dened in 364K, 365A and 366A, by saying that |u|
p
= if u L
0
L
p
. (For p = 1 this is already done
by the convention in 365A.) To see that | |
p
is now an extended Fatou norm, as dened in 369F, we note
that conditions (i)-(iii) and (v) there are true just because L
p
is a solid linear subspace on which | |
p
is a
Riesz norm, (iv) is true because | |
p
is a Fatou norm with the Levi property (363Ba, 365C, 366D), and (vi)
is true because S(A
f
) is included in L
p
and order-dense in L
0
(364L).
As usual, set q = p/(p 1) if 1 < p < , if p = 1, and 1 if p = . Then | |
q
is the associate extended
Fatou norm of | |
p
. PPP By 365Jb and 366C, |v|
q
= sup|u v|
1
: |u|
p
1 for every v L
q
. But as L
q
is order-dense in L
0
,
|v|
q
= sup
wL
q
,]w]v
|w|
q
= sup
_
[u[ [w[ : w L
q
, w [v[, |u|
p
1 = sup
_
[u[ [v[ : |u|
p
1
for every v L
0
. QQQ
369M Proposition Let (A, ) be a semi-nite measure algebra and an extended Fatou norm on
L
0
= L
0
(A, ). Then
(a) the embedding L

L
0
is continuous for the norm topology of L

and the topology of convergence

in measure on L
0
;
(b) : L
0
[0, ] is lower semi-continuous, that is, all the balls u : (u) are closed for the
topology of convergence in measure;
(c) if u
n
)
nN
is a sequence in L
0
which is order*-convergent to u L
0
(denition: 367A), then (u)
liminf
n
(u
n
).
proof (a) This is a special case of 367P.
(b) Set B

= u : (u) . If u L
0
B

, then
A = [u[ a : a A
f

is an upwards-directed set with supremum [u[, so there is an a A

f
such that (u a) > . ??? If u is in
the closure of B

for the topology of convergence in measure, then for every k N there is a v

k
B

such
that (a [[[u v
k
[ > 2
k
]]) 2
k
(see the formulae in 367M). Set
v
t
k
= [u[ inf
ik
[v
i
[
for each k, and v

= sup
kN
v
t
k
. Then (v
t
k
) (v
k
) for each k, and v
k
)
kN
is non-decreasing, so
(v

) . But
a [[[u[ v

> 2
k
]] a sup
ik
[[[u v
i
[ > 2
k
]]
has measure at most

i=k
2
i
for each k, so a [[[u[ v

> 0]] must be 0, that is, [u[ a v

and
([u[ a) ; contrary to the choice of a. XXX Thus u cannot belong to the closure of B

. As u is arbitrary,
B

is closed.
(c) If u
n
)
nN
order*-converges to u, it converges in measure (367Na). If > liminf
n
(u
n
), there is
a subsequence of u
n
)
nN
in B

, and (u) , by (b). As is arbitary, (u) liminf

n
(u
n
).
369N I now turn to another special case which we have already had occasion to consider in other
contexts.
Denition Let (A, ) be a measure algebra. Set
M
,1

= M
,1
(A, ) = L
1
(A, ) L

(A),
M
1,

= M
1,
(A, ) = L
1
(A, ) +L

(A),
and
369O Banach function spaces 393
|u|
,1
= max(|u|
1
, |u|

)
for u L
0
(A).
Remark I hope that the notation I have chosen here will not completely overload your short-term memory.
The idea is that in M
p,q
the symbol p is supposed to indicate the local nature of the space, that is, the
nature of u a where u M
p,q
and a < , while q indicates the nature of [u[ 1 for u M
p,q
. Thus
M
1,
is the space of u such that u a L
1
for every a A
f
, [u[ 1 L

; in M
1,0
we demand further
that [u[ 1 M
0
(366F); while in M
,1
we ask that [u[ 1 L
1
, u a L

for every a A
f
.
369O Proposition Let (A, ) be a semi-nite measure algebra.
(a) | |
,1
is an extended Fatou norm on L
0
= L
0
(A).
(b) Its associate | |
1,
may be dened by the formulae
|u|
1,
= min|v|
1
+|w|

: v L
1
, w L

, v +w = u
= min +
_
([u[ 1)
+
: 0
=
_

0
min(1, [[[u[ > ]])d
for every u L
0
, writing L
1
= L
1
(A, ), L

= L

(A).
(c)
u : u L
0
, |u|
1,
< = M
1,
= M
1,
(A, ),
u : u L
0
, |u|
,1
< = M
,1
= M
,1
(A, ).
(d) Writing A
f
= a : a < , S(A
f
) is norm-dense in M
,1
and S(A) is norm-dense in M
1,
.
(e) For any p [1, ],
|u|
1,
|u|
p
|u|
,1
for every u L
0
.
Remark By writing min rather than inf in the formulae of part (b) I mean to assert that the inma are
attained.
proof (a) This is easy; all we need to know is that | |
1
and | |

are extended Fatou norms.

(b) We have four functionals on L
0
to look at; let me give them names:

1
(u) = sup|u v|
1
: |v|
1,
1,

2
(u) = inf|u
t
|
1
+|u
tt
|

: u = u
t
+u
tt
,

3
(u) = inf
0
( +
_
([u[ 1)
+
),

4
(u) =
_

0
min(1, [[[u[ > ]])d.
(I write inf here to avoid the question of attainment for the moment.) Now we have the following.
(i)
1
(u)
2
(u). PPP If |v|
1,
1 and u = u
t
+u
tt
, then
|u v|
1
|u
t
v|
1
+|u
tt
v|
1
|u
t
|
1
|v|

+|u
tt
|

|v|
1
|u
t
|
1
+|u
tt
|

.
Taking the supremum over v and the innum over u
t
and u
tt
,
1
(u)
2
(u). QQQ
(ii)
2
(u)
4
(u). PPP If
4
(u) = this is trivial. Otherwise, take w such that |w|

1 and
u = [u[ w. Set
0
= inf : [[[u[ > ]] 1, and try
u
t
= w ([u[
0
1)
+
, u
tt
= w ([u[
0
1).
Then u = u
t
+u
tt
,
394 Function spaces 369O
|u
t
|
1
=
_

0
[[[u
t
[ > ]]d
=
_

0
[[[u[ > +
0
]]d
=
_

0
[[[u[ > ]]d =
_

0
min(1, [[[u[ > ]])d,
|u
tt
|

0
=
_

0
0
min(1, [[[u[ > ]])d,
so

2
(u) |u
t
|
1
+|u
tt
|

4
(u). QQQ
(iii)
4
(u)
3
(u). PPP For any 0,

4
(u) =
_

0
min(1, [[[u[ > ]])d +
_

min(1, [[[u[ > ]])d

+
_

0
[[[u[ > +]]d
= +
_

0
[[[u[ > +]]d
= +
_

0
[[([u[ 1)
+
> ]]d = +
_
([u[ 1)
+
.
Taking the inmum over ,
4
(u)
3
(u). QQQ
(iv)
3
(u)
1
(u).
PPP() It is enough to consider the case 0 <
1
(u) < , because if
1
(u) = 0 then u = 0 and evidently

3
(0) = 0, while if
1
(u) = the required inequality is trivial. Furthermore, since
3
(u) =
3
([u[) and

1
(u) =
1
([u[), it is enough to consider the case u 0.
() Note next that if a < , then |
1
max(1, a)
a|
,1
1, so that
_
a
u max(1, a)
1
(u).
() Set c = [[u > 2
1
(u)]]. If a c and a < , then
2
1
(u) a
_
a
u max(1, a)
1
(u),
so a
1
2
. As (A, ) is semi-nite, it follows that c
1
2
(322Eb).
() I may therefore write

0
= inf : 0, [[u > ]] 1.
Now [[u >
0
]] = sup
>
0
[[u > ]], so
[[u >
0
]] = sup
>
0
[[u > ]] 1.
() If
0
then
(u
0
1)
+
(
0
)[[u >
0
]] + (u 1)
+
,
so

0
+
_
(u
0
1)
+

0
+ (
0
) [[u >
0
]] +
_
(u 1)
+
+
_
(u 1)
+
.
If 0 <
0
then, for every [0,
0
[,
(u
0
1)
+
+[[u > +]] (u 1)
+
,
while [[u > +]] > 1, so
_
(u
0
1)
+
+ + +
_
(u 1)
+
;
369P Banach function spaces 395
taking the supremum over ,

0
+
_
(u
0
1)
+
+
_
(u 1)
+
.
Thus
0
+
_
(u
0
1)
+
=
3
(u).
() If
0
= 0, take v = [[u > 0]]; then |v|
,1
= [[u > 0]] 1 and

3
(u) =
_
u = |u v|
1

1
(u).
() If
0
> 0, set = [[u >
0
]]. Take any [0,
0
[. Then ([[u > ]] \ [[u >
0
]]) > 1 , so there
is a b [[u > ]] \ [[u >
0
]] such that 1 < b < . Set v = [[u >
0
]] +
1
b
b. Then |v|
,1
= 1 so

1
(u)
_
u v
_
(u
0
1)
+
+
0
+
1
b
b =
3
(u) (1 )(
0
).
As is arbitrary,
1
(u)
3
(u) in this case also. QQQ
(v) Thus
1
(u) =
2
(u) =
3
(u) =
4
(u) for every u L
0
, and I may write |u|
1,
for their common
value; as the associate of | |
,1
, | |
1,
is an extended Fatou norm. As for the attainment of the inma, the
argument of (iv-) above shows that, at least when 0 < |u|
1,
< , there is an
0
such that
0
+
_
([u[

0
)
+
= |u|
1,
. This omits the cases |u|
1,
0, ; but in either of these cases we can set
0
= 0 to see
that the inmum is attained for trivial reasons. For the other inmum, observe that the argument of (ii)
produces u
t
, u
tt
such that u = u
t
+u
tt
and |u
t
|
1
+|u
tt
|

4
(u).
(c) This is now obvious from the denition of | |
,1
and the characterization of | |
1,
in terms of | |
1
and | |

.
(d) To see that S = S(A) is norm-dense in M
1,
, we need only note that S is dense in L

and S L
1
is dense in L
1
; so that given v L
1
, w L

and > 0 there are v

t
, w
t
S such that
|(v +w) (v
t
+w
t
)|
1,
|v v
t
|
1
+|w w
t
|

.
As for M
,1
, if u 0 in M
,1
and r N, set v
r
= sup
kN
2
r
k[[u > 2
r
k]]; then each v
r
S
f
= S(A
f
),
|u v
r
|

2
r
, and v
r
)
rN
is a non-decreasing sequence with supremum u, so that lim
r
_
v
r
=
_
u
and lim
r
|uv
r
|
,1
= 0. Thus (S
f
)
+
is dense in (M
,1
)
+
. As usual, it follows that S
f
= (S
f
)
+
(S
f
)
+
is dense in M
,1
= (M
,1
)
+
(M
,1
)
+
.
(e)(i) If p = 1 or p = this is immediate from the denition of | |
,1
and the characterization of | |
1,
in (b). So suppose henceforth that 1 < p < .
(ii) If |u|
,1
1 then |u|
p
1. PPP Because |u|

1, [u[
p
[u[, so that
_
[u[
p
|u|
1
1 and
|u|
p
1. QQQ
On considering scalar multiples of u, we see at once that |u|
p
|u|
,1
for every u L
0
.
(ii) Now set q = p/(p 1). Then
|u|
p
= sup|u v|
1
: |v|
q
1
(369L)
sup|u v|
1
: |v|
,1
1 = |u|
1,
because | |
1,
is the associate of | |
,1
. This completes the proof.
369P In preparation for some ideas in 372, I go a little farther with M
1,0
, as dened in 366F.
Proposition Let (A, ) be a measure algebra.
(a) M
1,0
= M
1,0
(A, ) is a norm-closed solid linear subspace of M
1,
= M
1,
(A, ).
(b) The norm | |
1,
is order-continuous on M
1,0
.
(c) S(A
f
) and L
1
(A, ) are norm-dense and order-dense in M
1,0
.
proof (a) Of course M
1,0
, being a solid linear subspace of L
0
included in M
1,
, is a solid linear subspace of
M
1,
. To see that it is norm-closed, take any point u of its closure. Then for any > 0 there is a v M
1,0
396 Function spaces 369P
such that |u v|
1,
; now ([u v[ 1)
+
L
1
, so [[[u v[ > 2]] has nite measure; also [[[v[ > ]] has
nite measure, so
[[[u[ > 3]] [[[u v[ > 2]] [[[v[ > ]]
(364Fa) has nite measure. As is arbitrary, u M
1,0
; as u is arbitrary, M
1,0
is closed.
(b) Suppose that A M
1,0
is non-empty and downwards-directed and has inmum 0. Let > 0. Set
B = (u 1)
+
: u A. Then B L
1
(by 366Gc); B is non-empty and downwards-directed and
has inmum 0. Because | |
1
is order-continuous (365C), inf
vB
|v|
1
= 0 and there is a u A such that
|(u 1)
+
|
1
, so that |u|
1,
2. As is arbitrary, inf
uA
|u|
1,
= 0; as A is arbitrary, | |
1,
is
order-continuous on M
1,0
.
(c) By 366Gb, S(A
f
) is order-dense in M
1,0
. Because the norm of M
1,0
is order-continuous, S(A
f
) is
also norm-dense (354Ef). Now S(A
f
) L
1
M
1,0
, so L
1
must also be norm-dense and order-dense.
369Q Corollary Let (A, ) be a localizable measure algebra. Set M
1,
= M
1,
(A, ), etc.
(a) (M
1,
)

and (M
1,0
)

can both be identied with M

,1
.
(b) (M
,1
)

can be identied with M

1,
; M
1,
and M
,1
are perfect Riesz spaces.
proof Everything is covered by 369O and 369K except the identication of (M
1,0
)

with M
,1
. For this
I return to 369C. Of course M
1,0
is order-dense in L
0
, because it includes L
1
, or otherwise. Setting
V = v : v L
0
, u v L
1
for every u M
1,0
,
369C identies V with (M
1,0
)

. Of course M
,1
V just because M
1,0
M
1,
.
Also V M
,1
. PPP Let v V . (i) ??? If v / L

, then a
n
= [[[v[ > 4
n
]] ,= 0 for every n. For each n, choose
non-zero b
n
a
n
such that b
n
< (using the fact that (A, ) is semi-nite). Set u =

n=0
2
n
( b
n
)
1
b
n
;
then u L
1
M
1,0
, but
_
[v[ u 2
n
for every n N, so [v[ u / L
1
, which is impossible, because v V .
XXX Thus v L

. (ii) ??? If v / L
1
, then (again because (A, ) is semi-nite, so that [v[ = sup
aA
f [v[ a)
sup
aA
f
_
a
[v[ = . For each n N choose a
n
A
f
such that
_
a
n
[v[ 4
n
, and set u = sup
nN
2
n
a
n

M
1,0
; then
_
u[v[ 2
n
for each n, so again v / V . XXX Thus v L
1
. (iii) Putting these together, v M
,1
;
as v is arbitrary, V M
,1
. QQQ
So M
,1
= V can be identied with (M
1,0
)

.
369R The detailed formulae of 369O are of course special to the norms | |
1
, | |

, but the general

phenomenon is not.
Theorem Let (A, ) be a localizable measure algebra, and
1
,
2
two extended Fatou norms on L
0
= L
0
(A)
with associates
t
1
,
t
2
. Then we have an extended Fatou norm dened by the formula
(u) = min
1
(v) +
2
(w) : v, w L
0
, v +w = u
for every u L
0
, and its associate
t
is given by the formula

t
(u) = max(
t
1
(u),
t
2
(u))
for every u L
0
. Moreover, the corresponding function spaces are
L

= L

1
+L

2
, L

= L

1
L

2
.
proof (a) For the moment, dene by setting
(u) = inf
1
(v) +
2
(w) : v +w = u
for u L
0
. It is easy to check that, for u, u
t
L
0
and R,
(u +u
t
) (u) +(u
t
), (u) = [[(u), (u) (u
t
) if [u[ [u
t
[.
(For the last, remember that in this case u = u
t
z where |z|

1.)
(b) Take any non-empty, upwards-directed set A (L
0
)
+
, with supremum u
0
. Suppose that =
sup
uA
(u) < . For u A, n N set
C
un
= v : v L
0
, 0 v u
0
,
1
(v) +
2
(u v)
+
+ 2
n
.
369R Banach function spaces 397
Then
(i) every C
un
is non-empty (because (u) );
(ii) every C
un
is convex (because if v
1
, v
2
C
un
and [0, 1] and v = v
1
+ (1 )v
2
, then
(u v)
+
= ((u v
1
) + (1 )(u v
2
))
+
(u v
1
)
+
+ (1 )(u v
2
)
+
,
so

1
(v) +
2
(u v)
+

1
(v
1
) + (1 )
1
(v
2
) +
2
(u v
1
)
+
+ (1 )
2
(u v
2
)
+
+ 2
n
);
(iii) if u, u
t
A, m, n N and u u
t
, m n then C
u

n
C
um
;
(iv) every C
un
is closed for the topology of convergence in measure. PPP??? Suppose otherwise. Then we can
nd a v in the closure of C
un
for the topology of convergence in measure, but such that
1
(v) +
2
(uv)
+
>
+ 2
n
. In this case

1
(v) = sup
1
(v a) : a A
f
,
2
(u v)
+
= sup
2
((u v)
+
a) : a A
f
,
so there is an a A
f
such that

1
(v a) +
2
((u v)
+
a) > + 2
n
.
Now there is a sequence v
k
)
kN
in C
un
such that (a [[[v v
k
[ 2
k
]]) 2
k
for every k. Setting
v
t
k
= inf
ik
v
i
, w
k
= inf
ik
(u v
i
)
+
we have

1
(v
t
k
) +
2
(w
k
)
1
(v
k
) +
2
(u v
k
)
+
+ 2
n
for each k, and v
t
k
)
kN
, w
k
)
kN
are non-decreasing. So setting v

= sup
kN
vv
t
k
, w

= sup
kN
(uv)
+
w
k
,
we get

1
(v

) +
2
(w

) + 2
n
.
But v

v a and w

(u v)
+
a, so

1
(v a) +
2
((u v)
+
a) + 2
n
,
contrary to the choice of a. XXXQQQ
Applying 367V, we nd that

uA,nN
C
un
is non-empty. If v belongs to the intersection, then

1
(v) +
2
(u v)
+

for every u A; since (u v)
+
: u A is an upwards-directed set with supremum (u
0
v)
+
, and
2
is an
extended Fatou norm,

1
(v) +
2
(u
0
v)
+
.
(c) This shows both that the inmum in the denition of (u) is always attained (since this is trivial if
(u) = , and otherwise we consider A = [u[), and also that (supA) = sup
uA
(u) whenever A (L
0
)
+
is a non-empty upwards-directed set with a supremum. Thus satises conditions (i)-(iv) of 369F. Condition
(vi) there is trivial, since (for instance) (v)
1
(v) for every v. As for 369F(v), suppose that u > 0 in L
0
.
Take u
1
such that 0 < u
1
u and
t
1
(u
1
) 1, u
2
such that 0 < u
2
u
1
and
t
2
(u
2
) 1. In this case, if
u
2
= v +w, we must have

1
(v) +
2
(w) |v u
1
|
1
+|w u
2
|
1
|u
2
u
2
|
1
;
so that
(u) |u
2
u
2
|
1
> 0.
Thus all the conditions of 369F are satised, and is an extended Fatou norm on L
0
.
(d) The calculation of
t
is now very easy. Since surely we have
i
for both i, we must have
t

t
i
for both i. On the other hand, if u, z L
0
, then there are v, w such that u = v+w and (u) =
1
(v)+
2
(w),
so that
398 Function spaces 369R
|u z|
1
|v z|
1
+|w z|
1

1
(v)
t
1
(z) +
2
(w)
t
2
(z) (u) max(
t
1
(z),
t
2
(z));
as u is arbitrary,
t
(z) max(
t
1
(z),
t
2
(z)). So
t
= max(
t
1
,
t
2
), as claimed.
(e) Finally, it is obvious that
L

= z :
t
(z) < = z :
t
1
(z) < ,
t
2
(z) < = L

1
L

2
,
while the fact that the inmum in the denition of is always attained means that L

1
+L

2
, so that
we have equality here also.
369X Basic exercises >>>(a) Let A be a Dedekind -complete Boolean algebra. Show that the following
are equiveridical: (i) there is a function such that (A, ) is a semi-nite measure algebra; (ii) (L

separates the points of L

= L

(A); (iii) for every non-zero a A there is a completely additive functional

: A R such that a ,= 0; (iv) there is some order-dense Riesz subspace U of L
0
= L
0
(A) such that U

separates the points of U; (v) for every order-dense Riesz subspace U of L

0
there is an order-dense Riesz
subspace V of U such that V

separates the points of V .

(b) Let us say that a function : R ], ] is convex if (x + (1 )y) (x) + (1 )(y)
for all x, y I and [0, 1], interpreting 0 as 0, as usual. For any convex function : R ], ]
which is not always innite, set

(y) = sup
xR
xy (x) for every y I. (i) Show that

: R ], ]
is convex and lower semi-continuous. (Hint: 233Xh.) (ii) Show that if is lower semi-continuous then
=

. (Hint: It is easy to check that

. For the reverse inequality, set I = x : (x) < , and

consider x int I, x I int I and x / I separately; 233Ha is useful for the rst.)
>>>(c) For the purposes of this exercise and the next, say that a Youngs function is a non-negative
non-constant lower semi-continuous convex function : [0, [ [0, ] such that (0) = 0 and (x) is nite
for some x > 0. (Warning! the phrase Youngs function has other meanings.) (i) Show that in this case
is non-decreasing and continuous on the left and

, dened by saying that

(y) = sup
x0
xy (x) for
every y 0, is again a Youngs function. (ii) Show that

= . Say that and

are complementary.
(iii) Compute

in the cases () (x) = x () (x) = max(0, x 1) () (x) = x

2
() (x) = x
p
where
1 < p < .
>>>(d) Let , =

be complementary Youngs functions in the sense of 369Xc, and (A, ) a semi-nite

measure algebra. Set
B = u : u L
0
,
_

([u[) 1, C = v : v L
0
,
_

([v[) 1.
(For nite-valued ,

: (L
0
)
+
L
0
is given by 364I. Devise an appropriate convention for the case in which
takes the value .) (i) Show that B and C are order-closed solid convex sets, and that
_
[u v[ 2
for all u B, v C. (Hint: for order-closed, use 364Xg(iv).) (ii) Show that there is a unique extended
Fatou norm

on L
0
for which B is the unit ball. (iii) Show that if u L
0
B there is a v C such
that
_
[u v[ > 1. (Hint: start with the case in which u S(A)
+
.) (iv) Show that

, where

is the extended Fatou norm corresponding to and

t

is the associate of

, so that

and
t

can be
interpreted as equivalent norms on the same Banach space.
(U and V are complementary Orlicz spaces; I will call

Orlicz norms.)
(e) Let U be a Riesz space such that U

separates the points of U, and suppose that | | is a Fatou norm

on U. (i) Show that there is a localizable measure algebra (A, ) with an extended Fatou norm on L
0
(A)
such that U can be identied, as normed Riesz space, with an order-dense Riesz subspace of L

. (ii) Hence,
or otherwise, show that |u| = sup
fU

, |f|1
[f(u)[ for every u U. (iii) Show that if U is Dedekind
complete and has the Levi property, then U becomes identied with L

itself, and in particular is a Banach

lattice (cf. 354Xn).
(f ) Let (A, ) be a semi-nite measure algebra, and an extended Fatou norm on L
0
(A). Show that
the norm of L

is order-continuous i the norm topology of L

agrees with the topology of convergence in

measure on any order-bounded subset of L

.
369Xr Banach function spaces 399
(g) Let (A, ) be a -nite measure algebra of countable Maharam type, and an extended Fatou norm
on L
0
(A) such that the norm of L

is order-continuous. Show that L

is separable in its norm topology.

(h) Let (A, ) be an atomless semi-nite measure algebra. Show that |u|
1,
= max
_
a
[u[ : a A, a
1 for every u L
0
(A). (Hint: take a [[[u[ >
0
]] in part (b-iv) of the proof of 369O.)
(i) Let (A, ) be any semi-nite measure algebra. Show that if

is any Orlicz norm (369Xd), then there

is a > 0 such that |u|
1,

(u)
2
|u|
,1
for every u L
0
(A), so that M
,1

L

M
1,

.
(j) Let (A, ) be a semi-nite measure algebra. Show that the subspaces M
1,

, M
,1

of L
0
(A) can be
expressed as a complementary pair of Orlicz spaces, and that the norm | |
,1
can be represented as an
Orlicz norm, but | |
1,
cannot.
>>>(k) Let (A, ) be a measure algebra and U a Banach space. (i) Suppose that : A U is an additive
function such that |a| min(1, a) for every a A. Show that there is a unique bounded linear operator
T : M
1,

U such that T(a) = a for every a A. (ii) Suppose that : A
f
U is an additive function
such that |a| max(1, a) for every a A
f
. Show that there is a unique bounded linear operator
T : M
,1

U such that T(a) = a for every a A
f
.
(l) Let (A, ) and (B, ) be semi-nite measure algebras, and : A
f
B
f
a measure-preserving ring
homomorphism, as in 366H, with associated maps T : M
0

M
0

, P : M
1,0

M
1,0

. Show that |Tu|
,1
=
|u|
,1
for every u M
,1

, |Pv|
,1
|v|
,1
for every v M
,1

.
(m) Let (A, ) and (B, ) be measure algebras, and : A B a measure-preserving Boolean homomor-
phism. (i) Show that there is a unique Riesz homomorphism T : M
1,

M
1,

such that T(a) = (a)
for every a A and |Tu|
1,
= |u|
1,
for every u M
1,

. (ii) Now suppose that (A, ) is localiz-
able and is order-continuous. Show that there is a unique positive linear operator P : M
1,

M
1,

such that
_
a
Pv =
_
a
v for every a A
f
, v M
1,

, and that |Pv|

|v|

for every v L

(B),
|Pv|
1,
|v|
1,
for every v M
1,

. (Compare 365P.)
(n) Let (A, ) and (B, ) be semi-nite measure algebras, and : [0, [ [0, ] a Youngs function,
as in 369Xd; write

for the corresponding Orlicz norm on either L

0
(A) or L
0
(B). Let : A B be a
measure-preserving Boolean homomorphism, with associated map T : M
1,

M
1,

, as in 369Xm. (i)
Show that

(Tu) =

(u) for every u M

1,

. (ii) Show that if (A, ) is localizable, is order-continuous
and P : M
1,

M
1,

is the map of 369Xm(ii), then

(Pv)

(v) for every v M

1,

. (Hint: 365R.)
>>>(o) Let (A, ) be any semi-nite measure algebra and
1
,
2
two extended Fatou norms on L
0
(A). Show
that u max(
1
(u),
2
(u)) is an extended Fatou norm.
(p) Let (A, ) be a semi-nite measure algebra, and (

A, ) its localization (322P). Show that the Dedekind

completion of M
1,
(A, ) can be identied with M
1,
(

A, ).
(q) Let (A, ) be a localizable measure algebra. (i) Show that if B is any closed subalgebra of A such that
supb : b B, b < = 1 in A, we have an order-continuous positive linear operator P
B
: M
1,

M
1,
B
such that
_
b
P
B
u =
_
b
u whenever u M
1,

, b B and b < . (ii) Show that if B
n
)
nN
is a non-
decreasing sequence of closed subalgebras of A such that supb : b B
0
, b < = 1 in A, and B is the
closure of

nN
B
n
, then P
B
n
u)
nN
is order*-convergent to P
B
u for every u M
1,

. (Cf. 367K.)
(r) Let
1
and
2
be Youngs functions (369Xc) and (A, ) a semi-nite measure algebra. Set (x) =
max(
1
(x),
2
(x)) for x [0, [. (i) Show that is a Youngs function. (ii) Writing

2
,

for the
corresponding extended Fatou norms on L
0
(A) (369Xd), show that

max(

1
,

2
)
1
2

, so that
L

= L

1
L

2
and L

= L

1
+ L

2
, writing

for the Youngs function complementary to . (iii)

Repeat with =
1
+
2
in place of .
400 Function spaces 369Y
369Y Further exercises (a) Let (A, ) be a localizable measure algebra and A L
0
= L
0
(A) a
countable set. Show that the solid linear subspace U of L
0
generated by A is a perfect Riesz space. (Hint:
reduce to the case in which U is order-dense. If A = u
n
: n N, w (L
0
)
+
U nd v
n
(L
0
)
+
such that
_
v
n
u 2
n
4
n
_
v
n
[u
i
[ for every i n. Show that v = sup
nN
v
n
is dened in L
0
and corresponds
to a member of U

.)
(b) Let U be a Banach lattice and suppose that p [1, [ is such that |u +v|
p
= |u|
p
+|v|
p
whenever
u, v U and [u[ [v[ = 0. Show that U is isomorphic, as Banach lattice, to L
p

for some localizable measure
algebra (A, ). (Hint: start by using 354Yd to show that the norm of U is order-continuous, as in 354Yj.)
(c) Let : [0, [ [0, [ be a strictly increasing Youngs function (369Xc) such that (0) = 0 and
sup
t>0
(2t)/(t) is nite. Show that the associated Orlicz norms

(369Xd) are always order-continuous

on their function spaces.
(d) Let : [0, [ [0, ] be a Youngs function, and suppose that the corresponding Orlicz norm on
L
0
(A
L
), where (A
L
,
L
) is the measure algebra of Lebesgue measure on R, is order-continuous on its function
space L

. Show that there is an M 0 such that (2t) M(t) for every t 0.

(e) Let (A, ) be a semi-nite measure algebra and : [0, [ [0, [ a Youngs function such that the
Orlicz norm

is order-continuous on L

. Show that if T is a lter on L

, then T u L

for the
norm

i (i) T u for the topology of convergence in measure (ii) limsup

vJ

(v)

(u). (Compare
245Xk.)
(f ) Give an example of an extended Fatou norm on L
0
(A
L
), where (A
L
,
L
) is the measure algebra of
Lebesgue measure on [0, 1], such that (i) gives rise to an order-continuous norm on its function space L

(ii) there is a sequence u

n
)
nN
in L

, converging in measure to u L

, such that lim

n
(u
n
) = (u)
but u
n
)
nN
does not converge to u for the norm on L

.
(g) Let (A, ) be a semi-nite measure algebra, and an Orlicz norm on L
0
(A). Show that L

has the
Levi property, whether or not A is Dedekind complete.
(h) Let (A, ) be any measure algebra. Show that (M
1,0

)

can be identied with M

,1

. (Hint: show
that neither M
1,0
nor M
,1
is changed by moving rst to the semi-nite version of (A, ), as described in
322Xa, and then to its localization.)
(i) Give an example to show that the result of 369R may fail if (A, ) is only semi-nite, not localizable.
369 Notes and comments The representation theorems 369A-369D give a very concrete form to the
notion of perfect Riesz space: it is just one which can be expressed as a subspace of L
0
(A), for some
localizable measure algebra (A, ), in such a way that it is its own second dual, where the duality here is
between subspaces of L
0
, taking U
t
= v : u v L
1
for every u U. (I see that in this expression I
ought somewhere to mention that both U and U
t
are assumed to be order-dense in L
0
.) Indeed I believe
that the original perfect spaces were the vollkommene Raume of G.Kothe, which were subspaces of R
N
,
corresponding to the measure algebra TN with counting measure, so that U
t
or U

was v : u v
1
for
every u U.
I have presented Kakutanis theorem on the representation of L-spaces as a corollary of 369A and 369C.
As usual in such things, this is a reversal of the historical relationship; Kakutanis theorem was one of the
results which led to the general theory. If we take the trouble to re-work the argument of 369A in this
context, we nd that the L-space condition |u + v| = |u| + |v| whenever u, v 0 can be relaxed to
|u + v| = |u| + |v| whenever u v = 0 (369Yb). The complete list of localizable measure algebras
provided by Maharams theorem (332B, 332J) now gives us a complete list of L-spaces.
Just as perfect Riesz spaces come in dual pairs, so do some of the most important Banach lattices: those
with Fatou norms and the Levi property for which the order-continuous dual separates the points. (Note that
the dual of any space with a Riesz norm has these properties; see 356Da.) I leave the details of representing
such spaces to you (369Xe). The machinery of 369F-369K gives a solid basis for studying such pairs.
369 Notes Banach function spaces 401
Among the extended Fatou norms of 369F the Orlicz norms (369Xd, 369Yc-369Ye) form a signicant
subfamily. Because they are dened in a way which is to some extent independent of the measure algebra
involved, these spaces have some of the same properties as L
p
spaces in relation to measure-preserving
homomorphisms (369Xm-369Xn). In 373-374 I will elaborate on these ideas. Among the Orlicz spaces,
we have a largest and a smallest; these are just M
1,
= L
1
+L

and M
,1
= L
1
L

(369N-369O, 369Xi,
369Xj). Of course these two are particularly important.
There is an interesting phenomenon here. It is easy to see that | |
,1
= max(| |
1
, | |

) is an extended
Fatou norm and that the corresponding Banach lattice is L
1
L

; and that the same ideas work for any pair

of extended Fatou norms (369Xo). To check that the dual of L
1
L

is precisely the linear sum L

+ L
1
a little more is needed, and the generalization of this fact to other extended Fatou norms (369Q) seems to
go quite deep. In view of our ordinary expectation that properties of these normed function spaces should
be reected in perfect Riesz spaces in general, I mention that I believe I have found an example, dependent
on the continuum hypothesis, of two perfect Riesz subspaces U, V of R
N
such that their linear sum U +V
is not perfect.
402 Linear operators between function spaces
Chapter 37
Linear operators between function spaces
As everywhere in functional analysis, the function spaces of measure theory cannot be properly understood
without investigating linear operators between them. In this chapter I have collected a number of results
which rely on, or illuminate, the measure-theoretic aspects of the theory. 371 is devoted to a fundamental
property of linear operators on L-spaces, if considered abstractly, that is, of L
1
-spaces, if considered in the
language of Chapter 36, and to an introduction to the class T of operators which are norm-decreasing for both
| |
1
and | |

. This makes it possible to prove a version of (Birkhos) Ergodic Theorem for operators which
need not be positive (372D). In 372 I give various forms of this theorem, for linear operators between function
spaces, for measure-preserving Boolean homomorphisms between measure algebras, and for inverse-measure-
preserving functions between measure spaces, with an excursion into the theory of continued fractions. In
373 I make a fuller analysis of the class T , with a complete characterization of those u, v such that v = Tu
for some T T . Using this we can describe rearrangement-invariant function spaces and extended Fatou
norms (374). Returning to ideas left on one side in 364 and 368, I investigate positive linear operators
dened on L
0
spaces (375). In the nal section of the chapter (376), I look at operators which can be
dened in terms of kernels on product spaces.
371 The Chacon-Krengel theorem
The rst topic I wish to treat is a remarkable property of L-spaces: if U and V are L-spaces, then every
continuous linear operator T : U V is order-bounded, and |[T[| = |T| (371D). This generalizes in various
ways to other V (371B, 371C). I apply the result to a special type of operator between M
1,0
spaces which
will be conspicuous in the next section (371F-371H).
371A Lemma Let U be an L-space, V a Banach lattice and T : U V a bounded linear operator.
Take u 0 in U and set
B =

n
i=0
[Tu
i
[ : u
0
, . . . , u
n
U
+
,

n
i=0
u
i
= u V
+
.
Then B is upwards-directed and sup
vB
|v| |T||u|.
proof (a) Suppose that v, v
t
B. Then we have u
0
, . . . , u
m
, u
t
0
, . . . , u
t
n
U
+
such that

m
i=0
u
i
=

n
j=0
u
t
j
= u, v =

m
i=0
[Tu
i
[ and v
t
=

n
j=0
[Tu
t
j
[. Now there are v
ij
0 in U, for i m and j n, such
that u
i
=

n
j=0
v
ij
for i m and u
t
j
=

m
i=0
v
ij
for j n (352Fd). We have u =

m
i=0

n
j=0
v
ij
, so that
v
tt
=

m
i=0

n
j=0
[Tv
ij
[ B. But
v =

m
i=0
[Tu
i
[ =

m
i=0
[T(

n
j=0
v
ij
)[

m
i=0

m
j=0
[Tv
ij
[ = v
tt
,
and similarly v
t
v
tt
. As v and v
t
are arbitrary, B is upwards-directed.
(b) The other part is easy. If v B is expressed as

m
i=0
[Tu
i
[ where u
i
0,

m
i=0
u
i
= u then
|v|

m
i=0
|Tu
i
| |T|

m
i=0
|u
i
| = |T||u|
because U is an L-space.
371B Theorem Let U be an L-space and V a Dedekind complete Banach lattice U with a Fatou norm.
Then the Riesz space L

(U; V ) = L

(U; V ) is a closed linear subspace of the Banach space B(U; V ) and is

in itself a Banach lattice with a Fatou norm.
proof (a) I start by noting that L

(U; V ) = L

(U; V ) B(U; V ) just because V has a Riesz norm and U

is a Banach lattice with an order-continuous norm (355C, 355Kb).
(b) The rst new step is to check that |[T[| |T| for any T L

(U; V ). PPP Start with any u U

+
.
Set
B =

n
i=0
[Tu
i
[ : u
0
, . . . , u
n
U
+
,

n
i=0
u
i
= u V
+
,
371C The Chacon-Krengel theorem 403
as in 371A. If u
0
, . . . , u
n
0 are such that

n
i=0
u
i
= u, then [Tu
i
[ [T[u
i
for each i, so that

n
i=0
[Tu
i
[

n
i=0
[T[u
i
= [T[u; thus B is bounded above by [T[u and supB [T[u. On the other hand, if [v[ u in U,
then v
+
+v

+ (u [v[) = u, so [Tv
+
[ +[Tv

[ +[T(u [v[)[ B and

[Tv[ = [Tv
+
+Tv

[ [Tv
+
[ +[Tv

[ supB.
As v is arbitrary, [T[u supB and [T[u = supB. Consequently
|[T[u| | supB| = sup
wB
|w| |T||u|
because V has a Fatou norm and B is upwards-directed.
For general u U,
|[T[u| |[T[[u[| |T||[u[| = |T||u|.
This shows that |[T[| |T|. QQQ
(c) Now if [S[ [T[ in L

(U; V ), and u U, we must have

|Su| |[S[[u[| |[T[[u[| |[T[||[u[| |T||u|;
as u is arbitrary, |S| |T|. This shows that the norm of L

(U; V ), inherited from B(U; V ), is a Riesz

norm.
(d) Suppose next that T B(U; V ) belongs to the norm-closure of L

(U; V ). For each n N choose

T
n
L

(U; V ) such that |T T

n
| 2
n
. Set S
n
= [T
n+1
T
n
[ L

(U; V ) for each n. Then

|S
n
| = |T
n+1
T
n
| 3 2
n1
for each n, so S =

n=0
S
n
is dened in the Banach space B(U; V ). But if u U
+
, we surely have
Su =

n=0
S
n
u 0
in V . Moreover, if u U
+
and [v[ u, then for any n N
[T
n+1
v T
0
v[ = [

n
i=0
(T
i+1
T
i
)v[

n
i=0
S
i
u Su,
and T
0
v Su T
n+1
v T
0
v +Su; letting n , we see that
[T
0
[u Su T
0
v Su Tv T
0
v +Su [T
0
[u +Su.
So [Tv[ [T
0
[u +Su whenever [v[ u. As u is arbitrary, T L

(U; V ).
This shows that L

(U; V ) is closed in B(U; V ) and is therefore a Banach space in its own right; putting
this together with (b), we see that it is a Banach lattice.
(e) Finally, the norm of L

(U; V ) is a Fatou norm. PPP Let A L

(U; V )
+
be a non-empty, upwards-
directed set with supremum T
0
L

(U; V ). For any u U,

|T
0
u| = |[T
0
u[| |T
0
[u[| = | sup
TA
T[u[|
by 355Ed. But T[u[ : T A is upwards-directed and the norm of V is a Fatou norm, so
|T
0
u| sup
TA
|T[u[| sup
TA
|T||u|.
As u is arbitrary, |T
0
| sup
TA
|T|. As A is arbitrary, the norm of L

(U; V ) is Fatou. QQQ

371C Theorem Let U be an L-space and V a Dedekind complete Banach lattice with a Fatou norm
and the Levi property. Then B(U; V ) = L

(U; V ) = L

(U; V ) is a Dedekind complete Banach lattice with

a Fatou norm and the Levi property. In particular, [T[ is dened and |[T[| = |T| for every T B(U; V ).
proof (a) Let T : U V be any bounded linear operator. Then T L

(U; V ). PPP Take any u 0 in U.

Set
B =

n
i=0
[Tu
i
[ : u
0
, . . . , u
n
U
+
,

n
i=0
u
i
= u V
+
as in 371A. Then 371A tells us that B is upwards-directed and norm-bounded. Because V has the Levi
property, B is bounded above. But just as in part (b) of the proof of 371B, any upper bound of B is also
an upper bound of Tv : [v[ u. As u is arbitrary, T L

(U; V ). QQQ
404 Linear operators between function spaces 371C
(b) Accordingly L

(U; V ) = B(U; V ). By 371B, this is a Banach lattice with a Fatou norm, and equal
to L

(U; V ). To see that it also has the Levi property, let A L

(U; V ) be any non-empty norm-bounded

upwards-directed set. For u U
+
, Tu : T A is non-empty, norm-bounded and upwards-directed in V ,
so is bounded above in V . By 355Ed, A is bounded above in L

(U; V ).
371D Corollary Let U and V be L-spaces. Then L

(U; V ) = L

(U; V ) = B(U; V ) is a Dedekind

complete Banach lattice with a Fatou norm and the Levi property.
371E Remarks Note that both these theorems show that L

(U; V ) is a Banach lattice with properties

similar to those of V whenever U is an L-space. They can therefore be applied repeatedly, to give facts
about L

(U
1
; L

(U
2
; V )) where U
1
, U
2
are L-spaces and V is a Banach lattice, for instance. I hope that this
formula will recall some of those in the theory of bilinear maps and tensor products (see 253Xa-253Xb).
371F The class T
(0)
For the sake of applications in the next section, I introduce now a class of operators
of great intrinsic interest.
Denition Let (A, ), (B, ) be measure algebras. Recall that M
1,0
(A, ) is the space of those u
L
1
(A, ) + L

(A) such that [[[u[ > ]] < for every > 0 (366F-366G, 369P). Write T
(0)
= T
(0)
,
for
the set of all linear operators T : M
1,0
(A, ) M
1,0
(B, ) such that Tu L
1
(B, ) and |Tu|
1
|u|
1
for
every u L
1
(A, ), Tu L

(B) and |Tu|

|u|

for every u L

(A) M
1,0
(A, ).
371G Proposition Let (A, ) and (B, ) be measure algebras.
(a) T
(0)
= T
(0)
,
is a convex set in the unit ball of B(M
1,0
(A, ); M
1,0
(B, )). If T
0
: L
1
(A, ) L
1
(B, ) is
a linear operator of norm at most 1, and T
0
u L

(B) and |T
0
u|

|u|

for every u L
1
(A, )L

(A),
then T
0
has a unique extension to a member of T
(0)
.
(b) If T T
(0)
then T is order-bounded and [T[, taken in
L

(M
1,0
(A, ); M
1,0
(B, )) = L

(M
1,0
(A, ); M
1,0
(B, )),
also belongs to T
(0)
.
(c) If T T
(0)
then |Tu|
1,
|u|
1,
for every u M
1,0
(A, ).
(d) If T T
(0)
, p [1, [ and w L
p
(A, ) then Tw L
p
(B, ) and |Tw|
p
|w|
p
.
(e) If (C,

) is another measure algebra then ST T

(0)
,

for every T T
(0)
,
and every S T
(0)
,

.
proof I write M
1,0

, L
p

for M
1,0

, L
p
(B, ), etc.
(a)(i) If T T
(0)
and u M
1,0

then there are v L
1

, w L

such that u = v +w and |v|
1
+|w|

=
|u|
1,
(369Ob); so that
|Tu|
1,
|Tv|
1
+|Tw|

|v|
1
+|w|

|u|
1,
.
As u is arbitrary, T is in the unit ball of B(M
1,0

; M
1,0

).
(ii) Because the unit balls of B(L
1

; L
1

) and B(L

; L

) are convex, so is T
(0)
.
(iii) Now suppose that T
0
: L
1

L
1

is a linear operator of norm at most 1 such that |T
0
u|

|u|

for every u L
1

L

. By the argument of (i), T
0
is a bounded operator for the | |
1,
norms; since L
1

is
dense in M
1,0

(369Pc), T
0
has a unique extension to a bounded linear operator T : M
1,0

M
1,0

. Of course
|Tu|
1
= |T
0
u|
1
|u|
1
for every u L
1

.
Now suppose that u L

M
1,0

; set = |u|

. Let > 0, and set

v = (u
+
1)
+
(u

1)
+
;
then [v[ [u[ and |u v|

and v L
1

L

. Accordingly
|Tu Tv|
1,
|u v|
1,
, |Tv|

= |T
0
v|

|v|

.
So if we set w = ([Tu Tv[ 1)
+
L
1

, |w|
1
; while
371G The Chacon-Krengel theorem 405
[Tu[ [Tv[ +w +1 ( +)1 +w,
so
|([Tu[ ( +)1)
+
|
1
|w|
1
.
As is arbitrary, [Tu[ 1, that is, |Tu|

|u|

. As u is arbitrary, T T
(0)
.
(b) Because M
1,0

has an order-continuous norm (369Pb), L

(M
1,0

; M
1,0

) = L

(M
1,0

; M
1,0

) (355Kb).
Take any T T
(0)
and consider T
0
= TL
1

: L
1

L
1

. This is an operator of norm at most 1. By 371D,
T
0
is order-bounded, and |[T
0
[| 1, where [T
0
[ is taken in L

(L
1

; L
1

) = B(L
1

; L
1

). Now if u L
1

L

,
[[T
0
[u[ [T
0
[[u[ = sup
]u

]]u]
[T
0
u
t
[ |u|

1,
so |[T
0
[u|

|u|

. By (a), there is a unique S T

(0)
extending [T
0
[. Now Su
+
0 for every u L
1

, so
Su
+
0 for every u M
1,0

(since the function u (Su
+
)
+
Su
+
: M
1,0

M
1,0

is continuous and zero
on the dense set L
1

), that is, S is a positive operator; also S[u[ [Tu[ for every u L
1

, so Sv S[u[ [Tu[
whenever u, v M
1,0

and [u[ v. This means that T : M
1,0

M
1,0

is order-bounded. Because M
1,0

is
Dedekind complete (366Ga), [T[ is dened in L

(M
1,0

; M
1,0

).
If v 0 in L
1

, then
[T[v = sup
]u]v
Tu = sup
]u]v
T
0
u = [T
0
[v = Sv.
Thus [T[ agrees with S on L
1

. Because M
1,0

is a Banach lattice (or otherwise), [T[ is a bounded operator,
therefore continuous (2A4Fc), so [T[ = S T
(0)
, which is what we needed to know.
(c) We can express u as v +w where |v|
1
+|w|

= |u|
1,
; now w = u v M
1,0

, so we can speak of
Tw, and
|Tu|
1,
= |Tv +Tw|
1,
|Tv|
1
+|Tw|

|v|
1
+|w|

= |u|
1,
,
as required.
(d) (This is a modication of 244M.)
(i) Suppose that T, p, w are as described, and that in addition T is positive. The function t [t[
p
is
convex (233Xc), so we can nd families
q
)
qQ
,
q
)
qQ
of real numbers such that [t[
p
= sup
qQ

q
+
q
(tq)
for every t R (233Hb). Then [u[
p
= sup
qQ

q
1 +
q
(u q1) for every u L
0
. (The easiest way to
check this is perhaps to think of L
0
as a quotient of a space of functions, as in 364D; it is also a consequence
of 364Xg(iii).) We know that [w[
p
L
1

, so we may speak of T([w[
p
); while w M
1,0

(366Ga), so we may
speak of Tw.
For any q Q, 0
p

q
q
q
, that is, q
q

q
0, while
q
w[w[
p
(q
q

q
)1 and |(
q
w[w[
p
)
+
|

q
q

q
. Now this means that
T(
q
w [w[
p
) T(
q
w [w[
p
)
+
|T(
q
w [w[
p
)
+
|

1
|(
q
w [w[
p
)
+
|

1 (q
q

q
)1.
Turning this round again,

q
1 +
q
(Tw q1) T([w[
p
).
Taking the supremum over q, [Tw[
p
T([w[
p
), so that
_
[Tw[
p

_
[w[
p
(because |Tv|
1
|v|
1
for every
v L
1
). Thus Tw L
p
and |Tw|
p
|w|
p
.
(ii) For a general T T
(0)
, we have [T[ T
(0)
, by (b), and [Tw[ [T[[w[, so that |Tw|
p
|[T[[w[|
p

|w|
p
, as required.
(e) This is elementary, because
|STu|
1
|Tu|
1
|u|
1
, |STv|

|Tu|

|u|

whenever u L
1

, v L

M
1,0

.
406 Linear operators between function spaces 371H
371H Remark In the context of 366H, T

M
1,0

T
(0)
,
, while P

T
(0)
,
. Thus 366H(a-iv), 366H(b-iii)
are special cases of 371Gd.
371X Basic exercises >>>(a) Let U be an L-space, V a Banach lattice with an order-continuous norm
and T : U V a bounded linear operator. Let B be the unit ball of U. Show that [T[[B] T[B].
(b) Let U and V be Banach spaces. (i) Show that the space K
w
(U; V ) of weakly compact linear operators
from U to V (denition: 3A5Kb) is a closed linear subspace of B(U; V ). (ii) Show that if U is an L-space
and V is a Banach lattice with an order-continuous norm, then K
w
(U; V ) is a norm-closed Riesz subspace
of L

(U; V ).
(c) Let (A, ) be a semi-nite measure algebra and set U = L
1
(A, ). Show that L

(U; U) = B(U; U) is
a Banach lattice with a Fatou norm and the Levi property. Show that its norm is order-continuous i A is
nite. (Hint: consider operators u u a, where a A.)
>>>(d) Let U be a Banach lattice, and V a Dedekind complete M-space. Show that L

(U; V ) = B(U; V )
is a Banach lattice with a Fatou norm and the Levi property.
(e) Let U and V be Riesz spaces, of which V is Dedekind complete, and let T L

(U; V ). Dene
T
t
L

(V

; U

) by writing T
t
(h) = hT for h V

. (i) Show that [T[

t
[T
t
[ in L

(V

; U

). (ii) Show
that [T[
t
h = [T
t
[h for every h V

. (Hint: show that if u U

+
, h (V

)
+
then ([T
t
[h)(u) and h([T[u)
are both equal to sup

n
i=0
g
i
(Tu
i
) : [g
i
[ h, u
i
0,

n
i=0
u
i
= u.)
>>>(f ) Using 371D, but nothing about uniformly integrable sets beyond the denition (354P), show that if
U and V are L-spaces, A U is uniformly integrable in U, and T : U V is a bounded linear operator,
then T[A] is uniformly integrable in V .
371Y Further exercises (a) Let U and V be Banach spaces. (i) Show that the space K(U; V ) of
compact linear operators from U to V (denition: 3A5Ka) is a closed linear subspace of B(U; V ). (ii)
Show that if U is an L-space and V is a Banach lattice with an order-continuous norm, then K(U; V ) is a
norm-closed Riesz subspace of L

(U; V ). (See Krengel 63.)

(b) Let (A, ) be a measure algebra, U a Banach space, and T : L
1
(A, ) U a bounded linear operator.
Show that T is a compact linear operator i
1
a
T(a) : a A, 0 < a < is relatively compact in U.
(c) Let (A, ) be a probability algebra, and set L
1
= L
1
(A, ). Let a
n
)
nN
be a stochastically independent
sequence of elements of A of measure
1
2
, and dene T : L
1
R
N
by setting Tu(n) =
_
u 2
_
a
n
u for each
n. Show that T B(L
1
; ccc
0
) L

(L
1
; ccc
0
), where ccc
0
is the Banach lattice of sequences converging to 0. (See
272Yd.)
(d) Regarding T of 371Yc as a map from L
1
to

, show that [T
t
[ ,= [T[
t
in L

((

, L

(A)).
(e) (i) In
2
dene e
i
by setting e
i
(i) = 1, e
i
(j) = 0 if j ,= i. Show that if T L

(
2
;
2
) then
([T[e
i
[e
j
) = [(Te
i
[e
j
)[ for all i, j N. (ii) Show that for each n N there is an orthogonal (2
n
2
n
)-matrix
A
n
such that every coecient of A
n
has modulus 2
n/2
. (Hint: A
n+1
=
1

2
_
A
n
A
n
A
n
A
n
_
.) (iii) Show
that there is a linear isometry S :
2

2
such that [(Se
i
[e
j
)[ = 2
n/2
if 2
n
i, j < 2
n+1
. (iv) Show that
S / L

(
2
;
2
).
371 Notes and comments The Chacon-Krengel theorem, properly speaking (Chacon & Krengel 64),
is 371D in the case in which U = L
1
(), V = L
1
(); of course no new ideas are required in the generalizations
here, which I have copied from Fremlin 74a.
Anyone with a training in functional analysis will automatically seek to investigate properties of operators
T : U V in terms of properties of their adjoints T
t
: V

, as in 371Xe and 371Yd. When U is

372B The ergodic theorem 407
an L-space, then U

is a Dedekind complete M-space, and it is easy to see that this forces T

t
to be order-
bounded, for any Banach lattice V (371Xd). But since no important L-space is reexive, this approach
cannot reach 371B-371D without a new idea of some kind. It can however be adapted to the special case in
371Gb (Dunford & Schwartz 57, VIII.6.4).
In fact the results of 371B-371C are characteristic of L-spaces (Fremlin 74b). To see that they fail in
the simplest cases in which U is not an L-space and V is not an M-space, see 371Yc-371Ye.
372 The ergodic theorem
I come now to one of the most remarkable topics in measure theory. I cannot do it justice in the space I
have allowed for it here, but I can give the basic theorem (372D-372E) and a variety of the corollaries through
which it is regularly used (372F-372K), together with brief notes on one of its most famous and characteristic
applications (to continued fractions, 372M-372O) and on ergodic and mixing transformations (372P-
372R). In the rst half of the section (down to 372G) I express the arguments in the abstract language of
measure algebras and their associated function spaces, as developed in Chapter 36; the second half, from
372H onwards, contains translations of the results into the language of measure spaces and measurable
functions, the more traditional, and more readily applicable, forms.
372A Lemma Let U be a reexive Banach space and T : U U a bounded linear operator of norm at
most 1. Then
V = u +v Tu : u, v U, Tv = v
is dense in U.
proof Of course V is a linear subspace of U. ??? Suppose, if possible, that it is not dense. Then there is a
non-zero h U

such that h(v) = 0 for every v V (3A5Ad). Take u U such that h(u) ,= 0. Set
u
n
=
1
n+1

n
i=0
T
i
u
for each n N, taking T
0
to be the identity operator; because
|T
i
u| |T
i
||u| |T|
i
|u| |u|
for each i, |u
n
| |u| for every n. Note also that T
i+1
uT
i
u V for every i, so that h(T
i+1
uT
i
u) = 0;
accordingly h(T
i
u) = h(u) for every i, and h(u
n
) = u for every n.
Let T be any non-principal ultralter on N. Because U is reexive, v = lim
nJ
u
n
is dened in U for
the weak topology on U (3A5Gc). Now Tv = v. PPP For each n N,
Tu
n
u
n
=
1
n+1

n
i=0
(T
i+1
u T
i
u) =
1
n+1
(T
n+1
u u)
has norm at most
2
n+1
|u|. So Tu
n
u
n
)
nN
0 for the norm topology U and therefore for the weak
topology, and surely lim
nJ
Tu
n
u
n
= 0. On the other hand (because T is continuous for the weak
topology, 2A5If)
Tv = lim
nJ
Tu
n
= lim
nJ
(Tu
n
u
n
) + lim
nJ
u
n
= 0 +v = v,
where all the limits are taken for the weak topology. QQQ
But this means that v V , while
h(v) = lim
nJ
h(u
n
) = h(u) ,= 0,
contradicting the assumption that h V

. XXX
372B Lemma Let (A, ) be a measure algebra, and T : L
1
L
1
a positive linear operator of norm at
most 1, where L
1
= L
1
(A, ). Take any u L
1
and m N, and set
a = [[u > 0]] [[u +Tu > 0]] [[u +Tu +T
2
u > 0]] . . . [[u +Tu +. . . +T
m
u > 0]].
408 Linear operators between function spaces 372B
Then
_
a
u 0.
proof Set u
0
= u, u
1
= u + Tu, . . . , u
m
= u + Tu + . . . + T
m
u, v = sup
im
u
i
, so that a = [[v > 0]].
Consider u + T(v
+
). We have T(v
+
) Tv Tu
i
for every i m (because T is positive), so that
u + T(v
+
) u + Tu
i
= u
i+1
for i < m, and u + T(v
+
) sup
1im
u
i
. Also u + T(v
+
) u because
T(v
+
) 0, so u +T(v
+
) v. Accordingly
_
a
u
_
a
v
_
a
T(v
+
) =
_
v
+

_
a
T(v
+
) |v
+
|
1
|Tv
+
|
1
0
because |T| 1.
372C Maximal Ergodic Theorem Let (A, ) be a measure algebra, and T : L
1
L
1
a linear
operator, where L
1
= L
1
(A, ), such that |Tu|
1
|u|
1
for every u L
1
and |Tu|

|u|

for every
u L
1
L

(A). Set A
n
=
1
n+1

n
i=0
T
i
for each n N. Then for any u L
1
, u

= sup
nN
A
n
u is dened
in L
0
(A), and [[u

> ]] |u|
1
for every > 0.
proof (a) To begin with, suppose that T is positive and that u 0 in L
1
. Note that if v L
1
L

, then
|T
i
v|

|v|

for every i N, so |A
n
v|

|v|

for every n; in particular, A

n
(a) 1 for every n
and every a of nite measure.
For m N and > 0, set
a
m
= [[sup
im
A
i
u > ]].
Then a
m
|u|
1
. PPP Set a = a
m
, w = u a. Of course sup
im
A
i
u belongs to L
1
, so a is nite
and w L
1
. For any i m,
A
i
w = A
i
u A
i
(a) A
i
u 1,
so [[A
i
w > 0]] [[A
i
u > ]]. Accordingly a b, where
b = sup
im
[[A
i
w > 0]] = sup
im
[[w +Tw +. . . +T
i
w > 0]].
By 372B,
_
b
w 0. But this means that
a =
_
b
a =
_
b
u
_
b
w
_
b
u |u|
1
,
as claimed. QQQ
It follows that if we set c

= sup
nN
a
n
, c

1
|u|
1
for every > 0 and inf
>0
c

= 0. But
this is exactly the criterion in 364Mb for u

= sup
nN
A
n
u to be dened in L
0
. And [[u

> ]] = c

, so
[[u

> ]] |u|
1
for every > 0, as required.
(b) Now consider the case of general T, u. In this case T is order-bounded and |[T[| 1, where [T[ is
the modulus of T in L

(L
1
; L
1
) = B(L
1
; L
1
) (371D). If w L
1
L

, then
[[T[w[ [T[[w[ = sup
]w

]]w]
[Tw
t
[ |w|

1,
so |[T[w|

|w|

. Thus [T[ also satises the conditions of the theorem. Setting B

n
=
1
n+1

n
i=0
[T[
i
,
B
n
A
n
in L

(L
1
; L
1
) and B
n
[u[ A
n
u for every n. But by (a), v = sup
nN
B
n
[u[ is dened in L
0
and
[[v > ]] |[u[|
1
= |u|
1
for every > 0. Consequently u

= sup
nN
A
n
u is dened in L
0
and u

v, so
that [[u

> ]] |u|
1
for every > 0.
372D We are now ready for a very general form of the Ergodic Theorem. I express it in terms of the
space M
1,0
from 366F and the class T
(0)
of operators from 371F. If these formulae are unfamiliar, you may
like to glance at the statement of 372E before looking them up.
The Ergodic Theorem: rst form Let (A, ) be a measure algebra, and set M
1,0
= M
1,0
(A, ), T
(0)
=
T
(0)
,
B(M
1,0
; M
1,0
) as in 371F-371G. Take any T T
(0)
, and set A
n
=
1
n+1

n
i=0
T
i
: M
1,0
M
1,0
for
every n. Then for any u M
1,0
, A
n
u)
nN
is order*-convergent (denition: 367A) and | |
1,
-convergent
to a member Pu of M
1,0
. The operator P : M
1,0
M
1,0
is a projection onto the linear subspace u : u
M
1,0
, Tu = u, and P T
(0)
.
proof (a) It will be convenient to start with some elementary remarks. First, every A
n
belongs to T
(0)
,
by 371Ge and 371Ga. Next, A
n
u)
nN
is order-bounded in L
0
= L
0
(A) for any u M
1,0
; this is because if
372D The ergodic theorem 409
u = v +w, where v L
1
= L
1
(A, ) and w L

= L

(A), then A
n
v)
nN
and A
n
(v))
nN
are bounded
above, by 372C, while A
n
w)
nN
is norm- and order-bounded in L

. Accordingly I can uninhibitedly speak

of P

(u) = inf
nN
sup
in
A
i
u and P

(u) = sup
nN
inf
in
A
i
u for any u M
1,0
, these both being dened in
L
0
.
(b) Write V
1
for the set of those u M
1,0
such that A
n
u)
nN
is order*-convergent in L
0
; that is,
P

(u) = P

(u) (367Be). It is easy to see that V is a linear subspace of M

1,0
(use 367Ca and 367Cd). Also
it is closed for | |
1,
.
PPP We know that [T[, taken in L

(M
1,0
; M
1,0
), belongs to T
(0)
(371Gb); set B
n
=
1
n+1

n
i=0
[T[
i
for each
i.
Suppose that u
0
V
1
. Then for any > 0 there is a u V
1
such that |u
0
u|
1,

2
. Write
Pu = P

(u) = P

(u) for the order*-limit of A

n
u)
nN
. Express u
0
u as v +w where v L
1
, w L

and
|v|
1
+|w|

2
2
.
Set v

= sup
nN
B
n
[v[. Then [[v

> ]] 2, by 372C. Next, if w

= sup
nN
B
n
[w[, we surely have
w

2
2
1. Now
[A
n
u
0
A
n
u[ = [A
n
v +A
n
w[ B
n
[v[ +B
n
[w[ v

+w

for every n N, that is,

A
n
u v

A
n
u
0
A
n
u +v

+w

for every n. Because A

n
u)
nN
order*-converges to Pu,
Pu v

(u
0
) P

(u
0
) Pu +v

+w

,
and P

(u
0
) P

(u
0
) 2(v

+w

). On the other hand,

[[2(v

+w

) > 2 + 4
2
]] [[v

> ]] + [[w

> 2
2
]] = [[v

> ]] 2
(using 364Fa for the rst inequality). So
[[P

(u
0
) P

(u
0
) > 2(1 + 2)]] 2.
Since is arbitrary, A
n
u
0
)
nN
order*-converges to P

(u
0
) = P

(u
0
), and u
0
V
1
. As u
0
is arbitrary, V
1
is
closed. QQQ
(c) Similarly, the set V
2
of those u M
1,0
for which A
n
u)
nN
is norm-convergent is a linear subspace
of M
1,0
, and it also is closed. PPP This is a standard argument. If u
0
V
2
and > 0, there is a u V
2
such that |u
0
u|
1,
. There is an n N such that |A
i
u A
j
u|
1,
for all i, j n, and now
|A
i
u
0
A
j
u
0
|
1,
3 for all i, j n, because every A
i
has norm at most 1 in B(M
1,0
; M
1,0
) (371Gc).
As is arbitrary, A
i
u
0
)
nN
is Cauchy; because M
1,0
is complete, it is convergent, and u
0
V
2
. As u
0
is
arbitrary, V
2
is closed. QQQ
(d) Now let V be u+vTu : u M
1,0
L

, v M
1,0
, Tv = v. Then V V
1
V
2
. PPP If u M
1,0
L

,
then for any n N
A
n
(u Tu) =
1
n+1
(u T
n+1
u) 0
for | |

, and therefore is both order*-convergent and convergent for | |

1,
; so u Tu V
1
V
2
. On the
other hand, if Tv = v, then of course A
n
v = v for every n, so again v V
1
V
2
. QQQ
(e) Consequently L
2
= L
2
(A, ) V
1
V
2
. PPP L
2
V
1
V
2
is a linear subspace; but also it is closed for the
norm topology of L
2
, because the identity map from L
2
to M
1,0
is continuous (369Oe). We know also that
TL
2
is an operator of norm at most 1 from L
2
to itself (371Gd). Consequently W = u +v Tu : u, v
L
2
, Tv = v is dense in L
2
(372A). On the other hand, given u L
2
and > 0, there is a u
t
L
2
L

such
that |uu
t
|
2
(take u
t
= (u1)(1) for any large enough), and now |(uTu)(u
t
Tu
t
)|
2
2.
Thus W
t
= u
t
+v Tu
t
: u
t
L
2
L

, v L
2
, Tv = v is dense in L
2
. But W
t
V
1
V
2
, by (d) above.
Thus L
2
V
1
V
2
is dense in L
2
, and is therefore the whole of L
2
. QQQ
(f ) L
2
S(A
f
) is dense in M
1,0
, by 369Pc, so V
1
= V
2
= M
1,0
. This shows that A
n
u)
nN
is norm-
convergent and order*-convergent for every u M
1,0
. By 367D, the limits are the same; by 367F, A
n
u)
nN
is order*-convergent when regarded as a sequence in M
1,0
. Write Pu for the common value of the limits.
410 Linear operators between function spaces 372D
(g) Of course we now have
|Pu|

sup
nN
|A
n
u|

|u|

for every u L

M
1,0
, while
|Pu|
1
liminf |A
n
u|
1
|u|
1
for every u L
1
, by Fatous Lemma. So P T
(0)
. If u M
1,0
and Tu = u, then surely Pu = u,
because A
n
u = u for every u. On the other hand, for any u M
1,0
, TPu = Pu. PPP Because A
n
u)
nN
is
norm-convergent to Pu,
|TPu Pu|
1,
= lim
n
|TA
n
u A
n
u|
1,
= lim
n
1
n+1
|T
n+1
u u|
1,
= 0. QQQ
Thus, writing U = u : Tu = u, P[M
1,0
] = U and Pu = u for every u U.
372E The Ergodic Theorem: second form Let (A, ) be a measure algebra, and let T : L
1
L
1
,
where L
1
= L
1
(A, ), be a linear operator of norm at most 1 such that Tu L

= L

(A) and |Tu|

|u|

whenever u L
1
L

. Set A
n
=
1
n+1

n
i=0
T
i
: L
1
L
1
for every n. Then for any u L
1
, A
n
u)
nN
is order*-convergent to an element Pu of L
1
. The operator P : L
1
L
1
is a projection of norm at most 1
onto the linear subspace u : u L
1
, Tu = u.
proof By 371Ga, there is an extension of T to a member

T of T
(0)
. So 372D tells us that A
n
u)
nN
is
order*-convergent to some Pu L
1
for every u L
1
, and P : L
1
L
1
is a projection of norm at most 1,
because P is the restriction of a projection

P T
(0)
. Also we still have TPu = Pu for every u L
1
, and
Pu = u whenever Tu = u, so the set of values P[L
1
] of P must be exactly u : u L
1
, Tu = u.
Remark In 372D and 372E I have used the phrase order*-convergent from 367 without always being
specic about the partially ordered set in which it is to be interpreted. But, as remarked in 367F, the
notion is robust enough for the omission to be immaterial here. Since both M
1,0
and L
1
are solid linear
subspaces of L
0
, a sequence in M
1,0
is order*-convergent to a member of M
1,0
(when order*-convergence is
interpreted in the partially ordered set M
1,0
) i it is order*-convergent to the same point (when convergence
is interpreted in the set L
0
); and the same applies to L
1
in place of M
1,0
.
372F Corollary Let (A, ) be a measure algebra, and : A
f
A
f
a measure-preserving ring homo-
morphism, where A
f
= a : a < . Let T : M
1,0
M
1,0
be the corresponding Riesz homomorphism,
where M
1,0
= M
1,0
(A, ) (366H, in particular part (a-v)). Set A
n
=
1
n+1

n
i=0
T
i
for n N. Then for every
u M
1,0
, A
n
u)
nN
is order*-convergent and | |
1,
-convergent to some v such that Tv = v.
proof By 366H(a-iv), T T
(0)
, as dened in 371F. So the result follows at once from 372D.
372G Corollary Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean
homomorphism. Let T : L
1
L
1
be the corresponding Riesz homomorphism, where L
1
= L
1
(A, ). Set
A
n
=
1
n+1

n
i=0
T
i
for n N. Then for every u L
1
, A
n
u)
nN
is order*-convergent and | |
1
-convergent. If
we set Pu = lim
n
A
n
u for each u, P is the conditional expectation operator corresponding to the closed
subalgebra C = a : a = a of A.
proof (a) The rst part is just a special case of 372F; the point is that because (A, ) is totally nite,
L

(A) L
1
, so M
1,0
(A, ) = L
1
. Also (because 1 = 1) |u|

|u|
1
for every u L

, so the norm
| |
1,
is actually equal to | |
1
.
(b) For the last sentence, recall that C is a closed subalgebra of A (cf. 333R). By 372D or 372E, P is
a projection operator onto the subspace u : Tu = u. Now [[Tu > ]] = [[u > ]] (365Oc), so Tu = u i
[[u > ]] C for every R, that is, i u belongs to the canonical image of L
1
(C, C) in L
1
(365R). To
identify Pu further, observe that if u L
1
, a C then
_
a
Tu =
_
a
Tu =
_
a
u
(365Ob). Consequently
_
a
T
i
u =
_
a
u for every i N,
_
a
A
n
u =
_
a
u for every n N, and
_
a
Pu =
_
a
u
(because Pu is the limit of A
n
u)
nN
for | |
1
). But this is enough to dene Pu as the conditional expectation
of u on C (365R).
372K The ergodic theorem 411
372H The Ergodic Theorem is most often expressed in terms of transformations of measure spaces. In
the next few corollaries I will formulate such expressions. The translation is straightforward, in view of the
following.
Lemma Let (X, , ) be a measure space with measure algebra (A, ). For h L
0
= L
0
() write h

for
the corresponding member of L
0
= L
0
(A) (364Jc). Now let : X X be an inverse-measure-preserving
function, : A A the corresponding sequentially order-continuous measure-preserving homomorphism
dened by setting E

= (
1
[E])

for E (324M), and T : L

0
L
0
the Riesz homomorphism dened
by setting T(a) = (a) for a A (364R). Then Th

= (h)

for any h L
0
.
proof Let

h : X R be a -measurable function which is equal to h almost everywhere. Because
1
[E]
is negligible for every negligible set E, h =

h a.e., and

h is measurable, so h L
0
. For any R,
[[Th

> ]] = [[T

> ]] = [[

> ]] = x :

h(x) >

= x :

h((x)) >

= [[(

h)

> ]] = [[(h)

> ]].
372I Corollary Let (X, , ) be a measure space and : X X an inverse-measure-preserving
function. Let f be a real-valued function which is integrable over X. Then
g(x) = lim
n
1
n+1

n
i=0
f(
i
(x))
is dened for almost every x X, and g(x) = g(x) for almost every x.
proof Let (A, ) be the measure algebra of (X, , ), and : A A, T : L
0
(A) L
0
(A) the homomor-
phisms corresponding to , as in 372H. Set u = f

in L
1
(A, ). Then for any i N, T
i
u = (f
i
)

(372H),
so setting A
n
=
1
n+1

n
i=0
T
i
, A
n
u = g

n
, where g
n
(x) =
1
n+1

n
i=0
f(
i
(x)) whenever this is dened. Now
we know from 372E or 372F that A
n
u)
nN
is order*-convergent to some v such that Tv = v, so g
n
)
nN
must be convergent almost everywhere (367G), and taking g = lim
n
g
n
where this is dened, g

= v.
Accordingly (g)

= Tv = v = g

and g = g a.e., as claimed.

372J The following straightforward facts will be useful in the next corollary and elsewhere.
Lemma Let (X, , ) be a measure space with measure algebra (A, ). Let : X X be an inverse-
measure-preserving function and : A A the associated homomorphism, as in 372H. Set C = c : c
A, c = c, T = E : E ,
1
[E]E is negligible and T
0
= E : E ,
1
[E] = E. Then T and T
0
are -subalgebras of ; T
0
T, T = E : E , E

C, and C = E

: E T
0
.
proof It is easy to see that T and T
0
are -subalgebras of and that T
0
T = E : E

C. So we have
only to check that if c C there is an E T
0
such that E

= c. PPP Start with any F such that F

= c.
Now F
i
[F] is negligible for every i N, because (
i
[F])

=
i
c = c. So if we set
E =

nN

in

i
[F] = x : there is an n N such that
i
(x) F for every i n,
E

= c. On the other hand, it is easy to check that E T

0
. QQQ
372K Corollary Let (X, , ) be a probability space and : X X an inverse-measure-preserving
function. Let f be a real-valued function which is integrable over X. Then
g(x) = lim
n
1
n+1

n
i=0
f(
i
(x))
is dened for almost every x X; g = g a.e., and g is a conditional expectation of f on the -algebra
T = E : E ,
1
[E]E is negligible. If either f is -measurable and dened everywhere in X
or [E] is negligible for every negligible set E, then g is a conditional expectation of f on the -algebra
T
0
= E : E ,
1
[E] = E.
proof (a) We know by 372I that g is dened almost everywhere and that g = g a.e. In the language of
the proof of 372I, g

= v is the conditional expectation of u = f

on the closed subalgebra

C = a : a A, a = a = F

: F T = F

: F T
0
,
412 Linear operators between function spaces 372K
by 372G and 372J. So v must be expressible as h

where h : X R is T
0
-measurable and is a conditional
expectation of f on T
0
(and also on T). Since every set of measure zero belongs to T, g = h T-a.e., and
g is also a conditional expectation of f on T.
(b) Suppose now that f is dened everywhere and -measurable. Here I come to a technical obstruction.
The denition of conditional expectation in 233D asks for g to be T
0
-integrable, and since -negligible
sets do not need to be T
0
-negligible we have some more checking to do, to conrm that x : x
domg, g(x) = h(x) is T
0
-conegligible as well as -conegligible.
(i) For n N, set
n
=
n
[E] : E ; then
n
is a -subalgebra of , including T
0
. Set

nN

n
, still a -algebra including T
0
. Now any negligible set E

is T
0
-negligible. PPP For
each n N choose F
n
such that E =
n
[F
n
]. Because is inverse-measure-preserving, every F
n
is
negligible, so that
E

mN

nN,jm

j
[F
n
]
is negligible. Of course E =

mN

m
[F
m
] is included in E

. Now

1
[E

] =

mN

nN,jm

j1
[F
n
] =

m1

nN,jm

j
[F
n
] = E

because

nN,j1

j
[F
n
]

nN,j0

j
[F
n
].
So E

T
0
and E is included in a negligible member of T
0
, which is what we needed to know. QQQ
(ii) We are assuming that f is -measurable and dened everywhere, so that g
n
=
1
n+1

n
i=0
f

i
is -measurable and dened everywhere. If we set g

= limsup
n
g
n
, then g

: X [, ] is

-
measurable. PPP For any m N, f

i
is
m
-measurable for every i m, since x : f(
i
(x)) > =
m
[x :
f(
im
(x)) > ] for every . Accordingly
g

= limsup
n
1
n+1

n
i=m
f

i
is
m
-measurable. As m is arbitrary, g

is

-measurable. QQQ
Since h is surely

-measurable, and h = g

-a.e., (i) tells us that h = g

T
0
-a.e. But similarly
h = liminf
n
g
n
T
0
-a.e., so we must have h = g T
0
-a.e.; and g, like h, is a conditional expectation
of f on T
0
.
(c) Finally, suppose that [E] is negligible for every negligible set E. Then every -negligible set is T
0
-
negligible. PPP If E is -negligible, then [E],
2
[E] = [[E]], . . . are all negligible, so E

nN

n
[E] is
negligible, and there is a measurable negligible set F E

. Now F

mN

nm

n
[F] is a negligible
set in T
0
including E, so E is T
0
-negligible. QQQ Consequently g = h T
0
-a.e., and in this case also g is
a conditional expectation of f on T
0
.
372L Remark Parts (b)-(c) of the proof above are dominated by the technical question of the exact
denition of conditional expectation of f on T
0
, and it is natural to be impatient with such details. The
kind of example I am concerned about is the following. Let C [0, 1] be the Cantor set (134G), and
: [0, 1] [0, 1] a Borel measurable function such that [C] = [0, 1] and (x) = x for x [0, 1] C.
(For instance, we could take agreeing with the Cantor function on C (134H).) Because C is negligible,
is inverse-measure-preserving for Lebesgue measure , and if f is any Lebesgue integrable function then
g(x) = lim
n
1
n+1

n
i=0
f(
i
(x)) is dened and equal to f(x) for every x domf C. But for x C we
can, by manipulating , arrange for g(x) to be almost anything; and if f is undened on C then g will also
be undened on C. On the other hand, C is not T
0
-negligible, because the only member of T
0
including
C is [0, 1]. So we cannot be sure of being able to form
_
g d(T
0
).
If instead of Lebesgue measure itself we took its restriction
B
to the algebra of Borel subsets of [0, 1], then
would still be inverse-measure-preserving for
B
, but we should now have to worry about the possibility
that fC was non-measurable, so that gC came out to be non-measurable, even if everywhere dened, and
g was not
B
T
0
-virtually measurable.
372Mc The ergodic theorem 413
In the statement of 372K I have oered two ways of being sure that the problem does not arise: check
that [E] is negligible whenever E is negligible (so that all negligible sets are T
0
-negligible), or check
that f is dened everywhere and -measurable. Even if these conditions are not immediately satised in a
particular application, it may be possible to modify the problem so that they are. For instance, completing
the measure will leave inverse-measure-preserving (343Ac), will not change the integrable functions but will
make them all measurable (212F, 212Bc), and may enlarge T
0
enough to make a dierence. If our function
f is measurable (because the measure is complete, or otherwise) we can extend it to a measurable function
dened everywhere (121I) and the corresponding extension of g will be T
0
-integrable. Alternatively, if
the diculty seems to lie in the behaviour of rather than in the behaviour of f (as in the example above),
it may help to modify on a negligible set.
372M Continued fractions A particularly delightful application of the results above is to a question
which belongs as much to number theory as to analysis. It takes a bit of space to describe, but I hope you
will agree with me that it is well worth knowing in itself, and that it also illuminates some of the ideas
above.
(a) Set X = [0, 1] Q. For x X, set (x) = <
1
x
>, the fractional part of
1
x
, and k
1
(x) =
1
x
(x), the
integer part of
1
x
; then (x) X for each x X, so we may dene k
n
(x) = k
1
(
n1
(x)) for every n 1.
The strictly positive integers k
1
(x), k
2
(x), k
3
(x), . . . are the continued fraction coecients of x. Of
course k
n+1
(x) = k
n
((x)) for every n 1. Now dene p
n
(x))
nN
, q
n
(x))
nN
inductively by setting
p
0
(x) = 0, p
1
(x) = 1, p
n
(x) = p
n2
(x) +k
n
(x)p
n1
(x) for n 1,
q
0
(x) = 1, q
1
(x) = k
1
(x), q
n
(x) = q
n2
(x) +k
n
(x)q
n1
(x) for n 1.
The continued fraction approximations to x are the quotients p
n
(x)/q
n
(x).
(I do not discuss rational x, because for my purposes here these are merely distracting. But if we set
k
1
(0) = , (0) = 0 then the formulae above produce the conventional values for k
n
(x) for rational x [0, 1[.
As for the p
n
and q
n
, use the formulae above until you get to x = p
n
(x)/q
n
(x),
n
(x) = 0, k
n+1
(x) = ,
and then set p
m
(x) = p
n
(x), q
m
(x) = q
n
(x) for m n.)
(b) The point is that the quotients r
n
(x) = p
n
(x)/q
n
(x) are, relatively speaking, good rational approxi-
mations to x. (See 372Yf.) We always have r
n+1
(x) < x < r
n
(x) for every odd n 1 (372Xj). If x = 3,
then the rst few coecients are
k
1
= 7, k
2
= 15, k
3
= 1,
r
1
=
1
7
, r
2
=
15
106
, r
3
=
16
113
;
the rst and third of these corresponding to the classical approximations
22
7
,
355
113
. Or if we take
x = e 2, we get
k
1
= 1, k
2
= 2, k
3
= 1, k
4
= 1, k
5
= 4, k
6
= 1,
r
1
= 1, r
2
=
2
3
, r
3
=
3
4
, r
4
=
5
7
, r
5
=
23
32
, r
6
=
28
39
;
note that the obvious approximations
17
24
,
86
120
derived from the series for e are not in fact as close as the
even terms
5
7
,
28
39
above, and involve larger numbers.
(c) Now we need a variety of miscellaneous facts about these coecients, which I list here.
(i) For any x X, n 1 we have
p
n1
(x)q
n
(x) p
n
(x)q
n1
(x) = (1)
n
,
n
(x) =
p
n
(x)xq
n
(x)
xq
n1
(x)p
n1
(x)
(induce on n), so
x =
p
n
(x)+p
n1
(x)
n
(x)
q
n
(x)+q
n1
(x)
n
(x)
.
414 Linear operators between function spaces 372Mc
(ii) Another easy induction on n shows that for any nite string m = (m
1
, . . . , m
n
) of strictly positive
integers the set D
m
= x : x X, k
i
(x) = m
i
for 1 i n is an interval in X on which
n
is monotonic,
being strictly increasing if n is even and strictly decreasing if n is odd. (For the inductive step, note just
that
D
(m
1
,... ,m
n
)
= [
1
m
1
+1
,
1
m
1
]
1
[D
(m
2
,... ,m
n
)
].)
(iii) We also need to know that the intervals D
m
of (ii) are small; specically, that if m = (m
1
, . . . , m
n
),
the length of D
m
is at most 2
n+1
. PPP All the coecients p
i
, q
i
, for i n, take constant values p

i
, q

i
on
D
m
, since they are determined from the coecients k
i
which are constant on D
m
by denition. Now every
x D
m
is of the form (p

n
+ tp

n1
)/(q

n
+ tq

n1
) for some t X (see (i) above) and therefore lies between
p

n1
/q

n1
and p

n
/q

n
. But the distance between these is

n
q

n1
p

n1
q

n
q

n
q

n1

=
1
q

n
q

n1
,
by the rst formula in (i). Next, noting that q

i
q

i1
+q

i2
for each i 2, we see that q

i
q

i1
2q

i1
q

i2
for i 2, and therefore that q

n
q

n1
2
n1
, so that the length of D
m
is at most 2
n+1
. QQQ
372N Theorem Set X = [0, 1] Q, and dene : X X as in 372M. Then for every Lebesgue
integrable function f on X,
lim
n
1
n+1

n
i=0
f(
i
(x)) =
1
ln 2
_
1
0
f(t)
1+t
dt
for almost every x X.
proof (a) The integral just written, and the phrase almost every, refer of course to Lebesgue measure;
but the rst step is to introduce another measure, so I had better give a name
L
to Lebesgue measure on
X. Let be the indenite-integral measure on X dened by saying that E =
1
ln 2
_
E
1
1+x

L
(dx) whenever
this is dened. The coecient
1
ln 2
is of course chosen to make X = 1. Because
1
1+x
> 0 for every x X,
dom = dom
L
and has just the same negligible sets as
L
(234D); I can therefore safely use the terms
measurable set, almost everywhere and negligible without declaring which measure I have in mind each
time.
(b) Now is inverse-measure-preserving when regarded as a function from (X, ) to itself. PPP For each
k 1, set I
k
=
_
1
k+1
,
1
k
_
. On X I
k
, (x) =
1
x
k. Observe that I
k
: X I
k
X is bijective and
dierentiable relative to its domain in the sense of 262. Consider, for any measurable E X,
_
E
1
(y+k)(y+k+1)

L
(dy) =
_
I
k

1
[E]
1
((x)+k)((x)+k+1)
[
t
(x)[
L
(dx)
=
_
I
k

1
[E]
x
2
x+1
1
x
2

L
(dx) = ln2 (I
k

1
[E]),
using 263D (or more primitive results, of course). But

k=1
1
(y+k)(y+k+1)
=

k=1
1
y+k

1
y+k+1
=
1
y+1
for every y [0, 1], so
E =
1
ln 2

k=1
_
E
1
(y+k)(y+k+1)

L
(dy) =

k=1
(I
k

1
[E]) =
1
[E].
As E is arbitrary, is inverse-measure-preserving. QQQ
(c) The next thing we need to know is that if E X and
1
[E] = E then E is either negligible or
conegligible. PPP I use the sets D
m
of 372M(c-ii).
(i) For any string m = (m
1
, . . . , m
n
) of strictly positive integers, we have
372N The ergodic theorem 415
x =
p

n
+p

n1

n
(x)
q

n
+q

n1

n
(x)
for every x D
m
, where p

n
, etc., are dened from m as in 372M(c-iii). Recall also that
n
is strictly
monotonic on D
m
. So for any interval I [0, 1] (open, closed or half-open) with endpoints < ,

n
[I] D
m
will be of the form X J, where J is an interval with endpoints (p

n
+ p

n1
)/(q

n
+ q

n1
),
(p

n
+p

n1
)/(q

n
+q

n1
) in some order. This means that we can estimate
L
(
n
[I]D
m
)/
L
D
m
, because
it is the modulus of
p

n
+p

n1

n
+q

n1

n
+p

n1

n
+q

n1

n
q

n
+p

n1
q

n
+q

n1
=
()q

n
(q

n
+q

n1
)
(q

n
+q

n1
)(q

n
+q

n1
)

()q

n
q

n
+q

n1

1
2
( ).
Now look at
/ = E : E [0, 1] is Lebesgue measurable,
L
(
n
[E] D
m
)
1
2

L
E
L
D
m
.
Clearly the union of two disjoint members of / belongs to /. Because / contains every subinterval of [0, 1]
it includes the algebra c of subsets of [0, 1] consisting of nite unions of intervals. Next, the union of any
non-decreasing sequence in / belongs to /, and the intersection of a non-increasing sequence likewise. But
this means that / must include the -algebra generated by c (136G), that is, the Borel -algebra. But also,
if E / and H [0, 1] is negligible, then

L
(
n
[EH] D
m
) =
L
(
n
[E] D
m
)
1
2

L
E
L
D
m
=
1
2

L
(EH)
L
D
m
and EH /. And this means that every Lebesgue measurable subset of [0, 1] belongs to / (134Fb).
(ii) ??? Now suppose, if possible, that E is a measurable subset of X and that
1
[E] = E and E is
neither negligible nor conegligible in X. Set =
1
2

L
E > 0. By Lebesgues density theorem (223B) there
is some x X E such that lim
0
() = 0, where () =
1
2

L
(E [x , x + ]) for > 0. Take n so
large that () <
1
2
whenever 0 < 2
n+1
, and set m
i
= k
i
(x) for i n, so that x D
m
. Taking the
least such that D
m
[x , x +], we must have 2
n+1
, because the length of D
m
is at most 2
n+1
(372M(c-iii)), while
L
D
m
, because D
m
is an interval. Accordingly

L
(E D
m
)
L
(E [x , x +]) = 2() <
L
D
m
.
But we also have

L
(E D
m
) =
L
(
n
[E] D
m
)
L
D
m
,
by (i). XXX
This proves the result. QQQ
(d) The nal fact we need in preparation is that [E] is negligible for every negligible E X. This is
because is dierentiable relative to its domain (see 263D(ii)).
(e) Now let f be any
L
-integrable function. Because
1
1+x
1 for every x, f is also -integrable (235M);
consequently, using (b) above and 372K,
g(x) = lim
n
1
n+1

n
i=0
f(
i
(x))
is dened for almost every x X, and is a conditional expectation of f (with respect to the measure ) on
the -algebra T
0
= E : E is measurable,
1
[E] = E. But we have just seen that T
0
consists only of
negligible and conegligible sets, so g must be essentially constant; since
_
g d =
_
fd, we must have
lim
n
1
n+1
n

i=0
f(
i
(x)) =
_
fd =
1
ln2
_
1
0
f(t)
1+t

L
(dt)
for almost every x (using 235M to calculate
_
fd).
416 Linear operators between function spaces 372O
372O Corollary For almost every x [0, 1] Q,
lim
n
1
n
#(i : 1 i n, k
i
(x) = k) =
1
ln 2
(2 ln(k + 1) lnk ln(k + 2))
for every k 1, where k
1
(x), . . . are the continued fraction coecients of x.
proof In 372N, set f = (X [
1
k+1
,
1
k
]). Then (for i 1) f(
i
(x)) = 1 if k
i
(x) = k and zero otherwise. So
lim
n
1
n
#(i : 1 i n, k
i
(x) = k)
= lim
n
1
n
n

i=1
f(
i
(x)) = lim
n
1
n+1
n

i=0
f(
i
(x))
=
1
ln 2
_
1
0
f(t)
1+t
dt =
1
ln2
_
1/k
1/k+1
1
1+t
dt
=
1
ln 2
(ln(1 +
1
k
) ln(1 +
1
k+1
)) =
1
ln2
(2 ln(k + 1) lnk ln(k + 2)),
for almost every x X.
372P Mixing and ergodic transformations This seems an appropriate moment for some brief notes
on two special types of measure-preserving homomorphism or inverse-measure-preserving function.
Denitions (a) Let (A, ) be a probability algebra and : A A a measure-preserving Boolean homo-
morphism.
(i) is ergodic if a : a = a = 0, 1, that is, if the xed-point subalgebra of is trivial.
(ii) is mixing (sometimes strongly mixing) if lim
n
(
n
a b) = a b for all a, b A.
(b) Let (X, , ) be a probability space and : X X an inverse-measure-preserving function.
(i) is ergodic (also called metrically transitive, indecomposable) if every measurable set E such
that
1
[E] = E is either negligible or conegligible.
(ii) is mixing if lim
n
(F
n
[E]) = E F for all E, F .
(c) Remarks (i) The reason for introducing ergodic homomorphisms in this section is of course
372G/372K; if in 372G, or in 372K, is ergodic, then the limit Pu or g must be (essentially) constant,
being a conditional expectation on a trivial subalgebra.
(ii) In the denition (b-i) I should perhaps emphasize that we look only at measurable sets E. We
certainly expect that there will generally be many sets E for which
1
[E] = E, since any union of orbits
of will have this property.
(iii) Part (c) of the proof of 372N was devoted to showing that the function there was ergodic; see
also 372Xw. For another ergodic transformation see 372Xo. For examples of mixing transformations see
333P, 372Xm, 372Xn, 372Xr, 372Xu, 372Xv.
372Q The following facts are elementary.
Proposition (a) Let (A, ) be a probability algebra and : A A a measure-preserving Boolean homo-
morphism.
(i) If is mixing, it is ergodic.
(ii) Let T : L
0
= L
0
(A) L
0
be the Riesz homomorphism such that T(a) = a for every a A.
Then the following are equiveridical: () is ergodic; () the only u L
0
such that Tu = u are the multiples
of 1; () for every u L
1
= L
1
(A, ),
1
n+1

n
i=0
T
i
u)
nN
order*-converges to (
_
u)1.
(b) Let (X, , ) be a probability space, with measure algebra (A, ). Let : X X be an inverse-
measure-preserving function and : A A the associated homomorphism such that E

= (
1
[E])

for
every E .
372R The ergodic theorem 417
(i) is mixing i is, and in this case it is ergodic.
(ii) The following are equiveridical: () is ergodic; () is ergodic; () for every -integrable real-
valued function f,
1
n+1

n
i=0
f(
i
(x)))
nN
converges to
_
f for almost every x X.
proof (a)(i) If is mixing and a = a, then
0 = (a \ a) = lim
n
(
n
a \ a) = a (1 \ a),
so one of a, (1 \ a) must be zero, and a 0, 1. Thus is ergodic.
(ii)()() Tu = u i [[u > ]] = [[u > ]] for every ; if is ergodic, this means that [[u > ]] 0, 1
for every , and u must be of the form 1, where = inf : [[u > ]] = 0.
()() If () is true and u L
1
, then we know from 372G that
1
n+1

n
i=0
T
i
u)
nN
is order*-
convergent and | |
1
-convergent to some v such that Tv = v; by (), v is of the form 1; and
=
_
v = lim
n
1
n+1

n
i=0
_
T
i
u =
_
u.
()() Assuming (), take any a A such that a = a, and consider u = a. Then T
i
u = a for
every i, so
a = lim
n
1
n+1

n
i=0
T
i
u = (
_
u)1 = a 1,
and a must be either 0 or 1.
(b)(i) Simply translating the denitions, we see that is mixing i is. In this case is ergodic, as in
(a-i).
(ii)()() If a = a there is an E such that
1
[E] = E and E

= a, by 372L; now a = E 0, 1,
so a 0, 1.
()() Set u = f

L
1
. In the language of (a), T
i
u = (f
i
)

for each i, by 372H, so that

(
1
n+1

n
i=0
f
i
)

=
1
n+1

n
i=0
T
i
u
is order*-convergent to (
_
u)1 = (
_
f)1, and
1
n+1

n
i=0
f
i

_
f a.e.
()() If
1
[E] = E then, applying () to f = E, we see that E = E X a.e., so that E is
either negligible or conegligible.
372R There is a useful sucient condition for a homomorphism or function to be mixing.
Proposition (a) Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean homo-
morphism. If

nN

n
[A] = 0, 1, then is mixing.
(b) Let (X, , ) be a probability space, and : X X an inverse-measure-preserving function. Set
T = E : for every n N there is an F such that E =
n
[F].
If every element of T is either negligible or conegligible, is mixing.
proof (a) Let T : L
0
= L
0
(A) L
0
be the Riesz homomorphism associated with . Take any a, b A
and any non-principal ultralter T on N. Then T
n
(a))
nN
is a bounded sequence in the reexive space
L
2

= L
2
(A, ), so v = lim
nJ
T
n
(a) is dened for the weak topology of L
2

. Now for each n N set
B
n
=
n
[A]. This is a closed subalgebra of A (314F(a-i)), and contains
i
a for every i n. So if we identify
L
2
(B
n
, B
n
) with the corresponding subspace of L
2

(366I), it contains T
i
(a) for every i n; but also
it is norm-closed, therefore weakly closed (3A5Ee), so contains v. This means that [[v > ]] must belong to
B
n
for every and every n. But in this case [[v > ]]

nN
B
n
= 0, 1 for every , and v is of the form
1. Also
=
_
v = lim
nJ
_
T
n
(a) = a.
So
lim
nJ
(
n
a b) = lim
nJ
_
T
n
(a) b =
_
v b = b = a b.
418 Linear operators between function spaces 372R
But this is true of every non-principal ultralter T on N, so we must have lim
n
(
n
a b) = a b
(3A3Lc). As a and b are arbitrary, is mixing.
(b) The point is that if a

nN

n
[A], there is an E T such that E

= a. PPP For each n N there is

an a
n
A such that
n
a
n
= a; say a
n
= F

n
where F
n
. Then
n
[F
n
]

= a. Set
E =

mN

nm

n
[F
n
], E
k
=

mk

nm

(nk)
[F
n
]
for each k; then E

= a and

k
[E
k
] =

mk

nm

n
[F
n
] =

mN

nm

n
[F
n
] = E
for every k, so E T. QQQ
So

nN
A
n
= 0, 1 and and are mixing.
372X Basic exercises (a) Let U be any reexive Banach space, and T : U U an operator of norm
at most 1. Set A
n
=
1
n+1

n
i=0
T
i
for each n N. Show that Pu = lim
n
A
n
u is dened (as a limit for
the norm topology) for every u U, and that P : U U is a projection onto u : Tu = u. (Hint: show
that u : Pu is dened is a closed linear subspace of U containing Tu u for every u U.)
(This is a version of the mean ergodic theorem.)
>>>(b) Let (A, ) be a measure algebra, and T T
(0)
,
; set A
n
=
1
n+1

n
i=0
T
i
for n N. Take any p [1, [
and u L
p
= L
p
(A, ). Show that A
n
u)
nN
is order*-convergent and | |
p
-convergent to some v L
p
.
(Hint: put 372Xa together with 372D.)
(c) Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean homomorphism.
Let P : L
1
L
1
be the operator dened as in 365P/366Hb, where L
1
= L
1
(A, ), so that
_
a
Pu =
_
a
u
for u L
1
, a A. Set A
n
=
1
n+1

n
i=0
P
i
: L
1
L
1
for each i. Show that for any u L
1
, A
n
u)
nN
is
order*-convergent and | |
1
-convergent to the conditional expectation of u on the subalgebra a : a = a.
(d) Show that if f is any Lebesgue integrable function on R, and y R 0, then
lim
n
1
n+1

n
k=0
f(x +ky) = 0
for almost every x R.
(e) Let (X, , ) be a measure space and : X X an inverse-measure-preserving function. Set
T = E : E , (
1
[E]E) = 0, T
0
= E : E ,
1
[E] = E. (i) Show that T = EF : E
T
0
, F , F = 0. (ii) Show that a set A X is T
0
-negligible i
n
[A] is -negligible for every n N.
>>>(f ) Let be a Radon probability measure on R such that
_
[t[(dt) is nite (cf. 271C-271F). On X = R
N
let be the product measure obtained when each factor is given the measure . Dene : X X by setting
(x)(n) = x(n + 1) for x X, n N. (i) Show that is inverse-measure-preserving. (Hint: 254G. See also
372Xu below.) (iii) Set =
_
t(dt), the expectation of the distribution . By considering
1
n+1

n
i=0
f

i
,
where f(x) = x(0) for x X, show that lim
n
1
n+1

n
i=0
x(i) = for -almost every x X.
>>>(g) Use the Ergodic Theorem to prove Kolmogorovs Strong Law of Large Numbers (273I), as follows.
Given a complete probability space (, , ) and an independent identically distributed sequence F
n
)
nN
of measurable functions from to R, set X = R
N
and F() = F
n
())
nN
for . Show that if we give
each copy of R the distribution of F
0
then F is inverse-measure-preserving for and the product measure
on X. Now use 372Xf.
(h) Show that the continued fraction coecients of
1

2
are 1, 2, 1, 2, . . . .
>>>(i) For x X = [0, 1] Q let k
1
(x), k
2
(x), . . . be its continued-fraction coecients. Show that x
k
n+1
(x) 1)
nN
is a bijection between X and N
N
which is a homeomorphism if X is given its usual
topology (as a subset of R) and N
N
is given its usual product topology (each copy of N being given the
discrete topology).
372Xt The ergodic theorem 419
(j) For any irrational x [0, 1] let k
1
(x), k
2
(x), . . . be its continued-fraction coecients and p
n
(x), q
n
(x)
the numerators and denominators of its continued-fraction approximations, as described in 372M. Write
r
n
(x) = p
n
(x)/q
n
(x). (i) Show that x lies between r
n
(x) and r
n+1
(x) for every n N. (ii) Show that
r
n+1
(x) r
n
(x) = (1)
n
/q
n
(x)q
n+1
(x) for every n N. (iii) Show that [x r
n
(x)[ 1/q
n
(x)
2
k
n
(x) for
every n 1. (iv) Hence show that for almost every R, the set (p, q) : p Z, q 1, [
p
q
[ /q
2
is
innite for every > 0.
(k) Let (A, ) be an atomless probability algebra. Show that the following are equiveridical: (i) A is
homogeneous; (ii) there is an ergodic measure-preserving Boolean homomorphism : A A; (iii) there is
a mixing measure-preserving automorphism : A A. (Hint: 333P.)
(l) Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean homomorphism. (i)
Show that if n 1 then is mixing i
n
is mixing. (ii) Show that if n 1 and
n
is ergodic then is
ergodic. (iii) Show that if is an automorphism then it is ergodic, or mixing, i
1
is.
>>>(m) Consider the tent map

(x) = min(x, 1 x) for x [0, 1], [0, 2]. Show that

2
is inverse-
measure-preserving and mixing for Lebesgue measure on [0, 1]. (Hint: show that
n+1
2
(x) =
2
(<2
n
x>) for
n 1, and hence that (I
n
2
[J]) = I J whenever I is of the form [2
n
k, 2
n
(k + 1)] and J is an
interval.)
(n) Consider the quadratic map

(x) = x(1 x) for x [0, 1], [0, 4]. Show that

4
is inverse-
measure-preserving and mixing for the Radon measure on [0, 1] with density function t 1/
_
t(1 t).
(Hint: show that the transformation t sin
2 t
2
turns it into the tent map.) Show that for almost every x,
lim
n
1
n+1
#(i : i n,
i
4
(x) ) =
2

arcsin

for every [0, 1].

(o) Let (X, , ) be Lebesgue measure on [0, 1[, and x an irrational number [0, 1[. (i) Set (x) =
x +
1
for every x [0, 1[, where x +
1
is whichever of x + , x + 1 belongs to [0, 1[. Show that is
inverse-measure-preserving. (ii) Show that if I [0, 1[ is an interval then lim
n
1
n+1

n
i=0
I(
i
(x)) = I
for almost every x [0, 1[. (Hint: this is Weyls Equidistribution Theorem (281N).) (iii) Show that is
ergodic. (Hint: take the conditional expectation operator P of 372G, and look at P(I

) for intervals I.)

(iv) Show that
n
is ergodic for any n Z 0. (v) Show that is not mixing.
(p) Let (A, ) be a probability algebra and : A A a mixing measure-preserving homomorphism. Let
T : L
0
(A) L
0
(A) be the corresponding homomorphism. Let p, q [1, ] be such that
1
p
+
1
q
= 1. Show
that lim
n
_
T
n
u v =
_
u
_
v whenever u L
p
(A, ) and v L
q
(A, ). (Hint: start with u, v S(A).)
(q) Let (X, , ) be a probability space and : X X a mixing inverse-measure-preserving function.
Let p, q [1, ] be such that
1
p
+
1
q
= 1. Show that lim
n
_
f(
n
(x))g(x)dx =
_
f
_
g whenever f L
p
()
and g L
q
().
(r) Give [0, 1[ Lebesgue measure , and let k 2 be an integer. Dene : [0, 1[ [0, 1[ by setting
(x) = <kx>, the fractional part of kx. Show that is inverse-measure-preserving. Show that is mixing.
(Hint: if I = [k
n
i, k
n
(i + 1)[, J = [k
n
j, k
n
(j + 1)[ then (I
m
[J]) = I J for all m n.)
(s) Let (X, , ) be a probability space and : X X an ergodic inverse-measure-preserving function.
Let f be a -virtually measurable function dened almost everywhere on X such that
_
fd = . Show
that lim
n
1
n+1

n
i=0
f(
i
(x)) = for almost every x X. (Hint: look at the corresponding limits for
f
m
= f mX.)
(t) For irrational x [0, 1], write k
1
(x), k
2
(x), . . . for the continued-fraction coecients of x. Show that
lim
n
1
n

n
i=1
k
i
(x) = for almost every x. (Hint: take , as in 372N, and show that
_
k
1
d = .)
420 Linear operators between function spaces 372Xu
(u) Let (X, , ) be any probability space, and let be the product measure on X
N
. Dene : X
N
X
N
by setting (x)(n) = x(n+1). Show that is inverse-measure-preserving. Show that satises the conditions
of 372R, so is mixing.
(v) Let (X, , ) be any probability space, and let be the product measure on X
Z
. Dene : X
Z
X
Z
by setting (x)(n) = x(n + 1). Show that is inverse-measure-preserving. Show that is mixing. (Hint:
show that if C, C
t
are basic cylinder sets then (C
n
[C
t
]) = C C
t
for all n large enough.) Show that
does not satisfy the conditions of 372R. (Compare 333P.)
(w) In 372N, let T
1
be the family E : for every n N there is a measurable set F X such that

n
[F] = E. Show that every member of T
1
is either negligible or conegligible. (Hint: the argument of
part (c) of the proof of 372N still works.) Hence show that is mixing for the measure .
372Y Further exercises (a) In 372D, show that the null space of the limit operator P is precisely the
closure in M
1,0
of the subspace Tu u : u M
1,0
.
(b) Let (A, ) be a measure algebra and T T
(0)
,
. Take p ]1, [ and u L
p
(A, ), and set u

=
sup
nN
1
n+1

n
i=0
[T
i
u[. (i) Show that for any > 0,
[[u

> ]]
2

_
[[]u]>/2]]
[u[.
(Hint: apply 372C to ([u[
1
2
1)
+
.) (ii) Show that |u

|
p
2(
p
p1
)
1/p
|u|
p
. (Hint: show that
_
[[]u]>]]
[u[ =
[[[u[ > ]] +
_

[[[u[ > ]]d; see 365A. Use 366Xa to show that

|u

|
p
p
2p
_

0

p2
_

/2
[[[u[ > ]]dd + 2
p
|u|
p
p
,
and reverse the order of integration. Compare 275Yc.) (This is Wieners dominated ergodic theorem.)
(c) Let (A, ) be a probability algebra and T an operator in T
(0)
,
. Take u L
1
= L
1
(A, ) such that
h([u[) L
1
, where h(t) = t lnt for t 1, 0 for t 1, and

h is the corresponding function from L
0
(A) to
itself. Set u

= sup
nN
1
n+1

n
i=0
[T
i
u[. Show that u

L
1
. (Hint: use the method of 372Yb to show that
_

2
[[u

> ]]d 2
_

h(u).)
(d) In 372G, suppose that A is atomless. Show that there is always an a A such that a
1
2
and
inf
in

i
a ,= 0 for every n, so that (except in trivial cases) A
n
(a))
nN
will not be | |

-convergent.
(e) Show that an irrational x ]0, 1[ has an eventually periodic sequence of continued fraction coecients
i it is a solution of a quadratic equation with integral coecients.
(f ) In the language of 372M-372O and 372Xj, show the following. (i) For any x X, n 2, q
n
(x)q
n1
(x)
2
n1
, p
n
(x)p
n+1
(x) 2
n1
, so that q
n+1
(x)p
n
(x) 2
n1
and [1 x/r
n
(x)[ 2
n+1
, [ lnx lnr
n
(x)[
2
n+2
. Also [x r
n
(x)[ 1/q
n
(x)q
n+2
(x). (ii) For any x X, n 1, p
n+1
(x) = q
n
((x)) and
q
n
(x)

n1
i=0
r
ni
(
i
(x)) = 1. (iii) For any x X, n 1, [ lnq
n
(x) +

n1
i=0
ln
i
(x)[ 4. (iv) For al-
most every x X,
lim
n
1
n
lnq
n
(x) =
1
ln 2
_
1
0
ln t
1+t
dt =

2
12 ln 2
.
(Hint: 225Xi, 282Xo.) (v) For almost every x X, lim
n
1
n
ln[x r
n
(x)[ =
2
/6 ln2. (vi) For almost
every x X, 11
n
[x r
n
(x)[ 10
n
and 3
n
q
n
(x) 4
n
for all but nitely many n.
(g) In 372M, show that for any measurable set E X, lim
n

n
[E] = E. (Hint: recall that is
mixing for (372Xw). Hence show that lim
n
_

n
[E]
g d = E
_
g d for any integrable g. Apply this
to a Radon-Nikod ym derivative of
L
with respect to .) (I understand that this result is due to Gauss.)
372 Notes The ergodic theorem 421
(h) Let (X
i
,
i
,
i
))
iI
be any family of probability spaces, with product (X, , ). Suppose that for each
i I we are given an inverse-measure-preserving function
i
: X
i
X
i
. Show that there is a corresponding
inverse-measure-preserving function : X X given by setting (x)(i) =
i
(x(i)) for x X, i I. Show
that if each
i
is mixing so is .
(i) Give an example of an ergodic measure-preserving automorphism : [0, 1[ [0, 1[ such that
2
is not
ergodic. (Hint: set (x) =
1
2
(1 +
0
(2x)) for x <
1
2
, x
1
2
for x
1
2
. See also 387Xg.)
(j) Show that there is an ergodic : [0, 1] [0, 1] such that (
1
,
2
) ((
1
), (
2
)) : [0, 1]
2
[0, 1]
2
is
not ergodic. (Hint: 372Xo.)
(k) Let M be an r r matrix with integer coecients and non-zero determinant, where r 1. Let
: [0, 1[
r
[0, 1[
r
be the function such that (x) Mx Z
r
for every x [0, 1[
r
. Show that is inverse-
measure-preserving for Lebesgue measure on [0, 1[
r
.
372 Notes and comments I have chosen an entirely conventional route to the Ergodic Theorem here,
through the Mean Ergodic Theorem (372Xa) or, rather, the fundamental lemma underlying it (372A), and
the Maximal Ergodic Theorem (372B-372C). What is not to be found in every presentation is the generality
here. I speak of arbitrary T T
(0)
, the operators which are contractions both for | |
1
and for | |

, not
requiring T to be positive, let alone correspond to a measure-preserving homomorphism. (I do not mention
T
(0)
in the statement of 372C, but of course it is present in spirit.) The work we have done up to this
point puts this extra generality within easy reach, but as the rest of the section shows, it is not needed for
the principal examples. Only in 372Xc do I oer an application not associated with a measure-preserving
homomorphism or an inverse-measure-preserving function.
The Ergodic Theorem is an almost-everywhere pointwise convergence theorem, like the strong law(s)
of large numbers and the martingale theorem(s) (273, 275). Indeed Kolmogorovs form of the strong law
can be derived from the Ergodic Theorem (372Xg). There are some very strong family resemblances. For
instance, the Maximal Ergodic Theorem corresponds to the most basic of all the martingale inequalities
(275D). Consequently we have similar results, obtained by similar methods, concerning the domination of
sequences starting from members of L
p
(372Yb, 275Yc), though the inequalities are not identical. (Com-
pare also 372Yc with 275Yd.) There are some tantalising reections of these traits in results surrounding
Carlesons theorem on the pointwise convergence of square-integrable Fourier series (see 286 notes), but
Carlesons theorem seems to be much harder than the others. Other forms of the strong law (273D, 273H)
do not appear to t into quite the same pattern, but I note that here, as with the Ergodic Theorem, we
begin with a study of square-integrable functions (see part (e) of the proof of 372D).
After 372D, there is a contraction and concentration in the scope of the results, starting with a simple
replacement of M
1,0
with L
1
(372E). Of course it is almost as easy to prove 372D from 372E as the other
way about; I give precedence to 372D only because M
1,0
is the space naturally associated with the class
T
(0)
of operators to which these methods apply. Following this I turn to the special family of operators to
which the rest of the section is devoted, those associated with measure-preserving homomorphisms (372F),
generally on probability spaces (372G). This is the point at which we can begin to identify the limit as a
conditional expectation as well as an invariant element.
Next comes the translation into the language of measure spaces and inverse-measure-preserving functions,
all perfectly straightforward in view of the lemmas 372H (which was an exercise in 364) and 372J. These
turn 372F into 372I and 372G into the main part of 372K.
In 372K-372L I nd myself writing at some length about a technical problem. The root of the diculty is
in the denition of conditional expectation. Now it is generally accepted that any pure mathematician has
Humpty Dumptys privilege: When I use a word, it means just what I choose it to mean neither more
nor less. With any privilege come duties and responsibilities; in this context, the duty to be self-consistent,
and the responsibility to try to use terms in ways which will not mystify or mislead the unprepared reader.
Having written down a denition of conditional expectation in Volume 2, I must either stick to it, or go
back and change it, or very carefully explain exactly what modication I wish to make here. I dont wish
to suggest that absolute consistency in terminology or anything else is supreme among mathematical
virtues. Surely it is better to give local meanings to words, or tolerate ambiguities, than to suppress ideas
422 Linear operators between function spaces 372 Notes
which cannot be formulated eectively otherwise, and among ideas I wish to include the analogies and
resonances which a suitable language can suggest. But I do say that it is always best to be conscious of
what one is doing I go farther: one of the things which mathematics is for, is to raise our consciousness of
what our thoughts really are. So I believe it is right to pause occasionally over such questions.
In 372M-372O (see also 372Xj, 372Xt, 372Xw, 372Yf, 372Yg) I make an excursion into number theory.
This is a remarkable example of the power of advanced measure theory to give striking results in other
branches of mathematics. Everything here is derived from Billingsley 65, who goes farther than I have
space for, and gives references to more. Here let me point to 372Xi; almost accidentally, the construction
oers a useful formula for a homeomorphism between two of the most important spaces of descriptive set
theory, which will be important to us in Volume 4.
I end the section by introducing two terms, ergodic and mixing transformation, not because I wish to
use them for any new ideas (apart from the elementary 372R, these must wait for 385-386) but because it
may help if I immediately classify some of the inverse-measure-preserving functions we have seen (372Xm-
372Xo, 372Xr, 372Xt-372Xv). Of course in any application of any ergodic theorem it is of great importance
to be able to identify the limits promised by the theorem, and the point about an ergodic transformation
is just that our averages converge to constant limits (372Q). Actually proving that a given inverse-measure-
preserving function is ergodic is rarely quite trivial (see 372N, 372Xn, 372Xo), though a handful of standard
techniques cover a large number of cases, and it is usually obvious when a map is not ergodic, so that if an
invariant region does not leap to the eye one has a good hope of ergodicity.
I take the opportunity to mention two famous functions from [0, 1] to itself, the tent and quadratic
maps (372Xm, 372Xn). In the formulae

I include redundant parameters; this is because the real

importance of these functions lies in the way their behaviour depends, in bewildering complexity, on these
parameters. It is only at the extreme values = 2, = 4 that the methods of this volume can tell us
anything interesting.
373 Decreasing rearrangements
I take a section to discuss operators in the class T
(0)
of 371F-371H and 372 and two associated classes
T , T

(373A). These turn out to be intimately related to the idea of decreasing rearrangement (373C). In
373D-373F I give elementary properties of decreasing rearrangements; then in 373G-373O I show how they
may be used to characterize the set Tu : T T for a given u. The argument uses a natural topology on
the set T (373K). I conclude with remarks on the possible values of
_
Tu v for T T (373P-373Q) and
identications between T
(0)
,
, T
(0)
,
and T

,
(373R-373T).
373A Denition Let (A, ) and (B, ) be measure algebras. Recall that M
1,

= L
1
(A, ) + L

(A)
is the set of those u L
0
(A) such that ([u[ 1)
+
is integrable for some , its norm | |
1,
being dened
by the formulae
|u|
1,
= min|v|
1
+|w|

: v L
1
, w L

, v +w = u
= min +
_
([u[ 1)
+
: 0
(369Ob).
(a) T
,
will be the space of linear operators T : M
1,

M
1,

such that |Tu|
1
|u|
1
for every u L
1

and |Tu|

|u|

for every u L

(A). (Compare the denition of T

(0)
in 371F.)
(b) If B is Dedekind complete, so that M
1,

, being a solid linear subspace of the Dedekind complete
space L
0
(B), is Dedekind complete, T

,
will be T
,
L

(M
1,

; M
1,

).
373B Decreasing rearrangements 423
373B Proposition Let (A, ) and (B, ) be measure algebras.
(a) T = T
,
is a convex set in the unit ball of B(M
1,

; M
1,

).
(b) If T T then TM
1,0

belongs to T
(0)
,
, as dened in 371F. So if T T , p [1, [ and u L
p

then
Tu L
p

and |Tu|
p
|u|
p
.
(c) If B is Dedekind complete and T T , then T L

(M
1,

; M
1,

) and T
1
T whenever T
1

L

(M
1,

; M
1,

) and [T
1
[ [T[; in particular, [T[ T .
(d) If : A B is a measure-preserving Boolean homomorphism, then we have a corresponding operator
T T dened by saying that T(a) = (a) for every a A. If is order-continuous, then so is T.
(e) If (C,

) is another measure algebra and T T , S T

,

then ST T
,

.
proof (a) As 371G, parts (a-i) and (a-ii) of the proof.
(b) If u M
1,0

and > 0, then u is expressible as u
t
+u
tt
where |u
tt
|

and u
t
L
1

. (Set
u
tt
= (u
+
1) (u

1).)
So
([Tu[ 1)
+
[Tu Tu
tt
[ L
1

.
As is arbitrary, Tu M
1,0

; as u is arbitrary, TM
1,0

T
(0)
. Now the rest is a consequence of 371Gd.
(c) Because M
1,

is a solid linear subspace of L
0
(B), which is Dedekind complete because B is,
L

(M
1,

; M
1,

) is a Riesz space (355Ea).
Take any u 0 in M
1,

. Let 0 be such that (u 1)
+
L
1

. Because TL
1

belongs to
B(L
1

; L
1

) = L

(L
1

; L
1

) (371D), w
0
= supTv : v L
1

, 0 v (u 1)
+
is dened in L
1

. Now if
v M
1,

and 0 v u, we must have
Tv = T(v 1)
+
+T(v 1) w
0
+1 M
1,

.
Thus Tv : 0 v u is bounded above in M
1,

. As u is arbitrary, T L

(M
1,

; M
1,

) (355Ba).
Now take T
1
such that [T
1
[ [T[ in L

(M
1,

; M
1,

). 371D also tells us that |[TL
1

[| = |TL
1

|, so
that
|T
1
u|
1
= |[T
1
u[|
1
|[T
1
[[u[|
1
|[T[[u[|
1
= | sup
]v]]u]
Tv|
1
| sup
]v]]u]
v|
1
= |u|
1
for every u L
1

(using one of the formulae in 355Eb for the rst equality). At the same time, if u L

(A),
then
[T
1
u[ [T
1
[[u[ [T[[u[ = sup
]v]]u]
Tv
sup
]v]]u]
|Tv|

1 sup
]v]]u]
|v|

1 = |u|

1,
so |T
1
u|

|u|

. Thus T
1
T .
(d) By 365O and 363F, we have norm-preserving positive linear operators T
1
: L
1

L
1

and T

:
L

(A) L

(B) dened by saying that T

1
(a) = (a) whenever a < and T

(a) = (a) for every

a A. If u S(A
f
) = L
1

S(A) (365F), then T
1
u = T

u, because both T
1
and T

are linear and they

agree on a : a < . If u 0 in M
,1

= L
1

L

(A), there is a non-decreasing sequence u

n
)
nN
in
S(A
f
) such that u = sup
nN
u
n
and
lim
n
|u u
n
|
1
= lim
n
|u u
n
|

= 0
(see the proof of 369Od), so that
T
1
u = sup
nN
T
1
u
n
= sup
nN
T

u
n
= T

u.
Accordingly T
1
and T

agree on L
1

L

(A). But this means that if u M

1,

is expressed as v+w = v
t
+w
t
,
where v, v
t
L
1

and w, w
t
L

(A), we shall have

424 Linear operators between function spaces 373B
T
1
v
t
+T

w
t
= T
1
v +T

w +T
1
(v
t
v) T

(w w
t
) = T
1
v +T

w,
because v
t
v = w w
t
M
,1

. Accordingly we have an operator T : M
1,

M
1,

dened by setting
T(v +w) = T
1
v +T

w whenever v L
1

, w L

(A).
This formula makes it easy to check that T is linear and positive, and it clearly belongs to T .
To see that T is uniquely dened, observe that TL
1

and TL

(A) are uniquely dened by the values T

takes on S(A
f
), S(A) respectively, because these spaces are dense for the appropriate norms.
Now suppose that is order-continuous. Then T
1
and T

are also order-continuous (365Oa, 363Ff). If

A M
1,

is non-empty and downwards-directed and has inmum 0, take u
0
A and > 0 such that
(u
0
1)
+
L
1

. Set
A
1
= (u 1)
+
: u A, u u
0
, A

= u 1 : u A.
Then A
1
L
1

and A

(A) are both downwards-directed and have inmum 0, so inf T

1
[A
1
] =
inf T

[A

] = 0 in L
0
(B). But this means that inf(T
1
[A
1
] + T

[A

]) = 0 (351Dc). Now any w

T
1
[A
1
] + T

[A

] is expressible as T(u 1)
+
+ T(u
t
1) where u, u
t
A; because A is downwards-
directed, there is a v A such that v u u
t
, in which case Tv w. Accordingly T[A] must also have
inmum 0. As A is arbitrary, T is order-continuous.
(e) is obvious, as usual.
373C Decreasing rearrangements The following concept is fundamental to any understanding of
the class T . Let (A, ) be a measure algebra. Write M
0,

= M
0,
(A, ) for the set of those u L
0
(A)
such that [[[u[ > ]] is nite for some R. (See 369N for the ideology of this notation.) It is easy to see
that M
0,

is a solid linear subspace of L
0
(A). Let (A
L
,
L
) be the measure algebra of Lebesgue measure
on [0, [. For u M
0,

its decreasing rearrangement is u

M
0,

L
, dened by setting u

= g

, where
g(t) = inf : 0, [[[u[ > ]] t
for every t > 0. (This is always nite because inf
R
[[[u[ > ]] = 0, by 364A() and 321F.)
I will maintain this usage of the symbols A
L
,
L
, u

for the rest of this section.

373D Lemma Let (A, ) be a measure algebra.
(a) For any u M
0,

, its decreasing rearrangement u

may be dened by the formula

[[u

> ]] = [0, [[[u[ > ]][

for every 0,
that is,

L
[[u

> ]] = [[[u[ > ]] for every 0.

(b) If [u[ [v[ in M
0,

, then u

; in particular, [u[

= u

.
(c)(i) If u =

n
i=0

i
a
i
, where a
0
a
1
. . . a
n
and
i
0 for each i, then u

n
i=0

i
[0, a
i
[

.
(ii) If u =

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint and [
0
[ [
1
[ . . . [
n
[, then u

n
i=0
[
i
[[
i
,
i+1
[

, where
i
=

j<i
a
i
for i n + 1.
(d) If E ]0, [ is any Borel set, and u M
0

, then
L
[[u

E]] = [[[u[ E]].

(e) Let h : [0, [ [0, [ be a non-decreasing function such that h(0) = 0, and write

h for the
corresponding functions on L
0
(A)
+
and L
0
(A
L
)
+
(364I). Then (

h(u))

=

h(u

) whenever u 0 in M
0

. If
h is continuous on the left, (

h(u))

=

h(u

) whenever u 0 in M
0,

.
(f) If u M
0,

and 0, then
(u

1)
+
= (([u[ 1)
+
)

.
(g) If u M
0,

, then for any t > 0
_
t
0
u

= inf
0
t +
_
([u[ 1)
+
.
(h) If A (M
0,

)
+
is non-empty and upwards-directed and has supremum u
0
M
0,

, then u

0
=
sup
uA
u

.
373D Decreasing rearrangements 425
proof (a) Set
g(t) = inf : [[[u[ > ]] t
as in 373C. If 0,
g(t) > [[[u[ > ]] > t for some > [[[u[ > ]] > t
(because [[[u[ > ]] = sup
>
[[[u[ > ]]), so
[[u

> ]] = t : g(t) >

= [0, [[[u[ > ]][

.
Of course this formula denes u

.
(b) This is obvious, either from the denition in 373C or from (a) just above.
(c)(i) Setting v =

n
i=0

i
[0, a
i
[

, we have
[[v > ]] = 0 if
n

i=0

i

= [0, a
j
[

if
j1

i=0

i
<
j

i=0

i
= [0, a
0
[

if 0 <
0
,
and in all cases is equal to [0, [[[u[ > ]][

.
(ii) A similar argument applies. (If any a
j
has innite measure, then a
i
is irrelevant for i > j.)
(d) Fix > 0 for the moment, and consider
/ = E : E ], [ is a Borel set,
L
[[u

E]] = [[[u[ E]],

1 = ], [ : .
Then 1 / (by (a)), I J 1 for all I, J 1, E F / whenever E, F / and F E (because
u M
0

, so [[[u[ E]] < ), and

nN
E
n
/ whenever E
n
)
nN
is a non-decreasing sequence in /. So,
by the Monotone Class Theorem (136B), / includes the -algebra of subsets of ], [ generated by 1; but
this must contain E ], [ for every Borel set E R.
Accordingly, for any Borel set E ]0, [,

L
[[u

E]] = sup
nN

L
[[u

E ]2
n
, [ ]] = [[[u[ E]].
(e) For any > 0, E

= t : h(t) > is a Borel subset of ]0, [. If u M

0

then, using (d) above,

L
[[

h(u

) > ]] =
L
[[u

]] = [[u E

]] = [[

h(u) > ]] =
L
[[(

h(u))

> ]].
As both (

h(u))

and

h(u

) are equivalence classes of non-increasing functions, they must be equal.

If h is continuous on the left, then E

= ], [ for some , so we no longer need to use (d), and the

argument works for any u (M
0,

)
+
.
(f ) Apply (e) with h() = max(0, ).
(g) Express u

as g

, where
g(s) = inf : [[[u[ > ]] s
for every s > 0. Because g is non-increasing, it is easy to check that, for t > 0,
_
t
0
g = tg(t) +
_

0
max(0, g(s) g(t))ds t +
_

0
max(0, g(s) )ds
for every 0; so that
_
t
0
u

= min
0
t +
_
(u

1)
+
.
426 Linear operators between function spaces 373D
Now
_
(u

1)
+
=
_

0

L
[[(u

1)
+
> ]]d
=
_

0
[[([u[ 1)
+
> ]]d =
_
([u[ 1)
+
for every 0, using (f) and 365A, and
_
t
0
u

= min
0
t +
_
([u[ 1)
+
.
(h)
[[u
0
> ]] = (sup
uA
[[u > ]]) = sup
uA
[[u > ]]
for any > 0, using 364Mb and 321D. So
[[u

0
> ]] = [0, [[u
0
> ]][

= sup
uA
[0, [[u > ]][

= [[sup
uA
u

> ]]
for every , and u

0
= sup
uA
u

.
373E Theorem Let (A, ) be a measure algebra. Then
_
[u v[
_
u

for all u, v M
0,

.
proof (a) Consider rst the case u, v 0 in S(A). Then we may express u, v as

m
i=0

i
a
i
,

n
j=0

j
b
j
where a
0
a
1
. . . a
m
, b
0
. . . b
n
in A and
i
,
j
0 for all i, j (361Ec). Now u

, v

are given by
u

m
i=0

i
[0, a
i
[

, v

n
j=0

j
[0, b
j
[

(373Dc). So
_
u v =
m

i=0
n

j=0

j
(a
i
b
j
)
m

i=0
n

j=0

j
min( a
i
, b
j
)
=
m

i=0
n

j=0

L
([0, a
i
[ [0, b
j
[) =
_
u

.
(b) For the general case, we have non-decreasing sequences u
n
)
nN
, v
n
)
nN
in S(A)
+
with suprema [u[,
[v[ respectively (364Kd), so that
[u v[ = [u[ [v[ = sup
nN
[u[ v
n
= sup
m,nN
u
m
v
n
= sup
nN
u
n
v
n
and
_
[u v[ =
_
sup
nN
u
n
v
n
= sup
nN
_
u
n
v
n
sup
nN
_
u

n
v

n

_
u

,
using 373Db.
373F Theorem Let (A, ) be a measure algebra, and u any member of M
0,

.
(a) For any p [1, ], u L
p

i u

L
p

L
, and in this case |u|
p
= |u

|
p
.
(b)(i) u M
0

i u

M
0

L
;
(ii) u M
1,

i u

M
1,

L
, and in this case |u|
1,
= |u

|
1,
;
(iii) u M
1,0

i u

M
1,0

L
;
(iv) u M
,1

i u

M
,1

L
, and in this case |u|
,1
= |u

|
,1
.
proof (a)(i) Consider rst the case p = 1. In this case
_
[u[ =
_

0
[[[u[ > ]]d =
_

0

L
[[u

> ]]d =
_
u

.
(ii) If 1 < p < , then by 373De we have ([u[
p
)

= (u

)
p
, so that
|u|
p
p
=
_
[u[
p
=
_
([u[
p
)

=
_
(u

)
p
= |u

|
p
p
373H Decreasing rearrangements 427
if either |u|
p
or |u

|
p
is nite. (iii) As for p = ,
|u|

[[[u[ > ]] = 0 [[u

> ]] = 0 |u

.
(b)(i)
u M
0

[[[u[ > ]] < for every > 0

L
[[u

> ]] < for every > 0 u

M
0

L
.
(ii) For any 0,
_
([u[ 1)
+
=
_
(u

1)
+
as in the proof of 373Dg. So |u|
1,
= |u

|
1,
if either is nite, by the formula in 369Ob.
(iii) This follows from (i) and (ii), because M
1,0
= M
0
M
1,
.
(iv) Allowing as a value of an integral, we have
|u|
1,
= min +
_
([u[ 1)
+
: 0
= min +
_
(u

1)
+
: 0 = |u

|
1,
by 369Ob; in particular, u M
1,

i u

M
1,

L
.
373G Lemma Let (A, ) and (B, ) be measure algebras. If
either u M
1,

and T T
,
or u M
1,0

and T T
(0)
,
,
then
_
t
0
(Tu)

_
t
0
u

for every t 0.
proof Set T
1
= TL
1

, so that |T
1
| 1 in B(L
1

; L
1

), and [T
1
[ is dened in B(L
1

; L
1

), also with norm
at most 1. If 0, then we can express u as u
1
+ u
2
where [u
1
[ ([u[ 1)
+
and [u
2
[ 1. (Let
w L

(A) be such that |w|

1, u = [u[ w; set u
2
= w ([u[ 1).) So if
_
([u[ 1)
+
< ,
[Tu[ [Tu
1
[ +[Tu
2
[ [T
1
[[u
1
[ +1
and
_
([Tu[ 1)
+

_
[T
1
[[u
1
[
_
[u
1
[
_
([u[ 1)
+
.
The formula of 373Dg now tells us that
_
t
0
(Tu)

_
t
0
u

for every t.
373H Lemma Let (A, ) be a measure algebra, and : A
f
R an additive functional, where A
f
=
a : a < .
(a) The following are equiveridical:
() lim
t0
sup
at
[a[ = lim
t
1
t
sup
at
[a[ = 0,
() there is some u M
1,0
= M
1,0

such that a =
_
a
u for every a A
f
,
and in this case u is uniquely dened.
(b) Now suppose that (A, ) is localizable. Then the following are equiveridical:
() lim
t0
sup
at
[a[ = 0, limsup
t
1
t
sup
at
[a[ < ,
() there is some u M
1,
= M
1,

such that a =
_
a
u for every a A
f
,
and again this u is uniquely dened.
proof (a)(i) Assume (). For a, c A
f
, set
c
(a) = (a c). Then for each c A
f
, there is a unique
u
c
L
1

such that
c
a =
_
a
u
c
for every a A
f
(365Eb). Because u
c
is unique we must have u
c
= u
d
c
428 Linear operators between function spaces 373H
whenever c d A
f
. Next, given > 0, there is a t
0
0 such that [a[ a whenever a A
f
and
a t
0
; so that [[u
c
> ]] t
0
for every c A
f
, and e() = sup
cA
f [[u
+
c
> ]] is dened in A
f
. Of course
e() = [[u
+
e(1)
> ]] for every 1, so inf
R
e() = 0, and v
1
= sup
cA
f u
+
c
is dened in L
0
= L
0
(A)
(364Mb). Because [[v
1
> ]] = e() A
f
for each > 0, v
1
M
0

. For any a A
f
,
v
1
a = sup
cA
f u
+
c
a = u
+
a
,
so v
1
M
1,0
and
_
a
v
1
=
_
a
u
+
a
for every a A
f
.
Similarly, v
2
= sup
cA
f u

c
is dened in M
1,0
and
_
a
v
2
=
_
a
u

a
for every a A
f
. So we can set
u = v
1
v
2
M
1,0
and get
_
a
u =
_
a
u
a
= a
for every a A
f
. Thus () is true.
(ii) Assume (). If > 0, there is a > 0 such that
_
a
([u[ 1)
+
whenever a (365Ea), so
that [
_
a
u[ (1 + a) whenever a . As is arbitrary, lim
t0
sup
at
[
_
a
u[ = 0. Moreover, whenever
t > 0 and a t,
1
t
[
_
a
u[ +
1
t
_
([u[ 1)
+
. Thus
limsup
t
1
t
sup
at
[
_
a
u[ .
As is arbitrary, satises the conditions in ().
(iii) The uniqueness of u is a consequence of 366Gd.
(b) The argument for (b) uses the same ideas.
(i) Assume (). Again, for each c A
f
, we have u
c
L
1
such that
c
a =
_
a
u
c
for every a A
f
; again,
set e() = sup
cA
f [[u
+
c
> ]], which is dened because A is supposed to be Dedekind complete. This time,
there are t
0
, 0 such that [a[ a whenever a A
f
and a t
0
; so that [[u
c
> ]] t
0
for every
c A
f
, and e() < . Accordingly
inf

e() = inf

[[u
+
e()
> ]] = 0,
and once more v
1
= sup
cA
f u
+
c
is dened in L
0
= L
0
(A). As before, v
1
a = u
+
a
L
1
for any a A
f
,
Because [[v
1
> ]] = e() A
f
, v
1
M
1,
. Similarly, v
2
= sup
cA
f u

c
is dened in M
1,
, with v
2
a = u

a
for every a A
f
. So u = v
1
v
2
M
1,
, and
_
a
u =
_
a
u
a
= a
for every a A
f
.
(ii) Assume (). Take 0 such that =
_
([u[ 1)
+
is nite. If > 0, there is a > 0 such
that
_
a
([u[ 1)
+
whenever a , so that [
_
a
u[ + a whenever a . As is arbitrary,
lim
t0
sup
at
[
_
a
u[ = 0. Moreover, whenever t > 0 and a t, then
1
t
[
_
a
u[ +
1
t
_
([u[ 1)
+
. Thus
limsup
t
1
t
sup
at
[
_
a
u[ < ,
and the function a
_
a
u satises the conditions in ().
(iii) u is uniquely dened because u a must be u
a
, as dened in (i), for every a A
f
, and (A, ) is
semi-nite.
373I Lemma Suppose that u, v, w M
0,

L
are all equivalence classes of non-negative non-increasing
functions. If
_
t
0
u
_
t
0
v for every t 0, then
_
u w
_
v w.
proof For n N, i 4
n
set a
ni
= [[w > 2
n
i]]; set w
n
=

4
n
i=1
2
n
a
ni
. Then each a
ni
is of the form [0, t]

,
so
_
u w
n
=

4
n
i=1
2
n
_
a
ni
u

4
n
i=1
2
n
_
a
ni
v =
_
v w
n
.
Also w
n
)
nN
is a non-decreasing sequence with supremum w, so
_
u w = sup
nN
_
u w
n
sup
nN
_
v w
n
=
_
v w.
373L Decreasing rearrangements 429
373J Corollary Suppose that (A, ) and (B, ) are measure algebras and v M
0,

. If
either u M
1,0

and T T
(0)
,
or u M
1,

and T T
,
then
_
[Tu v[
_
u

.
proof Put 373E, 373G and 373I together.
373K The very weak operator topology of T Let (A, ) and (B, ) be two measure algebras. For
u M
1,

, w M
,1

set

uw
(S, T) = [
_
Su w
_
Tu w[ for all S, T T = T
,
.
Then
uw
is a pseudometric on T . I will call the topology generated by
uw
: u M
1,

, w M
,1


(2A3F) the very weak operator topology on T .
373L Theorem Let (A, ) be a measure algebra and (B, ) a localizable measure algebra. Then
T = T
,
is compact in its very weak operator topology.
proof Let T be an ultralter on T . If u M
1,

, w M
,1

then
[
_
Tu w[
_
u

<
for every T T (373J);
_
u

is nite because u

M
1,
and w

M
,1
(373F).
In particular,
_
Tu w : T T is bounded. Consequently h
u
(w) = lim
TJ
_
Tu w is dened in R
(2A3Se). Because w
_
Tu w is additive for every T T , so is h
u
. Also
[h
u
(w)[
_
u

|u

|
1,
|w

|
,1
= |u|
1,
|w|
,1
for every w M
,1

.
[h
u
(b)[
_
t
0
u

whenever b B
f
and b t. So
lim
t0
sup
bt
[h
u
(b)[ lim
t0
_
t
0
u

= 0,
limsup
t
1
t
sup
bt
[h
u
(b)[ limsup
t
1
t
_
t
0
u

< .
Of course b h
u
(b) is additive, so by 373H there is a unique Su M
1,

such that h
u
(b) =
_
b
Su for
every b B
f
. Since both h
u
and w
_
Su w are linear and continuous on M
,1

, and S(B
f
) is dense in
M
,1

(369Od),
_
Su w = h
u
(w) = lim
TJ
_
Tu w
for every w M
,1

. And this is true for every u M
1,

.
For any particular w M
,1

, all the maps u
_
Tu w are linear, so u
_
Su w also is; that is,
S : M
1,

M
1,

is linear.
Now S T . PPP () If u L
1

and b, c B
f
, then
_
b
Su
_
c
Su = lim
TJ
_
Tu (b c) sup
TJ
_
Tu (b c)
sup
TJ
|Tu|
1
|b c|

|u|
1
.
But, setting e = [[Su > 0]], we have
_
[Su[ =
_
e
Su
_
1\e
Su
= sup
bB
f
,be
_
b
Su + sup
cB
f
,c1\e
(Su) |u|
1
.
430 Linear operators between function spaces 373L
() If u L

(A), then
[
_
b
Su[ sup
TJ
[
_
Tu b[ sup
TJ
|Tu|

b |u|

b
for every b B
f
. So [[Su > |u|

]] = [[Su > |u|

]] = 0 and |Su|

|u|

. (Note that both parts of

this argument depend on knowing that (B, ) is semi-nite, so that we cannot be troubled by purely innite
elements of B.) QQQ
Of course we now have lim
TJ

uw
(T, S) = 0 for all u M
1,0

, w M
,1

, so that S = limT in T . As
T is arbitrary, T is compact (2A3R).
373M Corollary Let (A, ) be a measure algebra and (B, ) a localizable measure algebra, and u any
member of M
1,

. Then B = Tu : T T
,
is compact in M
1,

for the topology T
s
(M
1,

, M
,1

).
proof The point is just that the map T Tu : T
,
M
1,0

is continuous for the very weak operator
topology on T
,
and T
s
(M
1,

, M
,1

). So B is a continuous image of a compact set, therefore compact
(2A3Nb).
373N Corollary Let (A, ) be a measure algebra, (B, ) a localizable measure algebra and u any
member of M
1,

; set B = Tu : T T
,
. If v
n
)
nN
is any non-decreasing sequence in B, then sup
nN
v
n
is dened in M
1,

and belongs to B.
proof By 373M, v
n
)
nN
must have a cluster point v B for T
s
(M
1,

, M
,1

). Now for any b B
f
,
_
b
v
must be a cluster point of
_
n
v
n
)
nN
, because w
_
b
w is continuous for T
s
(M
1,

, M
,1

). But
_
b
v
n
)
nN
is a non-decreasing sequence, so its only possible cluster point is its supremum; thus
_
b
v = lim
n
_
b
v
n
.
Consequently v b must be the supremum of v
n
b : n N in L
1
. And this is true for every b B
f
;
as (B, ) is semi-nite, v is the supremum of v
n
)
nN
in L
0
(B) and in M
1,

.
373O Theorem Let (A, ), (B, ) be measure algebras and u M
1,

, v M
1,

. Then the following
are equiveridical:
(i) there is a T T
,
such that Tu = v,
(ii)
_
t
0
v

_
t
0
u

for every t 0.
In particular, given u M
1,

, there are S T
,
L
, T T

L
,
such that Su = u

, Tu

= u.
proof (i)(ii) is Lemma 373G. Accordingly I shall devote the rest of the proof to showing that (ii)(i).
(a) If (A, ), (B, ) are measure algebras and u M
1,

, v M
1,

, I will say that v u if there is
a T T
,
such that Tu = v, and that v u if v u and u v. (Properly speaking, I ought to write
(u, ) (v, ), because we could in principle have two dierent measures on the same algebra. But I do not
think any confusion is likely to arise in the argument which follows.) By 373Be, is transitive and is an
equivalence relation. Now we have the following facts.
(b) If (A, ) is a measure algebra and u
1
, u
2
M
1,

are such that [u
1
[ [u
2
[, then u
1
u
2
. PPP There
is a w L

(A) such that u

1
= w u
2
and |w|

1. Set Tv = w v for for v M

1,

; then T T
,
and Tu
2
= u
1
. QQQ So u [u[ for every u M
1,

.
(c) If (A, ) is a measure algebra and u 0 in S(A), then u u

. PPP If u = 0 this is trivial. Otherwise,

express u as

n
i=0

i
a
i
where a
0
, . . . , a
n
are disjoint and non-zero and
0
>
1
. . . >
n
> 0 R. If
a
i
= for any i, take m to be minimal subject to a
m
= ; otherwise, set m = n. Then u

m
i=0

i
[
i
,
i+1
[

, where
0
= 0,
j
=

j1
i=0
a
i
for 1 j m+ 1.
For i < m, and for i = m if a
m
< , dene h
i
: M
1,

R by setting
h
i
(v) =
1
a
i
_
a
i
v
for every v M
1,

. If a
m
= , then we need a dierent idea to dene h
m
, as follows. Let I be
a : a A, (aa
m
) < . Then I is an ideal of A not containing a
m
, so there is a Boolean homomorphism
: A 0, 1 such that a = 0 for a I and a
m
= 1 (311D). This induces a corresponding | |

-continuous
373O Decreasing rearrangements 431
linear operator h : L

(A) L

(0, 1)

= R, as in 363F. Now h(a) = 0 whenever a < , and accordingly
h(v) = 0 whenever v M
,1

, since S(A
f
) is dense in M
,1

for | |
,1
and therefore also for | |

. But this
means that h has a unique extension to a linear functional h
m
: M
1,

R such that h
m
(v) = 0 for every
v L
1

, while h
m
(a
m
) = 1 and [h(v)[ |v|

for every v L

(A).
Having dened h
i
for every i m, dene T : M
1,

M
1,

L
by setting
Tv =

m
i=0
h
i
(v)[
i
,
i+1
[

for every v M
1,

.
For any i m, v L
1

,
_

i+1

i
[Tv[ = [h
i
(v)[ a
i

_
a
i
[v[;
summing over i, |Tv|
1
|v|
1
. Similarly, for any i m, v L

(B), [h
i
(v)[ |v|

, so |Tv|

|v|

.
Thus T T
,
L
. Since u

= Tu, we conclude that u

u, as claimed. QQQ
(d) If (A, ) is a measure algebra and u 0 in M
1,

, then u

u. PPP Let u
n
)
nN
be a non-decreasing
sequence in S(A) with u
0
0 and sup
nN
u
n
= u. Then u

n
)
nN
is a non-decreasing sequence in M
1,

L
with
supremum u

, by 373Db and 373Dh. Also u

n
u
n
u for every n, by (b) and (c) of this proof. By 373N,
u

u. QQQ
(e) If (A, ) is a measure algebra and u 0 in S(A), then u u

. PPP The argument is very similar to

that of (c). Again, the result is trivial if u = 0; suppose that u > 0 and dene
i
, a
i
, m,
i
as before. This
time, set a
t
i
= a
i
for i < m, a
t
m
= sup
mjn
a
j
, u =

m
i=0

i
a
t
i
; then u u amd u

= u

. Set
h
i
(v) =
1

i+1

i
_

i+1

i
v
if i m,
i+1
< (that is, a
i
< ) and v M
1,

L
; and if a
m
= , set
h
m
(v) = lim
kJ
1
k
_
k
0
v
for some non-principal ultralter T on N. As before, we have
[h
i
(v)[ a
t
i

_

i+1

i
[v[,
whenever v L
1

L
, i m, while [h
i
(v)[ |v|

whenever v L

(A
L
) and i m. So we can dene
T T

L
,
by setting Tv =

m
i=0
h
i
(v)a
t
i
for every v M
1,

L
, and get
u u = Tu

. QQQ
(f ) If (A, ) is a measure algebra and u 0 in M
1,

, then u u

. PPP This time I seek to copy the ideas

of (d); there is a new obstacle to circumvent, since (A, ) might not be localizable. Set

0
= inf : 0, [[u > ]] < , e = [[u >
0
]].
Then e = sup
nN
[[u >
0
+ 2
n
]] is a countable supremum of elements of nite measure, so the principal
ideal A
e
, with its induced measure
e
, is -nite. Now let u
n
)
nN
be a non-decreasing sequence in S(A)
with u
0
0 and sup
nN
u
n
= u; set u = u e and u
n
= u
n
e, regarded as members of S(A
e
), for each
n. In this case
u
n
u

n
u

for every n. Because (A

e
,
e
) is -nite, therefore localizable, 373N tells us that u u

.
Let S T

L
,
e
be such that Su

= u. As in part (e), choose a non-principal ultralter T on N and set

h(v) = lim
kJ
1
k
_
k
0
v
for v M
1,

L
. Now dene T : M
1,

L
M
1,

by setting
Tv = Sv +h(v)(1 \ e),
432 Linear operators between function spaces 373O
here regarding Sv as a member of M
1,

. (I am taking it to be obvious that M
1,

e
can be identied with
w e : w M
1,

.) Then it is easy to see that T T

L
,
. Also u Tu

, because
h(u

) = inf :
L
[[u

> ]] < =
0
,
while u (1 \ e)
0
(1 \ e). So we get u Tu

. QQQ
(g) Now suppose that u, v 0 in M
1,

L
, that
_
t
0
u

_
t
0
v

for every t 0, and that v is of the form

n
i=1

i
a
i
where
1
> . . . >
n
> 0, a
1
, . . . , a
n
A
L
are disjoint and
L
a
i
< for each i. Then v u.
PPP Induce on n. If n = 0 then v = 0 and the result is trivial. For the inductive step to n 1, if v

we
have
v v

u,
using (b), (d) and (f) above. Otherwise, look at (t) =
1
t
_
t
0
u

for t > 0. We have

(t)
1
t
_
t
0
v

=
1
for t = a
1
, while lim
t
(t) <
1
, because (lim
t
(t))1 u

and v

1
1 and v

, u

.
Becaasue is continuoue, there is a such that () =
1
. Dene T
0
T

L
,
L
by setting
T
0
w = (
1

0
w)[0, [

+ (w [, [

)
for every w M
1,

L
. Then T
0
u

u, and
T
0
u

[0, [

= (
1

0
u

)[0, [

=
1
[0, [

.
We need to know that
_
t
0
T
0
u

_
t
0
v

for every t; this is because

_
t
0
T
0
u

=
1
t
_
t
0
v

whenever t ,
=
_

0
T
0
u

+
_
t

T
0
u

=
_
t
0
u

_
t
0
v

whenever t .
Set
u
1
= T
0
u

[, [

, v
1
= v

[, [

.
Then u

1
, v

1
are just translations of T
0
u

, v

to the left, so that

_
t
0
u

1
=
_
+t

T
0
u

=
_
+t
0
T
0
u

1

_
+t
0
v

1
=
_
+t

=
_
t
0
v

1
for every t 0. Also v
1
=

n
i=2

i
[
i1
,
i
[

where
i
=

i
j=1
a
j
for each j. So by the inductive
hypothesis, v
1
u
1
.
Let S T

L
,
L
be such that Su
1
= v
1
, and dene T T

L
,
L
by setting
Tw = w [0, [

+S(w [, [

) [, [

for every w M
1,

L
,
L
. Then TT
0
u

= v

, so v v

u, as required. QQQ
(h) We are nearly home. If u, v 0 in M
1,

L
and
_
t
0
v

_
t
0
u

for every t 0, then v u. PPP Let

v
n
)
nN
be a non-decreasing sequence in S(A
f
L
)
+
with supremum v. Then v

n
v

for each n, so (g) tells

us that v
n
u for every n. By 373N, for the last time, v u. QQQ
(i) Finally, suppose that (A, ) and (B, ) are arbitrary measure algebras and that u M
1,

, v M
1,

are such that
_
t
0
v

_
t
0
u

for every t 0. By (b), v [v[; by (f), [v[ [v[

; by 373Db, [v[

= v

; by (h)
of this proof, v

; by (d), u

= [u[

[u[; and by (b) again, [u[ u.

373P Theorem Let (A, ) be a measure algebra and (B, ) a semi-nite measure algebra. Then for
any u M
1,

and v M
0

, there is a T T = T
,
such that
_
Tu v =
_
u

.
373P Decreasing rearrangements 433
proof (a) It is convenient to dispose immediately of some elementary questions.
(i) We need only nd a T T such that
_
[Tu v[
_
u

. PPP Take v
0
L

(B) such that

[Tu v[ = v
0
Tu v and |v
0
|

1, and set T
1
w = v
0
Tw for w M
1,

; then T
1
T and
_
T
1
u v =
_
[Tu v[
_
u

_
T
1
u v
by 373J. QQQ
(ii) Consequently it will be enough to consider v 0, since of course
_
[Tu v[ =
_
[Tu [v[[, while
[v[

= v

.
(iii) It will be enough to consider u = u

. PPP If we can nd T T

L
,
such that
_
Tu

v =
_
(u

,
then we know from 373O that there is an S T
,
L
such that Su = u

, so that TS T and
_
TSu v =
_
(u

=
_
u

. QQQ
(iv) It will be enough to consider localizable (B, ). PPP Assuming that v 0, following (ii) above, set
e = [[v > 0]] = sup
nN
[[v > 2
n
]], and let
e
be the restriction of to the principal ideal B
e
generated by
e. Then if we write v for the member of L
0
(B
e
) corresponding to v (so that [[ v > ]] = [[v > ]] for every
> 0), v

= v

. Also (B
e
,
e
) is -nite, therefore localizable. Now if we can nd T T
,
e
such that
_
Tu v =
_
u

, then ST will belong to T

,
, where S : L
0
(B
e
) L
0
(B) is the canonical embedding
dened by the formula
[[Sw > ]] = [[w > ]] if 0,
= [[w > ]] (1 \ e) if < 0,
and
_
STu v =
_
Tu v =
_
u

=
_
u

. QQQ
(b) So let us suppose henceforth that =
L
, u = u

is the equivalence class of a non-increasing

non-negative function, v 0 and (B, ) is localizable.
For n, i N set
b
ni
= [[v > 2
n
i]],
ni
= b
ni
, c
ni
= b
ni
\ b
n,i+1
,
ni
= c
ni
=
ni

n,i+1
(because
ni
< if i > 0; this is really where I use the hypothesis that v M
0
). For n N set
K
n
= i : i 1,
ni
> 0,
T
n
w =

iK
n
_
1

ni
_

ni

n,i+1
w
_
c
ni
for w M
1,

L
; this is dened in L
0
(B) because K
n
is countable and c
ni
)
iN
is disjoint. Of course T
n
:
M
1,

L
L
0
(B) is linear. If w L

(A
L
) then
|T
n
w|

= sup
iK
n

ni
_

ni

n,i+1
w

|w|

,
and if w L
1

L
then
|T
n
w|
1
=

iK
n

ni
_

ni

n,i+1
w

c
ni
=

iK
n

ni

n,i+1
w

|w|
1
;
so T
n
w M
1,

for every w M
1,

L
, and T
n
T . It will be helpful to observe that
_
c
ni
T
n
w =
_

ni

n,i+1
w
whenever i 1, since if i / K
n
then both sides are 0.
Note next that for every n, i N,
b
ni
= b
n+1,2i
,
ni
=
n+1,2i
, c
ni
= c
n+1,2i
c
n+1,2i+1
,
ni
=
n+1,2i
+
n+1,2i+1
,
434 Linear operators between function spaces 373P
so that, for i 1,
_
c
ni
T
n
u =
_

ni

n,i+1
u =
_
c
ni
T
n+1
u.
This means that if T is any cluster point of T
n
)
nN
in T for the very weak operator topology (and such
a cluster point exists, by 373L),
_
c
mi
Tu must be a cluster point of
_
c
mi
T
n
u)
nN
, and therefore equal to
_
c
mi
T
m
u, for every m N, i 1.
Consequently, if m N,
_
[Tu v[

i=0
_
c
mi
[Tu[ v

i=0
2
m
i
_
c
mi
[Tu[
(because c
mi
[[v > 2
m
i]])

i=1
2
m
i[
_
c
mi
Tu[ =

i=1
2
m
i
_
c
mi
T
m
u
=

i=0
2
m
i
_

mi

m,i+1
u
_
u (v

2
m
1)
+
because
[
m,i+1
,
mi
]

[[v

2
m
(i + 1)]] = [[(v

2
m
1)
+
2
m
i]]
for each i N. But letting m , we have
_
[Tu v[ lim
m
_
u (v

2
m
1)
+
=
_
u v

because u (v

2
m
1)
+
)
mN
is a non-decreasing sequence with supremum u v

. In view of the
reductions in (a) above, this is enough to complete the proof.
373Q Corollary Let (A, ) be a measure algebra, (B, ) a semi-nite measure algebra, u M
1,

and
v M
0,

. Then
_
u

= sup
_
[Tu v[ : T T
,
= sup
_
Tu v : T T
,
.
proof There is a non-decreasing sequence c
n
)
nN
in B such that c
n
< for every n and v

= sup
nN
(v
c
n
)

. PPP For each rational q > 0, we can nd a countable non-empty set B

q
B such that
b [[[v[ > q]], b < for every b B
q
,
sup
bB
q
b = [[[v[ > q]]
(because (B, ) is semi-nite). Let b
n
)
nN
be a sequence running over

qQ,q>0
B
q
and set c
n
= sup
in
b
i
,
v
n
= v c
n
for each n. Then [v
n
[)
nN
and v

n
)
nN
are non-decreasing and sup
nN
v

n
v

in L
0
(A
L
).
But in fact sup
nN
v

n
= v

, because

L
[[v

> q]] = [[[v[ > q]] = sup

nN
[[v
n
> q]] = sup
nN

L
[[v

n
> q]] =
L
[[sup
nN
v

n
> q]]
for every rational q > 0, by 373Da. QQQ
For each n N we have a T
n
T
,
such that
_
T
n
u v
n
=
_
u

n
(373P). Set S
n
w = T
n
w c
n
for
n N, w M
1,

; then every S
n
belongs to T
,
, so
sup
_
Tu v : T T
,
sup
nN
_
S
n
u v = sup
nN
_
T
n
u v
n
= sup
nN
_
u

n
=
_
u

sup
_
[Tu v[ : T T
,
sup
_
Tu v : T T
,

by 373J, as usual.
373S Decreasing rearrangements 435
373R Order-continuous operators: Proposition Let (A, ) be a measure algebra, (B, ) a local-
izable measure algebra, and T
0
T
(0)
= T
(0)
,
. Then there is a T T

= T

,
extending T
0
. If (A, ) is
semi-nite, T is uniquely dened.
proof (a) Suppose rst that T
0
T
(0)
is non-negative, regarded as a member of L

(M
1,0

; M
1,0

). In this
case T
0
has an extension to an order-continuous positive linear operator T : M
1,

L
0
(B) dened by
saying that Tw = supT
0
u : u M
1,0

, 0 u w for every w 0 in M
1,

. PPP I use 355F. M
1,0

is a solid
linear subspace of M
1,

. T
0
is order-continuous when its codomain is taken to be M
1,0

, as noted in 371Gb,
and therefore if its codomain is taken to be L
0
(B), because M
1,0
is a solid linear subspace in L
0
, so the
embedding is order-continuous. If w 0 in M
1,

, let 0 be such that u
1
= (w 1)
+
is integrable. If
u M
1,0

and 0 u w, then (u 1)
+
u
1
, so
T
0
u = T
0
(u 1)
+
+T
0
(u 1) T
0
u
1
+1 L
0
(B).
Thus T
0
u : u M
1,0

, 0 u w is bounded above in L
0
(B), for any w 0 in M
1,

. L
0
(B) is Dedekind
complete, because (B, ) is localizable, so supT
0
u : 0 u w is dened in L
0
(B); and this is true for
every w (M
1,

)
+
. Thus the conditions of 355F are satised and we have the result. QQQ
(b) Now suppose that T
0
is any member of T
(0)
. Then T
0
has an extension to a member of T

. PPP [T
0
[,
T
+
0
=
1
2
([T
0
[ + T
0
) and T

0
=
1
2
([T
0
[ T
0
), taken in L

(M
1,0

; M
1,0

), all belong to T
(0)
(371G), so have
extensions S, S
1
and S
2
to order-continuous positive linear operators from M
1,

to L
0
(B) as dened in (a).
Now for any w L
1

,
|Sw|
1
= |[T
0
[w|
1
|w|
1
,
and for any w L

(A),
[Sw[ S[w[ = sup[T
0
[u : u M
1,0

, 0 u w |w|

1,
so |Sw|

|w|

. Thus S T ; similarly, S
1
and S
2
can be regarded as operators from M
1,

to M
1,

,
and as such belong to T . Next, for w 0 in M
1,

,
S
1
w +S
2
w = supT
+
0
u : u M
1,0

, 0 u w + supT

0
u : u M
1,0

, 0 u w
= supT
+
0
u +T

0
u : u M
1,0

, 0 u w = Sw.
But this means that
S = S
1
+S
2
[S
1
S
2
[
and T = S
1
S
2
T , by 373Bc; while of course T extends T
+
0
T

0
= T
0
. Finally, because S
1
and S
2
are
order-continuous, T L

(M
1,

; M
1,

), so T T

. QQQ
(c) If (A, ) is semi-nite, then M
1,0

is order-dense in M
1,

(because it includes L
1

, which is order-dense
in L
0
(A)); so that the extension T is unique, by 355F(iii).
373S Adjoints in T
(0)
: Theorem Let (A, ) and (B, ) be measure algebras, and T any member of
T
(0)
,
. Then there is a unique operator T
t
T
(0)
,
such that
_
a
T
t
(b) =
_
b
T(a) for every a A
f
, b B
f
,
and now
_
u T
t
v =
_
Tu v whenever u M
1,0

, v M
1,0

are such that
_
u

< .
proof (a) For each v M
1,0

we can dene T
t
v M
1,0

by the formula
_
a
T
t
v =
_
T(a) v
for every a A
f
. PPP Set a =
_
T(a) v for each a A
f
; because
_
(a)

< , the integral is dened

and nite (373J). Of course : A
f
R is additive because is additive and T, and
_
are linear. Also
lim
t0
sup
at
[a[ lim
t0
_
t
0
v

= 0,
lim
t
1
t
sup
at
[a[ lim
t
1
t
_
t
0
v

= 0
436 Linear operators between function spaces 373S
because v M
1,0

, so v

M
1,0

L
. By 373Ha, there is a unique T
t
v M
1,0

such that
_
a
T
t
v = a for every
a A
f
. QQQ
(b) Because the formula uniquely determines T
t
v, we see that T
t
: M
1,0

M
1,0

is linear. Now T
t
T
(0)
,
.
PPP (i) If v L
1

, then (because T
t
v M
1,0

) [T
t
v[ = sup
aA
f [T
t
v[ a, and
|T
t
v|
1
=
_
[T
t
v[ = sup
aA
f
_
a
[T
t
v[ = sup
b,cA
f
(
_
b
T
t
v
_
c
T
t
v)
= sup
b,cA
f
_
T(b c) v sup
b,cA
f
_
(b c)

=
_
v

= |v|
1
.
(ii) Now suppose that v L

(B) M
1,0

, and set = |v|

. ??? If a = [[[T
t
v[ > ]] ,= 0, then T
t
v ,= 0 so
v ,= 0 and > 0 and a < , because T
t
v M
1,0

. Set b = [[(T
t
v)
+
> ]], c = [[(T
t
v)

> ]]; then

a <
_
a
[T
t
v[ =
_
b
T
t
v
_
c
T
t
v =
_
T(b c) v
|T(b c)|
1
|b c|
1
= a,
which is impossible. XXX Thus [[[T
t
v[ > ]] = 0 and |T
t
v|

= |v|

.
Putting this together with (i), we see that T
t
T
(0)
,
. QQQ
(c) Let [T[ be the modulus of T in L

(M
1,0

; M
1,0

), so that [T[ T
(0)
,
, by 371Gb. If u 0 in M
1,0

, v 0
in M
1,0

are such that
_
u

< , let u
n
)
nN
be a non-decreasing sequence in S(A
f
)
+
with supremum
u. In this case [T[u = sup
nN
[T[u
n
, so
_
[T[u v = sup
nN
_
[T[u
n
v and
[
_
Tu v
_
Tu
n
v[
_
[T[(u u
n
) v 0
as n , because
_
[T[u v
_
u

< .
At the same time,
[
_
u T
t
v
_
u
n
T
t
v[
_
(u u
n
) [T
t
v[ 0
because
_
u [T
t
v[
_
u

< . So
_
Tu v = lim
n
_
Tu
n
v = lim
n
_
u
n
T
t
v =
_
u T
t
v,
the middle equality being valid because each u
n
is a linear combination of characteristic functions.
Because T and T
t
are linear, it follows at once that
_
u T
t
v =
_
Tu v whenever u M
1,0

, v M
1,0

are such that
_
u

< .
(d) Finally, to see that T
t
is uniquely dened by the formula in the statement of the theorem, observe
that this surely denes T
t
(b) for every b B
f
, by the remarks in (a). Consequently it denes T
t
on S(B
f
).
Since S(B
f
) is order-dense in M
1,0

, and any member of T
(0)
,
must belong to L

(M
1,0

; M
1,0

) (371Gb), the
restriction of T
t
to S(B
f
) determines T
t
(355J).
373T Corollary Let (A, ) and (B, ) be localizable measure algebras. Then for any T T

,
there is a
unique T
t
T

,
such that
_
uT
t
v =
_
Tuv whenever u M
1,

, v M
1,

are such that
_
u

< .
proof The restriction TM
1,0

belongs to T
(0)
,
(373Bb), so there is a unique S T
(0)
,
such that
_
u Sv =
_
Tuv whenever u M
1,0

, v M
1,0

are such that
_
u

< (373S). Now there is a unique T

t
T

,
extending S (373R). If u 0 in M
1,

, v 0 in M
1,

are such that
_
u

< , then
_
uT
t
v =
_
Tuv.
PPP If T 0, then both are
sup
_
u
0
T
t
v
0
: u
0
M
1,0

, v M
1,0

, 0 u
0
u, 0 v
0
v
373Xi Decreasing rearrangements 437
because both T and T
t
are (order-)continuous. In general, we can apply the same argument to T
+
and
T

, taken in L

(M
1,

; M
1,

), since these belong to T

,
, by 373B and 355H, and we shall surely have
T
t
= (T
+
)
t
(T

)
t
. QQQ As in 373S, it follows that
_
u T
t
v =
_
Tu v whenever u M
1,

, v M
1,

are
such that
_
u

< .
373U Corollary Let (A, ) and (B, ) be localizable measure algebras, and : A B an order-
continuous measure-preserving Boolean homomorphism. Then the associated map T T

,
(373Bd) has an
adjoint P T

,
dened by the formula
_
a
P(b) = (b a) for a A
f
, b B
f
.
proof The adjoint P = T
t
must have the property that
_
a
P(b) =
_
a P(b) =
_
T(a) b =
_
(a) b = (a b)
for every a A
f
, b B
f
. To see that this denes P uniquely, let S T

,
be any other operator with the
same property. By 373Hb, S(b) = P(b) for every b B
f
, so S and P agree on S(B
f
). Because both P
and S are supposed to belong to L

(M
1,

; M
1,

), and S(B
f
) is order-dense in M
1,

, S = P, by 355J.
373X Basic exercises (a) Let (A, ) and (B, ) be measure algebras, and : A B a ring ho-
momorphism such that a a for every a A. (i) Show that there is a unique T T
,
such that
T(a) = (a) for every a A, and that T is a Riesz homomorphism. (ii) Show that T is (sequentially)
order-continuous i is.
>>>(b) Let (A, ) and (B, ) be measure algebras, and : R R a convex function such that (0) 0.
Show that if T T
,
and T 0, then

(Tu) T(

(u)) whenever u M
1,

is such that

(u) M
1,

.
(Hint: 233J, 365Rb.)
(c) Let (A, ) be a measure algebra. Show that if w L

(A) and |w|

1 then u u w : M
1,


M
1,

belongs to T

,
.
(d) Let (A, ) and (B, ) be measure algebras. Show that if a
i
)
iI
, b
i
)
iI
are disjoint families in A, B
respectively, and T
i
)
iI
is any family in T
,
, and either I is countable or B is Dedekind complete, then
we have an operator T T
,
such that Tu b
i
= T
i
(u a
i
) b
i
for every u M
1,
,
, i I.
>>>(e) Let I, J be sets and write = , = for counting measure on I, J respectively. Show that
there is a natural one-to-one correspondence between T

,
and the set of matrices a
ij
)
iI,jJ
such that

iI
[a
ij
[ 1 for every j J,

jJ
[a
ij
[ 1 for every i I.
>>>(f ) Let (X, , ) and (Y, T, ) be -nite measure spaces, with measure algebras (A, ) and (B, ), and
product measure on X Y . Let h : X Y R be a measurable function such that
_
[h(x, y)[dx 1 for
-almost every y Y and
_
[h(x, y)[dy 1 for -almost every x X. Show that there is a corresponding
T T

,
dened by writing T(f

) = g

whenever f L
1
() +L

() and g(y) =
_
h(x, y)f(x)dx for almost
every y.
>>>(g) Let be Lebesgue measure on R, and (A, ) its measure algebra. Show that for any -integrable
function h with
_
[h[d 1 we have a corresponding T T

,
dened by setting T(f

) = (h f)

whenever
g L
1
() +L

(), writing h f for the convolution of h and f (255E). Explain how this may be regarded
as a special case of 373Xf.
>>>(h) Let (, , ) be a probability space and X L
0
() a non-negative real-valued random variable on
; let
X
be its distribution (271C). Write u = X

L
0
()

= L
0

, where (A, ) is the measure algebra of
(, , ). Show that
L
[[u

> ]] = Pr(X > ) for every , so that each of u

,
X
is uniquely determined
by the other.
(i) Let (A, ) and (B, ) be measure algebras, and : A B a measure-preserving Boolean homomor-
phism; let T : M
1,

M
1,

be the corresponding operator (373Bd). Show that (Tu)

= u

for every
u M
1,

.
438 Linear operators between function spaces 373Xj
(j) Let (A, ) be a totally nite measure algebra, and A a subset of L
1

. Show that the following
are equiveridical: (i) A is uniformly integrable; (ii) u

: u A is uniformly integrable in L
1

L
; (iii)
lim
t0
sup
uA
_
t
0
u

= 0.
(l) Let (A, ) be a measure algebra, and A (M
0

)
+
a non-empty downwards-directed set. Show that
(inf A)

= inf
uA
u

in L
0
(A
L
).
(k) Let (A, ) be a measure algebra. Show that |u|
1,
=
_
1
0
u

for every u M
1,
(A, ).
(m) Let (A, ) and (B, ) be measure algebras, and a Youngs function (369Xc). Write U
,
L
0
(A),
U
,
L
0
(B) for the corresponding Orlicz spaces. (i) Show that if T T
,
and u U
,
, then Tu U
,
and |Tu|

|u|

. (ii) Show that u U

,
i u

U
,
L
, and in this case |u|

= |u

.
>>>(n) Let (A, ) be a measure algebra and (B, ) a totally nite measure algebra. Show that if A L
1

is uniformly integrable, then Tu : u A, T T
,
is uniformly integrable in L
1

.
(o) (i) Give examples of u, v L
1
(A
L
) such that (u + v)

, u

+ v

. (ii) Show that if (A, ) is any

measure algebra and u, v M
0,

, then
_
t
0
(u +v)

_
t
0
u

+v

for every t 0.
(p) Let (A, ) and (B, ) be two measure algebras. For u M
1,0

, w M
,1

set

uw
(S, T) = [
_
(Su Tu) w[ for all S, T T
(0)
= T
(0)
,
.
The topology generated by the pseudometrics
uw
is the very weak operator topology on T
(0)
. Show
that T
(0)
is compact in this topology.
(q) Let (A, ) and (B, ) be measure algebras and let u M
1,0

. (i) Show that B = Tu : T T
(0)
,
is
compact for the topology T
s
(M
1,0

, M
,1

). (ii) Show that any non-decreasing sequence in B has a supremum
in L
0
(B) which belongs to B.
(r) Let (A, ) and (B, ) be measure algebras, and u M
1,0

, v M
1,0

. Show that the following are
equiveridical: (i) there is a T T
(0)
,
such that Tu = v; (ii)
_
t
0
u

_
t
0
v

for every t 0.
(s) Let (A, ) and (B, ) be measure algebras. Suppose that u
1
, u
2
M
1,

and v M
1,

are such
that
_
t
0
v

_
t
0
(u
1
+ u
2
)

for every t 0. Show that there are v

1
, v
2
M
1,

such that v
1
+ v
2
= v and
_
t
0
v

i

_
t
0
u

i
for both i, every t 0.
>>>(t) Set g(t) = t/(t + 1) for t 0, and set v = g

, u = [0, 1]

(A
L
). Show that
_
u

= 1 >
_
Tu v for every T T

L
,
L
.
(u) Let (A, ) and (B, ) be measure algebras, and for T T
(0)
,
dene T
t
T
(0)
,
as in 373S. Show that
T
tt
= T.
(v) Let (A, ) and (B, ) be measure algebras, and give T
(0)
,
, T
(0)
,
their very weak operator topologies
(373Xp). Show that the map T T
t
: T
(0)
,
T
(0)
,
is an isomorphism for the convex, order and topological
structures of the two spaces. (By the convex structure of a convex set C in a linear space I mean the
operation (x, y, t) tx + (1 t)y : C C [0, 1] C.)
373Y Further exercises (a) Let (A, ) be the measure algebra of Lebesgue measure on [0, 1]. Set
u = f

and v = g

in L
0
(A), where f(t) = t, g(t) = 1 2[t
1
2
[ for t [0, 1]. Show that u

= v

,
but that there is no measure-preserving Boolean homomorphism : A A such that T

v = u, writing
T

: L
0
(A) L
0
(A) for the operator induced by , as in 364R. (Hint: show that [[v > ]] : R does
not -generate A.)
373 Notes Decreasing rearrangements 439
(b) Let (A, ) be a totally nite homogeneous measure algebra of uncountable Maharam type. Let u,
v (M
1,

)
+
be such that u

= v

. Show that there is a measure-preserving automorphism : A A such

that T

u = v.
(c) Let u, v M
1,

L
be such that u = u

, v = v

and
_
t
0
v
_
t
0
u for every t 0. Show that there is
a non-negative T T

L
,
L
such that Tu = v and
_
t
0
Tw
_
t
0
w for every w 0 in M
1,
. Show that any
such T must belong to T

L
,
L
.
(d) Let (A, ) and (B, ) be measure algebras, and u M
1,

. (i) Suppose that w S(B
f
). Show
directly (without quoting the result of 373O, but possibly using some of the ideas of the proof) that for
every <
_
u

there is a T T
,
such that
_
Tuw . (ii) Suppose that (B, ) is localizable and that
v M
1,

Tu : T T
,
. Show that there is a w S(B
f
) such that
_
v w > sup
_
Tuw : T T
,
.
(Hint: use 373M and the Hahn-Banach theorem in the following form: if U is a linear space with the
topology T
s
(U, V ) dened by a linear subspace V of L(U; R), C U is a non-empty closed convex set, and
v U C, then there is an f V such that f(v) > sup
uC
f(u).) (iii) Hence prove 373O for localizable
(B, ). (iv) Now prove 373O for general (B, ).
(e) (i) Dene v L

(A
L
) as in 373Xt. Show that there is no T T

L
,
L
such that Tv = v

. (ii) Set
h(t) = 1 +max(0,
sin t
t
) for t > 0, w = h

(A
L
). Show that there is no T T

L
,
L
such that Tw

= w.
(f ) Let (A, ) be the measure algebra of Lebesgue measure on [0, 1]. Show that T
,
L
= T

,
L
can be
identied, as convex ordered space, with T

L
,
, and that this is a proper subset of T

L
,
.
(g) Show that the adjoint operation of 373T is not as a rule continuous for the very weak operator
topologies of T

,
, T

,
.
373 Notes and comments 373A-373B are just alternative expressions of concepts already treated in 371F-
371H. My use of the simpler formula T
,
symbolizes my view that T , rather than T
(0)
or T

, is the most
natural vehicle for these ideas; I used T
(0)
in 371-372 only because that made it possible to give theorems
which applied to all measure algebras, without demanding localizability (compare 371Gb with 373Bc).
The obvious examples of operators in T are those derived from measure-preserving Boolean homomor-
phisms, as in 373Bd, and their adjoints (373U). Note that the latter include conditional expectation oper-
ators. In return, we nd that operators in T share some of the characteristic properties of the operators
derived from Boolean homomorphisms (373Bb, 373Xb, 373Xm). Other examples are multiplication op-
erators (373Xc), operators obtained by piecing others together (373Xd) and kernel operators of the type
described in 373Xe-373Xf, including convolution operators (373Xg). (For a general theory of kernel opera-
tors, see 376 below.)
Most of the section is devoted to the relationships between the classes T of operators and the decreasing
rearrangements of 373C. If you like, the decreasing rearrangement u

of u describes the distribution of [u[

(373Xh); but for u / M
0
it loses some information (373Xt, 373Ye). It is important to be conscious that
even when u L
0
(A
L
), u

is not necessarily obtained by rearranging the elements of the algebra A

L
by a
measure-preserving automorphism (which would, of course, correspond to an automorphism of the measure
space ([0, [ ,
L
), by 344C). I will treat rearrangements of this narrower type in the next section; for the
moment, see 373Ya. Apart from this, the basic properties of decreasing rearrangements are straightforward
enough (373D-373F). The only obscure area concerns the relationship between (u + v)

and u

, v

(see
373Xo).
In 373G I embark on results involving both decreasing rearrangements and operators in T , leading to the
characterization of the sets Tu : T T in 373O. In one direction this is easy, and is the content of 373G.
In the other direction it depends on a deeper analysis, and the easiest method seems to be through studying
the very weak operator topology on T (373K-373L), even though this is an eective tool only when one
of the algebras involved is localizable (373L). A functional analyst is likely to feel that the method is both
natural and illuminating; but from the point of view of a measure theorist it is not perfectly satisfactory,
because it is essentially non-constructive. While it tells us that there are operators T T acting in the
required ways, it gives only the vaguest of hints concerning what they actually look like.
440 Linear operators between function spaces 373 Notes
Of course the very weak operator topology is interesting in its own right; and see also 373Xp-373Xq.
The proof of 373O can be thought of as consisting of three steps. Given that
_
t
0
v

_
t
0
u

for every t,
then I set out to show that v is expressible as T
1
v

(parts (c)-(d) of the proof), that v

is expressible as
T
2
u

(part (g)) and that u

is expressible as T
3
u (parts (e)-(f)), each T
i
belonging to an appropriate T .
In all three steps the general case follows easily from the case of u, v S(A), S(B). If we are willing to
use a more sophisticated version of the Hahn-Banach theorem than those given in 3A5A and 363R, there is
an alternative route (373Yd). I note that the central step above, from u

to v

, can be performed with an

order-continuous T
2
(373Yc), but that in general neither of the other steps can (373Ye), so that we cannot
use T

in place of T here.
A companion result to 373O, in that it also shows that Tu : T T is large enough to reach natural
bounds, is 373P; given u and v, we can nd T such that
_
Tu v is as large as possible. In this form the
result is valid only for v M
(0)
(373Xt). But if we do not demand that the supremum should be attained,
we can deal with other v (373Q).
We already know that every operator in T
(0)
is a dierence of order-continuous operators, just because
M
1,0
has an order-continuous norm (371Gb). It is therefore not surprising that members of T
(0)
can be
extended to members of T

, at least when the codomain M

1,

is Dedekind complete (373R). It is also very
natural to look for a correspondence between T
,
and T
,
, because if T T
,
we shall surely have an
adjoint operator (TL
1

)
t
from (L
1

)

to (L
1

)

, and we can hope that this will correspond to some member

of T
,
. But when we come to the details, the normed-space properties of a general member of T are not
enough (373Yf), and we need some kind of order-continuity. For members of T
(0)
this is automatically
present (373S), and now the canonical isomorphism between T
(0)
and T

gives us an isomorphism between

T

,
and T

,
when and are localizable (373T).
374 Rearrangement-invariant spaces
As is to be expected, many of the most important function spaces of analysis are symmetric in various
ways; in particular, they share the symmetries of the underlying measure algebras. The natural expression
of this is to say that they are rearrangement-invariant (374E). In fact it turns out that in many cases
they have the stronger property of T -invariance (374A). In this section I give a brief account of the most
important properties of these two kinds of invariance. In particular, T -invariance is related to a kind of
transfer mechanism, enabling us to associate function spaces on dierent measure algebras (374B-374D).
As for rearrangement-invariance, the salient fact is that on the most important measure algebras many
rearrangement-invariant spaces are T -invariant (374K, 374M).
374A T -invariance: Denitions Let (A, ) be a measure algebra. Recall that I write
M
1,

= L
1

+L

(A) L
0
(A),
M
,1

= L
1

L

(A),
M
0,

= u : u L
0
(A), inf
>0
[[[u[ > ]] < ,
(369N, 373C).
(a) I will say that a subset A of M
1,

is T -invariant if Tu A whenever u A and T T = T
,
(denition: 373Aa).
(b) I will say that an extended Fatou norm on L
0
is T -invariant or fully symmetric if (Tu) (u)
whenever u M
1,

and T T .
(c) As in 373, I will write (A
L
,
L
) for the measure algebra of Lebesgue measure on [0, [, and u

M
0,

L
for the decreasing rearrangement of any u belonging to any M
0,

(373C).
374C Rearrangement-invariant spaces 441
374B The rst step is to show that the associate of a T -invariant norm is T -invariant.
Theorem Let (A, ) be a semi-nite measure algebra and a T -invariant extended Fatou norm on L
0
(A).
Let L

be the Banach lattice dened from (369G), and

t
the associate extended Fatou norm (369H-369I).
Then
(i) M
,1

L

M
1,

;
(ii)
t
is also T -invariant, and
_
u

(u)
t
(v) for all u, v M
0,

.
proof (a) I check rst that L

M
0,

. PPP Take any u L
0
(A) M
0,

. There is surely some w > 0 in L

,
and we can suppose that w = a for some a of nite measure. Now, for any n N,
([u[ n1)

= n1 nw

in L
0
(A
L
), because [[[u[ > n]] = . So there is a T T
,
such that T([u[ n1) = nw, by 373O, and
(u) ([u[ n1) (T([u[ n1)) = (nw) = n(w).
As n is arbitrary, (u) = . As u is arbitrary, L

M
0,

. QQQ
(b) Next,
_
u

(u)
t
(v) for every u, v M
0,

. PPP If u M
1,

, then
_
u

= sup
_
[Tu v[ : T T
,

(373Q)
sup(Tu)
t
(v) : T T
,
= (u)
t
(v).
Generally, setting u
n
= [u[ n1, u

n
)
nN
is a non-decreasing sequence with supremum u

(373Db, 373Dh),
so
_
u

= sup
nN
_
u

n
v

sup
nN
(u
n
)
t
(v) = (u)
t
(v). QQQ
(c) Consequently, L

M
1,

. PPP If A = 0, this is trivial. Otherwise, take u L

. There is surely
some non-zero a such that
t
(a) < ; now, setting v = a,
_
a
0
u

=
_
u

(u)
t
(v) <
by (b) above. But this means that u

M
1,

, so that u M
1,

(373F(b-ii)). QQQ
(d) Next,
t
is T -invariant. PPP Suppose that v M
0,

, T T
,
, u L
0
(A) and (u) 1. Then
u M
1,

, by (c), so
_
[u Tv[
_
u

(u)
t
(v)
t
(v),
using 373J for the rst inequality. Taking the supremum over u, we see that
t
(Tv)
t
(v); as T and v are
arbitrary,
t
is T -invariant. QQQ
(e) Finally, putting (d) and (c) together, L

M
1,

, so that L

M
,1

, using 369J and 369O.
374C For any T -invariant extended Fatou norm on L
0
(A
L
) there are corresponding norms on L
0
(A)
for any semi-nite measure algebra, as follows.
Theorem Let be a T -invariant extended Fatou norm on L
0
(A
L
), and (A, ) a semi-nite measure algebra.
(a) There is a T -invariant extended Fatou norm on L
0
(A) dened by setting
(u) = (u

) if u M
0,

,
= if u L
0
(A) M
0,

.
(b) Writing
t
,
t
for the associates of and , we now have
442 Linear operators between function spaces 374C

t
(v) =
t
(v

) if v M
0,

,
= if v L
0
(A) M
0,

.
(c) If is an order-continuous norm on the Banach lattice L

, then is an order-continuous norm on L

.
proof (a)(i) The argument seems to run better if I use a dierent formula to dene : set
(u) = sup
_
[u Tw[ : T T

L
,
, w L
0
(A
L
),
t
(w) 1
for u L
0
(A). (By 374B, w M
1,

whenever
t
(w) 1, so there is no diculty in dening Tw.) Now
(u) = (u

) for every u M
0,

. PPP () If w L
0
(A
L
) and
t
(w) 1, then w M
1,

L
, by 374B(i), so
there is an S T

L
,
L
such that Sw = w

(373O). Accordingly
t
(w

)
t
(w) (because
t
is T -invariant,
by 374B); now
_
[u Tw[
_
u

(u

)
t
(w

) (u

)
t
(w) (u

);
as w is arbitrary, (u) (u

). () If w L
0
(A
L
) and
t
(w) 1, then
_
[u

w[
_
(u

(373E)
=
_
u

= sup
_
[u Tw[ : T T

L
,
L

(373Q)
(u).
But because is the associate of
t
(369I(ii)), this means that (u

) (u). QQQ
(ii) Now is an extended Fatou norm on L
0
(A). PPP Of the conditions in 369F, (i)-(iv) are true just
because (u) = sup
vB
_
[u v[ for some set B L
0
. As for (v) and (vi), observe that if u M
,1

then
u

M
,1

L
(373F(b-iv)), so that (u) = (u

) < , by 374B(i), while also

u ,= 0 = u

,= 0 = (u) = (u

) > 0.
As M
,1

is order-dense in L
0
(A) (this is where I use the hypothesis that (A, ) is semi-nite), 369F(v)-(vi)
are satised, and is an extended Fatou norm. QQQ
(iii) is T -invariant. PPP Take u M
1,

and T T
,
. There are S
0
T

L
,
and S
1
T
,
L
such that
S
0
u

= u, S
1
Tu = (Tu)

(373O); now S
1
TS
0
T

L
,
L
(373Be), so
(Tu) = ((Tu)

) = (S
1
TS
0
u

) (u

) = (u)
because is T -invariant. QQQ
(iv) We can now return to the denition of . I have already remarked that (u) = (u

) if u M
0,

.
For other u, we must have (u) = just because is a T -invariant extended Fatou norm (374B(i)). So the
denitions in the statement of the theorem and (i) above coincide.
(b) We surely have
t
(v) = if v L
0
(A) M
0,

, by 374B, because
t
, like , is a T -invariant extended
Fatou norm. So take v M
0,

.
(i) If u L
0
(A) and (u) 1, then
_
[v u[
_
v

t
(v

)(u

) =
t
(v

)(u)
t
(v

);
as u is arbitrary,
t
(v)
t
(v

).
(ii) If w L
0
(A
L
) and (w) 1, then
374D Rearrangement-invariant spaces 443
_
[v

w[
_
v

= sup
_
[v Tw[ : T T

L
,

(373Q)
sup
t
(v)(Tw) : T T

L
,
= sup
t
(v)((Tw)

) : T T

L
,

sup
t
(v)(STw) : T T

L
,
, S T
,
L

(because, given T, we can nd an S such that STw = (Tw)

, by 373O)
sup
t
(v)(Tw) : T T

L
,
L

t
(v).
As w is arbitrary,
t
(v

)
t
(v) and the two are equal. This completes the proof of (b).
(c)(i) The rst step is to note that L

M
0

. PPP??? Suppose that u L

M
0

, that is, that [[[u[ > ]] =
for some > 0. Then u

1 in L
0
(A
L
), so L

(A
L
) L

. For each n N, set v

n
= [n, [

. Then
v

n
= v
0
, so we can nd a T
n
T

L
,
L
such that T
n
v
n
= v
0
(373O), and (v
n
) (v
0
) for every n. But as
v
n
)
nN
is a decreasing sequence with inmum 0, this means that is not an order-continuous norm. XXXQQQ
(ii) Now suppose that A L

is non-empty and downwards-directed and has inmum 0. Then

inf
uA
[[u > ]] = 0 for every > 0 (put 364Nb and 321F together). But this means that B = u

: u A
must have inmum 0; since B is surely downwards-directed, inf
vB
(v) = 0, that is, inf
uA
(u) = 0. As A
is arbitrary, is an order-continuous norm.
374D What is more, every T -invariant extended Fatou norm can be represented in this way.
Theorem Let (A, ) be a semi-nite measure algebra, and a T -invariant extended Fatou norm on L
0
(A).
Then there is a T -invariant extended Fatou norm on L
0
(A
L
) such that (u) = (u

) for every u M
0,

.
proof I use the method of 374C. If A = 0 the result is trivial; assume that A ,= 0.
(a) Set
(w) = sup
_
[w Tv[ : T T
,
L
, v L
0
(A),
t
(v) 1
for w L
0
(A
L
). Note that
(w) = sup
_
w

: v L
0
(A),
t
(v) 1
for every w M
0,

L
, by 373J and 373Q again.
is an extended Fatou norm on L
0
(A
L
). PPP As in 374C, the conditions 369F(i)-(iv) are elementary. If
w > 0 in L
0
(A
L
), take any v L
0
(A) such that 0 <
t
(v) 1; then w

,= 0 so (w)
_
w

> 0.
So 369F(v) is satised. As for 369F(vi), if w > 0 in L
0
(A
L
), take a non-zero a A of nite measure such
that = (a) < . Let > 0, b A
L
be such that 0 <
L
b a and b w; then
(b) = sup

(v)1
_
(b)

sup

(v)1
_
(a)

(a) <
by 374B(ii). So (b) < and 369F(vi) is satised. Thus is an extended Fatou norm. QQQ
(b) is T -invariant. PPP If T T

L
,
L
and w M
1,

L
, then
(Tw) = sup

(v)1
_
(Tw)

sup

(v)1
_
w

= (w)
by 373G and 373I. QQQ
(c) (u

) = (u) for every u M

0,

. PPP We have
(u) = sup

(v)1
_
[u v[ sup

(v)1
_
u

(u),
using 369I, 373E and 374B. So
(u

) = sup

(v)1
_
u

= (u)
by the remark in (a) above. QQQ
444 Linear operators between function spaces 374E
374E I turn now to rearrangement-invariance. Let (A, ) be a measure algebra.
(a) I will say that a subset A of L
0
= L
0
(A) is rearrangement-invariant if T

u A whenever u A
and : A A is a measure-preserving Boolean automorphism, writing T

: L
0
L
0
for the isomorphism
corresponding to (364R).
(b) I will say that an extended Fatou norm on L
0
is rearrangement-invariant if (T

u) = (u)
whenever u L
0
and : A A is a measure-preserving automorphism.
374F Remarks (a) If (A, ) is a semi-nite measure algebra and : A A is a sequentially order-
continuous measure-preserving Boolean homomorphism, then T

M
1,

belongs to T
,
; this is obvious
from the denition of M
1,
= L
1
+ L

and the basic properties of T

(364R). Accordingly, any T -

invariant extended Fatou norm on L
0
(A) must be rearrangement-invariant, since (by 374B) we shall have
(u) = (T

(u)) = when u / M
1,

. Similarly, any T -invariant subset of M
1,

will be rearrangement-
invariant.
(b) I seek to describe cases in which rearrangement-invariance implies T -invariance. This happens only
for certain measure algebras; in order to shorten the statements of the main theorems I introduce a special
phrase.
374G Denition I say that a measure algebra (A, ) is quasi-homogeneous if for any non-zero a,
b A there is a measure-preserving Boolean automorphism : A A such that a b ,= 0.
374H Proposition Let (A, ) be a semi-nite measure algebra. Then the following are equiveridical:
(i) (A, ) is quasi-homogeneous;
(ii) either A is purely atomic and every atom of A has the same measure or there is a such that
the principal ideal A
a
is homogeneous, with Maharam type , for every a A of non-zero nite measure.
proof (i)(ii) Suppose that (A, ) is quasi-homogeneous.
() Suppose that A has an atom a. In this case, for any b A 0 there is an automorphism of
(A, ) such that a b ,= 0; now a must be an atom, so a = a b and a is an atom included in b. As b
is arbitrary, A is purely atomic; moreover, if b is an atom, then it must be equal to a and therefore of the
same measure as a, so all atoms of A have the same measure.
() Now suppose that A is atomless. In this case, if a A has nite non-zero measure, A
a
is homo-
geneous. PPP??? Otherwise, there are non-zero b, c a such that the principal ideals A
b
, A
c
are homogeneous
and of dierent Maharam types, by Maharams theorem (332B, 332H). But now there is supposed to be an
automorphism such that b c ,= 0, in which case A
b
, A
b
, A
bc
and A
c
must all have the same Maharam
type. XXXQQQ
Consequently, if a, b A are both of non-zero nite measure, the Maharam types of A
a
, A
ab
and A
b
must all be the same innite cardinal .
(ii)(i) Assume (ii), and take a, b A 0. If a b ,= 0 we can take to be the identity automorphism
and stop. So let us suppose that a b = 0.
() If A is purely atomic and every atom has the same measure, then there are atoms a
0
a, b
0
b.
Set
c = c if c a
0
b
0
or c (a
0
b
0
) = 0,
= c .(a
0
b
0
) otherwise.
Then it is easy to check that is a measure-preserving automorphism of A such that a
0
= b
0
, so that
a b ,= 0.
() If A
c
is Maharam-type-homogeneous with the same innite Maharam type for every non-zero
c of nite measure, set = min(1, a, b) > 0. Because A is atomless, there are a
0
a, b
0
b with
a
0
= b
0
= (331C). Now A
a
0
, A
b
0
are homogeneous with the same Maharam type and the same
374J Rearrangement-invariant spaces 445
magnitude, so by Maharams theorem (331I) there is a measure-preserving isomorphism
0
: A
a
0
A
b
0
.
Dene : A A by setting
c = (c \ (a
0
b
0
))
0
(c a
0
)
1
0
(c b
0
);
then it is easy to see that is a measure-preserving automorphism of A and that a b ,= 0.
Remark We shall return to these ideas in Chapter 38. In particular, the construction of from
0
in the
last part of the proof will be of great importance; in the language of 381G, = (

a
0
0
b
0
).
374I Corollary Let (A, ) be a quasi-homogeneous semi-nite measure algebra. Then
(a) whenever a, b A have the same nite measure, the principal ideals A
a
, A
b
are isomorphic as measure
algebras;
(b) there is a subgroup of the additive group R such that () a whenever a A and a < ()
whenever a A, and 0 a then there is a c a such that c = .
proof If A is purely atomic, with all its atoms of measure
0
, set =
0
Z, and the results are elementary.
If A is atomless, set = R; then (a) is a consequence of Maharams theorem, and (b) is a consequence of
331C, already used in the proof of 374H.
374J Lemma Let (A, ) be a quasi-homogeneous semi-nite measure algebra and u, v M
0,

. Let
Aut be the group of measure-preserving automorphisms of A. Then
_
u

= sup
Aut
_
[u T

v[,
where T

: L
0
(A) L
0
(A) is the isomorphism corresponding to .
proof (a) Suppose rst that u, v are non-negative and belong to S(A
f
), where A
f
is the ring a : a < ,
as usual. Then they can be expressed as u =

m
i=0

i
a
i
, v =

n
j=0

j
b
j
where
0
. . .
m
0,

0
. . .
n
0, a
0
, . . . , a
m
are disjoint and of nite measure, and b
0
, . . . , b
n
are disjoint and of nite
measure. Extending each list by a nal term having a coecient of 0, if need be, we may suppose that
sup
im
a
i
= sup
jn
b
j
.
Let (t
0
, . . . , t
s
) enumerate in ascending order the set
0

k
i=0
a
i
: k m

k
j=0
b
j
: k n.
Then every t
r
belongs to the subgroup of 374Ib, and t
s
=

m
i=0
a
i
=

n
j=0
b
j
. For 1 r s let k(r),
l(r) be minimal subject to the requirements t
r

k(r)
i=0
a
i
, t
r

l(r)
j=0
b
j
. Then a
i
=

k(r)=i
t
r
t
r1
,
so (using 374Ib) we can nd a disjoint family c
r
)
1rs
such that c
r
a
k(r)
and c
r
= t
r
t
r1
for each
r. Similarly, there is a disjoint family d
r
)
1rs
such that d
r
b
l(r)
and d
r
= t
r
t
r1
for each r. Now
the principal ideals A
c
r
, A
d
r
are isomorphic for every r, by 374Ia; let
r
: A
d
r
A
c
r
be measure-preserving
isomorphisms. Dene : A A by setting
a = (a \ sup
1rs
d
r
) sup
1rs

r
(a d
r
);
because
sup
rs
c
r
= sup
im
a
i
= sup
jn
b
j
= sup
rs
d
r
,
: A A is a measure-preserving automorphism.
Now
u =

s
r=1

k(r)
c
r
, v =

s
r=1

l(r)
d
r
,
u

s
r=1

k(r)
[t
r1
, t
r
[

, v

s
r=1

l(r)
[t
r1
, t
r
[

,
so
_
u T

v =

s
r=1

k(r)

l(r)
c
r
=

s
r=1

k(r)

l(r)
(t
r
t
r1
) =
_
u

.
(b) Now take any u
0
, v
0
M
0,

. Set
A = u : u S(A
f
), 0 u [u
0
[, B = v : v S(A
f
), 0 v [v
0
[.
446 Linear operators between function spaces 374J
Then A is an upwards-directed set with supremum [u
0
[, because (A, ) is semi-nite, so u

: u A is an
upwards-directed set with supremum [u
0
[

= u

0
(373Db, 373Dh). Similarly v

: v B is upwards-directed
and has supremum v

0
, so u

: u A, v B is upwards-directed and has supremum u

0
v

0
.
Consequently, if <
_
u

0
v

0
, there are u A, v B such that
_
u

. Now, by (a), there is a

Aut such that

_
u T

v
_
[u
0
[ T

[v
0
[ =
_
[u
0
T

v
0
[
because T

is a Riesz homomorphism. As is arbitrary,

_
u

0
v

0
sup
Aut
_
[u
0
T

v
0
[.
But the reverse inequality is immediate from 373J.
374K Theorem Let (A, ) be a quasi-homogeneous semi-nite measure algebra, and a rearrangement
-invariant extended Fatou norm on L
0
= L
0
(A). Then is T -invariant.
proof Write
t
for the associate of . Then 374J tells us that for any u, v M
0,

,
_
u

= sup
Aut
_
[T

u v[ sup
Aut
(T

u)
t
(v) = (u)
t
(v),
writing u

, v

for the decreasing rearrangements of u and v, and Aut for the group of measure-preserving
automorphisms of (A, ). But now, if u M
1,

and T T
,
,
(Tu) = sup
_
[Tu v[ :
t
(v) 1
(by 369I)
sup
_
u

:
t
(v) 1
(by 373J)
(u).
As T, u are arbitrary, is T -invariant.
374L Lemma Let (A, ) be a quasi-homogeneous semi-nite measure algebra. Suppose that u, v
(M
0,

)
+
are such that
_
u

= . Then there is a measure-preserving automorphism : A A such

that
_
u T

v = .
proof I take three cases separately.
(a) Suppose that A is purely atomic. Then it is surely innite, since otherwise
_
u

could not be
innite. Let be the common measure of its atoms. For each n N, set

n
= inf : 0, [[u > ]] 2
n
.
Then
n
)
nN
is non-increasing, and
[[u >
n
]] 2
n
[[u
n
]].
We can therefore choose a sequence a
n
)
nN
in A inductively so that
[[u >
n
]] a
n
[[u
n
]], a
n
= 2
n
, a
n
a
n+1
,
for each n. Now if
n+1
<
n
, 2
n
< [[u > ]] 2
n+1
. So if we set
u = |u|

[0, [

sup
nN

_
2
n
, 2
n+1

,
then u

u in L

(A
L
). Set a
n
= a
n+1
\ a
n
for each n; then a
n
)
nN
is disjoint and
a
n
= 2
n
, a
n
[[u
n+1
]]
for each n.
374L Rearrangement-invariant spaces 447
Similarly, we can nd a non-increasing sequence
n
)
nN
in [0, [ and a disjoint sequence b
n
)
nN
in A
such that
b
n
= 2
n
, b
n
[[v
n+1
]]
for each n, while
v

v = |v|

[0, [

sup
nN

_
2
n
, 2
n+1

in L

(A
L
).
Now we are supposing that
_
u

= , so we must have
=
_
u v = |u|

|v|

n=0
2
n

n
.
Because 2
n+1

n+1

n+1
2 2
n

n
for each n, we must have

n=0
2
2n+1

2n+1

2n+1
= .
At this point, recall that we are dealing with a purely atomic algebra in which every atom has measure
. Let A
n
, B
n
be the sets of atoms included in a
n
, b
n
for each n, and A =

nN
A
n
B
n
. Then
#(A
n
) = #(B
n
) = 2
n
for each n. We therefore have a bijection : A A such that [B
2n
] = A
2n
for
every n. (The point is that A

nN
A
2n
and A

nN
B
2n
are both countably innite.) Dene : A A
by setting
c = (c \ supA) sup
aA,ac
a.
Then is well-dened (because A is countable), and it is easy to check that it is a measure-preserving Boolean
automorphism (because it is just a permutation of the atoms); and b
2n
= a
2n
for every n. Consequently
_
u T

n=0

2n+1

2n+1
a
2n
=

n=0
2
2n

2n+1

2n+1
= .
So we have found a suitable automorphism.
(b) Next, consider the case in which (A, ) is atomless and of nite magnitude . Of course > 0. For
each n N set

n
= inf : 0, [[u > ]] 2
n
.
Then
u

sup
nN

n+1

_
2
n1
, 2
n

.
Also
[[u >
n
]] 2
n
[[u
n
]]
for each n, so we can choose inductively a decreasing sequence a
n
)
nN
such that
[[u >
n
]] a
n
[[u
n
]]
and a
n
= 2
n
for each n. Set a
n
= a
n
\ a
n+1
; then a
n
)
nN
is disjoint and a
n
= 2
n1
, a
n
[[u
n
]]
for each n.
In the same way, we can nd
n
)
nN
, b
n
)
nN
such that b
n
)
nN
is a disjoint sequence in A, b
n
= 2
n1

and b
n
[[v
n
]] for each n, and v

sup
nN

n+1

_
2
n1
, 2
n

.
Now all the principal ideals A
a
n
, A
b
n
are homogeneous and of the same Maharam type, so there are
measure-preserving isomorphisms
n
: A
b
n
A
a
n
. Dene : A A by setting c = sup
nN

n
(c a
n
);
then is a measure-preserving automorphism of A, and b
n
= a
n
for each n. Since u a
n

n
a
n
,
v b
n

n
b
n
for each n,
_
u T

n=0
2
n1

n
;
but on the other hand,
_
u

n=0
2
n1

n+1

n+1
2
_
u T

v.
So
_
u T

v = .
(c) Thirdly, consider the case in which A is atomless and not totally nite; take to be the common
Maharam type of all the principal ideals A
a
where 0 < a < . In this case, set
448 Linear operators between function spaces 374L

n
= inf : [[u > ]] 2
n
,
n
= inf : [[v > ]] 2
n

for each n Z. This time

u

sup
nZ

_
2
n
, 2
n+1
_

, v

sup
nZ

_
2
n
, 2
n+1
_

.
There are disjoint families a
n
)
nZ
, b
n
)
nZ
such that a
n
= b
n
= 2
n
for each n and
u sup
nZ

n+1
a
n
, v sup
nZ

n+1
b
n
.
(This time, start by xing a
0
such that a
0
= 1 and [[u >
0
]] a
0
[[u
0
]], and choose a
n+1
a
n
for
n 0, a
n1
a
n
for n 0.)
Set d

= sup
nZ
a
n
sup
nZ
b
n
. Then
d
1
= d

\ sup
nZ
a
2n
, d
2
= d

\ sup
nZ
b
2n
both have magnitude and Maharam type . So there is a measure-preserving isomorphism : A
d
2
A
d
1
(332J). At the same time, for each n Z there is a measure-preserving isomorphism
n
: A
b
2n
A
a
2n
. So
once again we can assemble these to form a measure-preserving automorphism : A A, dened by the
formula
c = (c \ d

) (c d
2
) sup
nZ

n
(c b
2n
).
Just as in (a) and (b) above,
_
u T

n=
2
2n

2n+1

2n+1
,
while
_
u

n=
2
n

n
is innite. Because
2
2n+2

2n+2

2n+2
+ 2
2n+1

2n+1

2n+1
6 2
2n

2n+1

2n+1
for every n,
_
u T

v
1
6
_
u

= .
Thus we have a suitable in any of the cases allowed by 374H.
374M Proposition Let (A, ) be a quasi-homogeneous localizable measure algebra, and U L
0
=
L
0
(A) a solid linear subspace which, regarded as a Riesz space, is perfect. If U is rearrangement-invariant
and M
,1

U M
1,

, then U is T -invariant.
proof Set V = v : u v L
1
for every u U, so that V is a solid linear subspace of L
0
which can be
identied with U

(369C), and U becomes u : u v L

1
for every v V ; note that M
,1

V M
1,

(using 369Q).
If u U
+
, v V
+
and : A A is a measure-preserving automorphism, then T

u U, so
_
vT

u < ;
by 374L,
_
u

is nite. But this means that if u U, v V and T T

,
,
_
[Tu v[
_
u

< .
As v is arbitrary, Tu U; as T and u are arbitrary, U is T -invariant.
374X Basic exercises >>>(a) Let (A, ) be a measure algebra and A M
1,

a T -invariant set. (i)
Show that A is solid. (ii) Show that if A is a linear subspace and not 0, then it includes M
,1

. (iii)
Show that if u A, v M
0,

and
_
t
0
v

_
t
0
u

for every t > 0, then v A. (iv) Show that if (B, ) is

any other measure algebra, then B = Tu : u A, T T
,
and C = v : v M
1,

, Tv A for every
T T
,
are T -invariant subsets of M
1,

, and that B C. Give two examples in which B C. Show
that if (A, ) = (A
L
,
L
) then B = C.
>>>(b) Let (A, ) be a measure algebra. Show that the extended Fatou norm | |
p
on L
0
(A) is T -invariant
for every p [1, ]. (Hint: 371Gd.)
374Yb Rearrangement-invariant spaces 449
(c) Let (A, ) and (B, ) be semi-nite measure algebras, and a Youngs function (369Xc). Let

be the corresponding Orlicz norms on L

0
(A), L
0
(B). Show that

(Tu)

(u) for every u L

0
(A),
T T
,
. (Hint: 369Xn, 373Xm.) In particular,

is T -invariant.
(d) Show that if (A, ) is a semi-nite measure algebra and is a T -invariant extended Fatou norm on
L
0
(A), then the Banach lattice L

dened from is T -invariant.

(e) Let (A, ) be a semi-nite measure algebra and a T -invariant extended Fatou norm on L
0
which is
an order-continuous norm on L

. Show that L

M
1,0

.
(f ) Let be a T -invariant extended Fatou norm on L
0
(A
L
) and (A, ), (B, ) two semi-nite measure
algebras. Let
1
,
2
be the extended Fatou norms on L
0
(A), L
0
(B) dened from by the method of 374C.
Show that
2
(Tu)
1
(u) whenever u M
1,

and T T
,
.
>>>(g) Let (A, ) be a semi-nite measure algebra, not 0, and set (u) = sup
0< a<
1

a
_
a
[u[ for
u L
0
(A). Show that is a T -invariant extended Fatou norm. Find examples of (A, ) for which is, and
is not, order-continuous on L

.
(h) Let (A, ) and (B, ) be semi-nite measure algebras and a T -invariant extended Fatou norm on
L
0
(A). (i) Show that there is a T -invariant extended Fatou norm on L
0
(B) dened by setting (v) =
sup(Tv) : T T
,
for v M
1,

. (ii) Show that
t
(w) = sup
t
(Tw) : T T
,
for every w M
1,

.
(iii) Show that when (A, ) = (A
L
,
L
) then (v) = (v

) for every v M
0,

. (iv) Show that when
(B, ) = (A
L
,
L
) then (u) = (u

) for every u M
0,

.
(i) Let (A, ) be a semi-nite measure algebra and an extended Fatou norm on L
0
= L
0
(A). Suppose
that L

is a T -invariant subset of L
0
. Show that there is a T -invariant extended Fatou norm which is
equivalent to in the sense that, for some M > 0, (u) M(u) M
2
(u) for every u L
0
. (Hint: show
rst that
_
u

< for every u L

and v L

, then that sup

(u)1,

(v)1
_
u

< .)
(j) Suppose that is a T -invariant extended Fatou norm on L
0
(A
L
), and that 0 < w = w

M
1,

L
. Let
(A, ) be any semi-nite measure algebra. Show that the function u (w u

) extends to a T -invariant
extended Fatou norm on L
0
(A). (Hint: (w u

) = sup(w Tu) : T T
,
L
for u M
1,

L
.) (When
= | |
p
these norms are called Lorentz norms; see Lindenstrauss & Tzafriri 79, p. 121.)
(k) Let (A, ) be TN with counting measure. Identify L
0
(A) with R
N
. Let U be u : u R
N
, n : u(n) ,=
0 is nite. Show that U is a perfect Riesz space, and is rearrangement-invariant but not T -invariant.
(l) Let (A, ) be an atomless quasi-homogeneous localizable measure algebra, and U L
0
(A) a rearrange-
ment-invariant solid linear subspace which is a perfect Riesz space. Show that U M
1,

and that U is T -
invariant. (Hint: assume U ,= 0. Show that (i) a U whenever a < (ii) V = v : vu L
1
u U
is rearrangement-invariant (iii) U, V M
1,
.)
374Y Further exercises (a) Let (A, ) be a localizable measure algebra and U M
1,

a non-zero
T -invariant Riesz subspace which, regarded as a Riesz space, is perfect. (i) Show that U includes M
,1

.
(ii) Show that its dual v : v L
0
, v u L
1
u U (which in this exercise I will denote by U

) is also
T -invariant, and is v : v M
0,

,
_
u

< u U. (iii) Show that for any localizable measure

algebra (B, ) the set V = v : v M
1,

, Tv U T T
,
is a perfect Riesz subspace of L
0
(B), and
that V

= v : v M
1,

, Tv U

T T
,
. (iv) Show that if, in (i)-(iii), (A, ) = (A
L
,
L
), then
V = v : v M
0,
, v

U. (v) Show that if, in (iii), (B, ) = (A

L
,
L
), then U = u : u M
0,
, u

V .
(b) Let (A, ) be a semi-nite measure algebra, and suppose that 1 q p < . Let w
pq
L
0
(A
L
) be
the equivalence class of the function t t
(qp)/p
. (i) Show that for any u L
0
(A),
_
w
pq
(u

)
q
= p
_

0
t
q1
( [[[u[ > t]])
q/p
dt.
450 Linear operators between function spaces 374Yb
(ii) Show that we have an extended Fatou norm | |
p,q
on L
0
(A) dened by setting
|u|
p,q
=
_
p
_

0
t
q1
( [[[u[ > t]])
q/p
dt
_
1/q
for every u L
0
(A). (Hint: use 374Xj with w = w
1/q
pq
, | | = | |
q
.) (iii) Show that if (B, ) is another
measure algebra and T T
,
, then |Tu|
p,q
|u|
p,q
for every u M
1,

. (iv) Show that | |
p,q
is an
order-continuous norm on L
| |
p,q
.
(c) Let (A, ) be a homogeneous measure algebra of uncountable Maharam type, and u, v 0 in M
0

such that u

= v

. Show that there is a measure-preserving automorphism of A such that T

u = v, where
T

: L
0
(A) L
0
(A) is the isomorphism corresponding to .
(d) In L
0
(A
L
) let u be the equivalence class of the function f(t) = te
t
. Show that there is no Boolean
automorphism of A
L
such that T

u = u

. (Hint: show that A

L
is -generated by [[u

> ]] : > 0.)

(e) Let (A, ) be a quasi-homogeneous semi-nite measure algebra and C L
0
(A) a solid convex order-
closed rearrangement-invariant set. Show that C M
1,

is T -invariant.
374 Notes and comments I gave this section the title rearrangement-invariant spaces because it looks
good on the Contents page, and it follows what has been common practice since Luxemburg 67b; but
actually I think that its T -invariance which matters, and that rearrangement-invariant spaces are signicant
largely because the important ones are T -invariant. The particular quality of T -invariance which I have
tried to bring out here is its transferability from one measure algebra (or measure space, of course) to
another. This is what I take at a relatively leisurely pace in 374B-374D and 374Xf, and then encapsulate
in 374Xh and 374Ya. The special place of the Lebesgue algebra (A
L
,
L
) arises from its being more or less
the simplest algebra over which every T -invariant set can be described; see 374Xa.
I dont think this section is particularly easy, and (as in 373) there are rather a lot of unattractive names
in it; but once one has achieved a reasonable familiarity with the concepts, the techniques used can be seen
to amount to half a dozen ideas non-trivial ideas, to be sure from 369 and 373. From 369 I take
concepts of duality: the symmetric relationship between a perfect Riesz space U L
0
and the representation
of its dual (369C-369D), and the notion of associate extended Fatou norms (369H-369K). From 373 I take
the idea of decreasing rearrangement and theorems guaranteeing the existence of useful members of T
,
(373O-373Q). The results of the present section all depend on repeated use of these facts, assembled in a
variety of patterns.
There is one new method here, but an easy one: the construction of measure-preserving automorphisms
by joining isomorphisms together, as in the proofs of 374H and 374J. I shall return to this idea, in greater
generality and more systematically investigated, in 381. I hope that the special cases here will give no
diculty.
While T -invariance is a similar phenomenon for both extended Fatou norms and perfect Riesz spaces
(see 374Xh, 374Ya), the former seem easier to deal with. The essential dierence is I think in 374B(i);
with a T -invariant extended Fatou norm, we are necessarily conned to M
1,
, the natural domain of the
methods used here. For perfect Riesz spaces we have examples like R
N
= L
0
(TN) and its dual, the space of
eventually-zero sequences (374Xk); these are rearrangement-invariant but not T -invariant, as I have dened
it. This problem does not arise over atomless algebras (374Xl).
I think it is obvious that for algebras which are not quasi-homogeneous (374G) rearrangement-invariance
is going to be of limited interest; there will be regions between which there is no communication by means
of measure-preserving automorphisms, and the best we can hope for is a discussion of quasi-homogeneous
components, if they exist, corresponding to the partition of unity used in the proof of 332J. There is a
special diculty concerning rearrangement-invariance in L
0
(A
L
): two elements can have the same decreas-
ing rearrangement without being rearrangements of each other in the strict sense (373Ya, 374Yd). The
phenomenon of 373Ya is specic to algebras of countable Maharam type (374Yc). You will see that some of
the labour of 374L is because we have to make room for the pieces to move in. 374J is easier just because
in that context we can settle for a supremum, rather than an actual innity, so the rearrangement needed
(part (a) of the proof) can be based on a region of nite measure.
375C Kwapiens theorem 451
375 Kwapiens theorem
In 368 and the rst part of 369 I examined maps from various types of Riesz space into L
0
spaces.
There are equally striking results about maps out of L
0
spaces. I start with some relatively elementary facts
about positive linear operators from L
0
spaces to Archimedean Riesz spaces in general (375A-375D), and
then turn to a remarkable analysis, due essentially to S.Kwapien, of the positive linear operators from a
general L
0
space to the L
0
space of a semi-nite measure algebra (375I), with a couple of simple corollaries.
375A Theorem Let A be a Dedekind -complete Boolean algebra and W an Archimedean Riesz space.
If T : L
0
(A) W is a positive linear operator, it is sequentially order-continuous.
proof (a) The rst step is to observe that if u
n
)
nN
is any non-increasing sequence in L
0
= L
0
(A)
with inmum 0, and > 0, then n(u
n
u
0
) : n N is bounded above in L
0
. PPP For k N set
a
k
= sup
nN
[[n(u
n
u
0
) > k]]; set a = inf
kN
a
k
. ??? Suppose, if possible, that a ,= 0. Because u
n
u
0
,
n(u
n
u
0
) nu
0
for every n and
a a
0
[[u
0
> 0]] = [[u
0
> 0]] = sup
nN
[[u
0
u
n
> 0]].
So there is some m N such that a
t
= a [[u
0
u
m
> 0]] ,= 0. Now, for any n m, any k N,
a
t
[[n(u
n
u
0
) > k]] [[u
0
u
m
> 0]] [[u
m
u
0
> 0]] = 0.
But a
t
sup
nN
[[n(u
n
u
0
) > k]], so in fact
a
t
sup
nm
[[n(u
n
u
0
) > k]] = [[v > k]],
where v = sup
nm
n(u
n
u
0
). And this means that inf
kN
[[v > k]] a
t
,= 0, which is impossible. XXX
Accordingly a = 0; by 364Ma, n(u
n
u
0
) : n N is bounded above. QQQ
(b) Now suppose that u
n
)
nN
is a non-increasing sequence in L
0
with inmum 0, and that w W is a
lower bound for Tu
n
: n N. Take any > 0. By (a), n(u
n
u
0
) : n N has an upper bound v in L
0
.
Because T is positive,
w Tu
n
= T(u
n
u
0
) +T(u
0
) T(
1
n
v) +T(u
0
) =
1
n
Tv +Tu
0
for every n 1. Because W is Archimedean, w Tu
0
. But this is true for every > 0, so (again because
W is Archimedean) w 0. As w is arbitrary, inf
nN
Tu
n
= 0. As u
n
)
nN
is arbitrary, T is sequentially
order-continuous (351Gb).
375B Proposition Let A be an atomless Dedekind -complete Boolean algebra. Then L
0
(A)

= 0.
proof ??? Suppose, if possible, that h : L
0
(A) R is a non-zero order-continuous positive linear functional.
Then there is a u > 0 in L
0
such that h(v) > 0 whenever 0 < v u (356H). Because A is atomless, there is
a disjoint sequence a
n
)
nN
such that a
n
[[u > 0]] for each n, so that u
n
= ua
n
> 0, while u
m
u
n
= 0
if m ,= n. Now however
v = sup
nN
n(h(u
n
))
1
u
n
is dened in L
0
, by 368K, and h(v) n for every n, which is impossible. XXX
375C Theorem Let A be a Dedekind complete Boolean algebra, W an Archimedean Riesz space, and
T : L
0
(A) W an order-continuous Riesz homomorphism. Then V = T[L
0
(A)] is an order-closed Riesz
subspace of W.
proof The kernel U of T is a band in L
0
= L
0
(A) (352Oe), and must be a projection band (353I), because
L
0
is Dedekind complete (364O). Since U + U

= L
0
, T[U] + T[U

] = V , that is, T[U

] = V ; since
U U

= 0, T is an isomorphism between U

and V . Now suppose that A V is upwards-directed and

has a least upper bound w W. Then B = u : u U

, Tu A is upwards-directed and T[B] = A. The

point is that B is bounded above in L
0
. PPP??? If not, then u
+
: u B cannot be bounded above, so there
is a u
0
> 0 in L
0
such that nu
0
= sup
uB
nu
0
u
+
for every n N (368A). Since B U

, u
0
U

and
Tu
0
> 0. But now, because T is an order-continuous Riesz homomorphism,
452 Linear operators between function spaces 375C
nTu
0
= sup
uB
T(nu
0
u
+
) = sup
vA
nTu
0
v
+
w
+
for every n N, which is impossible. XXXQQQ
Set u

= supB; then Tu

= sup A = w and w V . As A is arbitrary, V is order-closed.

375D Corollary Let W be an Archimedean Riesz space and V an order-dense Riesz subspace which is
isomorphic to L
0
(A) for some Dedekind complete Boolean algebra A. Then V = W.
proof Apply 375C to an isomorphism T : L
0
(A) V to see that V is order-closed in W.
375E I come now to the deepest result of this section, concerning positive linear operators from L
0
(A)
to L
0
(B) where B is a measure algebra. I approach through a couple of lemmas which are striking enough
in their own right.
The following temporary denition will be useful.
Denition Let A and B be Boolean algebras. I will say that a function : A B is a -subhomomor-
phism if
(a a
t
) = (a) (a
t
) for all a, a
t
A,
inf
nN
(a
n
) = 0 whenever a
n
)
nN
is a non-increasing sequence in A with inmum 0.
Now we have the following easy facts.
375F Lemma Let A and B be Boolean algebras and : A B a -subhomomorphism.
(a) (0) = 0, (a) (a
t
) whenever a a
t
, and (a) \ (a
t
) (a \ a
t
) for every a, a
t
A.
(b) If , are measures such that (A, ) and (B, ) are totally nite measure algebras, then for every
> 0 there is a > 0 such that (a) whenever a .
proof (a) This is elementary. Set every a
n
= 0 in the second clause of the denition 375E to see that
(0) = 0. The other two parts are immediate consequences of the rst clause.
(b) (Compare 232B, 327Bb.) ??? Suppose, if possible, otherwise. Then for every n N there is an a
n
A
such that a
n
2
n
and (a
n
) . Set c
n
= sup
in
a
i
for each n; then c
n
)
nN
is non-increasing and has
inmum 0 (since c
n
2
n+1
for each n), but (c
n
) for every n, so inf
nN
c
n
cannot be 0. XXX
375G Lemma Let (A, ) and (B, ) be totally nite measure algebras and : A B a -sub-
homomorphism. Then for every non-zero b
0
B there are a non-zero b b
0
and an m N such that
b inf
jm
(a
j
) = 0 whenever a
0
, . . . , a
m
A are disjoint.
proof (a) Suppose rst that A is atomless and that 1 = 1.
Set =
1
5
b
0
and let m 1 be such that (a) whenever a
1
m
. We need to know that
(1
1
m
)
m

1
2
; this is because (if m 2) lnmln(m1)
1
m
, so mln(1
1
m
) 1 ln2.
Set
C = inf
jm
(a
j
) : a
0
, . . . , a
m
A are disjoint.
??? Suppose, if possible, that b
0
supC. Then there are c
0
, . . . , c
k
C such that (b
0
sup
ik
c
i
) 4.
For each i k choose disjoint a
i0
, . . . , a
im
A such that c
i
= inf
jm
(a
ij
). Let D be the set of atoms of
the nite subalgebra of A generated by a
ij
: i k, j m, so that D is a nite partition of unity in A,
and every a
ij
is the join of the members of D it includes. Set p = #(D), and for each d D take a maximal
disjoint set E
d
e : e d, e =
1
pm
, so that (d \ supE
d
) <
1
pm
; set
d

= 1 \ sup(

dD
E
d
) = sup
dD
(d \ supE
d
),
so that d

is a multiple of
1
pm
and is less than
1
m
. Let E

be a disjoint set of elements of measure

1
pm
with
union d

, and take E = E

dD
E
d
, so that E is a partition of unity in A, e =
1
pm
for every e E, and
a
ij
\ d

is the join of the members of E it includes for every i k, j m.

Set
/ = K : K E, #(K) = p, M = #(/) =
(mp)!
p!(mpp)!
.
375H Kwapiens theorem 453
For every K /, (supK) =
1
m
so (supK) . So if we set
v =

KK
(supK),
_
v M. On the other hand,
(b
0
sup
ik
c
i
) 4, (d

) ,
so b
1
3, where
b
1
= b
0
sup
ik
c
i
\ (d

).
Accordingly
_
v
1
3
M b
1
and
b
2
= b
1
[[v <
1
2
M]]
is non-zero.
Because b
2
b
1
, there is an i k such that b
2
c
i
,= 0. Now
b
2
c
i
c
i
\ (d

) = inf
jm
(a
ij
) \ (d

) inf
jm
(a
ij
\ d

).
But every a
ij
\ d

is the join of the members of E it includes, so

b
2
c
i
inf
jm
(a
ij
\ d

) inf
jm
(supe : e E, e a
ij
)
= inf
jm
sup(e) : e E, e a
ij

= sup inf
jm
(e
j
) : e
0
, . . . , e
m
E and e
j
a
ij
for every j.
So there are e
0
, . . . , e
m
E such that e
j
a
ij
for each j and b
3
= b
2
inf
jm
(e
j
) ,= 0. Because
a
i0
, . . . , a
im
are disjoint, e
0
, . . . , e
m
are distinct; set J = e
0
, . . . , e
m
. Then whenever K / and KJ ,= ,
b
3
(supK).
So let us calculate the size of /
1
= K : K /, K J ,= . This is
M
(mpm1)!
p!(mppm1)!
= M
_
1
(mpp)(mpp1)...(mppm)
mp(mp1)...(mpm)
_
M
_
1 (
mpp
mp
)
m+1
_

1
2
M.
But this means that b
3
[[v
1
2
M]], while also b
3
[[v <
1
2
M]]; which is surely impossible. XXX
Accordingly b
0
, supC, and we can take b = b
0
\ supC.
(b) Now for the general case. Let A be the set of atoms of A, and set d = 1 \ supA. Then A
d
is atomless,
so there are a non-zero b
1
b
0
and an n N such that b
1
inf
jn
(a
j
) = 0 whenever a
0
, . . . , a
n
A
d
are
disjoint. PPP If d > 0 this follows from (a), if we apply it to A
d
and ( d)
1
A
d
. If d = 0 then we can
just take b
1
= b
0
, n = 0. QQQ
Let > 0 be such that (a) < b
1
whenever a . Let A
1
A be a nite set such that (supA
1
)
(supA) , and set r = #(A), d

= sup(A\ A
1
). Then d

so b = b
1
\ (d

) ,= 0. Try m = n + r.
If a
0
, . . . , a
m
are disjoint, then at most r of them can meet sup A
1
, so (re-ordering if necessary) we can
suppose that a
0
, . . . , a
n
are disjoint from supA
1
, in which case a
j
\ d

d for each j m. But in this case

(because b (d

) = 0)
b inf
jm
(a
j
) b inf
jn
(a
j
) = b inf
jn
(a
j
d) = 0
by the choice of n and b
1
.
Thus in the general case also we can nd appropriate b and m.
375H Lemma Let (A, ) and (B, ) be totally nite measure algebras and : A B a -subhomo-
morphism. Then for every non-zero b
0
B there are a non-zero b b
0
and a nite partition of unity C A
such that a b (a c) is a ring homomorphism for every c C.
proof By 375G, we can nd b
1
, m such that 0 ,= b
1
b
0
and b
1
inf
jm
(a
j
) = 0 whenever a
0
, . . . , a
m

A are disjoint. Do this with the smallest possible m. If m = 0 then b
1
(1) = 0, so we can take
454 Linear operators between function spaces 375H
b = b
1
, C = 1. Otherwise, because m is minimal, there must be disjoint c
1
, . . . , c
m
A such that
b = b
1
inf
1jm
(c
j
) ,= 0. Set c
0
= 1 \ sup
1jm
c
j
, C = c
0
, c
1
, . . . , c
m
; then C is a partition of unity
in A. Set
j
(a) = b (a c
j
) for each a A, j m. Then we always have
j
(a a
t
) =
j
(a)
j
(a
t
) for
all a, a
t
A, because is a subhomomorphism.
To see that every
j
is a ring homomorphism, we need only check that
j
(a a
t
) = 0 whenever a a
t
= 0.
(Compare 312H(iv).) In the case j = 0, we actually have
0
(a) = 0 for every a, because b (c
0
) =
b
1
inf
0jm
(c
j
) = 0 by the choice of b
1
and m. When 1 j m, if a a
t
= 0, then

j
(a)
j
(a
t
) = b
1
inf
1im,i,=j
(c
j
) (a) (a
t
)
is again 0, because a, a
t
, c
1
, . . . , c
j1
, c
j+1
, . . . , c
m
are disjoint. So we have a suitable pair b, C.
375I Theorem Let A be any Dedekind -complete Boolean algebra and (B, ) a semi-nite measure
algebra. Let T : L
0
(A) L
0
(B) be a positive linear operator. Then we can nd B, A
b
)
bB
such that B
is a partition of unity in B, each A
b
is a nite partition of unity in A, and u T(u a) b is a Riesz
homomorphism for every b B, a A
b
.
proof (a) Write B

for the set of potential members of B; that is, the set of those b B such that
there is a nite partition of unity A A such that T
ab
is a Riesz homomorphism for every a A, writing
T
ab
(u) = T(u a) b. If I can show that B

is order-dense in B, this will suce, since there will then

be a partition of unity B B

.
(b) So let b
0
be any non-zero member of B; I seek a non-zero member of B

included in b
0
. Of course
there is a non-zero b
1
b
0
with b
1
< . Let > 0 be such that b
2
= b
1
[[T(1) ]] is non-zero. Dene
: A [0, [ by setting a =
_
b
2
T(a) for every a A. Then is countably additive, because , T and
_
are all additive and sequentially order-continuous (using 375A). Set ^ = a : a = 0; then ^ is a -ideal
of A, and (C, ) is a totally nite measure algebra, where C = A/^ and a

= a for every a A (just as

in 321H).
(c) We have a function from C to the principal ideal B
b
2
dened by saying that a

= b
2
[[T(a) > 0]]
for every a A. PPP If a
1
, a
2
A are such that a

1
= a

2
in C, this means that a
1
.a
2
^; now
[[T(a
1
) > 0]] .[[T(a
2
) > 0]] [[[T(a
1
) T(a
2
)[ > 0]]
[[T([a
1
a
2
[) > 0]] = [[T(a
1
.a
2
) > 0]]
is disjoint from b
2
because
_
b
2
T(a
1
.a
2
) = 0. Accordingly b
2
[[T(a
1
) > 0]] = b
2
[[T(a
2
) > 0]] and we
can take this common value for (a

1
) = (a

2
). QQQ
(d) Now is a -subhomomorphism. PPP (i) For any a
1
, a
2
A we have
[[T(a
1
a
2
) > 0]] = [[T(a
1
) > 0]] [[T(a
2
) > 0]]
because
T(a
1
) T(a
2
) T(a
1
a
2
) T(a
1
) +T(a
2
).
So (c
1
c
2
) = (c
1
) (c
2
) for all c
1
, c
2
C. (ii) If c
n
)
nN
is a non-increasing sequence in C with inmum
0, choose a
n
A such that a

n
= c
n
for each n, and set a
n
= inf
in
a
i
\ inf
iN
a
i
for each n; then a

n
= c
n
so
(c
n
) = [[T( a
n
) > 0]] for each n, while a
n
)
nN
is non-increasing and inf
nN
a
n
= 0. ??? Suppose, if possible,
that b
t
= inf
nN
(c
n
) ,= 0; set =
1
2
b
t
. Then (b
2
[[T( a
n
) > 0]]) 2 for every n N. For each n,
take
n
> 0 such that (b
2
[[T( a
n
) >
n
]]) . Then u = sup
nN
n
1
n
a
n
is dened in L
0
(A) (because
sup
nN
[[n
1
n
a
n
> k]] a
m
if k max
im
i
1
i
, so inf
kN
sup
nN
[[n
1
n
a
n
> k]] = 0). But now
(b
2
[[Tu > n]]) (b
2
[[T( a
n
) >
n
]])
for every n, so inf
nN
[[Tu > n]] ,= 0, which is impossible. XXX Thus inf
nN
(c
n
) = 0; as c
n
)
nN
is arbitrary,
is a -subhomomorphism. QQQ
(e) By 375H, there are a non-zero b B
b
2
and a nite partition of unity C C such that d b (d c)
is a ring homomorphism for every c C. There is a partition of unity A A, of the same size as C, such
that C = a

: a A. Now T
ab
is a Riesz homomorphism for every a A. PPP It is surely a positive linear
375Xc Kwapiens theorem 455
operator. If u
1
, u
2
L
0
(A) and u
1
u
2
= 0, set e
i
= [[u
i
> 0]] for each i, so that e
1
e
2
= 0. Observe that
u
i
= sup
nN
u
i
ne
i
, so that
[[T
ab
u
i
> 0]] = sup
nN
[[T
ab
(u
i
ne
i
) > 0]] [[T
ab
(e
i
) > 0]] = b [[T(e
i
a) > 0]]
for both i (of course T
ab
, like T, is sequentially order-continuous). But this means that
[[T
ab
u
1
> 0]] [[T
ab
u
2
> 0]] b [[T(e
1
a) > 0]] [[T(e
2
a) > 0]]
= b (e

1
a

) (e

2
a

) = 0
because a

C, so d (d a

) is a ring homomorphism, while e

1
e

2
= 0. So T
ab
u
1
T
ab
u
2
= 0. As u
1
and u
2
are arbitrary, T
ab
is a Riesz homomorphism (352G(iv)). QQQ
(f ) Thus b B

. As b
0
is arbitrary, B

is order-dense, and were home.

375J Corollary Let A be a Dedekind -complete Boolean algebra and U a Dedekind complete Riesz
space such that U

separates the points of U. If T : L

0
(A) U is a positive linear operator, there is a
sequence T
n
)
nN
of Riesz homomorphisms from L
0
(A) to U such that T =

n=0
T
n
, in the sense that
Tu = sup
nN

n
i=0
T
i
u for every u 0 in L
0
(A).
proof By 369A, U can be embedded as an order-dense Riesz subspace of L
0
(B) for some localizable measure
algebra (B, ); being Dedekind complete, it is solid in L
0
(B) (353K). Regard T as an operator from L
0
(A)
to L
0
(B), and take B, A
b
)
bB
as in 375I. Note that L
0
(B) can be identied with

bB
L
0
(B
b
) (364S,
322K). For each b B let f
b
: A
b
N be an injection. If b B and n f
b
[A
b
], set T
nb
(u) = b T(ua);
otherwise set T
nb
= 0. Then T
nb
: L
0
(A) L
0
(B
b
) is a Riesz homomorphism; because A
b
is a nite partition
of unity,

n=0
T
nb
u = b Tu for every u L
0
(A). But this means that if we set T
n
u = T
nb
u)
bB
,
T
n
: L
0
(A)

bB
L
0
(B
b
)

= L
0
(B)
is a Riesz homomorphism for each n; and T =

n=0
T
n
. Of course every T
n
is an operator from L
0
(A) to
U because [T
n
u[ T[u[ U for every u L
0
(A).
375K Corollary (a) If A is a Dedekind -complete Boolean algebra, (B, ) is a semi-nite measure
algebra, and there is any non-zero positive linear operator from L
0
(A) to L
0
(B), then there is a non-trivial
sequentially order-continuous ring homomorphism from A to B.
(b) If (A, ) and (B, ) are homogeneous probability algebras and (A) > (B), then L

(L
0
(A); L
0
(B)) =
0.
proof (a) It is probably quickest to look at the proof of 375I: starting from a non-zero positive linear
operator T : L
0
(A) L
0
(B), we move to a non-zero -subhomomorphism : A/^ B and thence to
a non-zero ring homomorphism from A/^ to B, corresponding to a non-zero ring homomorphism from A
to B, which is sequentially order-continuous because it is dominated by . Alternatively, quoting 375I, we
have a non-zero Riesz homomorphism T
1
: L
0
(A) L
0
(B), and it is easy to check that a [[T(a) > 0]] is
a non-zero sequentially order-continuous ring homomorphism.
(b) Use (a) and 331J.
375X Basic exercises (a) Let A be a Dedekind complete Boolean algebra and W an Archimedean
Riesz space. Let T : L
0
(A) W be a positive linear operator. Show that T is order-continuous i
T : A W is order-continuous.
(b) Let A be an atomless Dedekind -complete Boolean algebra and W a Banach lattice. Show that the
only order-continuous positive linear operator from L
0
(A) to W is the zero operator.
(c) Let A be a Dedekind complete Boolean algebra and W an Archimedean Riesz space. Let T : L
0
(A)
W be an order-continuous Riesz homomorphism such that T[L
0
(A)] is order-dense in W. Show that T is
surjective.
456 Linear operators between function spaces 375Xd
(d) Let A and B be Boolean algebras and : A B a -subhomomorphism as dened in 375E. Show
that is sequentially order-continuous.
>>>(e) Let A be the measure algebra of Lebesgue measure on [0, 1] and G the regular open algebra of
R. (i) Show that there is no non-zero positive linear operator from L
0
(G) to L
0
(A). (Hint: suppose
T : L
0
(G) L
0
(A) were such an operator. Reduce to the case T(1) 1. Let b
n
)
nN
enumerate an
order-dense subset of G (316Yn). For each n N take non-zero b
t
n
b
n
such that
_
T(b
t
n
) 2
n2
_
T(1)
and consider T(sup
nN
b
t
n
).) (ii) Show that there is no non-zero positive linear operator from L
0
(A) to
L
0
(G). (Hint: suppose T : L
0
(A) L
0
(G) were such an operator. For each n N choose a
n
A,

n
> 0 such that a
n
2
n
and if b
n
[[T(1) > 0]] then b
n
[[T(a
n
) >
n
]] ,= 0. Consider Tu where
u =

n=0
n
1
n
a
n
.)
(f ) In 375J, show that for any u L
0
(A)
inf
nN
sup
mn
[[[Tu

m
i=0
T
i
u[ > 0]] = 0.
>>>(g) Prove directly, without quoting 375E-375K, that if A is a Dedekind -complete Boolean algebra
then every positive linear functional from L
0
(A) to R is a nite sum of Riesz homomorphisms.
375Y Further exercises (a) Show that the following are equiveridical: (i) there is a purely atomic
probability space (X, , ) such that = TX and x = 0 for every x X; (ii) there are a set X and
a Riesz homomorphism f : R
X
R which is not order-continuous; (iii) there are a Dedekind complete
Boolean algebra A and a positive linear operator f : L
0
(A) R which is not order-continuous; (iv) there
are a Dedekind complete Boolean algebra A and a sequentially order-continuous Boolean homomorphism
: A 0, 1 which is not order-continuous; (v) there are a Dedekind complete Riesz space U and
a sequentially order-continuous Riesz homomorphism f : U R which is not order-continuous; *(vi)
there are an atomless Dedekind complete Boolean algebra A and a sequentially order-continuous Boolean
homomorphism : A 0, 1 which is not order-continuous. (Compare 363S.)
(b) Give an example of an atomless Dedekind -complete Boolean algebra A such that L
0
(A)

,= 0.
(c) Let A, B be Dedekind -complete Boolean algebras of which B is weakly -distributive. Let
T : L
0
(A) L
0
(B) be a positive linear operator. Show that a [[T(a) > 0]] : A B is a -subho-
momorphism.
(d) Let (A, ) be a localizable measure algebra and U a Riesz space such that U

separates the points

of U. Suppose that T : L
0
(A) U is an order-continuous positive linear operator. Show that T[L
0
(A)] is
order-closed.
(e) Let A and B be Dedekind complete Boolean algebras, and : A B an -subhomomorphism such
that 1
A
= 1
B
. Show that there is a sequentially order-continuous Boolean homomorphism : A B such
that a a for every a A.
(f ) Let G be the regular open algebra of R, and L
0
= L
0
(G). Give an example of a non-zero positive
linear operator T : L
0
L
0
such that there is no non-zero Riesz homomorphism S : L
0
L
0
with S T.
375Z Problem Let G be the regular open algebra of R, and L
0
= L
0
(G). If T : L
0
L
0
is a positive
linear operator, must T[L
0
] be order-closed?
375 Notes and comments Both this section, and the earlier work on linear operators into L
0
spaces,
can be regarded as describing dierent aspects of a single fact: L
0
spaces are very large. The most explicit
statements of this principle are 368E and 375D: every Archimedean Riesz space can be embedded into a
Dedekind complete L
0
space, but no such L
0
space can be properly embedded as an order-dense Riesz
subspace of any other Archimedean Riesz space. Consequently there are many maps into L
0
spaces (368B).
But by the same token there are few maps out of them (375B, 375Kb), and those which do exist have a
variety of special properties (375A, 375I).
376A Kernel operators 457
The original version of Kwapiens theorem (Kwapien 73) was the special case of 375I in which A is the
Lebesgue measure algebra. The ideas of the proof here are mostly taken from Kalton, Peck & Roberts
84. I have based my account on the concept of subhomomorphism (375E); this seems to be an eective
tool when B is weakly (, )-distributive (375Yc), but less useful in other cases. The case B = 0, 1,
L
0
(B)

= R is not entirely trivial and is worth working through on its own (375Xg).
I mention 375C for the sake of its corollary 375D, but have to admit that I am not sure it is the best
possible result. For the cases of principal interest to a measure theorist, there is something a good deal
stronger (375Yd). Further questions concern possible relaxations of the hypotheses of 375B. One has a
straightforward resolution (375Yb); others can be reduced to the Banach-Ulam problem by the techniques
of 363S (375Ya).
376 Kernel operators
The theory of linear integral equations is in large part the theory of operators T dened from formulae
of the type
(Tf)(y) =
_
k(x, y)f(x)dx
for some function k of two variables. I make no attempt to study the general theory here. However, the
concepts developed in this book make it easy to discuss certain aspects of such operators dened between
the function spaces of measure theory, meaning spaces of equivalence classes of functions, and indeed allow
us to do some of the work in the abstract theory of Riesz spaces, omitting all formal mention of measures
(376D, 376H, 376P). I give a very brief account of two theorems characterizing kernel operators in the
abstract (376E, 376H), with corollaries to show the form these theorems can take in the ordinary language
of integral kernels (376J, 376N). To give an idea of the kind of results we can hope for in this area, I go a
bit farther with operators with domain L
1
(376Mb, 376P, 376S).
I take the opportunity to spell out versions of results from 253 in the language of this volume (376B-
376C).
376A Kernel operators To give an idea of where this section is going, I will try to describe the central
idea in a relatively concrete special case. Let (X, , ) and (Y, T, ) be -nite measure spaces; you can
take them both to be [0, 1] with Lebesgue measure if you like. Let be the product measure on X Y .
If k L
1
(), then
_
k(x, y)dx is dened for almost every y, by Fubinis theorem; so if f L

() then
g(y) =
_
k(x, y)f(x)dx is dened for almost every y. Also
_
g(y)dy =
_
k(x, y)f(x)dxdy
is dened, because (x, y) k(x, y)f(x) is -virtually measurable, dened -a.e. and is dominated by a
multiple of the integrable function k. Thus k denes a function from L

() to L
1
(). Changing f on a set
of measure 0 will not change g, so we can think of this as an operator from L

() to L
1
(); and of course we
can move immediately to the equivalence class of g in L
1
(), so getting an operator T
k
from L

() to L
1
().
This operator is plainly linear; also it is easy to check that T
k
T
]k]
, so that T
k
L

(L

(); L
1
()), and
that |T
k
|
_
[k[. Moreover, changing k on a -negligible set does not change T
k
, so that in fact we can
speak of T
w
for any w L
1
().
I think it is obvious, even before investigating them, that operators representable in this way will be
important. We can immediately ask what their properties will be and whether there is any straightforward
way of recognising them. We can look at the properties of the map w T
w
: L
1
() L

(L

(); L
1
()).
And we can ask what happens when L

() and L
1
() are replaced by other function spaces, dened by
extended Fatou norms or otherwise. Theorems 376E and 376H answer questions of this kind.
It turns out that the formula g(y) =
_
k(x, y)f(x)dx gives rise to a variety of technical problems, and
it is much easier to characterize Tu in terms of its action on the dual. In the language of the special case
above, if h L

(), then we shall have

_
k(x, y)f(x)h(y)d(x, y) =
_
g(y)h(y)dy;
458 Linear operators between function spaces 376A
since g

L
1
() is entirely determined by the integrals
_
g(y)h(y)dy as h runs over L

(), we can dene

the operator T in terms of the functional (f, h)
_
k(x, y)f(x)h(y)d(x, y). This enables us to extend the
results from the case of -nite spaces to general strictly localizable spaces; perhaps more to the point in the
present context, it gives them natural expressions in terms of function spaces dened from measure algebras
rather than measure spaces, as in 376E.
Before going farther along this road, however, I give a couple of results relating the theorems of 253 to
the methods of this volume.
376B The canonical map L
0
L
0
L
0
: Proposition Let (A, ) and (B, ) be semi-nite measure
algebras, and (C,

) their localizable measure algebra free product (325E). Then we have a bilinear map
(u, v) u v : L
0
(A) L
0
(B) L
0
(C) with the following properties.
(a) For any u L
0
(A), v L
0
(B), R,
[[u 1
B
> ]] = [[u > ]] 1
B
, [[1
A
v > ]] = 1
A
[[v > ]]
where for a A, b B I write a b for the corresponding member of A B (315M), identied with a
subalgebra of C (325Dc).
(b)(i) For any u L
0
(A)
+
, the map v u v : L
0
(B) L
0
(C) is an order-continuous multiplicative
Riesz homomorphism.
(ii) For any v L
0
(B)
+
, the map u u v : L
0
(A) L
0
(C) is an order-continuous multiplicative
Riesz homomorphism.
(c) In particular, [u v[ = [u[ [v[ for all u L
0
(A), v L
0
(B).
(d) For any u L
0
(A)
+
and v L
0
(B)
+
, [[u v > 0]] = [[u > 0]] [[v > 0]].
proof The canonical maps a a 1
B
, b 1
A
b from A, B to C are order-continuous Boolean
homomorphisms (325Da), so induce order-continuous multiplicative Riesz homomorphisms from L
0
(A) and
L
0
(B) to L
0
(C) (364R); write u, v for the images of u L
0
(A), v L
0
(B). Observe that [ u[ = [u[

,
[ v[ = [v[

and (1
A
)

= (1
B
)

= 1
C
. Now set u v = u v. The properties listed in (a)-(c) are just a
matter of putting the denition in 364Ra together with the fact that L
0
(C) is an f-algebra (364E). As for
[[u v > 0]] = [[ u v > 0]], this is (for non-negative u, v) just
[[ u > 0]] [[ v > 0]] = ([[u > 0]] 1
B
) (1
A
[[v > 0]]) = [[u > 0]] [[v > 0]].
376C For L
1
spaces we have a similar result, with additions corresponding to the Banach lattice
structures of the three spaces.
TheoremLet (A, ) and (B, ) be semi-nite measure algebras with localizable measure algebra free product
(C,

).
(a) If u L
1

= L
1
(A, ) and v L
1

= L
1
(B, ) then u v L
1

= L
1
(C,

) and
_
u v =
_
u
_
v, |u v|
1
= |u|
1
|v|
1
.
(b) Let W be a Banach space and : L
1

L
1

W a bounded bilinear map. Then there is a unique
bounded linear operator T : L
1

W such that T(uv) = (u, v) for all u L

1

and v L
1

, and |T| = ||.
(c) Suppose, in (b), that W is a Banach lattice. Then
(i) T is positive i (u, v) 0 for all u, v 0;
(ii) T is a Riesz homomorphism i u (u, v
0
) : L
1

W, v (u
0
, v) : L
1

W are Riesz
homomorphisms for all v
0
0 in L
1

and u
0
0 in L
1

.
proof (a) I refer to the proof of 325D. Let (X, , ) and (Y, T, ) be the Stone spaces of (A, ) and (B, )
(321K), so that (C,

) can be identied with the measure algebra of the c.l.d. product measure on X Y
(part (a) of the proof of 325D), and L
1

, L
1

, L
1

can be identied with L

1
(), L
1
() and L
1
() (365B). Now
if f L
0
() and g L
0
() then f g L
0
() (253Cb), and it is easy to check that (f g)

L
0
(

)
corresponds to f

as dened in 376B. (Look rst at the cases in which one of f, g is a constant function
with value 1.) By 253E, we have a canonical map (f

, g

) (f g)

from L
1
() L
1
() to L
1
(), with
_
f g =
_
f
_
g (253D); so that if u L
1

and v L
1

we must have u v L
1

, with
_
u v =
_
u
_
v.
As in 253E, it follows that |u v|
1
= |u|
1
|v|
1
.
376D Kernel operators 459
(b) In view of the situation described in (a) above, this is now just a translation of the same result about
L
1
(), L
1
() and L
1
(), which is Theorem 253F.
(c) Identifying the algebraic free product AB with its canonical image in C (325Dc), I write (AB)
f
for c : c A B,

c < , so that (A B)
f
is a subring of C. Recall that any member of A B is
expressible as sup
in
a
i
b
i
where a
0
, . . . , a
n
are disjoint (315Na); evidently this will belong to (A B)
f
i a
i
b
i
is nite for every i.
The next fact to lift from previous theorems is in part (e) of the proof of 253F: the linear span M of
(a b) : a A
f
, b B
f
is norm-dense in L
1

. Of course M can also be regarded as the linear span of

c : c (AB)
f
, or S(AB)
f
. (Strictly speaking, this last remark relies on 361J; the identity map from
(A B)
f
to C induces an injective Riesz homomorphism from S(A B)
f
into S(C) L
0
(C). To see that
c M for every c (AB)
f
, we need to know that c can be expressed as a disjoint union of members of
A B, as noted above.)
(i) If T is positive then of course (u, v) = T(u v) 0 whenever u, v 0, since u v 0. On
the other hand, if is non-negative on U
+
V
+
, then, in particular, T(a b) = (a, b) 0 whenever
a b < . Consequently T(c) 0 for every c (AB)
f
and Tw 0 whenever w 0 in M

= S(AB)
f
,
as in 361Ga.
Now this means that T[w[ 0 whenever w M. But as M is norm-dense in L
1

, w T[w[ is continuous
and W
+
is closed, it follows that T[w[ 0 for every w L
1

, that is, that T is positive.

(ii) If T is a Riesz homomorphism then of course u (u, v
0
) = T(u v
0
), v (u
0
, v) = T(u
0
v)
are Riesz homomorphisms for v
0
, u
0
0. On the other hand, if all these maps are Riesz homomorphisms,
then, in particular,
T(a b) T(a
t
b
t
) = (a, b) (a
t
, b
t
)
(a, b +b
t
) (a
t
, b +b
t
)
= (a a
t
, b +b
t
) = 0
whenever a, a
t
A
f
, b, b
t
B
f
and a a
t
= 0. Similarly, T(a b) T(a
t
b
t
) = 0 if b b
t
= 0. But this
means that Tc Tc
t
= 0 whenever c, c
t
(A B)
f
and c c
t
= 0. PPP Express c, c
t
as sup
im
a
i
b
i
,
sup
jn
a
t
j
b
t
j
where a
i
, a
t
j
, b
i
, b
t
j
all have nite measure. Now if i m, j n, (a
i
a
t
j
) (b
i
b
t
j
) =
(a
i
b
i
) (a
t
j
b
t
j
) = 0, so one of a
i
a
t
j
, b
i
b
t
j
must be zero, and in either case T(a
i
b
i
)T(a
t
j
b
t
j
) = 0.
Accordingly
Tc Tc
t
(
m

i=0
T(a
i
b
i
)) (
n

j=0
T(a
t
j
b
t
j
))

im,jn
T(a
i
b
i
) T(a
t
j
b
t
j
) = 0,
using 352Fa for the second inequality. QQQ
This implies that TM must be a Riesz homomorphism (361Gc), that is, T[w[ = [Tw[ for all w M.
Again because M is dense in L
1

, T[w[ = [Tw[ for every w L

1

, and T is a Riesz homomorphism.

376D Abstract integral operators: Denition The following concept will be used repeatedly in
the theorems below; it is perhaps worth giving it a name. Let U be a Riesz space and V a Dedekind
complete Riesz space, so that L

(U; V ) is a Dedekind complete Riesz space (355H). If f U

and v V
write P
fv
u = f(u)v for each u U; then P
fv
L

(U; V ). PPP If f 0 in U

and v 0 in V

then
P
fv
is a positive linear operator from U to V which is order-continuous because if A U is non-empty,
downwards-directed and has inmum 0, then (as V is Archimedean)
inf
uA
P
fv
(u) = inf
uA
f(u)v = 0.
Of course (f, g) P
fg
is bilinear, so P
fv
L

(U; V ) for every f U

, v V . QQQ Now I call a linear

operator from U to V an abstract integral operator if it is in the band of L

(U; V ) generated by
P
fv
: f U

, v V .
460 Linear operators between function spaces 376D
The rst result describes these operators when U, V are expressed as subspaces of L
0
(A), L
0
(B) for
measure algebras A, B and V is perfect.
376E Theorem Let (A, ) and (B, ) be semi-nite measure algebras, with localizable measure algebra
free product (C,

), and U L
0
(A), V L
0
(B) order-dense Riesz subspaces. Write W for the set of
those w L
0
(C) such that w (u v) is integrable for every u U, v V . Then we have an operator
w T
w
: W L

(U; V

) dened by setting
T
w
(u)(v) =
_
w (u v)
for every w W, u U and v V . The map w T
w
is a Riesz space isomorphism between W and the
band of abstract integral operators in L

(U; V

).
proof (a) The rst thing to check is that the formula oered does dene a member T
w
(u) of V

for any
w W, u U. PPP Of course T
w
(u) is a linear operator because
_
is linear and and are bilinear. It
belongs to V

because, writing g(v) =

_
[w[ ([u[ v), g is a positive linear operator and [T
w
(u)(v)[ g([v[)
for every v. (I am here using 376Bc to see that [w (u v)[ = [w[ ([u[ [v[).) Also g V

because
v [u[ v, w
t
[w[ w
t
and
_
are all order-continuous; so T
w
(u) also belongs to V

. QQQ
(b) Next, for any given w W, the map T
w
: U V

is linear (again because and are bilinear). It

is helpful to note that W is a solid linear subspace of L
0
(C). Now if w 0 in W, then T
w
L

(U; V

). PPP
If u, v 0 then u v 0, w (u v) 0 and T
w
(u)(v) 0; as v is arbitrary, T
w
(u) 0 whenever u 0;
as u is arbitrary, T
w
is positive. If A U is non-empty, downwards-directed and has inmum 0, then T
w
[A]
is downwards-directed, and for any v V
+
(inf T
w
[A])(v) = inf
uA
T
w
(u)(v) = inf
uA
_
w (u v) = 0
because u u v is order-continuous. So inf T
w
[A] = 0; as A is arbitrary, T
w
is order-continuous. QQQ
For general w W, we now have T
w
= T
w
+ T
w
L

(U; V

).
(c) Ths shows that w T
w
is a map from W to L

(U; V

). Running through the formulae once again,

it is linear, positive and order-continuous; this last because, given a non-empty downwards-directed C W
with inmum 0, then for any u U
+
, v V
+
(inf
wC
T
w
)(u)(v) inf
wC
_
w (u v) = 0
(because
_
and are order-continuous); as v is arbitrary, (inf
wC
T
w
)(u) = 0; as u is arbitrary, inf
wC
T
w
=
0.
(d) All this is easy, being nothing but a string of applications of the elementary properties of ,
and
_
. But I think a new idea is needed for the next fact: the map w T
w
: W L

(U; V

) is
a Riesz homomorphism. PPP Write D for the set of those d C such that T
w
T
w
= 0 whenever w,
w
t
W
+
, [[w > 0]] d and [[w
t
> 0]] 1
C
\ d. (i) If d
1
, d
2
D, w, w
t
W
+
, [[w > 0]] d
1
d
2
and
[[w
t
> 0]] (d
1
d
2
) = 0, then set w
1
= w d
1
, w
2
= w w
1
. In this case
[[w
1
> 0]] d
1
, [[w
2
> 0]] d
2
,
so
T
w
1
T
w
= T
w
2
T
w
= 0, T
w
T
w
(T
w
1
T
w
) + (T
w
2
T
w
) = 0.
As w, w
t
are arbitrary, d
1
d
2
D. Thus D is closed under . (ii) The symmetry of the denition of
D means that 1
C
\ d D whenever d D. (iii) Of course 0 D, just because T
w
= 0 if w W
+
and
[[w > 0]] = 0; so D is a subalgebra of C. (iv) If D D is non-empty and upwards-directed, with supremum c
in C, and if w, w
t
W
+
are such that [[w > 0]] c, [[w
t
> 0]] c = 0, then consider wd : d D. This is
upwards-directed, with supremum w; so T
w
= sup
dD
T
wd
, because the map q T
q
is order-continuous.
Also T
wd
T
w
= 0 for every d D, so T
w
T
w
= 0. As w, w
t
are arbitrary, c D; as D is arbitrary,
D is an order-closed subalgebra of C. (v) If a A and w, w
t
W
+
are such that [[w > 0]] a 1
B
,
[[w
t
> 0]] (a 1
B
) = 0, then any u U
+
is expressible as u
1
+ u
2
where u
1
= u a, u
2
= u (1
A
\ a).
Now
T
w
(u
2
)(v) =
_
w (u
2
v) =
_
w (a 1
B
) (u v) ((1
A
\ a) 1
B
) = 0
376E Kernel operators 461
for every v V , so T
w
(u
2
) = 0. Similarly, T
w
(u
1
) = 0. But this means that
(T
w
T
w
)(u) T
w
(u
2
) +T
w
(u
1
) = 0.
As u is arbitrary, T
w
T
w
= 0; as w and w
t
are arbitrary, a 1
B
D. (vi) Now suppose that b B and
that w, w
t
W
+
are such that [[w > 0]] 1
A
b, [[w
t
> 0]] (1
A
b) = 0. If u U
+
, v V
+
then
(T
w
T
w
)(u)(v)
_
w (u (v (1
B
\ b))) +
_
w
t
(u (v b)) = 0.
As u, v are arbitrary, T
w
T
w
= 0; as w and w
t
are arbitrary, 1
A
b D. (vii) This means that D is
an order-closed subalgebra of C including A B, and is therefore the whole of C (325D(c-ii)). (viii) Now
take any w, w
t
W such that w w
t
= 0, and consider c = [[w > 0]]. Then [[w
t
> 0]] 1
C
\ c and c D, so
T
w
T
w
= 0. This is what we need to be sure that w T
w
is a Riesz homomorphism (352G). QQQ
(e) The map w T
w
is injective. PPP (i) If w > 0 in W, then consider
A = a : a A, u U, a u, B = b : b B, v V, b v.
Because U and V are order-dense in L
0
(A) and L
0
(B) respectively, A and B are order-dense in A and B.
Also both are upwards-directed. So sup
aA,bB
a b = 1
C
and 0 <
_
w = sup
aA,bB
_
ab
w. Take a A,
b B such that
_
ab
w > 0; then there are u U, v V such that a u, b v, so that
T
w
(u)(v)
_
ab
w > 0
and T
w
> 0. (ii) For general non-zero w W, we now have [T
w
[ = T
]w]
> 0 so T
w
,= 0. QQQ
Thus w T
w
is an order-continuous injective Riesz homomorphism.
(f ) Write

W for T
w
: w W, so that

W is a Riesz subspace of L

(U; V

) isomorphic to W, and

W for the band it generates in L

(U; V

). Then

W is order-dense in

W. PPP Suppose that S > 0 in

W =

W

(353Ba). Then S /

W

, so there is a w W such that S T

w
> 0. Set w
1
= w 1
C
. Then
w = sup
nN
w nw
1
, so T
w
= sup
nN
T
w
nT
w
1
and R = S T
w
1
> 0.
Set U
1
= U L
1
(A, ). Because U is an order-dense Riesz subspace of L
0
(A), U
1
is an order-dense Riesz
subspace of L
1

= L
1
(A, ), therefore also norm-dense. Similarly V
1
= V L
1
(B, ) is a norm-dense Riesz
subspace of L
1

= L
1
(B, ). Dene
0
: U
1
V
1
R by setting
0
(u, v) = R(u)(v) for u U
1
, v V
1
. Then

0
is bilinear, and
[
0
(u, v)[ = [R(u)(v)[ [R(u)[([v[) R([u[)([v[) T
w
1
([u[)([v[)
=
_
w
1
([u[ [v[)
_
[u[ [v[ = |u|
1
|v|
1
for all u U
1
, v V
1
, because 0 R T
w
1
in L

(U; V

). Because U
1
, V
1
are norm-dense in L
1

, L
1

respectively,
0
has a unique extension to a continuous bilinear map : L
1

L
1

R. (To reduce this to
standard results on linear operators, think of R as a function from U
1
to V

1
; since every member of V

1
has a unique extension to a member of (L
1

)

, we get a corresponding function R

1
: U
1
(L
1

)

which is
continuous and linear, so has a unique extension to a continuous linear operator R
2
: L
1

(L
1

)

, and we
set (u, v) = R
2
(u)(v).)
By 376C, there is a unique h (L
1

= L
1
(C,

such that h(u v) = (u, v) for every u L

1

, v L
1

.
Because (C,

) is localizable, this h corresponds to a w

t
L

(C) (365Jc), and

_
w
t
(u v) = h(u v) =
0
(u, v) = R(u)(v)
for every u U
1
, v V
1
.
Because U
1
is norm-dense in L
1

, U
+
1
is dense in (L
1

)
+
, and similarly V
+
1
is dense in (L
1

)
+
, so U
+
1
V
+
1
is dense in (L
1

)
+
(L
1

)
+
; now
0
is non-negative on U
+
1
V
+
1
, so (being continuous) is non-negative on
(L
1

)
+
(L
1

)
+
. By 376Cc, h 0 in (L
1

and w
t
0 in L

(C). In the same way, because

0
(u, v) T
w
(u)(v)
for u U
+
1
and v V
+
1
, w
t
w
1
w in L
0
(C), so w
t
W. We have
T
w
(u)(v) =
_
w
t
(u v) = R(u)(v)
for all u U
1
, v V
1
. If u U
+
1
, then T
w
(u) and R(u) are both order-continuous, so must be identical, since
V
1
is order-dense in V . This means that T
w
and R agree on U
1
. But as both are themselves order-continuous
462 Linear operators between function spaces 376E
linear operators, and U
1
is order-dense in U, they must be equal.
Thus 0 < T
w
S in L

(U; V

). As S is arbitrary,

W is quasi-order-dense in

W, therefore order-dense
(353A). QQQ
(g) Because w T
w
: W

W is an injective Riesz homomorphism, we have an inverse map Q :

W L
0
(C), setting Q(T
w
) = w; this is a Riesz homomorphism, and it is order-continuous because W is
solid in L
0
(C), so that the embedding W L
0
(C) is order-continuous. By 368B, Q has an extension to
an order-continuous Riesz homomorphism

Q :

W L
0
(C). Because Q(S) > 0 whenever S > 0 in

W,

Q(S) > 0 whenever S > 0 in

W, so

Q is injective. Now

Q(S) W for every S

W. PPP It is enough
to look at non-negative S. In this case,

Q(S) must be sup

Q(T
w
) : w W, T
w
S = supC, where
C = w : T
w
S W. Take u U
+
, v V
+
. Then w (u v) : w C is upwards-directed, because
C is, and
sup
wC
_
w (u v) = sup
wC
T
w
(u)(v) S(u)(v) < .
So

Q(S) (u v) = sup
wC
w (u v) belongs to L
1

(365Df). As u and v are arbitrary,

Q(S) W. QQQ
(h) Of course this means that

W =

W and

Q = Q, that is, that w T
w
: W

W is a Riesz space
isomorphism.
(i) I have still to check on the identication of

W as the band Z of abstract integral operators in
L

(U; V

). Write P
fg
(u) = f(u)g for f U

, g V

and u U.
Set
U
#
= u : u L
0
(A), u u
t
L
1

for every u
t
U,
V
#
= v : v L
0
(B), v v
t
L
1

for every v
t
V .
From 369C we know that if we set f
u
(u
t
) =
_
uu
t
for u U
#
and u
t
U, then f
u
U

for every u U
#
,
and u f
u
is an isomorphism between U
#
and an order-dense Riesz subspace of U

. Similarly, setting
g
v
(v
t
) =
_
v v
t
for v V
#
and v
t
V , v g
v
is an isomorphism between V
#
and an order-dense Riesz
subspace of V

.
If u U
#
, v V
#
then
_
(u v) (u
t
v
t
) =
_
(u u
t
) (v v
t
) = (
_
u u
t
)(
_
v v
t
) = f
u
(u
t
)g
v
(v
t
)
for every u
t
U, v
t
V , so u v W and T
uv
= P
f
u
g
v
.
Now take f (U

)
+
and g (V

)
+
. Set A = u : u U
#
, u 0, f
u
f and B = v : v V
#
, v
0, g
v
g. These are upwards-directed, so C = u v : u A, v B is upwards-directed in L
0
(C).
Because f
u
: u U
#
is order-dense in U

, f = sup
uA
f
u
; by 355Ed, f(u
t
) = sup
uA
f
u
(u
t
) for every
u
t
U
+
. Similarly, g(v
t
) = sup
vB
f
v
(v
t
) for every v
t
V
+
.
??? Suppose, if possible, that C is not bounded above in L
0
(C). Because C and L
0
(C) are Dedekind
complete,
c = inf
nN
sup
uA,vB
[[u v n]]
must be non-zero (364Ma). Because U and V are order-dense in L
0
(A), L
0
(B) respectively,
1
A
= sup[[u
t
> 0]] : u
t
U, 1
B
= sup[[v
t
> 0]] : v
t
V ,
and there are u
t
U
+
, v
t
V
+
such that c [[u
t
> 0]] [[v
t
> 0]] ,= 0, so that
_
c
u
t
v
t
> 0. But now, for
any n N,
376E Kernel operators 463
f(u
t
)g(v
t
) sup
uA,vB
f
u
(u
t
)g
v
(v
t
)
= sup
uA,vB
_
(u v) (u
t
v
t
)
sup
uA,vB
_
((u v) nc) (u
t
v
t
)
=
_
sup
uA,vB
((u v) nc) (u
t
v
t
)
(because w
_
w (u
t
v
t
) is order-continuous)
=
_
(nc) (u
t
v
t
) = n
_
c
u
t
v
t
,
which is impossible. XXX
Thus C is bounded above in L
0
(C), and has a supremum w L
0
(C). If u
t
U
+
, v
t
V
+
then
_
w (u
t
v
t
) = sup
uA,vB
_
(u v) (u
t
v
t
)
= sup
uA,vB
f
u
(u
t
)g
v
(v
t
) = f(u
t
)g(v
t
) = P
fg
(u
t
)(v
t
).
Thus w W and
P
fg
= T
w


W

W.
And this is true for any non-negative f U

, g V

. Of course it follows that P

fg

W for every f U

,
g V

; as

W is a band, it must include Z.
(j) Finally,

W Z. PPP Since Z = Z

, it is enough to show that

W Z

= 0. Take any T > 0 in

W. There are u
t
0
U
+
, v
t
0
V
+
such that T(u
t
0
)(v
t
0
) > 0. So there is a v V
#
such that 0 g
v
T(u
t
0
)
and g
v
(v
t
0
) > 0, that is,
_
v v
t
0
> 0. Because V is order-dense in L
0
(B), there is a v
t
1
V such that
0 < v
t
1
v
t
0
[[v > 0]], so that
0 <
_
v v
t
1
= g
v
(v
t
1
) T(u
t
0
)(v
t
1
)
and [[v
t
1
> 0]] [[v > 0]].
Now consider the functional u
t
h(u
t
) = T(u
t
)(v
t
1
) : U R. This belongs to (U

)
+
and h(u
t
0
) > 0,
so there is a u U
#
such that 0 f
u
h and f
u
(u
t
0
) > 0. This time,
_
u u
t
0
> 0 so (because U is
order-dense in L
0
(A)) there is a u
t
1
U such that h(u
t
1
) > 0 and [[u
t
1
> 0]] [[u > 0]].
We can express T as T
w
where w W
+
. In this case, we have
_
w (u
t
1
v
t
1
) = T(u
t
1
)(v
t
1
) = h(u
t
1
) > 0,
so
0 ,= [[w > 0]] [[u
t
1
v
t
1
> 0]] = [[w > 0]] ([[u
t
1
> 0]] [[v
t
1
> 0]])
[[w > 0]] ([[u > 0]] [[v > 0]]) = [[w > 0]] [[u v > 0]],
and w (u v) > 0, so
T
w
P
f
u
g
v
= T
w
T
uv
= T
w(uv)
> 0.
Thus T / Z

. Accordingly

W Z

= 0 and

W Z

= Z. QQQ
Since we already know that Z

W, this completes the proof.

464 Linear operators between function spaces 376F
376F Corollary Let (A, ) and (B, ) be localizable measure spaces, with localizable measure algebra
free product (C,

). Let U L
0
(A), V L
0
(B) be perfect order-dense solid linear subspaces, and T : U V
a linear operator. Then the following are equiveridical:
(i) T is an abstract integral operator;
(ii) there is a w L
0
(C) such that
_
w (u v
t
) is dened and equal to
_
Tu v
t
whenever u U and
v
t
L
0
(B) is such that v
t
v is integrable for every v V .
proof Setting V
#
= v
t
: v
t
L
0
(B), v v
t
L
1
for every v V , we know that we can identify V
#
with
V

and V with (V
#
)

(369C). So the equivalence of (i) and (ii) is just 376E applied to V

#
in place of V .
376G Lemma Let U be a Riesz space, V an Archimedean Riesz space, T : U V a linear operator,
f (U

)
+
and e V
+
. Suppose that 0 Tu f(u)e for every u U
+
. Then if u
n
)
nN
is a sequence
in U such that lim
n
g(u
n
) = 0 whenever g U

and [g[ f, Tu
n
)
nN
order*-converges to 0 in V
(denition: 367A).
proof Let V
e
be the solid linear subspace of V generated by e; then Tu V
e
for every u U. We can identify
V
e
with an order-dense and norm-dense Riesz subspace of C(X), where X is a compact Hausdor space,
with e corresponding to X (353M). For x X, set g
x
(u) = (Tu)(x) for every u U; then 0 g
x
(u) f(u)
for u 0, so [g
x
[ f and lim
n
(Tu
n
)(x) = 0. As x is arbitrary, Tu
n
)
nN
order*-converges to 0 in C(X),
by 367L, and therefore in V
e
, because V
e
is order-dense in C(X) (367F). But V
e
, regarded as a subspace of
V , is solid, so 367F also tells us that Tu
n
)
nN
order*-converges to 0 in V .
376H Theorem Let U be a Riesz space and V a weakly (, )-distributive Dedekind complete Riesz
space (denition: 368N). Suppose that T L

(U; V ). Then the following are equiveridical:

(i) T is an abstract integral operator;
(ii) whenever u
n
)
nN
is an order-bounded sequence in U
+
and lim
n
f(u
n
) = 0 for every f U

,
then Tu
n
)
nN
order*-converges to 0 in V ;
(iii) whenever u
n
)
nN
is an order-bounded sequence in U and lim
n
f(u
n
) = 0 for every f U

, then
Tu
n
)
nN
order*-converges to 0 in V .
proof For f U

, v V and u U set P
fv
(u) = f(u)v. Write Z L

(U; V ) for the space of abstract

integral operators.
(a)(i)(iii) Suppose that T Z
+
, and that u
n
)
nN
is an order-bounded sequence in U such that
lim
n
f(u
n
) = 0 for every f U

. Note that P
fv
: f U
+
, v V
+
is upwards-directed, so that
T = supT P
fv
: f U
+
, v V
+
(352Va).
Take u

U
+
such that [u
n
[ u

for every n, and set w = inf

nN
sup
mn
Tu
m
, which is dened because
[Tu
n
[ Tu

for every n. Now w (T P

fv
)
+
(u

) for every f U
+
, v V
+
. PPP Setting T
1
= T P
fv
,
w
0
= (T P
fv
)
+
(u

) we have
Tu
n
T
1
u
n
[T T
1
[(u

) = (T P
fv
)
+
(u

) = w
0
for every n N, so Tu
n
w
0
+ T
1
u
n
. On the other hand, 0 T
1
u f(u)v for every u U
+
, so by 376G
we must have inf
nN
sup
mn
T
1
u
m
= 0. Accordingly
w w
0
+ inf
nN
sup
mn
T
1
u
m
= w
0
. QQQ
But as inf(T P
fv
)
+
: f U
+
, v V
+
= 0, w 0. Similarly (or applying the same argument to
u
n
)
nN
), sup
nN
inf
nN
Tu
n
0 and Tu
n
)
nN
order*-converges to zero.
For general T Z, this shows that T
+
u
n
)
nN
and T

u
n
)
nN
both order*-converge to 0, so Tu
n
)
nN
order*-converges to 0, by 367C. As u
n
)
nN
is arbitrary, (iii) is satised.
(b)(iii)(ii) is trivial.
(c)(ii)(i) ??? Now suppose, if possible, that (ii) is satised, but that T / Z. Because L

(U; V ) is
Dedekind complete (355H), Z is a projection band (353I), so T is expressible as T
1
+ T
2
where T
1
Z,
T
2
Z

and T
2
,= 0. At least one of T
+
2
, T

2
is non-zero; replacing T by T if need be, we may suppose
that T
+
2
> 0.
376I Kernel operators 465
Because T
+
2
, like T, belongs to L

(U; V ), its kernel U

0
is a band in U, which cannot be the whole of U,
and there is a u
0
> 0 in U

0
. In this case T
+
2
u
0
> 0; because T
+
2
(T

2
+ [T
1
[) = 0, there is a u
1
[0, u
0
]
such that T
+
2
(u
0
u
1
) + (T

2
+[T
1
[)(u
1
) , T
+
2
u
0
, so that
Tu
1
T
2
u
1
[T
1
[(u
1
) , 0
and Tu
1
,= 0. Now this means that the sequence (Tu
1
, Tu
1
, . . . ) is not order*-convergent to zero, so there
must be some f U

such that (f(u

1
), f(u
1
), . . . ) does not converge to 0, that is, f(u
1
) ,= 0; replacing f
by [f[ if necessary, we may suppose that f 0 and that f(u
1
) > 0.
By 356H, there is a u
2
such that 0 < u
2
u
1
and g(u
2
) = 0 whenever g U

and g f = 0. Because
0 < u
2
u
0
, u
2
U

0
and v
0
= T
+
2
u
2
> 0. Consider P
fv
0
Z. Because T
2
Z

, T
+
2
P
fv
0
= 0; set
S = P
fv
0
+T

2
, so that T
+
2
S = 0. Then
inf
u[0,u
2
]
T
+
2
(u
2
u) +Su = 0, sup
u[0,u
2
]
T
+
2
u Su = v
0
(use 355Ec for the rst equality, and then subtract both sides from v
0
). Now Su f(u)v
0
for every u 0,
so that for any > 0
sup
u[0,u
2
],f(u)
T
+
2
u Su (1 )v
0
and accordingly
sup
u[0,u
2
],f(u)
T
+
2
u = v
0
,
since the join of these two suprema is surely at least v
0
, while the second is at most v
0
. Note also that
v
0
= sup
u[0,u
2
],f(u)
T
+
2
u = sup
0u

uu
2
,f(u)
T
2
u
t
= sup
0u

u
2
,f(u

)
T
2
u
t
.
For k N set A
k
= u : 0 u u
2
, f(u) 2
k
. We know that
B
k
= sup
uI
T
2
u : I A
k
is nite
is an upwards-directed set with supremum v
0
for each k. Because V is weakly (, )-distributive, we can
nd a sequence v
t
k
)
kN
such that v
t
k
B
k
for every k and v
1
= inf
kN
v
t
k
> 0. For each k let I
k
A
k
be a
nite set such that v
t
k
= sup
uI
k
T
2
u.
Because each I
k
is nite, we can build a sequence u
t
n
)
nN
in [0, u
2
] enumerating each in turn, so that
lim
n
f(u
t
n
) = 0 (since f(u) 2
k
if u I
k
) while sup
mn
T
2
u
t
m
v
1
for every n (since u
t
m
: m n
always includes some I
k
). Now T
2
u
t
n
)
nN
does not order*-converge to 0.
However, lim
n
g(u
t
n
) = 0 for every g U

. PPP Express [g[ as g

1
+ g
2
where g
1
belongs to the band
of U

generated by f and g
2
f = 0 (353Hc). Then g
2
(u
t
n
) = g
2
(u
2
) = 0 for every n, by the choice of u
2
.
Also g
1
= sup
nN
g
1
nf (352Vb); so, given > 0, there is an m N such that (g
1
mf)
+
(u
2
) and
(g
1
mf)
+
(u
t
n
) for every n N. But this means that
[g(u
t
n
)[ [g[(u
t
n
) +mf(u
t
n
)
for every n, and limsup
n
[g(u
t
n
)[ ; as is arbitrary, lim
n
g(u
t
n
) = 0. QQQ
Now, however, part (a) of this proof tells us that T
1
u
t
n
)
nN
is order*-convergent to 0, because T
1
Z,
while Tu
t
n
)
nN
is order*-convergent to 0, by hypothesis; so T
2
u
t
n
)
nN
= Tu
t
n
T
1
u
t
n
)
nN
order*-converges
to 0. XXX
This contradiction shows that every operator satisfying the condition (ii) must be in Z.
376I The following elementary remark will be useful for the next corollary and also for Theorem 376S.
Lemma Let (X, , ) be a -nite measure space and U an order-dense solid linear subspace of L
0
(). Then
there is a non-decreasing sequence X
n
)
nN
of measurable subsets of X, with union X, such that X

n
U
for every n N.
proof Write A for the measure algebra of , so that L
0
() can be identied with L
0
(A) (364Jc). A = a :
a A 0, a U is order-dense in A, so includes a partition of unity a
i
)
iI
. Because is -nite, A is
ccc (322G) and I is countable, so we can take I to be a subset of N. Choose E
i
such that E

i
= a
i
for
i I; set E = X

iI
E
i
, X
n
= E

iI,in
E
i
for n N.
466 Linear operators between function spaces 376J
376J Corollary Let (X, , ) and (Y, T, ) be -nite measure spaces, with product measure on
X Y . Let U L
0
(), V L
0
() be perfect order-dense solid linear subspaces, and T : U V a linear
operator. Write U = f : f L
0
(), f

U, V
#
= h : h L
0
(), h

v L
1
for every v V . Then
the following are equiveridical:
(i) T is an abstract integral operator;
(ii) there is a k L
0
() such that
()
_
[k(x, y)f(x)h(y)[d(x, y) < for every f U, h V
#
,
() if f U and we set g(y) =
_
k(x, y)f(x)dx wherever this is dened, then g L
0
() and Tf

= g

;
(iii) T L

(U; V ) and whenever u

n
)
nN
is an order-bounded sequence in U
+
and lim
n
h(u
n
) = 0 for
every h U

, then Tu
n
)
nN
order*-converges to 0 in V .
Remark I write d(x, y) above to indicate integration with respect to the product measure . Recall that
in the terminology of 251, can be taken to be either the primitive or c.l.d. product measure (251K).
proof The idea is of course to identify L
0
() and L
0
() with L
0
(A) and L
0
(B), where (A, ) and (B, )
are the measure algebras of and , so that their localizable measure algebra free product can be identied
with the measure algebra of (325Eb), while V
#
= h

: h V
#
can be identied with V

.
(a)(i)(ii) By 376F, there is a w L
0
() such that
_
w(uv
t
) is dened and equal to
_
Tuv
t
for every
u U, v
t
V
#
. Express w as k

where k L
0
(). If f U and h V
#
then
_
[k(x, y)f(x)h(y)[d(x, y) =
_
[w (f

[ is nite, so (ii-) is satised.

Now take any f U, and set g(y) =
_
k(x, y)f(x)dx whenever this is dened in R. Write T for the set
of those F T such that F V
#
. Then for any F T, g is dened almost everywhere on F and gF is
-virtually measurable. PPP
_
k(x, y)f(x)F(y)d(x, y) is dened in R, so by Fubinis theorem (252B, 252C)
g
F
(y) =
_
k(x, y)f(x)F(y)dx is dened for almost every y, and is -virtually measurable; now gF = g
F
F.
QQQ Next, there is a sequence F
n
)
nN
in T with union Y , by 376I, because V is perfect and order-dense, so
V
#
must also be order-dense in L
0
().
For each n N, there is a measurable set F
t
n
F
n
domg such that gF
n
is measurable and F
n
F
t
n
is
negligible. Setting G =

nN
F
t
n
, G is conegligible and gG is measurable, so g L
0
().
If g L
0
() represents Tu L
0
(), then for any F T
_
F
g =
_
Tu (F)

=
_
F
g.
In particular, this is true whenever F T and F F
n
. So g and g agree almost everywhere on F
n
, for each
n, and g = g a.e. Thus g also represents Tu, as required in (ii-).
(b)(ii)(i) Set w = k

in L
0
(). If f U and h V
#
the hypothesis () tells us that (x, y)
k(x, y)f(x)h(y) is integrable (because it surely belongs to L
0
()). By Fubinis theorem,
_
k(x, y)f(x)h(y)d(x, y) =
_
g(y)h(y)dy
where g(y) =
_
k(x, y)f(x)dx for almost every y, so that Tf

= g

, by (). But this means that, setting

u = f

, v
t
= h

,
_
w (u v
t
) =
_
Tu v
t
;
and this is true for every u U, v
t
V
#
.
Thus T satises the condition 376F(ii), and is an abstract integral operator.
(b)(i)(iii) Because V is weakly (, )-distributive (368S), this is covered by 376H(i)(iii).
(c)(iii)(i) Suppose that T satises (iii). The point is that T
+
is order-continuous. PPP??? Otherwise,
let A U be a non-empty downwards-directed set, with inmum 0, such that v
0
= inf
uA
T
+
(u) > 0.
Let X
n
)
nN
be a non-decreasing sequence of sets of nite measure covering X, and set a
n
= X

n
for each
n. For each n, inf
uA
[[u > 2
n
]] = 0, so we can nd u
n
A such that (a
n
[[ u
n
> 2
n
]]) 2
n
. Set
u
n
= inf
in
u
i
for each n; then u
n
)
nN
is non-increasing and has inmum 0; also, [0, u
n
] meets A for each
n, so that v
0
supTu : 0 u u
n
for each n. Because V is weakly (, )-distributive, we can nd
a sequence I
n
)
nN
of nite sets such that I
n
[0, u
n
] for each n and v
1
= inf
nN
sup
uI
n
(Tu)
+
> 0.
Enumerating

nN
I
n
as u
t
n
)
nN
, as in part (d) of the proof of 376H, we see that u
t
n
)
nN
is order-bounded
376M Kernel operators 467
and lim
n
f(u
t
n
) = 0 for every f U

(indeed, u
t
n
)
nN
order*-converges to 0 in U), while Tu
t
n
)
nN
, 0
in V . XXXQQQ
Similarly, T

is order-continuous, so T L

(U; V ). Accordingly T is an abstract integral operator by

condition (ii) of 376H.
376K As an application of the ideas above, I give a result due to N.Dunford (376N) which was one
of the inspirations underlying the theory. Following the method of Zaanen 83, I begin with a couple of
elementary lemmas.
Lemma Let U and V be Riesz spaces. Then there is a Riesz space isomorphism T T
t
: L

(U; V

)
L

(V ; U

) dened by the formula

(T
t
v)(u) = (Tu)(v) for every u U, v V .
If for f U

, g V

we write P
fg
(u) = f(u)g for every u U, then P
fg
L

(U; V

) and P
t
fg
= P
gf
in
L

(V ; U

). Consequently T is an abstract integral operator i T

t
is.
proof All the ideas involved have already appeared. For positive T L

(U; V

) the functional (u, v)

(Tu)(v) is bilinear and order-continuous in each variable separately; so (just as in the rst part of the proof
of 376E) corresponds to a T
t
L

(V ; U

). The map T T
t
: L

(U; V

)
+
L

(V ; U

)
+
is evidently an
additive, order-preserving bijection, so extends to an isomorphism between L

(U; V

) and L

(V ; U

) given
by the same formula. I remarked in part (i) of the proof of 376E that every P
fg
belongs to L

(U; V

), and
the identication P
t
fg
= P
gf
is just a matter of checking the formulae. Of course it follows at once that the
bands of abstract integral operators must also be matched by the map T T
t
.
376L Lemma Let U be a Riesz space with an order-continuous norm. If w U
+
there is a g U

such that for every > 0 there is a > 0 such that |u| whenever 0 u w and g(u) .
proof (a) As remarked in 356D, U

= U

= U

. Set
A = v : v U and there is an f (U

)
+
such that f(u) > 0 whenever 0 < u [v[.
Then v
t
A whenever [v
t
[ [v[ A and v +v
t
A for all v, v
t
A (if f(u) > 0 whenever 0 < u [v[ and
f
t
(u) > 0 whenever 0 < u [v
t
[, then (f +f
t
)(u) > 0 whenever 0 < u [v +v
t
[); moreover, if v
0
> 0 in U,
there is a v A such that 0 < v v
0
. PPP Because U

= U

separates the points of U, there is a g > 0 in

U

such that g(v

0
) > 0; now by 356H there is a v ]0, v
0
] such that g is strictly positive on ]0, v], so that
v A. QQQ But this means that A is an order-dense solid linear subspace of U.
(b) In fact w A. PPP w = sup B, where B = A [0, w]. Because B is upwards-directed, w B (354Ea),
and there is a sequence u
t
n
)
nN
in B converging to w for the norm. For each n, choose f
n
(U

)
+
such
that f
n
(u) > 0 whenever 0 < u u
t
n
. Set
f =

n=0
1
2
n
(1+|f
n
|)
f
n
in U

= U

. Then whenever 0 < u w there is some n N such that u u

t
n
> 0, so that f
n
(u) > 0 and
f(u) > 0. So f witnesses that w A. QQQ
(c) Take g (U

)
+
such that g(u) > 0 whenever 0 < u w. This g serves. PPP??? Otherwise, there
is some > 0 such that for every n N we can nd a u
n
[0, w] with g(u
n
) 2
n
and |u
n
| . Set
v
n
= sup
in
u
i
; then 0 v
n
w, g(v
n
) 2
n+1
and |v
n
| for every n N. But v
n
)
nN
is non-
decreasing, so v = inf
nN
v
n
must be non-zero, while 0 v w and g(v) = 0; which is impossible. XXX
QQQ
Thus we have found an appropriate g.
376M Theorem (a) Let U be a Banach lattice with an order-continuous norm and V a Dedekind
complete M-space. Then every bounded linear operator from U to V is an abstract integral operator.
(b) Let U be an L-space and V a Banach lattice with order-continuous norm. Then every bounded linear
operator from U to V

is an abstract integral operator.

468 Linear operators between function spaces 376M
proof (a) By 355Kb and 355C, L

(U; V ) = L

(U; V ) B(U; V ); but since norm-bounded sets in V are

also order-bounded, Tu : [u[ u
0
is bounded above in V for every T B(U; V ) and u
0
U
+
, and
B(U; V ) = L

(U; V ).
I repeat ideas from the proof of 376H. (I cannot quote 376H directly as I am not assuming that V is
weakly (, )-distributive.) ??? Suppose, if possible, that B(U; V ) is not the band Z of abstract integral
operators. In this case there is a T > 0 in Z

. Take u
1
0 such that Tu
1
,= 0. Let f 0 in U

be such
that for every > 0 there is a > 0 such that |u| whenever 0 u u
1
and f(u) (376L). Set
v
0
= Tu
1
. Then, just as in part (d) of the proof of 376H,
sup
u[0,u
1
],f(u)
Tu = v
0
for every > 0. But there is a > 0 such that |T||u|
1
2
|v
0
| whenever 0 u u
1
and f(u) ; in
which case | sup
u[0,u
1
],f(u)
Tu|
1
2
|v
0
|, which is impossible. XXX
Thus Z = B(U; V ), as required.
(b) Because V has an order-continuous norm, V

= V

= V

; and the norm of V

is a Fatou norm with

the Levi property (356Da). So B(U; V

) = L

(U; V

), by 371C. By 376K, this is canonically isomorphic

to L

(V ; U

). Now U

= U

is an M-space (356Pb). By (a), every member of L

(V ; U

) is an abstract
integral operator; but the isomorphism between L

(V ; U

) and L

(U; V

) matches the abstract integral

operators in each space (376K), so every member of B(U; V

) is also an abstract integral operator, as claimed.

376N Corollary: Dunfords theorem Let (X, , ) and (Y, T, ) be -nite measure spaces and
T : L
1
() L
p
() a bounded linear operator, where 1 < p . Then there is a measurable function
k : X Y R such that Tf

= g

f
, where g
f
(y) =
_
k(x, y)f(x)dx almost everywhere, for every f L
1
().
proof Set q =
p
p1
if p is nite, 1 if p = . We can identify L
p
() with V

, where V = L
q
()

= L
p
()

(366Dc, 365Jc) has an order-continuous norm because 1 q < . By 376Mb, T is an abstract integral
operator. By 376F/376J, T is represented by a kernel, as claimed.
376O Under the right conditions, weakly compact operators are abstract integral operators.
Lemma Let U be a Riesz space, and W a solid linear subspace of U

. If C U is relatively compact
for the weak topology T
s
(U, W) (3A5E), then for every g W
+
and > 0 there is a u

U
+
such that
g([u[ u

)
+
for every u C.
proof Let W
g
be the solid linear subspace of W generated by g. Then W
g
is an Archimedean Riesz space
with order unit, so W

g
is a band in the L-space W

g
= W

g
(356Na), and is therefore an L-space in its own
right (354O). For u U, h W

g
set (Tu)(h) = h(u); then T is an order-continuous Riesz homomomorphism
from U to W

g
(356F).
Now W
g
is perfect. PPP I use 356K. W
g
is Dedekind complete because it is a solid linear subspace of the
Dedekind complete space U

. W

g
separates the points of W because T[U] does. If A W
g
is upwards-
directed and sup
hA
(h) is nite for every W

g
, then A acts on W

g
as a set of bounded linear functionals
which, by the Uniform Boundedness Theorem (3A5Ha), is uniformly bounded; that is, there is some M 0
such that sup
hA
[(h)[ M|| for every W

g
. Because g is the standard order unit of W
g
, we have
|| = [[(g) and [(h)[ M[[(g) for every W

g
, h A. In particular,
h(u) [h(u)[ = [(Tu)(h)[ M[Tu[(g) = MTu(g) = Mg(u)
for every h A, u U
+
. But this means that h Mg for every h A and A is bounded above in W
g
.
Thus all the conditions of 356K are satised and W
g
is perfect. QQQ
Accordingly T is continuous for the topologies T
s
(U, W) and T
s
(W

g
, W

g
), because every element of
W

g
corresponds to a member of W
g
W, so 3A5Ec applies.
Now we are supposing that C is relatively compact for T
s
(U, W), that is, is included in some compact set
C
t
; accordingly T[C
t
] is compact and T[C] is relatively compact for T
s
(W

g
, W

g
). Since W

g
is an L-space,
T[C] is uniformly integrable (356Q); consequently (ignoring the trivial case C = ) there are
0
, . . . ,
n

T[C] such that |([[ sup
in
[
i
[)
+
| for every T[C] (354Rb), so that ([[ sup
in
[
i
[)
+
(g) for
every T[C].
376R Kernel operators 469
Translating this back into terms of C itself, and recalling that T is a Riesz homomorphism, we see that
there are u
0
, . . . , u
n
C such that g([u[ sup
in
[u
i
[)
+
for every u C. Setting u

= sup
in
[u
i
[ we
have the result.
376P Theorem Let U be an L-space and V a perfect Riesz space. If T : U V is a linear operator
such that Tu : u U, |u| 1 is relatively compact for the weak topology T
s
(V, V

), then T is an
abstract integral operator.
proof (a) For any g 0 in V

, M
g
= sup
|u|1
g([Tu[) is nite. PPP By 376O, there is a v

V
+
such
that g([Tu[ v

)
+
1 whenever |u| 1; now M
g
g(v

) + 1. QQQ Considering |u|

1
u, we see that
g([Tu[) M
g
|u| for every u U.
Next, we nd that T L

(U; V ). PPP Take u U

+
. Set
B =

n
i=0
[Tu
i
[ : u
0
, . . . , u
n
U
+
,

n
i=0
u
i
= u V
+
.
Then B is upwards-directed. (Cf. 371A.) If g 0 in V

,
sup
vB
g(v) = sup
n

i=0
g([Tu
i
[) :
n

i=0
u
i
= u
sup
n

i=0
M
g
|u
i
| :
n

i=0
u
i
= u = M
g
|u|
is nite. By 356K, B is bounded above in V ; and of course any upper bound for B is also an upper bound
for Tu
t
: 0 u
t
u. As u is arbitrary, T is order-bounded. QQQ
Because U is a Banach lattice with an order-continuous norm, T L

(U; V ) (355Kb).
(b) Since we can identify L

(U; V ) with L

(U; V

), we have an adjoint operator T

t
L

(V

; U

), as
in 376K. Now if g 0 in V

and g
n
)
nN
is a sequence in [0, g] such that lim
n
g
n
(v) = 0 for every v V ,
T
t
g
n
)
nN
order*-converges to 0 in U

. PPP For any > 0, there is a v

V
+
such that g([Tu[ v

)
+

whenever |u| 1; consequently
|T
t
g
n
| = sup
|u|1
(T
t
g
n
)(u) = sup
|u|1
g
n
(Tu)
g
n
(v

) + sup
|u|1
g
n
([Tu[ v

)
+
g
n
(v

) + sup
|u|1
g([Tu[ v

)
+
g
n
(v

) +
for every n N. As lim
n
g
n
(v

) = 0, limsup
n
|T
t
g
n
| ; as is arbitrary, |T
t
g
n
|)
nN
0. But as
U

is an M-space (356Pb), it follows that T

t
g
n
)
nN
order*-converges to 0. QQQ
By 368Pc, V is weakly (, )-distributive. By 376H, T
t
is an abstract integral operator, so T also is, by
376K.
376Q Corollary Let (X, , ) and (Y, T, ) be -nite measure spaces and T : L
1
() L
1
() a weakly
compact bounded linear operator. Then there is a function k : X Y R such that Tf

= g

f
, where
g
f
(y) =
_
k(x, y)f(x)dx almost everywhere, for every f L
1
().
proof This follows from 376P and 376J, just as in 376N.
376R So far I have mentioned actual kernel functions k(x, y) only as a way of giving slightly more
concrete form to the abstract kernels of 376E. But of course they can provide new structures and insights.
I give one result as an example. The following lemma is useful.
Lemma Let (X, , ) be a measure space, (Y, T, ) a -nite measure space, and the c.l.d. product
measure on X Y . Suppose that k is a -integrable real-valued function. Then for any > 0 there is a
nite partition E
0
, . . . , E
n
of X into measurable sets such that |k k
1
|
1
, where
470 Linear operators between function spaces 376R
k
1
(x, y) =
1
E
i
_
E
i
k(t, y)dt whenever x E
i
, 0 < E
i
<
and the integral is dened in R,
= 0 in all other cases.
proof Once again I refer to the proof of 253F: there are sets H
0
, . . . , H
r
of nite measure in X, sets
F
0
, . . . , F
r
of nite measure in Y , and
0
, . . . ,
r
such that |k k
2
|
1

1
2
, where k
2
=

r
j=0

i
(H
j
F
j
).
Let E
0
, . . . , E
n
be the partition of X generated by H
i
: i r. Then for any i n,
_
E
i
Y
[k k
1
[ is
dened and is at most 2
_
E
i
Y
[k k
2
[. PPP If E
i
= 0, this is trivial, as both are zero. If E
i
= , then
again the result is elementary, since both k
1
and k
2
are zero on E
i
Y . So let us suppose that 0 < E
i
< .
In this case
_
E
i
k(t, y)dt must be dened for almost every y, by Fubinis theorem. So k
1
is dened almost
everywhere on E
i
Y , and
_
E
i
Y
[k k
1
[ =
_
Y
_
E
i
[k(x, y) k
1
(x, y)[dxdy.
Now take some xed y Y such that
=
1
E
i
_
E
i
k(t, y)dt
is dened. Then = k
1
(x, y) for every x E
i
. For every x E
i
, we must have k
2
(x, y) = where
=

j
: E
i
H
j
, y F
j
. But in this case, because
_
E
i
k(x, y) dx = 0, we have
_
E
i
max(0, k(x, y) )dx =
_
E
i
max(0, k(x, y))dx =
1
2
_
E
i
[k(x, y) k
1
(x, y)[dx.
If ,
_
E
i
max(0, k(x, y) )dx
_
E
i
max(0, k(x, y) )dx
_
E
i
[k(x, y) k
2
(x, y)[dx;
if ,
_
E
i
max(0, k(x, y))dx
_
E
i
max(0, k(x, y))dx
_
E
i
[k(x, y) k
2
(x, y)[dx;
in either case,
1
2
_
E
i
[k(x, y) k
1
(x, y)[dx
_
E
i
[k(x, y) k
2
(x, y)[dx.
This is true for almost every y, so integrating with respect to y we get the result. QQQ
Now, summing over i, we get
_
[k k
1
[ 2
_
[k k
2
[ ,
as required.
376S Theorem Let (X, , ) be a complete locally determined measure space, (Y, T, ) a -nite
measure space, and the c.l.d. product measure on XY . Let be an extended Fatou norm on L
0
() and
write L

for g : g L
0
(),
t
(g

) < , where
t
is the associate extended Fatou norm of (369H-369I).
Suppose that k L
0
() is such that k (f g) is integrable whenever f L
1
() and g L

. Then we
have a corresponding linear operator T : L
1
() L

dened by saying that

_
(Tf

) g

=
_
k (f g)
whenever f L
1
(), g L

.
For x X set k
x
(y) = k(x, y) whenever this is dened. Then k
x
L
0
() for almost every x; set
v
x
= k

x
L
0
() for such x. In this case x (v
x
) is measurable and dened and nite almost everywhere,
and |T| = ess sup
x
(v
x
).
Remarks The discussion of extended Fatou norms in 369 regarded them as functionals on spaces of the
form L
0
(A). I trust that no-one will be oended if I now speak of an extended Fatou norm on L
0
(), with
the associated function spaces L

, L

L
0
, taking for granted the identication in 364Jc.
Recall that (f g)(x, y) = f(x)g(y) for x domf, y domg (253B).
376S Kernel operators 471
By ess sup
x
(v
x
) I mean
infM : M 0, x : v
x
is dened and (v
x
) M is conegligible
(see 243D).
proof (a) To see that the formula (f, g)
_
k (f g) gives rise to an operator in L

(U; (L

), it is
perhaps quickest to repeat the argument of parts (a) and (b) of the proof of 376E. (We are not quite in
a position to quote 376E, as stated, because the localizable measure algebra free product there might be
strictly larger than the measure algebra of ; see 325B.) The rst step, of course, is to note that changing
f or g on a negligible set does not aect the integral
_
k (f g), so that we have a bilinear functional on
L
1
L

; and the other essential element is the fact that the maps f

(f Y )

, g

(X g)

are
order-continuous (put 325A and 364Rc together).
By 369K, we can identify (L

with L

, so that T becomes an operator in L

(U; L

). Note that it
must be norm-bounded (355C).
(b) By 376I, there is a non-decreasing sequence Y
n
)
nN
of measurable sets in Y , covering Y , such that
Y
n
L

for every n. Set X

0
= x : x X, k
x
L
0
(). Then X
0
is conegligible in X. PPP Let E be
any set of nite measure. Then for any n N, k (E Y
n
) is integrable, that is,
_
EY
n
k is dened and
nite; so by Fubinis theorem
_
Y
n
k
x
is dened and nite for almost every x E. Consequently, for almost
every x E, k
x
Y
n
L
0
() for every n N, that is, k
x
L
0
(), that is, x X
0
.
Thus EX
0
is negligible for every set E of nite measure. Because is complete and locally determined,
X
0
is conegligible. QQQ
This means that v
x
and (v
x
) are dened for almost every x.
(c) (v
x
) |T| for almost every x. PPP Take any E of nite measure, and n N. Then k(EY
n
)
is integrable. For each r N, there is a nite partition E
r0
, . . . , E
r,m(r)
of E into measurable sets such that
_
EY
n
[k k
(r)
[ 2
r
, where
k
(r)
(x, y) =
1
E
ri
_
E
ri
k(t, y)dt whenever y Y
n
, x E
ri
, E
ri
> 0
and the integral is dened in R,
= 0 otherwise
(376R). Now k
(r)
is also integrable over E Y
n
, so k
(r)
x
L
0
() for almost every x E, writing k
(r)
x
(y) =
k
(r)
(x, y), and we can speak of v
(r)
x
= (k
(r)
x
)

for almost every x. Note that k

(r)
x
= k
(r)
x
whenever x, x
t
belong
to the same E
ri
.
If E
ri
> 0, then v
(r)
x
must be dened for every x E
ri
. If v
t
L

is represented by g L

then
_
k (E
ri
(g Y
n
)) =
_
E
ri
Y
n
k(t, y)g(y)d(t, y)
= E
ri
_
k
(r)
(x, y)g(y)dy = E
ri
_
v
(r)
x
v
t
for any x E
ri
. But this means that
E
ri
_
v
(r)
x
v
t
=
_
T(E

ri
) v
t
Y

n
for every v
t
L

, so
v
(r)
x
=
1
E
ri
T(E

ri
) Y

n
, (v
(r)
x
)
1
E
ri
|T||E

ri
|
1
= |T|
for every x E
ri
. This is true whenever E
ri
> 0, so in fact (v
(r)
x
) |T| for almost every x E.
Because

rN
_
EY
n
[k k
(r)
[ < , we must have k(x, y) = lim
r
k
(r)
(x, y) for almost every (x, y)
E Y
n
. Consequently, for almost every x E, k(x, y) = lim
r
k
(r)
(x, y) for almost every y Y
n
, that is,
472 Linear operators between function spaces 376S
v
(r)
x
)
rN
order*-converges to v
x
Y

n
(in L
0
()) for almost every x E. But this means that, for almost
every x E,
(v
x
Y

n
) liminf
r
(v
(r)
x
) |T|
(369Mc). Now
(v
x
) = lim
n
(v
x
Y

n
) |T|
for almost every x E.
As in (b), this implies (since E is arbitrary) that (v
x
) |T| for almost every x X. QQQ
(d) I now show that x (v
x
) is measurable. PPP Take [0, [ and set A = x : x X
0
, (v
x
) .
Suppose that E < . Let G be a measurable envelope of AE (132Ed). Set

k(x, y) = k(x, y) when x G
and (x, y) domk, 0 otherwise. If f L
1
() and g L

, then
_

k(x, y)f(x)g(y)d(x, y) =
_
GY
k(x, y)f(x)g(y)d(x, y) =
_
G
f(x)
_
Y
k(x, y)g(y)dydx
is dened.
For x X
0
, set h(x) =
_
[

k(x, y)g(y)[dy. Then h is nite almost everywhere and measurable. For

x A E,
_
[

k(x, y)g(y)[dy =
_
[v
x
g

[
t
(g

).
So the measurable set G
t
= x : h(x)
t
(g

) includes A E, and (G G
t
) = 0. Consequently
[
_

k(x, y)f(x)g(y)d(x, y)[
_
G
[f(x)[h(x)dx |f|
1

t
(g

),
and this is true for every f L
1
(), g L

.
Now we have an operator

T : L
1
() L

dened by the formula

_
(

Tf

) g

=
_

k (f g) when f L
1
() and g L

,
and the formula just above tells us that [
_

Tu v
t
[ |u|
1

t
(v
t
) for every u L
1
(), v
t
L

; that is,
(

Tu) |u|
1
for every u L
1
(); that is, |

T| . But now (c) tells us that ( v

x
) for almost every
x X, where v
x
is the equivalence class of y

k(x, y), that is, v
x
= v
x
for x G X
0
, 0 for x X G.
So (v
x
) for almost every x G, and G A is negligible. But this means that AE is measurable. As
E is arbitrary, A is measurable; as is arbitrary, x (v
x
) is measurable. QQQ
(e) Finally, the ideas in (d) show that |T| ess sup
x
(v
x
). PPP Set M = ess sup
x
(v
x
). If f L
1
()
and g L

, then
_
[k(x, y)f(x)g(y)[d(x, y)
_
[f(x)[(v
x
)
t
(g

)dx M|f|
1

t
(g

);
as g is arbitrary, (Tf

) M|f|
1
; as f is arbitrary, |T| M. QQQ
376X Basic exercises >>>(a) Let be Lebesgue measure on R. Let h be a -integrable real-valued
function with |h|
1
1, and set k(x, y) = h(y x) whenever this is dened. Show that if f is in either L
1
()
or L

() then g(y) =
_
k(x, y)f(x)dx is dened for almost every y R, and that this formula gives rise to
an operator T T

,
as dened in 373A. (Hint: 255H.)
(b) Let (A, ) and (B, ) be semi-nite measure algebras with localizable measure algebra free product
(C,

), and take p [1, ]. Show that if u L

p
(A, ) and v L
p
(B, ) then u v L
p
(C,

) and
|u v|
p
= |u|
p
|v|
p
.
>>>(c) Let U, V , W be Riesz spaces, of which V and W are Dedekind complete, and suppose that T
L

(U; V ) and S L

(V ; W). Show that if either S or T is an abstract integral operator, so is ST.

(d) Let h be a Lebesgue integrable function on R, and f a square-integrable function. Suppose that
f
n
)
nN
is a sequence of measurable functions such that () [f
n
[ f for every n () lim
n
_
E
f
n
= 0 for
every measurable set E of nite measure. Show that lim
n
(h f
n
)(y) = 0 for almost every y R, where
h f
n
is the convolution of h and f
n
. (Hint: 376Xa, 376F.)
376Xn Kernel operators 473
(e) Let U and V be Riesz spaces, of which V is Dedekind complete. Suppose that W U

is a solid
linear subspace, and that T belongs to the band in L

(U; V ) generated by operators of the form u f(u)v,

where f W and v V . Show that whenever u
n
)
nN
is an order-bounded sequence in U such that
lim
n
f(u
n
) = 0 for every f W, then Tu
n
)
nN
order*-converges to 0 in V .
(f ) Let (A, ) be a semi-nite measure algebra and U L
0
= L
0
(A) an order-dense Riesz subspace such
that U

separates the points of U. Let u

n
)
nN
be an order-bounded sequence in U. Show that the following
are equiveridical: (i) lim
n
f([u
n
[) = 0 for every f U

; (ii) u
n
)
nN
0 for the topology of convergence
in measure on L
0
. (Hint: by 367T, condition (ii) is intrinsic to U, so we can replace (A, ) by a localizable
algebra and use the representation in 369D.)
(g) Let U be a Banach lattice with an order-continuous norm, and V a weakly (, )-distributive Riesz
space. Show that for T L

(U; V ) the following are equiveridical: (i) T belongs to the band in L

(U; V )
generated by operators of the form u f(u)v where f U

, v V ; (ii) Tu
n
)
nN
order*-converges to 0
in V whenever u
n
)
nN
is an order-bounded sequence in U
+
which is norm-convergent to 0; (iii) Tu
n
)
nN
order*-converges to 0 in V whenever u
n
)
nN
is an order-bounded sequence in U which is weakly convergent
to 0.
(h) Let (X, , ) and (Y, T, ) be -nite measure spaces, with product measure on X Y , and
measure algebras (A, ), (B, ). Suppose that k L
0
(). Show that the following are equiveridical: (i)()
if f L
1
() then g
f
(y) =
_
k(x, y)f(x)dx is dened for almost every y and g
f
L
1
() () there is an
operator T T

,
dened by setting Tf

= g

f
for every f L
1
(); (ii)
_
[k(x, y)[dy 1 for almost every
x X,
_
[k(x, y)[dx 1 for almost every y Y .
>>>(i) Let be Lebesgue measure on R. Give an example of a measurable function k : [0, 1]
2
R such
that, for any f L
2
(), g
f
(y) =
_
k(x, y)f(x)dx is dened for every y and |g
f
|
2
= |f|
2
, but k is not
integrable, so the linear isometry on L
2
= L
2
() dened by k does not belong to L

(L
2
; L
2
). (Hint: 371Ye.)
(j) Let (X, , ) be a -nite measure space and (Y, T, ) a complete locally determined measure space.
Let U L
0
(), V L
0
() be solid linear subspaces, of which V is order-dense; write V
#
= v : v
L
0
(), v v
t
is integrable for every v
t
V , U = f : f L
0
(), f

U, V = g : g L
0
(), g

V ,
V
#
= h : h L
0
(), h

V
#
. Let be the c.l.d. product measure on X Y , and k L
0
() a function
such that k(fg) is integrable for every f U, g V. (i) Show that for any f U, h
f
(y) =
_
k(x, y)f(x)dx
is dened for almost every y Y , and that h
f
V
#
. (ii) Show that we have a map T L

(U; V
#
) dened
either by writing Tf

= h

f
for every f U or by writing
_
(Tf

) g

=
_
k (f g) for every f U,
g V.
(k) Let (X, , ), (Y, T, ) and (Z, , ) be -nite measure spaces, and U, V , W perfect order-dense
solid linear subspaces of L
0
(), L
0
() and L
0
() respectively. Suppose that T : U V and S : V W
are abstract integral operators corresponding to kernels k
1
L
0
( ), k
2
L
0
( ), writing for
the (c.l.d. or primitive) product measure on X Y . Show that ST : U W is represented by the kernel
k L
0
( ) dened by setting k(x, z) =
_
k
1
(x, y)k
2
(y, z)dy whenever this integral is dened.
(l) Let U be a perfect Riesz space. Show that a set C U is relatively compact for T
s
(U, U

) i for
every g (U

)
+
, > 0 there is a u

U such that g([u[ u

)
+
for every u C. (Hint: 376O and the
proof of 356Q.)
>>>(m) Let be Lebesgue measure on [0, 1], and counting measure on [0, 1]. Set k(x, y) = 1 if x = y, 0
otherwise. Show that 376S fails in this context (with, e.g., = | |

).
(n) Suppose, in 376Xj, that U = L

for some extended Fatou norm on L

0
() and that V = L
1
(), so
that V
#
= L

(). Set k
y
(x) = k(x, y) whenever this is dened, u
y
= k

y
whenever k
y
L
0
(). Show that
u
y
L

for almost every y Y , and that the norm of T in B(L

; L

) is ess sup
y

t
(u
y
). (Hint: do the case
of totally nite Y rst.)
474 Linear operators between function spaces 376Y
376Y Further exercises (a) Let U, V and W be linear spaces (over any eld F) and : U V W
a bilinear map. Let W
0
be the linear subspace of W generated by [U V ]. Show that the following are
equiveridical: (i) for every linear space Z over F and every bilinear : U V Z, there is a (unique)
linear operator T : W
0
Z such that T = (ii) whenever u
0
, . . . , u
n
U are linearly independent
and v
0
, . . . , v
n
V are non-zero,

n
i=0
(u
i
, v
i
) ,= 0 (iii) whenever u
0
, . . . , u
n
U are non-zero and
v
0
, . . . , v
n
V are linearly independent,

n
i=0
(u
i
, v
i
) ,= 0 (iv) for any Hamel bases u
i
)
iI
, v
j
)
jJ
of U
and V , (u
i
, v
j
))
iI,jJ
is a Hamel basis of W
0
(v) for some pair u
i
)
iI
, v
j
)
jJ
of Hamel bases of U and
V , (u
i
, v
j
))
iI,jJ
is a Hamel basis of W
0
.
(b) Let (A, ), (B, ) be semi-nite measure algebras, and (C,

) their localizable measure algebra free

product. Show that : L
0
(A) L
0
(B) L
0
(C) satises the equivalent conditions of 376Ya.
(c) Let (X, , ) and (Y, T, ) be semi-nite measure spaces and the c.l.d. product measure on X Y .
Show that the map (f, g) f g : L
0
() L
0
() L
0
() induces a map (u, v) uv : L
0
() L
0
()
L
0
() possessing all the properties described in 376B and 376Ya.
(d) Let (A, ) be the measure algebra of 0, 1

1
with its usual measure, and b

)
<
1
the canonical
independent family of elements of measure
1
2
in A. Set U = L
2
(A, ) and V = v : v R

1
, : v() ,= 0 is
countable. Dene T : U V by setting Tu() = 2
_
b

u
_
u for <
1
, u U. Show that (i) Tu
n
)
nN
order*-converges to 0 in V whenever u
n
)
nN
is a sequence in U such that lim
n
f(u
n
) = 0 for every
f U

(ii) T / L

(U; V ).
(e) Let U be a Riesz space with the countable sup property (denition: 241Yd) such that U

separates the
points of U, and u
n
)
nN
a sequence in U. Show that the following are equiveridical: (i) lim
n
f(v[u
n
[) =
0 for every f U

, v U
+
; (ii) every subsequence of u
n
)
nN
has a sub-subsequence which is order*-
convergent to 0.
(f ) Let U be an Archimedean Riesz space and A a weakly (, )-distributive Dedekind complete Boolean
algebra. Suppose that T : U L
0
= L
0
(A) is a linear operator such that [Tu
n
[)
nN
order*-converges to 0
in L
0
whenever u
n
)
nN
is order-bounded and order*-convergent to 0 in U. Show that T L

c
(U; L
0
), so
that if U has the countable sup property then T L

(U; L
0
).
(g) Suppose that (Y, T, ) is a probability space in which T = TY , y = 0 for every y Y . (See 363S.)
Take X = Y and let be counting measure on X; let be the c.l.d. product measure on X Y , and set
k(x, y) = 1 if x = y, 0 otherwise. Show that we have an operator T : L

() L

() dened by setting
Tf = g

whenever f L

()

=

(X) and g(y) =

_
k(x, y)f(x)dx = f(y) for every y Y . Show that T
does not belong to L

(L

(); L

()) and in particular does not satisfy the condition (ii) of 376J.
(h) Give an example of an abstract integral operator T :
2
L
1
(), where is Lebesgue measure on
[0, 1], such that Te
n
)
nN
is not order*-convergent in L
1
(), where e
n
)
nN
is the standard orthonormal
sequence in
2
.
(i) Let U be an L-space and write G for the regular open algebra of R. Show that any bounded linear
operator from U to L

(G) is an abstract integral operator.

(j) Let U be an L-space and V a Banach lattice with an order-continuous norm. Let T L

(U; V ). Show
that the following are equiveridical: (i) T is an abstract integral operator; (ii) T[C] is norm-compact in V
whenever C is weakly compact in U. (Hint: start with the case in which C is order-bounded, and remember
that it is weakly sequentially compact.)
(k) Let (X, , ) be a semi-nite measure space. Show that 376S is valid for (X, , ) i (X, , ) has
locally determined negligible sets (213I).
(l) Let (X, , ) be a complete locally determined measure space and (Y, T, ), (Z, , ) two -nite
measure spaces. Suppose that , are extended Fatou norms on L
0
(), L
0
() respectively, and that T :
L
1
() L

is an abstract integral operator, with corresponding kernel k L

0
(), while S L

(L

; L

),
so that ST : L
1
() L

is an abstract integral operator (376Xc); let

k L
0
( ) be the corresponding
kernel. For x X set v
x
= k

x
when this is dened in L

, as in 376S, and similarly take w

x
=

k

x
L

.
Show that Sv
x
= w
x
for almost every x X.
376 Notes Kernel operators 475
(m) Set k(m, n) = 1/(n m +
1
2
) for m, n Z. (i) Show that

n=
k(m, n)
2
= 1 for every m Z.
(Hint: 282Xo.) (ii) Show that

n=
k(m, n)k(m
t
, n) = 0 for all distinct m, m
t
Z. (Hint: look at
k(m, n) k(m
t
, n).) (iii) Show that there is a norm-preserving linear operator T from
2
=
2
(Z) to itself
given by setting (Tu)(n) =

m=
k(m, n)u(m) for every n. (iv) Show that T
2
is the identity operator on

2
. (v) Show that T / L

(
2
;
2
). (Hint: consider

m,n=
[k(m, n)[x(m)x(n) where x(n) = 1/
_
[n[ ln[n[
for [n[ 2.) (T is a form of the Hilbert transform.)
(n) (i) Show that there is a compact linear operator from
2
to itself which is not in L

(
2
;
2
). (Hint:
start from the operator S of 371Ye.) (ii) Show that the identity operator on
2
is an abstract integral
operator.
376 Notes and comments I leave 376Yb to the exercises because I do not rely on it for any of the work
here, but of course it is an essential aspect of the map : L
0
(A) L
0
(B) L
0
(C) I discuss in this section.
The conditions in 376Ya are characterizations of the tensor product of two linear spaces, a construction of
great importance in abstract linear algebra (and, indeed, in modern applied linear algebra; it is by no means
trivial even in the nite-dimensional case). In particular, note that conditions (ii), (iii) of 376Ya apply to
arbitrary subspaces of U and V if they apply to U and V themselves.
The principal ideas used in 376B-376C have already been set out in 253 and 325. Here I do little more
than list the references. I remark however that it is quite striking that L
1
(C,

) should have no fewer than

three universal mapping theorems attached to it (376Ca, 376C(b-i) and 376C(b-ii)).
The real work of this section begins in 376E. As usual, much of the proof is taken up with relatively
straightforward verications, as in parts (a) and (b), while part (i) is just a manoeuvre to show that it
doesnt matter if A and B arent Dedekind complete, because C is. But I think that parts (d), (f) and (j)
have ideas in them. In particular, part (f) is a kind of application of the Radon-Nikod ym theorem (through
the identication of L
1
(C,

with L

(C)).
I have split 376E from 376H because the former demands the language of measure algebras, while the
latter can be put into the language of pure Riesz space theory. Asking for a weakly (, )-distributive space
V in 376H is a way of applying the ideas to V = L
0
as well as to Banach function spaces. (When V = L
0
,
indeed, variations on the hypotheses are possible, using 376Yf.) But it is a reminder of one of the directions
in which it is often possible to nd generalizations of ideas beginning in measure theory.
The condition lim
n
f(u
n
) = 0 for every f U

(376H(ii)) seems natural in this context, and gives

marginally greater generality than some alternatives (because it does the right thing when U

does not
separate the points of U), but it is not the only way of expressing the idea; see 376Xf and 376Ye. Note that
the conditions (ii) and (iii) of 376H are signicantly dierent. In 376H(iii) we could easily have [u
n
[ = u

for
every n; for instance, if u
n
= 2a
n
1 for some stochastically independent sequence a
n
)
nN
of elements
of measure
1
2
in a probability algebra (272Yd).
If you have studied compact linear operators between Banach spaces (denition: 3A5Ka), you will have
encountered the condition Tu
n
0 strongly whenever u
n
0 weakly. The conditions in 376H and 376J
are of this type. If a sequence u
n
)
nN
in a Riesz space U is order-bounded and order*-convergent to 0,
then lim
n
f(u
n
) = 0 for every f U

(367Xf). Visibly this latter condition is associated with weak

convergence, and order*-convergence is (in Banach lattices) closely related to norm convergence (367D-
367E). In the context of 376H, an abstract integral operator is one which transforms convergent sequences
of a weak type into convergent sequences of a stronger type. The relationship between the classes of (weakly)
compact operators and abstract integral operators is interesting, but outside the scope of this book; I leave
you with 376P-376Q and 376Y, and a pair of elementary examples to guard against extravagant conjecture
(376Yn).
376O belongs to an extensive general theory of weak compactness in perfect Riesz spaces, based on
adaptations of the concept of uniform integrability. I give the next step in 376Xl. For more information
see Fremlin 74a, chap. 8.
Note that 376Mb and 376P overlap when V

in 376Mb is reexive for instance, when V is an L

p
space
for some p ]1, [ since then every bounded linear operator from L
1
to V

must be weakly compact.

I give 376Yi as a hint that there may be more to be said about the case in which the codomain is not
perfect, indeed not weakly (, )-distributive. For more information on the representation of operators see
Dunford & Schwartz 57, particularly Table VI in the notes to Chapter VI.
476 Linear operators between function spaces 376 Notes
As soon as we leave formulations in terms of the spaces L
0
(A) and their subspaces, and return to the
original conception of a kernel operator in terms of integrating functions against sections of a kernel, we
are necessarily involved in the pathology of Fubinis theorem for general measure spaces. In general, the
repeated integrals
__
k(x, y)dxdy,
__
k(x, y)dydx need not be equal, and something has to give (376Xm).
Of course this particular worry disappears if the spaces are -nite, as in 376J. In 376S I take the trouble
to oer a more general condition, fairly near to the best possible result (376Yk), mostly as a reminder that
the techniques developed in Volume 2 do enable us sometimes to go beyond the -nite case. Note that
this is one of the many contexts in which anything we can prove about probability spaces will be true of all
-nite spaces; but that we cannot make the next step, to all strictly localizable spaces.
376S verges on the theory of integration of vector-valued functions, which I dont wish to enter here;
but it also seems to have a natural place in the context of this chapter. It is of course a special property
of L
1
spaces. The formula |T
k
| = ess sup
x
(k

x
) shows that |T
]k]
| = |T
k
|; now we know fron 376E that
T
]k]
= [T
k
[, so we get a special case of the Chacon-Krengel theorem (371D). Reversing the roles of X and Y ,
we nd ourselves with an operator from L

to L

(376Xn), which is the other standard context in which

|T| = |[T[| (371Xd). I include two exercises on L
2
spaces (376Xi, 376Ym) designed to emphasize the fact
that B(U; V ) is included in L

(U; V ) only in very special cases.

The history of the theory here is even more confusing than that of mathematics in general, because so
many of the ideas were developed in national schools in very imperfect contact with each other. My own
account gives no hint of how this material arose; I ought in particular to note that 376N is one of the oldest
results, coming (essentially) from Dunford 36. For further references, see Zaanen 83, chap. 13.
381B Automorphism groups of Boolean algebras 477
Chapter 38
Automorphism groups
As with any mathematical structure, every measure algebra has an associated symmetry group, the group
of all measure-preserving automorphisms. In this chapter I set out to describe some of the remarkable features
of these groups. I begin with a section on automorphism groups of general Dedekind complete Boolean
algebras (381), before continuing with applications of the ideas developed there to measure algebras (382);
the principal results of these sections concern the expression of a general automorphism as a product of
involutions (381N, 382D) and a description of the normal subgroups of automorphism groups of homogeneous
algebras (381S, 382H). I continue with a discussion of circumstances under which these automorphism groups
determine the underlying algebras and/or have few outer automorphisms (383).
One of the outstanding open problems of the subject is the isomorphism problem, the classication
of automorphisms of measure algebras up to conjugacy in the automorphism group. I oer a section on
entropy, the most important numerical invariant enabling us to distinguish some non-conjugate automor-
phisms (384). For Bernoulli shifts on the Lebesgue measure algebra (384Q-384S), the isomorphism problem
is solved by Ornsteins theorem; I present a complete proof of this theorem in 385-386. Finally, in 387,
I give Dyes theorem, describing the full subgroups generated by single automorphisms of measure algebras
of countable Maharam type.
381 Automorphism groups of Boolean algebras
My aim in this chapter is to describe the automorphism groups of measure algebras, but as usual I prefer
to begin with results which can be expressed in the language of general Boolean algebras, even though
on this occasion it is necessary to introduce some special terminology. I will however restrict myself to
the ideas which seem necessary for the theorems in the next two sections. I begin with some elementary
general results on the construction of automorphisms by piecing together fragments of others (381B-381C), a
discussion of the concept of an element supporting an automorphism (381D-381F, 381L), an introduction to
a cycle notation for certain automorphisms (381G-381I), and a version of Frolks theorem on the structure
of automorphisms (381J).
The principal theorems are 381N, giving a sucient condition for every automorphism to be a product of
involutions, and 381S, describing the normal subgroups of certain groups of automorphisms. Both depend
on Dedekind completeness of the Boolean algebra and on the concept of full subgroup of Aut A (381M);
the latter also requires the group to have many involutions (381P-381R). Both concepts are chosen with a
view to the next section, where the results will be applied to groups of measure-preserving automorphisms.
381A The group Aut A For any Boolean algebra A, I write Aut A for the set of automorphisms of A,
that is, the set of bijective Boolean homomorphisms : A A. This is a group, being a subgroup of the
group of all bijections from A to itself (use 312G). Note that every member of Aut A is order-continuous; this
is because it must be an isomorphism of the order structure of A, and is also a consequence of 313P(a-ii).
381B Lemma Let A be a Boolean algebra, and a
i
)
iI
, b
i
)
iI
two partitions of unity in A. Assume
either that I is nite
or that I is countable and A is Dedekind -complete
or that A is Dedekind complete.
Suppose that for each i I we have an isomorphism
i
: A
a
i
A
b
i
between the corresponding principal
ideals. The there is a unique Aut A such that c =
i
c whenever i I and c a
i
.
proof By 315F, we may identify A with each of the products

iI
A
a
i
,

iI
A
b
i
; now corresponds to
the isomorphism between the two products induced by the
i
.
478 Automorphism groups 381C
381C Corollary Let A be a homogeneous Boolean algebra, and A, B two partitions of unity in A,
neither containing 0. Let : A B be a bijection. Suppose
either that A, B are nite
or that A, B are countable and A is Dedekind -complete
or that A is Dedekind complete.
Then there is an automorphism of A extending .
proof For every a A, the principal ideals A
a
, A
a
are isomorphic to the whole algebra A, and therefore
to each other; let
a
: A
a
A
a
be an isomorphism. Now apply 381B.
381D Elements supporting an automorphism The following concept will be extremely useful. If
A is a Boolean algebra, Aut A and a A I will say that is supported by a, or a supports , if
d = d for every d 1 \ a. We have the following elementary facts.
381E Lemma Let A be a Boolean algebra.
(a) If Aut A is supported by a A then a = a and d a for every d a.
(b) If , Aut A are both supported by a A so are
1
and .
(c) If Aut A is supported by a A and Aut A then
1
is supported by a.
(d) If , Aut A are supported by a, b respectively and a b = 0, then = .
(e) Let Aut A and let A be the set of a A supporting . Then A is non-empty and closed under
arbitrary inma; also b A whenever b a A. In particular, if is supported by both a and b then it is
supported by a b.
(f) For any Aut A, a A, a supports i d a whenever d d = 0.
(g) If Aut A is supported by a A, and , Aut A are such that d = d for every d a, then

1
=
1
.
proof (a) (1 \ a) = 1 \ a, so a = a, and if d a then d a = a.
(b) If d a = 0 then d = d = d so
1
d = d = d.
(c) If d a = 0 then
1
d a = 0 so
1
d =
1
d and
1
d = d.
(d) If d a, then d = d so d = d. Also d a = a, so d = d; thus d = d. Similarly
d = d if d b. Finally, if d 1 \ (a b), then d = d = d. By the uniqueness assertion of 381B,
or otherwise, = .
(e) Of course 1 A, because 0 = 0; and it is also obvious that if b a A then b A. If B A is
non-empty and c = inf B is dened in A, then for any d 1 \ c we have
d = d \ c = sup
bB
d \ b,
so (because is order-continuous)
d = sup
bB
(d \ b) = sup
bB
d \ b = d.
So c supports .
(f )(i) If a supports and d d = 0, then
d \ a = (d \ a) d d d = 0,
so d a.
(ii) If a does not support , there is a c 1 \ a such that c ,= c. So one of c \ c, c \ c is non-zero.
If c \ c ,= 0, take this for d; then d , a and d d c \ c = 0. Otherwise, take d =
1
(c \ c); then
0 ,= d c, so d , a, while
d d = (c \
1
c) (c \ c) = 0.
(g) For d a,
1
d =
1
d = d, so
1
is supported by 1 \ a. By (d),
1
=
1
, so

1
=
1

1
=
1

1
=
1
.
381H Automorphism groups of Boolean algebras 479
381F Corollary Let A be a Boolean algebra and Aut A. If e A is such that there is any b e for
which b ,= b, then there is a non-zero a e such that a a = 0.
proof Because 1 \ e does not support , 381Ef tells us that there is a d , 1 e such that d d = 0; now
set a = d e.
381G Cyclic automorphisms: Denition Let A be a Boolean algebra.
(a) Suppose that a, b are disjoint members of A and that Aut A is such that a = b. I will write
(

b) for the member of Aut A dened by setting

d = d if d a,
=
1
d if d b,
= d if d 1 \ (a b).
Observe that in this case (if a ,= 0) is an involution, that is, has order 2 in the group Aut A; involutions
of this type I will call exchanging involutions.
(b) More generally, if a
1
, . . . , a
n
are disjoint elements of A and
i
Aut A are such that
i
a
i
= a
i+1
for
each i < n, then I will write
(

a
1
1
a
2
2
. . .

n1
a
n
)
for that Aut A such that
d =
i
d if 1 i < n, d a
i
,
=
1
1

1
2
. . .
1
n1
d if d a
n
,
= d if d 1 \ sup
in
a
i
.
(c) It will occasionally be convenient to use the same notation when each
i
is a Boolean isomorphism
between the principal ideals A
a
i
and A
a
i+1
, rather than an automorphism of the whole algebra A.
Remark The point of this notation is that we can expect to use the standard techniques for manipulating
cycles that are (I suppose) familiar to you from elementary group theory; the principal change is that we
have to keep track of the subscripted automorphisms

. The following results are typical.
381H Lemma Let A be a Boolean algebra.
(a) If = (

b) is an exchanging involution in Aut A, then

= (

b) = (

a) = (

1 a)
is supported by a b.
(b) If = (

b) is an exchanging involution in Aut A, then for any Aut A,

1
= (

1 b)
is another exchanging involution.
(c) If = (

b) and = (

d) are exchanging involutions, and a, b, c, d are all disjoint, then and

commute, and = = is also an exchanging involution, being (

a c

b d).
proof (a) Check the action of on the principal ideals A
a
, A
b
, A
1\(ab)
.
(b) a b = (a b) = 0 and

1
a = a = b,
so = (

1 b) is well-dened. Now check the action of on the principal ideals A

a
, A
b
, A
1\(ab)
.
(c) Check the action of on each of the principal ideals A
a
, . . . , A
e
, where e = 1 \ (a b c d).
480 Automorphism groups 381I
381I Remark I must emphasize that while, after a little practice, calculations of this kind become
easy and safe, they are absolutely dependent on all the cycles present involving only members of one list of
disjoint elements of A. If, for instance, a, b, c are disjoint, then
(

b)(

c) = (

c).
But if a c ,= 0 then there is no expression for the product in this language. Secondly, of course, we must
be scrupulous in checking, at every use of the notation (

a
1
1
. . . a
n
), that a
1
, . . . , a
n
are disjoint and that

i
a
i
= a
i+1
for i < n. Thirdly, a signicant problem can arise if the automorphisms involved dont match.
Consider for instance the product
= (

b)(

b).
Then we have d =
1
d if d a,
1
d if d b; is not necessarily expressible as a product of disjoint
cycles. Clearly there are indenitely complex variations possible on this theme. A possible formal expression
of a sucient condition to avoid these diculties is the following. Restrict yourself to calculations involving a
xed list a
1
, . . . , a
n
of disjoint elements of A for which you can describe a family of isomorphisms
ij
: A
a
i

A
a
j
such that
ii
is always the identity on A
a
i
,
jk

ij
=
ik
for all i, j, k, and whenever a
i
a
j
appears in
a cycle of the calculation, then agrees with
ij
on A
a
i
. Of course this would be intolerably unwieldy if it
were really necessary to exhibit all the
ij
every time. I believe however that it is usually easy enough to
form a mental picture of the actions of the isomorphisms involved suciently clear to oer condence that
such
ij
are indeed present; and in cases of doubt, then after performing the formal operations it is always
straightforward to check that the calculations are valid, by looking at the actions of the automorphisms on
each relevant principal ideal.
381J The following facts may help to put the concept of exchanging involution in its proper place.
Lemma Let A be a Dedekind complete Boolean algebra and any automorphism of A. Then there is a
partition of unity (a
t
, a
tt
, b
t
, b
tt
, c, e) of A such that
a
t
= b
t
, a
tt
= b
tt
, b
tt
= c, (b
t
c) = a
t
a
tt
, d = d for every d e.
proof (a) Let P be the set p : p A, p p = 0. Then P has a maximal element. PPP Of course P ,= , as
0 P. If Q P is non-empty and upwards-directed, set p = supQ, which is dened because A is Dedekind
complete; then p = sup[Q] (since , being an automorphism, is surely order-continuous). If q
1
, q
2
Q,
there is a q Q such that q
1
q
2
q, so q
1
q
2
q q = 0. By 313Bc, supQ sup[Q] = 0, that is,
p p = 0. This means that p P and is an upper bound for Q in P. As Q is arbitrary, Zorns Lemma
tells us that P has a maximal element. QQQ
(b) Let a be a maximal element of P and set b = a, c = b \ a, e = 1 \ (a b c).
(i) ??? Suppose, if possible, that there is some d e such that d ,= d. By 381F, there is a non-zero
d e such that d d = 0. Consider a = a
1
d; then a a. But
a a = (a b) (a d) ((b d)
1
d) = 0
because
((b d)
1
d) d (a c d) = 0.
So a is a member of P strictly greater than a, which is impossible. XXX
Thus d = d for every d e; in particular, e = e.
(ii) Because a b = 0, b c (a b) = 0, and (a, b, c, e) is a partition of unity. Because (a b e)
includes b c e, c 1 \ (b c e) = a.
(c) Set b
tt
=
1
c b, b
t
= b \ b
tt
, a
t
=
1
b
t
, a
tt
=
1
b
tt
. Then everything required in the statement of
the lemma has been covered, with the possible exception of (b
t
c) = a
t
a
tt
; but as b
t
c = 1 \ (a b
tt
e),
(b
t
c) = 1 \ (a b
tt
e) = 1 \ (b c e) = a = a
t
a
tt
.
So the proof is complete.
381N Automorphism groups of Boolean algebras 481
381K Lemma Let A be a Dedekind complete Boolean algebra. Then every involution in Aut A is an
exchanging involution in the sense of 381G.
proof Let be an involution and let a
t
, a
tt
, b
t
, b
tt
, c, e be a partition of unity as in 381J. Then
c =
2
a
tt
= a
tt
, c a
tt
= 0,
so a
tt
= b
tt
= c = 0 and d = d whenever d 1 \ (a
t
a
t
) = e, while a
t
a
t
= 0.
381L The support of an automorphism For Dedekind complete Boolean algebras A, the following
concept is very useful. For any Aut A, set supp = infa : is supported by a, the support of .
Note that supp supports (381Ee); it is the smallest element of A supporting . By 381Ef, supp is also
supd : d d = 0. Observe that in the language of 381J supp is just a
t
a
tt
b
t
b
tt
c = 1 \ e.
Note that supp
1
= supp , and that supp(
t
) supp supp
t
for all ,
t
Aut A, by 381Ed.
Moreover, if , Aut A, then
supp(
1
) = (supp).
PPP By 381Ec, (supp) supports
1
, so supp(
1
) (supp); but also
supp
1
(supp(
1
)),
so (supp) supp(
1
). QQQ
381M In order to apply the argument of the next theorem simultaneously to the groups Aut A here
and to the groups Aut

A of the next section, I introduce the following terminology.

Denition If A is a Boolean algebra, a subgroup G of Aut A is full if whenever a
i
)
iI
is a partition of
unity in A,
i
)
iI
is a family in G, and Aut A is such that a =
i
a
i
whenever i I and a a
i
, then
G.
381N Theorem Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A. Then
every member of G is expressible as the product of at most eight involutions belonging to G.
proof (a) For each n N, let 1
n
be the set of elements of G expressible as the product of n or fewer
involutions belonging to G(so that 1
0
consists of the identity alone). It will be helpful to note that
1
1
n
,

1
1
n
whenever 1
n
and G. In particular, for instance,
1

1
= (
1
)
1
1
2
whenever 1
1
and G.
Note that if a A, G and a a = 0, then the involution (

a) belongs to G, because it is made

up of the actions of ,
1
and on A
a
, A
a
and A
1\(aa)
respectively, and G is full.
(b) If G there are a A and involutions
1
,
2
G such that a a = 0 and
1

2
is supported by
a. PPP Take a
t
, a
tt
, b
t
, b
tt
, c, e from 381J, and set a = b
t
c. Then certainly a = a
t
a
tt
is disjoint from a.
Set

1
= (

b
tt

c),
2
= (

a
t
a
tt

b
t
b
tt
).
Then both
1
,
2
are involutions in G. Now trace the action of
1

2
on each of the principal ideals A
a

a
,
A
b
, A
e
; we get

2
d =
1
d = d if d a
t
a
tt
,

2
d =
1
d = d if d b
tt
,

2
d = d = d if d e.
(I write the identity operator into these formulae in the hope of suggesting how the three functions act in
each case.) So supp(
1

2
) b
t
c = a. QQQ
(c) Suppose that , G, a A, is supported by a and a a = 0. Then there are
1
,
2
,
3
,
t
G
and

b a such that (i)
1
,
2
,
3
are all involutions supported by a

b (ii)
t

b a \

b (iii)
1

3
is
supported by

b. PPP Take a
t
, a
tt
, b
t
, b
tt
, c, e from 381J; then
482 Automorphism groups 381N
a
t
a
tt
b
t
b
tt
c = supp a.
Set
b = b
t
c,

b
t
= b
t
, c = c,

b = b =

b
t
c a.
Note that b

b a a = 0. Set
t
=
1
, so that

b = b = (a
t
a
tt
)
is disjoint from b =

b.
Set

1
= (

b
tt

c)(

a
t

b
t
)(

a
tt

b
tt
),
so that (just as in (b) above)
1
is supported by b
t
c = b. Set

1
= (

b) 1
1
,
2
=
1
(

b)
1
.
Examining the action of
2
on A
b
, we see that

2
d =
1

1
d =
1
1

1

1
d = d if d b,
so that supp
2
b = 0; but as
supp
2
supp
1
supp(

b) supp
1
= b

b,
supp
2


b.
Now we have an alternative expression for
2
, as follows:

2
=
1
(

b)
1
=
1
(

b
t

b
t
)(

c)(

b
tt

c)(

a
t

b
t
)(

a
tt

b
tt
)
=
1
(

b
t

b
t
)(

a
t

b
t
)(

c)(

b
tt

c)(

a
tt

b
tt
)
(because (

a
t

b
t
) commutes with (

b
tt

c) and (

c))
=
1
(

b
t

b
t
)(

a
t

b
t
)(

1 b
tt

1 a
tt

2 c)
(of course this calculation depends on the fact that c, b
tt
, a
tt
, c are all disjoint)
=
1
(

b
t

b
t
)(

a
t

b
t
)(

b
tt

c)(

a
tt

2 c)(

b
tt

c)
=
1
(

b
t

b
t
)(

b
tt

c)(

a
tt

2 c)(

a
t

b
t
)(

b
tt

c)
(because (

a
t

b
t
) commutes with (

b
tt

c) and (

a
tt

2 c))
=
1

where

2
= (

b
t

b
t
)(

b
tt

c)(

a
tt

2 c),

3
= (

a
t

b
t
)(

b
tt

c)
are involutions in G, each being a product of disjoint involutions. Now
supp
1
b

b,
supp
2
b

b
t
b
tt
c a
tt
c,
supp
3
a
t
b
t
b
tt
c
are all included in a

b, and we already know that

supp(
1

3
) = supp
2


b. QQQ
381O Automorphism groups of Boolean algebras 483
(d) We are ready for the assault. Take any G. Choose
n
)
nN
, a
n
)
nN
, b
n
)
nN
,
1n
)
nN
,

2n
)
nN
,
3n
)
nN
,
n
)
nN
as follows. Start by dening a
0
,
1
,
2
from as in (b), so that
1
,
2
1
1
,
a
0
a
0
= 0 and
1

2
is supported by a
0
. Set
0
=
1

2
,
0
= and b
0
= a
0
. For the inductive step,
given
n
and
n
in G, a
n
supp
n
and b
n
=
n
a
n
disjoint from a
n
, use (c) to nd
1n
,
2n
,
3n
1
1
,
a
n+1
b
n
and
n+1
G such that
n+1
a
n+1
b
n
\ a
n+1
,
n+1
=
1n

2n

3n

n
is supported by a
n+1
, and
all the
in
are supported by a
n
a
n+1
; set b
n+1
=
n+1
a
n+1
, and continue.
Evidently b
n
)
nN
is now non-increasing, while a
n
)
nN
is disjoint; consequently each of the sequences
a
2k
a
2k+1
)
kN
, a
2k+1
a
2k+2
)
kN
is disjoint.
(e) For r = 0, r = 1 dene
r
G by setting

r
d =
2k+r

1
2k+r+1
d if k N, d a
2k+r
a
2k+r+1
,
= d if d sup
nr
a
n
= 0.
(The denition is valid because supp
n
a
n
for every n, and produces a member of G because G is full.)
Now
r
1
3
.
PPP For each n N,

1
n+1
=
1
3n

1
2n

1
1n
,
and all these automorphisms are supported by a
n
a
n+1
. So if we set

ir
d =
i,2k+r
d if d a
2k+r
a
2k+r+1
,
= d if d sup
nr
a
n
= 0,
each

ir
will be in 1
1
, and

r
=

1
3r

1
2r

1
1r
belongs to 1
3
. QQQ
(f ) Consequently
1

0
1
6
. But
1

0
=
0
. PPP If d a
0
, then

0
d =
0
d a
0
,
1

0
d =
0
d.
If d a
2k+1
, where k N, then

0
d =
1
2k+1
d a
2k+1
,
1

0
d =
2k+1

1
2k+1
d = d =
0
d.
If d a
2k+2
, then

0
d =
2k+2
d a
2k+2
,
1

0
d =
1
2k+2

2k+2
d = d =
0
d;
and nally, if d sup
nN
a
n
= 0,

0
d =
1
d =
1

0
d =
0
d = d.
Putting these together,
0
=
1

0
. QQQ
(g) Finally,
=
1
2

1
1

0
1
8
.
As is arbitrary, the theorem is proved.
381O Corollary Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A. Then
any G other than the identity is expressible as a product of nitely many involutions in G with supports
included in supp.
proof In fact the construction in the proof of 381N achieves this almost automatically. Alternatively,
setting a = supp, let A
a
be the principal ideal of A generated by a, and set
484 Automorphism groups 381O
G
a
= A
a
: G, supp a.
It is easy to check that G
a
is a full subgroup of Aut A
a
, so that A
a
is expressible as a product of involutions
belonging to G
a
. These are all of the form
i
A
a
where
i
G and supp
i
a, so that every
i
is also an
involution and is the product of the
i
.
381P Denition Let A be a Boolean algebra, and G a subgroup of the automorphism group Aut A. I
will say that G has many involutions if for every non-zero a A there is an involution G which is
supported by a.
381Q Lemma Let A be an atomless homogeneous Boolean algebra. Then Aut A has many involutions,
and in fact every non-zero element of A is the support of an exchanging involution.
proof If a A 0, then there is a b such that 0 ,= b a. By 381C there is a Aut A such that
b = a \ b; now (

a \ b) is an involution with support a.

381R Lemma Let A be a Dedekind complete Boolean algebra, and G a full subgroup of Aut A with
many involutions. Then every non-zero element of A is the support of an involution belonging to G.
proof By the denition 381P,
C = supp : G is an involution
is order-dense in A. So if a A 0 there is a disjoint B C such that sup B = a (313K). For each b B
let
b
G be an involution with support B. Dene G by setting d =
b
d for d b B, d = d if
d a = 0; then G is an involution with support a.
381S Theorem Let A be a Dedekind complete Boolean algebra, and G a full subgroup of Aut A with
many involutions. Then a subset H of G is a normal subgroup of G i it is of the form
: G, supp I
for some ideal I A which is G-invariant, that is, such that a I for every a I, G.
proof (a) I deal with the easy implication rst. Let I A be a G-invariant ideal and set H = :
G, supp I. Because the support of the identity automorphism is 0 I, H. If , H and
G, then
supp() supp supp I,
supp(
1
) = supp I,
supp(
1
) = (supp) I
and ,
1
,
1
all belong to H; so H G.
(b) For the rest of the proof, therefore, I suppose that H is a normal subgroup of G and seek to express
it in the given form. We can in fact describe the ideal I immediately, as follows. Set
J = a : a A, H whenever G is an involution, supp a;
then 0 J and a J whenever a b J. Also a J whenever a J, G. PPP If G is an involution,
supp a then
1
=
1
is an involution in G and
supp
1
=
1
(supp ) a,
so
1
H and =
1

1
H. As is arbitrary, a J. QQQ
I do not know how to prove directly that J is an ideal, so let us set
I = a
1
a
2
. . . a
n
: a
1
, . . . , a
n
J;
then I A, and a I for every a I, G.
381T Automorphism groups of Boolean algebras 485
(c) If a A, H and a a = 0 then a J. PPP If a = 0, this is trivial. Otherwise, let G be an
involution with supp a; say = (

c) where b c a. By 381R there is an involution

1
G such that
supp
1
= b; say
1
= (

b
t

1
b
tt
) where b
t
b
tt
= b. Set
c
t
= b
t
, c
tt
= b
tt
= c \ c
t
,

2
=
1

1
= (

b
t

1
b
tt
)(

c
t

1

1 c
tt
),
3
= (

b
t

c
t
),
=
1
2

2

1
H,
=
1
3

3

1
=
1
3

1
2

2

1
2

1

2
H.
Now
supp(
2

1
) = (supp
2
) = (b c) a
is disjoint from
supp
3
= b
t
c
t
a,
so
3
commutes with
2

1
, and
=
1
3

1
2

3

1
2

1

2
=
1
3

1
2

3

2
= (

b
t

c
t
)(

b
t

1
b
tt
)(

c
t

1

1 c
tt
)(

b
t

c
t
)(

b
t

1
b
tt
)(

c
t

1

1 c
tt
)
= (

b
t

c
t
)(

b
tt

c
tt
)
= .
So H. As is arbitrary, a J. QQQ
(d) If = (

b) is an involution in G and a J, then H. PPP By 381R again, there is an involution

G such that supp = a; because a J, H. Express as (

a
t

a
tt
) where a
t
a
tt
= a. Set b
t
= a
t
,
b
tt
= a
tt
, so that = (

a
t

b
t
)(

a
tt

b
tt
), and

1
=
1
= (

a
t

a
tt
)(

b
t

1 b
tt
) H.
As
1
(a
t
b
t
) = a
tt
b
tt
is disjoint from a
t
b
t
, a
t
b
t
J, by (c), and
1
= (

a
t

b
t
) H; similarly,
a
tt
b
tt
J, so
2
= (

a
tt

b
tt
) H and =
1

2
belongs to H. QQQ
(e) If G is an involution and supp I, then H. PPP Express as (

b). Let a
1
, . . . , a
n
J be
such that a b a
1
. . . a
n
. Set
c
j
= a a
j
\ sup
i<j
a
i
, b
j
= c
j
,
j
= (

c
j
b
j
)
for 1 j n; then every c
j
belongs to J, so every
j
belongs to H (by (d)) and =
1
. . .
n
H. QQQ
(f ) If G and supp I then H. PPP By 381O, is a product of involutions in G all with supports
included in supp; by (e), they all belong to H, so also does. QQQ
(g) We are nearly home. So far we know that I is a G-invariant ideal and that H whenever G,
supp I. On the other hand, supp I for every H. PPP Take a
t
, a
tt
, b
t
, b
tt
, c from Frolks theorem
(381J). Then
a
t
a
t
= b
t
b
t
= . . . = c c = 0,
so a
t
, . . . , c all belong to J, by (c), and supp = a
t
. . . c belongs to I. QQQ
So H is precisely the set of members of G with supports in I, as required.
381T Corollary Let A be a homogeneous Dedekind complete Boolean algebra. Then Aut A is simple.
486 Automorphism groups 381T
proof If A is 0 or 0, 1 this is trivial. Otherwise, let H be a normal subgroup of Aut A. Then by 381S
and 381Q there is an invariant ideal I of A such that H = : supp I. But if H is non-trivial so is I;
say a I 0. If a = 1 then certainly 1 I and H = Aut A. Otherwise, there is a Aut A such that
a = 1 \ a (as in 381C), so 1 \ a I, and again 1 I and H = Aut A.
Remark I ought to remark that in fact Aut A is simple for any homogeneous Dedekind -complete Boolean
algebra; see

St ep anek & Rubin 89, Theorem 5.9b.
381X Basic exercises >>>(a) Let X be a set and an algebra of subsets of X containing all singleton
sets. Show that Aut can be identied with the group of bijections : X X such that [E] ,

1
[E] for every E .
>>>(b) Let A be a Boolean algebra and G any subgroup of Aut A. Let H be the set of those Aut A
such that for every non-zero a A there are a non-zero b a and a G such that c = c for every c b.
Show that H is a full subgroup of Aut A, the smallest full subgroup of A including G.
>>>(c) Let A be a Dedekind complete Boolean algebra and G any subgroup of A. Show that an element
of Aut A belongs to the full subgroup of Aut A generated by G i there are a partition of unity a
i
)
iI
in A
and a family
i
)
iI
in G such that a =
i
a whenever i I and a a
i
.
(d) For a Boolean algebra A, let us say that a subgroup G of Aut A is countably full if whenever a
i
)
iI
is a countable partition of unity in A,
i
)
iI
is a family in G, and Aut A is such that a =
i
a
i
whenever i I and a a
i
, then G. Show that if A is a Dedekind complete Boolean algebra and G is
a countably full subgroup of Aut A, then every member of G is expressible as a product of at most eight
involutions belonging to G.
>>>(e) Let X be any set. Show that any automorphism of the Boolean algebra TX is expressible as a
product of at most two involutions.
(f ) Recall that in any group G, a commutator in G is an element of the form ghg
1
h
1
where g, h G.
Show that if A is a Dedekind complete Boolean algebra and G is a subgroup of Aut A with many involutions
then every involution in G is a commutator in G, so that every element of G is expressible as a product of
nitely many commutators.
(g) Give an example of a Dedekind complete Boolean algebra A such that not every member of Aut A is
a product of commutators in Aut A.
(h) Let A be a Dedekind complete Boolean algebra, and suppose that Aut A has many involutions. Show
that if H Aut A then every member of H is expressible as the product of at most eight involutions
belonging to H.
(i) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions.
Show that the partially ordered set H of normal subgroups of G is a distributive lattice, that is, HK
1
K
2
=
(H K
1
)(H K
2
), H(K
1
K
2
) = HK
1
HK
2
for all H, K
1
, K
2
H.
(j) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions.
Show that if H is a the normal subgroup of G generated by a nite subset of G, then it is the normal
subgroup generated by a single involution.
(k) Let A be a Dedekind complete Boolean algebra and G a full subgroup of Aut A with many involutions.
Show (i) that there is an involution G such that every member of G is expressible as a product of
conjugates of in G (ii) any proper normal subgroup of G is included in a maximal proper normal subgroup
of G.
(l) Let G be any group. Show that if , G are involutions then is conjugate to its inverse.
381 Notes Automorphism groups of Boolean algebras 487
381Y Further exercises (a) Find a Dedekind -complete Boolean algebra A with an automorphism
which cannot be expressed either as a product of nitely many involutions in Aut A, nor as a product of
nitely many commutators in Aut A. (This seems to require a certain amount of ingenuity.)
(b) Let X be a set and a countably generated -subalgebra of subsets of X. (i) Show that if f : X X
is a bijection such that = E : E X, f
1
[E] , then there are disjoint E
t
, E
tt
, F
t
, F
tt
, G
such that f
1
[E
t
] = F
t
, f
1
[E
tt
] = F
tt
, f
1
[F
tt
] = G, f
1
[F
t
G] = E
t
E
tt
, and f(x) = x for every
x X (E
t
E
tt
F
t
F
tt
G. (Hint: there is a sequence E
n
)
nN
in such that E
n
f
1
[E
n
] = for
every n and

nN
E
n
= x : f(x) ,= x.) (In particular, any member of Aut has a support.) (ii) Show
that every involution in Aut is an exchanging involution.
(c) Let X be a set and a countably generated -subalgebra of subsets of X. Show that any member of
Aut is expressible as a product of at most eight involutions in Aut .
(d) Let A be a homogeneous Boolean algebra which is isomorphic to the simple power A
N
. (For instance,
A could be the measure algebra of Lebesgue measure on R.) Show that any automorphism of A is the
product of at most ve exchanging involutions. (Cf.

St ep anek & Rubin 89, Corollary 5.9a(ii).)
381Z Problem In 381N, is eight best possible?
381 Notes and comments The ideas above are adapted from

St ep anek & Rubin 89 and Fathi 78.
Lemma 381J is a form of what is sometimes called Frolks theorem, following Frolk 68.
The two main results 381N and 381S, as written out above, both involve careful algebra. It seems
to me that we can distinguish two essential methods. (i) There are arguments involving nitely many
automorphisms, carefully pieced together from descriptions of their actions on dierent parts of the algebra,
as in (a)-(c) of the proof of 381N, and the whole of the proof of 381S; similar ideas can be used in 381Xf. It is
in these that I believe that the cycle notation introduced in 381G-381I can be of value. Generally the hope
is that we can use intuitions derived from the theory of permutation groups (that is, the case A = TX) to
guide us. (ii) There is the argument in 381N involving a sequence of automorphisms, designed to express an
automorphism
0
supported by an element a
0
as the product of an automorphism
0
supported by sup
kN
a
k
with an automorphism
1
supported by sup
k1
a
k
, so chosen that the actions of the
r
on sup
k1
a
k
cancel
out and we are left with
0
as the residue. (For an account of the origins of this idea see

St ep anek & Rubin
89.) Since we know of no automorphisms except those which can be derived from the original automorphism
, the method has to be to some extent constructive. The idea is that each
n+1
is not exactly a copy on
a
n+1
of the preceding
n
, but a modication of it by involutions. At each stage of the induction we have
to mention an auxiliary element
n
of G in order to be sure that there will be room (in b
n
=
n
a
n
) for the
next step, safely disjoint from the preceding a
k
. When we come to build the mutually cancelling pair
0
,
1
we nd that they incorporate the modiers
1n
,
2n
,
3n
, which can be assembled into the involutions

ir
in part (f) of the proof.
I note that the assumption of Dedekind completeness (as opposed to Dedekind -completeness) in 381N
is used only in parts (b) and (c) of the proof, when applying Frolks theorem. Consequently we have a slight
generalization possible (381Xd); but we do need the full hypothesis for the theorem as stated (381Ya). There
is however a very important special case, when A is a countably generated -algebra, for which we have a
version of Frolks theorem available for dierent reasons (381Yb), and can get a corresponding theorem to
match 381N (381Yc).
A natural question arising from 381T is: does every homogeneous Boolean algebra have a simple au-
tomorphism group? This leads into deep water. As remarked after 381T, every homogeneous Dedekind
-complete algebra has a simple automophism group. Using the continuum hypothesis, it is possible to
construct a homogeneous Boolean algebra which does not have a simple automorphism group; but as far as
I am aware no such construction is known which does not rely on some special axiom outside ordinary set
theory. See

St ep anek & Rubin 89, 5.
In 381Z I ask whether the number eight appearing in 381N is actually best possible. The argument is
complex enough to make it seem that there may be room for improvement see 381Xe and 381Yd. Ornstein
488 Automorphism groups 381 Notes
& Shields 73
1
present examples of automorphisms in the full subgroup G = Aut

A of measure-preserving
automorphisms of the Lebesgue probability algebra which are not conjugate (in G) to their inverses, and
therefore cannot be expressible as the product of two involutions in G.
382 Automorphism groups of measure algebras
I turn now to the group of measure-preserving automorphisms of a measure algebra, seeking to apply
the results of the last section. The principal theorems are 382D, which is a straightforward special case of
381N, and 382I, corresponding to 381T. I give another example of the use of 381S to describe the normal
subgroups of Aut

A (382J).
382A Denition Let (A, ) be a measure algebra. I will write Aut

A for the set of all measure-
preserving automorphisms of A. This is a group, being a subgroup of the group Aut A of all Boolean
automorphisms of A.
382B Lemma Let (A, ) be a measure algebra, and a
i
)
iI
, b
i
)
iI
two partitions of unity in A. Assume
either that I is countable
or that (A, ) is localizable.
Suppose that for each i I we have a measure-preserving isomorphism
i
: A
a
i
A
b
i
between the
corresponding principal ideals. Then there is a unique Aut

A such that c =
i
c whenever i I and
c a
i
.
proof (Compare 381B.) By 322K, we may identify A with each of the simple products

iI
A
a
i
,

iI
A
b
i
;
now corresponds to the isomorphism between the two products induced by the
i
.
382C Corollary If (A, ) is a localizable measure algebra, then, in the language of 381M, Aut

A is a
full subgroup of Aut A.
382D Theorem Let (A, ) be a localizable measure algebra. Then every measure-preserving automor-
phism of A is expressible as the product of at most eight measure-preserving involutions.
proof This is immediate from 382C and 381N.
382E Lemma If (A, ) is a homogeneous semi-nite measure algebra, it is -nite, therefore localizable.
proof If A = 0, this is trivial. Otherwise there is an a A such that 0 < a < . The principal ideal
A
a
is ccc (322G), so A also is, and (A, ) must be -nite (by 322G again).
382F Lemma Let (A, ) be a homogeneous semi-nite measure algebra.
(a) If a
i
)
iI
, b
i
)
iI
are partitions of unity in A with a
i
= b
i
for every i, there is a Aut

A such
that a
i
= b
i
for each i.
(b) If (A, ) is totally nite, then whenever a
i
)
iI
, b
i
)
iI
are disjoint families in A with a
i
= b
i
for
every i, there is a Aut

A such that a
i
= b
i
for each i.
proof (a) By 382E, (A, ) is -nite, therefore localizable. For each i I, the principal ideals A
a
i
, A
b
i
are homogeneous, of the same measure and the same Maharam type (being (A) if a
i
,= 0, 0 if a
i
= 0).
Because they are ccc, they are of the same magnitude, as dened in 332G, and there is a measure-preserving
isomorphism
i
: A
a
i
A
b
i
(332J). By 382B there is a measure-preserving automorphism : A A such
that d =
i
d for every i I, d a
i
; and this serves.
(b) Set a

= 1 \ sup
iI
a
i
, b

= 1 \ sup
iI
b
i
. We must have
a

= 1

iI
a
i
= 1

iI
b
i
= b

,
so adding a

, b

to the families we obtain partitions of unity to which we can apply the result of (a).
1
I am indebted to G.Hjorth for the reference.
382J Automorphism groups of measure algebras 489
382G Lemma (a) If (A, ) is an atomless semi-nite measure algebra, then Aut A and Aut

A have
many involutions.
(b) If (A, ) is an atomless localizable measure algebra, then every element of A is the support of some
involution in Aut

A.
proof (a) If a A 0, then by 332A there is a non-zero b a, of nite measure, such that the principal
ideal A
b
is (Maharam-type-)homogeneous. Now because A is atomless, there is a c b such that c =
1
2
b
(331C), so that A
c
and A
b\c
are isomorphic measure algebras. If : A
c
A
b\c
is any measure-preserving
isomorphism, then = (

b \ c) is an involution in Aut

A (and therefore in Aut A) supported by a.
(b) Use 382C, (a) and 381R.
382H Corollary Let (A, ) be an atomless localizable measure algebra. Then
(a) the lattice of normal subgroups of Aut A is isomorphic to the lattice of Aut A-invariant ideals of A;
(b) the lattice of normal subgroups of Aut

A is isomorphic to the lattice of Aut

A-invariant ideals of A.
proof Use 381S. Taking G to be either Aut A or Aut

A, and 1 to be the family of G-invariant ideals in
A, we have a map I H
I
= : G, supp I from 1 to the family H of normal subgroups of G. Of
course this map is order-preserving; 381S tells us that it is surjective; and 382Gb tells us that it is injective
and its inverse is order-preserving, since if a I J there is a G with supp = a, so that H
I
H
J
.
Thus we have an order-isomorphism between H and 1.
382I Normal subgroups of Aut A and Aut

A 381S provides the machinery for a full description of
the normal subgroups of Aut A and Aut

A when (A, ) is an atomless localizable measure algebra, as we
know that they correspond exactly to the invariant ideals of A. The general case is complicated. But the
following are simple enough.
Theorem Let (A, ) be a homogeneous semi-nite measure algebra.
(a) Aut A is simple.
(b) If (A, ) is totally nite, Aut

A is simple.
(c) If (A, ) is not totally nite, Aut

A has exactly one non-trivial proper normal subgroup.
proof (a) This is a special case of 381T.
(b)-(c) The point is that the only possible Aut

A-invariant ideals of A are 0, A
f
and A. PPP If A is
0 or 0, 1 this is trivial. Otherwise, A is atomless. Let I A be an invariant ideal.
(i) If I , A
f
, take a I with a = . By 382E, A is -nite, so a has the same magnitude as 1. By
332I, there is a partition of unity e
n
)
nN
in A with e
n
= 1 for every n; setting b = sup
nN
e
2n
, b
t
= 1 \ b, we
see that both b and b
t
are of innite measure. Similarly we can divide a into c, c
t
, both of innite measure.
Now by 332J the principal ideals A
b
, A
b
, A
c
, A
1\c
are all isomorphic as measure algebras, so that there are
automorphisms , Aut

A such that
c = b, c = b
t
.
But this means that both b and b
t
belong to I, so that 1 = b b
t
I and I = A.
(ii) If I A
f
and I ,= 0, take any non-zero a I. If b is any member of A, then (because A is
atomless) b can be partitioned into b
0
, . . . , b
n
, all of measure at most a. Then for each i there is a b
t
i
a
such that b
t
i
= b
i
; since this common measure is nite, (1 \ b
t
i
) = (1 \ b
i
). By 332J and 382Fa, there is
a
i
Aut

A such that
i
b
t
i
= b
i
, so that b
i
belongs to I. Accordingly b I. As b is arbitrary, I = A
f
.
Thus the only invariant ideals of A are 0, A
f
and A. QQQ
By 382Hb we therefore have either one, two or three normal subgroups of Aut

A, according to whether
1 is zero, nite and not zero, or innite.
Remark For the Lebesgue probability algebra, (b) is due to Fathi 78. The extension to algebras of
uncountable Maharam type is from Choksi & Prasad 82.
382J The language of 352 oers a way of describing another case.
490 Automorphism groups 382J
Proposition Let (A, ) be an atomless totally nite measure algebra. For each innite cardinal , let e

be
the Maharam-type- component of A, and let K be : e

,= 0. Let H be the lattice of normal subgroups

of Aut

A. Then
(i) if K is nite, H is isomorphic, as partially ordered set, to TK;
(ii) if K is innite, then H is isomorphic, as partially ordered set, to the lattice of solid linear subspaces
of

.
proof (a) Let 1 be the family of Aut

A-invariant ideals of A, so that H

= 1, by 382Hb. For a, b A,
say that a _ b if there is some k N such that (a e

) k (b e

) for every K. Then an ideal

I of A is Aut

A-invariant i a I whenever a _ b I. PPP () Suppose that I is Aut

A invariant and
that b I, (a e

) k (b e

) for every K. Then for each we can nd a

1
, . . . , a
k
such that
a e

= sup
ik
a
i
and a
i
(b e

) for every i. Now there are measure-preserving automorphisms

i
of the principal ideal A
e

such that
i
a
i
b. Setting
i
d = sup
K

i
(d e

) for every d A, and

a
i
= sup
K
a
i
, we have
i
Aut

A and
i
a
i
b, so a
i
I for each i; also a = sup
ik
a
i
, so a I. ()
On the other hand, if a A and Aut

A, then
(a e

) = (a e

) = (a e

)
for every K, because e

= e

, so that a _ a. So if I satises the condition, [I] I for every

Aut

A and I 1. QQQ
(b) Consequently, for I 1 and K, e

I i there is some a I such that a a

,= 0, since in this
case e

_ a. (This is where I use the hypothesis that (A, ) is totally nite.) It follows that if K is nite,
any I 1 is the principal ideal generated by supe

: e

I. Conversely, of course, all such ideals are

Aut

A-invariant. Thus 1 is in a natural order-preserving correspondence with TK, and H

= TK.
(c) Now suppose that K is innite; enumerate it as
n
)
nN
. Dene : A

by setting a =
(a e

n
)/ (e

n
))
nN
for a A; so that
a _ b i there is some k such that a kb,
a (a b) a +b 2(a b)
for all a, b A, while (1
A
) is the standard order unit 1 of

. Let | be the family of solid linear subspaces

of

and dene functions I V

I
: 1 |, U J
U
: | 1 by saying
V
I
= f : f

, [f[ ka for some a I, k N,

J
U
= a : a A, a U.
The properties of just listed ensure that V
I
|, J
U
1 for every I 1, U |. Of course both I V
I
and U J
U
are order-preserving. If I 1, then
J
V
I
= a : b I, a _ b = I,
Finally, V
J
U
= U for every U |. PPP
V
J
U
= f : a A, k N, [f[ ka U U
because U is a solid linear subspace. But also, given g U, there is an a A such that (a e

n
) =
min(1, [g(n)[) (e

n
) for every n (because A is atomless); in which case
a [g[ max(1, |g|

)a
so a J
U
and g V
J
U
. Thus U = V
J
U
. QQQ So the functions I V
I
and U J
U
are the two halves of an
order-isomorphism between 1 and |, and H

= 1

= |, as claimed.
382X Basic exercises >>>(a) Let (A, ) be a localizable measure algebra. For each innite cardinal ,
let e

be the Maharam-type- component of A. (i) Show that Aut

A is a simple group i either there is
just one innite cardinal such that e

,= 0, that e

has nite measure and all the atoms of A (if any) have
dierent measures or A is purely atomic and there is just one pair of atoms of the same measure or A is
purely atomic and all its atoms have dierent measures. (ii) Show that Aut A is a simple group i either
382Ye Automorphism groups of measure algebras 491
(A, ) is -nite and there is just one innite cardinal such that e

,= 0 and A has at most one atom or

A is purely atomic and has at most two atoms.
(b) Let (A, ) be a localizable measure algebra. (i) Show that Aut

A is simple i it is isomorphic to one
of the groups , Z
2
or Aut

where is an innite cardinal and (B

) is the measure algebra of the

usual measure on 0, 1

. (ii) Show that Aut A is simple i it is isomorphic to one of the groups , Z

2
or
Aut B

.
(c) Show that if (A, ) is a semi-nite measure algebra of magnitude greater than c, its automorphism
group Aut

A is not simple.
(d) Let (A, ) be an atomless localizable measure algebra. For each innite cardinal write e

for the
Maharam-type- component of A. For , Aut

A show that belongs to the normal subgroup of Aut

A
generated by i there is a k N such that
mag(e

supp) k mag(e

supp) for every innite cardinal ,

writing mag a for the magnitude of a, and setting k = if k > 0 and is an innite cardinal.
>>>(e) Let (A, ) be the measure algebra of Lebesgue measure on R. For n N set e
n
= [n, n]

A. Let
G Aut

A be the group consisting of measure-preserving automorphisms such that supp e
n
for some
n. Show that G is simple. (Hint: show that G is the union of an increasing sequence of simple subgroups.)
(f ) Let (A, ) be an atomless totally nite measure algebra. Let H be the lattice of normal subgroups of
Aut A. Show that H is isomorphic, as partially ordered set, to TK for some countable set K.
(g) Let (A, ) be an atomless localizable measure algebra which is not -nite, and suppose that (A
a
) =
(A
b
) whenever a, b A and 0 < a b < . Let be the magnitude of A. (i) Show that the lattice
H of normal subgroups of Aut

A is well-ordered, with least member , next member : (a) <
whenever a < , and one member H

for each innite cardinal less than or equal to , setting

H

= : Aut

A, mag(a) whenever mag a ,
where mag a is the magnitude of a. (ii) Show that the lattice H
t
of normal subgroups of Aut A is well-
ordered, with least member and one member H
t

for each innite cardinal less than or equal to ,

setting
H
t

= : Aut A, mag(a) whenever mag a .

382Y Further exercises (a) Let (A, ) be an atomless totally nite measure algebra. Show that
Aut

A, Aut A have the same (cardinal) number of normal subgroups.
(b) Let X be a set. Show that Aut TX has one normal subgroup if #(X) 1, two if #(X) = 2, three if
#(X) = 3 or 5 #(X) , four if #(X) = 4 or #(X) =
1
.
(c) Let (A, ) be a totally nite measure algebra and : A A a measure-preserving homomorphism.
Take any a A and set c = sup
n1

n
a. (i) Show that a c and that c = sup
mn

m
a for every n N.
(ii) Set a
n
=
n
a \ sup
1i<n

i
a for n 1. Show that

n=1
n (a a
n
) = c. (Hint: For j < k set
a
jk
=
k
a
j
a \ sup
j<i<k

i
a =
j
(a a
kj
). Show that, for any n, a
jk
)
jn<k
is disjoint and has union
sup
in

i
a.)
(d) Let (X, , ) be a totally nite measure space and f : X X an inverse-measure-preserving function.
Take E and set F = x : n 1, f
n
(x) E. (i) Show that E F is negligible. (ii) For x E F set
k
x
= minn : n 1, f
n
(x) E. Show that
_
E
k
x
(dx) = F. (This is a simple form of the Recurrence
Theorem.)
(e) Let (A, ) be a totally nite measure algebra and : A A a measure-preserving Boolean homo-
morphism. Show that there are a measure algebra (B, ), a measure-preserving automorphism : B B
492 Automorphism groups 382Ye
and a closed subalgebra C of B such that [C] C and (C, C, C) is isomorphic to (A, , ). (Hint: Take
B
1
A
N
to be the subalgebra consisting of sequences a = (
0
,
1
, . . . ) such that

n=0
(
n+1
.
n
) < .
For a B
1
set
1
a = lim
n

n
,
1
a = (
1
,
2
, . . . ). Let 1 be the ideal a :
1
a = 0 and let B be the
quotient B
1
/1; dene : B B by setting a

= (
1
a)

. For a A let a

B be the equivalence class of

the sequence
n
a)
nN
; set C = a

: a A.) (This is an abstract version of a construction known as the

natural extension of an inverse-measure-preserving function; see Petersen 83, 1.3G.)
(f ) Let (X, , ) be a measure space in which is countably generated as -algebra, and write Aut

for
the group of automorphisms : such that (E) = E for every E . Show that every member
of Aut

is expressible as a product of at most eight involutions belonging to Aut

. (Hint: 381Yc.)
382 Notes and comments This section is short because there are no substantial new techniques to be
developed. 382D is simply a matter of checking that the hypotheses of 381N are satised (and these hy-
potheses were of course chosen with 382D in mind), and 382I is similarly direct from 381S. 382I-382J, 382Xd
and 382Xg are variations on a theme. In a general Boolean algebra A with a group G of automorphisms,
we have a transitive, reexive relation _
G
dened by saying that a _
G
b if there are
1
, . . . ,
k
G such
that a sup
ik

i
b; the point about localizable measure algebras is that the functions Maharam type and
magnitude enable us to describe this relation when G = Aut

A, and the essence of 381S is that in that
context belongs to the normal subgroup of G generated by i supp _
G
supp.
Most of the work of this chapter is focused on atomless measure algebras. There are various extra
complications which appear if we allow atoms. The most striking are in the next section; here I mention
only 382Xa and 382Yb.
383 Outer automorphisms
Continuing with the investigation of the abstract group-theoretic nature of the automorphism groups
Aut A, Aut

A, I devote a section to some remarkable results concerning isomorphisms between them.
Under any of a variety of conditions, any isomorphism between two groups Aut A, Aut B must correspond
to an isomorphism between the underlying Boolean algebras (383E, 383F, 383J, 383M); consequently Aut A
has few, or no, outer automorphisms (383G, 383K, 383O). I organise the section around a single general
result (383D).
383A Lemma Let A be a Boolean algebra and G a subgroup of Aut A which has many involutions
(denition: 381P). Then for every non-zero a A there is an automorphism G, of order 4, which is
supported by a.
proof Let G be an involution supported by a. Let b a be such that b ,= b. Then at least one
of b \ b, b \ b = (b \ b) is non-zero, so in fact both are. Let be an involution supported by b \ b.
Then =
1
is an involution supported by b \ b, so commutes with , and the product is an
involution. But this means that = has order 4, and of course it is supported by a because and
both are.
383B A note on supports Since in this section we shall be looking at more than one automorphism
group at a time, I shall need to call on the following elementary extension of a fact in 381Ec. Let A and
B be Boolean algebras, and : A B a Boolean isomorphism. If Aut A is supported by a A, then

1
Aut B is supported by a. (Use the same argument as in 381Ec.) Accordingly, if a is the support
of then a will be the support of
1
, as in 381L.
383C Lemma Let A and Bbe two Boolean algebras, and G a subgroup of Aut A with many involutions.
If
1
,
2
: A B are distinct isomorphisms, then there is a G such that
1

1
1
,=
2

1
2
.
proof Because
1
,=
2
, =
1
2

1
is not the identity automorphism on A, and there is some non-zero a A
such that a a = 0. Let G be an involution supported by a; then
1
is supported by a, so cannot
be equal to , and
1

1
1
,=
2

1
2
.
383D Outer automorphisms 493
383D Theorem Let A and B be Dedekind complete Boolean algebras and G and H subgroups of
Aut A, Aut B respectively, both having many involutions. Let q : G H be an isomorphism. Then there
is a unique Boolean isomorphism : A B such that q() =
1
for every G.
proof (a) The rst half of the proof is devoted to setting up some structures in the group G. Let G be
any involution. Set
C

= : G, = ,
the centralizer of in G;
U

= : C

, =
1
,
1
=
1
for every C

,
the set of involutions in C

commuting with all their conjugates in C

, together with the identity,

V

= : G, = for every U

,
the centralizer of U

in G,
S

=
2
: V

,
W

= : G, = for every S

,
the centralizer of S

in G.
(b) The point of this list is to provide a purely group-theoretic construction corresponding to the support
of in A. In the next few paragraphs of the proof (down to (f)), I set out to describe the objects just
introduced in terms of their action on A. First, note that is an exchanging involution (381K); express it
as (

a
t

a
tt
), so that the support of is a

= a
t
a
tt
.
(c) I start with two elementary properties of C

:
(i) (a

) = a

for every C

. PPP As remarked in 381L, the support of =

1
is (a

), so this
must be a

. QQQ
(ii) If C

and is not supported by a

, there is a non-zero d 1 \ a

such that d d = 0, by
381Ef.
(d) Now for the properties of U

:
(i) If U

, then is supported by a

.
PPP()??? Suppose rst that there is a d 1 \ a

such that d (d
2
d) = 0. Let G be an involution
supported by d. Then supp supp = 0, so C

. There is a c d such that c ,= c, so

1
c =
2
c =
2
c,
because d (c
2
c) = 0, while

1
c =
2

1
c,
because d
1
c = 0; but this means that
1
c ,=
1
c, so and
1
do not commute, and
/ U

. XXX
()??? Suppose that
2
is not supported by a

. Then, as remarked in (b), there is a non-zero

d 1 \ a

such that
2
d d = 0. Now d ,
2
d, so d , d; set d
t
= d \ d. Then d
t
d
t
= d
t

2
d
t
= 0 and
0 ,= d
t
1 \ a

; but this is impossible, by (). XXX

() Thus
2
d = d for every d 1 \ a

. ??? Suppose, if possible, that is not supported by a

. Then
there is a non-zero d 1 \ a

such that d d = 0. By 383A, there is a G, of order 4, supported

by d. Because d a

= 0, C

. Because ,=
1
, there is a c d such that c ,=
1
c; but now
c d =
1
c d = 0, so

1
c =
2
c = c ,=
1
c =
2

1
c =
1
c,
and does not commute with its conjugate
1
, contradicting the assumption that U

. XXX
So is supported by a

, as claimed. QQQ
(ii) If u A and u = u, then
u
U

, where
494 Automorphism groups 383D

u
d = d if d u,
u
d = d if d u = 0,
that is,
u
= (

a
t
u

a
tt
u). PPP For any Aut A,

1
= (

(a
t
u)

1 (a
tt
u))
(381Hb). () Accordingly

1
= (

a
tt
u

a
t
u) =
u
and
u
C

. () If C

, then
=
1
= (

a
t

1 a
tt
) = (

a
t

a
tt
).
So

1
= (

(a
t
u)

1 (a
tt
u)) = (

a
t
u

a
tt
u) =
u
.
Now if v = v then
u

v
=
uZv
=
v

u
; in particular,
u

u
=
u

u
. As is arbitrary,
u
U

. QQQ
In particular, of course, =
1
belongs to U

.
(e) The two parts of (d) lead directly to the properties we need of V

.
(i) V

, because U

. Consequently a

= a

for every V

.
(ii) If V

then d d d for every d a

. PPP??? Suppose, if possible, otherwise. Set u

0
= d d,
so that u
0
= u
0
, and u = u
0
\ u
0
,= 0; also u a

= a

. Since u
0
= u
0
= u
0
, u = u. Set
v = u a
t
, so that u = v v and v ,= v. Because u v (u
0
u) = 0,

u
v = v ,= v =
u
v,
which is impossible. XXXQQQ
(iii) It follows that
2
d = d whenever V

and d a

. PPP Let e be the support of . Recall that

e = supc : c c = 0 (381L), so that d e = supc : c d, c c = 0. Now if c a

and c c = 0, we
know that c c c, so in fact c c. This shows that (d e) (d e). Also, because = , by
(i), we have

2
(d e) (d e) = (d e)
2
(d e) = d e.
Of course
2
(d \ e) = d \ e, so
2
d d. This is true for every d a

. But as also
2
a

= a

= a

,
2
d = d
for every d a

. QQQ
(iv) The nal thing we need to know about V

is that V

whenever G and supp a

= 0;
this is immediate from (d-i) above.
(f ) From (e-iii), we see that if S

then supp a

= 0. But we also see from (e-iv) that if

0 ,= c 1 \ a

there is an involution in S

supported by c; for there is a member of G, of order 4,

supported by c, and now V

so
2
S

, while
2
is an involution.
(g) Consequently, W

is just the set of members of G supported by a

. PPP (i) If supp a

and S

,
then supp a

= 0, as noted in (e), so = ; as is arbitrary, W

. (ii) If supp , a

, then take a
non-zero d 1 \ a

such that d d = 0. Let S

be an involution supported by d; then if c d is such

that c ,= c,
c ,= c = c,
and ,= so / W

. QQQ
(h) We can now return to consider the isomorphism q : G H. If G is an involution, then q() H
is an involution, and it is easy to check that
q[C

] = C
q()
,
q[U

] = U
q()
,
q[V

] = V
q()
,
383E Outer automorphisms 495
q[S

] = S
q()
,
q[W

] = W
q()
,
dening C
q()
, . . . , W
q()
H as in (a) above. So we see that, for any G,
supp supp suppq() suppq().
(i) Dene : A B by writing
a = supsuppq() : G is an involution and supp a
for every a A. Evidently is order-preserving. Now if a A, G is an involution and supp , a,
suppq() , a. PPPThere is a G, of order 4, supported by supp \ a. Now
2
is an involution supported by
supp, so suppq(
2
) suppq(). On the other hand, if
t
G is an involution supported by a, then V

and
2
S

, so q(
2
) S
q(

)
and suppq(
2
) suppq(
t
) = 0. As
t
is arbitrary, supp q(
2
) a = 0; so
suppq() \ a suppq(
2
) ,= 0. QQQ
(j) In the same way, we can dene

: B A by setting

b = supsuppq
1
() : H is an involution and supp b
for every b B. Now

a = a for every a A. PPP () If 0 ,= u a, there is an involution G supported

by u. Now q() is an involution in H supported by a, so
u

a u suppq
1
q() = supp ,= 0.
As u is arbitrary, a

a. () If H is an involution supported by a, then = q

1
() is an involution
in G with suppq() = supp a, so supp a, by (i) above; as is arbitrary,

a a. QQQ
Similarly,

b = b for every b B. But this means that and

are the two halves of an order-

isomorphism between A and B. By 312L, both are Boolean homomorphisms.
(k) If G is an involution, then (supp) = suppq(). PPP By the denition of , suppq() (supp).
On the other hand,
suppq() =

(suppq()) (suppq
1
q()) = (supp). QQQ
Similarly, if H is an involution,
1
(supp ) =

(supp ) = suppq
1
().
(l) We are nearly home. Let us conrm that q() =
1
for every G. PPP??? Otherwise, =
q()
1

1
is not the identity automorphism on B, and there is a non-zero b B such that b b = 0,
that is,
1
b q()b = 0. Let H be an involution supported by b. Then q
1
() is supported by
1
b,
by (j), so
1
b supports q
1
()
1
and
1
b supports q(q
1
()
1
) = q()q()
1
. On the other
hand, q()b also supports q()q()
1
, which is not the identity automorphism; so these two elements of B
cannot be disjoint. XXXQQQ
(m) Finally, is unique by 383C.
Remark The ideas of the proof here are taken from Eigen 82.
383E The rest of this section may be regarded as a series of corollaries of this theorem. But I think it
will be apparent that they are very substantial results.
TheoremLet A and Bbe atomless homogeneous Boolean algebras, and q : Aut A Aut Ban isomorphism.
Then there is a unique Boolean isomorphism : A B such that q() =
1
for every Aut A.
proof (a) Let

A be the Dedekind completion of A (314U). Then every Aut A has a unique extension to
a Boolean homomorphism

:

A

A (314Tb). Because the extension is unique, we must have ()=

for all , Aut A; consequently,

and

1
are inverses of each other, and

Aut

A for each Aut A;

moreover,

is a group homomorphism. Of course it is injective, so we have a subgroup G =

:
Aut A of Aut

A which is isomorphic to Aut A. Clearly

G = : Aut

A, u A for every u A.
496 Automorphism groups 383E
If a

A is non-zero, then there is a non-zero u a belonging to A. Because A is atomless and homogeneous,
there is an involution Aut A supported by u (381Q); now G is an involution supported by a. As a
is arbitrary, G has many involutions.
Similarly, writing

B for the Dedekind completion of B, we have a subgroup H =

: Aut B of
Aut

B isomorphic to Aut B, and with many involutions. Let q : G H be the corresponding isomorphism,
so that q(

) =

q() for every Aut A.
By 383D, there is a Boolean isomorphism

:

A

B such that q() =

1
for every G.
(b) If u A, then

u B. PPP It is enough to consider the case u / 0, 1, since surely

0 = 0,

1 = 1.
Take any w B which is neither 0 nor 1; then there is an involution in Aut B with support w (381Q
again); the corresponding member of H is still an involution with support w. Its image q
1
() in G is
an involution with support a =

1
w

A; of course 0 ,= a ,= 1. Take non-zero u
1
, u
3
A such that u
1
a
and u
3
1 \ a; set u
2
= 1 \ (u
1
u
3
). Because A is homogeneous, there are , G such that u
1
= u,
u
1
= u
1
, u
2
= u
3
; set
2
= . Then we have
u = u
1
(supp q
1
()) = supp( q
1
()
1
) (u
1
u
2
) = u u
2
,
u =
2
u
1

2
(supp q
1
()) = supp(
2
q
1
()
1
2
) u
2
u
2
= u u
3
,
so
(supp q
1
())
2
(supp q
1
()) = u,
and

u =

((supp q
1
()))

(
2
(supp q
1
()))
=

(supp q
1
()
1
)

(supp
2
q
1
()
1
2
)
=

(supp q
1
( q() q()
1
))

(supp q
1
( q(
2
) q(
2
)
1
))
=

1
(supp( q() q()
1
))

1
(supp( q(
2
) q(
2
)
1
))
= supp( q() q()
1
) supp( q(
2
) q(
2
)
1
)
= q()(supp) q(
2
)(supp) = q()w q(
2
)w B
because both q() and q(
2
) belong to H. QQQ
Similarly,

1
v A for every v B, and =

A is an isomorphism between A and B.
We now have
q() = q(

)B = (

1
)B =
1
for every Aut A. Finally, is unique by 383C, as before.
383F Corollary If A and Bare atomless homogeneous Boolean algebras with isomorphic automorphism
groups, they are isomorphic as Boolean algebras.
Remark Of course a one-element Boolean algebra 0 and a two-element Boolean algebra 0, 1 have
isomorphic automorphism groups without being isomorphic.
383G Corollary If A is a homogeneous Boolean algebra, then Aut A has no outer automorphisms.
proof If A = 0, 1 this is trivial. Otherwise, A is atomless, so if q is any automorphism of Aut A, there
is a Boolean isomorphism : A A such that q() =
1
for every Aut A, and q is an inner
automorphism.
383H Denitions Complementary to the notion of many involutions is the following concept.
(a) A Boolean algebra A is rigid if the only automorphism of A is the identity automorphism.
(b) A Boolean algebra A is nowhere rigid if no non-trivial principal ideal of A is rigid.
383N Outer automorphisms 497
383I Lemma Let A be a Boolean algebra. Then the following are equiveridical:
(i) A is nowhere rigid;
(ii) for every a A 0 there is a Aut A, not the identity, supported by a;
(iii) for every a A 0 there are distinct non-zero b, c a such that the principal ideals A
b
, A
c
they
generate are isomorphic;
(iv) the automorphism group Aut A has many involutions.
proof (a)(ii)(i) If a A 0, let Aut A be a non-trivial automophism supported by a; then A
a
is a non-trivial automorphism of the principal ideal A
a
, so A
a
is not rigid.
(b)(i)(iii) There is a non-trivial automorphism of A
a
; now if b A
a
is such that b = c ,= b, A
b
is
isomorphic to [A
b
] = A
c
.
(c)(iii)(iv) Take any non-zero a A. By (iii), there are distinct b, c a such that A
b
, A
c
are
isomorphic. At least one of b \ c, c \ b is non-zero; suppose the former. Let : A
b
A
c
be an isomorphism,
and set d = b \ c, d
t
= (b \ c); then d
t
c, so d
t
d = 0, and = (

d
t
) is an involution supported by a.
(d)(iv)(ii) is trivial.
383J Theorem Let A and B be nowhere rigid Dedekind complete Boolean algebras and q : Aut A
Aut B an isomorphism. Then there is a Boolean isomorphism : A B such that q() =
1
for every
Aut A.
proof Put 383I(i)(iv) and 383D together.
383K Corollary Let A be a nowhere rigid Dedekind complete Boolean algebra. Then Aut A has no
outer automorphisms.
383L Examples I note the following examples of nowhere rigid algebras.
(a) An atomless homogeneous Boolean algebra is nowhere rigid.
(b) Any principal ideal of a nowhere rigid Boolean algebra is nowhere rigid.
(c) A simple product of nowhere rigid Boolean algebras is nowhere rigid.
(d) Any atomless semi-nite measure algebra is nowhere rigid.
(e) A free product of nowhere rigid Boolean algebras is nowhere rigid.
(f) The Dedekind completion of a nowhere rigid Boolean algebra is nowhere rigid.
Indeed, the diculty is to nd an atomless Boolean algebra which is not nowhere rigid; for a variety of
constructions of rigid algebras, see Bekkali & Bonnet 89.
383M Theorem Let (A, ) and (B, ) be atomless localizable measure algebras, and Aut

A, Aut

Bthe
corresponding groups of measure-preserving automorphisms. Let q : Aut

A Aut

B be an isomorphism.
Then there is a Boolean isomorphism : A B such that q() =
1
for every Aut

A.
proof The point is just that Aut

A has many involutions. PPP Let a A 0. Then there is a non-
zero b a such that the principal ideal A
b
is Maharam-type-homogeneous. Take c b, d b \ c such that
c = d = min(1,
1
2
b) (331C). The principal ideals A
c
, A
d
are now isomorphic as measure algebras (331I);
let : A
c
A
d
be a measure-preserving isomorphism. Then (

d) Aut

A is an involution supported by
a. QQQ
Similarly, Aut

B has many involutions, and the result follows at once from 383D.
383N To make proper use of the last theorem we need the following result.
Proposition Let (A, ) and (B, ) be localizable measure algebras and : A B a Boolean isomorphism.
For each innite cardinal let e

be the Maharam-type- component of A (332G) and for each ]0, [

let A

be the set of atoms of A of measure . Then the following are equiveridical:

(i) for every Aut

A,
1
Aut

B;
498 Automorphism groups 383N
(ii)() for every innite cardinal there is an

> 0 such that (a) =

a for every a e

() for
every ]0, [ there is an

> 0 such that (a) =

a for every a A

.
proof (a)(i)(ii)() Let be an innite cardinal. The point is that if a, a
t
e

and a = a
t
<
then (a) = (a
t
). PPP The principal ideals A
a
, A
a
are isomorphic as measure algebras; moreover, by
332J, the principal ideals A
e

\a
, A
e

\a
are isomorphic. We therefore have a Aut

A such that a = a
t
.
Consequently a = a
t
, where =
1
Aut

B, and (a) = (a
t
). QQQ
If e

= 0 we can take

= 1. Otherwise x on some c
0
e

such that 0 < c

0
< ; take b c
0
such
that 0 < b < , and set c =
1
b,

= b/ c. Then we shall have (a) = (c) =

a whenever
a e

and a = c. But we can nd for any n 1 a partition c

n1
, . . . , c
nn
of c into elements of measure
1
n
c; since (c
ni
) = (c
nj
) for all i, j n, we must have (c
ni
) =
1
n
(c) =

c
ni
for all i. So if a e

and a =
1
n
c, (a) = (c
n1
) =

a. Now suppose that a e

and a =
k
n
c for some k, n 1; then a
can be partitioned into k sets of measure
1
n
c, so in this case also (a) =

a. Finally, for any a e

, set
D = d : d a, d is a rational multiple of c,
and let D
t
D be a maximal upwards-directed set. Then supD
t
= a, so [D
t
] is an upwards-directed set
with supremum a, and
(a) = sup
dD
(d) = sup
dD

d =

a.
() Let ]0, [. If A

= take

= 1. Otherwise, x on any c A

and set

= (c)/. If
a A

then there is a Aut

A exchanging the atoms a, c, so that
1
Aut

B exchanges the atoms
a, c, and
(a) = (c) =

a.
(b)(ii)(i) Now suppose that the conditions () and () are satised, that Aut

A and that a A.
For each innite cardinal , we have e

= e

, so
((e

a)) =

((e

a)) =

(e

a) = ((e

a)).
Similarly, if we write a

= supA

, then for each ]0, [ we have [A

] = A

, a

= a

, and for c a

we have
c = #(e : e A

, e c);
so
((a

a)) =

#(e : e A

, e a)
=

#(e : e A

, e a)
=

eA

,ea
(e) = ((a

a)).
Putting these together,
(a) =

is an innite cardinal
((e

a)) +

]0,[
((a

a))
=

is an innite cardinal
((e

a)) +

]0,[
((a

a))
= (a).
But this means that
(
1
b) = (
1
b) = b
for every b B, and
1
is measure-preserving, as required by (i).
383P Outer automorphisms 499
383O Corollary If (A, ) is an atomless totally nite measure algebra, Aut

A has no outer automor-
phisms.
proof Let q : Aut

A Aut

A be any automorphism. By 383M, there is a corresponding Aut A such
that q() =
1
for every Aut

A. By 383N, there is for each innite cardinal an

> 0 such that

(a) =

a whenever a e

, the Maharam-type- component of A. But since e

= e

and e

<
for every , we must have

= 1 whenever e

,= 0; as A is atomless,
(a) =

is an innite cardinal
((a e

))
=

is an innite cardinal

(a e

)
=

is an innite cardinal
(a e

) = a
for every a A. Thus Aut

A and q is an inner automorphism.
383P The results above are satisfying and complete in their own terms, but leave open a number of
obvious questions concerning whether some of the hypotheses can be relaxed. Atoms can produce a variety
of complications (see 383Ya-383Yd below). To show that we really do need to assume that our algebras are
Dedekind complete or localizable, I oer the following.
Example (a) There are an atomless localizable measure algebra (A, ) and an atomless semi-nite measure
algebra (B, ) such that Aut A

= Aut B, Aut

A

= Aut

B but A and B are not isomorphic.
proof Let (A
0
,
0
) be an atomless homogeneous probability algebra; for instance, the measure algebra
of Lebesgue measure on the unit interval. Let A be the simple product Boolean algebra A

1
0
, and the
corresponding measure (322K); then (A, ) is an atomless localizable measure algebra. In A let I be the set
a : a A and the principal ideal A
a
is ccc;
then I is an ideal of A, the -ideal generated by the elements of nite measure (cf. 322G). Set
B = a : a A, either a I or 1 \ a I.
Then B is a -subalgebra of A, so if we set = B then (B, ) is a measure algebra in its own right.
The denition of I makes it plain that it is invariant under all Boolean automorphisms of A; so B is also
invariant under all automorphisms, and we have a homomorphism q() = B : Aut A Aut B. On
the other hand, because B is order-dense in A, and A is Dedekind complete, every automorphism of B can
be extended to an automorphism of A (see part (a) of the proof of 383E). So q is actually an isomorphism
between Aut A and Aut B. Moreover, still because B is order-dense, q() is measure-preserving i is
measure-preserving, so Aut

A is isomorphic to Aut

B. But of course there is no Boolean isomorphism, let
alone a measure algebra isomorphism, between A and B, because A is Dedekind complete while B is not.
Remark Thus the hypothesis Dedekind complete in 383D and 383J (and localizable in 383M), and the
hypothesis homogeneous in 383E-383F, are essential.
(b) There is an atomless semi-nite measure algebra (C,

) such that Aut C has an outer automorphism.

proof In fact we can take C to be the simple product of A and B above. I claim that the isomorphism
between Aut A and Aut B gives rise to an outer automorphism of Aut C; this seems very natural, but I think
there is a fair bit to check, so I take the argument in easy stages.
(i) We may identify the Dedekind completion of C = A B with A A. For Aut C, we have a
corresponding

Aut(A A). Now B A is invariant under

. PPP Consider rst (0, 1) = (a
1
, b
1
) C.
The corresponding principal ideal C
(a
1
,b
1
)

= A
a
1
B
b
1
of C must be isomorphic to the principal ideal
C
(0,1)

= B; so that if (a, b) C and (a, b) (a
1
, b
1
), then just one of the principal ideals C
(a,b)

= A
a
B
b
,
C
(a
1
\a,b
1
\b)

= A
a
1
\a
B
b
1
\b
is ccc. But this can only happen if A
a
1
is ccc and B
b
1
is not; that is, if a
1
and
1 \ b
1
belong to I. Consequently

(0, a) (a
1
, b
1
) belongs to BA for every a A. We also nd that
(1, 0) = (1, 1) \ (0, 1) = (1 \ a
1
, 1 \ b
1
) BA.
500 Automorphism groups 383P
Now if b I, then
C
(b,0)

= C
(b,0)

= A
b
is ccc and
(b, 0) I I BA;
while
(1 \ b, 0) = (1 \ a
1
, 1 \ b
1
) \ (b, 0) BA.
This shows that (b, 0) BA for every b B. So

(b, a) =

(b, 0)

(0, a) BA
for every b B, a A. QQQ
(ii) Let : A A A A be the involution dened by setting (a, b) = (b, a) for all a, b A. Take
Aut C and consider =

1
Aut(A A). If c = (a, b) C, then
1
c = (b, a) B A, so

1
c BA, by (i), and c AB = C. This shows that C is a homomorphism from C to itself. Of
course
1
=

1
has the same property. So we have a map q : Aut C Aut C given by setting
q() =

1
C
for Aut C. Evidently q is an automorphism.
(iii) ??? Suppose, if possible, that q were an inner automorphism. Let Aut C be such that q() =

1
for every Aut C. Then

1
=

q() =

1
for every Aut C. Since G =

: Aut C is a subgroup of Aut(A A) with many involutions, the

uniqueness assertion of 383D tells us that = . But
[C] = BA ,= C = [C] = [C],
so this cannot be. XXX
Thus q is the required outer automorphism of Aut C.
Remark Thus the hypothesis homogeneous in 383E, and the hypothesis Dedekind complete in 383J, are
necessary.
383Q Example Let be Lebesgue measure on R, and (A, ) its measure algebra. Then Aut

A has an
outer automorphism. PPP Set f(x) = 2x for x R. Then E f
1
[E] =
1
2
E is a Boolean automorphism of
the domain of , and (
1
2
E) =
1
2
E for every E (263A, or otherwise). So we have a corresponding
Aut A dened by setting E

= (
1
2
E)

for every E , and (a) =

1
2
a for every a A. By
383N, we have an automorphism q of Aut

A dened by setting q() =
1
for every measure-preserving
automorphism . But q is now an outer automorphism of Aut

A, because (by 383D) the only possible
automorphism of A corresponding to q is , and is not measure-preserving. QQQ
Thus the hypothesis totally nite in 383O cannot be omitted.
383X Basic exercises (a) Let A be a Boolean algebra. Show that the following are equiveridical: (i)
A is nowhere rigid; (ii) for every a A 0 and n N there are disjoint non-zero b
0
, . . . , b
n
a such that
the principal ideals A
b
i
they generate are all isomorphic; (iii) for every a A 0 and n 1 there is a
Aut A, of order n, supported by a.
(b) Let A be a nowhere rigid Boolean algebra. Show that its Dedekind completion is nowhere rigid.
(c) Let A be an atomless homogeneous Boolean algebra and B a nowhere rigid Boolean algebra, and
suppose that Aut A is isomorphic to Aut B. Show that there is an invariant order-dense subalgebra of B
which is isomorphic to A.
383 Notes Outer automorphisms 501
(d) Let A and B be nowhere rigid Boolean algebras. Show that if Aut A and Aut B are isomorphic, then
the Dedekind completions

A and

B are isomorphic.
(e) Find two non-isomorphic atomless totally nite measure algebras (A, ), (B, ) such that Aut

A and
Aut

B are isomorphic. (This is easy.)
(f ) Let (A, ) and (B, ) be semi-nite measure algebras and : A B a Boolean isomorphism. Show
that the following are equiveridical: (i) for every Aut

A,
1
Aut

B; (ii)() for every innite
cardinal there is an

> 0 such that (a) =

a whenever a A and the principal ideal A

a
is
Maharam-type-homogeneous with Maharam type ; () for every ]0, [ there is an

> 0 such that

(a) =

a whenever a A is an atom of measure .

(g) Let q : Aut C Aut C be the automorphism of 383Pb. Show that q() is measure-preserving whenever
is measure-preserving, so that q Aut

C is an outer automorphism of Aut

C.
383Y Further exercises (a) Let (A, ) and (B, ) be localizable measure algebras such that Aut A

=
Aut B. Show that either A

= B or one of A, B has just one atom and the other is atomless.

(b) Let (A, ), (B, ) be localizable measure algebras such that Aut

A

= Aut

B. Show that either
(A, )

= (B, ) or there is some ]0, [ such that one of A, B has just one atom of measure and the
other has none or there are ,
t
]0, [ such that the number of atoms of A of measure is equal to the
number of atoms of B of measure
t
, but not to the number of atoms of A of measure
t
.
(c) Let (A, ) be a localizable measure algebra. Show that there is an outer automorphism of Aut A i A
has exactly six atoms.
(d) Let (A, ) be a localizable measure algebra. For each innite cardinal let e

be the Maharam-type-
component of A and for each ]0, [ let A

be the set of atoms of A of measure . Show that there is an

outer automorphism of Aut

A i
either there is an innite cardinal such that e

=
or there are distinct , ]0, [ such that #(A

) = #(A

) 2
or there is a ]0, [ such that #(A

) = 6
or there are , ]0, [ such that #(A

) = 2 < #(A

) < .
383 Notes and comments Let me recapitulate the results above. If A and B are Boolean algebras,
any isomorphism between Aut A and Aut B corresponds to an isomorphism between A and B if either A
and B are atomless and homogeneous (383E) or they are nowhere rigid and Dedekind complete (383J). If
(A, ) and (B, ) are atomless localizable measure algebras, then any automorphism between Aut

A and
Aut

B corresponds to an isomorphism between A and B (383M) which if = is totally nite will be
measure-preserving (383O).
These results may appear a little less surprising if I remark that the elementary Boolean algebras TX
give rise to some of the same phenomena. The automorphism group of TX can be identied with the
group S
X
of all permutations of X, and this has no outer automorphisms unless X has just six elements.
Some of the ideas of the fundamental theorem 383D can be traced through in the purely atomic case also,
though of course there are signicant changes to be made, and some serious complications arise, of which the
most striking surround the remarkable fact that S
6
does have an outer automorphism (Burnside 11, 162;
Rotman 84, Theorem 7.8). I have not attempted to incorporate these in the main results. For localizable
measure algebras, where the only rigid parts are atoms, the complications are superable, and I think I have
listed them all (383Ya-383Yd).
502 Automorphism groups 384 intro.
384 Entropy
Perhaps the most glaring problem associated with the theory of measure-preserving homomorphisms and
automorphisms is the fact that we have no generally eective method of determining when two homomor-
phisms are the same, in the sense that two structures (A, , ) and (B, , ) are isomorphic, where (A, )
and (B, ) are measure algebras and : A A, : B B are Boolean homomorphisms. Of course the
rst part of the problem is to decide whether (A, ) and (B, ) are isomorphic; but this is solved (at least
for localizable algebras) by Maharams theorem (see 332J). The diculty lies in the homomorphisms. Even
when we know that (A, ) and (B, ) are both isomorphic to the Lebesgue measure algebra, the extraordi-
nary variety of constructions of homomorphisms corresponding in part to the variety of measure spaces
with such measure algebras, each with its own natural inverse-measure-preserving functions means that
the question of which are isomorphic to each other is continually being raised. In this section I give the most
elementary ideas associated with the concept of entropy, up to the Kolmogorov-Sina theorem (384P). This
is an invariant which can be attached to any measure-preserving homomorphism on a probability algebra,
and therefore provides a useful method for distinguishing non-isomorphic homomorphisms.
The main work of the section deals with homomorphisms on measure algebras, but as many of the most
important ones arise from inverse-measure-preserving functions on measure spaces I comment on the extra
problems arising in the isomorphism problem for such functions (384T-384V). I should remark that some of
the lemmas will be repeated in stronger forms in the next section.
384A Notation Throughout this section, I will use the letter q to denote the function from [0, [ to R
dened by saying that q(t) = t lnt = t ln
1
t
if t > 0, q(0) = 0.
0 0.5 1 1.5
-0.5
The function q
We shall need the following straightforward facts concerning q.
(a) q is continuous on [0, [ and dierentiable on ]0, [; q
t
(t) = 1 lnt and q
tt
(t) =
1
t
for t > 0.
Because q
tt
is negative, q is concave, that is, q is convex. q has a unique maximum at (
1
e
,
1
e
).
(b) If s 0, t > 0 then q
t
(s +t) q
t
(t); consequently
q(s +t) = q(s) +
_
t
0
q
t
(s +)d q(s) +q(t)
for s, t 0. It follows that q(

n
i=0
s
i
)

n
i=0
q(s
i
) for all s
0
, . . . , s
n
0 and (because q is continuous)
q(

i=0
s
i
)

i=0
q(s
i
) for every non-negative summable series s
i
)
iN
.
(c) If s, t 0 then q(st) = sq(t) +tq(s); more generally, if n 1 and s
i
0 for i n then
q(

n
i=0
s
i
) =

n
j=0
q(s
j
)

i,=j
s
i
.
(d) The function t q(t) +q(1 t) has a unique maximum at (
1
2
, ln2). (
d
dt
(q(t) +q(1 t)) = ln
1t
t
.) It
follows that for every > 0 there is a > 0 such that [t
1
2
[ whenever q(t) +q(1 t) ln2 .
(e) If 0 t
1
2
, then q(1 t) q(t). PPP Set f(t) = q(t) q(1 t). Then
f
tt
(t) =
1
t
+
1
1t
=
2t1
t(1t)
0
for 0 < t
1
2
, while f(0) = f(
1
2
) = 0, so f(t) 0 for 0 t
1
2
. QQQ
384G Entropy 503
(f ) (i) If A is a Dedekind -complete Boolean algebra, I will write q for the function from L
0
(A)
+
to
L
0
(A) dened from q (364I). Note that because 0 q(t) 1 for t [0, 1], 0 q(u) 1 if 0 u 1.
(ii) By (b), q(u + v) q(u) + q(v) for all u, v 0 in L
0
(A). (Represent A as the measure algebra of a
measure space, so that q(f

) = (qf)

, as in 364Jb.)
(iii) Similarly, if u, v L
0
(A)
+
, then q(u v) = u q(v) +v q(u).
384B Lemma Let (A, ) be a probability algebra, B a closed subalgebra of A, and P : L
1
(A, )
L
1
(A, ) the corresponding conditional expectation operator (365R). Then
_
q(u) q(
_
u) and P( q(u))
q(Pu) for every u L

(A)
+
.
proof Apply the remarks in 365Rb to q. ( q(u) L

L
1
for every u (L

)
+
because q is bounded on
every bounded interval in [0, [.)
384C Denition Let (A, ) be a probability algebra. If A is a partition of unity in A, its entropy is
H(A) =

aA
q( a), where q is the function dened in 384A.
Remarks (a) In the denition of partition of unity (311Gc) I allowed 0 to belong to the family. In the
present context this is a mild irritant, and when convenient I shall remove 0 from the partitions of unity
considered here (as in 384F below). But because q(0) = 0, it makes no dierence; H(A) = H(A 0)
whenever A is a partition of unity. So if you wish you can read partition of unity in this section to mean
partition of unity not containing 0, if you are willing to make an occasional amendment in a formula. In
important cases, in fact, A is of the form a
i
: i I or a
i
: i I 0, where a
i
)
iI
is an indexed
partition of unity, with a
i
a
j
= 0 for i ,= j, but no restriction in the number of i with a
i
= 0; in this case,
we still have H(A) =

iI
q( a
i
).
(b) Many authors prefer to use log
2
in place of ln. This makes sense in terms of one of the intuitive
approaches to entropy as the information associated with a partition. See Petersen 83, 5.1.
384D Denition Let (A, ) be a probability algebra, B a closed subalgebra of A and A a partition of
unity in A. Let P : L
1
(A, ) L
1
(A, ) be the conditional expectation operator associated with B. Then
the conditional entropy of A on B is
H(A[B) =

aA
_
q(Pa),
where q is dened as in 384A.
384E Elementary remarks (a) In the formula

aA
_
q(Pa),
we have 0 P(a) 1 for every a, so q(Pa) 0 and every term in the sum is non-negative; accordingly
H(A[B) is well-dened in [0, ].
(b) H(A) = H(A[0, 1), since if B = 0, 1 then P(a) = a1, so that
_
q(Pa) = q( a). If A B,
H(A[B) = 0, since P(a) = a, q(Pa) = 0 for every a.
384F Denition If A is a Boolean algebra and A, B A are partitions of unity, I write A B for the
partition of unity a b : a A, b B 0. (See 384Xq.)
384G Lemma Let (A, ) be a probability algebra and B a closed subalgebra. Let A A be a partition
of unity.
(a) If B is another partition of unity in A, then
H(A[B) H(A B[B) H(A[B) +H(B[B).
(b) If B is purely atomic and D is the set of its atoms, then H(A D) = H(D) +H(A[B).
(c) If C B is another closed subalgebra of A, then H(A[C) H(A[B). In particular, H(A) H(A[B).
(d) Suppose that B
n
)
nN
is a non-decreasing sequence of closed subalgebras of A such that B =

nN
B
n
. If H(A) < then
504 Automorphism groups 384G
H(A[B) = lim
n
H(A[B
n
).
In particular, if A B then lim
n
H(A[B
n
) = 0.
proof Write P for the conditional expectation operator on L
1
(A, ) associated with B.
(a)(i) If B is innite, enumerate it as b
j
)
jN
; if it is nite, enumerate it as b
j
)
jn
and set b
j
= 0 for
j > n. For any a A,
a =

j=0
(a b
j
), P(a) =

j=0
P(a b
j
),
q(Pa) = lim
n
q(
n

j=0
P(a b
j
))
lim
n
n

j=0
q(P(a b
j
)) =

j=0
q(P(a b
j
))
where all the innite sums are to be regarded as order*-limits of the corresponding nite sums (see 367),
and the middle inequality is a consequence of 384Af(ii). Accordingly
H(A B[B) =

aA,bB,ab,=0
_
q(P(a b))
=

aA

j=0
_
q(P(a b
j
))

aA
_
q(Pa
i
) = H(A[B).
(ii) Suppose for the moment that A and B are both nite. For a A set u
a
= P(a). If a, b A we
have 0 u
ab
u
b
in L
0
(B), so we may choose v
ab
L
0
(B) such that 0 v
ab
1 and u
ab
= v
ab
u
b
.
For any b B,

aA
u
ab
= u
b
(because

aA
(a b) = b), so u
b


aA
v
ab
= u
b
. Since
[[[ q(u
b
)[ > 0]] [[u
b
> 0]], q(u
b
)

aA
v
ab
= q(u
b
).
For any a A,
q(u
a
) = q(

bB
u
ab
) = q(

bB
u
b
v
ab
) = q(P(

bB
b v
ab
))
(because v
ab
L
0
(B) for every b, so P(b v
ab
) = P(b) v
ab
)
P( q(

bB
b v
ab
))
(384B)
= P(

bB
b q(v
ab
))
(because B is disjoint)
=

bB
u
b
q(v
ab
)
(because q(v
ab
) L
0
(B) for every b).
Putting these together,
H(A B[B) =

aA,bB
_
q(u
ab
) =

aA,bB
_
q(u
b
v
ab
)
=

aA,bB
_
u
b
q(v
ab
) +

aA,bB
_
v
ab
q(u
b
)

aA
_
q(u
a
) +

bB
_
q(u
b
) = H(A[B) +H(B[B).
384G Entropy 505
(iii) For general partitions of unity A and B, take any nite set C A B. Then C a b : a
A
0
, b B
0
where A
0
A and B
0
B are nite. Set
A
t
= A
0
1 \ supA
0
, B
t
= B
0
1 \ supB
0
,
so that A
t
and B
t
are nite partitions of unity and C A
t
B
t
. Now

cC
_
q(Pc)

cA

_
q(Pc) = H(A
t
B
t
[B) H(A
t
[B) +H(B
t
[B)
(by (ii))
H(A
t
A[B) +H(B
t
B[B)
(by (i))
= H(A[B) +H(B[B).
As C is arbitrary,
H(A B[B) =

cAB
_
q(Pc) H(A[B) +H(B[B).
(b) It follows from 384Ab that

dD
q( (a d)) q( a) for any a A.
Now, because B is purely atomic and D is its set of atoms,
P(a) =

dD
(ad)
d
d, q(P(a)) =

dD
q(
(ad)
d
)d
for every a A,
H(A[B) =

aA,dD
q(
(ad)
d
) d.
Putting these together,
H(A D) =

aA,dD
q( (a d)) =

aA,dD
q(
(ad)
d
) d +
(ad)
d
q( d)
= H(A[B) +

dD
q( d) = H(A[B) +H(D).
(c) Write P
C
for the conditional expectation operator corresponding to C. If a A,
q(P
C
a) = q(P
C
Pa) P
C
q(Pa)
by 384B. So
H(A[C) =

aA
_
q(P
C
a)

aA
_
P
C
q(Pa) =

aA
_
q(Pa) = H(A[B).
Taking C = 0, 1, we get H(A) H(A[B).
(d) Let P
n
be the conditional expectation operator corresponding to B
n
, for each n. Fix a A.
Then P(a) is the order*-limit of P
n
(a))
nN
, by Levys martingale theorem (367Kb). Consequently
(because q is continuous) q(P
n
a))
nN
is order*-convergent to q(Pa) for every a A (367I). Also, because
0 P
n
a 1 for every n, 0 q(P
n
a)
1
e
1 for every n. By the Dominated Convergence Theorem
(367J), lim
n
_
q(P
n
a) =
_
q(Pa).
By 384B, we also have
0
_
q(P
n
a) q(
_
P
n
(a)) = q(
_
a) = q( a)
for every a A and n N; since also
0
_
q(Pa) = q( a),
we have [
_
q(P
n
a)
_
q(Pa)[ q( a) for every a A, n N.
506 Automorphism groups 384G
Now we are supposing that H(A) is nite. Given > 0, we can nd a nite set I A such that

aA\I
q( a) , and an n
0
N such that

aI
[
_
q(P
n
a)
_
q(Pa)[
for every n n
0
; in which case

aA\I
[
_
q(P
n
a)
_
q(Pa)[

aA\I
q( a)
and [H(A[B
n
) H(A[B)[ 2 for every n n
0
. As is arbitrary, H(A[B) = lim
n
H(A[B
n
).
384H Corollary Let (A, ) be a probability algebra and A, B two partitions of unity in A. Then
H(A) H(A B) H(A) +H(B).
proof Take B = 0, 1 in 384Ga.
384I Lemma Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean
homomorphism. If A A is a partition of unity, then H([A]) = H(A).
proof

aA
q( a) =

aA
q( a).
384J Lemma Let (A, ) be a measure algebra. Let A be the set of its atoms. Then the following are
equiveridical:
(i) either A is not purely atomic or A is purely atomic and H(A) = ;
(ii) there is a partition of unity B A such that H(B) = ;
(iii) for every R there is a nite partition of unity C A such that H(C) .
proof (i)(ii) We need examine only the case in which A is not purely atomic. Let a A be a non-zero
element such that the principal ideal A
a
is atomless. By 331C we can choose inductively a disjoint sequence
a
n
)
nN
such that a
n
a and a
n
= 2
n1
a. Now, for each n N, choose a disjoint set B
n
such that
#(B
n
) = 2
2
n
, b a
n
and b = 2
2
n
a
n
for each b B
n
.
Set
B =

nN
B
n
1 \ a.
Then B is a partition of unity in A and
H(B)

n=0

bB
n
q( B
n
) =

n=0
2
2
n
q
_
a
2
n+1+2
n
_
=

n=0
a
2
n+1
ln
_
2
n+1+2
n
a
_

n=0
a
2
n+1
2
n
ln2 = .
(ii)(iii) Enumerate B as b
i
)
iN
. For each n N, C
n
= b
i
: i n1 \ sup
in
b
i
is a nite partition
of unity, and
lim
n
H(C
n
) lim
n

n
i=0
q( b
i
) = H(B) = .
(iii)(i) We need only consider the case in which A is purely atomic and A is its set of atoms. In this
case, A C = A for every partition of unity C A, so H(C) H(A) for every C (384H), and H(A) must
be innite.
384K Denition Let A be a Boolean algebra. If : A A is an order-continuous Boolean homo-
morphism, A A is a partition of unity and n 1, write D
n
(A, ) for the partition of unity generated by

i
a : a A, 0 i < n, that is, inf
i<n

i
a
i
: a
i
A for every i < n0. Observe that D
1
(A, ) = A0
and
D
n+1
(A, ) = D
n
(A, )
n
[A] = A [D
n
(A, )]
for every n 1.
384N Entropy 507
384L Lemma Let (A, ) be a probability algebra and : A A a measure-preserving Boolean
homomorphism. Let A A be a partition of unity. Then lim
n
1
n
H(D
n
(A, )) = inf
n1
1
n
H(D
n
(A, ))
is dened in [0, ].
proof (a) Set
0
= 0,
n
= H(D
n
(A, )) for n 1. Then
m+n

m
+
n
for all m, n 0. PPP If m,
n 1, D
m+n
(A, ) = D
m
(A, )
m
[D
n
(A, )]. So 384Ga tells us that
H(D
m+n
(A, )) H(D
m
(A, )) +H(
m
[D
n
(A, )]) = H(D
m
(A, )) +H(D
n
(A, ))
because is measure-preserving. QQQ
(b) If
1
= then of course H(D
n
(A, )) H(A) = for every n, by 384H, so inf
n1
1
n
H(D
n
(A, )) =
= lim
n
1
n
H(D
n
(A, )). Otherwise,
n
n
1
is nite for every n. Set = inf
n1
1
n

n
. If > 0 there
is an m 1 such that
1
m

m
+ . Set M = max
j<m

j
. Now, for any n m, there are k 1, j < m
such that n = km+j, so that

n
k
m
+
j
,
1
n

n

k
n

m
+
M
n

1
m

m
+
M
n
.
Accordingly limsup
n
1
n

n
+. As is arbitrary,
liminf
n
1
n

n
limsup
n
1
n

n

and lim
n
1
n

n
= is dened in [0, ].
Remark See also 384Yb and 385Nc below.
384M Denition Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean
homomorphism. For any partition of unity A A, set
h(, A) = inf
n1
1
n
H(D
n
(A, )) = lim
n
1
n
H(D
n
(A, ))
(384L). Now the entropy of is
h() = suph(, A) : A A is a nite partition of unity.
Remarks (a) We always have
h(, A) H(D
1
(A, )) =H(A).
(b) Observe that if is the identity automorphism then D
n
(A, ) = A 0 for every A and n, so that
h() = 0.
384N Lemma Let (A, ) be a probability algebra and A, B two partitions of unity in A. Let : A A
be a measure-preserving Boolean homomorphism. Then h(, A) h(, B) +H(A[B), where B is the closed
subalgebra of A generated by B.
proof We may suppose that 0 / B, since removing 0 from B changes neither D
n
(B, ) nor B. For each
n N, set A
n
=
n
[A], B
n
=
n
[B]. Let B
n
=
n
[B] be the closed subalgebra of A generated by B
n
,
and B

n
the closed subalgebra of A generated by D
n
(B, ). Then H(A
n
[B
n
) = H(A[B) for each n. PPP The
point is that, because B is purely atomic and B is its set of atoms,
H(A[B) =

aA,bB
q(
(ab)
b
) b
as in the proof of 384Gb. Similarly,
H(A
n
[B
n
) =

aA,bB
q(
(
n
a
n
b)
(
n
b)
) (
n
b) = H(A[B). QQQ
Accordingly, for any n 1,
508 Automorphism groups 384N
H(D
n
(A, )[B

n
)
n1

i=0
H(A
i
[B

n
)
(by 384Ga)

n1

i=0
H(A
i
[B
i
)
(by 384Gc)
= nH(A[B).
Now
h(, A) = lim
n
1
n
H(D
n
(A, )) limsup
n
1
n
H(D
n
(A, ) D
n
(B, ))
(384Ga)
limsup
n
1
n
H(D
n
(B, )) +
1
n
H(D
n
(A, )[B

n
)
(384Gb)
h(, B) +H(A[B).
Remark Compare 385Nd below.
384O Lemma Let (A, ) be a probability algebra, : A A a measure-preserving Boolean homomor-
phism, and A A a partition of unity such that H(A) < . Then h(, A) h().
proof If A is nite, this is immediate from the denition of h(); so suppose that A is innite. Enumerate
A as a
i
)
iN
. For each n N let B
n
be the subalgebra of A generated by a
0
, . . . , a
n
; then B
n
has atoms
B
n
= a
0
, . . . , a
n
, b
n
where b
n
= sup
i>n
a
i
. Now
H(A[B
n
) =

i=n+1
q(

n
)
n
where
i
= a
i
,
n
= b
n
, because if P
n
is the conditional expectation associated with B
n
then P
n
(a
i
) =
a
i
if i n, P
n
(a
i
) =

i

n
b
n
if i > n. But

i=n+1

n
q(

n
) =

i=n+1
q(
i
)

i

n
q(
n
)
=

i=n+1
q(
i
) q(
n
) 0 as n
because

i=0
q(
i
) < and lim
n

n
= 0. So by 384N and 384Gd we get
h(, A) h(, B
n
) +H(A[B
n
) h() +H(A[B
n
) h()
as n , and h(, A) h().
Remark Compare 385Nb below.
384P Theorem (Kolmogorov 58, Sina 59) Let (A, ) be a probability algebra, and : A A a
measure-preserving Boolean homomorphism.
(i) Suppose that A A is a partition of unity such that H(A) < and the closed subalgebra of A
generated by

nN

n
[A] is A itself. Then h() = h(, A).
(ii) Suppose that is an automorphism, and that A A is a partition of unity such that H(A) < and
the closed subalgebra of A generated by

nZ

n
[A] is A itself. Then h() = h(, A).
384R Entropy 509
proof I take the two arguments together. In both cases, by 384O, we have h(, A) h(), so I have to
show that if B A is any nite partition of unity, then h(, B) h(, A). For (i), let A
n
be the partition
of unity generated by

0j<n

j
[A]; for (ii), let A
n
be the partition of unity generated by

nj<n

j
[A].
Then h(, A
n
) = h(, A) for every n. PPP In case (i), we have D
m
(A
n
, ) = D
m+n
(A, ) for every m, so that
lim
m
1
m
H(D
m
(A
n
, )) = lim
m
1
m
H(D
m+n
(A, ))
= lim
m
1
m
H(D
m
(A, )).
In case (ii), we have D
m
(A
n
, ) =
n
[D
m+2n
(A, )] for every m, so that
lim
m
1
m
H(D
m
(A
n
, )) = lim
m
1
m
H(D
m+2n
(A, ))
= lim
m
1
m
H(D
m
(A, )). QQQ
Let A
n
be the purely atomic closed subalgebra of A generated by A
n
; our hypothesis is that the
closed subalgebra generated by

nN
A
n
is A itself, that is, that

nN
A
n
is dense. But this means that
lim
n
H(B[A
n
) = 0 (384Gd). Since
h(, B) h(, A
n
) +H(B[A
n
) = h(, A) +H(B[A
n
)
for every n (384N), we have the result.
384Q Bernoulli shifts Let (A, ) be a probability algebra, and : A A a measure-preserving
Boolean homomorphism.
(a) is a one-sided Bernoulli shift if there is a closed subalgebra A
0
in A such that (i)
k
[A
0
])
kN
is
independent (that is, (inf
jk

j
a
j
) =

k
j=0
a
j
for all a
0
, . . . , a
k
A
0
; see 325L) (ii) the closed subalgebra
of A generated by

kN

k
[A
0
] is A itself. In this case A
0
is a root algebra for .
(b) is a two-sided Bernoulli shift if it is an automorphism and there is a closed subalgebra A
0
in A
such that (i)
k
[A
0
])
kZ
is independent (ii) the closed subalgebra of A generated by

kZ

k
[A
0
] is A itself.
In this case A
0
is a root algebra for .
It is important to be aware that a Bernoulli shift can have many, and (in the case of a two-sided shift)
very dierent, root algebras; this is the subject of 386 below.
384R Theorem Let (A, ) be a probability algebra and : A B a Bernoulli shift, either one- or
two-sided, with root algebra A
0
.
(i) If A
0
is purely atomic, then h() = H(A), where A is the set of atoms of A
0
.
(ii) If A
0
is not purely atomic, then h() = .
proof (a) The point is that for any partition of unity C A
0
, h(, C) = H(C). PPP For any n 1, D
n
(C, )
is the partition of unity consisting of elements of the form inf
j<n

j
c
j
, where c
0
, . . . , c
n1
C. So
H(D
n
(C, )) =

c
0
,... ,c
n1
C
q( ( inf
j<n

j
c
j
)) =

c
0
,... ,c
n1
C
q(
n1

j=0
c
j
))
=

c
0
,... ,c
n1
C
n1

j=0
q( c
j
)

i,=j
c
i
(384Ac)
=
n1

j=0

cC
q( c) = nH(C).
510 Automorphism groups 384R
So
h(, C) = lim
n
1
n
H(D
n
(C, )) = H(C). QQQ
(b) If A
0
is purely atomic and H(A) < , the result can now be read o from 384P, because the closed
subalgebra of A generated by A is A
0
and the closed subalgebra of A generated by

kN

k
[A] or

kZ

k
[A]
is A; so h() = h(, A) = H(A).
(c) Otherwise, 384J tells us that there are nite partitions of unity C A
0
such that H(C) is arbitrarily
large. Since h() h(, C) = H(C) for any such C, by (a) and the denition of h(), h() must be innite,
as claimed.
384S Remarks (a) The standard construction of a Bernoulli shift is from a product space, as follows.
If (X, ,
0
) is any probability space, write for the product measure on X
N
; let (A, ) be the measure
algebra of , and A
0
A the set of equivalence classes of sets of the form x : x(0) E where E , so
that (A
0
, A
0
) can be identied with the measure algebra of
0
. We have an inverse-measure-preserving
function f : X
N
X
N
dened by setting
f(x)(n) = x(n + 1) for every x X
N
, n N,
and f induces, as usual, a measure-preserving homomorphism : A A. Now is a one-sided Bernoulli
shift with root algebra A
0
. PPP (i) If a
0
, . . . , a
k
A
0
, express each a
j
as x : x(0) E
j

, where E
j
.
Now

j
a
j
= (f
j
x : x(0) E
j
)

= x : x(j) E
j

for each j, so
(inf
jk

j
a
j
) = (

jk
x : x(j) E
j
) =

k
j=0

0
E
j
=

k
j=0
a
j
.
Thus
k
[A
0
])
kN
is independent. (ii) The closed subalgebra A
t
of A generated by

kN

k
[A
0
] must contain
x : x(k) E

for every k N, E , so must contain W

for every W in the -algebra generated by sets

of the form x : x(k) E; but every set measured by is equivalent to such a set W. So A
t
= A. QQQ
(b) The same method gives us two-sided Bernoulli shifts. Again let (X, ,
0
) be a probability space,
and this time write for the product measure on X
Z
; again let (A, ) be the measure algebra of , and
A
0
A the set of equivalence classes of sets of the form x : x(0) E where E , so that (A
0
, A
0
) can
once more be identied with the measure algebra of
0
. This time, we have a measure space automorphism
f : X
Z
X
Z
dened by setting
f(x)(n) = x(n + 1) for every x X
Z
, n Z,
and f induces a measure-preserving automorphism : A A. The arguments used above show that is a
two-sided Bernoulli shift with root algebra A
0
.
(c) I remarked above that a Bernoulli shift will normally have many root algebras. But it is important
to know that, up to isomorphism, any root algebra is associated with just one Bernoulli shift of each type.
PPP(i) Given a probability algebra (A
0
,
0
) then we can identify it with the measure algebra of a probability
space (X, ,
0
) (321J), and now the constructions of (a) and (b) provide Bernoulli shifts with root algebras
isomorphic to (A
0
,
0
).
(ii) Let (A, ) and (B, ) be probability algebras with one-sided Bernoulli shifts , with root algebras
A
0
, B
0
, and suppose that
0
: A
0
B
0
is a measure-preserving isomorphism. Then (A, ) can be identied
with the probability algebra free product of
k
[A
0
])
kN
(325L), while (B, ) can be identied with the
probability algebra free product of
k
[B
0
])
kN
. For each k N,
k

0
(
k
)
1
is a measure-preserving
isomorphism between
k
[A
0
] and
k
[B
0
]. Assembling these, we have a measure-preserving isomorphism
: A B such that a =
k

0
(
k
)
1
a whenever k N and a
k
[A
0
], that is,
k
a =
k

0
a for every
a A
0
, k N. Of course extends
0
.
If we set
384T Entropy 511
C = a : a A, a = a,
then C is a closed subalgebra of A. If a A
0
and k N, then
(
k
a) =
k+1
a =
k+1

0
a = (
k

0
a) = (
k
a),
so
k
a C. Thus
k
[A
0
] C for every k N, and C = A.
This means that : A B is such that =
1
; is an isomorphism between the structures (A, , )
and (B, , ) extending the isomorphism
0
from A
0
to B
0
.
(iii) Now suppose that (A, ) and (B, ) are probability algebras with two-sided Bernoulli shifts ,
with root algebras A
0
, B
0
, and suppose that
0
: A
0
B
0
is a measure-preserving isomorphism. Repeating
(ii) word for word, but changing each N into Z, we nd that
0
has an extension to a measure-preserving
isomorphism : A B such that = , so that once more the structures (A, , ) and (B, , ) are
isomorphic. QQQ
(d) The classic problem to which the theory of this section was directed was the following: suppose we
have two two-sided Bernoulli shifts and , one based on a root algebra with two atoms of measure
1
2
and
the other on a root algebra with three atoms of measure
1
3
; are they isomorphic? The Kolmogorov-Sina
theorem tells us that they are not, because h() = ln2 and h() = ln3 are dierent. The question of which
Bernoulli shifts are isomorphic is addressed, and (for countably-generated algebras) solved, in 386 below.
(e) We shall need to know that any Bernoulli shift (either one- or two-sided) is ergodic. In fact, it is
mixing. PPP Let (A, ) be a probability algebra and : A A a Bernoulli shift with root algebra A
0
. Let B
be the subalgebra of A generated by

kN

k
[A
0
] (if is one-sided) or by

kZ

k
[A
0
] (if is two-sided).
If b, c B, there is some n N such that both belong to the algebra B
n
generated by

jn

j
[A
0
] (if is
one-sided) or by

]j]n

j
[A
0
] (if is two-sided). If now k > 2n,
k
b belongs to the algebra generated by

j>n

j
[A
0
]. But this is independent of B
n
(cf. 325Xf, 272K), so
(c
k
b) = c (
k
b) = c b.
And this is true for every k n. Generally, if b, c A and > 0, there are b
t
, c
t
B such that (b .b
t
) ,
(c .c
t
) , so that
limsup
k
[ (c
k
b) c b[ limsup
k
[ (c
t

k
b
t
) c
t
b
t
[
+ (c .c
t
) + (
k
b .
k
b
t
) +[ c b c
t
b
t
[
0 + + +[ c c
t
[ +[ b b
t
[
4.
As , b, and c are arbitrary, is mixing. By 372Qa, it is ergodic. QQQ
(f ) The following elementary remark will be useful. If (A, ) is a probability algebra, : A A is a
measure-preserving automorphism, and A
0
A is a closed subalgebra such that
k
[A
0
])
kN
is independent,
then
k
[A
0
])
kZ
is independent. PPP If J Z is nite and a
j
)
jJ
is a family in A
0
, take n N such that
n j for every j J; then
(inf
jJ

j
a
j
) = (inf
jJ

n+j
a
j
) =

jJ
a
j
. QQQ
384T Isomorphic homomorphisms (a) In this section I have spoken of isomorphic homomorphisms
without oering a formal denition. I hope that my intention was indeed obvious, and that the next sentence
will merely conrm what you have already assumed. If (A
1
,
1
) and (A
2
,
2
) are measure algebras, and

1
: A
1
A
2
,
2
: A
2
A
2
are functions, then I say that (A
1
,
1
,
1
) and (A
2
,
2
,
2
) are isomorphic if
there is a measure-preserving isomorphism : A
1
A
2
such that
2
=
1

1
. In this context, using
Maharams theorem or otherwise, we can expect to be able to decide whether (A
1
,
1
) and (A
2
,
2
) are or
are not isomorphic; and if they are, we have a good hope of being able to describe a measure-preserving
isomorphism : A
1
A
2
. In this case, of course, (A
2
,
2
,
2
) will be isomorphic to (A
1
,
1
,
t
2
) where

t
2
=
1

2
. So now we have to decide whether (A
1
,
1
,
1
) is isomorphic to (A
1
,
1
,
t
2
); and when
1
,
2
512 Automorphism groups 384T
are measure-preserving Boolean automorphisms, this is just the question of whether
1
,
t
2
are conjugate in
the group Aut

1
(A
1
) of measure-preserving automorphisms of A
1
. Thus the isomorphism problem, as stated
here, is very close to the classical group-theoretic problem of identifying the conjugacy classes in Aut

(A)
for a measure algebra (A, ). But we also want to look at measure-preserving homomorphisms which are
not automorphisms, so there would be something left even if the conjugacy problem were solved. (In eect,
we are studying conjugacy in the semigroup of all measure-preserving Boolean homomorphisms, not just in
its group of invertible elements.)
The point of the calculation of the entropy of a homomorphism is that it is an invariant under this
kind of isomorphism; so that if
1
,
2
have dierent entropies then (A
1
,
1
,
1
) and (A
2
,
2
,
2
) cannot be
isomorphic. Of course the properties of being ergodic or mixing (see 372P) are also invariant.
(b) All the main work of this section has been done in terms of measure algebras; part of my purpose
in this volume has been to insist that this is often the right way to proceed, and to establish a language
which makes the arguments smooth and natural. But of course a large proportion of the most important
homomorphisms arise in the context of measure spaces, and I take a moment to discuss such applications.
Suppose that we have two quadruples (X
1
,
1
,
1
, f
1
), (X
2
,
2
,
2
, f
2
) where, for each i, (X
i
,
i
,
i
) is a
measure space and f
i
: X
i
X
i
is an inverse-measure-preserving function. Then we have associated
structures (A
1
,
1
,
1
), (A
2
,
2
,
2
) where (A
i
,
i
) is the measure algebra of (X
i
,
i
,
i
) and
i
: A
i
A
i
is the measure-preserving homomorphism dened by the usual formula
i
E

= f
1
i
[E]

. Now we can call

(X
1
,
1
,
1
, f
1
) and (X
2
,
2
,
2
, f
2
) isomorphic if there is a measure space isomorphism g : X
1
X
2
such
that f
2
= gf
1
g
1
. In this case (A
1
,
1
,
1
) and (A
2
,
2
,
2
) are isomorphic under the obvious isomorphism
(E

) = g[E]

for every E
1
.
It is not the case that if the (A
i
,
i
,
i
) are isomorphic, then the (X
i
,
i
,
i
, f
i
) are; in fact we do not
even need to have an isomorphism of the measure spaces (for instance, one could be Lebesgue measure, and
the other the Stone space of the Lebesgue measure algebra). Even when (A
1
,
1
,
1
) and (A
2
,
2
,
2
) are
actually identical, f
1
and f
2
need not be isomorphic. There are two examples in 343 of a probability space
(X, , ) with a measure space automorphism f : X X such that f(x) ,= x for every x X but the
corresponding automorphism on the measure algebra is the identity (343I, 343J); writing f
0
for the identity
map from X to itself, (X, , , f
0
) and (X, , , f) are non-isomorphic but give rise to the same (A, , ).
(c) Even with Lebesgue measure, we can have a problem in a formal sense. Take (X, , ) to be [0, 1]
with Lebesgue measure, and set f(0) = 1, f(1) = 0, f(x) = x for x ]0, 1[; then f is not isomorphic to the
identity function on X, but induces the identity automorphism on the measure algebra. But in this case we
can sort things out just by discarding the negligible set 0, 1, and for Lebesgue measure such a procedure
is eective in a wide variety of situations. To formalize it I oer the following denition.
384U Denition Let (X
1
,
1
,
1
) and (X
2
,
2
,
2
) be measure spaces, and f
1
: X
1
X
1
, f
2
: X
2
X
2
two inverse-measure-preserving functions. I will say that the structures (X
1
,
1
,
1
, f
1
) and (X
2
,
2
,
2
, f
2
)
are almost isomorphic if there are conegligible sets X
t
i
X
i
such that f
i
[X
t
i
] X
t
i
for both i and
the structures (X
t
i
,
t
i
,
t
i
, f
t
i
) are isomorphic in the sense of 384Tb, where
t
i
is the algebra of relatively
measurable subsets of X
t
i
,
t
i
is the subspace measure on X
t
i
and f
t
i
= f
i
X
t
i
.
384V I leave the elementary properties of this notion to the exercises (384Xn-384Xp), but I spell out
the result for which the denition is devised. I phrase it in the language of 342-343; if the terms are not
immediately familiar, start by imagining that both (X
i
,
i
,
i
) are measurable subspaces of R endowed with
some Radon measure (342J, 343H), or indeed that both are [0, 1] with Lebesgue measure.
Proposition Let (X
1
,
1
,
1
) and (X
2
,
2
,
2
) be perfect, complete, strictly localizable and countably
separated measure spaces, and (A
1
,
1
), (A
2
,
2
) their measure algebras. Suppose that f
1
: X
1
X
1
,
f
2
: X
2
X
2
are inverse-measure-preserving functions and that
1
: A
1
A
1
,
2
: A
2
A
2
are the
measure-preserving Boolean homomorphisms they induce. If (A
1
,
1
,
1
) and (A
2
,
2
,
2
) are isomorphic,
then (X
1
,
1
,
1
, f
1
) and (X
2
,
2
,
2
, f
2
) are almost isomorphic.
proof Because (A
1
,
1
) and (A
2
,
2
) are isomorphic, we surely have
1
X
1
=
2
X
2
. If both are zero, we
can take X
t
1
= X
t
2
= and stop; so let us suppose that
1
X
1
> 0. Let : A
1
A
2
be a measure-
preserving automorphism such that
2
=
1

1
. Because both
1
and
2
are complete and strictly
384Xi Entropy 513
localizable and compact (343K), there are inverse-measure-preserving functions g
1
: X
1
X
2
, g
2
: X
2
X
1
representing
1
, respectively (343B). Now g
1
g
2
: X
2
X
2
, g
2
g
1
: X
1
X
1
, f
2
g
1
: X
1
X
2
and
g
1
f
1
: X
1
X
2
represent, respectively, the identity automorphism on A
2
, the identity automorphism on
A
1
, the homomorphism
1

2
=
1

1
: A
2
A
1
and the homomorphism
1

1
again. Next, because
both
1
and
2
are countably separated, the sets E
1
= x : g
2
g
1
(x) = x, H = x : f
2
g
1
(x) = g
1
f
1
(x)
and E
2
= y : g
1
g
2
(y) = y are all conegligible (343F). As in part (b) of the proof of 344I, g
1
E
1
and
g
2
E
2
are the two halves of a bijection, a measure space isomorphism if E
1
and E
2
are given their subspace
measures. Set G
0
= E
1
H, and for n N set G
n+1
= G
n
f
1
1
[G
n
]. Then every G
n
is conegligible, so
X
t
1
=

nN
G
n
is conegligible. Because X
t
1
is a conegligible subset of E
1
, h = g
1
X
t
1
is a measure space
isomorphism between X
t
1
and X
t
2
= g
1
[X
t
1
], which is conegligible in X
2
. Because f
1
[G
n+1
] G
n
for each n,
f
1
[X
t
1
] X
t
1
. Because X
t
1
H, g
1
f
1
(x) = f
2
g
1
(x) for every x X
t
1
. Next, if y X
t
2
, g
2
(y) X
t
1
, so
f
2
(y) = f
2
g
1
g
2
(y) = g
1
f
1
g
2
(y) g
1
[f
1
[X
t
1
]] g
1
[X
t
1
] = X
t
2
.
Accordingly we have f
t
2
= hf
t
1
h
1
, where f
t
i
= f
i
X
t
i
for both i.
Thus h is an isomorphism between (X
t
1
, f
t
1
) and (X
t
2
, f
t
2
), and (X
1
,
1
,
1
, f
1
) and (X
2
,
2
,
2
, f
2
) are
almost isomorphic.
384X Basic exercises (a) Let (A, ) be a probability algebra and A A a partition of unity. Show
that if #(A) = n then H(A) lnn.
>>>(b) Let (A, ) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A,
enumerated as a
n
)
nN
. Set a

n
= sup
i>n
a
i
, A
n
= a
0
, . . . , a
n
, a

n
for each n. Show that H(A
n
[B)
H(A
n+1
[B) for every n, and that H(A[B) = lim
n
H(A
n
[B).
(c) Let (A, ) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A. Show
that H(A[B) = 0 i A B.
(d) Let (A, ) be a probability algebra, B a closed subalgebra of A and A a partition of unity in A. Show
that H(A[B) = H(A) i (a b) = a b for every a A, b B. (Hint: for only if, start with the case
B = 0, b, 1 \ b, 1 and use 384Gc.)
(e) Let (A, ) be a probability algebra and A, B two partitions of unity in A. Show that H(A B) =
H(A) + H(B) i (a b) = a b for all a A, b B. Show that H(A B) = H(A) i every member of
A is included in some member of B, that is, i A = A B.
(f ) Let (A
i
,
i
))
iI
be a family of probability algebras, with probability algebra free product (C,

)
(325K). Suppose that
i
: A
i
A
i
is a measure-preserving Boolean homomorphism for each i I, and that
: C C is the measure-preserving Boolean homomorphism they induce. Show that h() =

iI
h(
i
).
(Hint: use 384Gb and 384Gd to show that h() is the supremum of h(, A) as A runs over the nite
partitions of unity in

iI
A
i
. Use this to reduce to the case I = 0, 1. Now show that if A
i
A
i
is a
nite partition of unity for each i, and A = a
0
a
1
: a
0
A
0
, a
1
A
1
, then H(A) = H(A
0
) +H(A
1
), so
that h(, A) = h(
0
, A
0
) +h(
1
, A
1
).)
>>>(g) Let (A, ) be a probability algebra and : A A a measure-preserving automorphism. Show that
h(
1
) = h().
(h) Let (A, ) be a probability algebra and : A A a measure-preserving Boolean homomorphism.
Show that h(
k
) = kh() for any k N. (Hint: if A A is a partition of unity, h(
k
, A) h(
k
, D
k
(A, )) =
kh(, A).)
>>>(i) Let (A, ) be a probability algebra and : A A a measure-preserving Boolean homomorphism.
(i) Suppose there is a partition of unity A A such that () (a b) = a b for every a A, b A ()
A is the closed subalgebra of itself generated by

nN

n
[A]. Show that is a one-sided Bernoulli shift, and
that h() = H(A). (ii) Suppose that is a one-sided Bernoulli shift of nite entropy. Show that there is a
partition of unity satisfying () and ().
514 Automorphism groups 384Xj
>>>(j) Let (A, ) be the measure algebra of Lebesgue measure on [0, 1[. Fix an integer k 2, and dene
f : [0, 1[ [0, 1[ by setting f(x) = <kx>, the fractional part of kx, for every x [0, 1[; let : A B be the
corresponding homomorphism. (Cf. 372Xr.) Show that is a one-sided Bernoulli shift and that h() = lnk.
(Hint: in 384Xi, set A = a
0
, . . . , a
k1
where a
i
=
_
i
k
,
i+1
k
_

for i < k.)

>>>(k) Let (A, ) be the measure algebra of Lebesgue measure on [0, 1]. Set f(x) = 2 min(x, 1 x) for
x [0, 1] (see 372Xm). Show that the corresponding homomorphism : A A is a one-sided Bernoulli
shift and that h() = ln2. (Hint: in 384Xi, set A = a, 1 \ a where a = [0,
1
2
]

.)
(l) Let (A, ) be a probability algebra and : A A a two-sided Bernoulli shift. Show that
1
is a two-
sided Bernoulli shift and that there is a measure-preserving involution : A A such that
1
=
1
,
so that is a product of two involutions in Aut

(A).
(m) Let (A
i
,
i
))
iI
be a family of probability algebras, and (C,

) their probability algebra free product.

Suppose that for each i I we have a measure-preserving Boolean homomorphism
i
: A
i
A
i
, and that
: C C is the measure-preserving homomorphism induced by
i
)
iI
(325Xd). (i) Show that if every
i
is a one-sided Bernoulli shift so is . (ii) Show that if every
i
is a two-sided Boolean shift so is .
(n) Show that the relation almost isomorphic to (384U) is an equivalence relation.
(o) Show that the concept of almost isomorphism described in 384U is not changed if we amend the
denition to require that the subspaces X
t
1
, X
t
2
should be measurable.
(p) Show that if (X
1
,
1
,
1
, f
1
) and (X
2
,
2
,
2
, f
2
) are almost isomorphic quadruples as described in
384U, then (A
1
,
1
,
1
) and (A
2
,
2
,
2
) are isomorphic, where for each i (A
i
,
i
) is the measure algebra of
(X
i
,
i
,
i
) and
i
: A
i
A
i
is the measure-preserving Boolean homomorphism derived from f
i
: X
i
X
i
.
(q) Let (A, ) be a probability algebra, and write / for the set of partitions of unity in A not containing
0, ordered by saying that A B if every member of B is included in some member of A. (i) Show that /
is a Dedekind complete lattice, and can be identied with the lattice of purely atomic closed subalgebras
of A. Show that for A, B /, A B, as dened in 384F, is supA, B in /. (ii) Show that H(A
B) + H(A B) H(A) + H(B) for all A, B /, where , are the lattice operations on /. (iii) Set
/
1
= A : A /, H(A) < . For A, B /
1
set (A, B) = 2H(A B) H(A) H(B). Show that
is a metric on /
1
(the entropy metric). (iv) Show that if : A A is a measure-preserving Boolean
homomorphism, then [h(, A) h(, B)[ (A, B) for all A, B /
1
. (iv) Show that the lattice operations
, are -continuous on /
1
. (v) Show that H : /
1
[0, [ is order-continuous. (vi) Show that if B is
any closed subalgebra of A, then A H(A[B) is order-continuous on /
1
.
384Y Further exercises (a) Let (A, ) be a probability algebra, and write P for the lattice of closed
subalgebras of A. Show that if A is any partition of unity in A of nite entropy, then the order-preserving
function B H(A[B) : P ], 0] is order-continuous.
(b) Let (A, ) be a probability algebra, A a partition of unity in A of nite entropy, and : A A a
measure-preserving Boolean homomorphism. Show that h(, A) = lim
n
H(A[B
n
), where B
n
is the closed
subalgebra of A generated by

1in

i
[A]. (Hint: use 384Gb to show that H(A[B
n
) = H(D
n+1
(A, ))
H(D
n
(A, )) and observe that lim
n
H(A[B
n
) is dened.)
(c) Let (A, ) be a probability algebra and : A A a measure-preserving Boolean homomorphism.
Suppose that there is a partition of unity A of nite entropy such that the closed subalgebra of A generated
by

i1

i
[A] is A. Show that h() = 0. (Hint: use 384Yb and 384Pa.)
(d) Let be Lebesgue measure on [0, 1[, and take any ]0, 1[. Let f : [0, 1[ [0, 1[ be the measure
space automorphism dened by saying that f(x) is to be one of x +, x +1. Let (A, ) be the measure
algebra of ([0, 1[ , ) and : A A the measure-preserving automorphism corresponding to f. Show that
h() = 0. (Hint: if Q, use 384Xh; otherwise use 384Yc with A = a, 1 \ a where a =
_
0,
1
2
_

.)
385A More about entropy 515
(e) Set X = [0, 1] Q, let be the measure on X dened by setting E =
1
ln 2
_
E
1
1+x
dx for every
Lebesgue measurable set E X, and for x X let f(x) be the fractional part <
1
x
> of
1
x
. Recall that f
is inverse-measure-preserving for (372N). Let (A, ) be the measure algebra of (X, ) and : A A the
homomorphism corresponding to f. Show that h() =
2
/6 ln2. (Hint: use the Kolmogorov-Sina theorem
and 372Yf(v).)
(f ) Consider the triplets ([0, 1[ ,
1
, f
1
) and ([0, 1],
2
, f
2
) where
1
,
2
are Lebesgue measure on [0, 1[,
[0, 1] respectively, f
1
(x) = <2x> for each x [0, 1[, and f
2
(x) = 2 min(x, 1 x) for each x [0, 1]. Show
that these structures are almost isomorphic in the sense of 384U, and give a formula for an isomorphism.
(g) Let (A, ) be a probability algebra, and /
1
the set of partitions of unity of nite entropy not containing
0, as in 384Xq. Show that /
1
is complete under the entropy metric. (Hint: show that if A
n
)
nN
is a non-
decreasing sequence in /
1
and sup
nN
H(A
n
) < , then the closed subalgebra of A generated by

nN
A
n
is purely atomic.)
384 Notes and comments In preparing this section I have been heavily inuenced by Petersen 83.
I have taken almost the shortest possible route to Theorem 384P, the original application of the theory,
ignoring both the many extensions of these ideas and their intuitive underpinning in the concept of the
quantity of information carried by a partition. For both of these I refer you to Petersen 83. The
techniques described there are I think suciently powerful to make possible the calculation of the entropy
of any of the measure-preserving homomorphisms which have yet appeared in this treatise.
Of course the idea of entropy of a partition, or of a homomorphism, can be translated into the language of
probability spaces and inverse-measure-preserving functions; if (X, , ) is a probability space, with measure
algebra (A, ), then partitions of unity in A correspond (subject to decisions on how to treat negligible sets)
to countable partitions of X into measurable sets, and an inverse-measure-preserving function from X to
itself gives rise to a measure-preserving homomorphism
f
: A A; so we can dene the entropy of f to
be h(
f
). The whole point of the language I have sought to develop in this volume is that we can do this
when and if we choose; in particular, we are not limited to those homomorphisms which are representable
by inverse-measure-preserving functions. But of course a large proportion of the most important examples
do arise in this way (see 384Xj, 384Xk). The same two examples are instructive from another point of view:
the case k = 2 of 384Xj is (almost) isomorphic to the tent map of 384Xk. The similarity is obvious, but
exhibiting an actual isomorphism is I think another matter (384Yf).
I must say almost isomorphic here because the doubling map on [0, 1[ is everywhere two-to-one, while
the tent map is not, so they cannot be isomorphic in any exact sense. This is the problem grappled with
in 384T-384V. In some moods I would say that a dislike of such contortions is a sign of civilized taste.
Certainly it is part of my motivation for working with measure algebras whenever possible. But I have to
say also that new ideas in this topic arise more often than not from actual measure spaces, and that it is
absolutely necessary to be able to operate in the more concrete context.
385 More about entropy
In preparation for the next two sections, I present a number of basic facts concerning measure-preserving
homomorphisms and entropy. Compared with the work to follow, they are mostly fairly elementary, but the
Halmos-Rokhlin-Kakutani lemma (385E) and the Shannon-McMillan-Breiman theorem (385G), in their full
strengths, go farther than one might expect.
385A Periodic and aperiodic parts If X is a set and f : X X is a bijection, then the orbits

x
= f
n
(x) : n Z of f are described by their cardinalities, and X has a natural decomposition into
the sets X
i
= x : #(
x
) = i for 1 i . Corresponding to this is a partition of unity of a Dedekind
complete Boolean algebra with an automorphism, as follows.
Denition If A is a Boolean algebra, a Boolean homomorphism : A A is periodic, with period n 2,
if A ,= 0,
n
is the identity operator and whenever b A 0 and 1 i < n there is a c b such that
516 Automorphism groups 385A

i
c ,= c. is periodic with period 1 i it is the identity operator. is aperiodic if for every non-zero
b A, n 1 there is a c b such that
n
c ,= c. I remark immediately that if is aperiodic, so is
n
for
every n 1. Note that if A = 0 then the trivial automorphism of A is counted both as aperiodic and as
periodic with period 1.
385B Lemma Let A be a Dedekind complete Boolean algebra and : A A a Boolean homomorphism
which is periodic with period n 2. Then is a Boolean automorphism and there is an a A such that
(a, a,
2
a, . . . ,
n1
a) is a partition of unity in A; that is (in the language of 381G) is of the form
(

a
1
a
2
. . .

a
n
) where (a
1
, . . . , a
n
) is a partition of unity in A.
proof Because
n
: A A is injective and surjective, so is , and is a Boolean automorphism, therefore
order-continuous. Set D = d : d A,
i
d d = 0 whenever 1 i < n. Then the supremum of any
upwards-directed subset of D is dened in A (because A is Dedekind complete) and belongs to D (because

i
is order-continuous for every i use 313Bc); so D has a maximal element a, by Zorns lemma. ??? If
(a, a, . . . ,
n1
a) is not a partition of unity in A, set b
0
= 1 \ (a a . . .
n1
a). Then
b
0
= 1 \ (a . . .
n1
a a) = b
0
.
Now we can choose b
1
, . . . , b
n1
such that 0 ,= b
i
b
i1
and
i
b
i
b
i
= 0 for 1 i < n. PPP Given that
b
i1
,= 0, where 1 i < n, then (by the denition of periodic homomorphism) there is a c b
i1
such
that
i
c ,= c. If c \
i
c ,= 0, take b
i
= c \
i
c. Otherwise,
i
c \ c ,= 0 so c \
i
c ,= 0 and we can take
b
i
= c \
i
c. QQQ At the end of this induction, set d = b
n1
; then d
i
d = 0 for 1 i < n, so d D; also
d
i
a =
i
d a = 0 for 1 i < n (because
i
d
i
b
0
= b
0
for every i), so a d D, which is supposed to
be impossible. XXX
Thus (a, a, . . . ,
n1
a) is a partition of unity, as required.
385C Proposition Let A be a Dedekind complete Boolean algebra and : A A an injective order-
continuous Boolean homomorphism. Then there is a partition of unity c
i
)
1i
in A such that c
i
= c
i
for
every i and A
c
n
is periodic with period n whenever n is nite and c
n
,= 0, while A
c

is aperiodic.
Remark As usual, I write A
a
for the principal ideal of A generated by a.
proof For each n 1, set
B
n
= b : b A,
n
c = c for every c b, b
n
= sup B
n
.
Because is a Boolean homomorphism, B
n
is an ideal of A; because is order-continuous, b
n
B
n
, that
is, B
n
is the principal ideal A
b
n
. Also b
n
= b
n
. PPP If b B
n
and c b, then
n1
c
n
b = b, so

n1
(
n
c) =
n
(
n1
c) =
n1
c.
But
n1
, like , is injective, so
n
c = c. As c is arbitrary, b B
n
. As is order-continuous,
b
n
= sup
bB
n
b b
n
.
But this means that
i+1
b
n

i
b
n
for every i, so
b
n
=
n
b
n
b
n
b
n
and b
n
= b
n
. QQQ
Set
c
n
= b
n
\ sup
1i<n
b
i
for n N, c

= 1 \ sup
n1
b
n
= 1 \ sup
n1
c
n
.
Then c
i
)
1i
serves. PPP Because b
n
= b
n
for every n, c
i
= c
i
for every i . So c c
i
= c
i
for
every c c
i
, and A
c
i
is a Boolean homomorphism from A
c
i
to itself, for every i . If 1 i < and
c c
i
, then c b
i
so
i
c = c; accordingly (A
c
i
)
i
= A
c
i
. If 1 j < i and c c
i
is non-zero, then
c , b
j
, so there is a non-zero d c such that
j
d ,= d. This shows both that A
c
i
is periodic with period i
if 1 i < and that A
c

is aperiodic. So we have an appropriate partition c

i
)
i
. QQQ
385D The following elementary fact is essential.
385E More about entropy 517
Lemma Let (A, ) be a totally nite measure algebra and : A A a measure-preserving Boolean
homomorphism. Then
(sup
nN

n
a) = sup
nN

n
a, a sup
n1

n
a
for every a A.
proof Set c = sup
nN

n
a. Then c = sup
n1

n
, because is order-continuous (324Kb), so c c; as
is measure-preserving and c < , c = c. Now
a c = c = sup
nN

n+1
a = sup
n1

n
a.
Remark See 382Yc-382Yd.
385E The Halmos-Rokhlin-Kakutani lemma Let (A, ) be a totally nite measure algebra and
: A A a measure-preserving Boolean homomorphism. Set C = c : c = c. Then C is a closed
subalgebra of A, and the following are equiveridical:
(i) is aperiodic;
(ii) A is relatively atomless over C (denition: 331A);
(iii) whenever n 1 and 0 <
1
n
there is an a A such that a, a,
2
a, . . . ,
n1
a are disjoint and
(a c) = c for every c C;
(iv) whenever n 1, 0 <
1
n
and B A is nite, there is an a A such that a, a,
2
a, . . . ,
n1
a
are disjoint and (a b) = b for every b B.
proof Of course C is a subalgebra of A because is a Boolean homomorphism, and is (order-)closed because
is (order-)continuous (324Kb).
(i)(ii) ??? Suppose, if possible, that is aperiodic, but there is a non-zero a A such that A
a
= a c :
c C. Let n 1 be such that b = a
n
a ,= 0 (385D). If d b there is some c C such that d = a c and

n
d =
n
a c b c d;
because is measure-preserving,
n
d = d. But this means that is not aperiodic. XXX
(ii)(iii) Set =
1
n
(
1
n
) > 0. By 331B, there is a d A such that (c d) = c for every c C. Set
d
k
=
k
d \ sup
i<k

i
d for k N. Note that
d
j+k
=
j+k
d \ sup
i<j+k

i
d
j+k
d \ sup
i<k

j+i
d =
j
d
k
whenever j, k N. Next,
i
d
j
d
k
sup
mi
d
m
for any i, j, k N such that i +j ,= k. PPP () If k i this
is obvious. () If i < k < i +j then

i
d
j
d
k

i
d
j

i
d
ki
=
i
(d
j
d
ki
) = 0.
() If i +j < k, then

i
d
j
d
k

i+j
d d
k
= 0. QQQ
Setting c

= sup
iN
d
i
= sup
iN

i
d, we have c

= c

, by 385D, so that c

C and (d \ c

) = (1 \ c

);
but as d c

, c

= 1.
Set a

= sup
mN
d
mn
(the mn here is a product, not a double subscript!), d

= sup
i<n
d
i
= sup
i<n

i
d.
Then
(c d

n1
i=0
(c
i
d) =

n1
i=0

i
(c d) = n (c d) = n c
for every c C. Next,
i
d
mn
d
mn+i
for all m and i, so
sup
i<n

i
a

= sup
iN
d
i
= 1.
Consequently
c

n1
i=0
(c
i
a

) = n (c a

),
(c a

\ d

) (c a

) (c d

) (
1
n
n) c = c
for every c C.
518 Automorphism groups 385E
By 331B again (applied to the principal ideal of A generated by a

\ d

) there is an a a

\ d

such that
(a c) = c for every c C. For 0 < i < n,

i
a

= sup
k,lN

i
d
kn
d
ln
sup
mi
d
m
d

,
so
i
a a = 0; accordingly a, a, . . . ,
n1
a are all disjoint and (iii) is satised.
(iii)(iv) Note that A is certainly atomless, since for every k 1 we can nd a c A such that
c, c, . . . ,
k1
c are disjoint and c =
1
k+1
, so that we have a partition of unity consisting of sets of measure
1
k+1
. Let B
t
be the set of atoms of the (nite) subalgebra of A generated by B, and m = #(B
t
). Let > 0
and r, k N be such that
3 (1 n) b for every b B
t
, m( 1)
2
< r
2
, k 1.
By (iii), there is a c A such that c, c, . . . ,
nr(k+1)1
c are disjoint and (sup
i<nr(k+1)

i
c) = 1 . For
j < r, set e
j
= sup
n(k+1)ji<n(k+1)(j+1)

i
c, d
j
= sup
i<k

n(k+1)j+ni
c. Observe that d
j
, d
j
, . . . ,
n1
d
j
are
disjoint, and that
i
d
j
e
j
for i < 2n. Set e = sup
j<r
e
j
= sup
i<nr(k+1)

i
c, so that e = 1 .
Suppose we choose d A by the following random process. Take s(0), . . . , s(r 1) independently in
0, . . . , n 1, so that Pr(s(j) = l) =
1
n
for each l < n, and set d = sup
j<r

s(j)
d
j
. Because we certainly
have
i

s(j)
d
j
e
j
whenever i < n, d, d, . . . ,
n1
d will be disjoint. Now for any b A,
Pr
_
(d b)
1
n
( b 3)
_
<
1
m
.
PPP We can express the random variable (d b) as X =

r1
j=0
X
j
, where X
j
= (
s(j)
d
j
b). Then the X
j
are independent random variables. For each j, X
j
takes values between 0 and k c
1
nr
, and has expectation
1
n
(e
t
j
b), where
e
t
j
= sup
i<n

i
d
j
= sup
n(k+1)ji<n(k+1)j+nk

i
c.
So X has expectation
1
n
(e
t
b) where e
t
= sup
j<r
e
t
j
. Now
e
j
\ e
t
j
= sup
n(k+1)j+nki<n(k+1)(j+1)

i
c
has measure n c
n 1
nr(k+1)
for each j, so (e \ e
t
)
1
k+1
and (1 \ e
t
) 2; thus E(X)
1
n
( b 2), while
Var(X) =

r1
j=0
Var(X
j
) r
_
1
nr
_
2
=
( 1)
2
n
2
r
.
But this means that
( 1)
2
n
2
r

_

n
_
2
Pr
_
X
1
n
( b 3)
_
,
and
Pr
_
X
1
n
( b 3)
_

( 1)
2
r
2
<
1
m
by the choice of r. QQQ
This is true for every b B
t
, while #(B
t
) = m. There must therefore be some choice of s(0), . . . , s(r 1)
such that, taking d

= sup
j<r

s(j)
d
j
,
(d

b)
1
n
( b 3) b
for every b B
t
, while d

, d

, . . . ,
n1
d

are disjoint. Because A is atomless, there is a d d

such that
(d b) = b for every b B
t
. Since every member of B is a disjoint union of members of B
t
, (d b) = b
for every b B.
(iv)(i) If a A 0, 1 and n 1 then (iv) tells us that there is a b A such that b, b, . . . ,
n
b are
all disjoint and (1 \ sup
in

i
b) < a. Now there must be some i < n such that d =
i
b a ,= 0, in which
case
d
n
d
i
b
i+n
b =
i
(b
n
b) = 0,
385G More about entropy 519
and
n
d ,= d. As n and a are arbitrary, is aperiodic.
385F Corollary An ergodic measure-preserving Boolean homomorphism on an atomless totally nite
measure algebra is aperiodic.
proof This is (ii)(i) of 385E in the case C = 0, 1.
385G I turn now to a celebrated result which is a kind of strong law of large numbers.
The Shannon-McMillan-Breiman theorem Let (A, ) be a probability algebra, : A A a measure-
preserving Boolean homomorphism, and A A a partition of unity of nite entropy. For each n 1,
set
w
n
=
1
n

dD
n
(A,)
ln(
1
d
)d,
where D
n
(A, ) is the partition of unity generated by
i
a : a A, i < n, as in 384K. Then w
n
)
nN
is norm-convergent in L
1
= L
1
(A, ) to w say; moreover, w
n
)
nN
is order*-convergent to w (denition:
367A). If T : L
0
(A) L
0
(A) is the Riesz homomorphism dened by , so that T(a) = (a) for every
a A (364R), then Tw = w.
proof (Petersen 83) We may suppose that 0 / A.
(a) For each n N, let B
n
be the subalgebra of A generated by
i
a : a A, 1 i n, B
n
the set of
its atoms, and P
n
the corresponding conditional expectation operator on L
1
= L
1
(A) (365R). Let B be the
closed subalgebra of A generated by

nN
B
n
, and P the corresponding conditional expectation operator.
Observe that B
n
= [D
n
(A, )] and that, in the language of 384F, D
n+1
(A, ) = A B
n
. Let C be the
xed-point algebra c : c A, c = c and Q the associated conditional expectation. Set L
0
= L
0
(A), and
let

ln be the function from v : [[v > 0]] = 1 to L
0
corresponding to ln : ]0, [ R (364I).
(b) It will save a moment later if I note an elementary fact here: if v L
1
, then
1
n
T
n
v)
n1
is order*-
convergent and | |
1
-convergent to 0. PPP We know from the ergodic theorem (372G) that v
n
)
nN
is order*-
convergent and | |
1
-convergent to Qv, where v
n
=
1
n+1

n
i=0
T
i
v. Now
1
n
T
n
v =
n+1
n
v
n
v
n1
is order*-
convergent and | |
1
-convergent to Qv Qv = 0 (using 367C for order*-convergent). QQQ
(c) Set
v
n
=

aA
P
n
(a) a =

aA,bB
n
(ab)
b
(a b).
By Levys martingale theorem (275I, 367Kb),
v
n
a)
nN
= P
n
(a) a)
nN
is order*-convergent to P(a) a for every a A; consequently v
n
)
nN
order*-converges to v =

aA
P(a) a. It follows that

lnv
n
)
nN
order*-converges to

lnv. PPP The point is that, for any a A,
n N, a [[P
n
(a) > 0]], so that [[v
n
> 0]] = 1 for every n, and

lnv
n
is dened. Similarly,

lnv is dened,
and

lnv
n
)
nN
order*-converges to

lnv by 367I. QQQ As 0 v
n
1 for every n, v
n
)
nN
v for | |
1
, by
the Dominated Convergence Theorem (367J).
Next,

lnv
n
)
nN
is order-bounded in L
1
. PPP Of course

lnv
n
0 for every n, because P
n
(a) P
n
(1)
1 for each a, so v
n
1. To see that

lnv
n
: n N is bounded below in L
1
, we use an idea from the
fundamental martingale inequality 275D. Set v

= inf
nN
v
n
. For > 0, a A and n N set
b
an
() = [[P
n
(a) < ]] inf
i<n
[[P
i
(a) ]],
so that
[[v

< ]] = sup
aA,nN
a b
an
().
Now b
an
() B
n
, so
(a b
an
()) =
_
b
an
()
a =
_
b
an
()
P
n
(a) (b
an
()),
and
520 Automorphism groups 385G
(a [[v

< ]]) min( a,

n=0
(a b
an
()))
min( a,

n=0
b
an
()) min( a, ).
Letting 0, (a [[v

= 0]]) = 0 for every a A, so [[v

> 0]] = 1, and

lnv

is dened. Moreover,
(a [[

lnv

> ln]]) = (a [[v

< ]]) min( a, )

for every a A, > 0; that is,
(a [[

lnv

> ]]) min( a, e

)
for every a A, R. Accordingly
_
(

lnv

) =
_

0
[[

lnv

> ]]d =

aA
_

0
(a [[

lnv

> ]])d

aA
_

0
min( a, e

)d
=

aA
_
_
ln(1/ a)
0
a d +
_

ln(1/ a)
e

d
_
=

aA
_
ln(
1
a
) a +e
ln a
_
=

aA
ln(
1
a
) a +

aA
a = H(A) + 1 <
because A has nite entropy. But this means that

lnv

belongs to L
1
, and of course it is a lower bound for

lnv
n
: n N. QQQ
By 367J again,

lnv L
1
and

lnv
n
)
nN


lnv for | |
1
.
(d) Fix n N for the moment. For each d D
n+1
(A, ) let d
t
be the unique element of B
n
such that
d d
t
. Then
(n + 1)w
n+1
=

dD
n+1
(A,)
ln(
1
d

)d

dD
n+1
(A,)
ln(
d
d

)d
=

bB
n
ln(
1
b
)b

aA,bB
n
,ab,=0
ln(
(ab)
b
)(a b)
=

dD
n
(A,)
ln(
1
(d)
)(d)

lnv
n
= T(nw
n
)

lnv
n
.
Inducing on n, starting from
w
1
=

aA
ln(
1
a
)a =

lnv
0
,
we get
nw
n
=

n1
i=0
T
i
(

lnv
ni1
), w
n
=
1
n

n1
i=0
T
i
(

lnv
ni1
)
for every n 1.
(e) Set w
t
n
=
1
n

n1
i=0
T
i
(

lnv) for n 1. By the Ergodic Theorem, w
t
n
)
n1
is order*-convergent and
| |
1
-convergent to w = Q(

lnv), and Tw = w. To estimate w
n
w
t
n
, set u

n
= sup
kn
[

lnv
k


lnv[ for
385I More about entropy 521
each n N. Then u

n
)
nN
is a non-increasing sequence, u

0
L
1
(by (c) above), and inf
nN
u

n
= 0 because

lnv
n
)
nN
order*-converges to

lnv. Now, whenever n > m N,
[w
n
w
t
n
[
1
n
n1

i=0
T
i
[

lnv

lnv
ni1
[
=
1
n
_
nm1

i=0
T
i
[

lnv

lnv
ni1
[ +
n1

i=nm
T
i
[

lnv

lnv
ni1
[
_

1
n
_
nm1

i=0
T
i
u

m
+
m1

j=0
T
n1j
[

lnv

lnv
j
[
_

1
nm
_
nm1

i=0
T
i
u

m
+
m1

j=0
T
n1j
u

0
_
=
1
nm
nm1

i=0
T
i
u

m
+
1
nm
T
nm
m1

j=0
T
m1j
u

1
nm
nm1

i=0
T
i
u

m
+
1
nm
T
nm
u
m
,
setting u
m
=

m1
j=0
T
m1j
u

0
.
Holding m xed and letting n , we know that
1
nm

nm1
i=0
T
i
u

m
is order*-convergent and | |
1
-convergent to Qu

m
. As for the other term,
1
nm
T
nm
u
m
is order*-convergent
and | |
1
-convergent to 0, by (b). What this means is that
limsup
n
[w
n
w
t
n
[ Qu

m
,
limsup
n
|w
n
w
t
n
|
1
|Qu

m
|
1
for every m N. Since Qu

m
)
mN
is surely a non-decreasing sequence with inmum 0,
limsup
n
[w
n
w
t
n
[ = 0, limsup
n
|w
n
w
t
n
|
1
= 0.
Since w
t
n
is order*-convergent and | |
1
-convergent to w, so is w
n
.
385H Corollary If, in 385G, is ergodic, then w
n
)
nN
is order*-convergent and | |
1
-convergent to
h(, A)1.
proof Because the limit w in 385G has Tw = w, it must be of the form 1, because is ergodic. Now
=
_
w must be
lim
n
_
w
n
= lim
n
1
n

dD
n
(A,)
ln(
1
d
) d = lim
n
1
n

dD
n
(A,)
q( d)
(where q is the function of 384A)
= lim
n
1
n
H(D
n
(A, )) = h(, A).
385I Denition Set p(t) = t lnt for t > 0, p(0) = 0; for any Dedekind -complete Boolean algebra A,
let p : L
0
(A)
+
L
0
(A) be the corresponding function, as in 364I. (Thus p = q where q is the function of
384A.)
522 Automorphism groups 385J
385J Lemma (Csisz ar 67, Kullback 67) Let (A, ) be a probability algebra, and u a member of
L
1
(A, )
+
such that
_
u = 1. Then
(
_
[u 1[)
2
2
_
p(u).
proof Set a = [[u < 1]], = a, =
_
a
u, b = 1 \ a. Then b = 1 and
_
b
u = 1 . Surely < 1.
If = 0 then u = 1 and the result is trivial; so let us suppose that 0 < < 1. Because the function p is
convex,
_
a
p(u) a p(
1
a
_
a
u) = p(

) = p() ln,
(using 233Ib/365Rb for the rst inequality), and similarly
_
b
p(u) p(1 ) (1 ) ln(1 ).
Also
_
[u 1[ =
_
a
(1 u) +
_
b
(u 1) = + (1 ) (1 ) = 2( ),
so
_
p(u)
1
2
(
_
[u 1[)
2
p() ln +p(1 ) (1 ) ln(1 ) 2( )
2
= ()
say. Now is continuous on [0, 1] and arbitrarily often dierentiable on ]0, 1[,
() = 0,

t
(t) = lnt ln ln(1 t) + ln(1 ) + 4( t) for t ]0, 1[,

t
() = 0,

tt
(t) =
1
t
+
1
1t
4 0 for t ]0, 1[.
So (t) 0 for t [0, 1] and, in particular, () 0; but this means that
_
p(u)
1
2
(
_
[u 1[)
2
0,
that is, (
_
[u 1[)
2
2
_
p(u), as claimed.
385K Corollary Whenever (A, ) is a probability algebra and A, B are partitions of unity of nite
entropy,

aA,bB
[ (a b) a b[
_
2(H(A) +H(B) H(A B)).
proof Replacing A, B by A 0 and B 0 if necessary, we may suppose that neither A nor B contains
0. Let (C,

) be the probability algebra free product of (A, ) with itself (325E, 325K). Set
u =

aA,bB
(ab)
a b
(a b) L
0
(C);
then u is non-negative and integrable and
_
u =

aA,bB
(a b) = 1. Now
385N More about entropy 523
_
p(u) =

aA,bB
(a b) ln
(ab)
a b
= H(A B)

aA,bB
(a b) ln a

aA,bB
(a b) ln b
= H(A B)

aA
a ln a

bB
b ln b
= H(A) +H(B) H(A B).
On the other hand,
_
[u 1[ =

aA,bB
a b[
(ab)
a b
1[ =

aA,bB
[ (a b) a b[.
So what we are seeking to prove is that
_
[u 1[
_
2
_
p(u),
which is 385J.
385L The next six lemmas are notes on some more or less elementary facts which will be used at various
points in the next section. The rst two are nearly trivial.
Lemma Let (A, ) be a probability algebra and a
i
)
iI
, b
i
)
iI
two partitions of unity in A. Then
(sup
iI
a
i
b
i
) = 1
1
2

iI
(a
i
.b
i
).
proof
(sup
iI
a
i
b
i
) =

iI
(a
i
b
i
) =

iI
1
2
( a
i
+ b
i
(a
i
.b
i
))
= 1
1
2

iI
(a
i
.b
i
).
385M Lemma Let (A, ) be a totally nite measure algebra, B
k
)
kN
a non-decreasing sequence of
subsets of A such that 0 B
0
, and c
i
)
iI
a partition of unity in A such that c
i

kN
B
k
for every i I.
Then
lim
k
sup
iI
(c
i
, B
k
) = 0,
writing (c, B) = inf
bB
(c .b) for c A and non-empty B A, as in 3A4I.
proof Let > 0. Then J = j : j I, c
j
is nite. For each j J, lim
k
(c
i
, B
k
) = 0, by 3A4I,
while
(c
i
, B
k
) (c
i
.0) = c
i

for every i I J. So
limsup
k
sup
iI
(c
i
, B
k
) max(, limsup
k
sup
iJ
(c
i
, B
k
)) = .
As is arbitrary, we have the result.
385N Lemma Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean
homomorphism. Let A, B and C be partitions of unity in A.
(a) H(A B C) +H(C) H(B C) +H(A C).
(b) h(, A) h(, A B) h(, A) +h(, B) h(, A) +H(B).
(c) If H(A) < ,
524 Automorphism groups 385N
h(, A) = inf
nN
H(D
n+1
(A, )) H(D
n
(A, ))
= lim
n
H(D
n+1
(A, )) H(D
n
(A, )).
(d) If H(A) < and B is any closed subalgebra of A such that [B] B, then h(, A) h(B) +
H(A[B).
proof (a) Let C be the closed subalgebra of A generated by C, so that C is purely atomic and C is the set
of its atoms. Then
H(A B C) +H(C) = H(A B[C) + 2H(C)
H(A[C) +H(B[C) + 2H(C) = H(A C) +H(B C)
by 384Gb and 384Ga.
(b) We need only observe that D
n
(AB, ) = D
n
(A, ) D
n
(B, ) for every n N, being the partition
of unity generated by
i
a : i < n, a A
i
b : i < n, b B. Consequently
h(, A) = lim
n
1
n
H(D
n
(A, )) lim
n
1
n
H(D
n
(A, ) D
n
(B, ))
= lim
n
1
n
H(D
n
(A B, )) = h(, A B)
lim
n
1
n
(H(D
n
(A, ) +H(D
n
(B, ))) = h(, A) +h(, B)
h(, A) +H(B)
as remarked in 384M.
(c) Set
n
= H(D
n+1
(A, )) H(D
n
(A, )) for each n N. By 384H,
n
0. From (a) we see that

n+1
= H(A [D
n+1
(A, )]) H(A [D
n
(A, )])
H([D
n+1
(A, )]) H([D
n
(A, ]) =
n
for every n N. So lim
n

n
= inf
nN

n
; write for the common value. Now
h(, A) = lim
n
1
n
H(D
n
(A, )) = lim
n
1
n

n1
i=0

i
=
(273Ca).
(d) Let P : L
1

L
1

be the conditional expectation operator corresponding to B. Let b
k
)
kN
be a
sequence running over [[P(a) > q]] : a A, q Q, so that b
k
B for every k, and for each k N
let B
k
B be the subalgebra generated by b
i
: i k; let P
k
be the conditional expectation operator
corresponding to B
k
. Writing B

B for

kN
B
k
, and P

for the corresponding conditional expectation

operator, then P(a) L
0
(B

), so P

(a) = P(a), for every a A. So

H(A[B) =

aA
_
q(Pa) = H(A[B

) = lim
k
H(A[B
k
),
by 384Gd.
For each k, let B
k
be the set of atoms of B
k
. Then
h(, A) h(, B
k
) +H(A[B
k
) h(B) +H(A[B
k
)
by 384N and the denition of h(B). So
h(, A) h(B) + lim
k
H(A[B
k
) = h(B) +H(A[B).
385O More about entropy 525
385O Lemma Let (A, ) be a probability algebra and B a closed subalgebra.
(a) There is a function h : A B such that (a .h(a)) = (a, B) for every a A and h(a) h(a
t
) = 0
whenever a a
t
= 0.
(b) If A is a partition of unity in A, then H(A[B)

aA
q((a, B)), where q is the function of 384A.
(c) If B is atomless and a
i
)
iN
is a partition of unity in A, then there is a partition of unity b
i
)
iN
in
B such that b
i
= a
i
and (b
i
.a
i
) 2(a
i
, B) for every i N.
proof (a) Let P : L
1

L
1

be the conditional expectation operator associated with B. For any c B,
_
[P(a) c[ =
_
1\c
P(a) + c
_
c
P(a) =
_
1\c
a + c
_
c
a
= (a \ c) + c (a c) = (a .c).
If a A set h(a) = [[P(a) >
1
2
]]. Then [P(a) h(a)[ [P(a) c[ for any c B, so
(a, B) = inf
cB
(a .c) = inf
cB
_
[P(a) c[
=
_
[P(a) h(a)[ = (a .h(a)).
If a a
t
= 0, then
P(a) +P(a
t
) = P(a a
t
) 1,
so
h(a) h(a
t
) = [[P(a) >
1
2
]] [[P(a
t
) >
1
2
]] [[P(a) +P(a
t
) > 1]] = 0,
by 364D(b-i).
(b) By 384Ae, q(1t) q(t) whenever 0 t
1
2
. Consequently q(t) q(min(t, 1t)) for every t [0, 1],
and q(u) q(u (1 u)) whenever u L
0
(A) and 0 u 1. Fix a A for the moment. We have
q(P(a)) q(P(a) (1 P(a)) = q([P(a) h(a)[).
Consequently
_
q(Pa)
_
q([P(a) h(a)[) q
_
_
[P(a) h(a)[
_
(because q is concave)
= q((a, B)).
Summing over a,
H(A[B) =

aA
_
q(Pa)

aA
q((a, B)).
(c) Set b
t
i
= h(a
i
) for each i N. Then b
t
i
)
iN
is disjoint. Next, for each i N, take b
tt
i
B such that
b
tt
i
b
t
i
and b
tt
i
= min( a
i
, b
tt
i
); then b
tt
i
)
iN
is disjoint and b
tt
i
a
i
for every i. We can therefore nd a
partition of unity b
i
)
iN
such that b
i
b
tt
i
and b
i
= a
i
for every i. (Use 331C to choose d
i
)
iN
inductively
so that d
i
1 \ (sup
j<i
d
j
sup
jN
b
tt
j
) and d
i
= a
i
b
tt
i
for each i, and set b
i
= b
tt
i
d
i
.)
Take any i N. If b
t
i
> a
i
, then
(a
i
.b
i
) = (a
i
.b
tt
i
) (a
i
.b
t
i
) + (b
t
i
.b
tt
i
)
= (a
i
.b
t
i
) + b
t
i
a
i
2 (a
i
.b
t
i
) = 2(a
i
, B).
If b
t
i
a
i
, then
(a
i
.b
i
) (a
i
.b
t
i
) + (b
t
i
.b
i
)
= (a
i
.b
t
i
) + a
i
b
t
i
2 (a
i
.b
t
i
) = 2(a
i
, B).
526 Automorphism groups 385P
385P Lemma Let (A, ) be a probability algebra and : A A a measure-preserving automorphism.
Suppose that B A. For k N, let B
k
be the closed subalgebra of A generated by
j
b : b B, [j[ k,
and let B be the closed subalgebra of A generated by
j
b : b B, j Z. Then
(a) B is the topological closure

kN
B
k
.
(b) [B] = B.
(c) If C is any closed subalgebra of A such that [C] = C, and a B
k
, then
(a, C) (2k + 1)

bB
(b, C).
proof (a) Because B
k
)
kN
is non-decreasing,

kN
B
k
is a subalgebra of A, so its closure also is (323J),
and must be B.
(b) Of course
1
[B
k+1
] is a closed subalgebra of A containing
j
b whenever [j[ k and b B, so
includes B
k
; thus [B
k
] B
k+1
B for every k, and
[B] = [

kN
B
k
]

kN
[B
k
] B B
because is continuous (324Kb). Similarly,
1
[B] B and [B] = B.
(c) For each b B, choose c
b
C such that (b .c
b
) = (c
b
, C) (385Oa). Set
e = sup
]j]k
sup
bB

j
(b .c
b
);
then
e (2k + 1)

bB
(b .c
b
) = (2k + 1)

bB
(b, C).
Now
B
t
= d : d A, c C such that d \ e = c \ e
is a subalgebra of A. By 314Fa, applied to the order-continuous homomorphism c c \ e : C A
1\e
,
c \ e : c C is an order-closed subalgebra of the principal ideal A
1\e
; by 313Id, applied to the order-
continuous function d d \ e : A A
1\e
, B
t
is order-closed. If b B and [j[ k, then
j
b .
j
c
b
e, so

j
b B
t
; accordingly B
t
B
k
. Now a B
k
, so there is a c C such that a .c e, and
(a, C) (a .c) e (2k + 1)

bB
(b, C),
as claimed.
385Q Lemma Let (A, ) be a probability algebra and suppose either that A is not purely atomic or
that it is purely atomic and H(D
0
) = , where D
0
is the set of atoms of A. Then whenever A A is a
partition of unity and H(A) , there is a partition of unity B, rening A, such that H(B) = .
proof (a) By 384J, there is a partition of unity D
1
such that H(D
1
) = . Set D = D
1
A; then we
still have H(D) = . Enumerate D as d
i
)
iN
. Choose B
k
)
kN
inductively, as follows. B
0
= A. Given
that B
k
is a partition of unity, then if H(B
k
d
k
, 1 \ d
k
) , set B
k+1
= B
k
d
k
, 1 \ d
k
; otherwise set
B
k+1
= B
k
.
Let B be the closed subalgebra of A generated by

kN
B
k
. Note that, for each d D,
c : c A, d c or d c = 0
is a closed subalgebra of A including every B
k
, so includes B. If b B 0, there is surely some d D
such that b d ,= 0, so b d; thus B must be purely atomic. Let B be the set of atoms of B. Because
A = B
0
B, B renes A.
(b) H(B) . PPP For each k N, let B
k
be the closed subalgebra of A generated by B
k
, so that
B =

kN
B
k
. Suppose that b
0
, . . . , b
n
are distinct members of B. Then for each k N we can nd disjoint
b
0k
, . . . , b
nk
B
k
such that (b
ik
.b
i
) (b
i
, B
k
) for every i n (385Oa). Accordingly b
i
= lim
k
b
ik
for each i, and

n
i=0
q( b
i
) = lim
k

n
i=0
q( b
ik
) sup
kN
H(B
k
) .
As b
0
, . . . , b
n
are arbitrary, H(B) . QQQ
385Yd More about entropy 527
(c) H(B) . PPP??? Suppose otherwise. We know that
lim
k
H(d
k
, 1 \ d
k
) = lim
k
q( d
k
) +q(1 d
k
) = 0.
Let m N be such that H(B) +H(d
k
, 1 \ d
k
) for every k m. Because B renes B
k
, we must have
H(B
k
d
k
, 1 \ d
k
) H(B
k
) +H(d
k
, 1 \ d
k
) ,
so that B
k+1
= B
k
d
k
, 1 \ d
k
for every k m. But this means that d
k
B for every k m, so that
> H(B)

k=m
q( d
k
) = ,
which is impossible. XXXQQQ
Thus B has the required properties.
385X Basic exercises >>>(a) Let A be a Boolean algebra, not 0, and : A A an automorphism;
set C = c : c = c. Show that is periodic, with period n 1, i A
c
has order n in the group Aut A
c
whenever c C 0. Show that is aperiodic i A
c
has innite order in the group Aut A
c
whenever
c C 0.
(b) In 385C, show that the family c
i
)
1i
is uniquely determined.
>>>(c) Show that for a Bernoulli shift , the Shannon-McMillan-Breiman theorem is a special case of the
Ergodic Theorem. (Hint: w
n
=
1
n

n1
i=0
T
i
w
1
.)
(d) Let (A, ) be a totally nite measure algebra, B
k
)
kN
a non-decreasing sequence of subsets of A such
that 0 B
0
, and c
i
)
iI
a partition of unity in A. Show that
lim
k

iI
(c
i
, B
k
) =

iI
(c
i
, B)
where B =

kN
B
k
.
>>>(e) Let (A, ) be a probability algebra, : A A a measure-preserving Boolean homomorphism and
A a partition of unity in A. Show that h(, D
n
(A, )) = h(, A) = h(, [A]) for any n 1.
(f ) Let (A, ) be a totally nite measure algebra and : A A a measure-preserving Boolean ho-
momorphism. Suppose that B A. For k N, let B
k
be the closed subalgebra of A generated by

j
b : b B, j k, and let B be the closed subalgebra of A generated by
j
b : b B, j N. Show that
B =

kN
B
k
, [B] B,
and that if C is any closed subalgebra of A such that [C] C, and a B
k
, then (a, C) (k+1)

bB
(b, C).
385Y Further exercises (a) Give an example to show that the word injective in the statement of
385C is essential.
(b) Let (A, ) be a totally nite measure algebra and : A A an aperiodic measure-preserving Boolean
homomorphism. Set C = c : c = c. Show that whenever n 1, 0 <
1
n
and B A is nite, there is
an a A such that a, a,
2
a, . . . ,
n1
a are disjoint and (a b c) = (b c) for every b B, c C.
(c) Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean homomorphism.
Let P be the set of all closed subalgebras of A which are invariant under , ordered by inclusion. Show that
B h(B) : P [0, ] is order-preserving and order-continuous on the left, in the sense that if Q P
is non-empty and upwards-directed then h( supQ) = sup
BQ
h(B).
(d) Let (A, ) be a probability algebra, and : A A a measure-preserving Boolean homomorphism of
nite entropy. Let P be the set of all -invariant closed subalgebras of A. Show that B h(B) : P
[0, [ is order-continuous. (Hint: if Q P is non-empty and downwards-directed, then for any partition of
unity A A, H(A[ inf Q) = sup
BQ
H(A[B).)
528 Automorphism groups 385 Notes
385 Notes and comments I have taken the trouble to give sharp forms of the Halmos-Rokhlin-Kakutani
lemma (385E) and the Cziszar-Kullback inequality (385J); while it is possible to get through the principal
results of the next two sections with rather less, the formulae become better focused if we have the exact
expressions available. Of course one can always go farther still (385Yb). Ornsteins theorem in 386 (though
not Sinas, as stated there) can be deduced from the special case of the Shannon-McMillan-Breiman theorem
(385G) in which the homomorphism is a Bernoulli shift, which can be deduced from the Ergodic Theorem
(385Xc).
Lemma 385D is the starting point of the theory of recurrence; the next steps are in 382Yc-382Yd and
387E-387F.
386 Ornsteins theorem
I come now to the most important of the handful of theorems known which enable us to describe auto-
morphisms of measure algebras up to isomorphism: two two-sided Bernoulli shifts (on algebras of countable
Maharam type) of the same entropy are isomorphic (386I, 386K). This is hard work. It requires both delicate
- analysis and substantial skill with the manipulation of measure-preserving homomorphisms. The proof
is based on two dicult lemmas (386C and 386F), and includes Sinas theorem (386E, 386L), describing
the Bernoulli shifts which arise as factors of a given ergodic automorphism.
386A The following denitions oer a language in which to express the ideas of this section.
Denitions Let (A, ) be a probability algebra and : A A a measure-preserving Boolean homomor-
phism.
(a) A Bernoulli partition for is a partition of unity a
i
)
iI
such that
(inf
jk

j
a
i(j)
) =

k
j=0
a
i(j)
whenever i(0), . . . , i(k) I.
(b) If is an automorphism, a Bernoulli partition a
i
)
iI
for is (two-sidedly) generating if the closed
subalgebra generated by
j
a
i
: i I, j Z is A itself.
(c) A factor of (A, , ) is a triple (B, B, B) where Bis a closed subalgebra of A such that [B] = B.
386B Remarks Let (A, ) be a probability algebra, : A A a measure-preserving Boolean homo-
morphism and a
i
)
iI
a Bernoulli partition for .
(a)
k
[A
0
])
kN
is independent, where A
0
is the closed subalgebra of A generated by a
i
: i I. PPP
Suppose that c
j

j
[A
0
] for j k. Then each
j
c
j
A
0
is expressible as sup
iI
j
a
i
for some I
j
I. Now
(inf
jk
c
j
) = ( sup
i
0
I
0
,... ,i
k
I
k
inf
jk

j
a
i
j
)
=

i
0
I
0
,... ,i
k
I
k
(inf
jk

j
a
i
j
) =

i
0
I
0
,... ,i
k
I
k
k

j=0
a
i
j
=
k

j=0

iI
j
a
i
=
k

j=0
(sup
iI
j
a
i
) =
k

j=0
c
j
.
As c
0
, . . . , c
k
are arbitrary,
k
[A
0
])
kN
is independent. QQQ
(b) If is an automorphism, then
k
[A
0
])
kZ
is independent, by 384Sf.
(c) Setting A = a
i
: i I 0, we have h(, A) = H(A), as in part (a) of the proof of 384R, so
h() H(A).
386C Ornsteins theorem 529
(d) If H(A) > 0, then A is atomless. PPP As A contains at least two elements of non-zero measure,
= max
aA
a < 1. Because a
i
)
iI
is a Bernoulli partition, every member of D
k
(A, ) has measure at
most
k
, for any k N. Thus any atom of A could have measure at most inf
kN

k
= 0. QQQ
(e) If B is any closed subalgebra of A such that [B] B, then h(B) h(), just because h(B) is
calculated from the action of on a smaller set of partitions. If C
+
is the closed subalgebra of A generated
by
j
a
i
: i I, j N, then [C
+
] C
+
(compare 385Pb), and C
+
is a one-sided Bernoulli shift with
root algebra A
0
and entropy H(A), so that
H(A) = h(C
+
)
by the Kolmogorov-Sina theorem (384P, 384R).
(f ) If is an automorphism, and C is the closed subalgebra of A generated by
j
a
i
: i I, j Z, then
[C] = C (385Pb) and C is a two-sided Bernoulli shift with root algebra A
0
.
(g) Thus every Bernoulli partition for gives rise to a factor of (A, , ) which is a one-sided Bernoulli
shift, and if is an automorphism we can extend this to the corresponding two-sided Bernoulli shift. If
has a generating Bernoulli partition then it is itself a Bernoulli shift.
(h) Now suppose that (B, ) is another probability algebra, : B B is a measure-preserving Boolean
homomorphism, and b
i
)
iI
is a Bernoulli partition for such that b
i
= a
i
for every i. We have a unique
measure-preserving Boolean homomorphism
+
: C
+
B such that
+
(
j
a
i
) =
j
b
i
for every i I, j N.
(Apply 324P.) Now
+
=
+
. (The set a :
+
a =
+
a is a closed subalgebra of C
+
containing every

j
a
i
.)
(i) If, in (g) above, and are both automorphisms, then the same arguments show that we have a
unique measure-preserving Boolean homomorphism : C B such that a
i
= b
i
for every i I and
= .
386C Lemma Let (A, ) be an atomless probability algebra and : A A an ergodic measure-
preserving automorphism. Let a
i
)
iN
be a partition of unity in A, of nite entropy, and
i
)
iN
a sequence
of non-negative real numbers such that

i=0

i
= 1,

i=0
q(
i
) h(),
where q is the function of 384A. Then for any > 0 we can nd a partition a
t
i
)
iN
of unity in A such that
(i) i : a
t
i
,= 0 is nite,
(ii)

i=0
[
i
a
t
i
[ ,
(iii)

i=0
(a
t
i
.a
i
) + 6
_

i=0
[ a
i

i
[ +
_
2(H(A) h(, A))
where A = a
i
: i N 0,
(iv) H(A
t
) h(, A
t
) +
where A
t
= a
t
i
: i N 0.
proof (a) Of course h(, A) H(A), by 384Ma, so the square root
_
2(H(A) h(, A)) gives no diculty.
Set =
_

i=0
[ a
i

i
[ +
_
2(H(A) h(, A)), = min(
1
4
,
1
24
).
There is a sequence
i
)
iN
of non-negative real numbers such that i :
i
> 0 is nite,

i=0

i
= 1,

i=0
[
i

i
[ 2
2
and

i=0
q(
i
) h(). PPP Take k N such that

i=k

i

2
, and set
i
=
i
for i < k,

k
=

i=k

i
and
i
= 0 for i > k; then q(
k
)

i=k
q(
i
) (384Ab), so

i=0
q(
i
)

i=0
q(
i
) h(),
while

i=0
[
i

i
[
k
+

i=k

i
2
2
. QQQ
Because

i=0
q(
i
) is nite, there is a partition of unity C in A, of nite entropy, such that

i=0
q(
i
)
h(, C) + 3; replacing C by C A if need be (note that C A still has nite entropy, by 384H), we may
suppose that C renes A.
530 Automorphism groups 386C
There is a sequence
t
i
)
iN
of non-negative real numbers such that

i=0

t
i
= 1, i :
t
i
> 0 is nite,

i=0
[
t
i

i
[ 4
2
and

i=0
q(
t
i
) = h(, C) + 3.
PPP Take k N such that
i
= 0 for i > k. Take r 1 such that
2
ln(
r

2
) h(, C) + 3 and set

i
= (1
2
)
i
for i k,
=
1
r

2
for k + 1 i k +r,
= 0 for i > k +r.
Then

i=0
[
i

i
[ = 2
2
,

i=0
[
i

i
[ 4
2
,

k+r
i=0
q(
i
) h(, C) + 3
2
ln(
r

2
) = rq(

2
r
)

k+r
i=0
q(
i
).
Now the function

k+r
i=0
q(
i
+ (1 )
i
) : [0, 1] R
is continuous, so there is some [0, 1] such that

k+r
i=0
q(
i
+ (1 )
i
) = h(, C) + 3,
and we can set
t
i
=
i
+ (1 )
i
for every i; of course

i=0
[
t
i

i
[

i=0
[
i

i
[ + (1 )

i=0
[
i

i
[ 4
2
. QQQ
Set M = i :
t
i
,= 0, so that M is nite.
(b) Let ]0, ] be so small that
(i) [q(s) q(t)[

1+#(M)
whenever s, t [0, 1] and [s t[ 3,
(ii)

cC
q(min( c, 2)) ,
(iii)
1
6
.
(Actually, (iii) is a consequence of (i). For (ii) we must of course rely on the fact that

cC
q( c) is nite.)
Let be the probability measure on M dened by saying that i =
t
i
for every i M, and the
product measure on M
N
. Dene X
ij
: M
N
0, 1, for i M and j N, and Y
j
: M
N
R, for j N, by
setting
X
ij
() = 1 if (j) = i,
= 0 otherwise,
Y
j
() = ln(
t
(j)
) for every M
N
.
Then, for each i M, X
ij
)
jN
is an independent sequence of random variables, all with expectation
t
i
,
and Y
j
)
jN
is also an independent sequence of random variables, all with expectation

iM

t
i
ln
t
i
=

i=0
q(
t
i
) = h(, C) 3.
Let n 1 be so large that
(iv) [[w
n
h(, C)1 ]] < , where
w
n
=
1
n

dD
n
(C,)
ln(
1
d
)d;
(v)
Pr
_
iM
[
1
n

n1
j=0
X
ij

t
i
[
_
1 ,
386C Ornsteins theorem 531
Pr
_
[
1
n

n1
j=0
Y
j
+h(, C) + 3[
_
1 ;
(vi) e
n
2,
1
n+1
, q(
1
n+1
) +q(
n
n+1
) ;
these will be true for all suciently large n, using the Shannon-McMillan-Breiman theorem (385H) for (iv)
and the strong law of large numbers (in any of the forms 273D, 273H or 273I) for (v).
(c) There is a family b
ji
)
j<n,iM
such that
() for each j < n, b
ji
)
iM
is a partition of unity in A,
() (inf
j<n
b
j,i(j)
) =

n1
j=0

t
i(j)
for every i(0), . . . , i(n 1) M,
()

iM
(b
ji

j
a
i
) 1
2
4
2
for every j < n.
PPP Construct b
ji
)
iM
for j = n 1, n 2, . . . , 0, as follows. Given b
ji
, for k < j < n, such that
(inf
jk

j
a
i(j)
inf
k<j<n
b
j,i(j)
) = (inf
jk

j
a
i(j)
)

n1
j=k+1

t
i(j)
for every i(0), . . . , i(n1) M (of course this hypothesis is trivial for k = n1), let B
k
be the set of atoms
of the (nite) subalgebra of A generated by b
ji
: i M, k < j < n. Then (b d) = b d for every
b B
k
, d D
k+1
(A, ).
Now

i=0

cD
k
(A,)
[ (
k
a
i
c)
t
i
c[

i=0

cD
k
(A,)
[ (
k
a
i
c) a
i
c[ +

i=0
[ a
i

t
i
[

cD
k
(A,)
c

i=0
[
i

t
i
[ +

i=0
[ a
i

i
[ +

i=0

cD
k
(A,)
[ (
k
a
i
c) a
i
c[
4
2
+

i=0
[ a
i

i
[ +
_
2
_
H(
k
[A]) +H(D
k
(A, )) H(D
k+1
(A, ))
_
(by 385K, because D
k+1
(A, ) =
k
[A] D
k
(A, ))
4
2
+

i=0
[ a
i

i
[ +
_
2(H(A) h(, A))
(because h(, A) H(D
k+1
(A, )) H(D
k
(A, )), by 385Nc)
=
2
+ 4
2
.
Choose a partition of unity b
ki
)
iM
such that, for each c D
k
(A, ), b B
k
, i M,
(b
ki
b c) =
t
i
(b c),
if (
k
a
i
b c)
t
i
(b c) then b
ki
b c
k
a
i
,
if (
k
a
i
b c)
t
i
(b c) then
k
a
i
b c b
ki
.
(This is where I use the hypothesis that A is atomless.) Note that in these formulae we always have
(b c) = b c, (
k
a
i
b c) = (
k
a
i
c) b.
Consequently
532 Automorphism groups 386C

iM
(
k
a
i
b
ki
) =

bB
k

cD
k
(A,)

iM
(b c (
k
a
i
b
ki
))
=

bB
k

cD
k
(A,)

i=0
min( (b c
k
a
i
),
t
i
(b c))

bB
k

cD
k
(A,)

i=0
(b c
k
a
i
) [ (b c
k
a
i
)
t
i
(b c)[
= 1

bB
k

cD
k
(A,)

i=0
[ (b c
k
a
i
)
t
i
(b c)[
= 1

bB
k

cD
k
(A,)

i=0
b [ (c
k
a
i
)
t
i
c[
= 1

cD
k
(A,)

i=0
[ (c
k
a
i
)
t
i
c[ 1
2
4
2
.
Also we have
(b
ki
b c) =
t
i
b c
for every b B
k
, c D
k
(A, ) and i M, so the (downwards) induction proceeds. QQQ
(d) Let B be the set of atoms of the algebra generated by b
ji
: j < n, i M. For b B, d D
n
(C, )
set
I
bd
= j : j < n, i M, b b
ji
, d
j
a
i
.
Then, for any j < n,
supb d : b B, d D
n
(C, ), j I
bd
= sup
iM
b
ji

j
a
i
,
because C renes A, so every
j
a
i
is a supremum of members of D
n
(C, ). Accordingly

bB,dD
n
(C,)
#(I
bd
) (b d) =
n1

j=0

iM
(b
ji

j
a
i
) n(1
2
4
2
).
Set
e
0
= supb d : b B, d D
n
(C, ), #(I
bd
) n(1 4);
then e
0
1 .
(e) Let B
t
B be the set of those b B such that
b e
n(h(,C)+2)
,

iM
[
t
i

1
n
#(j : j < n, b b
ji
)[ .
Then (supB
t
) 1 2. PPP Set
B
t
1
= b : b B, b e
n(h(,C)+2)

= b : b B, h(, C) + 2 +
1
n
ln( b) 0
= inf
j<n
b
i,i(j)
: i(0), . . . , i(n 1) M, h(, C) + 2 +
1
n
n1

j=0
ln
t
i(j)
0.
Then
386C Ornsteins theorem 533
(supB
t
1
) = Pr(h(, C) + 2 +
1
n
n1

j=0
Y
j
0)
Pr([h(, C) + 3 +
1
n
n1

j=0
Y
j
[ ) 1
by the choice of n. On the other hand, setting
B
t
2
= b : b B,

iM
[
t
i

1
n
#(j : j < n, b b
ji
)[
= inf
j<n
b
i,i(j)
: i(0), . . . , i(n 1) M,

iM
[
t
i

1
n
#(j : i(j) = i)[ ,
we have
(supB
t
2
) = Pr(

iM
[
t
i

1
n

n1
j=0
X
ij
[ ) 1
by the other half of clause (b-v). Since B
t
= B
t
1
B
t
2
, (supB) 1 2. QQQ
Let D
t
0
be the set of those d D
n
(C, ) such that
1
n
ln(
1
d
) h(, C) +, i.e., d e
n(h(,C)+)
;
by (b-iv), (supD
t
0
) > 1 . Let D
t
D
t
0
be a nite set such that (supD
t
) 1 . If d D
t
, b B
t
then
d e
n(h(,C)+)
e
n
b 2 b.
Since (supD
t
) 1 2 (supB
t
) (remember that
1
4
, #(D
t
) #(B
t
).
Set e
1
= e
0
supB
t
, so that e
1
1 3, and
D
tt
= d : d D
t
, (d e
1
)
1
2
d;
then
(sup(D
t
D
tt
)) 2 (1 \ e
1
) 2 + 6,
so
(supD
tt
) 1 2 6 1 2 7.
(f ) If d
1
, . . . , d
k
D
tt
are distinct,
(sup
1ik
d
i
e
1
)
k
2
inf
ik
d
i
k sup
bB
b,
and
#(b : b B
t
, b e
0
sup
1ik
d
i
,= 0) k.
By the Marriage Lemma (3A1K), there is an injective function f
0
: D
tt
B
t
such that d f
0
(d) e
0
,= 0
for every d D
tt
. Because #(D
t
) #(B
t
), we can extend f
0
to an injective function f : D
t
B
t
.
(g) By the Halmos-Rokhlin-Kakutani lemma, in the strong form 385E(iv), there is an a A such that
a,
1
a, . . . ,
n+1
a are disjoint and (a d) =
1
n+1
d for every d D
t
1. Set e = sup
j
(a d) : j <
n, d D
t
. Because
j
(a d))
j<n,dD
is disjoint,
e =

n1
j=0

dD

(a d) =
n
n+1

dD

d (1 )
2
1 2.
(h) For i M, set
a
t
i
= sup
j
(a d) : j < n, d D
t
, f(d) b
ji
.
534 Automorphism groups 386C
Then the a
t
i
are disjoint. PPP Suppose that i, i
t
M are distinct. If j, j
t
< n and d, d
t
D
t
and f(d) b
ji
,
f(d
t
) b
j

i
, then either j ,= j
t
or j = j
t
. In the former case,

j
(a d)
j

(a d
t
)
j
a
j

a = 0.
In the latter case, b
ji
b
j

i
= 0, so f(d) ,= f(d
t
) and d ,= d
t
and

j
(a d)
j

(a d
t
)
j
(d d
t
) = 0. QQQ
Observe that
sup
iM
a
t
i
= sup
j<n,dD

j
(a d) = e
because if j < n and d D
t
there must be some i M such that f(d) b
ji
. Take any m N M and set
a
t
m
= 1 \ e, a
t
i
= 0 for i N (M m); then a
t
i
)
iN
is a partition of unity. Now

iM
[ a
t
i

t
i
[

iM

t
i
[1 n (a supD
t
)[ +

iM
[ a
t
i
n
t
i
(a supD
t
)[
1
n
n+1
(supD
t
)
+

iM
[
n1

j=0

dD

f(d) b
ji
(
j
(a d)) n
t
i

dD

(a d)[
1 (1 )
2
+

dD

iM
[ (a d) #(j : j < n, f(d) b
ji
) n
t
i
(a d)[
1 (1 )
2
+

dD

(a d)n
(see the denition of B
t
in (d) above)
2 +n a 3.
So

i=0
[ a
t
i

i
[ a
t
m
+

iM
[ a
t
i

t
i
[ +

i=0
[
t
i

i
[
2 + 3 + 4
2
6 .
We shall later want to know that [ a
t
i

t
i
[ 3 for every i; for i M this is covered by the formulae
above, for i = m it is true because a
t
m
= 1 e 2 (see (g)), and for other i it is trivial.
(i) The next step is to show that

i=0
(a
t
i
a
i
) 1 3 12. PPP It is enough to consider the case in
which 3 + 12 < 1. We know that
sup
iN
a
t
i
a
i
sup
j
(a d) : j < n, d D
t
,
i M such that f(d) b
ji
and d
j
a
i

= sup
j
(a d) : d D
t
, j I
f(d),d

(see (d) for the denition of I

bd
) has measure at least

dD

#(I
f(d),d
) (a d).
For any d D
tt
, we arranged that d f(d) e
0
,= 0. This means that there must be some b B and
d
t
D
n
(C, ) such that d f(d) b f(d
t
) ,= 0 and #(I
bd
) n(1 4); of course b = f(d) and d
t
= d,
so that #(I
f(d),d
) must be at least n(1 4). Accordingly

i=0
(a
t
i
a
i
)

dD

n(1 4) (a d) = n(1 4)
1
n+1
(supD
tt
)
(1 )(1 4)(1 2 7) 1 3 12. QQQ
386C Ornsteins theorem 535
But this means that

i=0
(a
t
i
.a
i
) = 2(1

i=0
(a
t
i
a
i
)) 6 + 24 + 6
(using 385L for the equality).
(j) Finally, we need to estimate H(A
t
) and h(, A
t
), where A
t
= a
t
i
: i N 0. For the former, we
have H(A
t
) h(, C) + 4. PPP [ a
t
i

t
i
[ 3 for every i, by (h) above. So by (b-i),
H(A
t
) =

iMm]
q( a
t
i
) +

i=0
q(
t
i
) = h(, C) + 4. QQQ
(k) Consider the partition of unity
A
tt
= A
t
a, 1 \ a.
Let D be the closed subalgebra of A generated by
j
c : j Z, c A
tt
.
(i) a d D for every d D
t
. PPP Of course a e D, because 1 \ e = a
t
m
. If d
t
D
t
and d
t
,= d, then
(because f is injective) f(d) ,= f(d
t
); there must therefore be some k < n and distinct i, i
t
M such that
f(d) b
ki
and f(d
t
) b
ki
. But this means that
k
(a d) a
t
i
and
k
(a d
t
) a
t
i
, so that a d
k
a
t
i
and a d
t

k
a
t
i
= 0.
What this means is that if we set

d = a e inf
k
a
t
i
: k < n, i M, a d
k
a
t
i
,
we get a member of D (because every a
t
i
D, and [D] = D) including a d and disjoint from a d
t
whenever d
t
D
t
and d
t
,= d. But as a
j
a = 0 if 0 < j < n, a e must be supa d
t
: d
t
D
t
, and
a d =

d belongs to D. QQQ
(ii) Consequently c e D for every c C. PPP We have
c e = supc
j
(a d) : j < n, d D
t

= sup
j
(
j
c a d) : j < n, d D
t

= sup
j
(a d) : j < n, d D
t
, d
j
c
(because if d D
t
and j < n then either d
j
c or d
j
c = 0)
D
because a d D for every d D
t
and
1
[D] = D. QQQ
(iii) It follows that h(, A
tt
) h(, C) . PPP For any c C,
(c, D) (c .(c e)) = (c \ e) min( c, 2)
1
3
.
So
h(, C) h(D) +H(C[D)
(385Nd, because [D] = D)
h(, A
tt
) +

cC
q((c, D))
(by the Kolmogorov-Sina theorem (384P) and 385Ob)
h(, A
tt
) +

cC
q(min( c, 2))
(because q is monotonic on [0,
1
3
])
h(, A
tt
) +
by the choice of . QQQ
(iv) Finally, h(, A
t
) h(, C) 2. PPP Using 385Nb,
536 Automorphism groups 386C
h(, C) h(, A
tt
) h(, A
t
) +H(a, 1 \ a)
= h(, A
t
) +q( a) +q(1 a)
= h(, A
t
) +q(
1
n+1
) +q(
n
n+1
) h(, A
t
) +
by the choice of n. QQQ
(l) Putting these together,
H(A
t
) h(, C) + 4 h(, A
t
) + 6 h(, A
t
) +,
and the proof is complete.
386D Corollary Let (A, ) be an atomless probability algebra and : A A an ergodic measure-
preserving automorphism. Let a
i
)
iN
be a partition of unity in A, of nite entropy, and
i
)
iN
a sequence
of non-negative real numbers such that

i=0

i
= 1,

i=0
q(
i
) h().
Then for any > 0 we can nd a Bernoulli partition a

i
)
iN
for such that a

i
=
i
for every i N and

i=0
(a

i
.a
i
) + 6
_

i=0
[ a
i

i
[ +
_
2(H(A) h(, A)),
writing A = a
i
: i N 0.
proof (a) Set =
_

i=0
[ a
i

i
[ +
_
2(H(A) h(, A)). Let
n
)
nN
be a sequence of strictly positive
real numbers such that

n=0

n
+ 6
_

n
+

2
n
.
Using 386C, we can choose inductively, for n N, partitions of unity a
ni
)
iN
such that, for each n N,

i=0
[
i
a
ni
[
n
,
H(A
n
) h(, A
n
) +
n
<
(writing A
n
= a
ni
: i N 0),

i=0
(a
n+1,i
.a
ni
)
n+1
+ 6
_

n
+

2
n
,
while

i=0
(a
0i
.a
i
)
0
+ 6.
On completing the induction, we see that

n=0

i=0
(a
n+1,i
.a
ni
)

n=1

n
+

n=0
6
_

n
+

2
n
< .
In particular, given i N,

n=0
(a
n+1,i
.a
ni
) is nite, so a
ni
)
nN
is a Cauchy sequence in the complete
metric space A (323Gc), and has a limit a

i
, with
a

i
= lim
n
a
ni
=
i
(323C). If i ,= j,
a

i
a

j
= lim
n
a
ni
a
nj
= 0
(using 323B), so a

i
)
iN
is disjoint; since

i=0
a

i
=

i=0

i
= 1,
a

i
)
iN
is a partition of unity. We also have
386F Ornsteins theorem 537

i=0
(a

i
.a
i
)

i=0
(a
0i
.a
i
) +

n=0

i=0
(a
n+1,i
.a
ni
)

0
+ 6 +

n=1

n
+

n=0
6
_

n
+

2
n
+ 6.
(b) Now take any i(0), . . . , i(k) N. For each j < k, n N,
H(
j
[A
n
]) +H(D
j
(A
n
, )) H(D
j+1
(A
n
, )) H(A
n
) h(, A
n
)
n
(using 385Nc). But this means that

dD
j
(A
n
,)

i=0
[ (d
j
a
ni
) d a
ni
[

2
n
,
by 385K. A fortiori,
[ (d
j
a
ni
) d a
ni
[

2
n
for every d D
j
(A
n
, ), i N. Inducing on r, we see that
[ (inf
jr

j
a
n,i(j)
)

r
j=0
a
n,i(j)
[ r

2
n
0
as n , for any r k. Because , and are all continuous (323C, 323B, 324Kb),
(inf
jk

j
a

i(j)
) = lim
n
(inf
jk

j
a
n,i(j)
)
= lim
n

jk
a
n,i(j)
=

jk

i(j)
.
As i(0), . . . , i(k) are arbitrary, a

i
)
iN
is a Bernoulli partition for .
386E Sinas theorem (atomic case) (Sina 62) Let (A, ) be an atomless probability algebra and
: A A an ergodic measure-preserving automorphism. Let
i
)
iN
be a sequence of non-negative real
numbers such that

i=0

i
= 1 and

i=0
q(
i
) h(). Then there is a Bernoulli partition a

i
)
iN
for
such that a

i
=
i
for every i N.
proof Apply 386D from any starting point, e.g., a
0
= 1, a
i
= 0 for i > 0.
386F Lemma Let (A, ) be an atomless probability algebra and a measure-preserving automorphism
of A. Let b
i
)
iN
, c
i
)
iN
be Bernoulli partitions for , of the same nite entropy, and write B, C for the
closed subalgebras of A generated by
j
b
i
: i N, j Z and
j
c
i
: i N, j Z. Suppose that C B.
Then for any > 0 we can nd a Bernoulli partition d
i
)
iN
for such that
(i) d
i
C for every i N,
(ii) d
i
= b
i
for every i N,
(iii) (c
i
.c
i
) for every i N,
where : B C is the measure-preserving Boolean homomorphism such that b
i
= d
i
for every i and
= (386Bi).
proof (a) Set B = b
i
: i N 0, C = c
i
: i N 0. If only one c
i
is non-zero, then H(C) = 0, so
H(B) = 0 and B = 0, 1, in which case B = C and we take d
i
= b
i
and stop. Otherwise, C is atomless
(386Bd).
For k N, let B
k
B be the closed subalgebra of A generated by
j
b
i
: i k, [j[ k. Because
C B, there is an m N such that
(c
i
, B
m
)
1
4
for every i N
(385M). Let , > 0 be such that
538 Automorphism groups 386F
+ 6
4

2

4(2m+1)
, min(

4
,
1
6
),

i=0
q(min(2, c
i
)) .
(The last is achievable because

i=0
q( c
i
) is nite.) Let r m be such that
(c
i
, B
r
) for every i N.
Let n r be such that
2r+1
2n+2
, c
i
for every i > n.
(b) Let b
t
i
)
iN
be a partition of unity in C such that b
t
i
= b
i
for every i N. Let U be the set of
atoms of the subalgebra of B generated by
j
b
i
: i n, [j[ n
j
c
i
: i n, [j[ n, and V the set of
atoms of the subalgebra of C generated by
j
b
t
i
: i n, [j[ n
j
c
i
: i n, [j[ n. For each v V ,
choose a disjoint family d
vu
)
uU
in C such that sup
uU
d
vu
= v and d
vu
= (v u) for every u U. By
385E(iv), there is an a C such that a, a, . . . ,
2n
a are disjoint and (a d
vu
) =
1
2n+2
(d
vu
) for every
u U and v V . (C is a Bernoulli shift, therefore ergodic, by 384Se, therefore aperiodic, by 385F.) Set
e = sup
]j]n

j
a, e = sup
]j]nr

j
a; then
e = (2(n r) + 1) a = 1
2r+1
2n+2
.
Let C
e
be the principal ideal of C generated by e.
(c) The family
j
(a d
vu
))
]j]n,uU,vV
is disjoint. PPP All we have to note is that the families
d
vu
)
uU,vV
and

j
a)
]j]n
=
n
(
n+j
a))
]j]n
are disjoint. QQQ Consequently, if we set

b
i
= sup
]j]n
sup
vV
sup
uU,u
j
b
i

j
(a d
vu
) C
for i N,

b
i
)
iN
is disjoint, since a given triple (j, u, v) can contribute to at most one

b
i
.
Of course

b
i
sup
]j]n

j
a
i
= e for every i. If i n, we also have

b
i
= e b
i
. PPP For [j[ n,
j
b
i
is
a union of members of U, so

b
i
=
n

j=n

vV

uU,u
j
b
i
(
j
(a d
vu
))
(because
j
(a d
vu
))
]j]n,uU,vV
is disjoint)
=
n

j=n

vV

uU,u
j
b
i
(a d
vu
) =
1
2n+2
n

j=n

vV

uU,u
j
b
i
d
vu
(by the choice of a)
=
1
2n+2
n

j=n

vV

uU,u
j
b
i
(v u)
(by the choice of d
vu
)
=
1
2n+2
n

j=n

uU,u
j
b
i
u =
1
2n+2
n

j=n
(
j
b
i
)
(because
j
b
i
is a disjoint union of members of U when i n, [j[ n)
=
2n+1
2n+2
b
i
= e b
i
. QQQ
Again because C is atomless, we can choose a partition of unity b

i
)
iN
C such that b

i
= b
i
for every i,
while b

b
i
and b

i
e =

b
i
for i n.
386F Ornsteins theorem 539
(d) Let E be the nite subalgebra of B generated by
j
b
i
: i n, [j[ r
j
c
i
: i n, [j[ r.
Dene : E C
e
by setting
b = sup
]j]nr
sup
vV
sup
uU,u
j
b

j
(a d
vu
)
for b E.
(i) is a Boolean homomorphism. PPP The point is that if [j[ n r and b E, then
j
b belongs to
the algebra generated by
k
b
i
: i n, [k[ n
k
c
i
: i n, [k[ n, so is a union of members of U.
Since each map
b
j
(a d
vu
) if u
j
b, 0 otherwise
is a Boolean homomorphism from E to the principal ideal generated by
j
(a d
vu
), and

j
(a d
vu
))
]j]nr,uU,vV
is a partition of unity in C
e
, also is a Boolean homomorphism. QQQ
(ii) (b) b for every b E. PPP (Compare (c) above.)
(b) =
nr

j=n+r

vV

uU,u
j
b

j
(a d
vu
)
=
1
2n+2
nr

j=n+r

vV

uU,u
j
b
(v u) =
2n2r+1
2n+2
b b. QQQ
(iii) (
k
b
i
) = e
k
b

i
for i n, [k[ r. PPP Of course
k
b
i
E. If [j[ n r, then [j +k[ n, so

j
a (
k
b
i
) = sup
vV
sup
uU,u
j+k
b
i

j
(a d
vu
)
=
k
_
sup
vV
sup
uU,u
j+k
b
i

jk
(a d
vu
)
_
=
k
(
jk
a

b
i
) =
j
a
k
(e b

i
) =
j
a
k
b

i
because
j
a
k
e. Taking the supremum of these pieces we have
(
k
b
i
) = sup
]j]nr

j
a (
k
b
i
) = sup
]j]nr

j
a
k
b

i
= e
k
b

i
. QQQ
It follows that
(
k
(1 \ sup
il
b
i
)) = e
k
(1 \ sup
il
b

i
))
if l n, [k[ r.
(iv) Finally, c
i
= c
i
e for every i n. PPP If [j[ n r, v V then either v
j
c
i
or v
j
c
i
= 0.
In the former case,
d
vu
= v u = 0 whenever u U and u ,
j
c
i
,
so that
v = sup
uU
d
vu
= sup
uU,u
j
c
i
d
vu
;
in the latter case, d
vu
= v u = 0 whenever u
j
c
i
. So we have
v
j
c
i
= sup
uU,u
j
c
i
d
vu
for every v V , and
540 Automorphism groups 386F
c
i
= sup
]j]nr
sup
vV
sup
uU,u
j
c
i

j
(a d
vu
)
= sup
]j]nr

j
(a sup
vV
sup
uU,u
j
c
i
d
vu
)
= sup
]j]nr

j
(a sup
vV
(v
j
c
i
))
= sup
]j]nr

j
(a
j
c
i
) = c
i
sup
]j]nr

j
a = c
i
e. QQQ
(e) Let B

be the closed subalgebra of A generated by

j
b

i
: i N, [j[ Z. Then for every b B
r
there is a b

such that b = b

e. PPP The set of b for which this is true is a subalgebra of A containing

k
b
i
for i r and [k[ r, by (d-iii). QQQ It follows that
(c
i
, B

) 2 for i N.
PPP If i > n this is trivial, because c
i
, by the choice of n. Otherwise, c
i
E. Take b B
r
such that
(b .c
i
) = (c
i
, B
r
) (385Oa). Let b

be such that b = b

e. Then
(c
i
, B

k
) (c
i
.b

) 1 e + ( e (c
i
.b

))
=
2r+1
2n+2
+ (( e c
i
) .b) =
2r+1
2n+2
+ (c
i
.b)
(by (d-iv))
=
2r+1
2n+2
+ ((c
i
.b))
2r+1
2n+2
+ (c
i
.b)
(by (d-ii))
2
by the choice of n. QQQ
(f ) Set B

= b

i
: i N 0. Then H(B

) = h(, C) h(, B

) +. PPP
H(B

) = H(B) = H(C)
(because b

i
= b
i
for every i, and we supposed from the beginning that H(C) = H(B))
= h(, C)
(because C is a Bernoulli partition, see 386Bc)
h(B

r
) +H(C[B

r
)
(385Nd)
h(B

) +

i=0
q((c
i
, B

r
))
(by the denition of h(B

), and 385Ob)
h(, B

) +

i=0
q(min(2, c
i
))
(by the Kolmogorov-Sina theorem, 384P(ii), and (e) above, recalling that
1
6
, so that q is monotonic on
[0, 2])
h(, B

) +
by the choice of . QQQ
386G Ornsteins theorem 541
(g) By 386D, applied to C and the partition b

i
)
iN
of unity in C and the sequence
i
)
iN
= b

i
)
iN
,
we have a Bernoulli partition d
i
)
iN
in C such that d
i
= b

i
= b
i
for every i N and

i=0
(d
i
.b

i
) + 6
4

2

4(2m+1)
.
Let D C be the closed subalgebra of A generated by
j
d
i
: i N, j Z. Then (B, B, b
i
)
iN
) is
isomorphic to (D, D, d
i
)
iN
), with an isomorphism : B D such that = and b
i
= d
i
for every
i N (386Bi).
(h) Set
e

= e \ sup
]j]m,iN

j
(d
i
.b

i
).
Then (
j
b
i
) e

= (
j
b
i
) e

whenever i m and [j[ m. PPP

(
j
b
i
) e

=
j
(b
i
) e

=
j
d
i
e

=
j
b

i
e

=
j
b

i
e e

= (
j
b
i
) e

by (d-iii), because i, [j[ r n. QQQ Since b b e

: A A
e
, b b e

: E A
e
are Boolean
homomorphisms, b e

= b e

for every b B
m
.
Now (c
i
.c
i
) for every i N. PPP If i > n then of course
(c
i
.c
i
) 2 c
i
2 .
If i n, then (by the choice of m) there is a b B
m
such that (c
i
, b)
1
4
. So
c
i
.c
i
(c
i
.b) (b .b) (b .c
i
) (c
i
.c
i
)
(c
i
.b) (1 \ e

) (b .c
i
)
(using the denition of e

and (d-iv)) has measure at most

(c
i
.b) + (1 \ e

) + (b .c
i
)
(by (d-ii), since b and c both belong to E)
2 (c
i
.b) + (1 \ e) + (2m+ 1)

i=0
(d
i
.b

i
)

2
+
2r+1
2n+2
+

4
,
as required. QQQ
386G Lemma Let (A, ) be an atomless probability algebra and a measure-preserving automorphism
of A. Let b
i
)
iN
, c
i
)
iN
be Bernoulli partitions for , of the same nite entropy, and write B, C for the
closed subalgebras generated by
j
b
i
: i N, j Z and
j
c
i
: i N, j Z. Suppose that C B. Then
for any > 0 we can nd a Bernoulli partition d
i
)
iN
for such that
(i) d
i
= c
i
for every i N,
(ii) (d
i
.c
i
) for every i N,
(iii) writing D for the closed subalgebra of A generated by
j
d
i
: i N, j Z, (b
i
, D) for every
i N.
proof (a) By 386F, there is a Bernoulli partition b

i
)
iN
for such that b

i
C for every i N, b

i
= b
i
for every i N, and (c
i
.c
i
)
1
4
for every i N, where : B C is the measure-preserving Boolean
homomorphism such that b
i
= b

i
for every i and = . Note that this implies that
1
=
1
,
and generally that
j
=
j
for every j Z; so [B] C is the closed subalgebra of A generated by

j
b
i
: i N, j Z =
j
b

i
: i N, j Z (324L), and is invariant under the action of and
1
.
Let m N be such that
(c
i
, B
m
)
1
4
for every i N,
542 Automorphism groups 386G
where B
m
is the closed subalgebra of A generated by
j
b
i
: i N, [j[ m (385M). Let ]0, ] be such
that
(2m+ 1)

i=0
min(, 2 b
i
)
1
4
.
By 386F again, applied to C, there is a Bernoulli partition c

i
)
iN
for such that c

i
[B], c

i
= c
i
and (b

i
.b

i
) for every i N, where : C C is the measure-preserving Boolean homomorphism
such that c
i
= c

i
for every i N and = . Once again, [C] will be the closed subalgebra of A
generated by
j
c
i
: i N, j Z =
j
c

i
: i N, j Z; because every c

i
belongs to [B], [C] [B].
(b) Now (c

i
.c
i
) for every i N. PPP There is a b B
m
such that (c
i
.b)
1
4
. We know that
[B
m
] is the closed subalgebra of A generated by
j
b
i
: i N, [j[ m =
j
b

i
: i N, [j[ m, and
contains b. Because
(b) .b sup
iN,]j]m
(
j
b

i
) .
j
b

i
= sup
]j]m

j
(sup
iN
b

i
.b

i
),
we have
(b .b) (2m+ 1)

i=0
(b

i
.b

i
)
(2m+ 1)

i=0
min(, 2 b
i
)
1
4
.
But this means that
(c

i
.c
i
) = (c
i
.c
i
) (c
i
.b) + (b .b) + (b .c
i
)
(c
i
.b) +

4
+ (b .c
i
) (c
i
.c
i
) + (c
i
.b) +

2

4
+ (c
i
.b) +

2
. QQQ
(c) Set d
i
=
1
c

i
for each i; this is well-dened because is injective and c

i
[B]. Write D for the
closed subalgebra of A or of B generated by
j
d
i
: i N, j Z =
1

j
c
i
: i N, j Z; then
D =
1
[[C]], by 324L again, because
1
: [B] B is a measure-preserving homomorphism. Then
d
i
= c

i
= c
i
for every i N, and d
i
)
iN
is a Bernoulli partition for . PPP If i(0), . . . , i(n) N, then
( inf
jn

j
d
i(j)
) = ( inf
jn

1
c

i(j)
) = (( inf
jn

1
c

i(j)
))
= ( inf
jn

j
c

i(j)
)) =

jn
c

i(j)
=

jn
d
i(j)
. QQQ
Next,
(c
i
.d
i
) = (c
i
.d
i
) = (c
i
.c

i
)
for every i, by (b). Finally, if i N, then b

i
belongs to [C], while D =
1
[[C]], so
(b
i
, D) = (b
i
, [C]) (b
i
.b

i
) = (b

i
.b

i
) .
This completes the proof.
386H Lemma Let (A, ) be an atomless probability algebra and a measure-preserving automorphism
of A. Let b
i
)
iN
, c
i
)
iN
be Bernoulli partitions for , of the same nite entropy, and write B, C for the
closed subalgebras generated by
j
b
i
: i N, j Z and
j
c
i
: i N, j Z. Suppose that C B. Then
for any > 0 we can nd a Bernoulli partition d
i
)
iN
for such that
(i) d
i
= c
i
for every i N,
(ii) (d
i
.c
i
) for every i N,
(iii) the closed subalgebra of A generated by
j
d
i
: i N, j Z is B.
386H Ornsteins theorem 543
proof (a) To begin with (down to the end of (c) below) suppose that A = B. Choose
n
)
nN
,
n
)
nN
,
r
n
)
nN
and d
ni
)
iN
)
nN
inductively, as follows. Start with d
0i
= c
i
for every i, r
0
= 0. Given that d
ni
)
iN
is a Bernoulli partition with d
ni
= c
i
for every i, take
n
> 0 such that
(2r
m
+ 1)
n
2
n
for every m n,

n
> 0 such that

n
2
n1
,

i=0
min(
n
, 2 c
i
)
n
,
and use 386G to nd a Bernoulli partition d
n+1,i
)
iN
for such that
d
n+1,i
= c
i
, (d
n+1,i
.d
ni
)
n
, (b
i
, D
(n+1)
) 2
n1
for every i N, where D
(n+1)
is the closed subalgebra of A generated by
j
d
n+1,i
: i N, j Z. Let r
n+1
be such that
(b
i
, D
(n+1)
r
n+1
) 2
n
for every i N, where D
(n+1)
r
n+1
is the closed subalgebra of A generated by
j
d
n+1,i
: i N, [j[ r
n+1
.
Continue.
(b) For any i N,

n=0
(d
n+1,i
.d
ni
)

n=0

n
,
so d
ni
)
nN
has a limit d
i
in A. Of course
(c
i
.d
i
)

n=0
(d
n+1,i
.d
ni
)
for every i. We must have
d
i
= lim
n
d
ni
= c
i
for each i, and if i ,= j then
d
i
d
j
= lim
n
d
ni
d
nj
= 0;
since

i=0
c
i
= 1, d
i
)
iN
is a partition of unity in A. For any i(0), . . . , i(k) in N,
(inf
jk

j
d
i(j)
) = lim
n
(inf
jk

j
d
n,i(j)
) = lim
n

k
j=0
d
n,i(j)
=

k
j=0
d
i(j)
,
so d
i
)
iN
is a Bernoulli partition.
(c) Let D be the closed subalgebra of A generated by
j
d
i
: i N, j Z. Then b
j
D for every j N.
PPP Fix m N. Then (b
j
, D
(m+1)
r
m+1
) 2
m
, so there is a b D
(m+1)
r
m+1
such that (b
j
.b) 2
m
. Now

i=0
(d
m+1,i
, D)

i=0
(d
m+1,i
.d
i
)

i=0

k=m+1
(d
k+1,i
.d
ki
)

k=m+1

i=0
min(2 c
i
,
k
)

k=m+1

k
.
So
(b, D) (2r
m+1
+ 1)

i=0
(d
m+1,i
, D)

k=m+1
(2r
m+1
+ 1)
k

k=m+1
2
k
= 2
m
,
and
(b
j
, D) (b
j
.b) +(b, D) 2
m
+ 2
m
= 2 2
m
.
As m is arbitrary, (b
j
, D) = 0 and b
j
D. QQQ
(d) This completes the proof if A = B. For the general case, apply the arguments above to (B, B, B).
544 Automorphism groups 386I
386I Ornsteins theorem (nite entropy case) Let (A, ) and (B, ) be probability algebras, and
: A A, : B B two-sided Bernoulli shifts of the same nite entropy. Then (A, , ) and (B, , )
are isomorphic.
proof (a) Let a
i
)
iN
, b
i
)
iN
be (two-sided) generating Bernoulli partitions in A, B respectively. By the
Kolmogorov-Sina theorem, a
i
)
iN
and b
i
)
iN
both have entropy equal to h() = h(). If this entropy is
zero, then A and B are both 0, 1, and the result is trivial; so let us assume that h() > 0, so that A is
atomless (386Bd).
(b) By Sinas theorem (386E), there is a Bernoulli partition c
i
)
iN
for such that c
i
= b
i
for every
i N. By 386H, there is a Bernoulli partition d
i
)
iN
for such that d
i
= c
i
for every i and the
closed subalgebra of A generated by
j
d
i
: i N, j Z is A. But now (A, , , d
i
)
iN
) is isomorphic to
(B, , , b
i
)
iN
), so (A, , ) and (B, , ) are isomorphic.
386J Using the same methods, we can extend the last result to the case of Bernoulli shifts of innite
entropy. The rst step uses the ideas of 386C, as follows.
Lemma Let (A, ) be a measure algebra and : A A an ergodic measure-preserving automorphism.
Suppose that a
i
)
iI
is a nite Bernoulli partition for , with #(I) = r and a
i
= 1/r for every i I, and
that h() ln2r. Then for any > 0 there is a Bernoulli partition b
ij
)
iI,j0,1]
for such that
(a
i
.(b
i0
b
i1
)) , b
i0
= b
i1
=
1
2r
for every i I.
proof (a) Let > 0 be such that
+ 6

4 .
Let > 0 be such that
< ln2,

8 ,
and
[t
1
2
[ whenever t [0, 1] and q(t) +q(1 t) ln2 4
(384Ad). We have
H(A) = rq(
1
r
) = lnr,
and d = r
n
whenever n N, d D
n
(A, ).
Note that A is atomless. PPP??? If a A is an atom, then (because is ergodic) sup
jN

j
a = 1, and A is
purely atomic, with nitely many atoms all of the same size as a; but this means that H(C) ln(
1
a
) for
every partition of unity C A, so that
h(, C) = lim
n
1
n
H(D
n
(C, ), ) lim
n
1
n
ln(
1
a
) = 0
for every partition of unity C, and
0 = h() ln2r ln2. XXXQQQ
(b) There is a nite partition of unity C A such that
h(, C) = ln2r ,
and C renes A = a
i
: i I 0. PPP Because h() ln2r, there is a nite partition of unity C
t
such
that h(, C
t
) ln2r ; replacing C
t
by C
t
A if need be, we may suppose that C
t
renes A; take such
a C
t
of minimal size. Because H(C
t
) h(, C
t
) > H(A), there must be distinct c
0
, c
1
C
t
included in
the same member of A. Because A is atomless, the principal ideal generated by c
1
has a closed subalgebra
isomorphic, as measure algebra, to the measure algebra of Lebesgue measure on [0, 1], up to a scalar multiple
of the measure; and in particular there is a family d
t
)
t[0,1]
such that d
s
d
t
whenever s t, d
1
= c
1
and
d
t
= t c
1
for every t [0, 1]. Let D
t
be the partition of unity
386J Ornsteins theorem 545
(C
t
c
0
, c
1
) c
0
d
t
, c
1
\ d
t

for each t [0, 1]. Then

h(, D
1
) = h(, (C
t
c
0
, c
1
) c
0
c
1
) < ln2r ,
by the minimality of #(C
t
), while
h(, D
0
) = h(, C
t
) ln2r .
Using 384N, we also have, for any s, t [0, 1] such that [s t[
1
e
,
h(, D
s
) h(, D
t
) H(D
s
[D
t
)
(where D
t
is the closed subalgebra generated by D
t
)
q((c
0
d
s
, D
t
)) +q((c
1
\ d
s
, D
t
))
(by 385Ob, because D
s
D
t
c
0
d
s
, c
1
\ d
s
)
q( ((c
0
d
s
) .(c
0
d
t
))) +q( ((c
1
\ d
s
) .(c
1
\ d
t
)))
= 2q( (d
s
.d
t
)) = 2q([s t[ c
1
)
because q is monotonic on [0, [s t[ c
1
]. But this means that t h(, D
t
) is continuous and there must be
some t such that h(, D
t
) = ln2r ; take C = D
t
. QQQ
(c) Let > 0 be such that
,
1
6
, q(2) +q(1 2) ,

cC
q(min(2, c)) .
Let n N be such that
1
n+1
, q(
1
n+1
) +q(
n
n+1
) ,
[[w
n
h(, C)1 ]] ,
where
w
n
=
1
n

dD
n
(C,)
ln(
1
d
)d.
(The Shannon-McMillan-Breiman theorem, 385H, assures us that any suciently large n has these proper-
ties.)
(d) Let D be the set of those d D
n
(C, ) such that
d (2r)
n
, i.e.,
1
n
ln(
1
d
) ln2r.
Then (supD) 1 , by the choice of n, because h(, C) = ln2r . Note that every member of
D is included in some member of D
n
(A, ), because C renes A. If b D
n
(A, ), then b = r
n
, so
#(d : d D, d b) 2
n
; we can therefore nd a function f : D 0, 1
n
such that f is injective on
d : d D, d b for every b D
n
(A, ).
(e) By 385E(iv), as usual, there is an a A such that a,
1
a, . . . ,
n+1
a are disjoint and (a d) =
1
n+1
d for every d D
n
(C, ). Set
e = sup
dD,j<n

j
(a d);
then
e =

n1
j=0

dD
(a d) =
n
n+1
(supD) (1 )
2
1 2.
(f ) Set
546 Automorphism groups 386J
c

= sup
j
(a d) : j < n, d D, f(d)(j) = 1.
(I am identifying members of 0, 1
n
with functions from 0, . . . , n 1 to 0, 1.) Set
A

= A c

, 1 \ c

, A
t
= A

a, 1 \ a e, 1 \ e,
and let A
t
be the closed subalgebra of A generated by
j
a : a A
t
, j Z. Then a d A
t
for every
d D. PPP Set

d = upr(a d, A
t
) = infc : c A
t
, c a d A
t
.
Let b be the element of D
n
(A, ) including d. Because a, b, e A
t
,

d a b e = sup
d

D
a b d
t
= supa d
t
: d
t
D, d
t
b.
Now if d
t
D, d
t
b and d
t
,= d, then f(d
t
) ,= f(d). Let j be such that f(d
t
)(j) ,= f(d)(j); then
j
(a d)
is included in one of c

, 1 \ c

and
j
(a d
t
) in the other. This means that one of
j
c

, 1 \
j
c

is a member
of A
t
including a d and disjoint from a d
t
, so that

d d
t
= 0. Thus

d must be actually equal to a d, and
a d A
t
. QQQ
Next, c e A
t
for every c C. PPP
j
(a d))
j<n, dD
is a disjoint family in A
t
with supremum e.
But whenever d D, j < n we must have d
j
c
t
for some c
t
C, so either d
j
c or d
j
c = 0; thus

j
(a d) must be either included in c or disjoint from it. Accordingly
c e = sup
j
(a d) : j < n, d D, d
j
c A
t
. QQQ
Consequently h(, A
t
) ln2r 2. PPP For any c C,
(c, A
t
) (c .(c e)) = (c \ e) min( c, 2)
1
3
,
so
ln2r = h(, C) h(A
t
) +H(C[A
t
)
(385Nd)
h(, A
t
) +

cC
q((c, A
t
))
(by the Kolmogorov-Sina theorem and 385Ob)
h(, A
t
) +

cC
q(min( c, 2)) h(, A
t
) +
by the choice of . QQQ
Finally, h(, A

) ln2r 4. PPP
ln2r 2 h(, A
t
) h(, A

) +H(a, 1 \ a) +H(e, 1 \ e)
(applying 385Nb twice)
= h(, A

) +q( a) +q(1 a) +q( e) +q(1 e)

h(, A

) +q(
1
n
) +q(
n
n+1
) +q(2) +q(1 2)
h(, A

) + + = h(, A

) + 2. QQQ
(g) It follows that H(c

, 1 \ c

) ln2 4. PPP We have

ln2r 4 h(, A

) H(A

) H(A) +H(c

, 1 \ c

) = lnr +H(c

, 1 \ c

),
so H(c

, 1 \ c

) ln2 4. QQQ Thus q( c

) +q(1 c

) ln2 4; by the choice of , [ c

1
2
[ .
Next,

iI
[ (a
i
c

)
1
2r
[ +[ (a
i
\ c

)
1
2r
[ 3.
386K Ornsteins theorem 547
PPP By 385K,

iI
[ (a
i
c

)
1
r
c

[ +[ (a
i
\ c

)
1
r
(1 \ c

)[

_
2(H(A) +H(c

, 1 \ c

) H(A

))

_
2(lnr + ln2 ln2r + 4) =
_
8 .
So

iI
[ (a
i
c

)
1
2r
[ +[ (a
i
\ c

)
1
2r
[

iI
_
[ (a
i
c

)
1
r
c

[ +
1
r
[ c

1
2
[
+[ (a
i
\ c

)
1
r
(1 \ c

)[ +
1
r
[ (1 \ c

)
1
2
[
_
+[ c

1
2
[ +[ (1 \ c

)
1
2
[ 3. QQQ
(h) Now apply 386D to the partition of unity A

, indexed as a

ij
)
iI,j0,1]
, where a

i1
= a
i
c

and
a

i0
= a
i
\ c

, and
ij
)
iI,j0,1]
, where
ij
=
1
2r
for all i, j. We have

iI,j0,1]
[ a

ij

ij
[ 3
by (g), while
H(A

) h(, A

) lnr + ln2 ln2r + 4 = 4,

so

iI,j0,1]
[ a

ij

ij
[ +
_
2(H(A

) h(, A

)) 3 +

8 4.
Also

iI,j0,1]
q(
ij
) = ln2r h().
So 386D tells us that there is a Bernoulli partition b
ij
)
iI,j0,1]
for such that b

ij
=
1
2r
for all i, j and

iI,j0,1]
(b
ij
.a

ij
) + 6

4 .
Now of course

iI
(a
i
.(b
i0
b
i1
))

iI
((a
i
c

) .b
i1
) + ((a
i
\ c

) .b
i0
)
=

iI,j0,1]
(a

ij
.b
ij
) ,
as required.
386K Ornsteins theorem (innite entropy case) Let (A, ) be a probability algebra of countable
Maharam type, and : A A a two-sided Bernoulli shift of innite entropy. Then (A, , ) is isomorphic
to (B, , ), where (B, ) is the measure algebra of the usual measure on [0, 1]
Z
, and is the standard
Bernoulli shift on B (384Sb).
proof (a) We have to nd a root algebra E for which is isomorphic to the measure algebra of the usual
measure on [0, 1]. The materials we have to start with are a root algebra A
0
of A such that either A
0
is not
purely atomic or H(A
0
) = , where A
0
is the set of atoms of A
0
.
Because A has countable Maharam type, there is a sequence d
n
)
nN
in A
0
such that d
n
: n N is
dense in the metric of A
0
.
548 Automorphism groups 386K
(b) There is a sequence C
n
)
nN
of partitions of unity in A
0
such that C
n+1
renes C
n
, H(C
n
) = nln2
and d
n
is a union of members of C
n+1
for every n. PPP We have
= supH(C) : C A
0
is a partition of unity
(384J). Choose the C
n
inductively, as follows. Start with C
0
= 0, 1. Given C
n
with H(C
n
) = nln2, set
C
t
n
= C
n
d
n
, 1 \ d
n
; then
H(C
t
n
) H(C
n
) +H(d
n
, 1 \ d
n
) (n + 1) ln2.
By 385Q, there is a partition of unity C
n+1
, rening C
t
n
, such that H(C
n+1
) = (n + 1) ln2. Continue. QQQ
(c) For each n N, let C
n
be the closed subalgebra of A generated by
j
a : a C
n
, j Z. Then
C
n
)
nN
is increasing, and d
n
C
n+1
, [C
n
] = C
n
and
h(C
n
) = h(, C
n
) = H(C
n
) = nln2
for every n.
Choose inductively, for each n N,
n
> 0, r
n
N and a Bernoulli partition b
n
)
0,1]
n in C
n
, as
follows. Start with b
n
= 1. (See 3A1H for the notation I am using here.) Given that b
n
)
0,1]
n is a
Bernoulli partition for which generates C
n
, in the sense that C
n
is the closed subalgebra of A generated
by
j
b
n
: 0, 1
n
, j Z, and b
n
= 2
n
for every , take
n
> 0 such that
(2r
m
+ 1)
n
2
n
for every m < n.
We know that
h(C
n+1
) = (n + 1) ln2 = ln(2 2
n
).
So we can apply 386J to (C
n+1
, C
n+1
) to see that there is a Bernoulli partition b
t
n
)
0,1]
n+1 for such
that
b
t
n
C
n+1
, b
t
n
= 2
n1
for every 0, 1
n+1
,
(b
n
.(b
t
n,

0
b
t
n,

1
)) 2
n

n
for every 0, 1
n
. By 386H (with B = C = C
n+1
), there is a Bernoulli partition b
n+1,
)
0,1]
n+1 for
C
n+1
such that the closed subalgebra generated by
j
b
n+1,
: 0, 1
n+1
, j Z is C
n+1
, b
n+1,
=
2
n1
for every 0, 1
n+1
, and

0,1]
n+1
(b
n+1,
.b
t
n
)
n
.
For each k N, let B
(n+1)
k
be the closed subalgebra of C
n+1
generated by
j
b
n+1,
: 0, 1
n+1
, [j[ k.
Since d
m
C
m+1
C
n+1
for every m n, there is an r
n
N such that
(d
m
, B
(n+1)
r
n
) 2
n
for every m n.
Continue.
(d) Fix m n N for the moment. For 0, 1
m
, set
b
n
= supb
n
: 0, 1
n
, extends .
(If n = m, then of course is the unique member of 0, 1
m
extending itself, so this formula is safe.) Then
b
n
= 2
n
#( : 0, 1
n
, extends ) = 2
n
2
nm
= 2
m
.
Next, if ,
t
0, 1
m
are distinct, there is no member of 0, 1
n
extending both, so b
n
b
n
= 0; thus
b
n
)
0,1]
m is a partition of unity. If (0), . . . , (k) 0, 1
m
, then
386K Ornsteins theorem 549
(inf
jk

j
b
m,(j)
) = ( sup
(0),... ,(k)0,1]
n
(j)(j)jk
inf
jk

j
b
n,(j)
)
=

(0),... ,(k)0,1]
n
(j)(j)jk
(inf
jk

j
b
n,(j)
)
=

(0),... ,(k)0,1]
n
(j)(j)jk
(2
n
)
k+1
= (2
nm
)
k+1
(2
n
)
k+1
= (2
m
)
k+1
=
k

j=0
b
n,(j)
,
so b
n
)
0,1]
m is a Bernoulli partition.
(e) If m n N, then

0,1]
m
(b
n
.b
n+1,
) 2
n
whenever m n N. PPP We have

0,1]
m
(b
n
.b
n+1,
)

0,1]
n
(b
n
.b
n+1,
)
=

0,1]
n
(b
n
.(b
n+1,

0
b
n+1,

1
))

0,1]
n
(b
n
.(b
t
n,

0
b
t
n,

1
)) +

0,1]
n+1
(b
t
n
.b
n+1,
)

0,1]
n
2
n

n
+
n
= 2
n
. QQQ
(f ) In particular, for any m N and 0, 1
m
,

n=m
(b
n
.b
n+1,
)

n=m
2
n
< .
So we can dene b

= lim
n
b
n
in A. We have
b

= lim
n
b
n
= 2
m
;
and if ,
t
0, 1
m
are distinct, then
b

= lim
n
b
n
b
n
= 0,
so b

)
0,1]
m is a partition of unity in A. If (0), . . . , (k) 0, 1
m
, then
(inf
jk

j
b
(j)
) = lim
n
(inf
jk

j
b
n,(j)
)
= lim
n
k

j=0
b
n,(j)
=
k

j=0
b
(j)
,
so b

)
0,1]
m is a Bernoulli partition for . If 0, 1
m
, then
b

0
b

1
= lim
n
b
n,

0
b
n,

1
= lim
n
b
n,
= b

.
(g) Let E be the closed subalgebra of A generated by

mN
b

: 0, 1
m
. Then E is atomless and
countably generated, so (E, E) is isomorphic to the measure algebra of Lebesgue measure on [0, 1]. Now
(inf
jk

j
e
j
) =

k
j=0
e
j
for all e
0
, . . . , e
k
E. PPP Let > 0. For m N, let E
m
be the subalgebra of E
550 Automorphism groups 386K
generated by b

: 0, 1
m
. E
m
)
mN
is non-decreasing, so

mN
E
m
is a closed subalgebra of A, and
must be E. Now the function
(a
0
, . . . , a
k
) (inf
jk

j
a
j
)

k
j=0
a
j
: A
k+1
R
is continuous and zero on E
k+1
m
for every m, by 386Bb, so is zero on E
k+1
, and in particular is zero at
(e
0
, . . . , e
k
), as required. QQQ
By 384Sf,
j
[E])
jZ
is independent.
(h) Let B

be the closed subalgebra of A generated by

j
b

mN
0, 1
m
, j Z; then E B

,
so B

is the closed subalgebra of A generated by

jZ

j
[E]. It follows from (e) that, for any m N,

0,1]
m
(b
m
, B

0,1]
m
(b
m
, b

0,1]
m

n=m
(b
n
, b
n+1,
) 2

n=m

n
.
So if b B
(m+1)
r
m
,
(b, B

) (2r
m
+ 1)

0,1]
m+1
(b
m+1,
, B

)
2(2r
m
+ 1)

n=m+1

n
2

n=m+1
2
n
= 2
m+1
.
It follows that, whenever m n in N,
(d
m
, B

) (d
m
, B
(n+1)
r
n
) + 2
n+1
2
n
+ 2
n+1
by the choice of r
n
. Letting n , we see that (d
m
, B

) = 0, that is, d
m
B

, for every m N. But

this means that A
0
B

, by the choice of d
m
)
mN
. Accordingly
j
[A
0
] B

for every j and B

must be
the whole of A.
(i) Thus is a two-sided Bernoulli shift with root algebra E; by 384Sc, (A, , ) is isomorphic to (B, , ).
386L Corollary: Sinas theorem (general case) Let (A, ) be an atomless probability algebra, and
: A A a measure-preserving automorphism. Let (B, ) be a probability algebra of countable Maharam
type, and : B B a one- or two-sided Bernoulli shift with h() h(). Then (B, , ) is isomorphic to
a factor of (A, , ).
proof (a) To begin with (down to the end of (b)) suppose that is two-sided. Let B
0
be a root algebra
for . If B
0
is purely atomic, then there is a generating Bernoulli partition b
i
)
iN
for of entropy h().
By 386E, there is a Bernoulli partition c
i
)
iN
for such that c
i
= b
i
for every i. Let C be the closed
subalgebra of A generated by
j
c
i
: i N, j Z. Now (C, C, C) is a factor of (A, , ) isomorphic to
(B, , ).
(b) If B
0
is not purely atomic, then there is still a partition of unity b
i
)
iN
in B
0
of innite entropy.
Again, let C be the closed subalgebra of A generated by
j
c
i
: i N, j Z, where c
i
)
iN
is a Bernoulli
partition for such that c
i
= b
i
for every i. Now C is a Bernoulli shift of innite entropy and C has
countable Maharam type, so 386K tells us that there is a closed subalgebra C
0
C such that
k
[C
0
])
kN
is independent and (C
0
, C
0
) is isomorphic to the measure algebra of Lebesgue measure on [0, 1]. But
(B
0
, B
0
) is a probability algebra of countable Maharam type, so is isomorphic to a closed subalgebra
C
1
of C
0
(332N). Of course
k
[C
1
])
kN
is independent, so if we take C

1
to be the closed subalgebra of A
generated by

kZ

k
[C
1
], C

1
will be a two-sided Bernoulli shift isomorphic to (384Sf).
(c) If is a one-sided Bernoulli shift, then 384Sa and 384Sc show that (B, , ) can be represented
in terms of a product measure on a space X
N
and the standard shift operator on X
N
. Now this extends
naturally to the standard two-sided Bernoulli shift represented by the product measure on X
Z
, as described
in 384Sb; so that (B, , ) becomes represented as a factor of (B
t
,
t
,
t
) where
t
is a two-sided Bernoulli
387A Dyes theorem 551
shift with the same entropy as (since the entropy is determined by the root algebra, by 384R). By (a)-(b),
(B
t
,
t
,
t
) is isomorphic to a factor of (A, , ), so (B, , ) also is.
Remark Thus (A, , ) has factors which are Bernoulli shifts based on root algebras of all countably-
generated types permitted by the entropy of .
386X Basic exercises (a) Let (A, ) be a probability algebra, and : A A a one- or two-sided
Bernoulli shift. Show that
n
is a Bernoulli shift for any n 1. (Hint: if A
0
is a root algebra for , the
closed subalgebra generated by

j<n

j
[A
0
] is a root algebra for
n
.)
(b) Let (A, ) be a probability algebra, B a closed subalgebra of A and Aut

A a measure-preserving
automorphism such that [B] = B. Show that if is ergodic or mixing, so is B.
(c) Let (A, ) be a probability algebra of countable Maharam type, and : A A a two-sided Bernoulli
shift. Show that for any n 1 there is a Bernoulli shift : A A such that
n
= . (Hint: construct
a Bernoulli shift such that h() =
1
n
h(), and use 384Xh and Ornsteins theorem to show that is
isomorphic to
n
.)
(d) Let
i
)
iN
,
i
)
iN
be non-negative real sequences such that

i=0

i
=

i=0

i
= 1 and

i=0
q(
i
)
=

i=0
q(
i
). Let
0
,
0
be the measures on N dened by the formulae

0
E =

iE

i
,
0
E =

iE

i
for E N. Set X = N
Z
and let , be the product measures on X derived from
0
and
0
. Show that there
is a bijection f : X X such that is precisely the image measure f
1
and f is translation-invariant,
that is, f(x) = f(x) for every x X, where (n) = n + 1 for every n Z.
386Y Further exercises (a) Suppose that (A, , ) and (B, , ) are probability algebras with one-
sided Bernoulli shifts, and that they are isomorphic. Show that they have isomorphic root algebras. (Hint:
apply the results of 333 to (A, , [A]).)
386 Notes and comments The arguments here are expanded from Smorodinsky 71 and Ornstein 74.
I have sought the most direct path to 386I and 386K; of course there is a great deal more to be said (386Xc
is a hint), and, in particular, extensions of the methods here provide powerful theorems enabling us to show
that automorphisms are Bernoulli shifts. (See Ornstein 74.)
387 Dyes theorem
I have repeatedly said that any satisfactory classication theorem for automorphisms of measure algebras
remains elusive. There is however a classication, at least for the Lebesgue measure algebra, of the orbit
structures corresponding to measure-preserving automorphisms; in fact, they are dened by the xed-point
subalgebras, which I described in 333. We have to work hard for this result, but the ideas are instructive.
387A Full subgroups Recall the denition of full subgroup (381M): if A is a Dedekind complete
Boolean algebra, then a subgroup G of its automorphism group is full if whenever a
i
)
iI
is a partition of
unity in A, and Aut A is such that for each i I there is a
i
G such that b =
i
b for every b a
i
,
then G. I take the opportunity to note two facts: (i) if A is a Dedekind complete Boolean algebra and
c A, then : Aut A, c = c is full (ii) if (A, ) is a localizable measure algebra, then the set of
measure-preserving automorphisms of A is a full subgroup of Aut A (382C).
Evidently the intersection of any family of full subgroups of Aut A is full, so we can speak of the full
subgroup G

generated by a given automorphism . From the facts cited above (or otherwise) we see that
(i)
t
if Aut A and C is the closed subalgebra c : c = c, then c = c for every G

and c C (ii)
t
if (A, ) is a localizable measure algebra and Aut A is measure-preserving, then every member of G

is
measure-preserving.
Now the subgroups G

are easy to describe, as follows.

552 Automorphism groups 387B
387B Proposition Let A be a Dedekind complete Boolean algebra and , Boolean automorphisms
of A. Then the following are equiveridical:
(i) belongs to the full subgroup of Aut A generated by ;
(ii) for every non-zero a A there are a non-zero b a and an n Z such that c =
n
c for every c b;
(iii) there is a partition of unity a
n
)
nZ
in A such that c =
n
c whenever n Z and c a
n
.
proof (i)(ii) Let G be the set of those automorphisms satisfying (ii). Then G is a full subgroup of Aut A.
PPP () Suppose that
1
,
2
G and a A 0. Then there are a non-zero b
1
a and an m Z such that

2
c =
m
c for every c b
1
. Next, there are a non-zero b
2

2
b
1
and an n Z such that
1
c =
n
c for
every c b
2
. Set b =
1
2
b
2
, so that 0 ,= b a and

2
c =
n

2
c =
n

m
c =
m+n
c
for every c b. As a is arbitrary,
1

2
G. () Suppose that G and a A\ 0. Then there are a
non-zero b
1

1
a and an n Z such that c =
n
c for every c b
1
. Set b = b
1
, so that 0 ,= b a; then
for any c b,

n
c =
n

1
c =
1
c.
As a is arbitrary,
1
G. () Of course G, so G is a subgroup of Aut A. () Now suppose that a
i
)
iI
is a partition of unity in A and that Aut A is such that for every i I there is a
i
G such that
c =
i
c for every c a
i
. If a A0, then there is an i I such that a a
i
,= 0, and there are a non-zero
b a a
i
and an n Z such that c =
i
c =
n
c for every c b. As a is arbitrary, G. As a
i
)
iI
is
arbitrary, G is full. QQQ
Since G is a full subgroup of Aut A containing , it includes the full subgroup generated by and, in
particular, G.
(ii)(iii) For n Z, let B
n
be the set of those b A such that c =
n
c for every c b. Set b
n
= sup B
n
for each n; then if c b
n
,
c = (sup
bB
n
b c) = sup
bB
n
(b c) = sup
bB
n

n
(b c) =
n
c.
Set
a
n
= b
n
\ sup
0i<n
b
i
if n N,
= b
n
\ sup
i>n
b
i
if n Z N;
then a
n
)
nZ
is disjoint,
sup
nZ
a
n
= sup
nZ
b
n
= sup(

nZ
B
n
) = 1,
and c =
n
c for every c a
n
, n Z; so (iii) is true.
(iii)(i) is trivial.
387C Orbit structures I said that this section was directed to a classication of orbit structures,
without saying what these might be. In fact what I will do is to classify the full subgroups generated by
measure-preserving automorphisms of the Lebesgue measure algebra. One aspect of the relation with orbits
is the following.
Proposition Let (X, , ) be a localizable countably separated measure space (denition: 343D), with
measure algebra (A, ). Suppose that f and g are measure space automorphisms from X to itself, inducing
measure-preserving automorphisms , of A. Then the following are equiveridical:
(i) belongs to the full subgroup of Aut A generated by ;
(ii) for almost every x X, there is an n Z such that g(x) = f
n
(x);
(iii) for almost every x X, g
n
(x) : n Z f
n
(x) : n Z.
proof (i)(ii) Let H
k
)
kN
be a sequence in which separates the points of X; we may suppose that
H
0
= X. By 387B, there is a partition of unity a
n
)
nZ
in A such that c =
n
c for every c a
n
, n Z.
For each n Z let E
n
be such that E

n
= a
n
; then Y
0
=

nZ
E
n
is conegligible. The transformation
f
n
induces
n
, so for any k N and n Z the set
387E Dyes theorem 553
F
nk
= x : f
n
(x) E
n
H
k
, g(x) / E
n
H
k

x : g(x) E
n
H
k
, f
n
(x) / E
n
H
k

is negligible, and Y = g
1
[Y
0
]

nZ,kN
F
nk
is conegligible. Now, for any x Y , there is some n such
that g(x) E
n
, so that f(x) E
n
and k : g(x) H
k
= k : f
n
(x) H
k
and g(x) = f
n
(x). As Y is
conegligible, (ii) is satised.
(ii)(iii) For x X, set
x
= f
n
(x) : n Z; we are supposing that A
0
= x : g(x) /
x
is negligible.
Set A =

nZ
g
n
[A
0
], so that A is negligible and g
n
(x) X A for every x X A, n Z.
Suppose that x X A and n N. Then g
n
(x)
x
. PPP Induce on n. Of course g
0
(x) = x
x
. For
the inductive step to n + 1, g
n
(x)
x
A
0
, so there is a k Z such that g
n
(x) = f
k
(x). At the same
time, there is an i Z such that g(g
n
(x)) = f
i
(g
n
(x)), so that g
n+1
(x) = f
i+k
(x)
x
. Thus the induction
continues. QQQ
Consequently g
n
(x)
x
whenever x X A and n N. PPP Since g
n
(x) X A, there is a k Z
such that x = g
n
g
n
(x) = f
k
g
n
(x) and g
n
(x) = f
k
(x)
x
. QQQ
Thus g
n
(x) : n Z
x
for every x in the conegligible set X A.
(iii)(ii) is trivial.
(ii)(i) Set
E
n
= x : g(x) = f
n
(x) = X

kN
(g
1
H
k
f
n
[H
k
]),
for n Z. Then (ii) tells us that

nZ
E
n
is conegligible, so

nZ
g[E
n
] is conegligible. But also each E
n
is measurable, so g[E
n
] also is, and we can set a
n
= g[E
n
]

. Now for y g[E

n
], y = f
n
(g
1
(y)), that is,
g
1
(y) = f
n
(y); so a =
n
a for every a a
n
. Since sup
nZ
a
n
= 1 in A, belongs to the full subgroup
generated by .
Remark Of course the requirement countably separated is essential here; for other measure spaces we can
have and actually equal without g(x) and f(x) being related for any particular x (see 343I and 343J).
387D Corollary Under the hypotheses of 387C, and generate the same full subgroup of Aut A i
f
n
(x) : n Z = g
n
(x) : n Z for almost every x X.
387E Induced automorphisms of principal ideals: Proposition Let (A, ) be a totally nite mea-
sure algebra and : A A a measure-preserving automorphism. For b A write A
b
for the principal ideal
of A generated by b.
(a) For any b A we have a measure-preserving automorphism
b
: A
b
A
b
dened by saying that

b
d =
n
d whenever n 1 and d b
n
b \ sup
1i<n

i
b.
(b) If c A and c = c then
b
(b c) = b c for every b A.
(c) If c b A, then
c
= (
b
)
c
, where (
b
)
c
is the automorphism of A
c
induced by
b
: A
b
A
b
.
(d) If 0 ,= a b A and n 1 there are a non-zero a
t
a and an m 1 such that
n
b
d =
m
d for every
d a
t
.
(e) If is aperiodic, so is
b
for every b A.
proof (a) For n 1 set
d
n
= b
n
b \ sup
1i<n

i
b.
If 1 m < n then
d
n

n
b \
m
b, d
m

m
b
so d
m
d
n
= 0. Also
d
m
b,
nm
d
n
b =
nm
(d
n

(nm)
b) = 0
so

n
d
n

m
d
m
=
m
(
nm
d
n
d
m
) = 0.
554 Automorphism groups 387E
Thirdly,
sup
n1
d
n
= sup
n1
b
n
b = b
by 385D, applied to
1
.
It follows that d
n
)
n1
is a partition of unity in A
b
. Since
n
d
n
)
n1
is also a disjoint family in A
b
, and

n=1

n
d
n
=

n=1
d
n
= b,
it is another partition of unity. So we have an automorphism
b
: A
b
A
b
dened by setting
b
d =
n
d if
d d
n
(381B). Of course
b
is measure-preserving, by 382C.
(b) If c = c, then
n
(d
n
c) c for every n, so
b
(b c) c; because
b
(b c) = (b c),
b
(b c) =
b c.
(c) Set D = d : d A
c
,
c
d = (
b
)
c
d. Then D is order-dense in A
c
. PPP Take any non-zero a A
c
. Since
c sup
n1

n
c, there is an n N such that a
t
= a
n
c \ sup
1i<n

i
c is non-zero. Next, there is a non-
zero d a
t
such that for every m n either d
m
b or d
m
b = 0. Enumerate m : m n, d
m
b
in ascending order as (m
0
, . . . , m
k
) (note that as a
t
b
n
b, we must have m
0
= 0 and m
k
= n). Set
d
i
=
m
i
d for i k, so that
d
0
= d,
m
i+1
m
i
d
i
= d
i+1
b,
while

j
d
i
=
m
i
+j
d 1 \ b
for 1 j < m
i+1
m
i
; that is, d
i+1
=
b
d
i
for i < k. Thus

k
b
d =
m
k
d =
n
d c,
while

i
b
d = d
i
=
m
i
d
m
i
a
t
1 \ c
for every i < k, and
(
b
)
c
d =
n
d =
c
d,
so that d D. As a is arbitrary, D is order-dense. QQQ
Because
c
and (
b
)
c
are both order-continuous Boolean homomorphisms on A
c
, and every member of
A
c
is a supremum of some subfamily of D (313K), they must be equal.
(d) Induce on n. We know that there are a non-zero c a and an i 1 such that
b
d =
i
d for every
d c. So if n = 1 we can just take a
t
= c and m = i. For the inductive step to n + 1, take a
t
and m 1
such that 0 ,= a
t

b
c and
n
b
d =
m
d for every d a
t
; then
0 ,=
1
b
a
t
c a
and
n+1
b
d =
m+i
d for every d
1
b
a
t
.
(e) If 0 ,= a b and n 1, take a non-zero a
t
a and m 1 such that
m
d =
n
b
d for every d a
t
((d)
above). Now (because is aperiodic) there is a d a
t
such that
m
d ,= d, so
n
b
d ,= d. As n and a are
arbitrary,
b
is aperiodic.
387F Similar ideas lead us to the following fact.
Lemma Let (A, ) be a totally nite measure algebra, and : A A a measure-preserving automorphism;
let C be the closed subalgebra c : c = c. Let d
i
)
iI
, e
i
)
iI
be two disjoint families in A such that
(c d
i
) = (c e
i
) for every i I and c C. Then there is a G

, the full subgroup of Aut A generated

by , such that d
i
= e
i
for every i I.
proof Adding d

= 1 \ sup
iI
d
i
, e

= 1 \ sup
iI
e
i
to the respective families, we may suppose that d
i
)
iI
,
e
i
)
iI
are partitions of unity. Dene a
n
)
nN
inductively by the formula
a
n
= sup
iI
(d
i
sup
m<n
a
m
)
n
(e
i
\ sup
m<n

m
a
m
).
387Gc Dyes theorem 555
Then a
n
d
i
a
m
= 0 whenever m < n and i I, so a
n
)
nN
is disjoint. Also

n
a
n
sup
iI
e
i
\ sup
m<n

m
a
m
for each n, so
n
a
n
)
nN
is disjoint. Note that as
n
(a
n
d
j
) e
j
for each j,

n
a
n
e
i
= sup
jI

n
(a
n
d
j
) e
i
= sup
jI

n
(a
n
d
j
) e
j
e
i
=
n
(a
n
d
i
) e
i
=
n
(a
n
d
i
)
for every i I, n N.
??? Suppose, if possible, that a = 1 \ sup
nN
a
n
is non-zero. Then there is an i I such that a d
i
,= 0.
Set c = sup
nN

n
(a d
i
); then c c so c C. Now

n=0
(c e
i

n
a
n
) =

n=0
(c
n
(a
n
d
i
)) =

n=0
(
n
(c a
n
d
i
))
=

n=0
(c a
n
d
i
) = (c d
i
\ a) < (c d
i
) = (c e
i
).
So b = c e
i
\ sup
nN

n
a
n
is non-zero, and there is an n N such that b
n
(a d
i
) is non-zero. But look
at a
t
=
n
(b
n
(a d
i
)). We have 0 ,= a
t
a d
i
, so a
t
d
i
\ sup
m<n
a
m
; while

n
a
t
b e
i
\ sup
m<n

m
a
m
.
But this means that a
t
a
n
, which is absurd. XXX
This shows that a
n
)
nN
is a partition of unity in A. Since

n=0
(
n
a
n
) =

n=0
a
n
= 1,

n
a
n
)
nN
is also a partition of unity. We can therefore dene G

by setting d =
n
d whenever n N
and d a
n
. Now, for any i I,
d
i
= sup
nN
(d
i
a
n
) = sup
nN

n
(d
i
a
n
) = sup
nN
e
i
a
n
= e
i
.
So we have found a suitable .
387G von Neumann transformations: Denitions (a) Let A be a Boolean algebra and Aut A
an automorphism. is weakly von Neumann if there is a sequence a
n
)
nN
in A such that a
0
= 1 and,
for every n, a
n+1

2
n
a
n+1
= 0, a
n+1

2
n
a
n+1
= a
n
. In this case, is von Neumann if a
n
)
nN
can be
chosen in such a way that
m
a
n
: m, n N -generates A, and relatively von Neumann if a
n
)
nN
can be chosen so that
m
a
n
: m, n N c : c = c -generates A.
(b) There is another way of looking at automorphisms of this type which will be useful. If A is a
Boolean algebra and : A A an automorphism, then a dyadic cycle system for is a nite or innite
family d
mi
)
mn,i<2
m or d
mi
)
mN,i<2
m such that () for each m, d
mi
)
i<2
m is a partition of unity such
that d
mi
= d
m,i+1
whenever i < 2
m
1 (so that d
m,2
m
1
must be d
m0
) () d
m0
= d
m+1,0
d
m+1,2
m for
every m < n (in the nite case) or for every m N (in the innite case). An easy induction on m shows
that if k m then
d
ki
= supd
mj
: j < 2
m
, j i mod 2
k

for every i < 2

k
.
Conversely, if d is such that
j
d)
j<2
n is a partition of unity in A, then we can form a nite dyadic cycle
system d
mi
)
mn,i<2
m by setting d
mi
= sup
j
d : j < 2
n
, j i mod 2
m
whenever m n and j < 2
m
.
(c) Now an automorphism : A A is weakly von Neumann i it has an innite dyadic cycle system
d
mi
)
mN,i<2
m. (The a
m
of (a) correspond to the d
m0
of (b); starting from the denition in (a), you must
check rst, by induction on m, that
i
a
m
)
i<2
m is a partition of unity in A.) is von Neumann i it has a
dyadic cycle system d
mi
)
mN,i<2
m which -generates A.
556 Automorphism groups 387H
387H Example The following is the basic example of a von Neumann transformation in a sense,
the only example of a measure-preserving von Neumann transformation. Let be the usual measure on
X = 0, 1
N
, its domain, and (A, ) its measure algebra. Dene f : X X by setting
f(x)(n) = 1 x(n) if x(i) = 0 for every i < n,
= x(n) otherwise.
Then f is a homeomorphism and a measure space automorphism. PPP (i) To see that f is a homeomorphism,
perhaps the easiest way is to look at g, where
g(x)(n) = 1 x(n) if x(i) = 1 for every i < n,
= x(n) otherwise,
and check that f and g are both continuous and that fg, gf are both the identity function. (ii) To see that
f is inverse-measure-preserving, it is enough to check that x : f(x)(i) = z(i) for every i n = 2
n1
for
every n N, z X (254G). But
x : f(x)(i) = z(i) for every i n = x : x(i) = g(z)(i) for every i n.
(iii) Similarly, g is inverse-measure-preserving, so f is a measure space automorphism. QQQ
If n N, x X then
f
2
k
(x)(n) = 1 x(n) if n k and x(i) = 0 whenever k i < n,
= x(n) otherwise.
(Induce on k. For the inductive step, observe that if we identify X with 0, 1 X then f
2
(, y) = (, f(y))
for every 0, 1 and y X.)
Let : A A be the corresponding automorphism, setting E

= f
1
[E]

for E . Then is
a von Neumann transformation. PPP Set E
n
= x : x X, x(i) = 1 for every i < n, a
n
= E

n
. Then
f
2
n
[E
n+1
] = x : x(i) = 1 for i < n, x(n) = 0, so a
n+1
,
2
n
a
n+1
split a
n
for each n, and a
n
)
nN
witnesses that is weakly von Neumann. Next, inducing on n, we nd that f
i
[E
n
] : i < 2
n
runs over
the basic cylinder sets of the form x : x(i) = z(i) for every i < n determined by coordinates less than
n. Since the equivalence classes of such sets -generate A (see part (a) of the proof of 331K), is a von
Neumann transformation. QQQ
For another way of looking at the functions f and g, see 445Xq in Volume 4.
387I We are now ready to approach the main results of this section.
Lemma Let (A, ) be a totally nite measure algebra and : A A an aperiodic measure-preserving
automorphism. Let C be the closed subalgebra c : c = c. Then for any a A there is a b a such that
(b c) =
1
2
(a c) for every c C and
b
is a weakly von Neumann transformation, writing
b
for the
induced automorphism of the principal ideal A
b
, as in 387E.
Remark On rst reading, there is something to be said for supposing here that is ergodic, that is, that
C = 0, 1.
proof Set
n
=
1
2
(1 + 2
n
) for each n N, so that
n
)
nN
is strictly decreasing, with
0
= 1 and
lim
n

n
=
1
2
. Now there are b
n
)
nN
, d
ni
)
nN,i<2
n such that, for each n N,
b
n+1
b
n
a, (b
n
c) =
n
(a c) for every c C,
d
ni
)
i<2
n is disjoint, sup
i<2
n d
ni
= b
n
,

b
n
d
ni
= d
n,i+1
for every i < 2
n
1,
b
n+1
d
ni
= d
n+1,i
d
n+1,i+2
n for every i < 2
n
.
PPP Start with b
0
= d
00
= a. To construct b
n+1
and d
n+1,i
)
i<2
n+1, given d
ni
)
i<2
n, note rst that (because

b
n
is measure-preserving and
b
n
(b
n
c) = b
n
c) (d
n0
c) = (d
ni
c) whenever c C, i < 2
n
, so
387J Dyes theorem 557
(d
n0
c) = 2
n
(b
n
c) = 2
n

n
(a c)
for every c C, and
d
n0
= b
n
\ sup
i<2
n
1

b
n
d
ni
=
b
n
d
n,2
n
1
=
2
n
b
n
d
n0
.
Now
b
n
is aperiodic (387Ee) so
2
n
b
n
also is (385A), and there is a d
n+1,0
d
n0
such that

2
n
b
n
d
n+1,0
d
n+1,0
= 0, (d
n+1,0
c) = 2
n1

n+1
(a c) for every c C
(applying 385E(iii) to
2
n
b
n
A
d
n0
, with =
n+1
/2
n
). Set d
n+1,j
=
j
b
n
d
n+1,0
for each j < 2
n+1
. Be-
cause
2
n
b
n
d
n+1,0
d
n0
\ d
n+1,0
, while
j
b
n
d
n0
)
j<2
n is disjoint,
j
b
n
d
n+1,0
)
j<2
n+1 is disjoint. Set b
n+1
=
sup
i<2
n+1
i
b
n
d
n+1,0
; then b
n+1
b
n
and (b
n+1
c) =
n+1
(a c) for every c C. For j < 2
n+1
,
d
n+1,j
d
ni
where i is either j or j 2
n
, so b
n+1
d
ni
= d
n+1,i
d
n+1,i+2
n.
For i < 2
n+1
1,

b
n
d
n+1,i
= d
n+1,i+1
b
n+1
,
so we must also have

b
n+1
d
n+1,i
= (
b
n
)
b
n+1
d
n+1,i
= d
n+1,i+1
(using 387Ec). Thus the induction continues. QQQ
Set
b = inf
nN
b
n
, e
ni
= b d
ni
for n N, i < 2
n
.
Because b
n
)
nN
is non-increasing,
(b c) = lim
n
(b
n
c) =
1
2
(a c)
for every c C. Next,
e
ni
= b b
n+1
d
ni
= b (d
n+1,i
d
n+1,i+2
n) = e
n+1,i
e
n+1,i+2
n
whenever i < 2
n
.
If m n, j < 2
m
then
b
n
d
mj
= supd
ni
: i < 2
n
, i j mod 2
m

(induce on n). So
(b
n
d
mj
) = 2
nm
d
n0
= 2
m

n
;
taking the limit as n , e
mj
= 2
m
b. Next,

b
n
(b
n
d
mj
) = supd
n,i+1
: i < 2
n
, i j mod 2
m

= supd
ni
: i < 2
n
, i j + 1 mod 2
m
= b
n
d
m,j+1
,
here interpreting d
n,2
n as d
n0
, d
m,2
m as d
m0
. Consequently
b
e
mj
e
m,j+1
. PPP??? Otherwise, there are a
non-zero e d
mj
b and k 1 such that
i
e b = 0 for 1 i < k and
k
e b \ d
m,j+1
. Take n m so
large that e > k (b
n
\ b), so that
e
t
= e \ sup
1i<k

i
(b
n
\ b) ,= 0;
now
i
e
t
b
n
= 0 for 1 i < k, while
k
e
t
b
n
, and

b
n
e
t
=
k
e
t
1 \ d
m,j+1
.
But this means that
b
n
(b
n
d
mj
) , d
m,j+1
, which is impossible. XXXQQQ
Since
b
e
mj
= e
m,j+1
, we must have
b
e
mj
= e
m,j+1
. And this is true whenever m N and j < 2
m
,
if we identify e
m,2
m with e
m0
. Thus e
mi
)
mN,i<2
m is a dyadic cycle system for
b
and
b
is a weakly von
Neumann transformation.
387J Lemma Let (A, ) be a totally nite measure algebra and , two measure-preserving automor-
phisms of A. Suppose that belongs to the full subgroup G

generated by and that there is a b A such

that sup
nZ

n
b = 1 and the induced automorphisms
b
,
b
on A
b
are equal. Then G

= G

.
558 Automorphism groups 387J
proof (a) The rst fact to note is that if 0 ,= b
t
b, n Z and
n
b
t
b, then there are m Z, b
tt
b
t
such
that b
tt
,= 0 and
n
d =
m
d for every d b
tt
. PPP () If n = 0 take b
tt
= b
t
, m = 0. () Next, suppose that
n > 0. Take a non-zero b
1
b
t
such that
i
b
1
is either included in b or disjoint from b for every i n (e.g.,
an atom of the nite subalgebra generated by b
t

i
b : 0 i n). Enumerate i : i n,
i
b
1
b in
ascending order as (l
0
, . . . , l
k
). Then
l
j
d =
j
b
d whenever d b
1
and j k (compare part (c) of the proof
of 387E); in particular,
n
d =
k
b
d =
k
b
d for every d b
1
. But now 387Ed tells us that there must be a
non-zero b
tt
b
1
and an m N such that

m
d =
k
b
d =
n
d
for every d b
tt
. () If n < 0, then set e
t
=
n
b
t
, so that e
t
,
n
e
t
b. By (), there are a non-zero e
tt
e
t
and an s N such that
n
d =
s
d for every d e
tt
. Setting b
tt
=
n
e
tt
=
s
e
tt
, we have 0 ,= b
tt
b
t
and

n
d =
s
d for every d b
tt
. QQQ
(b) Now take any non-zero a A. Then there are m, n Z such that a
1
= a
m
b ,= 0, a
2
=
a
1

n
b ,= 0. Set b
1
=
m

1
a
2
. Because G

, there are a non-zero b

2
b
1
and a k Z such that

m
d =
k
d for every d b
2
. Now

k
b
2
=
n

m
b
2

n

m
b
1
=
n
a
2
b.
By (a), there are a non-zero b
3
b
2
and an r Z such that
k
d =
r
d for every d b
3
. Consider a
t
=
m
b
3
.
Then
0 ,= a
t

m
b
1
=
1
a
2
a
1
a;
and, for d a
t
,
m
d b
3
b
2
, so that
d =
n
(
n

m
)
m
d =
n

m
d =
n+rm
d.
As a is arbitrary, this shows that G

, so that G

and the two are equal.

387K Lemma Let (A, ) be a totally nite measure algebra, : A A an aperiodic measure-
preserving automorphism, and any member of the full subgroup G

of Aut A generated by . Suppose that

d
mi
)
mn,i<2
m is a nite dyadic cycle system for . Then there is a weakly von Neumann transformation ,
with dyadic cycle system d
t
mi
)
mN,i<2
m, such that G

= G

, a = a whenever a d
n0
= 0, and d
t
mi
= d
mi
whenever m n and i < 2
m
.
proof Write C for the closed subalgebra c : c = c. By 387I there is a b d
n0
such that (b c) =
1
2
(d
n0
c) for every c C and
b
: A
b
A
b
is a weakly von Neumann transformation. Let e
ki
)
kN,i<2
k
be a dyadic cycle system for
b
.
If we dene
1
Aut A by setting

1
d =
b
d for d b,
1
d =
1\b
d for d 1 \ b,
then
1
G

. Next, for any c C,

(
2
n
+1
b c) =
2
n
+1
(b c) = (b c) =
1
2
(d
n0
c) = ((d
n0
\ b) c)
because
2
n
+1
G

, so
2
n
+1
c = c. By 387F, there is a
2
G

such that
2
(d
n0
\ b) =
2
n
+1
b. Set

3
=
2
n
+1

1
2

2
n
+1

1
, so that
3
G

and

3
b =
2
n
+1

1
2

2
n
+1
b =
2
n
+1
(d
n0
\ b).
Thus
3
b and
2
(d
n0
\ b) are disjoint and have union
2
n
+1
d
n0
= d
n1
(if n = 0, we must read d
01
as
d
00
= 1). Accordingly we can dene G

by setting
d =
3
d if d b,
=
2
d if d d
n0
\ b,
= d if d d
n0
= 0.
Since d
n0
= d
n1
, we have d
ni
= d
ni
for every i < 2
n
, and therefore
i
d
m0
= d
mi
whenever m n
and i < 2
m
. Looking at
2
n
, we have
387L Dyes theorem 559

2
n
d
n0
=
2
n
d
n0
= d
n0
,
2
n
b =
2
n
1

3
b = d
n0
\ b,
so that
2
n
(d
n0
\ b) = b and
2
n+1
b = b. Accordingly

2
n+1
d =
2
n
1

2
n
1

3
d =
1
d =
b
d
for every d b. Also sup
i<2
n+1
i
b = 1, so 387J tells us that G

= G

.
Now dene a
m
)
mN
as follows. For m n, a
m
= d
m0
; for m > n, a
m
= e
mn1,0
. Then for m < n we
have

2
m
a
m+1
=
2
m
d
m+1,0
=
2
m
d
m+1,0
= d
m+1,2
m = a
m
\ a
m+1
,
for m = n we have

2
n
a
n+1
=
2
n
e
00
=
2
n
b = d
n0
\ b = a
n
\ a
n+1
,
and for m > n we have

2
m
a
m+1
= (
2
n+1
)
2
mn1
e
mn,0
= (
b
)
2
mn1
e
mn,0
= e
mn,2
mn1 = e
mn1,0
\ e
mn,0
= a
m
\ a
m+1
.
Thus a
m
)
mN
witnesses that is a weakly von Neumann transformation. If d
t
mi
=
i
a
m
for m N, i < 2
m
then d
t
mi
)
mN,i<2
m will be a dyadic cycle system for and d
t
mi
= d
mi
for m n, as required.
387L Lemma Let (A, ) be a totally nite measure space and C a closed subalgebra of A such that A
is relatively atomless over C. For a A write C
a
= a c : c C.
(a) Suppose that b A, w C and > 0 are such that (b c) c whenever c C and c w. Then
there is an e A such that e b w and (e c) = c whenever c C
w
.
(b) Suppose that k 1 and that (b
0
, . . . , b
r
) is a nite partition of unity in A. Then there is a partition
E of unity in A such that
(e c) =
1
k
c for every e E, c C,
#(e : e E, i r, b
i
e / C
e
) r + 1.
proof (a) Set a = b w and consider the principal ideal A
a
generated by A. We know that (A
a
, A
a
) is
a totally nite measure algebra (322H), and that C
a
is a closed subalgebra of A
a
(333Bc); and it is easy to
see that A
a
is relatively atomless over C
a
.
Let : C
w
C
a
be the Boolean homomorphism dened by setting c = c b for c C
w
. If c C
w
and
c = 0, then c C and c (c b) = 0, so c = 0; thus is injective; since it is certainly surjective, it is
a Boolean isomorphism. We can therefore dene a functional =
1
: C
a
[0, [, and we shall have
d d for every d C
a
. By 331B, there is an e A
a
such that d = (d e) for every d C
a
, that is,
c = (c e) for every c C
w
, as required.
(b)(i) Write D for the set of all those e A such that (c e) =
1
k
c for every c C and b
i
e C
e
for
every i r. Then whenever a A and >
r+1
k
is such that (a c) = c for every c C, there is an e D
such that e a. PPP For d A, c C set
d
(c) = (d c), so that
d
: C [0, [ is a completely additive
functional. For i r set v
i
= [[ C > k
ab
i
]], in the notation of 326P; so that v
i
C and c k(a b
i
c)
whenever c C and c v
i
, while c k (a b
i
c) whenever c C and c v
i
= 0. Setting v = inf
ir
v
i
,
we have
k v = k (a v) =

r
i=0
k(a b
i
v) (r + 1) v.
Since k > r + 1, v = 0. So if we now set w
i
= (inf
j<i
v
j
) \ v
i
for i r (starting with w
0
= 1 \ v
0
),
(w
0
, . . . , w
r
) is a partition of unity in C, and c k (a b
i
c) whenever c C and c w
i
.
By (a), we can nd for each i r an e
i
A such that e
i
a b
i
w
i
and (c e
i
) =
1
k
c whenever c C
and c w
i
. Set e = sup
ir
e
i
, so that e a,
e b
i
= e w
i
b
i
= e
i
= e w
i
C
e
for each i, and
560 Automorphism groups 387L
(c e) =
r

i=0
(c e
i
) =
r

i=0
(c w
i
e
i
)
=
r

i=0
1
k
(c w
i
) =
1
k
c
for every c C. So e has all the properties required. QQQ
(ii) Let E
0
D be a maximal disjoint family, and set m = #(E
0
), a = 1 \ supE
0
. Then
(a c) = c

eE
0
(c e) = (1
m
k
) c
for every c C, while a does not include any member of D. By (i), 1
m
k

r+1
k
, that is, m k r 1.
Applying (a) repeatedly, with w = 1 and =
1
k
, we can nd disjoint d
0
, . . . , d
km1
a such that
(c d
i
) =
1
k
c for every c C and i < k m. So if we set E = E
0
d
i
: i < k m we shall have a
partition of unity with the properties required.
387M Lemma Let (A, ) be a totally nite measure algebra and : A A an aperiodic measure-
preserving automorphism. Write C for the closed subalgebra c : c = c. Suppose that is a member of
the full subgroup G

of Aut A generated by with a nite dyadic cycle system d

mi
)
mn,i<2
m, and that
a A and > 0. Then there is a G

such that
(i) has a dyadic cycle system d
t
mi
)
mk,i<2
m, with k n and d
t
mi
= d
mi
for m n, i < 2
m
;
(ii) d = d if d d
n0
= 0;
(iii) there is an a
t
in the subalgebra of A generated by C d
t
ki
: i < 2
k
such that (a .a
t
) .
proof (a) Take k n so large that 2
k
2
n
2
2
n
1. Let D be the subalgebra of the principal ideal
A
d
n1
generated by d
n1

j
a : j < 2
n
; then D has atoms b
0
, . . . , b
r
where r < 2
2
n
. (If n = 0, take
d
01
= d
00
= 1.) Applying 387L to the closed subalgebra C
d
n1
of A
d
n1
, we can nd a partition of unity E of
A
d
n1
such that
(e c) = 2
k
c for every e E, c C
E
1
= e : e E, there is some i r such that b
i
e / C
e

has cardinal at most r + 1 2

2
n
. Of course e = 2
k
for every e E, so #(E) = 2
kn
and (supE
1
)
2
k
2
2
n
2
n
. Write e

for supE
1
.
(b) For e E set e
t
=
2
n
1
e; then e
t
: e E is a disjoint family, of cardinal 2
kn
; enumerate it as
v
i
)
i<2
kn. Note that
sup
i<2
kn v
i
=
2
n
1
(supE) = d
n0
,
(v
i
c) = (
2
n
+1
v
i
c) = 2
k
c
for every c C, i < 2
kn
. There is therefore a
1
G

such that

1
v
i
=
2
n
+1
v
i+1
for i < 2
kn
1,
1
v
2
kn
1
=
2
n
+1
v
0
(387F). We have

1
d
n0
=
1
( sup
i<2
kn
v
i
) = sup
i<2
kn

1
v
i
= sup
i<2
kn
1

2
n
+1
v
i+1

2
n
+1
v
0
= sup
i<2
kn

2
n
+1
v
i
=
2
n
+1
d
n0
= d
n1
= d
n0
.
So we may dene G

by setting
d =
1
d if d d
n0
,
= d if d d
n0
= 0.
387N Dyes theorem 561
(c) For each i < 2
kn
,

2
n
v
i
=
2
n
1

1
v
i
= v
i+1
(identifying v
2
kn with v
0
). Moreover,
j
v
i
d
nl
whenever i < 2
kn
and j l mod 2
n
. So
j
v
0
)
j<2
k is a
partition of unity in A. What this means is that if we set
d
t
mj
= sup
i
v
0
: i < 2
k
, i j mod 2
m

for m k, then d
t
mj
)
mk,j<2
m is a dyadic cycle system for , with d
t
mj
= d
mj
if m n, j < 2
m
.
(d) Let B be the subalgebra of A generated by Cd
t
kj
: j < 2
k
. Recall the denition of v
i
: i < 2
kn

as
2
n
1
e : e E; this implies that
v
i
: i < 2
kn
=
1
v
i
: i < 2
kn
=
2
n
+1
v
i
: i < 2
kn
= E,
so that

j+1
v
i
: i < 2
kn
=
j
e : e E
for j < 2
n
, and
B d
t
kj
: j < 2
k
=
j
v
i
: i < 2
kn
, j < 2
n
=
j
e : e E, j < 2
n
.
Set E
0
= E E
1
. For e E
0
and i r there is a c
ei
C such that e b
i
= e c
ei
. Set
K = (i, j) : 1 i r, j < 2
n
, b
i

j
a,
a
t
= sup
j
e c
ei
: e E
0
, (i, j) K.
Then a
t
is a supremum of (nitely many) members of B, so belongs to B. If (i, j) K and e E
0
, then

j
e c
ei
=
j
(e c
ei
) =
j
(e b
i
) a,
so a
t
a. Next, d
n1

j
(a \ a
t
) e

for each j < 2

n
. PPP Set
I = i : i r, (i, j) K = i : b
i

j
a;
then d
n1

j
a = sup
iI
b
i
. Now, for each i I,
b
i
= sup
eE
(b
i
e) sup
eE
0
(e c
ei
) e

,
so that
d
n1

j
a = sup
iI
b
i
sup
eE
0
,iI
(e c
ei
) e

= (d
n1

j
a
t
) e

. QQQ
But this means that
(d
n,j+1
a \ a
t
) = (d
n1

j
(a \ a
t
)) e

2
n

for every j < 2

n
(interpreting d
n,2
n as d
n0
, as usual), and
(a .a
t
) =

2
n
j=1
(d
nj
a \ a
t
) ,
so that the nal condition of the lemma is satised.
387N Theorem Let (A, ) be a totally nite measure algebra, with Maharam type , and : A A
an aperiodic measure-preserving automorphism. Then there is a relatively von Neumann transformation
: A A such that and generate the same full subgroups of Aut A.
proof (a) The idea is to construct as the limit of a sequence
n
)
nN
of weakly von Neumann transforma-
tions such that G

n
= G

. Each
n
will have a dyadic cycle system d
nmi
)
mN,i<2
m; there will be a strictly
increasing sequence k
n
)
nN
such that
d
n+1,m,i
= d
n,m,i
whenever m k
n
, i < 2
m
,

n+1
a =
n
a whenever a d
n,k
n
,0
= 0.
Interpolated between the
n
will be a second sequence
n
)
nN
in G

, with associated (nite) dyadic cycle

systems d
t
nmi
)
mk

n
,i<2
m.
562 Automorphism groups 387N
(b) Before starting on the inductive construction we must x on a countable set B A which -generates
A, and a sequence b
n
)
nN
in B such that every member of B recurs conally often in the sequence. (For
instance, take the sequence of rst members of an enumeration of BN.) As usual, I write C for the closed
subalgebra c : c = c. The induction begins with
0
= , k
t
0
= 0, d
t
000
= 1. Given
n
G

and its dyadic

cycle system d
t
nmi
)
mk

n
,i<2
m, use 387K to nd a weakly von Neumann transformation
n
, with dyadic
cycle system d
nmi
)
mN,i<2
m, such that G

n
= G

, d
nmi
= d
t
nmi
for m k
t
n
and i < 2
m
, and
n
a =
n
a
whenever a d
t
n,k

n
,0
= 0.
(c) Given the weakly von Neumann transformation
n
, with its dyadic cycle system d
nmi
)
mN,i<2
m,
such that G

n
= G

, then we have a partition of unity e

nj
)
jZ
such that a =
j
n
a whenever j Z
and a e
nj
(387C(ii)). Take r
n
such that e
n
2
n
, where e
n
= sup
]j]>r
n
e
nj
, and k
n
> k
t
n
such that
2
k
n
(2r
n
+ 1) 2
n
. Set
e

n
= sup
]j]r
n

j
n
d
n,k
n
,0
,
so that e

n
2
n+1
.
Now use 387M to nd a
n+1
G

, with a dyadic cycle system d

t
n+1,m,i
)
mk

n+1
,i<2
m, such that
k
t
n+1
k
n
, d
t
n+1,m,i
= d
nmi
if m k
n
,
n+1
a =
n
a if a d
n,k
n
,0
= 0, and there is a b
t
n
in the algebra
generated by C d
t
n+1,m,i
: m k
t
n+1
, i < 2
m
such that (b
n
.b
t
n
) 2
n
. Continue.
(d) The eect of this construction is to ensure that if l < n in N then
d
lmi
= d
nmi
whenever m k
l
, i < 2
m
,

n
a =
l
a whenever a d
l,k
l
,0
= 0,
b
t
l
belongs to the subalgebra generated by C d
nmi
: m k
n
, i < 2
m
,
and, of course, d
n,k
n
,0
d
l,k
l
,0
. Since k
n
)
nN
is strictly increasing, inf
nN
d
n,k
n
,0
= 0. Now, for each n N,
d
n,k
n
,1
=
n
d
n,k
n
,0
=
n+1
d
n,k
n
,0

n+1
d
n+1,k
n+1
,0
= d
n+1,k
n+1
,1
,
so setting
a
0
= 1 \ d
0,k
0
,0
, a
n+1
= d
n,k
n
,0
\ d
n+1,k
n+1
,0
for each n,
we have

0
a
0
= 1 \ d
0,k
0
,1
,
n+1
a
n+1
= d
n,k
n
,1
\ d
n+1,k
n+1
,1
for each n,
and
n
a
n
)
nN
is a partition of unity. There is therefore a Aut A dened by setting a =
n
a if a a
n
;
because G

is full, G

.
(e) If m n, then a
m
d
n,k
n
,0
= 0, so
n
a =
m
a = a for every a a
m
. Thus
n
a = a for every
a sup
mn
a
m
= 1 \ d
n,k
n
,0
. In particular, d
nmi
= d
n,m,i+1
whenever m k
n
, 1 i < 2
m
(counting
d
n,m,2
m as d
nm0
, as usual); so that in fact d
nmi
= d
n,m,i+1
whenever m k
n
, i < 2
m
.
For each n, we have d
nmi
= d
t
nmi
= d
n+1,m,i
whenever m k
n
and i < 2
m
. We therefore have a family
d

mi
)
mN,i<2
m dened by saying that d

mi
= d
nmi
whenever n N, m k
n
and i < 2
m
. Now, for any
m N, there is a k
n
> m, so that d

mi
)
i<2
m = d
nmi
)
i<2
m is a partition of unity; and
d

mi
= d
nmi
= d
n,m+1,i
d
n,m+1,i+2
m = d

m+1,i
d

m+1,i+2
m
for each i < 2
m
. Moreover,
d

m,i
=
n
d
nmi
= d
n,m,i+1
= d

m,i+1
at least for 1 i < 2
m
(counting d

m,2
m as d

m,0
, as usual), so that in fact d

mi
= d

m,i+1
for every i < 2
m
.
Thus d

mi
)
mN,i<2
m is a dyadic cycle system for , and is a weakly von Neumann transformation.
Writing B for the closed subalgebra of A generated by C d

mi
: m N, i < 2
m
, then
C d
t
nmi
: m k
t
n
, i < 2
m
= C d
n+1,m,i
: m k
t
n
, i < 2
m

= C d

mi
: m k
t
n
, i < 2
m
B
387O Dyes theorem 563
for any n N. So b
t
n
B for every n. If b B and > 0, there is an n N such that 2
n
and b
n
= b,
so that (b .b
t
n
) ; as every b
t
n
belongs to B, and B is closed, b B; as b is arbitrary, and B -generates
A, B = A. Thus is a von Neumann transformation.
(f ) If n N and d e

n
= 0, then
j
d =
j
n
d and
j
d =
j
n
d whenever 0 j r
n
. PPP Induce on
j. For j = 0 the result is trivial. For the inductive step to j + 1 r
n
, note that if d
t
d
n,k
n
,1
= 0 then

1
n
d
t
d
n,k
n
,0
= 0, so

1
d
t
=
1

n
(
1
n
d
t
) =
1
(
1
n
d
t
) =
1
n
d
t
.
Now we have

j+1
d = (
j
n
d) =
n
(
j
n
d) =
j+1
n
d
because

j
n
d d
n,k
n
,0
=
j
n
(d
j
n
d
n,k
n
,0
) = 0,
while

j1
d =
1
(
j
n
d) =
1
n
(
j
n
d) =
j1
n
d
because

j
n
d d
n,k
n
,1
=
j
n
(d
j+1
n
d
n,k
n
,0
) = 0. QQQ
Thus
j
d =
j
n
d whenever [j[ r
n
.
(g) Finally, G

= G

. PPP I remarked in (d) that G

, so that G

. To see that G

, take
any non-zero a A. Because (e

n
e
n
) 2
n+1
for each n, there is an n such that a
t
= a \ (e

n
e
n
) ,= 0.
Now there is some j Z such that a
tt
= a
t
e
nj
,= 0; since a
t
e
n
= 0, [j[ r
n
. If d a
tt
, then d =
j
n
d,
by the denition of e
nj
. But also
j
n
d =
j
d, by (f), because d e

n
= 0. So d =
j
d for every d a
tt
. As
a is arbitrary, G

and G

. QQQ
This completes the proof.
387O Theorem Let (A
1
,
1
) and (A
2
,
2
) be totally nite measure algebras of countable Maharam
type, and
1
: A
1
A
1
,
2
: A
2
A
2
measure-preserving automorphisms. For each i, let C
i
be the closed
subalgebra c : c A
i
,
i
c = c and G

i
the full subgroup of Aut A
i
generated by
i
. If (A
1
,
1
, C
1
) and
(A
2
,
2
, C
2
) are isomorphic, so are (A
1
,
1
, G

1
) and (A
2
,
2
, G

2
).
proof (a) It is enough to consider the case in which (A
1
,
1
, C
1
) and (A
2
,
2
, C
2
) are actually equal; I
therefore delete the subscripts and speak of a structure (A, , C), with two automorphisms
1
,
2
of A both
with xed-point subalgebra C.
(b) Suppose rst that A is relatively atomless over C, that is, that both the
i
are aperiodic (385E). In
this case, 387N tells us that there are von Neumann transformations
1
and
2
of A such that G

1
= G

1
and
G

2
= G

2
. But (A, ,
1
) and (A, ,
2
) are isomorphic. PPP Let d
mi
)
mN,i<2
m and d
t
mi
)
mN,i<2
m be dyadic
cycle systems for
1
,
2
respectively such that C d
mi
: m N, i < 2
m
and C d
t
mi
: m N, i < 2
m

both -generate A.
Writing B
1
, B
2
for the subalgebras of A generated by C d
mi
: m N, i < 2
m
and C d
t
mi
: m
N, i < 2
m
respectively, it is easy to see that these algebras are isomorphic: we just set
0
c = c for c C,

0
d
mi
= d
t
mi
for i < 2
m
to obtain a measure-preserving isomorphism
0
: B
1
B
2
. Because these are
topologically dense subalgebras of A, there is a unique extension of
0
to a measure-preserving automorphism
: A A (324O). Next, we see that

1
c = c =
2
c for every c C,

1
d
t
mi
=
1
d
mi
= d
m,i+1
= d
t
m,i+1
=
2
d
t
m,i+1
for m N, i < 2
m
(as usual, taking d
m,2
m to be d
m0
and d
t
m,2
m to be d
t
m0
). But this means that

1
b =
2
b for every b B
2
, so (again because B
2
is dense in A)
1

1
=
2
. Thus is an isomorphism
between (A, ,
1
) and (A, ,
2
). QQQ
Of course is now also an isomorphism between (A, , G

1
) = (A, , G

1
) and (A, , G

2
) = (A, , G

2
).
564 Automorphism groups 387O
(c) Next, consider the case in which
1
is periodic, with period n, for some n 1. In this case
2
G

1
.
PPP Let (d
0
, . . . , d
n1
) be a partition of unity in A such that
1
d
i
= d
i+1
for i < n 1 and
1
d
n1
= d
0
(385B). If d d
j
, then c = sup
i<n

i
1
d C and d = d
j
c; so any member of A is of the form sup
j<n
d
j
c
j
for some family c
0
, . . . , c
n1
in C.
If a A\ 0, take i, j < n such that a
t
= a d
i

1
2
d
j
,= 0. Then any d a
t
is of the form
d
i
c
1
=
1
2
(d
j
c
2
) = c
2

1
2
d
j
for some c
1
, c
2
C; setting c = c
1
c
2
, we have
d = d
i
c,
2
d = d
j
c =
ji
1
d.
As a is arbitrary, this shows that
2
G

1
. QQQ
Now sup
nZ

n
2
d
0
belongs to C and includes d
0
, so must be 1. Finally, if we look at the two induced
automorphisms (
1
)
d
0
, (
2
)
d
0
on A
d
0
, they must both be the identity, by 387Eb, because every element of
A
d
0
is of the form d
0
c for some c C. So 387J tells us that G

1
= G

2
.
(d) For the general case, we see from 385C that there is a partition of unity c
i
)
1i
in C such that

1
A
c

is aperiodic and if i is nite and c

i
,= 0 then
1
A
c
i
is periodic with period i. For each i, let H
i
be
A
c
i
: G

1
; then H
i
is a full subgroup of Aut A
c
i
, and
G

1
= : Aut A, A
c
i
H
i
whenever 1 i .
Similarly, writing H
t
i
= A
c
i
: G

2
,
G

2
= : Aut A, A
c
i
H
t
i
whenever 1 i .
Note also that H
i
, H
t
i
are the full subgroups of Aut A
c
i
generated by
1
A
c
i
,
2
A
c
i
respectively. By (b)
and (c), H
i
= H
t
i
for nite i, while there is a measure-preserving automorphism : A
c

A
c

such that
H

1
= H
t

. Now we can dene a measure-preserving automorphism

1
: A A by setting
1
a = a
if a c

,
1
a = a if a c

= 0, and we shall have

1
G

1
1
= G

2
. Thus (A, , G

1
) and (A, , G

2
) are
isomorphic, as claimed.
387X Basic exercises >>>(a) Let (X, , ) be a localizable measure space, with measure algebra (A, ).
Suppose that and are automorphisms of A, and that is represented (in the sense of 343A) by a
measure space automorphism f : X X. Show that the following are equiveridical: (i) belongs to
the full subgroup of Aut A generated by ; (ii) there is a function g : X X, representing , such that
g(x) f
n
(x) : n Z for every x X; (iii) there is a measure space automorphism g : X X, representing
, such that g
n
(x) : n Z f
n
(x) : n Z for every x X.
(b) Let (X, , ) be a totally nite measure space, with measure algebra (A, ); let f : X X be a
measure space isomorphism, and : A A the corresponding measure-preserving automorphism. Take
E and b = E

A; let
b
: A
b
A
b
be the corresponding induced automorphism, as in 387E. (i)
Set E
1
= x : x E, n 1, f
n
(x) E; show that (E E
1
) = 0, so that E

1
= b. (ii) For x E
1
set n
x
= minn : n 1, f
n
x
(x) E, g
E
(x) = f
n
x
(x). Show that g
E
: E
1
E induces
b
: A
b
A
b
.
(iii) Show that there is a measurable set G E
1
such that E
1
G is negligible and g
E
G = g
G
is an
automorphism for the induced measure on G. (Hint: try G = E

nN

m,rn
f
m
[E] f
r
[E].) (iv) Show
that, in (iii), g
n
G
(x) : n Z = G f
n
(x) : n Z for every x G. (v) Find an expression of 387Ec in
this language.
(c) Let (A, ) be a totally nite measure algebra, and : A A a measure-preserving automorphism.
Let us say that a pseudo-cycle for is a partition of unity a
i
)
i<n
, where n 1, such that a
i
= a
i+1
for
i < n 1 (so that a
n1
= a
0
). (i) Show that if we have pseudo-cycles a
i
)
i<n
and b
j
)
j<m
, where m is a
multiple of n, then we have a pseudo-cycle c
j
)
j<m
with c
0
a
0
, so that a
i
= supc
j
: j < m, j i mod
n for every i < n. (ii) Show that is weakly von Neumann i it has a pseudo-cycle of length 2
n
for any
n N.
(d) Let (A
1
,
1
) and (A
2
,
2
) be probability algebras, and
1
: A
1
A
2
and
2
: A
2
A
2
measure-
preserving von Neumann automorphisms. Show that there is a measure-preserving Boolean isomorphism
: A
1
A
2
such that
2
=
2

1
.
387 Notes Dyes theorem 565
(e) Let (A, ) be an atomless probability algebra of countable Maharam type, and Aut

A the group of
measure-preserving Boolean automorphisms of A. Let Aut

A be a von Neumann transformation. (i)
Show that for any ultralter T on N there is a
J
Aut

A dened by the formula
J
(a) = lim
nJ

n
a for
every a A, the limit being taken in the measure-algebra topology. (ii) Show that
J
: T is an ultralter
on N is a subgroup of Aut

A homeomorphic to Z
N
2
. (Hint: 387H.)
(f ) Let A be a Boolean algebra and : A A a weakly von Neumann automorphism. Show that
n
is
a weakly von Neumann automorphism for every n Z 0. (Hint: consider n = 2, n = 1, odd n 3
separately. The formula of 387H may be useful.)
(g) Let A be a Boolean algebra and : A A a von Neumann automorphism. (i) Show that is ergodic
(in the sense that its xed-point algebra is trivial) but
2
is not ergodic. (ii) Show that
2
is relatively von
Neumann. (iii) Show that
n
is von Neumann for every odd n Z 0.
387Y Further exercises (a) Let A be a Dedekind complete Boolean algebra and G a semigroup of
order-continuous Boolean homomorphisms from A to itself. Let us say that G is full if whenever : A A
is an order-continuous Boolean homomorphism, and there is a partition of unity a
i
)
iI
in A such that for
every i I there is a
i
G such that a =
i
a for every a a
i
, then G. Show that if and
are order-continuous Boolean homomorphisms from A to itself, then the following are equiveridical: (i)
belongs to the full semigroup generated by ; (ii) for every non-zero a A there are a non-zero b a and an
n N such that d =
n
d for every d b; (iii) there is a partition of unity a
n
)
nN
in A such that a =
n
a
for every n N, a a
n
.
(b) Let A be a Dedekind complete Boolean algebra and Z its Stone space. Let , : A A be
Boolean automorphisms, and f, g : Z Z the corresponding homeomorphisms. Show that the following
are equiveridical: (i) belongs to the full subgroup of Aut A generated by ; (ii) z : g(z) / f
n
(z) : n Z
is nowhere dense in Z; (iii) z : g
n
(z) : n Z , f
n
(z) : n Z is nowhere dense in Z.
(c) Let X be a set, a -algebra of subsets of X, and 1 a -ideal of such that the quotient algebra
A = /1 is Dedekind complete and there is a countable subset of separating the points of X. Suppose
that f and g are automorphisms of the structure (X, , 1) inducing , Aut A. Show that the following
are equiveridical: (i) belongs to the full subgroup of Aut A generated by ; (ii) x : x X, f(x) / g
n
(x) :
n Z 1; (iii) x : x X, f
n
(x) : n Z , g
n
(x) : n Z 1.
(d) Let (A, ) be a totally nite measure algebra and : A A a relatively von Neumann transformation.
Show that is aperiodic and has zero entropy.
(e) Let (A
1
,
1
) and (A
2
,
2
) be atomless probability algebras, and (A, ) their probability algebra free
product. Let
1
: A
1
A
1
be a measure-preserving von Neumann automorphism and
2
: A
2
A
2
a
mixing measure-preserving automorphism. Let : A A be the measure-preserving automorphism such
that (a
1
a
2
) =
1
(a
1
)
2
(a
2
) for all a
1
A
1
, a
2
A
2
. Show that is an ergodic weakly von Neumann
automorphism which is not a relatively von Neumann automorphism.
(f ) Let be Lebesgue measure on [0, 1]
2
, and (A, ) its measure algebra; let C be the closed subalgebra of
elements expressible as (E [0, 1])

, where E [0, 1] is measurable. Suppose that : A A is a measure-

preserving automorphism such that C = c : c = c. Show that there is a family f
x
)
x[0,1]
of ergodic
measure space automorphisms of [0, 1] such that (x, y) (x, f
x
(y)) is a measure space automorphism of
[0, 1]
2
representing .
387 Notes and comments Dyes theorem (Dye 59) is actually Theorem 387O in the case in which
1
,
2
are ergodic, that is, in which C
1
and C
2
are both trivial. I take the trouble to give the generalized form here
(a simplied version of that in Krieger 76) because it seems a natural target, once we have a classication
of the relevant structures (A, , C) (333R). The essential mathematical ideas are the same in both cases.
You can nd the special case worked out in Hajian Ito & Kakutani 75, from which I have taken the
argument used here; and you may nd it useful to go through the version above, to check what kind of
566 Automorphism groups 387 Notes
simplications arise if each C is taken to be 0, 1. Essentially the dierence will be that every aperiodic
turns into ergodic (with an occasional atomless thrown in) and 331B turns into 331C. As far as I know,
there is no simplication available in the structure of the argument; of course the details become a bit easier,
but with the possible exception of 387L-387M I think there is little dierence.
Of course modifying a general argument to give a simpler proof of a special case is a standard exercise in
this kind of mathematics. What is much more interesting is the reverse process. What kinds of theorem about
ergodic automorphisms will in fact be true of all automorphisms? A variety of very powerful approaches
to such questions have been developed in the last half-century, and I hope to describe some of the ideas in
Volumes 4 and 5. The methods used in this section are relatively straightforward and do not require any
deep theoretical underpinning beyond Maharams lemma 331B. But an alternative approach can be found
using 387Yf: in eect (at least for the Lebesgue measure algebra) any measure-preserving automorphism can
be disintegrated into ergodic measure space automorphisms (the bre maps f
x
of 387Yf). It is sometimes
possible to guess which theorems about ergodic transformations are uniformisable in the sense that they can
be applied to such a family f
x
)
x[0,1]
, in a systematic way, to provide a structure which can be interpreted
on the product measure. The details tend to be complex, which is one of the reasons why I do not attempt
to work through them here; but such disintegrations can be a most valuable aid to intuition.
Dyes theorem itself is less signicant than many results which I have omitted; I include it partly because
it is relatively easy and partly because some of the ideas on the way are very important indeed. In particular,
the concept of recurrence is vital. I have already oered a couple of relevant exercises (382Yc-382Yd). The
fundamental lemmas introduced in this section are 387E and 387F.
In this section I use von Neumann transformations as an auxiliary tool: the point is, rst, that two von
Neumann transformations are isomorphic that is, the von Neumann transformations on a given totally
nite measure algebra (A, ) (necessarily isomorphic to the Lebesgue measure algebra, since we must have A
atomless and (A) = ) form a conjugacy class in the group Aut

A of measure-preserving automorphisms;
and next, that for any ergodic measure-preserving automorphism (on an atomless totally nite algebra
of countable Maharam type) there is a von Neumann transformation such that G

= G

(387N). But
I think they are remarkable in themselves. A (weakly) von Neumann transformation has a pseudo-cycle
(387Xc) for every power of 2. For some purposes, existence is all we need to know; but in the arguments of
387K-387N we need to keep track of named pseudo-cycles in what I call dyadic cycle systems (387G).
In this volume I have systematically preferred arguments which deal directly with measure algebras,
rather than with measure spaces. I believe that such arguments can have a simplicity and clarity which
repays the extra eort of dealing with more abstract structures. But undoubtedly it is necessary, if you are
to have any hope of going farther in the subject, to develop methods of transferring intuitions and theorems
between the two contexts. I oer 387Xb as an example. The description there of induced automorphism
requires a certain amount of manoeuvering around negligible sets, but in 387Xb(iv) gives a valuably graphic
description. In the same way, 387C, 387Xa and 387Yb provide alternative ways of looking at full subgroups.
There are contexts in which it is useful to know whether an element of the full subgroup generated by
actually belongs to the full semigroup generated by (387Ya); for instance, this happens in 387F.
391C Kelleys theorem 567
Chapter 39
Measurable algebras
In the nal chapter of this volume, I present results connected with the following question: which algebras
can appear as the underlying Boolean algebras of measure algebras? Put in this form, there is a trivial answer
(391A). The proper question is rather: which algebras can appear as the underlying Boolean algebras of
semi-nite measure algebras? This is easily reducible to the question: which algebras can appear as the
underlying Boolean algebras of probability algebras? Now in one sense Maharams theorem (332) gives us
the answer exactly: they are the countable simple products of the measure algebras of 0, 1

for cardinals
. But if we approach from another direction, things are more interesting. Probability algebras share a very
large number of very special properties. Can we nd a selection of these properties which will be sucient
to force an abstract Boolean algebra to be a probability algebra when endowed with a suitable functional?
No fully satisfying answer to this question is known. But in exploring the possibilities we encounter
some interesting and important ideas. In 391 I discuss algebras which have strictly positive additive real-
valued functionals; for such algebras, weak (, )-distributivity is necessary and sucient for the existence
of a measure; so we are led to look for conditions sucient to ensure that there is a strictly positive
additive functional. A slightly dierent approach lies through the concept of submeasure. Submeasures
arise naturally in the theories of topological Boolean algebras, topological Riesz spaces and vector measures
(see the second half of 393), and on any given algebra there is a strictly positive uniformly exhaustive
submeasure i there is a strictly positive additive functional; this is the Kalton-Roberts theorem (392F). It
is unknown whether the word uniformly can be dropped; this is one of the forms of the Control Measure
Problem, which I investigate at length in 393. In 394, I look at a characterization in terms of the special
properties which the automorphism group of a measure algebra must have (Kawadas theorem, 394Q). 395
complements the previous section by looking briey at the subgroups of an automorphism group Aut A
which can appear as groups of measure-preserving automorphisms.
391 Kelleys theorem
In this section I introduce the notion of measurable algebra (391B), which will be the subject of the
whole chapter once the trivial construction of 391A has been dealt with. I show that for weakly (, )-
distributive algebras countable additivity can be left to look after itself, and all we need to nd is a strictly
positive nitely additive functional (391D). I give Kelleys criterion for the existence of such a functional
(391H-391J), and a version of Gaifmans example of a ccc Boolean algebra without such a functional (391N).
391A Proposition Let A be any Dedekind -complete Boolean algebra. Then there is a function
: A [0, ] such that (A, ) is a measure algebra.
proof Set 0 = 0, a = for a A 0.
391B Denition I will call a Boolean algebra A measurable if there is a functional : A [0, [
such that (A, ) is a totally nite measure algebra.
In this case, if ,= 0, then it has a scalar multiple with total mass 1. So a Boolean algebra A is measurable
i either it is 0 or there is a functional such that (A, ) is a probability algebra.
391C Proposition Let A be a Boolean algebra.
(a) The following are equiveridical: (i) there is a functional : A [0, ] such that (A, ) is a semi-nite
measure algebra; (ii) A is Dedekind -complete and a : a A, A
a
is measurable is order-dense in A,
writing A
a
for the principal ideal generated by a.
(b) The following are equiveridical: (i) there is a functional : A [0, ] such that (A, ) is a localizable
measure algebra; (ii) A is Dedekind complete and a : a A, A
a
is measurable is order-dense in A.
proof (a) (i)(ii): if (A, ) is a semi-nite measure algebra, then A
f
= a : a < is order-dense in A
and A
a
is measurable for every a A
f
.
568 Measurable algebras 391C
(ii)(i): setting D = a : a A, A
a
is measurable, D is order-dense, so there is a partition of unity
C D (313K). For each c C, choose
c
such that (A
c
,
c
) is a totally nite measure algebra. Set
a =

cC

c
(a c) for every a A; then it is easy to check that (A, ) is a semi-nite measure algebra.
(b) Follows immediately.
391D Theorem Let A be a Boolean algebra. Then the following are equiveridical:
(i) A is measurable;
(ii) A is Dedekind -complete and weakly (, )-distributive, and there is a strictly positive nitely
additive functional : A [0, [.
Remark An additive functional on a Boolean algebra A is strictly positive if a > 0 for every non-zero
a A.
proof (i)(ii) Put the denition together with 322C(b)-(c) (for Dedekind completeness) and 322F (for
weak (, )-distributivity).
(ii)(i) Given that (ii) is satised, let be the completely additive part of , as dened in 326Yq: that
is,
a = infsup
dD
d : D is a non-empty upwards-directed set with supremum a
for every a A. Then (A, ) is a totally nite measure algebra.
PPP() Of course 0 a 1 for every a A, so is a function from A to [0, [.
() Suppose that a
n
)
nN
is a disjoint sequence in A with supremum a. Let > 0. For each n,
choose an upwards-directed set D
n
, with supremum a
n
, such that sup
dD
n
d a
n
+ 2
n
; choose an
upwards-directed set D, with supremum a, such that sup
dD
d a +.
Set
D

= sup
in
d
i
: n N, d
i
D
i
for every i n.
Then D

is upwards-directed, d a for every d D

, and any upper bound for D

must be an upper bound

for D
n
for every n, so sup D

= a. This means that

a sup
dD

d = sup(sup
in
d
i
) : d
i
D
i
for i n
= sup
n

i=0
d
i
: d
i
D
i
for i n
(because d
i
d
j
a
i
a
j
= 0 if i ,= j, d
i
D
i
, d
j
D
j
)
=

i=0
sup
dD
i
d

i=0
a
i
+ 2
i
= 2 +

i=0
a
i
.
Next, set D
t
n
= a
n
d : d D for each n. Then D
t
n
is upwards-directed and has supremum a
n
a = a
n
(313Ba), so

i=0
a
i

i=0
sup
dD

i
d = sup

n
i=0
(d
i
a
i
) : n N, d
i
D for i n.
But, given n N and d
i
D for i n, there is a d D such that sup
in
d
i
d (because D is upwards-
directed), so that (because a
0
, . . . , a
n
are disjoint)

n
i=0
(d
i
a
i
) d a +.
Accordingly

i=0
a
i
a +.
As is arbitrary,

i=0
a
i
= a; as a
n
)
nN
is arbitrary, is countably additive.
() Now suppose that a A and a = 0. Then for every n N there is an upwards-directed set
D
n
, with supremum a, such that d 2
n
for every d D
n
. Set B
n
= a \ d : d D
n
, so that each
B
n
is downwards-directed and has inmum 0 (313A). Because A is weakly (, )-distributive, there is a
391E Kelleys theorem 569
set B A, also with inmum 0, such that for every b B, n N there is a b
t
B
n
with b
t
b. If
B = then A = 0 and surely a = 0. Otherwise, set D = a \ b : b B. Then D is upwards-directed
and has supremum a. But if d D, then for every n N there is a b
t
B
n
such that b
t
a \ d, that is,
d a \ b
t
D
n
, and d 2
n
. Accordingly d = 0; but we are supposing that is strictly positive, so this
means that D = 0 and a must be 0.
This shows that is strictly positive.
() Thus is a strictly positive countably additive functional; as we are also assuming that A is
Dedekind -complete, (A, ) is a totally nite measure algebra. QQQ
Accordingly A is measurable.
391E Thus we are led naturally to the question: which Boolean algebras carry strictly positive nitely
additive functionals? The Hahn-Banach theorem, suitably applied, gives some sort of answer to this question.
For the sake of applications later on, I give two general results on the existence of additive functionals related
to given functionals.
Theorem Let A be a Boolean algebra, not 0, and : A [0, 1] a functional. Then the following are
equiveridical:
(i) there is a nitely additive functional : A [0, 1] such that 1 = 1 and a a for every a A;
(ii) whenever a
i
)
iI
is a nite indexed family in A, and

iI
a
i
m1 in S = S(A) (denition: 361A),
where m N, then

iI
a
i
m.
proof (a)(i)(ii) If : A [0, 1] is a nitely additive functional such that 1 = 1 and a a for every
a A, let h : S R be the positive linear functional corresponding to (361G). Now if a
i
)
iI
is a nite
family in A and

iI
a
i
m1, then

iI
a
i

iI
a
i
=

iI
h(a
i
)
= h(

iI
a
i
) h(m1) = m.
As a
i
)
iI
is arbitrary, (ii) is true.
(b)(ii)(i) Now suppose that satises (ii). For u S, set
p(u) = inf

n
i=0

i
a
i
: a
0
, . . . , a
n
A,
0
, . . . ,
n
0,

n
i=0

i
a
i
u.
Then it is easy to check that p(u +v) p(u) +p(v) for all u, v S, and that p(u) = p(u) for all u S,
0. Also p(1) 1. PPP??? If not, there are a
0
, . . . , a
n
A and
0
, . . . ,
n
0 such that 1

n
i=0

i
a
i
but

n
i=0

i
a
i
< 1. Increasing each
i
slightly if necessary, we may suppose that every
i
is rational; let
m 1 and k
0
, . . . , k
n
N be such that
i
= k
i
/m for each i n.
Set K = (i, j) : 0 i n, 1 j k
i
, and for (i, j) K set a
ij
= a
i
. Then

(i,j)K
a
ij
=

n
i=0
k
i
a
i
= m

n
i=0

i
a
i
m1,
but

(i,j)K
a
ij
=

n
i=0
k
i
a
i
= m

n
i=0

i
a
i
< m,
which is supposed to be impossible. XXXQQQ
By the Hahn-Banach theorem, in the form 3A5Aa, there is a linear functional h : S R such that
h(1) = p(1) 1 and h(u) p(u) for every u S. In particular, h(a) b whenever a b A. Set
a = h(a) for a A; then : A [0, [ is an additive functional, 1 1 and a b whenever a b
in A. We do not know whether is positive, but if we dene
+
as in 362Ab, we shall have a non-negative
additive functional such that

+
a = sup
ba
b a
for every a A, and
1 1
+
1 1 = 1,
so
+
witnesses that (i) is true.
570 Measurable algebras 391F
391F Theorem Let A be a Boolean algebra, not 0, and : A [0, 1] a functional, where A A.
Then the following are equiveridical:
(i) there is a non-negative nitely additive functional : A [0, 1] such that 1 = 1 and a a for
every a A;
(ii) whenever a
i
)
iI
is a nite indexed family in A, there is a set J I such that #(J)

iI
a
i
and
inf
iJ
a
i
,= 0.
Remark In (ii) here, we may have to interpret the inmum of the empty set in A as 1.
proof We apply 391E to the functional , where a = 1 (1 \ a) for a A.
(a) If : A [0, 1] is a non-negative nitely additive functional such that 1 = 1, then
a a 1 a 1 a (1 \ a) (1 \ a).
So (i) here is true of i 391E(i) is true of .
(b) Suppose that (ii) here is true of , and that a
i
)
iI
is a nite family in A such that

iI
a
i
m1,
while

iI
a
i
= . Then

iI
(1 \ a
i
) =

iI
(1 a
i
) = #(I) ,
so there is a set J I such that #(J) #(I) and inf
iJ
(1 \ a
i
) = c ,= 0. Now c a
i
= 0 for i J, so
mc

iI
(a
i
c) =

iI\J
(a
i
c) #(I J)c
and m #(I) #(J) . As a
i
)
iI
is arbitrary, 391E(ii) is true of .
(c) Suppose that 391E(ii) is true of , and that a
i
)
iI
is a nite family in A. Set
=

iI
(1 \ a
i
) = #(I)

iI
a
i
and let k be the least integer greater than . Since

iI
(1 \ a
i
) < k,

iI
(1 \ a
i
) , k1, that is,

iI
a
i
, (#(I) k)1. But this means that there must be some J I such that #(J) > #(I) k and
inf
iJ
a
i
,= 0. Now

iI
a
i
= #(I) #(I) (k 1) #(J).
As a
i
)
iI
is arbitrary, (ii) here is true of .
(d) Since we know that 391E(i) 391E(ii), we can conclude that (i) and (ii) here are equivalent.
391G Corollary Let A be a Boolean algebra, B a subalgebra of A, and
0
: B R a non-negative
nitely additive functional. Then there is a non-negative nitely additive functional : A R extending

0
.
proof (a) Suppose rst that
0
1 = 1. Set b =
0
b for every b B. Then must satisfy the condition (ii)
of 391F when regarded as a functional dened on a subset of B; but this means that it satises the same
condition when regarded as a functional dened on a subset of A. So there is a non-negative nitely additive
functional : A R such that 1 = 1 and b
0
b for every b B. In this case
b = 1 (1 \ b) 1
0
(1 \ b) =
0
b b
for every b B, so extends
0
.
(b) For the general case, if
0
1 = 0 then
0
must be the zero functional on B, so we can take to be the
zero functional on A; and if
0
1 = > 0, we apply (a) to
1

0
.
391H Denition Let A be a Boolean algebra, and A A 0 any non-empty set. The intersection
number of A is the largest 0 such that whenever a
i
)
iI
is a nite family in A, with I ,= , there is a
J I such that #(J) #(I) and inf
iJ
a
i
,= 0.
Remarks (a) It is essential to note that in the denition above the a
i
)
iI
are indexed families, with
repetitions allowed; see 391Xh.
(b) I spoke perhaps rather glibly of the largest such that . . . ; you may prefer to write
391N Kelleys theorem 571
= infsup
,=J0,... ,n],inf
jJ
a
j
,=0
#(J)
n+1
: a
0
, . . . , a
n
A.
391I Proposition Let A be a Boolean algebra and A A 0 any non-empty set. Write C for the set
of non-negative nitely additive functionals : A [0, 1] such that 1 = 1. Then the intersection number
of A is precisely max
C
inf
aA
a.
proof Write for the intersection number of A, and
t
for sup
C
inf
aA
a.
(a) For any <
t
, we can nd a C such that a for every a A. So if we set a = for every
a A, satises condition (i) of 391F. But this means that if a
i
)
iI
is any nite family in A, there must
be a J I such that inf
iJ
a
i
,= 0 and #(J) #(I). Accordingly ; as is arbitrary,
t
.
(b) Dene : A [0, 1] by setting a = for every a A. If a
i
)
iI
is a nite indexed family in A,
there is a J I such that #(J) #(I) and inf
iJ
a
i
,= 0; but #(I) =

iI
a
i
, so this means that
condition (ii) of 391F is satised. So there is a C such that a for every a A; and witnesses not
only that
t
, but that the supremum is a maximum.
391J Theorem Let A be a Boolean algebra. Then the following are equiveridical:
(i) there is a strictly positive nitely additive functional on A;
(ii) either A = 0 or A0 is expressible as a countable union of sets with non-zero intersection numbers.
proof (i)(ii) If there is a strictly positive nitely additive functional on A, and A ,= 0, set A
n
=
a : a 2
n
1 for every n N; then (applying 391I to the functional
1
1
) we see that every A
n
has
intersection number at least 2
n
, while A 0 =

nN
A
n
because is strictly positive, so (ii) is satised.
(ii)(i) If A 0 is expressible as

nN
A
n
, where each A
n
has intersection number
n
> 0, then for
each n choose a nitely additive functional
n
on A such that
n
1 = 1,
n
a
n
for every a A
n
. Setting
a =

n=0
2
n

n
a for every a A, is a strictly positive additive functional on A, and (i) is true.
391K Corollary Let A be a Boolean algebra. Then A is measurable i it is Dedekind -complete and
weakly (, )-distributive and either A = 0 or A 0 is expressible as a countable union of sets with
non-zero intersection numbers.
proof Put 391D and 391J together.
391L 391J-391K are due to Kelley 59; condition (ii) of 391J is called Kelleys criterion. It
provides some sort of answer to the question which Boolean algebras carry strictly positive nitely additive
functionals?, but leaves quite open the possibility that there is some more abstract criterion which is also
necessary and sucient. It is indeed a non-trivial exercise to nd any ccc Boolean algebra which does not
carry a strictly positive nitely additive functional. The rst example published seems to have been that of
Gaifman 64; I present this in a modied form, following Comfort & Negrepontis 82. First, a denition:
Denition Let A be a Boolean algebra. A set A A is linked if a b ,= 0 for all a, b A. A is -linked
if A 0 is a countable union of linked sets.
Remark -linkedness is one of a very long list of chain conditions which have been studied; there is an
extensive investigation in Comfort & Negrepontis 82. I hope to look at some of them in Volume 5. For
the moment I will introduce them one by one as the occasion arises.
391M Lemma A -linked Boolean algebra is ccc.
proof If A is -linked, take a sequence A
n
)
nN
of linked sets with union A 0. Then any disjoint subset
of A can meet each A
n
in at most one element, so must be countable.
391N Example There is a -linked Boolean algebra A such that there is no strictly positive nitely
additive real-valued functional dened on A.
572 Measurable algebras 391N
proof (a) Write 1 for the set of half-open intervals of the form [q, q
t
[ where q, q
t
Q and q < q
t
. Then
1 is countable; enumerate it as I
n
)
nN
. For each n N let
n
1 be a disjoint family of subintervals of
I
n
such that #(
n
) = (n + 1)
2
. Give 0, 1
R
its usual product topology, so that it is a compact Hausdor
space (3A3K); for x 0, 1
R
set Q
x
= t : t R, x(t) = 1. Now set
X =

nN
x : x 0, 1
R
, #(J : J
n
, Q
x
J ,= ) n + 1.
Take A to be the algebra of open-and-closed subsets of X.
(b) I show rst that there is no strictly positive nitely additive functional on A. PPP??? If there were,
there would be a sequence A
k
)
kN
with union A such that every A
k
had strictly positive intersection
number (391J). For each t R, write e
t
= x : x X, x(t) = 1, so that e
t
A. Note that if we set
x
t
(t) = 1, x
t
(s) = 0 for s ,= t, then Q
x
= t so x
t
X, x
t
e
t
and e
t
,= ; so there must be some k such
that e
t
A
k
. Set T
k
= t : e
t
A
k
for each k; then

kN
T
k
= R. By Baires theorem (3A3G) there is a
k N such that G = int T
k
,= .
Let > 0 be the intersection number of A
k
. There must be some n 1/ such that I
n
G. Since
I T
k
, there is for every J
n
a point t
J
J T
k
, so that e
t
J
A
k
.
Consider the family e
t
J
)
J
n
. Because the intersection number of A
k
is , there must be a set /
n
,
of cardinal at least #(
n
) = (n + 1)
2
> n + 1, such that inf
JK
e
t
J
,= , that is, X

JK
e
t
J
,= . But
if x

JK
e
t
J
then Q
x
contains t
J
for every J /, so
#(J : J
n
, Q
x
J ,= ) #(/) > n + 1
and x / X. XXXQQQ
Thus A does not satisfy Kelleys criterion and there is no strictly positive nitely additive functional
dened on A.
(c)(i) Write F for the set of functions f such that domf is a nite subset of R and f(t) 0, 1 for
every t domf; for f F write
Q
f
= t : t domf, f(t) = 1,
H
f
= x : x 0, 1
R
, x(t) = f(t) for every t domf,
c
f
= X H
f
A.
(ii) For k N write K
k
for the set of all disjoint nite families / of non-empty half-open intervals with
rational endpoints such that #(/) = k and whenever n 2k and J
n
then every member of / is either
included in J or disjoint from J. (This allows to belong to K
0
.) For / K
k
let F
K
be the set of those
f F such that Q
f
=

/ domf and Q
f
K has just one member for every K /. Set
A
K
= a : a A, there is some f F
K
such that ,= c
f
a.
I claim that a b ,= for any a, b A
K
. PPP There are f, g F
K
such that a c
f
,= , b c
g
,= . Now if
t domf domg,
f(t) = 1 t

/ g(t) = 1;
that is, f and g agree on domfdomg, and therefore there is a function h = fg, with domain domfdomg,
extending both. Of course h F. Dene z 0, 1
R
by setting z(t) = h(t) if t domh, 0 otherwise, so that
Q
z
= Q
h
= Q
f
Q
g
and z extends both f and g. Note that Q
z

/, and #(Q
z
) #(Q
f
) +#(Q
g
) = 2k.
Now we are supposing that c
f
is non-empty; say x c
f
. Then Q
x
Q
f
, so Q
x
meets every member of
/. If n 2k, then
J : J
n
, Q
x
J ,= J : J
n
, J

/ ,=
because for K /, J
n
either K J or K J = . But this means that
#(J : J
n
, Q
z
J ,= ) = #(J : J
n
, J
_
/ ,= )
#(J : J
n
, Q
x
J ,= ) n + 1
for every n 2k; while also
391Yb Kelleys theorem 573
#(J : J
n
, Q
z
J ,= ) #(Q
z
) 2k n + 1
for every n 2k. So z X. As z extends both f and g, z c
f
c
g
a b and a b ,= . QQQ
Thus every A
K
is linked.
(iii) For every non-zero a A there is some k N, / K
k
such that a A
K
. PPP Take any x a.
Because a is a relatively open subset of X, there is an open G 0, 1
R
such that a = G X; now, by the
denition of the topology of 0, 1
R
, there is an f F such that x H
f
G, so that ,= c
f
a.
Set k = #(Q
f
). Then we can choose a disjoint family K
t
)
tQ
f
in 1 such that, for each t, K
t
domf = t
and, for every J

n2k

n
, either K
t
J or K
t
J = . Set / = K
t
: t Q
f
; then / K
k
, f F
K
and a A
K
. QQQ
But

kN
K
k
is countable, being a subset of the family of nite sets in the countable set 1, so the A
K
form a countable family of linked sets covering A 0, and A is -linked.
391X Basic exercises For this series of exercises, I will call a Boolean algebra A chargeable if there
is a strictly positive nitely additive functional : A [0, [.
(a) Show that a chargeable Boolean algebra is ccc, so is Dedekind complete i it is Dedekind -complete.
(b) Show (i) that any subalgebra of a chargeable Boolean algebra is chargeable (ii) that a countable
simple product of chargeable Boolean algebras is chargeable (iii) that any free product of chargeable Boolean
algebras is chargeable.
(c) (i) Let A be a Boolean algebra with a chargeable order-dense subalgebra. Show that A is chargeable.
(ii) Show that the Dedekind completion of a chargeable Boolean algebra is chargeable.
(d) (i) Show that the algebra of open-and-closed subsets of 0, 1
I
is chargeable for any set I. (ii) Show
that the regular open algebra of R is chargeable.
(e) (i) Show that any principal ideal of a chargeable Boolean algebra is chargeable. (ii) Let A be a
chargeable Boolean algebra and 1 an order-closed ideal of A. Show that A/1 is chargeable.
>>>(f ) Show that a Boolean algebra is chargeable i it is isomorphic to a subalgebra of a measurable
algebra. (Hint: if A is a Boolean algebra and : A [0, [ is a strictly positive additive functional, set
(a, b) = (a .b) for a, b A. Show that is a metric on A and that the metric completion of A under
is a measurable algebra under the natural operations (cf. 324O).)
(g) Explain how to use the Hahn-Banach theorem to prove 391G directly, without passing through 391F.
(Hint: S(B) can be regarded as a subspace of S(A).)
>>>(h) Take X = 0, 1, 2, 3, A = TX, A = 0, 1, 0, 2, 0, 3, 1, 2, 3. Show that the intersection
number of A is
3
5
. (Hint: use 391I.) Show that if a
0
, . . . , a
n
are distinct members of A then there is a set
J 0, . . . , n, with #(J)
2
3
(n + 1), such that inf
jJ
a
j
,= 0.
(i) Let A be a Boolean algebra. For non-empty A A 0 write (A) for the intersection number of A.
Show that for any non-empty A A 0, (A) = sup(I) : I is a non-empty nite subset of A.
(j) (i) Show that any subalgebra of a -linked Boolean algebra is -linked. (ii) Show that if A is a Boolean
algebra with a -linked order-dense subalgebra, then A is -linked. (iii) Show that the simple product of
a countable family of -linked Boolean algebras is -linked. (iv) Show that a principal ideal of a -linked
Boolean algebra is -linked.
391Y Further exercises (a) Show that in 391D and 391K we can replace weakly (, )-distributive
by weakly -distributive.
(b) Show that TN is -linked and chargeable but that the quotient algebra TN/[N]
<
is not ccc, therefore
neither -linked nor chargeable.
574 Measurable algebras 391Yc
(c) (i) Show that if X is a separable topological space, then its regular open algebra is chargeable. (ii) Let
X
i
)
iI
be any family of topological spaces with chargeable regular open algebras. Show that their product
has a chargeable regular open algebra.
(d) Show that any -linked Boolean algebra can have cardinal at most c. (Hint: take a sequence A
n
)
nN
of linked sets covering A 0. Set A

n
= a : b A
n
, b a. Show that if a , b there is an n such that
a A

n
and b / A

n
.)
(e) Show that the free product of c or fewer -linked Boolean algebras is -linked. (Hint: let A
i
)
iI
be
a family of -linked Boolean algebras, where I R. For each i I, let A
in
)
nN
be a sequence of linked
sets with union A
i
0. For q
0
< q
1
< . . . < q
r
Q and n
1
, . . . , n
r
N let B
q,n
be the set of elements
expressible as inf
iJ

i
(a
i
) where J I [q
0
, q
r
[ is a set meeting each interval [q
k1
, q
k
[ in at most one point
and a
i
A
i,n
k
if i J [q
k1
, q
k
[; show that B
q,n
is linked.)
(f ) Let X be the space of 391N. (i) Show that X is closed in 0, 1
R
, therefore compact. (ii) Show that
the regular open algebra of X is -linked but not chargeable.
(g) Let A be the algebra of 391N. Show that A is not weakly (, )-distributive. (Hint: show that for
any t R there is a strictly decreasing sequence t
i
)
iN
, converging to t, such that x x(t
i
))
iN
is a
surjection from X onto 0, 1
N
, and hence that the algebra of open-and-closed subsets of 0, 1
N
can be
regularly embedded into A; now use 316Xq and 316Xm.)
(h) If A is a Boolean algebra, a set A A is m-linked, where m 2, if a
1
a
2
. . . a
m
,= 0 for all
a
1
, . . . , a
m
A; and A is -m-linked if A 0 can be expressed as the union of a sequence of m-linked
sets. (Thus linked is 2-linked and -linked is -2-linked.) Show that the algebra of 391N is -m-linked
for every m 2. (Hint: in part (c) of the proof, replace each 2k by mk.)
391Z Problem Must a Dedekind complete -linked weakly (, )-distributive Boolean algebra be
measurable?
391 Notes and comments By the standards of this volume, this is an easy section; I note that I have
hardly called on anything after Chapter 32, except for a reference to the construction S(A) in 361. I do
ask for a bit of functional analysis (the Hahn-Banach theorem) in 391E. Kelleys criterion (391J) is a little
unsatisfying. It is undoubtedly useful (see part (b) of the proof of 391N, or 392F below), but at the same
time the structure of the criterion a special sequence of subsets of A is rather close to the structure of
the conclusion; after all, one is, or can be represented by, a function from A 0 to N, while the other is a
function from A to R. Also the actual intersection number of a family A A 0 can be hard to calculate;
as often as not, the best method is to look at the additive functionals on A (see 391Xh).
I take the trouble to show that Gaifmans example (391N) is -linked in order to show that even conditions
very much stronger than ccc are not sucient to guarantee the existence of suitable functionals. (See also
391Yh.) For other examples see Comfort & Negrepontis 82. But I note that none of the standard
examples is weakly (, )-distributive (see 391Yg), and I believe it is open, in the formal sense, whether any
weakly (, )-distributive -linked Boolean algebra must carry a strictly positive nitely additive functional
(391Z). But I conjecture that the answer is no.
Since every -linked algebra has cardinal at most c (391Yd), not every measurable algebra is -linked.
In fact it is known that a measurable algebra of cardinal c or less is -m-linked (391Yh) for every m 2
(Dow & Stepr ans 93).
392D Submeasures 575
392 Submeasures
In 391 I looked at what we can deduce if a Boolean algebra carries a strictly positive nitely additive
functional. There are important contexts in which we nd ourselves with a subadditive, rather than additive,
functional, and this is what I wish to investigate here. It turns out that, once we have found the right
hypotheses, such functionals can also provide a criterion for measurability of an algebra (392J). The argument
runs through a new idea, using a result in nite combinatorics (392D).
392A Denition Let A be a Boolean algebra. A submeasure on A is a functional : A [0, [ such
that
(a b) a +b for all a, b A;
a b whenever a b;
0 = 0.
(In this context I do not allow as a value of a submeasure.) Any positive nitely additive functional is
a submeasure (326Ba, 326Bf).
392B Denitions Let A be a Boolean algebra and : A [0, [ a submeasure. Then
(a) is strictly positive if a > 0 for every a ,= 0;
(b) is exhaustive if lim
n
a
n
= 0 for every disjoint sequence a
n
)
nN
in A;
(c) is uniformly exhaustive if for every > 0 there is an n N such that there is no disjoint family
a
0
, . . . , a
n
with a
i
for every i n.
392C Proposition Let A be a Boolean algebra.
(a) If there is an exhaustive strictly positive submeasure on A, then A is ccc.
(b) A uniformly exhaustive submeasure on A is exhaustive.
(c) Any positive linear functional on A is a uniformly exhaustive submeasure.
proof These are all elementary. If : A [0, [ is an exhaustive strictly positive submeasure, and a
i
)
iI
is a disjoint family in A 0, then i : a
i
2
n
must be nite for each n, so I is countable. (Cf. 322G.)
If : A [0, [ is a uniformly exhaustive submeasure and a
n
)
nN
is disjoint in A, then i : a
i
2
n
is
nite for each n, so lim
i
a
i
= 0. If : A [0, [ is a positive linear functional, and > 0, then take
n
1

1; if a
0
, . . . , a
n
are disjoint, then

n
i=0
a
i
1, so min
in
a
i
< .
392D Lemma Suppose that k, l, m N are such that 3 k l m and 18mk l
2
. Let L, M be sets
of sizes l, m respectively. Then there is a set R M L such that (i) each vertical section of R has just
three members (ii) #(R[E]) #(E) whenever E [M]
k
; so that for every E [M]
k
there is an injective
function f : E L such that (x, f(x)) R for every x E.
recall that [M]
k
= I : I M, #(I) k (3A1J).
proof (a) We need to know that n! 3
n
n
n
for every n N; this is immediate from the inequality

n
i=2
lni
_
n
1
lnxdx = nlnn n + 1 for every n 2.
(b) Let be the set of those R M L such that each vertical section of R has just three members, so
that
#() = #([L]
3
)
m
=
_
l!
3!(l3)!
_
m
.
(I write [X]
j
for the set of subsets of X with j members.) Let us regard as a probability space with the
uniform probability.
If F [L]
n
, where 3 n k, and x M, then
Pr(R[x] F) =
#([F]
3
)
#([L]
3
)
(because R[x] is a random member of [L]
3
)
576 Measurable algebras 392D
=
n(n1)(n2)
l(l1)(l2)

n
3
l
3
.
So if E [M]
n
and F [L]
n
, then
Pr(R[E] F) =

xE
Pr(R[x] F)
(because the sets R[x] are chosen independently)

n
3n
l
3n
.
Accordingly
Pr(there is an E M such that #(R[E]) < #(E) k)
Pr(there is an E M such that #(R[E]) #(E) k)
= Pr(there is an E M such that 3 #(R[E]) #(E) k)
(because if E ,= then #(R[E]) 3)

n=3

E[M]
n

F[L]
n
Pr(R[E] F)
k

n=3
#([M]
n
)#([L]
n
)
n
3n
l
3n
=
k

n=3
m!
n!(mn)!
l!
n!(ln)!
n
3n
l
3n

k

n=3
m
n
l
n
n
3n
n!n!l
3n

k

n=3
m
n
n
n
3
2n
l
2n
(using (a))
=
k

n=3
_
9mn
l
2
_
n

n=3
1
2
n
< 1.
There must therefore be some R such that #(R[E]) #(E) whenever E M and #(E) k.
(c) If now E [M]
k
, the restriction R
E
= R (E L) has the property that #(R
E
[I]) #(I) for
every I E. By Halls Marriage Lemma (3A1K) there is an injective function f : E L such that
(x, f(x)) R
E
R for every x E.
Remark Of course this argument can be widely generalized; see references in Kalton & Roberts 83.
392E Lemma Let A be a Boolean algebra and : A [0, [ a uniformly exhaustive submeasure.
Then for any ]0, 1] the set A = a : a has intersection number greater than 0.
proof (a) If 1 = 0 this is trivial, so we may assume that 1 > 0; since neither the hypothesis nor
the conclusion is aected if we multiply by a positive scalar, we may suppose that 1 = 1. Because
is uniformly exhaustive, there is an r 1 such that whenever c
i
)
iI
is a disjoint family in A then
#(i : c
i
>
1
5
) r, so that

iI
c
i
r +
1
5
#(I). Set = /5r, =
1
74

2
, so that

1
18
( )
2

1
18
(
2
2) = 4.
(b) Let a
i
)
iI
be a non-empty nite family in A. Let m be any multiple of #(I) greater than or equal
to 1/. Then there are integers k, l such that
3
k
m
4
1
18
( )
2
,
l
m
,
in which case
3 k l m, 18mk m
2
( )
2
l
2
.
392G Submeasures 577
(c) Take a set M of the form I S where #(S) = m/#(I), so that #(M) = m. For x = (i, s) M set
d
x
= a
i
. Let L be a set with l members. By 392D, there is a set R M L such that every vertical section
of R has just three members and whenever E [M]
k
there is an injective function f
E
: E L such that
(x, f
E
(x)) R for every x E.
For E M set
b
E
= inf
xE
d
x
\ sup
xM\E
d
x
,
so that b
E
)
EM
is a partition of unity in A. For x M, j L set
c
xj
= supb
E
: x E [M]
k
, f
E
(x) = j.
If x, y are distinct members of M and j L then
c
xj
c
yj
= supb
E
: x, y E [M]
k
, f
E
(x) = f
E
(y) = j = 0,
because every f
E
is injective. Set
m
j
= #(x : x M, c
xj
,= 0)
for each j L. Note that c
xj
= 0 if (x, j) / R, so

jL
m
j
#(R) = 3m.
We have

xM
c
xj
r +
1
5
m
j
for each j, by the choice of r; so

xM,jL
c
xj
rl +
1
5

jL
m
j
rl +
3
5
m
(r +
3
5
)m =
4
5
m < m
by the choice of l and . There must therefore be some x M such that
(sup
jL
c
xj
)

jL
c
xj
< d
x
,
and d
x
cannot be included in
sup
jL
c
xj
= supb
E
: x E [M]
k
.
But as supb
E
: x E M is just d
x
, there must be an E M, of cardinal greater than k, such that
b
E
,= 0.
Recall now that M = I S, and that
k 3m = 3#(I)#(S).
The set J = i : s, (i, s) E must therefore have more than 3#(I) members, since E J S. But also
d
(i,s)
= a
i
for each (i, s) E, so that inf
iJ
a
i
b
E
,= 0.
(d) As a
i
)
iI
is arbitrary, the intersection number of A is at least 3 > 0.
392F Theorem Let A be a Boolean algebra with a strictly positive uniformly exhaustive submeasure.
Then A has a strictly positive nitely additive functional.
proof If A = 0 this is trivial. Otherwise, let : A [0, [ be a strictly positive uniformly exhaustive
submeasure. For each n, A
n
= a : a 2
n
1 has intersection number greater than 0, and

nN
A
n
=
A 0 because is strictly positive; so A has a strictly positive nitely additive functional, by Kelleys
theorem (391J).
392G Since positive additive functionals are uniformly exhaustive submeasures, the condition of this
theorem is necessary as well as sucient. Thus we have a description, in terms of submeasures, of a condition
equivalent to one part of the criterion for measurability of an algebra in 391D. The language of submeasures
also provides a formulation of another part of this criterion, as follows.
578 Measurable algebras 392G
Denition Let A be a Boolean algebra. A Maharam submeasure or continuous outer measure on A
is a submeasure : A [0, [ such that lim
n
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence
in A with inmum 0.
392H Lemma Let A be a Boolean algebra and a Maharam submeasure on A.
(a) is sequentially order-continuous.
(b) is countably subadditive, that is, whenever a
n
)
nN
is a sequence in A and a A is such that
a = sup
nN
a a
n
, then a

n=0
a
n
.
(c) If A is Dedekind -complete, then is exhaustive.
proof (a) (Of course is an order-preserving function, by the denition of submeasure; so we can apply the
ordinary denition of sequentially order-continuous in 313Hb.) (i) If a
n
)
nN
is a non-decreasing sequence in
A with supremum a, then a
n
\ a)
nN
is a non-increasing sequence with inmum 0, so lim
n
(a
n
\ a) = 0;
but as
a
n
a a
n
+(a \ a
n
)
for every n, it follows that a = lim
n
a
n
. (ii) If a
n
)
nN
is a non-increasing sequence in A with inmum
a, then
a a
n
a +(a
n
\ a) a
as n .
(b) Set b
n
= sup
in
a a
i
; then b
n

n
i=0
a
i
for each n (inducing on n), so that
a = lim
n
b
n

i=0
a
i
.
(c) If a
n
)
nN
is a disjoint sequence in A, set b
n
= sup
in
a
i
for each n; then inf
nN
b
n
= 0, so
limsup
n
a
n
lim
n
b
n
= 0.
392I Proposition Let A be a Dedekind -complete Boolean algebra and a strictly positive Maharam
submeasure on A. Then A is ccc, Dedekind complete and weakly (, )-distributive.
proof By 392Hc, is exhaustive; by 392Ca, A is ccc; by 316Fa, A is Dedekind complete.
Now suppose that we have a sequence A
n
)
nN
of non-empty downwards-directed subsets of A, all with
inmum 0. Let A be the set
a : a A, n N a
t
A
n
such that a
t
a.
By 316Fc and 392Ha, is order-continuous, so inf
aA
n
a = 0 for each n. Given > 0, we can choose
a
n
)
nN
such that a
n
A
n
and a
n
2
n
for each n; now a = sup
nN
a
n
A and a

n=0
a
n
2.
Thus inf
aA
a = 0. Since is strictly positive, inf A = 0. As A
n
)
nN
is arbitrary, A is weakly (, )-
distributive.
392J Theorem Let A be a Boolean algebra. Then it is measurable i it is Dedekind -complete and
carries a uniformly exhaustive strictly positive Maharam submeasure.
proof If A is measurable, it surely satises the conditions, since any totally nite measure on A is also
a uniformly exhaustive strictly positive Maharam submeasure. If A satises the conditions, then it is
weakly (, )-distributive, by 392I, and carries a strictly positive nitely additive functional, by 392F; so is
measurable, by 391D.
392X Basic exercises (a) Show that the rst two clauses of the denition 392A can be replaced by
a (a b) a +b whenever a b = 0.
(b) Let A be any Boolean algebra and a submeasure on A. (i) Show that the following are equiveridical:
() is order-continuous; () whenever A A is non-empty, downwards-directed and has inmum 0, then
inf
aA
a = 0. (ii) Show that in this case is exhaustive. (Hint: if a
n
)
nN
is disjoint, then

nN
b : b a
i
for every i n has inmum 0.)
393 intro. The Control Measure Problem 579
>>>(c) Let A be the nite-conite algebra on an uncountable set (316Yk). (i) Set
1
0 = 0,
1
a = 1
for a A 0. Show that
1
is a strictly positive Maharam submeasure but is not exhaustive. (ii) Set

2
a = 0 for nite a, 1 for conite a. Show that
2
is a uniformly exhaustive Maharam submeasure but is
not order-continuous.
>>>(d) Let A be a Boolean algebra and a submeasure on A. Set I = a : a = 0. Show that (i) I is
an ideal of A (ii) there is a submeasure on A/I dened by setting a

= a for every a A (iii) if is

exhaustive, so is (iv) if is uniformly exhaustive, so is (v) if is a Maharam submeasure, I is a -ideal
(vi) if is a Maharam submeasure and A is Dedekind -complete, is a Maharam submeasure.
(e) Let A be a Dedekind complete Boolean algebra and an order-continuous submeasure on A. Show
that has a unique support a A such that A
a
is strictly positive and A
1\a
is identically zero.
(f ) Let A be a Dedekind -complete Boolean algebra and a uniformly exhaustive Maharam submeasure
on A. Show that there is a non-negative countably additive functional on A such that a : a = 0 = a :
a = 0. (Hint: 392Xd(vi).)
392 Notes and comments Much of this section is a matter of generalizing earlier arguments. Thus 392C
and 392H ought by now to be very easy, while 392I uses the methods of 322F-322G, and 392Xb recalls the
elementary theory of -additive functionals.
The new ideas are in the combinatorics of 392D-392E. I have cast 392D in the form of an argument
in probability theory. Of course there is nothing here but simple counting, since the probability measure
simply puts the same mass on each point of , and every statement of the form Pr(R. . . ) . . . is just
a matter of counting the elements R of with the given property. But I think many of us nd that the
probabilistic language makes the calculations more natural; in particular, we can use intuitions associated
with the notion of independence of events. Indeed I strongly recommend the method. It has been used to
very great eect in the last fty years in a wide variety of combinatorial problems.
392F/392J constitute the Kalton-Roberts theorem (Kalton & Roberts 83).
It is not known whether every exhaustive submeasure is uniformly exhaustive; this is the Control Measure
Problem, which I will treat in the next section.
393 The Control Measure Problem
I come now to a discussion of a classic problem of measure theory. Its importance derives to a great extent
from the variety of forms in which it appears, and this section is devoted primarily to a description of some of
these forms, with the arguments to show that they are indeed all the same problem, in that a solution to one
of the questions will provide solutions to all the others. The syntax of the exposition seems to be simplest
if I present each formulation as a statement CM
n
; the corresponding question being is CM
n
true?, and
the proof that all the questions are really the same question becomes a proof that CM
m
CM
n
for all
m, n.
The propositions CM

are listed in 393A, 393H, 393J, 393L and 393P, while the arguments that they are
equiveridical form the rest of the material down to and including 393R. CM
1
-CM
3B
all involve submeasures
in one way or another, and much of the section amounts to a theory of submeasures. On the way I mention a
description of the open-and-closed algebra of 0, 1
N
(393F). The propositions CM
4
, CM
5
and CM
6
concern
topologies on Boolean algebras and L
0
spaces, and vector measures. They can be expressed more or less
adequately with very little of the surrounding theories, but for completeness I include a basic theorem on
vector measures (393S). At the end of the section I present two of the many examples of submeasures which
have been described.
580 Measurable algebras 393A
393A The problem The language I introduced in the last section is already sucient for more than
one formulation of the problem. Consider the statements
(CM
1
) If A is a Dedekind complete Boolean algebra and is a strictly positive Maharam submeasure
on A, then A is measurable.
(CM
t
1
) Let A be a Dedekind -complete Boolean algebra and : A [0, [ a Maharam submea-
sure. Then there is a non-negative countably additive functional : A [0, [ such that, for
a A, a = 0 a = 0.
(CM
tt
1
) Let X be a set, a -algebra of subsets of X, and : [0, [ a Maharam submeasure.
Then there is a totally nite measure , with domain , such that, for E , E = 0
E = 0.
(CM
2
) Every exhaustive submeasure is uniformly exhaustive.
(CM
t
2
) Every exhaustive submeasure on the algebra of open-and-closed subsets of 0, 1
N
is uni-
formly exhaustive.
The following lemmas assemble the new ideas we need in order to prove that these statements are equiv-
eridical.
393B Lemma Let A be a Boolean algebra and a strictly positive submeasure on A.
(a) We have a metric on A dened by the formula
(a, b) = (a .b)
for all a, b A.
(b) The Boolean operations , , ., \ , and the function : A R, are all uniformly continuous for
.
(c) The metric space completion (

A, ) of (A, ) is a Boolean algebra under the natural continuous

extensions of the Boolean operations, and has a unique continuous extension to

A which is again a
strictly positive submeasure.
(d) If is exhaustive, then

A is Dedekind complete and ccc and is a Maharam submeasure.
(e) If is additive, then (

A, ) is a totally nite measure algebra.

proof (a)-(b) This is just a generalization of 323A-323B; essentially the same formulae can be used. For
the triangle inequality for , we have a .c (a .b) (b .c), so
(a, c) = (a .c) (a .b) +(b .c) = (a, b) +(b, c).
For the uniform continuity of the Boolean operations, we have
(b c) .(b
t
c
t
) (b .b
t
) (c .c
t
)
so that
(b c, b
t
c
t
) (b, b
t
) +(c, c
t
)
for each of the operations = , , \ and .. For the uniform continuity of the function itself, we have
b c +(b \ c) c +(b, c),
so that [b c[ (b, c).
(c) AA is a dense subset of

A

A, so the Boolean operations on A, regarded as uniformly continuous

functions from AA to A

A, have unique extensions to continuous binary operations on

A (3A4G). If we
look at
A = (a, b, c) : a .(b .c) = (a .b) .c,
this is a closed subset of

A

A

A, because the maps (a, b, c) a .(b .c), (a, b, c) (a .b) .c are
continuous and the topology of

A is Hausdor; since A includes the dense set A A A, it is the whole of

A, that is, a .(b .c) = (a .b) .c for all a, b, c

A. All the other identities we need to show that

A is a Boolean algebra can be conrmed by the same method. Of course A is now a subalgebra of

A.
393C The Control Measure Problem 581
Because : A [0, [ is uniformly continuous, it has a unique continuous extension :

A [0, [. We
have
0 = 0, a (a b) a + b, a = (a, 0)
for every a, b A and therefore for every a, b

A, so is a submeasure on

A, and
a = 0 = (a, 0) = 0 = a = 0,
so is strictly positive.
(d) Now suppose that is exhaustive.
(i) The point is that any non-increasing sequence a
n
)
nN
in

A is a Cauchy sequence for the metric .
PPP Let > 0. For each n N, choose b
n
A such that (a
n
, b
n
) 2
n
, and set c
n
= inf
in
b
i
. Then
(a
n
, c
n
) = (inf
in
a
i
, inf
in
b
i
)

n
i=0
(a
i
, b
i
) 2
for every n. Choose n(k))
kN
inductively so that, for each k N, n(k + 1) n(k) and
(c
n(k)
\ c
n(k+1)
) sup
in(k)
(c
n(k)
\ c
i
) .
Then c
n(k)
\ c
n(k+1)
)
kN
is a disjoint sequence in A, so
limsup
k
sup
in(k)
(a
n(k)
, a
i
) 4 + limsup
k
sup
in(k)
(c
n(k)
, c
i
)
= 4 + limsup
k
sup
in(k)
(c
n(k)
\ c
i
)
4 + limsup
k
(c
n(k)
\ c
n(k+1)
) + = 5.
As is arbitrary, a
n
)
nN
is Cauchy. QQQ
(ii) It follows that

A is Dedekind -complete. PPP If a
n
)
nN
is any sequence in

A, b
n
)
nN
=
inf
in
a
i
)
nN
is a Cauchy sequence with a limit b

A. For any k N,
(b \ a
k
) = lim
n
(b
n
\ a
k
) = 0,
so b a
k
, because is strictly positive. While if c

A is a lower bound for a
n
: n N, we have c b
n
for every n, so
(c \ b) = lim
n
(c \ b
n
) = 0
and c b. Thus b = inf
nN
a
n
; as a
n
)
nN
is arbitrary,

A is Dedekind -complete (314Bc). QQQ
(iii) We also nd that is a Maharam submeasure, because if a
n
)
nN
is a non-increasing sequence in

A with inmum 0, it must have a limit a which (as in (ii) just above) must be its inmum, that is, a = 0;
consequently
lim
n
a
n
= a = 0.
(iv) It follows at once that is exhaustive (392Hc), so that

A is ccc (392Ca) and Dedekind complete
(316Fa).
(e) If is additive, then (being nite-valued) it must be exhaustive, so that

A is Dedekind complete and
is a Maharam submeasure. But now observe that the function
(a, b) (a b) + (a b) a b :

A

A R
is continuous (by (b)) and zero on the dense set A A, so is zero everywhere on

A

A. In particular,
(a b) = a+ b whenever a b = 0 in

A, and is additive. Since it is also countably subadditive (392Hb)
it is countably additive, and (

A, ) is a totally nite measure algebra.

393C Lemma Let A be a Boolean algebra and an exhaustive submeasure on A. Then there are a
Dedekind complete Boolean algebra B, a strictly positive Maharam submeasure on B, and a Boolean
homomorphism : A B such that = . If is additive, then (B, ) is a totally nite measure algebra.
582 Measurable algebras 393C
proof (a) Set I = a : a = 0. Then I A; let B
0
be the Boolean quotient algebra A/I (312K). If a
1
,
a
2
A and a

1
a

2
in B, then a
1
\ a
2
I, so
a
1
a
2
+(a
1
\ a
2
) = a
2
.
So if a

1
= a

2
we must have a
1
= a
2
; there is therefore a functional
0
: B
0
[0, [ such that
0
a

= a
for every a A. We have just seen that
0
a

1

0
a

2
if a

1
a

2
; now of course 0

= 0 = 0,

0
(a

1
a

2
) =
0
(a
1
a
2
)

= (a
1
a
2
) a
1
+a
2
=
0
a

1
+
0
a

2
for all a
1
, a
2
A, so
0
is a submeasure. If b B 0, then b = a

where a / I, so
0
b = a > 0; thus
0
is strictly positive.
If b
n
)
nN
is a disjoint sequence in B
0
, express it as a

n
)
nN
and set a
n
= a
n
\ sup
i<n
a
i
for each n, so
that a
n
)
nN
is disjoint and
lim
n

0
b
n
= lim
n
a
n
= 0
because is exhaustive. So
0
is exhaustive.
(b) Let B be the metric completion of B
0
under the metric associated with
0
(393Bc) and the asso-
ciated strictly positive submeasure; because
0
is exhaustive, B is Dedekind complete and is a Maharam
submeasure (393Bd).
Writing a = a

, interpreted as a member of B, for each a A, is a Boolean homomorphism and

a =
0
a

= a
for every a A.
(c) If is additive, then (a
1
a
2
)+(a
1
a
2
) = a
1
+a
2
for all a
1
, a
2
A, so
0
(b
1
b
2
)+
0
(b
1
b
2
) =

0
b
1
+
0
b
2
for all b
1
, b
2
B
0
, and
0
is additive; by 393Bd, (B, ) is a totally nite measure algebra. This
completes the proof.
393D Denition Let A be a Boolean algebra and ,
t
two submeasures on A. Then is absolutely
continuous with respect to
t
if for every > 0 there is a > 0 such that a whenever
t
a .
393E Lemma Let A be a Dedekind -complete Boolean algebra and ,
t
two Maharam submeasures
on A such that a = 0 whenever
t
a = 0. Then is absolutely continuous with respect to
t
.
proof ??? Otherwise, we can nd a sequence a
n
)
nN
in A such that
t
a
n
2
n
for every n, but =
inf
nN
a
n
> 0. Set b
n
= sup
in
a
i
for each n, b = inf
nN
b
n
. Then
t
b
n

i=n
2
i
2
n+1
for each n
(392Hb), so
t
b = 0; but b
n
for each n, so b (392Ha), contrary to the hypothesis. XXX
393F Lemma Let B be the algebra of open-and-closed subsets of 0, 1
N
. Then a Boolean algebra A
is isomorphic to B i it is atomless, countable and not 0.
proof (a) I must check that B has the declared properties. The point is that it is the subalgebra B
t
of TX
generated by b
i
: i N, where I write X = 0, 1
N
, b
i
= x : x X, x(i) = 1. PPP Of course b
i
and its
complement x : x(i) = 0 are open, so b
i
B for each i, and B
t
B. In the other direction, the open
cylinder sets of X are all of the form c
z
= x : x(i) = z(i) for every i J, where J I and z 0, 1
J
;
now
c
z
= X

z(i)=1
b
i

z(i)=0
b
i
B
t
.
If b B then b is expressible as a union of such cylinder sets, because it is open; but also it is compact, so
is the union of nitely many of them, and must belong to B
t
. Thus B = B
t
, as claimed. QQQ
Because B = B
t
is generated by a countable set, it is countable (331Gc). Next, if b B is non-empty,
there are a nite J I and a z 0, 1
J
such that c
z
b; now if we take any i N J, and look at the
two extensions z
0
, z
1
of z to J i, then c
z
0
, c
z
1
are disjoint non-empty members of B included in b. So
B is atomless.
(b) Now suppose that A is another algebra with the same properties. Enumerate A as a
n
)
nN
. For each
n N let B
n
be the nite subalgebra of B generated by b
i
: i < n (so that B
0
= 0, 1). Then B
n
)
nN
393G The Control Measure Problem 583
is an increasing sequence of subalgebras of B with union B; also b b
n
, b \ b
n
are non-zero for every n N,
b B
n
.
Choose nite subalgebras A
n
A and isomorphisms
n
: A
n
B
n
as follows. A
0
= 0, 1,
0
0 = 0,

0
1 = 1. Given A
n
and
n
, let A
n
be the set of atoms of A
n
. For a A
n
, choose a
t
A such that
if a
n
a, a
n
\ a are both non-zero, then a
t
= a
n
a;
otherwise, a
t
a is any element such that a
t
, a \ a
t
are both non-zero.
(This is where I use the hypothesis that A is atomless.) Set a
t
n
= sup
aA
n
a
t
. Then we see that a a
t
n
, a \ a
t
n
are non-zero for every a A
n
and therefore for every non-zero a A
n
, that is, that
supa : a A
n
, a a
t
n
= 0, infa : a A
n
, a a
t
n
= 1.
Let A
n+1
be the subalgebra of A generated by A
n
a
t
n
. Since we have
supb : b B
n
, b b
n
= 0, infb : b B
n
, b b
n
= 1,
there is a (unique) extension of
n
: A
n
B
n
to a homomorphism
n+1
: A
n+1
B
n+1
such that

n+1
a
t
n
= b
n
(312N). Since we similarly have an extension of
1
n
to a homomorphism from B
n+1
to
A
n+1
with b
n
= a
t
n
, and since
n+1
,
n+1
must be the respective identity homomorphisms,
n+1
is an
isomorphism, and the induction continues.
Since
n+1
extends
n
for each n, these isomorphisms join together to give us an isomorphism
:

nN
A
n

nN
B
n
= B.
Observe next that the construction ensures that a
n
A
n+1
for each n, since a
n
a is either 0 or a or a
t
n
a
for every a A
n
, and in all cases belongs to A
n+1
. So

nN
A
n
contains every a
n
and (by the choice of
a
n
)
nN
) must be the whole of A. Thus : A B witnesses that A

= B.
393G Theorem If one of CM
1
, CM
t
1
, CM
tt
1
, CM
2
, CM
t
2
is true, so are the others.
proof CM
1
CM
t
1
Assume that CM
1
is true, and that A and are as in the statement of CM
t
1
. Set
I = a : a = 0; because is countably subadditive, I is a -ideal, B = A/I is Dedekind -complete
and the quotient map a a

is sequentially order-continuous (313Qb, 314C). Now we have a functional

: B [0, [ dened by saying that a

= a for every a A, and is a strictly positive submeasure

(as in part (a) of the proof of 393C). In fact is a Maharam submeasure. PPP Suppose that b
n
)
nN
is a
non-increasing sequence in B with inmum 0. Choose a
n
A such that a

n
= b
n
for each n, and set
c
n
= inf
in
a
i
inf
iN
a
i
;
then c
n
)
nN
is non-increasing and has inmum 0, while c

n
= b
n
for each n, so
lim
n
b
n
= lim
n
c
n
= 0.
As b
n
)
nN
is arbitrary, is a Maharam submeasure. QQQ
Now CM
1
tells us that B must be a measurable algebra, that is, carries a totally nite measure . Setting
a = a

for a A, we see that is a non-negative countably additive functional, and that

a = 0 a

= 0 a = 0,
as required.
CM
t
1
CM
tt
1
is trivial; allowing for the change in notation, CM
tt
1
is just a special case of CM
t
1
.
CM
tt
1
CM
1
Assume CM
tt
1
, and let A be a Dedekind complete Boolean algebra with a strictly positive
Maharam submeasure . Then we can express A as /1 for some -algebra of subsets of a set X and a
-ideal 1 of (314M). Set E = E

for E ; then X, , satisfy the conditions of CM

tt
1
, so there is a
totally nite measure on X, with domain , such that
E = 0 E = 0 E

= 0 E

= 0.
Consequently we can identify A with the measure algebra of (X, , ), and A is measurable.
CM
1
CM
2
Now suppose that CM
1
is true, and that is an exhaustive submeasure on a Boolean
algebra A. By 393B, we can nd a Dedekind complete Boolean algebra B with a strictly positive Maharam
584 Measurable algebras 393G
submeasure and a Boolean homomorphism : A B such that a = a for every a A. CM
1
assures
us that B is measurable; let be a totally nite measure on B.
Let > 0. By 393E, is absolutely continuous with respect to , so there is a > 0 such that b
whenever a . Take n
1

1. If a
0
, . . . , a
n
are disjoint in A, then a
0
, . . . , a
n
are disjoint in B,
so there is some i such that a
i
, in which case a
i
= a
i
. As is arbitrary, is uniformly
exhaustive; as A and are arbitrary, CM
2
is true.
CM
2
CM
t
2
is trivial.
CM
t
2
CM
1
Suppose that CM
t
2
is true, that is, every exhaustive submeasure on the algebra B of open-
and-closed subsets of 0, 1
N
is uniformly exhaustive, and that is a strictly positive Maharam submeasure
on a Dedekind complete Boolean algebra A. Let E be the set of atoms of A, and set a = 1 \ supE, so that
the principal ideal A
a
is atomless, and A
a
is still a strictly positive Maharam submeasure.
??? Suppose, if possible, that A
a
is not uniformly exhaustive. Then there is a family a
ni
)
inN
in A
a
such that a
ni
)
in
is disjoint for each n but inf
inN
a
ni
> 0. There is a countable atomless subalgebra D
of A
a
containing every a
ni
. PPP For each d A
a
x on a d
t
d such that d
t
and d \ d
t
are both non-zero.
Dene D
n
)
nN
by setting
D
0
= 0, a a
ni
: i n N,
D
n+1
= D
n
d
1
d
2
: d
1
, d
2
D
n
a \ d : d D
n
d
t
: d D
n
.
Then every D
n
is countable, so D =

nN
D
n
is countable. Because D includes D
0
, it contains all the a
ni
;
also d
1
d
2
, 1 \ d
1
and d
t
1
belong to D for every d
1
, d
2
D, so D is an atomless subalgebra. QQQ
Because is an exhaustive submeasure on A, D is an exhaustive submeasure on D. Because every a
ni
belongs to D and inf
inN
a
ni
> 0, D is not uniformly exhaustive.
Let B be the algebra of open-and-closed subsets of 0, 1
N
. By 393F, there is a Boolean isomorphism
: B D. Set b = b for every b B; then (B, ) is isomorphic to (D, D), and, in particular, is an
exhaustive submeasure on B which is not uniformly exhaustive, which we are supposing to be impossible.
XXX
Thus A
a
has a uniformly exhaustive Maharam submeasure; since it is surely Dedekind complete, it is
measurable (392J). Let
1
be a totally nite measure on A
a
. Next, because lim
n
e
n
= 0 for any sequence
e
n
)
nN
of distinct elements of E, and e > 0 for every e E, E must be countable, and there is a summable
family
e
)
eE
of strictly positive real numbers. Setting c =
1
(c a) +

eE,ec

e
, we get a totally
nite measure on A, and A is measurable.
393H Variations on CM
2
The following modications of CM
2
are interesting because they display
some general properties of exhaustive submeasures. Consider the statements
(CM
3A
) If A is a Boolean algebra and is a non-zero exhaustive submeasure on A, then there is a
non-zero nitely additive functional on A such that 0 a a for every a A.
(CM
t
3A
) If A is a Boolean algebra and is a non-zero exhaustive submeasure on A, then there is
a non-zero non-negative nitely additive functional on A such that is absolutely continuous
with respect to .
(CM
3B
) If A is a Boolean algebra and is an exhaustive submeasure on A, then there is a non-
negative nitely additive functional on A such that is absolutely continuous with respect to
.
393I Proposition If one of CM
1
, . . . , CM
t
2
, CM
3A
, CM
t
3A
, CM
3B
is true, so are the others.
proof CM
1
CM
3A
& CM
3B
Assume CM
1
, and let be a non-zero exhaustive submeasure on a
Boolean algebra A. By 393C, there are a Dedekind complete Boolean algebra B, a strictly positive Maharam
submeasure on B, and a Boolean homomorphism : A B such that = . Since 1 = 1 ,= 0,
B ,= 0. CM
1
tells us that B is measurable; let be a strictly positive countably additive functional on B,
and set
1
= , so that
1
is a nitely additive functional on A. By 393E, is absolutely continuous with
respect to and is absolutely continuous with respect to ; it follows from the latter that is absolutely
continuous with respect to
1
.
393K The Control Measure Problem 585
Set =
1
2
1/ 1, and consider D = b : b B, b d. ??? If D is order-dense in B, there is a disjoint
C D such that supC = 1. Because B is measurable, it is ccc, and C is countable. If C is innite,
enumerate it as c
n
)
nN
; if it is nite, enumerate it as c
i
)
in
and set c
i
= 0 for i > n. In either case,
c
n
)
nN
is a disjoint sequence in C with supremum 1, so that
1

n=0
c
n

n=0
c
n
= 1 < 1
by the choice of , which is absurd. XXX
There is therefore a non-zero b B such that c > c for every non-zero c b. Set

2
a = (b a) (b a) a = a
for a A; then
2
is a non-negative nitely additive functional on A,
2
, and
2
1 = b ,= 0, so
2
is
non-zero.
As A and are arbitrary, CM
3A
and CM
3B
are both true.
(The statement of CM
3B
does not assume to be non-zero, but of course the case = 0 is trivial.)
CM
3A
CM
t
3A
is trivial.
CM
t
3A
CM
1
Assume CM
t
3A
, and let be a strictly positive Maharam submeasure on a Dedekind
complete Boolean algebra A. Note that A is ccc (392Ca). Now C = a : a A, A
a
is measurable is
order-dense in A. PPP Take any a A 0. Set
a
b = (a b) for b A. It is easy to check that
a
is a non-
zero Maharam submeasure on A; in particular, it is exhaustive. So there is a non-zero non-negative nitely
additive functional which is absolutely continuous with respect to
a
. If b
n
)
nN
is a non-increasing
sequence in A with inmum 0, then lim
n

a
b
n
= 0, so lim
n
b
n
= 0; accordingly is countably
additive (326Ga), therefore completely additive, because A is ccc (326L). Set I = b : b = 0; then I is an
order-closed ideal, so contains its supremum b
0
say. Since b = 0 whenever
a
b = 0, b
0
1 \ a, and c a,
where c = 1 \ b
0
. But A
c
I = 0, so A
c
is strictly positive, and witnesses that A
c
is measurable. Also
c = 1 ,= 0, so c is a non-zero member of C included in a. As a is arbitrary, C is order-dense. QQQ
By 391Cb, there is a function
1
such that (A,
1
) is a localizable measure algebra. Because A is ccc, A
is measurable (322G).
CM
3B
CM
1
Assume CM
3B
, and let A be a Dedekind complete Boolean algebra and a strictly
positive Maharam submeasure on A. By CM
3B
, there is a non-negative nitely additive functional on A
such that is absolutely continuous with respect to . But this means that
a = 0 = a = 0 = a = 0,
that is, is strictly positive. Also A is Dedekind complete and weakly (, )-distributive, by 392I. So A is
measurable, by 391D. As A and are arbitrary, CM
1
is true.
393J The rst published version of the Control Measure Problem (Maharam 47), in the form is
CM
1
true?, was a re-formulation of a question about topologies on Boolean algebras, which I now describe.
Consider the statement
(CM
4
) Let A be a ccc Dedekind complete Boolean algebra with a Hausdor topology T such that
(i) the Boolean operation : AA A is continuous at (0, 0) (ii) if a
n
)
nN
is a non-increasing
sequence in A with inmum 0, then a
n
)
nN
converges to 0 for T. Then A is measurable.
393K Proposition CM
4
is true i CM
1
, . . . , CM
3B
are true.
proof CM
1
CM
4
Assume CM
1
, and take A, T as in the statement of CM
4
.
(i) For any e A 0, there is a Maharam submeasure on A such that e > 0.
PPP() Choose a sequence G
n
)
nN
of neighbourhoods of 0, as follows. Because T is Hausdor, there is a
neighbourhood G
0
of 0 not containing e. Given G
n
, choose a neighbourhood G
n+1
of 0 such that G
n+1
G
n
and a b c G
n
whenever a, b, c G
n+1
. (Take neighbourhoods H, H
t
of 0 such that a b G
n
for a,
b H, b c H for b, c H
t
and set G
n+1
= H H
t
G
n
.) Dene
0
: A [0, 1] by setting
586 Measurable algebras 393K

0
a = 1 if a / G
0
,
= 2
n
if a G
n
G
n+1
,
= 0 if a

nN
G
n
.
Then whenever a
0
, . . . , a
r
A, n N and

r
i=0

0
a
i
< 2
n
, sup
ir
a
i
G
n
. To see this, induce on r. If
r = 0 then we have
0
a
0
< 2
n
so a
0
G
n+1
G
n
. For the inductive step to r 1, there must be a k r
such that

i<k

0
a
i
< 2
n1
,

k<in

0
a
i
< 2
n1
(allowing k = 0 or k = n, in which case one of the
sums will be zero). (If

r
i=0

0
a
i
< 2
n1
, take k = n; otherwise, take k to be the least number such that

k
i=0

0
a
i
2
n1
.) By the inductive hypothesis, and because 0 certainly belongs to G
n+1
, b = sup
i<k
a
i
and c = sup
k<ir
a
i
both belong to G
n+1
; but also
0
a
k
< 2
n
so a
k
G
n+1
. Accordingly, by the choice of
G
n+1
,
sup
ir
a
i
= b a
k
c
belongs to G
n
, and the induction continues.
() Set

1
a = inf

r
i=0

0
a
i
: a
0
, . . . , a
r
A, a = sup
ir
a
i

for every a A. It is easy to see that

1
(a b)
1
a+
1
b for all a, b A; also a G
n
whenever
1
a < 2
n
,
so, in particular,
1
e 1, because e / G
0
.
Set
a = inf
1
b : a e b e
for every a A. Then of course 0 a b whenever a b, and
0
1
0
0
0 = 0,
so 0 = 0. If a, b A and > 0, there are a
t
, b
t
such that a e a
t
e, b e b
t
e,
1
a
t
a + and

1
b
t
b +; so that (a b) e a
t
b
t
e and
(a b)
1
(a
t
b
t
)
1
a
t
+
1
b
t
a +b + 2.
As , a and b are arbitrary, is a submeasure. Next, if a
i
)
iN
is any non-increasing sequence in A with
inmum 0, a
i
e)
iN
is another, so converges to 0 for T. If n N there is an m such that a
i
e G
n
for
every i m, so that
a
i

1
(a
i
e)
0
(a
i
e) 2
n
for every i m. As n is arbitrary, lim
i
a
i
= 0; as a
i
)
iN
is arbitrary, is a Maharam submeasure.
Finally,
e =
1
e 1,
so e ,= 0. QQQ
(ii) Write C for the set of those c A such that A
c
is a measurable algebra. Then C is order-dense in A.
PPP Take any e A 0. By (i), there is a Maharam submeasure such that e > 0. Set I = a : a = 0,
a

= supI. Because A is ccc, there is a sequence a

n
)
nN
in I such that a

= sup
nN
a
n
; because is
countably subadditive, a

= 0. Consider c = e \ a

. Then A
c
is a strictly positive Maharam submeasure,
so CM
1
assures us that A
c
is measurable; also c ,= 0 because c = e ,= 0, so 0 ,= c e and c C. As e is
arbitrary, C is order-dense. QQQ
Now A is a ccc, Dedekind complete Boolean algebra in which the measurable principal ideals are order-
dense, so A is measurable, just as in the proof of CM
t
3A
CM
1
(393I)
CM
4
CM
1
Now assume that CM
4
is true, and that A is a Dedekind complete Boolean algebra with
a strictly positive Maharam submeasure . Then the associated metric on A (393B) denes a topology T
satisfying the conditions of CM
4
. PPP The continuity of is mentioned in 393B. If a
n
)
nN
is a non-increasing
sequence with inmum 0, then (a
n
, 0) = a
n
0 so a
n
0 for T. QQQ By CM
4
, A is measurable.
393N The Control Measure Problem 587
393L My own rst encounter with the Control Measure Problem was in the course of investigating
topological Riesz spaces. For various questions depending on this problem, see Fremlin 75. I give the
following as a sample.
(CM
5
) Let A be a ccc Dedekind complete Boolean algebra, and suppose that there is a Haus-
dor linear space topology T on L
0
(A) such that for every neighbourhood G of 0 there is a
neighbourhood H of 0 such that u G whenever v H and [u[ [v[. Then A is measurable.
393M Proposition Let A be a Dedekind -complete Boolean algebra and : A [0, [ a strictly
positive Maharam submeasure. For u L
0
= L
0
(A) set
(u) = inf : 0, [[[u[ > ]] .
Then denes a metrizable linear space topology T on L
0
such that for every neighbourhood G of 0 there
is a neighbourhood H of 0 such that u G whenever v H and [u[ [v[.
proof (a) The point is that
(u +v) (u) +(v), (u) (u) if [[ 1, lim
0
(u) = 0
for every u, v L
0
. PPP (i) It will save a moment if we observe that whenever > (u) there is an
such that [[[u[ > ]] , so that
[[[u[ > ]] [[[u[ > ]] .
Also, because is sequentially order-continuous,
[[[u[ > (u)]] = lim
n
[[[u[ > (u) + 2
n
]] lim
n
(u) + 2
n
= (u).
(ii) So
[[[u +v[ > (u) +(v)]] [[[u[ +[v[ > (u) +(v)]]) ([[[u[ > (u)]] [[[v[ > (v)]])
(364Fa)
[[[u[ > (u)]] +[[[v[ > (v)]] (u) +(v),
and (u +v) (u) +(v). (iii) If [[ 1 then
[[[u[ > (u)]] [[[u[ > (u)]] (u),
and (u) (u). (iv) lim
n
[[[u[ > n]] = 0 because [[[u[ > n]])
nN
is a non-increasing sequence with
inmum 0. So if > 0, there is an n 1 such that [[[u[ > n]] , in which case [[[u[ > ]] whenever
[[
1
n
, so that (u) whenever [[
1
n
. As is arbitrary, lim
0
(u) = 0. QQQ
(b) Accordingly we have a metric (u, v) (uv) which denes a linear space topology T on L
0
(2A5B).
Now let G be an open set containing 0. Then there is an > 0 such that H = u : (u) < is included in
G. If v H and [u[ [v[, then
[[[u[ > (v)]] [[[v[ > (v)]] (v),
so (u) (v) and u H G. So T satises all the conditions.
393N Proposition CM
5
is true i CM
1
, . . . , CM
4
are true.
proof CM
4
CM
5
Assume CM
4
. Let A, T be as in the statement of CM
5
. Let S be the topology on A
induced by T and the function : A L
0
; that is, S =
1
[G] : G T. Then S satises the conditions
of CM
4
. PPP (i) Because T is Hausdor and is injective, S is Hausdor. (ii) If 0 G S, there is an
H T such that G =
1
[H]. Now 0 (the zero of L
0
) belongs to H, so there is an open set H
1
containing
0 such that u H whenever v H
1
and [u[ [v[. Next, addition on L
0
is continuous for T, so there is an
open set H
2
containing 0 such that u + v H
1
whenever u, v H
2
. Consider G
t
=
1
[H
2
]. This is an
open set in A containing 0, and if a, b G
t
then
[(a b)[ a +b H
2
+H
2
H
1
,
588 Measurable algebras 393N
so (a b) H and a b G. As G is arbitrary, is continuous at (0, 0). (iii) If a
n
)
nN
is a non-increasing
sequence in A with inmum 0, u
0
= sup
nN
na
n
is dened in L
0
(use the criterion of 364Ma:
inf
mN
sup
nN
[[na
n
> m]] = inf
mN
a
m+1
= 0.)
If 0 G S, take H T such that G =
1
[H], and H
1
T such that 0 H
1
and u H whenever
v H
1
, [u[ [v[. Because scalar multiplication is continuous for T, there is a k 1 such that
1
k
u
0
H
1
.
For any n k, a
n

1
k
u
0
so a
n
H and a
n
G. As G is arbitrary, a
n
)
nN
0 for S. As a
n
)
nN
is
arbitrary, condition (ii) in the statement of CM
4
is satised. QQQ
Since CM
4
is true, A must be a measurable algebra.
CM
5
CM
1
Assume that CM
5
is true, and let A be a Dedekind complete Boolean algebra with a
strictly positive Maharam submeasure . By 393M, L
0
= L
0
(A) has a topology satisfying the conditions of
CM
5
, so A is measurable.
393O The phrase control measure derives, in fact, from none of the formulations above; it belongs to
the theory of vector measures, as follows.
Denitions (a) Let A be a Dedekind -complete Boolean algebra and U a Hausdor linear topological
space. (The idea is intended to apply, in particular, when A is a -algebra of subsets of a set.) A function
: A U is a vector measure if

n=0
a
n
= lim
n

n
i=0
a
i
is dened in U and equal to (sup
nN
a
n
)
for every disjoint sequence a
n
)
nN
in A.
(b) In this case, a non-negative countably additive functional : A [0, [ is a control measure for
if a = 0 whenever a = 0.
393P Now I can formulate the last of this string of statements equiveridical to CM
1
. Consider the
statement
(CM
6
) If A is a Dedekind -complete Boolean algebra, U is a metrizable linear topological space,
and : A U is a vector measure, then has a control measure.
393Q Lemma If A is a Dedekind -complete Boolean algebra, U a Hausdor linear topological space,
and : A U a vector measure, then lim
n
a
n
= 0 whenever a
n
)
nN
is a non-increasing sequence in A
with inmum 0.
proof a
n
=

i=n
(a
i
a
i+1
) for every n.
393R Proposition CM
6
is true i CM
1
, . . . , CM
5
are true.
proof CM
t
1
CM
6
(i) Assume CM
t
1
, and let A, U and be as in the statement of CM
6
. Then there is a
functional : U [0, [ such that, for u, v U, (u + v) (u) + (v), (u) (u) whenever [[ 1,
lim
0
(u) = 0, (u) = 0 i u = 0, and the topology of U is dened by the metric (u, v) (u v)
(2A5Cb). For a A set a = sup
ba
min(1, (b)).
(ii) Consider the functional : A [0, [. ()
0 = (0) = (0) = 0.
(To see that 0 = 0 set a
n
= 0 for every n in the denition 393Oa.) () If a b then of course a b. ()
If a, b A and c a b, then
min(1, (c)) = min(1, ((c a) +(c \ a)))
min(1, ((c a)) +((c \ a))) a +b.
As c is arbitrary, (a b) a + b. Thus is a submeasure. () ??? Suppose, if possible, that there is
a non-decreasing sequence a
n
)
nN
in A, with inmum 0, such that (a
n
))
nN
does not converge to 0.
Because (a
n
))
nN
is non-increasing, =
1
3
inf
nN
a
n
is greater than 0. Now for each n N there are
m > n, b such that b a
n
\ a
m
and (b) . PPP As a
n
3 there is a c a
n
such that (c) 2. Now
(a
m
c)
mN
converges to 0 in U, by 393Q, so there is an m > n such that
*393S The Control Measure Problem 589
((c \ a
m
)) = (c (a
m
c)) (c) ((a
m
c)) ,
and we can take b = c \ a
m
. QQQ
We may therefore choose b
k
)
kN
, n(k))
kN
such that n(k+1) > n(k), b
k
a
n(k)
\ a
n(k+1)
and (b
k
)
for every k. But b
k
)
kN
is disjoint, so we ought to be able to form

k=0
b
k
= (sup
kN
b
k
) in U, and
lim
k
b
k
= 0 in U, that is, lim
k
(b
k
) = 0. XXX
Thus lim
n
a
n
= 0 for every non-increasing sequence a
n
)
nN
with inmum 0, and is a Maharam
submeasure.
(iii) By CM
t
1
, there is a non-negative countably additive functional : A [0, [ such that a = 0
whenever a = 0. In particular, if a = 0, then (a) = 0 and a = 0. So is a control measure for . As
A, U and are arbitrary, CM
6
is true.
CM
6
CM
1
Assume CM
6
, and let A be a Dedekind complete Boolean algebra with a strictly positive
Maharam submeasure . Give L
0
= L
0
(A) the topology of 393M. Then : A L
0
is a vector measure in
the sense of 393O. PPP If a
n
)
nN
is a disjoint sequence in A with supremum a, set b
n
= sup
in
a
i
, so that
b
n
=

n
i=0
a
i
for each n. We have (a \ b
n
) 0, so that
(a b
n
)) = min(1, (a \ b
n
)) 0,
where is the functional of 393M, and a =

i=0
a
i
in L
0
. QQQ
CM
6
now assures us that there is a non-negative countably additive functional on A such that
a = 0 = a = 0 = a = 0,
so that is strictly positive and A is measurable. As A and are arbitrary, CM
1
is true.
*393S I must not go any farther without remarking that the generality of the phrase metrizable linear
topological space in CM
6
is essential. If we look only at normed spaces we do not need to know anything
about the Control Measure Problem, as the following theorem shows.
Theorem Let A be a Dedekind -complete Boolean algebra, U a normed space and : A U a vector
measure. Then has a control measure.
proof (a) Since U can certainly be embedded in a Banach space

U (3A5Ib), and as will still be a vector
measure when regarded as a map from A to

U, we may assume from the beginning that U itself is complete.
(b) is bounded (that is, sup
aA
|a| is nite). PPP??? Suppose, if possible, otherwise. Choose a
n
)
nN
inductively, as follows. a
0
= 1. Given that sup
aa
n
|a| = , choose b a
n
such that |b| |a
n
| + 1.
Then |(a
n
\ b)| 1. Also
sup
aa
n
|a| sup
aa
n
|(a b)| +|(a \ b)|,
so at least one of sup
ab
|a|, sup
aa
n
\b
|a| must be innite. We may therefore take a
n+1
to be either b
or a
n
\ b and such that sup
aa
n+1
|a| = . Observe that in either case we shall have |(a
n
\ a
n+1
)| 1.
Continue.
At the end of the induction we shall have a disjoint sequence a
n
\ a
n+1
)
nN
such that |(a
n
\ a
n+1
)| 1
for every n, so that

n=0
(a
n
\ a
n+1
) cannot be dened in U; which is impossible. XXXQQQ
(c) Accordingly we have a bounded linear operator T : L

U, where L

= L

(A), such that T =

(363Ea).
Now the key to the proof is the following fact: if u
n
)
nN
is a disjoint order-bounded sequence in (L

)
+
,
Tu
n
)
nN
0 in U. PPP Let be such that u
n
1 for every n. Let > 0, and let k be the integer
part of /. For n N, i k set a
ni
= [[u
n
> (i + 1)]]; then a
ni
)
nN
is disjoint for each i, and if we set
v
n
=

k
i=0
a
ni
, we get v
n
u
n
v
n
+1, so |u
n
v
n
|

.
Because a
ni
)
nN
is disjoint,

n=0
a
ni
is dened in U, and a
ni
)
nN
0, for each i k. Consequently
Tv
n
=

k
i=0
a
ni
0
as n . But
|Tu
n
Tv
n
| |T||u
n
v
n
|

|T|
590 Measurable algebras *393S
for each n, so limsup
n
|Tu
n
| |T|. As is arbitrary, lim
n
|Tu
n
| = 0. QQQ
(d) Consider the adjoint operator T
t
: U

(L

. Recall that L

is an M-space (363B) so that its

dual is an L-space (356N). Write
A = T
t
g : g U

, |g| 1 (L

= (L

.
If u L

, then
sup
fA
[f(u)[ = sup
|g|1
[(T

g)(u)[ = sup
|g|1
[g(Tu)[ = |Tu|.
Now A is uniformly integrable. PPP I use the criterion of 356O. Of course |f| |T
t
| for every f A, so A
is norm-bounded. If u
n
)
nN
is an order-bounded disjoint sequence in (L

)
+
, then
sup
fA
[f(u
n
)[ = |Tu
n
| 0
as n . So A is uniformly integrable. QQQ
(e) Next, A (L

c
. PPP If f A, it is of the form T
t
g for some g U

, that is,
f(a) = (T
t
g)(a) = gT(a) = g(a)
for every a A. If now a
n
)
nN
is a disjoint sequence in A with supremum a,
f(a) = g((sup
nN
a
n
)) = g(

n=0
a
n
) =

n=0
g(a
n
) =

n=0
f(a
n
).
So f is countably additive. By 363K, f (L

c
. QQQ
(f ) Because A is uniformly integrable, there is for each m N an f
m
0 in (L

such that |([f[

f
m
)
+
| 2
m
for every f A; moreover, we can suppose that f
m
is of the form sup
ik
m
[f
mi
[ where every
f
mi
belongs to A (354R(b-iii)), so that f
m
(L

c
and
m
= f
m
is countably additive. Set

m
= 1 +
m
1 for each m, =

m=0
1
2
m

m
;
then : A [0, [ is a non-negative countably additive functional.
Now is a control measure for . PPP If a = 0, then
m
a = 0, that is, f
m
(a) = 0, for every m N. But
this means that if g U

and |g| 1,
[g(a)[ = [(T
t
g)(a)[ f
m
(a) +|([T
t
g[ f
m
)
+
| 2
m
for every m, by the choice of f
m
; so that g(a) = 0. As g is arbitrary, a = 0; as a is arbitrary, is a control
measure for . QQQ
393T This concludes the list of positive results I wish to present in this section. I now devote a
few pages to signicant examples of submeasures. The rst example is a classic formulation (taken from
Talagrand 80) which shows in clear relief some of the fundamental ways in which submeasures dier from
measures.
Examples (a) Fix n 1, and let I be the set 0, 1, . . . , 2n 1, X = [I]
n
, so that X is a nite set. For
each i I set A
i
= a : i a X. For E X set
E =
1
n+1
inf#(J) : J I, E
_
iJ
A
i

=
1
n+1
inf#(J) : a J ,= for every a E.
It is elementary to check that : TX [0, [ is a strictly positive submeasure, therefore (because TX is
nite) a Maharam submeasure.
The essential properties of are twofold: (i) X = 1; (ii) for any non-negative additive functional such
that E E for every E X, X
2
n+1
. PPP (i)() If J I and #(J) n, there is an a [I J]
n
,
so that a X

iJ
A
i
and X ,

iJ
A
i
. This means that X cannot be covered by fewer than n + 1 of
the sets A
i
, so that X must be at least 1. () On the other hand, if J I is any set of cardinal n + 1,
a J ,= for every a X, so that X =

iJ
A
i
and X 1. (ii) Every member of X belongs to just n of
the sets A
i
, so
393U The Control Measure Problem 591
nX =

iI
A
i

#(I)
n+1
=
2n
n+1
,
and X
2
n+1
. QQQ
(b) This example shows at least that any proof of CM
3A
cannot work through any generally valid in-
equality of the form if is a Maharam submeasure there is an additive functional with . If
we take a sequence of these spaces we can form a result which in one direction is stronger, as follows.
For each n 1 set I
n
= 0, . . . , 2n1 and dene X
n
= [I
n
]
n
, A
ni
= a : i a X
n
,
n
: TX
n
[0, 1]
as in (a) above. Set Z =

n=1
X
n
, K = (n, i) : n 1, i I
n
and
C
ni
= x : x Z, x(n) A
ni

for (n, i) K. Now dene : TZ [0, 1] by setting

W = inf

(n,i)J
1
n+1
: J K, W

(n,i)J
C
ni

for every W Z. Then is an outer measure on Z, by arguments we have been familiar with since 114D.
Also Z = 1. PPP (i) Because (for instance) X
1
is covered by the two sets A
10
and A
11
, Z is covered by C
10
and C
11
, so that Z 1. (ii) If J K and

(n,i)J
1
n+1
< 1, set J
n
= i : (n, i) J for each n; then
#(J
n
) < n+1, so we can choose an x(n) X
n

iJ
n
A
ni
. This denes a sequence x Z such that x / C
ni
for every (n, i) J, and Z ,=

(n,i)J
C
ni
. As J is arbitrary, Z 1. QQQ
Finally, if is any subalgebra of TZ containing every C
ni
, and : [0, [ a non-negative nitely
additive functional such that E E for every E , then = 0. PPP For each n, every point of x belongs
to n dierent C
ni
, just as in (a) above; so that
nZ

iI
n
C
ni

2n
n+1
.
As this is true for every n 1, Z = 0. QQQ
(c) A non-zero submeasure on a Boolean algebra A is called pathological if the only additive functional
such that 0 a a for every a A is the zero functional. Thus the submeasure of (b) above is
pathological, and CM
3A
can be read an exhaustive submeasure cannot be pathological.
(d) It is important to note that Fubinis theorem fails catastrophically for submeasures. As a simple
example, consider the space (X, ) of (a) above, for some xed n 1. If we dene : T(X X) [0, 1] by
setting
W = inf

i=0
E
i
F
i
: W

iN
E
i
F
i

for each W X X, then we obtain an outer measure, just as if itself were a measure; but (X X)
4n/(n +1)
2
is small compared with (X)
2
. PPP Give 0, . . . , 2n 1 = Z
2n
its usual group operation +
2n
of
addition mod 2n. Then
X X

i<2n
(A
i
A
i
)

i<2n
(A
i
A
i+
2n
1
),
because if a, b X either a b ,= and there is some i such that a, b both belong to A
i
, or b =
0, . . . , 2n 1 a and there is some i such that i a, i +
2n
1 b. So
(X X)

i<2n
A
i
(A
i
+A
i+
2n
1
) =
4n
(n+1)
2
. QQQ
393U For the next example, I present a much deeper idea from Roberts 93.
Example Let B be the algebra of open-and-closed subsets of 0, 1
N
. Then for any > 0 we can nd a
submeasure : B [0, 1] such that
(i) for every n N there is a disjoint sequence S
n0
, . . . , S
nn
in B such that S
ni
= 1 for every i n;
(ii) if E
n
)
nN
is any disjoint sequence in B then limsup
n
E
n
.
proof (a) For each n N let I
n
be the nite set 0, . . . , n, given its discrete topology; set X =

nN
I
n
,
with the product topology; let C be the algebra of subsets of X generated by sets of the form S
ij
= x :
592 Measurable algebras 393U
x(i) = j, where i N and j i. Note that X is compact and Hausdor and that every member of C is
open-and-closed (because all the S
ij
are). Also C is atomless, countable and non-zero, so is isomorphic to
B, by 393F. It will therefore be enough if I can describe a submeasure : C [0, 1] with the properties (i)
and (ii) above, and this is what I will do.
(b) For each n N let /
n
be the set of non-empty members of C determined by coordinates in 0, . . . , n;
note that /
n
is nite. For k l N, say that E C is (k, l)-thin if for every A /
k
there is an A
t
/
l
such that A
t
A E. Note that if k
t
k l l
t
then /
k
/
k
and /
l
/
l
, so any (k, l)-thin set is also
(k
t
, l
t
)-thin.
Say that every E C is (k, 0)-small for every k N, and that for k, r N a set E C is (k, r+1)-small if
there is some l k such that E is (k, l)-thin and (l, r)-small. Observe that E is (k, 1)-small i it is (k, l)-thin
for some l k, that is, there is no member of /
k
included in E. Observe also that if E is (k, r)-small then
it is (k
t
, r)-small for every k
t
k.
Write o = S
ij
: j i N.
(c) Suppose that E C and k l m are such that E is both (k, l)-thin and (l, m)-thin. Then whenever
A /
k
, S o and A S ,= , there is an A
t
/
m
such that A
t
A E and A
t
S ,= . PPP Take S = S
ni
where i n. (i) If n l, then A S /
l
; because E is (l, m)-thin, there is an A
t
/
m
such that
A
t
(A S) E. (ii) If n > l, there is an A
t
/
l
such that A
t
A E, because E is (k, l)-thin; now
A
t
/
m
, and A
t
S is non-empty because A
t
is determined by coordinates less than n. QQQ
(d) It follows that if S o, k N, A /
k
, A S ,= , r N and E
0
, . . . , E
r1
are (k, 2r)-small, then
AS is not covered by E
0
, . . . , E
r1
. PPP Induce on r. The case r = 0 demands only that AS should not
be covered by the empty sequence, that is, A S ,= , which is one of the hypotheses. For the inductive
step to r +1, we know that for each j r there are l
j
, m
j
such that k l
j
m
j
and E
j
is (k, l
j
)-thin and
(l
j
, m
j
)-thin and (m
j
, 2r)-small. Rearranging E
0
, . . . , E
r
if necessary we may suppose that m
r
m
j
for
every j r; set m = m
r
. By (c), there is an A
t
/
m
such that A
t
S ,= and A
t
A E
r
. Now every
E
j
, for j < r, is (m
j
, 2r)-small, therefore (m, 2r)-small, so by the inductive hypothesis A
t
S ,

j<r
E
j
.
Accordingly A S ,

jr
E
j
and the induction continues. QQQ
(e) Now suppose that E
n
)
nN
is a disjoint sequence in C. Then for any k N there are l, n

N such
that E
n
is (k, l)-thin for every n n

. PPP Consider G
n
=

jn
E
j
for each n N. Then every G
n
is open
and

nN
G
n
= . If A /
k
, then A, with its subspace topology, is compact, so Baires theorem (3A3G)
tells us that there is an n
A
such that G
n
A
A is not dense in A; let l
A
be such that A G
n
A
includes a
member of /
l
A
. Set n

= maxn
A
: A /
k
, l = maxl
A
: A /
k
. If n n

, A /
k
there is an
A
t
/
l
A
/
l
such that
A
t
A G
n
A
A E
n
.
As A is arbitrary, E
n
is (k, l)-thin. QQQ
It follows at once that for any r N we can nd n

r
, k
0
< k
1
< . . . < k
r
N such that E
n
is (k
j
, k
j+1
)-thin
for every j < r and n n

r
; so that E
n
is (0, r)-small for every n n

r
.
(f ) Take an integer r 1/. Let | be the set of (0, 2r)-small members of C. Set
E =
1
r
minm : E E
1
. . . E
m
for some E
1
, . . . , E
m
|
if E can be covered by r or fewer members of |, 1 otherwise. It is easy to check that : C [0, 1] is a
submeasure. By (d), no member of o can be covered by r or fewer members of |, so S
ni
= 1 whenever
i n N. By (e), if E
n
)
nN
is any disjoint sequence in C, E
n
belongs to | for all but nitely many n, so
that E
n

1
r
for all but nitely many n. Thus has the required properties.
393X Basic exercises (a) Let A be a Boolean algebra and a strictly positive Maharam submeasure
on A. Show that a subset F of A is closed for the topology dened by i the limit of any order*-convergent
sequence in F belongs to F.
(b) Let A be a Dedekind -complete Boolean algebra with a strictly positive Maharam submeasure .
Show that A is complete under the metric dened from (393B).
393Yd The Control Measure Problem 593
(c) Let A be a Dedekind complete Boolean algebra and ,
t
two strictly positive Maharam submeasures
on A. Show that the corresponding metrics on A (393B) are uniformly equivalent.
(d) Let A be a countable Boolean algebra, not 0. Show that A is isomorphic to an order-closed
subalgebra of the algebra B of open-and-closed subsets of 0, 1
N
. (Hint: show that A B

= B.)
(e) Suppose that CM
1
-CM
6
are false. Show that there is a Dedekind complete Boolean algebra A, with
a strictly positive Maharam submeasure, such that (i) the only countably additive real-valued functional on
A is the zero functional (ii) (A) = .
(f ) Let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure. Show
that if (A) then the associated topology on A (393B) is separable and A is -linked.
(g) Consider the statement
(CM
t
6
) If X is a set, a -algebra of subsets of X, U a metrizable linear topological space, and
: U a vector measure, then has a control measure.
Show that CM
t
6
CM
6
.
(h) Consider the outer measures
n
, of 393Ta-b. Give every X
n
its discrete topology and Z the
corresponding product topology. Let c be the algebra of open-and-closed subsets of Z. (i) Show that c is
the subalgebra of TZ generated by sets of the form x : x Z, x(n) = a for n 1, a X
n
. (ii) Take
n 1 and set A = a : a X
n
,

ia
i is even. Show that
n
A =
n
(X A) =
n
n+1
. (iii) Show that for
any < 1 there is a disjoint sequence E
n
)
nN
in c such that E
n
for every n.
393Y Further exercises (a) Let A be a Dedekind complete Boolean algebra with a strictly positive
Maharam submeasure. Show that if (A) then A is -m-linked (391Yh) for every m.
(b) Let A be a Boolean algebra with just four elements, and let N
+
be the set of submeasures on A,
N = N
+
N

the set of functionals from A to R expressible as the dierence of two submeasures. (i) Show
that N is just the three-dimensional space of functionals : A R such that 0 = 0. (ii) Show that there
is a partial order on N under which N is a partially ordered linear space with positive cone N
+
, but that
N is now not a lattice.
(c) Let A be a Dedekind complete Boolean algebra. Show that there is at most one Hausdor topology T
on A such that (i) the Boolean operations , . are continuous (ii) for every open set G containing 0 there
is an open set H containing 0 such that a G whenever a b H (iii) 0 A whenever A A is non-empty
and downwards-directed and has inmum 0. (Hint: Given such a topology T, show that the construction
in 393K produces enough order-continuous submeasures on A to dene T. Now if is any order-continuous
submeasure, let a be its support; because A
a
is ccc, there is a T-continuous submeasure with support a; use
393E to see that is continuous.)
(d) Let A be a Boolean algebra. Let T be the topology on A such that the closed sets for T are just those
sets F A such that a F whenever a
n
)
nN
is a sequence in F which order*-converges to a (367Yc). Show
that a a c, a a c, a a .c, a a \ c, a c \ a are all T-continuous, for any c A. Write | for the
family of open sets G containing 0 such that [0, a] G for every a G; for A A set A
t
= a b : a, b A.
Show that G G
t
| for every G |.
Suppose now that A is ccc and Dedekind complete and that T is Hausdor. Show that (i) A is weakly
(, )-distributive; (ii) whenever A A and a A there is a sequence in A which order*-converges to a (see
367Yl); (iii) whenever 0 G T there is an H | such that H G (hint: H = inta : [0, a] G |);
(iv) for any a A 0 there is a G | such that a / G
t
; (v) for every G | there is an H | such that
H
tt
G
t
; (vi) for any a A 0 there is a sequence G
n
)
nN
in | such that a , sup(

nN
G
n
); (vii) there
is a sequence G
n
)
nN
in | such that

nN
G
n
= 0; (viii) for every G | there is an H | such that
H
t
G; (ix) there is a strictly positive Maharam submeasure on A (hint: 393K).
(See Balcar Glowczy nski & Jech 98.)
594 Measurable algebras 393Ye
(e) Consider the statement
(CM
t
4
) Let A be a ccc Dedekind complete Boolean algebra with a Hausdor topology T such that
if a
n
)
nN
is a sequence in A which order*-converges to a A, then a
n
)
nN
converges to 0 for
the topology T. Then A is measurable.
Show that CM
t
4
is true i CM
1
is true.
(f ) Let A be a Dedekind complete Boolean algebra with a strictly positive Maharam submeasure, and let
S, T be the associated metrizable topologies on A, L
0
(A) (393B, 393N). Write d(A), d(L
0
) for the topological
densities of these spaces (331Yf). Show that
max((A), ) = max(d(A), ) = max(d(L
0
), ).
(g) Let A be a Dedekind -complete Boolean algebra, : A [0, [ a non-negative nitely additive
functional, and a Maharam submeasure on A such that a = 0 whenever a = 0. Show that is absolutely
continuous with respect to and therefore also with respect to the countably additive part of as dened
in 362Bc.
(h) Let A be a Dedekind -complete Boolean algebra, U a Hausdor linear topological space and : A U
a vector measure. Suppose that there is a bounded nitely additive functional : A R such that a = 0
whenever a = 0. Show that has a control measure.
(i) A linear topological space U is locally convex if the topology has a base consisting of convex sets.
Show that if A is a Dedekind -complete Boolean algebra, U is a metrizable locally convex linear topological
space, and : A U is a vector measure, then has a control measure.
(h) Consider the outer measure of 393Tb. (i) Show that there is a non-decreasing sequence W
n
)
nN
of
subsets of Z such that (

nN
W
n
) > sup
nN
W
n
. (ii) Show that if W
n
)
nN
is a non-decreasing sequence
of open subsets of Z then (

nN
W
n
) = sup
nN
W
n
.
393 Notes and comments This section is long in pages, but I hope has been reasonably easy reading,
because so many of the arguments amount to re-written versions of ideas already presented. Thus 393Ba-
b can be thought of as an elementary extension of 323A-323B, and the completion procedure in 393Bc
uses some of the same ideas as the proof of 325C. The notion of absolute continuity for submeasures is
taken directly from that for measures, and 393E is a simple translation of 232Ba. Other fragments of the
argument for the Radon-Nikod ym theorem appear in the proof that CM
1
implies CM
3A
(393I) and in 393S.
The argument of 393K takes a fair bit of space, but is mostly a recapitulation of the proof that a uniformity
can be dened from pseudometrics (Engelking 89, 8.1.10; Bourbaki 66, IX.1.4). If there is a new idea
there, it is in the formula
a = inf
1
b : a e b e
in the middle of the proof. (Contrast this with the formula
a = sup
ba
min(1, (b))
in the proof of 393R.) 393M is just a matter of picking the right description of the topology of convergence
in measure to generalize.
I do think it is a little surprising that CM
3A
and CM
3B
should be equiveridical; one is asking for an
additive functional dominated by , and the other for an additive functional weakly dominating . Of
course it is not exactly the same in both cases. It is tempting to look for a theory of spaces of submeasures
corresponding to the theory of spaces of additive functionals (326), but it seems that, at the very least,
there are new obstacles (393Yb).
The arguments of this section tend to be based on hypotheses of the form CM
n
is true; I hope that
this will not lead to any presumption that the Control Measure Problem has a positive answer. I am myself
inclined to believe that all the statements CM
n
are false, and that one day all this material will have to
be re-written with the arguments in reverse, proving CM
n
CM
m
instead of CM
m
CM
n
, or using
reductio ad absurdum at every point where it is not used above.
394A Kawadas theorem 595
I include CM
t
2
because it seems to make the problem more accessible, and indeed it is signicant for
theoretical reasons. In this form it is possible to show that the Control Measure Problem cannot depend on
any of the special axioms that have been introduced in the last thirty-ve years; that is, if there is a solution
using such an axiom as the continuum hypothesis, or Martins axiom, or the axiom of constructibility,
then there must be a solution using only the ordinary axioms of Zermelo-Fraenkel set theory (including
the axiom of choice). For elucidation of this remark, which I fear may be somewhat cryptic, look for
Shoenelds Absoluteness Theorem in a book on mathematical logic and forcing (e.g., Kunen 80). An
incidental benet from including CM
t
2
is that it has forced me, at last, to spell out the characterization of
the algebra of open-and-closed subsets of 0, 1
N
(393F).
I mention CM
tt
1
and CM
t
6
(393Xg) as well as CM
t
1
and CM
6
because some natural versions of the Control
Measure Problem arise in the context of -algebras of sets, rather than Dedekind -complete Boolean
algebras, and a reminder that this makes no dierence to the problem may be helpful.
This is not a book about vector measures, but having introduced vector measures in CM
6
it would be
disgraceful to leave you unaware that there is no control measure problem for measures taking values in
normed spaces, and since the proof is no more than an assembly of ideas already covered in their scattered
locations, it seems right to set it out (393S).
I have said already that my own prejudice is in favour of believing that all the statements CM

are false;
that is, that there is a Dedekind complete Boolean algebra A, with a strictly positive Maharam submeasure,
such that A is not measurable. Ordinarily such a presumption makes a problem seem less interesting; only
rarely is there as much honour and prot in constructing a recondite counter-example as there is in proving
a substantial theorem. But in this case I believe that it could be the starting point of a new theory of
submeasurable algebras, being Dedekind complete Boolean algebras carrying strictly positive Maharam
submeasures. These would be ccc weakly (, )-distributive algebras (392I), some of them -m-linked
for every m (393Ya), carrying metrics for which the Boolean operations were uniformly continuous and
order-closed sets were closed (393B); there would be associated theories of topologies of convergence in
submeasure on algebras and L
0
spaces (393B, 393M). We already have a remarkable characterization of
such algebras in terms of order*-convergence (393Yd). But the real prize would be a new algebra for use in
forcing, conceivably leading to new models of set theory.
The examples in 393T and 393U are there for dierent purposes. In 393Ta we have submeasures on nite
algebras, which are therefore necessarily uniformly exhaustive, in which dominated measures are cruelly
dominated; the submeasures are almost pathological, and can readily be assembled into a submeasure
which is pathological in the strict sense (393Tb), but is now very far from being exhaustive (393Xh). Of
course the method of 393Tb is extraordinarily crude, but as far as I know nothing else works either (see
393Td).
I called 393T a classic construction; I am sure that whatever resolution is at last found for the Control
Measure Problem, 393Ta at least will always be of interest. It is less clear that 393U will endure in the same
way, but for the moment it is the best example known of a submeasure which is almost exhaustive while
being far from uniformly exhaustive.
394 Kawadas theorem
I now describe a completely dierent characterization of (homogeneous) measurable algebras, based on
the special nature of their automorphism groups. The argument depends on the notion of non-paradoxical
group of automorphisms; this is an idea of great importance in other contexts, and I therefore aim at a fairly
thorough development, with proofs which are adaptable to other circumstances.
394A Denitions Let A be a Dedekind complete Boolean algebra, and G a subgroup of Aut A. For a,
b A I will say that an isomorphism : A
a
A
b
between the corresponding principal ideals belongs to the
full local semigroup generated by G if there are a partition of unity a
i
)
iI
in A
a
and a family
i
)
iI
in G such that c =
i
c whenever i I and c a
i
. If such an isomorphism exists I will say that a and b
are G--equidecomposable.
I will write a

G
b to mean that there is a b
t
b such that a and b
t
are G--equidecomposable.
596 Measurable algebras 394A
For any function f with domain A, I will say that f is G-invariant if f(a) = f(a) whenever a A and
G.
394B The notion of full local semigroup is of course an extension of the idea of full subgroup (381M,
387A). The word semigroup is justied by (c) of the following lemma, and the word full by (e).
Lemma Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. Write G

for the full

local semigroup generated by G.
(a) Suppose that a, b A and that : A
a
A
b
is an isomorphism. Then the following are equiveridical:
(i) G

;
(ii) for every non-zero c
0
a there are a non-zero c
1
c
0
and a G such that c = c for every
c c
1
;
(iii) for every non-zero c
0
a there are a non-zero c
1
c
0
and a G

such that c = c for every

c c
1
.
(b) If a, b A and : A
a
A
b
belongs to G

, then
1
: A
b
A
a
also belongs to G

.
(c) Suppose that a, b, a
t
, b
t
A and that : A
a
A
a
, : A
b
A
b
belong to G

. Then G

; its
domain is A
c
where c =
1
(b a
t
), and its set of values is A
c
where c
t
= (b a
t
).
(d) If a, b A and : A
a
A
b
belongs to G

, then A
c
G

for any c a.
(e) Suppose that a, b A and that : A
a
A
b
is an isomorphism such that there are a partition of
unity a
i
)
iI
in A
a
and a family
i
)
iI
in G

such that c =
i
c whenever i I and c a
i
. Then G

.
proof (a) (Compare 387B.)
(i)(iii) is trivial, since of course G G

.
(iii)(ii) Suppose that satises (iii), and that 0 ,= c
0
a. Then we can nd a G

and a non-zero
c
1
c
0
such that agrees with on A
c
1
. Suppose that dom = A
d
, where necessarily d c
1
. Then there
are a partition of unity d
i
)
iI
in A
d
and a family
i
)
iI
such that c =
i
c whenever c d
i
. There is
some i I such that c
2
= c
1
d
i
,= 0, and we see that c = c =
i
c for every c c
2
. As c
0
is arbitrary,
satises (ii).
(ii)(i) If satises (ii), set
D = d : d a, there is some G such that c = c for every c d.
The hypothesis is that D is order-dense in A, so there is a partition of unity a
i
)
iI
of A
a
lying within D
(313K); for each i I take
i
G such that c =
i
c for c a
i
; then a
i
)
iI
and
i
)
iI
witness that
G

.
(b) This is elementary; if a
i
)
iI
,
i
)
iI
witness that G

, then a
i
)
iI
=
i
a
i
)
iI
,
1
i
)
iI
witness
that
1
G

.
(c) I ought to start by computing the domain of :
d dom() d dom, d dom
d a, d b d
1
(a
t
b) = c.
So the domain of is indeed A
c
; now A
c
is an isomorphism between A
c
and A
c
, where c = a
t
b A
b
,
so is an isomorphism between A
c
and A
c
= A
c
. Let a
i
)
iI
, b
j
)
jJ
be partitions of unity in A
a
,
A
b
respectively, and
i
)
iI
,
j
)
jJ
families in G such that d =
i
d for d a
i
, e =
j
e for e b
j
. Set
c
ij
= a
i

1
i
b
j
; then c
ij
)
iI,jJ
is a partition of unity in A
c
and d =
j

i
d for d c
ij
, so G

(because all the

j

i
belong to G).
(d) This is nearly trivial; use the denition of G

or the criteria of (a), or apply (c) with the identity

map on A
c
as one of the factors.
(e) This follows at once from the criterion (a-iii) above, or otherwise.
394C Lemma Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. Write G

for the full local semigroup generated by G.

394D Kawadas theorem 597
(a) For a, b A, a

G
b i there is a G

such that a dom and a b.

(b)(i)

G
is transitive and reexive;
(ii) if a

G
b and b

G
a then a and b are G--equidecomposable.
(c) G--equidecomposability is an equivalence relation on A.
(d) If a
i
)
iI
and b
i
)
iI
are families in A, of which b
i
)
iI
is disjoint, and a
i

G
b
i
for every i I, then
sup
iI
a
i

G
sup
iI
b
i
.
proof (a) This is immediate from the denition of G--equidecomposable and 394Bd.
(b)(i) a

G
a because the identity homomorphism belongs to G

. If a

G
b

G
c there are , G

such that a b, b c so that a c; as G

(394Bc), a

G
c.
(ii) (This is of course a Schroder-Bernstein theorem, and the proof is the usual one.) Take , G

such that a b, b a. Set a

0
= a, b
0
= b, a
n+1
= b
n
and b
n+1
= a
n
for each n. Then a
n
)
nN
,
b
n
)
nN
are non-increasing sequences; set a

= inf
nN
a
n
, b

= inf
nN
b
n
. For each n,
A
a
2n
\a
2n+1
: A
a
2n
\a
2n+1
A
b
2n+1
\b
2n+2
,
A
b
2n
\b
2n+1
: A
b
2n
\b
2n+1
A
a
2n+1
\a
2n+2
are isomorphisms, while
A
a

: A
a

A
b

is another. So we can dene an isomorphism : A

a
A
b
by setting
c = c if c a

sup
nN
a
2n
\ a
2n+1
,
=
1
c if c sup
nN
a
2n+1
\ a
2n+2
.
By 394Be, G

, so a and b are G--equidecomposable.

(c) This is easy to prove directly from the results in 394B, but also follows at once from (b-i); any
transitive reexive relation gives rise to an equivalence relation.
(d) We may suppose that I is well-ordered by a relation . For i I, set a
t
i
= a
i
\ sup
j<i
a
j
. Set
a = sup
iI
a
i
= sup
iI
a
t
i
, b = sup
iI
b
i
. For each i I, we have a b
t
i
b
i
and a
i
G

such that
i
a
t
i
= b
t
i
.
Set b
t
= sup
iI
b
t
i
b; then we have an isomorphism : A
a
A
b
dened by setting d =
i
d if d a
t
i
,
and G

, so a and b
t
are G--equidecomposable and a

G
b.
394D Theorem Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. Then the
following are equiveridical:
(i) there is an a ,= 1 such that a is G--equidecomposable with 1;
(ii) there is a disjoint sequence a
n
)
nN
of non-zero elements of A which are all G--equidecomposable;
(iii) there are non-zero G--equidecomposable a, b, c A such that a b = 0 and a b c;
(iv) there are G--equidecomposable a, b A such that a b.
proof Write G

for the full local semigroup generated by G.

(i)(ii) Assume (i). There is a G

such that 1 = a. Set a

n
=
n
(1 \ a) for each n N; because
every
n
belongs to G

(counting
0
as the identity operator on A, and using 394Bc), with dom
n
= A, a
n
is G--equidecomposable with a
0
= 1 \ a for every n. Also a
n
=
n
1 \
n+1
1 for each n, while
n
1)
nN
is
non-increasing, so a
n
)
nN
is disjoint. Thus (ii) is true.
(ii)(iii) Assume (ii). Set a = sup
nN
a
2n
, b = sup
nN
a
2n+1
, c = sup
nN
a
n
, so that a b = 0 and
a b = c. For each n we have a
n
G

such that
n
a
0
= a
n
. So if we set
d = sup
nN

1
2n
(d a
2n
) for d a,
belongs to G

(using 394B) and witnesses that a and c are G--equidecomposable. Similarly, b and c are
G--equidecomposable, so (iii) is true.
(iii)(iv) is trivial.
598 Measurable algebras 394D
(iv)(i) Take G

such that b = a. Set

d = (d b) (d \ b)
for d A; then G

witnesses that 1 is G--equidecomposable with a (1 \ b) ,= 1.

394E Denition Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A. I will
say that G is fully non-paradoxical if the statements of 394D are false; that is, if one of the following
equiveridical statements is true:
(i) if a is G--equidecomposable with 1 then a = 1;
(ii) there is no disjoint sequence a
n
)
nN
of non-zero elements of A which are all G--equide-
composable;
(iii) there are no non-zero G--equidecomposable a, b, c A such that a b = 0 and a b c;
(iv) if a b A and a, b are G--equidecomposable then a = b.
Note that if G is fully non-paradoxical, and H is a subgroup of Aut A such that H G, then H is also fully
non-paradoxical, because if a

G
b then a

H
b, so that a and b are H--equidecomposable whenever they
are G--equidecomposable.
394F Proposition Let (A, ) be a totally nite measure algebra, and G = Aut

A the group of
measure-preserving automorphisms of A. Then G is fully non-paradoxical.
proof If : A A
a
belongs to the full local subgroup generated by G, then we have a partition of unity
a
i
)
iI
and a family
i
)
iI
in G such that a
i
=
i
a
i
for every i; but this means that
a =

iI

i
a
i
=

iI

i
a
i
=

iI
a
i
= 1.
As 1 < , we can conclude that a = 1, so that G satises the condition (i) of 394E.
394G The xed-point subalgebra of a group Let A be a Boolean algebra and G a subgroup of
Aut A.
(a) By the xed-point subalgebra of G I mean
C = c : c A, c = c for every G.
(I looked briey at this construction in 333R, and again in 387 in the special case of a group generated by
a single element.) This is a subalgebra of A, and is order-closed, because every G is order-continuous.
(b) Now suppose that A is Dedekind complete. In this case C is Dedekind complete (314Ea), and we
have, for any a A, an element upr(a, C) of C, dened by setting
upr(a, C) = infc : a c C
(314V). Now upr(a, C) = supa : G. PPP Set c
1
= upr(a, C), c
2
= supa : G. (i) Because
a c
1
C, a c
1
= c
1
for every G, and c
2
c
1
. (ii) For any G,
c
2
= sup
G
a = sup
G
a = c
2
because G = : G. So c
2
C; since also a c
2
, c
1
c
2
, and c
1
= c
2
, as claimed. QQQ
(c) Again supposing that A is Dedekind complete, write G

for the full local semigroup generated by

G. Then (a c) = a c whenever G

, a dom and c C. PPP We have a = sup

iI

i
a
i
, where
a = sup
iI
a
i
and
i
G for every i. Now
(a c) = sup
iI

i
(a
i
c) = sup
iI

i
a
i
c = a c. QQQ
Consequently upr(a, C) = upr(a, C) whenever G

and a dom. PPP For c C,

a c a c = a (a c) = a a c = a a c. QQQ
It follows that upr(a, C) upr(b, C) whenever a

G
b.
394I Kawadas theorem 599
(d) Still supposing that A is Dedekind complete, we also nd that if a

G
b and c C then a c

G
b c.
PPP There is a G

such that a b; now (a c) = a c b c. QQQ Hence, or otherwise, a c and b c

are G--equidecomposable whenever a and b are G--equidecomposable and c C.
(e) Of course the case C = 0, 1 is particularly signicant; when this happens I will call G ergodic.
Thus an automorphism is ergodic in the sense of 372P i the group
n
: n Z it generates is ergodic.
394H I now embark on a series of lemmas leading to the main theorem (394N).
Lemma Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A.
Write C for the xed-point subalgebra of G. Take any a, b A. Set c
0
= supc : c C, a c

G
b; then
a c
0

G
b and b \ c
0

G
a.
proof Enumerate G as

)
<
, where = #(G). Dene a

)
<
, b

)
<
inductively, setting
a

= (a \ sup
<
a

)
1

(b \ sup
<
b

), b

.
Then a

)
<
is a disjoint family in A
a
and b

)
<
is a disjoint family in A
b
, and sup
<
a

is G--
equidecomposable with sup
<
b

. Set a
t
= a \ sup
<
a

, b
t
= b \ sup
<
b

,
c
0
= 1 \ upr(a
t
, C) = supc : c C, c a
t
= 0.
Then
a c
0
sup
<
a

G
b,
so c
0
c
0
.
Now b
t
c
0
. PPP??? Otherwise, because c
0
= 1 \ sup
<

a
t
(394Gb), there must be a < such that

a
t
b
t
,= 0. But in this case d = a
t

1

b
t
,= 0, and we have
d (a \ sup
<
a

)
1

(b \ sup
<
b

),
so that d a

, which is absurd. XXXQQQ Consequently

b \ c
0
sup
<
b

G
a.
Now take any c C such that a c

G
b, and consider c
t
= c \ c
0
. Then b
t
c
t
= 0, that is, b c
t
=
sup
<
b

c
t
, which is G--equidecomposable with sup
<
a

c
t
= (a \ a
t
) c
t
(394Gd). But now
a c
t
= a c c
t

G
b c
t

G
(a c
t
) \ (a
t
c
t
);
because G is fully non-paradoxical, a
t
c
t
must be 0, that is, c
t
c
0
and c
t
= 0. As c
t
is arbitrary, c
0
c
0
and c
0
= c
0
. So c
0
has the required properties.
Remark By analogy with the notation I used in discussing the Hahn decomposition of countably additive
functionals (326O), we might denote c
0
as [[a

G
b]], or perhaps [[a

G
b]]
C
, the region (in C) where a

G
b.
The same notation would write upr(a, C) as [[a ,= 0]]
C
.
394I The construction I wish to use depends essentially on L
0
spaces as described in 364. The next
step is the following.
Lemma Let A be a Dedekind complete Boolean algebra, not 0, and G a fully non-paradoxical subgroup
of Aut A. Let C be the xed-point subalgebra of G. Suppose that a, b A and that upr(a, C) = 1. Then
there are non-negative u, v L
0
= L
0
(C) such that
[[u n]] = maxc : c C, there is a disjoint family d
i
)
i<n
such that c a

G
d
i
b for every i < n,
[[v n]] = maxc : c C, there is a family d
i
)
i<n
such that d
i

G
a for every i < n and b c sup
i<n
d
i

for every n N. Moreover, we can arrange that

(i) [[u N]] = [[v N]] = 1,
600 Measurable algebras 394I
(ii) [[v > 0]] = upr(b, C),
(iii) u v u +1.
Remark By writing max in the formulae above, I mean to imply that the elements [[u n]], [[v n]] belong
to the sets described.
proof (a) Choose c
n
)
nN
, b
n
)
nN
as follows. Given b
i
)
i<n
, set b
t
n
= b \ sup
i<n
b
i
,
c
n
= supc : a c

G
b
t
n
,
so that a c
n

G
b
t
n
(394H); choose b
n
b
t
n
such that a c
n
is G--equidecomposable with b
n
, and continue.
Then b
n
)
nN
is a disjoint sequence in A
b
and c
n
)
nN
is a non-increasing sequence in C.
For each n, we have b
t
n
\ c
n

G
a, by 394H; while a c ,

G
b
t
n
whenever c C and c , c
n
. Note also that,
because upr(a, C) = 1,
c
n
= upr(a c
n
, C) = upr(b
n
, C) upr(b
t
n
, C).
(b) Now c

= inf
nN
c
n
= 0. PPP b
n
c

)
nN
is a disjoint sequence, all G--equidecomposable with
a c

, so a c

= 0; because upr(a, C) = 1, it follows that c

= 0. QQQ Accordingly, if we set u =

sup
nN
(n +1)c
n
, u L
0
and [[u n]] = c
n1
for n 1. The construction ensures that [[u N]], as dened
in 364H, is equal to 1.
(c) Consider next c
t
0
= upr(b, C), c
t
n
= c
n1
upr(b
t
n
, C) for n 1. Then c
t
n
)
nN
is a non-increasing
sequence with zero inmum, so again we can dene v L
0
by setting v = sup
nN
(n + 1)c
t
n
. Once again,
[[v N]] = 1, and [[v n]] = 1 \ c
t
n
for each n.
Of course [[v > 0]] = c
t
0
= upr(b, C). Because c
n
c
t
n
c
n1
,
(n + 1)c
n
(n + 1)c
t
n
nc
n1
+1
for each n 1, and u v u +1.
(d) Now set
C
n
= c : c C, there is a disjoint family d
i
)
i<n
such that c a

G
d
i
b for every i < n.
Then c
n
= max C
n+1
.
PPP() Because c
n
c
n1
. . . c
0
, a c
n

G
b
i
for every i n, so that b
i
)
in
witnesses that c
n

C
n+1
.
() Suppose that c C
n+1
; let d
i
)
in
be a disjoint family such that c a

G
d
i
b for every i. Set
c
t
= c \ c
n
. For each i < n, b
i

G
a, so
b
i
c
t

G
a c
t

G
d
i
c
t
,
while also
b
t
n
c
t

G
a c
t

G
d
n
c
t
.
Take d d
n
c
t
such that b
t
n
c
t
is G--equidecomposable with d. Then
b c
t
= (b
t
n
c
t
) sup
i<n
(b
i
c
t
)

G
d sup
i<n
(d
i
c
t
) b c
t
.
Because G is fully non-paradoxical, d sup
i<n
(d
i
c
t
) must be exactly b c
t
, so d must be the whole of
d
n
c
t
, and
a c
t

G
d
n
c
t
= d

G
b
t
n
.
But this means that c
t
c
n
. Thus c
t
= 0 and c c
n
. So c
n
= supC
n+1
= max C
n+1
. QQQ
Accordingly
[[u n]] = c
n1
= max C
n
for n 1. For n = 0 we have [[u 0]] = 1 = max C
0
. So [[u n]] = max C
n
for every n, as required.
(e) Similarly, if we set
394K Kawadas theorem 601
C
t
n
= c : c C, there is a family d
i
)
i<n
such that d
i

G
a for every i < n and b c sup
i<n
d
i

then 1 \ c
t
n
= max C
t
n
for every n.
PPP() If n = 0, then of course (interpreting sup as 0) 1 \ c
t
0
C
t
0
because b c
t
0
. For each n N, set

b
n
= b
n
(b
t
n
\ c
n
) = (b
n
c
n
) (b
t
n
\ c
n
).
Because b
n

G
a and b
t
n
\ c
n

G
a, we have b
n
c
n

G
a c
n
and b
t
n
\ c
n

G
a \ c
n
, so

b
n

G
a (394Cd). If
we look at
sup
i<n

b
i
sup
i<n
b
i
(b
t
n1
\ c
n1
),
we see that, for n 1,
b \ sup
i<n

b
i
b
t
n
c
n1
c
t
n
,
so that b \ c
t
n
sup
i<n

b
i
and

b
i
: i < n witnesses that 1 \ c
t
n
C
t
n
.
() Now take any c C
t
n
and a corresponding family d
i
)
i<n
such that d
i

G
a for every i < n and
b c sup
i<n
d
i
.
Set c
t
= c c
t
n
. For each i < n,
c
t
d
i

G
c
t
a

G
b
i
because c
t
c
i
. So (by 394Cd, as usual)
c
t
b

G
c
t
sup
i<n
b
i
c
t
b,
and (again because G is fully non-paradoxical) c
t
b = c
t
sup
i<n
b
i
, that is, c
t
b
t
n
= 0. But c
t
c
t
n

upr(b
t
n
, C), so c
t
must be 0, which means that c 1 \ c
t
n
. As c is arbitrary, 1 \ c
t
n
= sup C
t
n
= max C
t
n
. QQQ
Thus [[v n]] = supC
t
n
, as declared.
394J Notation In the context of 394I, I will write b : a| for u, b : a| for v.
394K Lemma Let A be a Dedekind complete Boolean algebra, not 0, and G a fully non-paradoxical
subgroup of Aut A with xed-point subalgebra C. Suppose that a, b, b
1
, b
2
A and that upr(a, C) = 1.
(a) 0 : a| = 0 : a| = 0, 1 : a| 1 and 1 : 1| = 1.
(b) If b
1

G
b
2
then b
1
: a| b
2
: a| and b
1
: a| b
2
: a|.
(c) b
1
b
2
: a| b
1
: a| +b
2
: a|.
(d) If b
1
b
2
= 0, b
1
: a| +b
2
: a| b
1
b
2
: a|.
(e) If c C is such that a c is a relative atom over C (denition: 331A), then c [[b : a| b : a| = 0]].
proof (a)-(b) are immediate from the denitions and the basic properties of

G
, . . . | and . . . |, as listed
in 394C, 394E and 394I.
(c) For j, k N, set c
jk
= [[b
1
: a| = j]] [[b
2
: a| = k]]. Then
c
jk
[[b
1
b
2
: a| j +k]] [[b
1
: a| +b
2
: a| = j +k]].
PPP We may suppose that c
jk
,= 0. Of course
c
jk
[[b
1
: a| +b
2
: a| = j +k]].
Next, there are sets J, J
t
A such that d

G
a for every d J J
t
, #(J) j, #(J
t
) k, supJ b
1
c
jk
and supJ
t
b
2
c
jk
. So sup(J J
t
) (b
1
b
2
) c
jk
and J J
t
witnesses that c
jk
[[b
1
b
2
: a| j +k]].
QQQ
Accordingly
c
jk
[[b
1
: a| +b
2
: a| b
1
b
2
: a| 0]].
Now as sup
j,kN
c
jk
= 1, we must have b
1
b
2
: a| b
1
: a| +b
2
: a|.
602 Measurable algebras 394K
(d) This time, set c
jk
= [[b
1
: a| = j]] [[b
2
: a| = k]] for j, k N. Then
c
jk
[[b
1
b
2
: a| j +k]] [[b
1
: a| +b
2
: a| = j +k]]
for every j, k N. PPP Once again, we surely have
c
jk
[[b
1
: a| +b
2
: a| = j +k]].
Next, we can nd a family d
i
)
i<j+k
such that
d
i
)
i<j
is disjoint, a c
jk

G
d
i
b
1
for every i < k,
d
i
)
ji<j+k
is disjoint, a c
jk

G
d
i
b
2
for j i < j +k.
As b
1
b
2
= 0, the whole family d
i
)
i<j+k
is disjoint and witnesses that c
jk
[[b
1
b
2
: a| j +k]]. QQQ
So
c
jk
[[b
1
b
2
: a| b
1
: a| b
2
: a| 0]]
Since sup
j,kN
c
jk
= 1, as before, we must have b
1
b
2
: a| b
1
: a| +b
2
: a|.
(e) ??? Otherwise, there must be some k N such that
c
0
= c [[b : a| = k]] [[b : a| > k]] ,= 0.
Let d
i
)
i<k
be a disjoint family in A
b
such that a c
0

G
d
i
for each i; cutting the d
i
down if necessary, we
may suppose that a c
0
is G--equidecomposable with d
i
for each i. As c
0
, [[b : a| k]], b c
0
, sup
i<k
d
i
;
set d = b c
0
\ sup
i<k
d
i
,= 0. If c
t
C is non-zero and c
t
c
0
, then a c
t

G
d
i
for every i < k, while
c
t
, [[b : a| k + 1]], so a c
t
,

G
d; by 394H, d c
0

G
a and d = d c
0

G
a c
0
. There is therefore a
non-zero a a c
0
such that a

G
d. But now remember that a c is supposed to be a relative atom over
C, so a = a c for some c C such that c c
0
. In this case, a c

G
d
i
for every i < k and also a c

G
d,
so 0 ,= c [[b : a| k + 1]], which is absurd. XXX
394L Lemma Let A be a Dedekind complete Boolean algebra, not 0, and G a fully non-paradoxical
subgroup of Aut A with xed-point subalgebra C. Suppose that a
1
, a
2
, b A and that upr(a
1
, C) =
upr(a
2
, C) = 1. Then
b : a
2
| b : a
1
| a
1
: a
2
|, b : a
2
| b : a
1
| a
1
: a
2
|.
proof I use the same method as in 394K. As usual, write G

for the full local semigroup generated by G.

(a) For j, k N set
c
j,k
= [[b : a
1
| = j]] [[a
1
: a
2
| = k]].
Then
c
j,k
[[b : a
1
| a
1
: a
2
| = jk]] [[b : a
2
| jk]].
PPP Write c for c
j,k
. As in parts (c) and (d) of the proof of 394K, it is elementary that c is included in
[[b : a
1
| a
1
: a
2
| = jk]]; what we need to check is that c [[b : a
2
| jk]]. Again, we may suppose that
c ,= 0. There are families d
i
)
i<j
, d

l
)
l<k
such that
d
i
)
i<j
is disjoint, a
1
c

G
d
i
b for every i < j,
d

l
)
l<k
is disjoint, a
2
c

G
d

l
a
1
for every l < k.
For each i < j, let
i
G

be such that
i
(a
1
c) d
i
. If i < j and l < k, then
a
2
c

G
d

l
c

G

i
(d

l
c)
i
(a
1
c) d
i
b.
Also
i
(d

l
c))
i<j,l<k
is disjoint because
i
(a
1
c))
i<j
and d

l
)
l<k
are, so witnesses that c is included in
[[b : a
2
| jk]]. QQQ
Now, just as in 394K, it follows from the fact that sup
j,kN
c
j,k
= 1 that b : a
1
| a
1
: a
2
| b : a
2
|.
(b) For j, k N set
c
j,k
= [[b : a
1
| = j]] [[a
1
: a
2
| = k]].
394M Kawadas theorem 603
Then
c
j,k
[[b : a
1
| a
1
: a
2
| = jk]] [[b : a
2
| jk]].
PPP Write c for c
j,k
. Then c [[b : a
1
| a
1
: a
2
| = jk]]. There are families d
i
)
i<j
, d

l
)
l<k
such that
d
i

G
a
1
for every i < j, d

G
a
2
for every l < k, b c sup
i<j
d
i
and a
1
c sup
l<k
d

l
. For each i < j,
let d
t
i
a
1
be G--equidecomposable with d
i
, and take
i
G

such that
i
d
t
i
= d
i
. Then

i
(d
t
i
d

l
)

G
d

G
a
2
for every i < j, l < k,
sup
i<j,l<k

i
(d
t
i
d

l
) = sup
i<j

i
(d
t
i
sup
l<k
d

l
) sup
i<j

i
(d
t
i
c)
= sup
i<j
d
i
c b c.
So
i
(d
t
i
d

l
))
i<j,l<k
witnesses that c [[b : a
2
| jk]]. QQQ
Once again, it follows easily that b : a
1
| a
1
: a
2
| b : a
2
|.
394M Lemma Let A be a Dedekind complete Boolean algebra, not 0, and G a subgroup of Aut A
with xed-point subalgebra C.
(a) For any a A, there is a b a such that b

G
a \ b and a
t
= a \ upr(b, C) is a relative atom over C.
(b) Now suppose that G is fully non-paradoxical. Then for any > 0 there is an a A such that
upr(a, C) = 1 and b : a| b : a| +1 : a| for every b A.
proof (a) Set B = d : d a, d

G
a \ d and let D B be a maximal subset such that upr(d, C) upr(d
t
, C)
= 0 for all distinct d, d
t
D. Set b = supD. For any d D, d

G
a \ d, so
b upr(d, C) = d upr(d, C)

G
(a \ d) upr(d, C) = (a \ b) upr(d, C) a \ b
by 394Gc. By 394H,
b = b sup
dD
upr(d, C)

G
a \ b.
??? Suppose, if possible, that a
t
= a \ upr(b, C) is not a relative atom over C. Let d
0
a
t
be an element
not expressible as a
t
c for any c C; then d
0
,= a upr(d
0
, C) and there must be a G such that
d
1
= d
0
a \ d
0
is non-zero. In this case
d
1

G

1
d
1
d
0
a \ d
1
,
so d
1
B; but also
d
1
upr(d, C) d
1
upr(b, C) = 0,
so upr(d
1
, C) upr(d, C) = 0, for every d D, and we ought to have put d
1
into D. XXX
Thus b has the required properties.
(b)(i) For every n N we can nd a
n
A and c
n
C such that upr(a
n
, C) = 1, a
n
c
n
is a relative
atom over C, and 1 : a
n
| 2
n
c
n
. PPP Induce on n. The induction starts with a
0
= c
0
= 1, because
1 : 1| = 1. For the inductive step, having found a
n
and c
n
, let d a
n
c
n
be such that d

G
a
n
c
n
\ d
and a
n
c
n
\ upr(d, C) is a relative atom over C, as in (a). Set c
n+1
= upr(d, C), a
n+1
= (a
n
\ c
n+1
) d;
then
upr(a
n+1
, C) = upr(a
n
\ c
n+1
, C) upr(d, C)
= (upr(a
n
, C) \ c
n+1
) c
n+1
= (1 \ c
n+1
) c
n+1
= 1
by 314Vb-314Vc and the inductive hypothesis.
We have c
n+1
d

G
c
n+1
a
n
\ d, so
c
n+1
a
n+1
= d a
n
, c
n+1
a
n+1

G
a
n
\ d,
and a
n
: a
n+1
| 2c
n+1
; by 394L,
1 : a
n+1
| 1 : a
n
| a
n
: a
n+1
| 2
n
c
n
2c
n+1
= 2
n+1
c
n+1
.
If
604 Measurable algebras 394M
b a
n+1
\ c
n+1
= (a
n
\ c
n
) (a
n
c
n
\ c
n+1
),
then, because both terms on the right are relative atoms over C, there are c
t
, c
tt
C such that
b = (b a
n
\ c
n
) (b a
n
c
n
\ c
n+1
)
= (c
t
a
n
\ c
n
) (c
tt
a
n
c
n
\ c
n+1
) = c a
n+1
\ c
n+1
where c = (c
t
\ c
n
) (c
tt
c
n
) belongs to C. So a
n+1
\ c
n+1
is a relative atom over C.
Thus the induction continues. QQQ
(ii) Now suppose that > 0. Take n such that 2
n
, and consider a
n
, c
n
taken from (i) above.
Let b A. Set
c = [[b : a
n
| b : a
n
| 1 : a
n
| > 0]] C.
Since we know that
1 : a
n
| 2
n
2
n
c
n
= c
n
, b : a
n
| b : a
n
| +1,
we must have c c
n
= 0. But this means that a
n
c is a relative atom over C. By 394Ke, c is included in
[[b : a
n
| b : a
n
| = 0]]; as also 1 : a
n
| 1 (394Ka), c must be zero, that is, b : a
n
| b : a
n
|+1 : a
n
|.
394N We are at last ready for the theorem.
Theorem Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A
with xed-point subalgebra C. Then there is a unique function : A L

(C) such that

(i) is additive, non-negative and order-continuous;
(ii) [[a > 0]] = upr(a, C) for every a A; in particular, a = 0 i a = 0;
(iii) 1 = 1;
(iv) (a c) = a c for every a A, c C; in particular, c = c for every c C;
(v) If a, b A are G--equidecomposable, then a = b; in particular, is G-invariant.
proof If A = 0 this is trivial; so I suppose henceforth that A ,= 0.
(a) Set A

= a : a A, upr(a, C) = 1 and for a A

, b A set

a
(b) =
b : a|
1 : a|
L
0
= L
0
(C);
the rst thing to note is that because 1 : a| 1, we can always do the divisions to obtain elements
a
(b)
of L
0
(A) (364P). Set
b = inf
aA

a
b
for b A. (Note that L
0
(C) is Dedekind complete (364O), so the inmum is dened.)
(b) The formulae of 394K tell us that, for a A

and b
1
, b
2
A,

a
0 = 0,
a
b
1

a
b
2
if b
1
b
2
,

a
(b
1
b
2
)
a
b
1
+
a
b
2
,

a
1 1.
It follows at once that
0 = 0, b
1
b
2
if b
1
b
2
,
1 1.
(c) For each n N there is an e
n
A

such that b : e
n
| b : e
n
| +2
n
1 : e
n
| for every b A (394M).
Now
e
n
b
a
b + 2
n
b : a| for every a A

, b A. PPPa : e
n
| a : e
n
| + 2
n
1 : e
n
|, so
394N Kawadas theorem 605
a : e
n
| 1 : a| a : e
n
| 1 : a| + 2
n
1 : e
n
| 1 : a|
1 : e
n
| + 2
n
1 : e
n
| 1 : a|
(by 394L); accordingly
b : e
n
| 1 : a| b : a| a : e
n
| 1 : a|
(by the other half of 394L)
b : a| 1 : e
n
| + 2
n
b : a| 1 : e
n
| 1 : a|
and, dividing by 1 : a| 1 : e
n
|, we get
e
n
b
a
b + 2
n
b : a|. QQQ
(d) Now is additive. PPP Taking e
n
)
nN
from (c), observe rst that
inf
nN

e
n
b
a
b + inf
nN
2
n
b : a| =
a
b
for every a A

, b A, so that b = inf
nN

e
n
b for every b. Now suppose that b
1
, b
2
A and b
1
b
2
= 0.
Then, for any n N,
b
1
: e
n
| +b
2
: e
n
| b
1
: e
n
| +b
2
: e
n
| + 2
n+1
1 : e
n
|
b
1
b
2
: e
n
| + 2
n+1
1 : e
n
|
(by 394Kd)
b
1
b
2
: e
n
| + 2
n+1
1 : e
n
|.
Dividing by 1 : e
n
|, we have
b
1
+b
2

e
n
b
1
+
e
n
b
2

e
n
(b
1
b
2
) + 2
n+1
1.
Taking the inmum over n, we get
b
1
+b
2
(b
1
b
2
).
In the other direction, if a, a
t
A

and n N,
(b
1
b
2
)
e
n
(b
1
b
2
)
e
n
(b
1
) +
e
n
(b
2
)

a
(b
1
) + 2
n
b
1
: a| +
a
(b
2
) + 2
n
b
2
: a
t
|.
As n is arbitrary, (b
1
b
2
)
a
(b
1
) +
a
(b
2
); as a and a
t
are arbitrary, (b
1
b
2
) b
1
+b
2
(using 351Dc).
As b
1
and b
2
are arbitrary, is additive. QQQ
We see also that 1 : e
n
| (1 + 2
n
)1 : e
n
|, so that
e
n
1 (1 + 2
n
)1 for each n; since we already
know that 1 1, we have 1 = 1 exactly.
(e) If c C then
[[c > 0]] [[
1
c > 0]] [[c : 1| > 0]] = upr(c, C) = c
(394I(ii)). It follows that
(b c) b c b c
for any b A, c C. Similarly, (b \ c) b (1 \ c); adding, we must have equality in both, and
(b c) = b c.
Rather late, I point out that
0 a 1 = 1 L

= L

(C)
for every a A, so that a L

for every a.
(f ) If b A\ 0, then
[[b > 0]] [[
1
b > 0]] [[b : 1| > 0]] = upr(b, C)
606 Measurable algebras 394N
by 394I(ii) again. ??? Suppose, if possible, that [[b > 0]] ,= upr(b, C). Set c
0
= upr(b, C) \ [[b > 0]], a
0
=
b (1 \ upr(b, C)) A

. Let k 1 be such that c

1
= c
0
[[1 : a
0
| k]] ,= 0. Then a
0
c
1
= b c
1
, so
a
0
c
1
= (a
0
c
1
) = (b c
1
) = b c
1
= 0.
By 364Nb, there is an a A

such that c
1
, [[
a
a
0
c
1

1
k
]], that is, c
2
= c
1
[[
a
a
0
<
1
k
]] ,= 0. Now
c
2
[[1 : a| ka
0
: a| > 0]] [[1 : a
0
| a
0
: a| ka
0
: a| > 0]] [[1 : a
0
| > k]],
which is impossible, as c
2
c
1
. XXX
Thus [[b > 0]] = upr(b, C). In particular, b = 0 i b = 0.
(g) If b, b
t
A and b

G
b
t
, then b b
t
. PPP For every a A

, b : a| b
t
: a| (394Kb) so
a
b
a
b
t
.
QQQ So if b, b
t
A and c = [[b b
t
> 0]], b
t
c

G
b. PPP??? Otherwise, by 394H, there is a non-zero c
t
c
such that b c
t

G
b
t
. But in this case b c
t
= (b c
t
) b
t
and c
t
[[b
t
b 0]]. XXXQQQ
(h) If a
i
)
iI
is any disjoint family in A with supremum a, a =

iI
a
i
, where the sum is to be
interpreted as sup
JI is nite

iJ
a
i
. PPP Induce on #(I). If #(I) is nite, this is just nite additivity ((d)
above). For the inductive step to #(I) = , we may suppose that I is actually equal to the cardinal .
Of course
a (sup
J
a

) =

J
a

for every nite J , so (because L

(C) is Dedekiond complete) u =

<
a

is dened, and u a.
For < , set b

= sup
<
a

. By the inductive hypothesis,

b

<
a

= sup
J is nite

J
a

u.
At the same time, if J is nite, there is some < such that J , so that

J
a

; accordingly
sup
<
b

= u.
??? Suppose, if possible, that u < a; set v = a u. Take > 0 such that c
0
= [[v > ]] ,= 0. Let <
be such that c
1
= c
0
\ [[u b

> ]] is non-zero (cf. 364Nb). Now u b

(a \ b

), so
c
1
[[(a \ b

) > 0]] = upr(a \ b

, C),
and c
1
(a \ b

) ,= 0; there is therefore an
t
such that d = c
1
a

,= 0. Since d u b

and
c
1
[[u b

]] [[v > ]], [[v d > 0]] c

1
.
Choose d

)
<
inductively, as follows. Given that d

)
<
is a disjoint family in A
a\d
such that d

is G-
-equidecomposable with a

c
1
for every < , then e

= sup
<
d

is G--equidecomposable with b

c
1
,
so that e

, and
[[(a \ (d e

)) a

> 0]] = [[a d e

> 0]] [[a d b

> 0]]
= [[a d b
+1
> 0]] [[v d > 0]] c
1
.
By (g), a

c
1

G
a \ (d e

); take d

a \ (d e

) G--equidecomposable with a

c
1
, and continue.
At the end of this induction, we have a disjoint family d

)
<
in A
a\d
such that d

is G--equidecom-
posable with a

c
1
for every . But this means that a
t
= sup
<
d

is G--equidecomposable with a c
1
,
while a
t
(a \ d) c
1
; since d c
1
,= 0, G cannot be fully non-paradoxical. XXX
Thus a = u =

<
a

and the induction continues. QQQ

(i) It follows that is order-continuous. PPP () If B A is non-empty and upwards-directed and has
supremum e, then

bB
A
b
is order-dense in A
e
, so includes a partition of unity A of A
e
; now (h) tells us
that
e =

aA
a sup
bB
b.
Since of course b e for every b B, e = sup
bB
b. () If B A is non-empty and downwards-directed
and has inmum e, then, using (), we see that
1 e = (1 \ e) = sup
bB
(1 \ b) = sup
bB
1 b,
so that e = inf
bB
b. QQQ
(j) I still have to show that is unique. Let
t
: A L

be any non-negative order-continuous G-invariant

additive function such that
t
c = c for every c C.
394P Kawadas theorem 607
(i) Just as in (e) of this proof, but more easily, we see that
t
(b c) =
t
b c for every b A, c C.
(ii) If a
i
)
iI
is a disjoint family in A with supremum a, then sup
iJ
a
i
)
JI is nite
is an upwards-
directed family with supremum a, so that

t
a = sup
JI is nite

t
(sup
iJ
a
i
) = sup
JI is nite

iJ

t
a
i
=

iI

t
a
i
.
(iii)
t
a =
t
b whenever a and b are G--equidecomposable. PPP Take a partition a
i
)
iI
of a and a
family
i
)
iI
in G such that
i
a
i
)
iI
is a partition of b. Then

t
a =

iI

t
a
i
=

iI

t

i
a
i
=
t
b. QQQ
Consequently
t
a
t
b whenever a

G
b.
(iv) Take a A

, b A and for j, k N set c

jk
= [[1 : a| = j]] [[b : a| = k]]. Then
b : a| c
jk

t
b 1 : a| c
jk
.
PPP If c
jk
= 0 this is trivial; suppose c
jk
,= 0. Now we have sets I, J such that #(I) = j, #(J) k,
a c
jk

G
d for every d I, e

G
a for every e J, I is disjoint, and b c
jk
supJ. So

t
b 1 : a| c
jk
= j
t
b c
jk
= j
t
(b c
jk
) j

eJ

t
(e c
jk
)
jk
t
(a c
jk
) k

dI

t
(d c
jk
) k
t
c
jk
= kc
jk
= b : a| c
jk
. QQQ
Summing over j and k, b : a|
t
b 1 : a|, that is,
a
b
t
b. Taking the inmum over a, b
t
b. But
also
b = 1 (1 \ b) 1
t
(1 \ b) =
t
b,
so b =
t
b. As b is arbitrary, =
t
. This completes the proof.
394O We have reached the summit. The rest of the section is a list of easy corollaries.
Theorem Let A be a Dedekind complete Boolean algebra, not 0, and G a fully non-paradoxical subgroup
of Aut A. Then there is a G-invariant additive functional : A [0, 1] such that 1 = 1.
proof Let C be the xed-point subalgebra of G, and : A L

(C) the function of 394N. By 311D, there

is a ring homomorphism
0
: C 0, 1 such that
0
1 = 1; now
0
can also be regarded as an additive
functional from C to R. Let f
0
: L

(C) R be the corresponding positive linear functional (363K). Set

= f
0
. Then is order-preserving and additive because f
0
and are, 1 = f
0
(1) =
0
1 > 0, and is
G-invariant because is.
394P Theorem Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup
of Aut A with xed-point subalgebra C. Then the following are equiveridical:
(i) A is a measurable algebra;
(ii) C is a measurable algebra;
(iii) there is a strictly positive G-invariant countably additive real-valued functional on A.
proof (iii)(i)(ii) are trivial. For (ii)(iii), let : A L

(C) be the function of 394N, and : C R

a strictly positive countably additive functional. Let f : L

(C) R be the corresponding linear operator;

then f is sequentially order-continuous (363K again). Set = f. Then is additive and order-preserving
and sequentially order-continuous because f and are. It is also strictly positive, because if a A 0
then a > 0 (394N(ii)), that is, there is some > 0 such that [[a > ]] ,= 0, so that
a [[a > ]] > 0.
Finally, is G-invariant because is.
608 Measurable algebras 394Q
394Q Corollary: Kawadas Theorem Let A be a Dedekind complete Boolean algebra such that
Aut A is ergodic and fully non-paradoxical. Then A is measurable.
proof This is the case C = 0, 1 of 394P.
394R Thus the existence of an ergodic fully non-paradoxical subgroup is a sucient condition for
a Dedekind complete Boolean algebra to be measurable. It is not quite necessary, because if a measure
algebra A is not homogeneous then its automorphism group is not ergodic. But for homogeneous algebras
the condition is necessary as well as sucient, by the following result.
Proposition If (A, ) is a homogeneous totally nite measure algebra, the group G = Aut

A of measure-
preserving automorphisms of A is ergodic.
proof If A = 0, 1 this is trivial. Otherwise, A is atomless. If a A 0, 1, set = min( a, (1 \ a)) > 0;
then there are b a, d 1 \ a such that b = d = . By 382Fb, there is a G such that b = d, so that
a ,= a. As a is arbitrary, the xed-point subalgebra of G is 0, 1.
394X Basic exercises (a) Re-write the section on the assumption that every group G is ergodic,
that is, C is always 0, 1, so that L
0
(C) may be identied with R, the functions . . . | and . . . | become
real-valued, the functionals
a
(394N) become submeasures and becomes a measure.
(b) Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A with xed-point subalgebra
C. Suppose that c
i
)
iI
is a partition of unity in C and that a, b A are such that a c
i

G
b for every
i I. Show that a

G
b.
(c) Let A be a Dedekind complete Boolean algebra and G a subgroup of Aut A with xed-point subalgebra
C. Show that A is relatively atomless over C i the full local semigroup generated by G has many involutions
(denition: 381P).
(d) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with
xed-point subalgebra C. Show that the following are equiveridical: (i) there is a strictly positive real-valued
additive functional on A; (ii) there is a strictly positive real-valued additive functional on C; (iii) there is a
strictly positive G-invariant real-valued additive functional on A.
(e) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A
with xed-point subalgebra C. Show that the following are equiveridical: (i) there is a non-zero completely
additive functional on A; (ii) there is a non-zero completely additive functional on C; (iii) there is a non-zero
G-invariant completely additive functional on A.
(f ) Let A be a ccc Dedekind complete Boolean algebra. Show that it is a measurable algebra i there is
a fully non-paradoxical subgroup G of Aut A such that the xed-point subalgebra of G is purely atomic.
(g) Let (A, ) be a localizable measure algebra. Show that the following are equiveridical: (i) Aut

A is
ergodic; (ii) A is quasi-homogeneous in the sense of 374G; (iii) whenever a, b A and a = b then the
principal ideals A
a
, A
b
are isomorphic.
(h) Let (A, ) be a localizable measure algebra. Show that Aut

A is fully non-paradoxical i (i) for every
innite cardinal , the Maharam-type- component of A (denition: 332G) has nite measure (ii) for every
]0, [ there are only nitely many atoms of measure .
394Y Further exercises (a) Let A be a Dedekind complete Boolean algebra, G a subgroup of Aut A,
and G

the full local semigroup generated by G. For , G

, say that if extends . (i) Show that

every member of G

can be extended to a maximal member of G

. (ii) Show that G is fully non-paradoxical

i every maximal member of G

is actually a Boolean automorphism of A.

(b) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with
xed-point subalgebra C. Show that A is ccc i C is ccc. (Hint: if C is ccc, L

(C) has the countable sup

property (363Yb).)
394 Notes Kawadas theorem 609
(c) Let A be a Dedekind complete Boolean algebra and G a fully non-paradoxical subgroup of Aut A with
xed-point subalgebra C. Show that A is weakly (, )-distributive i C is.
(d) Let A be a Dedekind complete Boolean algebra, G an ergodic subgroup of A, and G

the full local

semigroup generated by G. Suppose that there is a non-zero a A for which there is no G

such that
a a. Show that there is a measure such that (A, ) is a localizable measure algebra. (Hint: show that
A
a
is a measurable algebra.)
394Z Problem Suppose that A is a Dedekind complete Boolean algebra, not 0, and G a subgroup
of Aut A such that whenever a
i
)
in
is a nite partition of unity in A and we are given
i
,
t
i
G for every
i n, then the elements
0
a
0
,
t
a
0
,
1
a
1
,
t
1
a
1
, . . . ,
t
n
a
n
are not all disjoint. Must there be a non-zero
non-negative G-invariant nitely additive functional on A?
(See Tarskis theorem in the notes below.)
394 Notes and comments Regarded as a sucient condition for measurability, Kawadas theorem suers
from the obvious defect that it is going to be rather rarely that we can verify the existence of an ergodic
fully non-paradoxical group of automorphisms without having some quite dierent reason for supposing that
our algebra is measurable. If we think of it as a criterion for the existence of a G-invariant measure, rather
than as a criterion for measurability in the abstract, it seems to make better sense. But if we know from
the start that the algebra A is measurable, the argument short-circuits, as we shall see in 395.
I take the trouble to include the in every G--equidecomposable, G

and

G
because there are
important variations on the concept, in which the partitions a
i
)
iI
of 394A are required to be nite or
countable. Indeed Tarskis theorem relies on one of these. I spell it out because it is close to Kawadas in
spirit, though there are signicant dierences in the ideas needed in the proof:
Let X be a set and G a subgroup of Aut TX. Then the following are equiveridical: (i) there is
a G-invariant additive functional : TX [0, 1] such that A = 1; (ii) there are no A
0
, . . . , A
n
,

0
, . . . ,
n
,
t
0
, . . . ,
t
n
such that A
0
, . . . , A
n
are subsets of X covering X,
0
, . . . ,
t
n
all belong
to G, and
0
A
0
,
t
0
A
0
,
1
A
1
,
t
1
A
1
, . . . ,
t
n
A
n
are all disjoint.
For a proof, see 449I in Volume 4; for an illuminating discussion of this theorem, see Wagon 85, Chapter
9. But it seems to be unknown whether the natural translation of this result is valid in all Dedekind
complete Boolean algebras (394Z). Note that we are looking for theorems which do not depend on any
special properties of the group G or the Boolean algebra A. For abelian or amenable groups, or weakly
(, )-distributive algebras, for instance, much more can be done, as described in 395Ya and 449.
The methods of this section can, however, be used to prove similar results for countable groups of auto-
morphisms on Dedekind -complete Boolean algebras; I will return to such questions in 448.
As noted, Kawada (Kawada 44) treated the case in which the group G of automorphisms is ergodic,
that is, the xed-point subalgebra C is trivial. Under this hypothesis the proof is of course very much
simpler. (You may nd it useful to reconstruct the original version, as suggested in 394Xa.) I give the more
general argument partly for the sake of 394O, partly to separate out the steps which really need ergodicity
from those which depend only on non-paradoxicality, partly to prepare the ground for the countable version
in the next volume, partly to show o the power of the construction in 364, and partly to get you used
to Boolean-valued arguments. A bolder use of language could indeed simplify some formulae slightly by
writing (for instance) [[ka
0
: a| < 1 : a|]] in place of [[1 : a| ka
0
: a| > 0]] (see part (f) of the proof of
394N). As in 387, the dierences involved in the extension to non-ergodic groups are, in a sense, just a
matter of technique; but this time the technique is more obtrusive. The presentation here owes a good deal
to Nadkarni 90 and something to Becker & Kechris 96.
610 Measurable algebras 395 intro.
395 The Hajian-Ito theorem
In the notes to the last section, I said that the argument there short-circuits if we are told that we are
dealing with a measurable algebra. The point is that in this case there is a much simpler criterion for the
existence of a G-invariant measure (395B(ii)), with a proof which is independent of 394 in all its non-trivial
parts, which makes it easy to prove that non-paradoxicality is sucient as well as necessary.
395A Lemma Let (A, ) be a localizable measure algebra.
(a) Let Aut A be a Boolean automorphism (not necessarily measure-preserving). Let T

be the
corresponding Riesz homomorphism from L
0
= L
0
(A) to itself (364R). Then there is a unique w

(L
0
)
+
such that
_
w

v =
_
T

v for every v (L
0
)
+
.
(b) If , Aut A then w

= w

1w

.
(c) For each Aut A we have a norm-preserving isomorphism U

from L
2
= L
2
(A, ) to itself dened
by setting
U

v = T

1
for every v L
2
, and U

= U

for all , Aut A.

proof (a) Applying 365T with a = (a), we see that there is a unique w

(L
0
)
+
such that
_
a
w

= (a)
for every a A. If we look at
W = v : v (L
0
)
+
,
_
v w

=
_
T

v,
we see that W contains a for every a A, that v + v
t
W and v W whenever v, v
t
W and 0,
and that sup
nN
v
n
W whenever v
n
)
nN
is a non-decreasing sequence in W which is bounded above in
L
0
. By 364Kd, W = (L
0
)
+
, as required.
(b) For any v (L
0
)
+
,
_
w

v =
_
T

v =
_
T

v
(364Re)
=
_
w

v =
_
T

(T

1w

v)
(recalling that T

is multiplicative)
=
_
w

1w

v.
As v is arbitrary (and (A, ) is semi-nite), w

= w

1w

.
(c)(i) For any v L
0
,
_
(T

1)
2
=
_
T

v
2
w

1 =
_
T

1T

v
2
=
_
v
2
.
So U

v L
2
and |U

v|
2
= |v|
2
whenever v L
2
, and U

is a norm-preserving operator on L
2
.
(ii) Now consider U

. For any v L
2
, we have
U

v = T

(T

1)

1
= T

v
_
T

1 w

1
(using 364Rd)
= T

1
(by (b) above)
= U

v.
395B The Hajian-Ito theorem 611
So U

= U

.
(iii) Writing for the identity operator on A, we see that T

is the identity operator on L

0
, w

= 1
and U

is the identity operator on L

2
. Since U

1U

= U

1 = U

, U

: L
2
L
2
is an isomorphism,
with inverse U

1, for every Aut A.

395B Theorem (Hajian & Ito 69) Let A be a measurable algebra and G a subgroup of Aut A. Then
the following are equiveridical:
(i) there is a G-invariant functional such that (A, ) is a totally nite measure algebra;
(ii) whenever a A 0 and
n
)
nN
is a sequence in G,
n
a)
nN
is not disjoint;
(iii) G is fully non-paradoxical (denition: 394E).
proof (a) (i)(iii) by the argument of 394F, and (iii)(ii) by the criterion (ii) of 394E. So for the rest of
the proof I assume that (ii) is true and seek to prove (i).
(b) Let be such that (A, ) is a totally nite measure algebra. If a A 0, then inf
G
(a) > 0.
PPP??? Otherwise, let
n
)
nN
be a sequence in G such that
n
a 2
n
for each n N. Set b
n
= sup
kn

k
a
for each n; then inf
nN
b
n
= 0, so that
inf
nN
b
n
= 0, lim
n
(
n
a) = 0
for every Aut A. Choose n
i
)
iN
inductively so that
(
1
n
i

n
j
a) 2
j2
a
whenever i < j. Set
c = a \ sup
i<j

1
n
i

n
j
a.
Because

j=1

j1
i=0
(
1
n
i

n
j
a) < a,
c ,= 0, while
n
i
c
n
j
c = 0 whenever i < j, contrary to the hypothesis (ii). XXXQQQ
(c) For each G, dene w

L
0
= L
0
(A) and U

: L
2
L
2
as in 395A, where L
2
= L
2
(A, ).
If a A 0, then inf
G
_
a

> 0. PPP??? Otherwise, there is a sequence

n
)
nN
in G such that
_
a
v
n
4
n2
a for every n, where v
n
=

w

n
. In this case, (a [[v
n
2
n
]]) 2
n2
a for every n, so
that b = a \ sup
nN
[[v
n
2
n
]] is non-zero. But now
(
n
b) =
_
b
w

n
=
_
b
v
2
n
4
n
b 0
as n , contradicting (b) above. XXXQQQ
(d) Write e = 1 for the standard weak order unit of L
0
or L
2
. Let C L
2
be the convex hull of
U

e : G. Then C and its norm closure C are G-invariant in the sense that U

v C, U

v
t
C
whenever v C, v
t
C and G. By 3A5Ld, there is a unique u
0
C such that |u
0
|
2
|u|
2
for every
u C. Now if G, U

u
0
C, while |U

u
0
|
2
= |u
0
|
2
; so U

u
0
= u
0
. Also, if a A 0,
_
a
u
0
inf
uC
_
a
u = inf
uC
_
a
u
(because u
_
a
u is | |
2
-continuous)
= inf
G
_
a
U

e = inf
G
_
a
T

1
= inf
G
_
a

1 = inf
G
_
a

> 0
by (c). So [[u
0
> 0]] = 1.
(e) For a A, set a =
_
a
u
2
0
. Because u
0
L
2
, is a non-negative countably additive functional on A;
because [[u
2
0
> 0]] = [[u
0
> 0]] = 1, is strictly positive, and (A, ) is a totally nite measure algebra. Finally,
is G-invariant. PPP If a A and G, then
612 Measurable algebras 395B
(a) =
_
a
u
2
0
=
_
u
2
0
(a) =
_
T

(T

1u
2
0
a)
=
_
w

1u
2
0
a =
_
a
(T

1u
0

)
2
=
_
a
(U

1u
0
)
2
=
_
a
u
2
0
= a. QQQ
So (i) is true.
395C Remark If A is a Boolean algebra and G a subgroup of Aut A, a non-zero element a of A is called
weakly wandering if there is a sequence
n
)
nN
in G such that
n
a)
nN
is disjoint. Thus condition (ii)
of 395B may be read as there is no weakly wandering element of A.
395X Basic exercises (a) Let (A, ) be a totally nite measure algebra, and : A A an order-
continuous Boolean homomorphism. Let T

: L
0
(A) L
0
(A) be the corresponding Riesz homomorphism.
Show that there is a unique w

L
1
(A, ) such that
_
T

v =
_
v w

for every v L
0
(A)
+
.
(b) In 395A, show that the map U

: Aut A B(L
2
; L
2
) is injective.
(c) Let (A, ) be a probability algebra, Aut

A an ergodic measure-preserving automorphism, and
Aut A an automorphism. Suppose that
1
is measure-preserving. Show that is measure-preserving.
(Hint: compare w

, w

and w

in 395A.)
(d) Let A be a measurable algebra and G a subgroup of Aut A. Suppose that there is a strictly positive
G-invariant nitely additive functional on A. Show that there is a G-invariant such that (A, ) is a totally
nite measure algebra.
395Y Further exercises (a) Let A be a weakly (, )-distributive Dedekind complete Boolean algebra
and G a subgroup of Aut A. For a, b A, say that a and b are G-equidecomposable if there are nite
partitions of unity a
i
)
iI
in A
a
and b
i
)
iI
in A
b
, and a family
i
)
iI
in G, such that
i
a
i
= b
i
for every
i I. Show that the following are equiveridical: (i) G is fully non-paradoxical in the sense of 394E; (ii) if
a
n
)
nN
is a disjoint sequence of mutually G-equidecomposable elements of A, they must all be 0.
395 Notes and comments I have separated these few pages from 394 partly because 394 was already
up to full weight and partly in order that the ideas here should not be entirely overshadowed by those of
the earlier section. It will be evident that the construction of the U

in 395A, providing us with a faithful

representation, acting on a Hilbert space, of the whole group Aut A, is a basic tool for the study of that
group.
3A1E Set Theory 613
Appendix to Volume 3
Useful Facts
This volume assumes a fairly wide-ranging competence in analysis, a solid understanding of elementary
set theory and some straightforward Boolean algebra. As in previous volumes, I start with a few pages of
revision in set theory, but the absolutely essential material is in 3A2, on commutative rings, which is the
basis of the treatment of Boolean rings in 311. I then give three sections of results in analysis: topological
spaces (3A3), uniform spaces (3A4) and normed spaces (3A5). Finally, I add six sentences on group
theory (3A6).
3A1 Set Theory
3A1A The axioms of set theory This treatise is based on arguments within, or in principle reducible
to, ZFC, meaning Zermelo-Fraenkel set theory, including the Axiom of Choice. For discussions of this
system, see, for instance, Krivine 71, Jech 78 or Kunen 80. As I remarked in 2A1, I believe that it is
helpful, as a matter of general principle, to distinguish between results dependent on the axiom of choice and
those which can be proved without it, or with some relatively weak axiom such as countable choice. (See
134C.) In Volumes 1 and 2, such a distinction is useful in appreciating the special features of dierent ideas.
In the present volume, however, most of the principal theorems require something close to the full axiom of
choice, and there are few areas where it seems at present appropriate to work with anything weaker. Indeed,
at many points we shall approach questions which are, or may be, undecidable in ZFC; but I postpone
discussion of these to Volume 5. In particular, I specically exclude, for the time being, results dependent
on such axioms as the continuum hypothesis.
3A1B Denition Let X be a set. By an enumeration of X I mean a bijection f : X where
= #(X) (2A1Kb); more often than not I shall express such a function in the form x

)
<
. In this case
I say that the function f, or the family x

)
<
, enumerates X. You will see that I am tacitly assuming
that #(X) is always dened, that is, that the axiom of choice is true.
3A1C Calculation of cardinalities The following formulae are basic.
(a) For any sets X and Y , #(X Y ) max(, #(X), #(Y )). (Enderton 77, p. 64; Jech 78, p. 42;
Krivine 71, p. 33; Kunen 80, 10.13.)
(b) For any r N and any family X
i
)
ir
of sets, #(

r
i=0
X
i
) max(, max
ir
#(X
i
)). (Induce on r.)
(c) For any family X
i
)
iI
of sets, #(

iI
X
i
) max(, #(I), sup
iI
#(X
i
)). (Jech 78, p. 43; Krivine
71, p. 33; Kunen 80, 10.21.)
(d) For any set X, the set [X]
<
of nite subsets of X has cardinal at most max(, #(X)). (There is a
surjection from

rN
X
r
onto [X]
<
. For the notation [X]
<
see 3A1J below.)
3A1D Cardinal exponentiation For a cardinal , I write 2

for #(T). So 2

= c, and
+
2

for
every . (Enderton 77, p. 132; Lipschutz 64, p. 139; Jech 78, p. 24; Krivine 71, p. 25; Halmos 60,
p. 93.)
3A1E Denition The class of innite initial ordinals, or cardinals, is a subclass of the class On of
all ordinals, so is itself well-ordered; being unbounded, it is a proper class; consequently there is a unique
increasing enumeration of it as

)
On
. We have
0
= ,
+1
=
+

for every (compare 2A1Fc),

<

for non-zero limit ordinals . (Enderton 77, pp. 213-214; Jech 78, p. 25; Krivine 71, p.
31.)
614 Appendix 3A1F
3A1F Conal sets (a) If P is any partially ordered set (denition: 2A1A), a subset Q of P is conal
with P if for every p P there is a q Q such that p q.
(b) If P is any partially ordered set, the conality of P, cf(P), is the least cardinal of any conal subset
of P. Note that cf(P) = 0 i P = , and that cf(P) = 1 i P has a greatest element.
(c) Observe that if P is upwards-directed and cf(P) is nite, then cf(P) is either 0 or 1; for if Q is a nite,
non-empty conal set then it has an upper bound, which must be the greatest element of P.
(d) If P is a totally ordered set of conality , then there is a strictly increasing family p

)
<
in P such
that p

: < is conal with P. PPP If = 0 then P = and this is trivial. Otherwise, let Q be a conal
subset of P of cardinal , and q

: < an enumeration of Q. Dene p

)
<
inductively, as follows.
Start with p
0
= q
0
. Given p

)
<
, where < , then if p

< q

for every < , take p

= q

; otherwise,
because #() < , p

: < cannot be conal with P, so there is a p

P such that p

, p

for
every < , that is, p

< p

for every < . Note that there is some < such that q

, so that
q

. Continue.
Now p

)
<
is a strictly increasing family in P such that q

for every ; it follows at once that

p

: < is conal with P. QQQ

(e) In particular, for a totally ordered set P, cf(P) = i there is a conal strictly increasing sequence
in P.
3A1G Zorns Lemma In Volume 2 I used Zorns Lemma only once or twice, giving the arguments in
detail. In the present volume I feel that continuing in such a manner would often be tedious; but nevertheless
the arguments are not always quite obvious, at least until you have gained a good deal of experience. I
therefore take a paragraph to comment on some of the standard forms in which they appear.
The statement of Zorns Lemma, as quoted in 2A1M, refers to arbitrary partially ordered sets P. A large
proportion of the applications can in fact be represented more or less naturally by taking P to be a family
P of sets ordered by ; in such a case, it will be sucient to check that (i) P is not empty (ii)

Q P
for every non-empty totally ordered Q P. More often than not, this will in fact be true for all non-empty
upwards-directed sets Q P, and the line of the argument is sometimes clearer if phrased in this form.
Within this class of partially ordered sets, we can distinguish a special subclass. If A is any set and
any relation on A, we can consider the collection P of sets I A such that a b for all distinct a, b I.
In this case we need look no farther before declaring P has a maximal element; for necessarily belongs
to P, and if Q is any upwards-directed subset of P, then

Q P. PPP If a, b are distinct elements of

Q,
there are I
1
, I
2
Q such that a I
1
, b I
2
; because Q is upwards-directed, there is an I Q such that
I
1
I
2
I, so that a, b are distinct members of I P, and a b. QQQ So

Q is an upper bound of Q in P;
as Q is arbitrary, P satises the conditions of Zorns Lemma, and must have a maximal element.
Another important type of partially ordered set in this context is a family of functions, ordered by
saying that f g if g is an extension of f. In this case, for any non-empty upwards-directed , we
shall have a function h dened by saying that
domh =

f
domf, h(x) = f(x) whenever f , x domf,
and the usual attack is to seek to prove that any such h belongs to .
I nd that at least once I wish to use Zorns Lemma upside down: that is, I have a non-empty partially
ordered set P in which every non-empty totally ordered subset has a lower bound. In this case, of course,
P has a minimal element. The point is that the denition of partial order is symmetric, so that (P, ) is
a partially ordered set whenever (P, ) is; and we can seek to apply Zorns Lemma to either.
3A1H Natural numbers and nite ordinals I remarked in 2A1De that the rst few ordinals
, , , , , , , , . . .
may be identied with the natural numbers 0, 1, 2, 3, . . . ; the idea being that n = 0, 1, . . . , n 1 is a set
with n elements. If we do this, then the set N of natural numbers becomes identied with the rst innite
ordinal . This convention makes it possible to present a number of arguments in a particularly elegant
3A2A Rings 615
form. A typical example is in 344H. There I wish to describe an inductive construction for a family K
z
)
zS
where S =

nN
0, 1
n
. If we think of n as the set of its predecessors, then z 0, 1
n
becomes a function
from n to 0, 1; since the set n has just n members, this corresponds well to the idea of z as the list of
its n coordinates, except that it would now be natural to index them as z(0), . . . , z(n 1) rather than as

1
, . . . ,
n
, which was the language I favoured in Volume 2. An extension of z to a member of 0, 1
n+1
is
of the form v = z

i where v(k) = z(k) for k < n and v(n) = i. If w 0, 1

N
, then we can identify the
initial segment (w(0), w(1), . . . , w(n 1)) of its rst n coordinates with the restriction wn of w to the set
n = 0, . . . , n 1.
3A1I Denitions (a) If P and Q are lattices (2A1Ad), a lattice homomorphism from P to Q is a
function f : P Q such that f(p p
t
) = f(p) f(p
t
) and f(p p
t
) = f(p) f(p
t
) for all p, p
t
P. Such
a homomorphism is surely order-preserving (313H), for if p p
t
in P then f(p
t
) = f(p p
t
) = f(p) f(p
t
)
and f(p) f(p
t
).
(b) If P is a lattice, a sublattice of P is a set Q P such that p q and p q belong to Q for all p,
q Q.
(c) A lattice P is distributive if
(p q) r = (p r) (q r), (p q) r = (p r) (q r)
for all p, q, r P.
3A1J Subsets of given size The following concepts are used often enough for a special notation to
be helpful. If X is a set and is a cardinal, write
[X]

= A : A X, #(A) = ,
[X]

= A : A X, #(A) ,
[X]
<
= A : A X, #(A) < .
Thus
[X]
0
= [X]
0
= [X]
<1
= ,
[X]
2
is the set of doubleton subsets of X, [X]
<
is the set of nite subsets of X, [X]

is the set of countable

subsets of X, and so on.
3A1K The next result is one of the fundamental theorems of combinatorics. In this volume it is used
in the proofs of Ornsteins theorem (386) and the Kalton-Roberts theorem (392).
Halls Marriage Lemma Suppose that X and Y are nite sets and R X Y is a relation such that
#(R[I]) #(I) for every I X. Then there is an injective function f : X Y such that (x, f(x)) R for
every x X.
Remark Recall that R[I] is the set y : x I, (x, y) R (1A1Bc).
proof Bollob as 79, p. 54, Theorem 7; Anderson 87, 2.2.1; Bose & Manvel 84, 10.2.
3A2 Rings
I give a very brief outline of the indispensable parts of the elementary theory of (commutative) rings. I
assume that you have seen at least a little group theory.
3A2A Denition A ring is a triple (R, +, .) such that
(R, +) is a commutative group; its identity will always be denoted 0 or 0
R
;
(R, .) is a semigroup, that is, ab R for all a, b R and a(bc) = (ab)c for all a, b, c R;
a(b +c) = ab +ac, (a +b)c = ac +bc for all a, b, c R.
A commutative ring is one in which multiplication is commutative, that is, ab = ba for all a, b R.
616 Appendix 3A2B
3A2B Elementary facts Let R be a ring.
(a) a0 = 0a = 0 for every a R. PPP
a0 = a(0 + 0) = a0 +a0, 0a = (0 + 0)a = 0a + 0a;
because (R, +) is a group, we may subtract a0 or 0a from each side of the appropriate equation to see that
0 = a0, 0 = 0a. QQQ
(b) (a)b = a(b) = (ab) for all a, b R. PPP
ab + ((a)b) = (a + (a))b = 0b = 0 = a0 = a(b + (b)) = ab +a(b);
subtracting ab from each term, we get (a)b = (ab) = a(b). QQQ
3A2C Subrings If R is a ring, a subring of R is a set S R such that 0 S and a +b, ab, a belong
to S for all a, b S. In this case S, together with the addition and multiplication induced by those of R, is
a ring in its own right.
3A2D Homomorphisms (a) Let R, S be two rings. A function : R S is a ring homomorphism
if (a +b) = (a) +(b) and (ab) = (a)(b) for all a, b R. The kernel of is a : a R, (a) = 0
S
.
(b) Note that if : R S is a ring homomorphism, then it is also a group homomorphism from (R, +)
to (S, +), so that (0
R
) = 0
S
and (a) = (a) for every a R; moreover, [R] is a subring of S, and
is injective i its kernel is 0
R
.
(c) If R, S and T are rings, and : R S, : S T are ring homomorphisms, then : R T is a
ring homomorphism, because
()(a b) = ((a b)) = ((a) (b)) = ((a)) ((b))
for all a, b R, taking to be either addition or multiplication. If is bijective, then
1
: S R is a ring
homomorphism, because

1
(c d) =
1
((
1
(c)) (
1
(d))) =
1
(
1
(c)
1
(d)) =
1
(c)
1
(d)
for all c, d S, again taking to be either addition or multiplication.
3A2E Ideals (a) Let R be a ring. An ideal of R is a subring I of R such that ab I and ba I
whenever a I and b R. this case we write I R.
Note that R and 0 are always ideals of R.
(b) If R and S are rings and : R S is a ring homomorphism, then the kernel I of is an ideal of R.
PPP (i) Because is a group homomorphism, I is a subgroup of (R, +). (ii) If a I, b R then
(ab) = (a)(b) = 0
S
(b) = 0
S
, (ba) = (b)(a) = (b)0
S
= 0
S
so ab, ba I. QQQ
3A2F Quotient rings (a) Let R be a ring and I an ideal of R. A coset of I is a set of the form
a + I = a + x : x I where a R. (Because + is commutative, we do not need to distinguish between
left cosets a +I and right cosets I +a.) Let R/I be the set of cosets of I in R.
(b) For A, B R/I, set
A+B = x +y : x A, y B, A B = xy +z : x A, y B, z I.
Then A + B, A B both belong to R/I; moreover, if A = a + I and B = b + I, then A + B = (a + b) + I
and A B = ab +I. PPP (i)
3A2G Rings 617
A+B = (a +I) + (b +I)
= (a +x) + (b +y) : x, y I
= (a +b) + (x +y) : x, y I
(because addition is associative and commutative)
(a +b) +z : z I = (a +b) +I
(because I +I I)
= (a +b) + (z + 0) : z I
(a +I) + (b +I) = A+B
because 0 I. (ii)
A B = (a +x)(b +y) +z : x, y, z I
= ab + (ay +xb +z) : x, y, z I
ab +w : w I = ab +I
(because ay, xb I for all x, y I, and I is closed under addition)
= (a + 0)(b + 0) +w : w I
A B. QQQ
(c) It is now an elementary exercise to check that (R/I, +, ) is a ring, with zero 0 + I = I and additive
inverses (a +I) = (a) +I.
(d) Moreover, the map a a +I : R R/I is a ring homomorphism.
(e) Note that for a, b R, the following are equiveridical: (i) a b+I; (ii) b a+I; (iii) (a+I)(b+I) ,= ;
(iv) a+I = b+I; (v) ab I. Thus the cosets of I are just the equivalence classes in R under the equivalence
relation a b a + I = b + I; accordingly I shall generally write a

for a + I, if there seems no room

for confusion. In particular, the kernel of the canonical map from R to R/I is just a : a +I = I = I = 0

.
(f ) If R is commutative so is R/I, since
a

= (ab)

= (ba)

= b

for all a, b R.
3A2G Factoring homomorphisms through quotient rings: Proposition Let R and S be rings,
I an ideal of R, and : R S a homomorphism such that I is included in the kernel of . Then we have
a ring homomorphism : R/I S such that (a

) = (a) for every a R. is injective i I is precisely

the kernel of .
proof If a, b R and a

= b

in R/I, then ab I (3A2Fe), so (a)(b) = (ab) = 0, and (a) = (b).

This means that the formula oered does indeed dene a function from R/I to S. Now if a, b R and
is either multiplication or addition,
(a

) = ((a b)

) = (a b) = (a) (b) = (a

) (b

).
So is a ring homomorphism.
The kernel of is a

: (a) = 0, which is 0 i (a) = 0 a

= 0 a I.
618 Appendix 3A2H
3A2H Product rings (a) Let R
i
)
iI
be any family of rings. Set R =

iI
R
i
and for a, b R dene
a +b, ab R by setting
(a +b)(i) = a(i) +b(i), (ab)(i) = a(i)b(i)
for every i I. It is easy to check from the denition in 3A2A that R is a ring; its zero is given by the
formula
0
R
(i) = 0
R
i
for every i I,
and its additive inverses by the formula
(a)(i) = a(i) for every i I.
(b) Now let S be any other ring. Then it is easy to see that a function : S R is a ring homomorphism
i s (s)(i) : S R
i
is a ring homomorphism for every i I.
(c) Note that R is commutative i R
i
is commutative for every i.
3A3 General topology
In 2A3, I looked at a selection of topics in general topology in some detail, giving proofs; the point was
that an ordinary elementary course in the subject would surely go far beyond what we needed there, and
at the same time might omit some of the results I wished to quote. It seemed therefore worth taking a bit
of space to cover the requisite material, giving readers the option of delaying a proper study of the subject
until a convenient opportunity arose. In the context of the present volume, this approach is probably no
longer appropriate, since we need a much greater proportion of the fundamental ideas, and by the time you
have reached familiarity with the topics here you will be well able to nd your way about one of the many
excellent textbooks on the subject. This time round, therefore, I give most of the results without proofs (as
in 2A1 and 3A1), hoping that some of the references I oer will be accessible in all senses. I do, however,
give a full set of denitions, partly to avoid ambiguity (since even in this relatively mature subject, there
are some awkward divergences remaining in the usage of dierent authors), and partly because many of the
proofs are easy enough for even a novice to ll in with a bit of thought, once the meaning of the words is
clear. In fact this happens so often that I will mark with a those points where a proof needs an idea not
implicit in the preceding work.
3A3A Taxonomy of topological spaces I begin with the handful of denitions we need in order to
classify the dierent types of topological space used in this volume. A couple have already been introduced
in Volume 2, but I repeat them because the list would look so odd without them.
Denitions Let (X, T) be a topological space.
(a) X is Hausdor or T
2
if for any distinct points x, y X there are disjoint open sets G, H X such
that x G and y H.
(b) X is regular if whenever F X is closed and x X F there are disjoint open sets G, H X such
that x G and F H. (Note that in this denition I do not require X to be Hausdor, following James
87 but not Engelking 89, Bourbaki 66, Dugundji 66, Schubert 68 or Gaal 64.)
(c) X is completely regular if whenever F X is closed and x X F there is a continuous function
f : X [0, 1] such that f(x) = 1 and f(y) = 0 for every y F. (Note that many authors restrict the
phrase completely regular to Hausdor spaces. The terms Tychono space and T
3
1
2
space are also used
for Hausdor completely regular spaces.)
(d) X is zero-dimensional if whenever G X is an open set and x G then there is an open-and-closed
set H such that x H G.
3A3D General topology 619
(e) X is extremally disconnected if the closure of every open set in X is open.
(f ) X is compact if every open cover of X has a nite subcover.
(g) X is locally compact if for every x X there is a set K X such that x int K and K is compact
(in its subspace topology, as dened in 2A3C).
(h) If every subset of X is open, we call T the discrete topology on X.
3A3B Elementary relationships (a) A completely regular space is regular. (Engelking 89, p. 39;
Dugundji 66, p. 154; Schubert 68, p. 104.)
(b) A locally compact Hausdor space is completely regular, therefore regular. (Engelking 89, 3.3.1;
Dugundji 66, p. 238; Gaal 64, p. 149.)
(c) A compact Hausdor space is locally compact, therefore completely regular and regular.
(d) A regular extremally disconnected space is zero-dimensional. (Engelking 89, 6.2.25.)
(e) Any topology dened by pseudometrics (2A3F), in particular the weak topology of a normed space
(2A5I), is completely regular, therefore regular. (Bourbaki 66, IX.1.5; Dugundji 66, p. 200.)
(f ) If X is a completely regular Hausdor space (in particular, if X is (locally) compact and Hausdor),
and x, y are distinct points in X, then there is a continuous function f : X R such that f(x) ,= f(y).
(Apply 3A3Ac with F = y, which is closed because X is Hausdor.)
(g) An open set in a locally compact Hausdor space is locally compact in its subspace topology. (En-
gelking 89, 3.3.8; Bourbaki 66, I.9.7.)
3A3C Continuous functions Let (X, T) and (Y, S) be topological spaces.
(a) If f : X Y is a function and x X, we say that f is continuous at x if x int f
1
[H] whenever
H Y is an open set containing f(x).
(b) Now a function f : X Y is continuous i it is continuous at every point of X. (Bourbaki 66,
I.2.1; Dugundji 66, p. 80; Schubert 68, p. 24; Gaal 64, p. 183; James 87, p. 26.)
(c) If f : X Y is continuous at x X, and A X is such that x A, then f(x) f[A]. (Bourbaki
66, I.2.1; Schubert 68, p. 23.)
(d) If f : X Y is continuous, then f[A] f[A] for every A X. (Engelking 89, 1.4.1; Bourbaki
66, I.2.1; Dugundji 66, p. 80; Schubert 68, p. 24; Gaal 64, p. 184; James 87, p. 27.)
(e) A function f : X Y is a homeomorphism if it is a continuous bijection and its inverse is also
continuous; that is, if S = f[G] : G T and T = f
1
[H] : H S.
(f ) A function f : X [, ] is lower semi-continuous if x : x X, f(x) > is open for every
R. (Cf. 225H.)
3A3D Compact spaces Any extended series of applications of general topology is likely to involve
some new features of compactness. I start with the easy bits, continuing from 2A3Nb.
(a) The rst is just a denition of compactness in terms of closed sets instead of open sets. A family T of
sets has the nite intersection property if

T
0
is non-empty for every nite T
0
T. Now a topological
space X is compact i

T , = whenever T is a family of closed subsets of X with the nite intersection
property. (Engelking 89, 3.1.1; Bourbaki 66, I.9.1; Dugundji 66, p., 223; Schubert 68, p. 68; Gaal
64, p. 127.)
620 Appendix 3A3Db
(b) A marginal generalization of this is the following. Let X be a topological space and T a family of
closed subsets of X with the nite intersection property. If T contains a compact set then

T ,= . (Apply
(a) to K F : F T where K T is compact.)
(c) In a Hausdor space, compact subsets are closed. (Engelking 89, 3.1.8; Bourbaki 66, I.9.4;
Dugundji 66, p. 226; Schubert 68, p. 70; Gaal 64, p. 138; James 87, p. 77.)
(d) If X is compact, Y is Hausdor and : X Y is continuous and injective, then is a homeomorphism
between X and [X] (where [X] is given the subspace topology). (Engelking 89, 3.1.12; Bourbaki 66,
I.9.4; Dugundji 66, p. 226; Schubert 68, p. 71; Gaal 64, p. 207.)
(e) Let X be a regular topological space and A a subset of X. Then the following are equiveridical: (i)
A is relatively compact in X (that is, A is included in some compact subset of X, as in 2A3Na); (ii) A is
compact; (iii) every ultralter on X which contains A has a limit in X. PPP(ii)(i) is trivial, and (i)(iii)
is a consequence of 2A3R; neither of these requires X to be regular. Now assume (iii) and let T be an
ultralter on X containing A. Set
H = B : B X, there is an open set G T such that A G B.
Then H does not contain and B
1
B
2
H whenever B
1
, B
2
H, so H is a lter on X, and it contains
A. Let H

H be an ultralter (2A1O). By hypothesis, H

has a limit x say. Because A H

, X A is
an open set not belonging to H

, and cannot be a neighbourhood of x; thus x must belong to A. Let G be

an open set containing x. Then there is an open set H such that x H H G (this is where I use the
hypothesis that X is regular). Because H

x, H H

so XH does not belong to H

and therefore does

not belong to H. But X H is open, so by the denition of H it cannot belong to T. As T is an ultralter,
H T and G T. As G is arbitrary, T x. As T is arbitrary, A is compact (2A3R). Thus (iii)(ii). QQQ
3A3E Dense sets Recall that a set D in a topological space X is dense if D = X, and that X is
separable if it has a countable dense subset (2A3Ud).
(a) If X is a topological space, D X is dense and G X is dense and open, then G D is dense.
(Engelking 89, 1.3.6.) Consequently the intersection of nitely many dense open sets is always dense.
(b) If X and Y are topological spaces, D A X, D is dense in A and f : X Y is a continuous
function, then f[D] is dense in f[A]. (Use 3A3Cd.)
3A3F Meager sets Let X be a topological space.
(a) A set A X is nowhere dense or rare if int A = , that is, int(X A) = X A is dense, that is, for
every non-empty open set G there is a non-empty open set H G A.
(b) A set M X is meager, or of rst category, if it is expressible as the union of a sequence of
nowhere dense sets.
(c) Any subset of a nowhere dense set is nowhere dense; the union of nitely many nowhere dense sets is
nowhere dense. (3A3Ea.)
(d) Any subset of a meager set is meager; the union of countably many meager sets is meager. (314L.)
3A3G Baires theorem for locally compact Hausdor spaces Let X be a locally compact Haus-
dor space and G
n
)
nN
a sequence of dense open subsets of X. Then

nN
G
n
is dense. (Engelking
89, 3.9.4; Bourbaki 66, IX.5.3; Dugundji 66, p. 249; Schubert 68, p. 148.)
3A3H Corollary (a) Let X be a compact Hausdor space. Then a non-empty open subset of X cannot
be meager. (Dugundji 66, p. 250; Schubert 68, p. 147.)
(b) Let X be a non-empty locally compact Hausdor space. If A
n
)
nN
is a sequence of sets covering X,
then there is some n N such that int A
n
is non-empty. (Dugundji 66, p. 250.)
3A3L General topology 621
3A3I Product spaces (a) Denition Let X
i
)
iI
be a family of topological spaces, and X =

iI
X
i
their Cartesian product. We say that a set G X is open for the product topology if for every x G
there are a nite J I and a family G
j
)
jJ
such that every G
j
is an open set in the corresponding X
j
and
y : y X, y(j) G
j
for every j J
contains x and is included in G.
(Of course we must check that this does indeed dene a topology; see Engelking 89, 2.3.1; Bourbaki
66, I.4.1; Schubert 68, p. 38; Gaal 64, p. 144.)
(b) If X
i
)
iI
is a family of topological spaces, with product X, and Y another topological space, a
function : Y X is continuous i
i
is continuous for every i I, where
i
(x) = x(i) for x X, i I.
(Engelking 89, 2.3.6; Bourbaki 66, I.4.1; Dugundji 66, p. 101; Schubert 68, p. 62; James 87, p. 31.)
(c) Let X
i
)
iI
be any family of non-empty topological spaces, with product X. If T is a lter on X and
x X, then T x i
i
[[T]] x(i) for every i, where
i
(y) = y(i) for y X, and
i
[[T]] is the image
lter on X
i
(2A1Ib). (Bourbaki 66, I.7.6; Schubert 68, p. 61; James 87, p. 32.
(d) The product of any family of Hausdor spaces is Hausdor. (Engelking 89, 2.3.11; Bourbaki 66,
I.8.2; Schubert 68, p. 62; James 87, p. 87.)
(e) Let X
i
)
iI
be any family of topological spaces. If D
i
is a dense subset of X
i
for each i, then

iI
D
i
is dense in

iI
X
i
. (Engelking 89, 2.3.5.).
(f ) Let X
i
)
iI
be any family of topological spaces. If F
i
is a closed subset of X
i
for each i, then

iI
F
i
is closed in

iI
X
i
. (Engelking 89, 2.3.4; Bourbaki 66I.4.3.)
(g) Let (X
i
, T
i
))
iI
be a family of topological spaces with product (X, T). Suppose that each T
i
is dened
by a family P
i
of pseudometrics on X
i
(2A3F). Then T is dened by the family P =
i
: i I, P
i
of
pseudometrics on X, where I write
i
(x, y) = (
i
(x),
i
(y)) whenever i I, P
i
and x, y X, taking

i
to be the coordinate map from X to X
i
, as in (b)-(c). PPP (Compare 2A3Tb). (i) It is easy to check that
every
i
is a pseudometric on X. Write T
P
for the topology generated by P. (ii) If x G T
P
, let P
t
P
and > 0 be such that P
t
is nite and y : (y, x) for every P
t
is included in G. Express P
t
as

j
: j J, P
t
j
where J I is nite and P
t
j
P
j
is nite for each j J. Set
G
j
= t : t X
j
, (t,
j
(x)) < for every P
t
j

for every j J. Then G

t
= y :
j
(y) G
j
for every j J contains x, belongs to T and is included in G.
As x is arbitrary, G T; as G is arbitrary, T
P
T. (iii) Now every
i
is (T
P
, T
i
)-continuous, by 2A3H; by
(b) above, the identity map from X to itself is (T
P
, T)-continuous, that is, T
P
T and T
P
= T, as claimed.
QQQ
3A3J Tychonos theorem The product of any family of compact topological spaces is compact.
proof Engelking 89, 3.2.4; Bourbaki 66, I.9.5; Dugundji 66, p. 224; Schubert 68, p. 72; Gaal 64,
p. 146 and p. 272; James 87, p. 67.
3A3K The spaces 0, 1
I
, R
I
For any set I, we can think of 0, 1
I
as the product

iI
X
i
where
X
i
= 0, 1 for each i. If we endow each X
i
with its discrete topology, the product topology is the usual
topology on 0, 1
I
. Being a product of Hausdor spaces, it is Hausdor; by Tychonos theorem, it is
compact. A subset G of 0, 1
I
is open i for every x G there is a nite J I such that y : y
0, 1
I
, yJ = xJ G.
Similarly, the usual topology of R
I
is the product topology when each factor is given its Euclidean
topology (cf. 2A3Tc).
3A3L Cluster points of lters (a) Let X be a topological space and T a lter on X. A point x of
X is a cluster point (or adherence point or accumulation point) of T if x A for every A T.
622 Appendix 3A3Lb
(b) For any topological space X, lter T on X and x X, x is a cluster point of T i there is a lter
( T such that ( x. (Engelking 89, 1.6.8; Bourbaki 66, I.7.2; Gaal 64, p. 260; James 87, p. 22.)
(c) If
n
)
nN
is a sequence in R, R and lim
n1

n
= for every non-principal ultralter H
on N (denition: 2A3Sb), then lim
n

n
= . PPP??? If
n
)
nN
, , there is some > 0 such that
I = n : [
n
[ is innite. Now T
0
= F : F N, I F is nite is a lter on N, so there is an
ultralter T T
0
. But now cannot be lim
nJ

n
. XXXQQQ
3A3M Topology bases (a) If X is a set and / is any family of subsets of X, the topology generated
by / is the smallest topology on X including /.
(b) If X is a set and T is a topology on X, a base for T is a set | T such that whenever x G T
there is a U | such that x U G; that is, such that T =

( : ( |. In this case, of course, |

generates T.
(c) If X is a set and c is a family of subsets of X, then c is a base for a topology on X i (i) whenever
E
1
, E
2
c and x E
1
E
2
then there is an E c such that x E E
1
E
2
(ii)

c = X. (Engelking
89, p. 12.)
3A3N Uniform convergence (a) Let X be a set, (Y, ) a metric space and f
n
)
nN
a sequence of
functions from X to Y . We say that f
n
)
nN
converges uniformly to a function f : X Y if for every
> 0 there is an n
0
N such that (f
n
(x), f(x)) whenever n n
0
and x X.
(b) Let X be a topological space and (Y, ) a metric space. Suppose that f
n
)
nN
is a sequence of
continuous functions from X to Y converging uniformly to f : X Y . Then f is continuous. (Engelking
89, 1.4.7/4.2.19; Gaal 64, p. 202.)
3A3O One-point compactications Let (X, T) be a locally compact Hausdor space. Take any
object x

not belonging to X and set X

= X x

. Let T

be the family of those sets H X

such
that H X T and either x

/ H or X H is compact (for T). Then T

is the unique compact Hausdor

topology on X

inducing T as the subspace topology on X; (X

, T

) is the one-point compactication

or Alexandro compactication of (X, T). (Engelking 89, 3.5.11; Bourbaki 66, I.9.8; Dugundji 66,
p. 246.)
3A3P Miscellaneous denitions Let X be a topological space.
(a) A subset F of X is a zero set or functionally closed if it is of the form f
1
[0] for some continuous
function f : X R. A subset G of X is a cozero set or functionally open if its complement is a zero
set.
(b) An isolated point of X is a point x X such that the singleton set x is open.
3A4 Uniformities
I continue the work of 3A3 with some notes on uniformities, so as to be able to discuss completeness
and the extension of uniformly continuous functions in non-metrizable contexts (3A4F-3A4H). As in 3A3,
most of the individual steps are elementary; I mark exceptions with a .
3A4A Uniformities (a) Let X be any set. A uniformity on X is a lter J on X X such that
(i) (x, x) W for every x X, W J;
(ii) for every W J, W
1
= (y, x) : (x, y) W J;
(iii) for every W J, there is a V J such that
V

V = (x, z) : y, (x, y) V &(y, z) V W.
It is convenient to allow the special case X = , J = , even though this is not properly speaking a lter.
The pair (X, J) is now a uniform space.
3A4E Uniformities 623
(b) If J is a uniformity on a set X, the associated topology T is the set of sets G X such that
for every x G there is a W J such that W[x] =y : (x, y) W G.
(Engelking 89, 8.1.1; Bourbaki 66, II.1.2; Gaal 64, p. 48; Schubert 68, p. 115; James 87, p. 101.)
(c) We say that a uniformity is Hausdor if the associated topology is Hausdor.
(d) If U is a linear topological space, then it has an associated uniformity
J = W : W U U, there is an open set G containing 0
such that (u, v) W whenever u v G.
(Schaefer 66, I.1.4.)
3A4B Uniformities and pseudometrics (a) If P is a family of pseudometrics on a set X, then the
associated uniformity is the smallest uniformity on X containing all the sets W(; ) = (x, y) : (x, y) <
as runs through P, through ]0, [. (Engelking 89, 8.1.18; Bourbaki 66, IX.1.2.)
(b) If J is the uniformity dened by a family P of pseudometrics, then the topology associated with J
is the topology dened from P (2A3F). (Dugundji 66, p. 203.)
(c) A uniformity J is metrizable if it can be dened by a single metric.
(d) If U is a linear space with a topology dened from a family of functionals : U [0, [ such that
(u + v) (u) + (v), (u) (u) when [[ 1, and lim
0
(u) = 0 (2A5B), the uniformity dened
from the topology (3A4Ad) coincides with the uniformity dened from the pseudometrics

(u, v) = (uv).
(Immediate from the denitions.)
3A4C Uniform continuity (a) If (X, J) and (Y, 1) are uniform spaces, a function : X Y is
uniformly continuous if (x, y) : ((x), (y)) V J for every V 1.
(b) The composition of uniformly continuous functions is uniformly continuous. (Bourbaki 66, II.2.1;
Schubert 68, p. 118.)
(c) If uniformities J, 1 on sets X, Y are dened by non-empty families P, of pseudometrics, then a
function : X Y is uniformly continuous i for every , > 0 there are
0
, . . . ,
n
P and > 0
such that ((x), (x
t
)) whenever x, x
t
X and max
in

i
(x, x
t
) . (Elementary verication.)
(d) A uniformly continuous function is continuous for the associated topologies. (Bourbaki 66, II.2.1;
Schubert 68, p. 118; James 87, p. 102.)
3A4D Subspaces (a) If (X, J) is a uniform space and Y is any subset of X, then J
Y
= W(Y Y ) :
W J is a uniformity on Y ; it is the subspace uniformity. (Bourbaki 66, II.2.4; Schubert 68, p.
122.)
(b) If J denes a topology T on X, then the topology dened by J
Y
is the subspace topology on Y , as
dened in 2A3C. (Schubert 68, p. 122; James 87, p. 103.)
(c) If J is dened by a family P of pseudometrics on X, then J
Y
is dened by Y Y : P.
(Elementary verication.)
3A4E Product uniformities (a) If (X, |) and (Y, 1) are uniform spaces, the product uniformity
is the smallest uniformity J on X Y containing all sets of the form
((x, y), (x
t
, y
t
)) : (x, x
t
) U, (y, y
t
) V
as U runs through | and V through 1. (Engelking 89, 8.2; Bourbaki 66, II.2.6; Schubert 68, p. 124;
James 87, p. 93.)
624 Appendix 3A4Eb
(b) If |, 1 are dened from families P, of pseudometrics, then J will be dened by the family
: P

: , writing
((x, y), (x
t
, y
t
)) = (x, x
t
),

((x, y), (x
t
, y
t
)) = (y, y
t
)
as in 2A3Tb. (Elementary verication.)
(c) If (X, |), (Y, 1) and (Z, J) are uniform spaces, a map : Z X Y is uniformly continuous i the
coordinate maps
1
: Z X and
2
: Z Y are uniformly continuous. (Engelking 89, 8.2.1; Bourbaki
66, II.2.6; Schubert 68, p. 125; James 87, p. 93.)
3A4F Completeness (a) If J is a uniformity on a set X, a lter T on X is Cauchy if for every
W J there is an F T such that F F W.
Any convergent lter in a uniform space is Cauchy. (Bourbaki 66, II.3.1; Gaal 64, p. 276; Schubert
68, p. 134; James 87, p. 109.)
(b) A uniform space is complete if every Cauchy lter is convergent.
(c) If J is dened from a family P of pseudometrics, then a lter T on X is Cauchy i for every P
and > 0 there is an F T such that (x, y) for all x, y F; equivalently, for every P, > 0 there
is an x X such that U(x; ; ) T. (Elementary verication.)
(d) A complete subspace of a Hausdor uniform space is closed. (Engelking 89, 8.3.6; Bourbaki 66,
II.3.4; Schubert 68, p. 135; James 87, p. 148.)
(e) A metric space is complete i every Cauchy sequence converges (cf. 2A4Db). (Schubert 68, p. 141;
Gaal 64, p. 276; James 87, p. 150.)
(f ) If (X, ) is a complete metric space, D X a dense subset, (Y, ) a metric space and f : X Y is
an isometry (that is, (f(x), f(x
t
)) = (x, x
t
) for all x, x
t
X), then f[X] is precisely the closure of f[D]
in Y . (For it must be complete, and and we use (d).)
(g) If U is a linear space with a linear space topology and the associated uniformity (3A4Ad), then a
lter T on U is Cauchy i for every open set G containing 0 there is an F T such that F F G (cf.
2A5F). (Immediate from the denitions.)
3A4G Extension of uniformly continuous functions: Theorem If (X, J) is a uniform space,
(Y, 1) is a complete uniform space, D X is a dense subset of X, and : D Y is uniformly continuous
(for the subspace uniformity of D), then there is a uniformly continuous

: X Y extending . If Y
is Hausdor, the extension is unique. (Engelking 89, 8.3.10; Bourbaki 66, II.3.6; Gaal 64, p. 300;
Schubert 68, p. 137; James 87, p. 152.)
In particular, if (X, ) is a metric space, (Y, ) is a complete metric space, D X is a dense subset, and
: D Y is an isometry, then there is a unique isometry

: X Y extending .
3A4H Completions (a) Theorem If (X, J) is any Hausdor uniform space, then we can nd a
complete Hausdor uniform space (

X,

J) in which X is embedded as a dense subspace; moreover, any two
such spaces are essentially unique. (Engelking 89, 8.3.12; Bourbaki 66, II.3.7; Gaal 64, p. 297 & p.
300; Schubert 68, p. 139; James 87, p. 156.)
(b) Such a space (

X,

J) is called a completion of (X, J). Because it is unique up to isomorphism as
a uniform space, we may call it the completion.
(c) If J is the uniformity dened by a metric on a set X, then there is a unique extension of to a
metric on

X dening the uniformity

J. (Bourbaki 66, IX.1.3.)
3A4I A note on metric spaces I mention some elementary facts which are used in 386. Let (X, )
be a metric space. If x X and A X is non-empty, set
3A5C Normed spaces 625
(x, A) = inf
yA
(x, y).
Then (x, A) = 0 i x A (2A3Kb). If B X is another non-empty set, then
(x, B) (x, A) + sup
yA
(y, B).
In particular, (x, A) = (x, A). If A
n
)
nN
is a non-decreasing sequence of non-empty sets with union A,
then
(x, A) = lim
n
(x, A
n
).
3A5 Normed spaces
I run as quickly as possible over the results, nearly all of them standard elements of any introductory
course in functional analysis, which I nd myself calling on in this volume. As in the corresponding section
of Volume 2 (2A4), a large proportion of these are valid for both real and complex normed spaces, but
as the present volume is almost exclusively concerned with real linear spaces I leave this unsaid, except in
3A5L, and if in doubt you should suppose that scalars belong to the eld R.
3A5A The Hahn-Banach theorem: analytic forms This is one of the central ideas of functional
analysis, both nite- and innite-dimensional, and appears in a remarkable variety of forms. I list those
formulations which I wish to quote, starting with those which are more or less analytic, according to the
classication of Bourbaki 87. Recall that if U is a normed space I write U

for the Banach space of

bounded linear functionals on U.
(a) Let U be a linear space and p : U [0, [ a functional such that p(u + v) p(u) + p(v) and
p(u) = p(u) whenever u, v U and 0. Then for any u
0
U there is a linear functional f : U R
such that f(u
0
) = p(u
0
) and f(u) p(u) for every u U. Rudin 91, 3.2; Dunford & Schwartz 57,
II.3.10.
(b) Let U be a normed space and V a linear subspace of U. Then for any f V

there is a g U

,
extending f, with |g| = |f|. (363R; Bourbaki 87, II.3.2; Rudin 91, 3.3; Dunford & Schwartz 57,
II.3.11; Lang 93, p. 69; Wilansky 64, p. 66; Taylor 64, 3.7-B & 4.3-A.)
(c) If U is a normed space and u U there is an f U

such that |f| 1 and f(u) = |u|. (Bourbaki

87, II.3.2; Rudin 91, 3.3; Dunford & Schwartz 57, II.3.14; Wilansky 64, p. 67; Taylor 64, 3.7-C &
4.3-B.)
(d) If U is a normed space and V U is a linear subspace which is not dense, then there is a non-zero
f U

such that f(v) = 0 for every v V . (Rudin 91, 3.5; Dunford & Schwartz 57, II.3.12; Taylor
64, 4.3-D.)
(e) If U is a normed space, U

separates the points of U. (Rudin 91, 3.4; Lang 93, p. 70; Dunford &
Schwartz 57, II.3.14.)
3A5B Cones (a) Let U be a linear space. A convex cone (with apex 0) is a set C U such that
u + v C whenever u, v C and , 0. The intersection of any family of convex cones is a convex
cone, so for every subset A of U there is a smallest convex cone including A.
(b) Let U be a normed space. Then the closure of a convex cone is a convex cone, and the closure of a
linear subspace is a linear subspace. (Bourbaki 87, II.2.6; Dunford & Schwartz 57, V.2.1.)
3A5C Hahn-Banach theorem: geometric forms (a) Let U be a normed space and C U a convex
set such that |u| 1 for every u C. Then there is an f U

such that |f| 1 and f(u) 1 for every

u C. (Dunford & Schwartz 57, V.1.12.)
626 Appendix 3A5Cb
(b) Let U be a normed space and B U a non-empty convex set such that 0 / B. Then there is an
f U

such that inf

uB
f(u) > 0. (Bourbaki 87, II.4.1; Rudin 91, 3.4; Lang 93, p. 70; Dunford &
Schwartz 57, V.2.12.)
(c) Let U be a normed space, B a closed convex subset of U containing 0, and u a point of U B. Then
there is an f U

such that f(u) > 1 and f(v) 1 for every v B. (Apply (b) to B u to nd a g U

such that g(u) < inf

vB
g(v) and now set f =
1

g where g(u) < < inf

vB
g(v)).
3A5D Separation from nitely-generated cones Let U be a linear space over R and u, v
0
, . . . , v
n
points of U such that u does not belong to the convex cone generated by v
0
, . . . , v
n
. Then there is a linear
functional f : U R such that f(v
i
) 0 for every i and f(u) < 0.
proof (a) If U is nite-dimensional this is covered by Gale 60, p. 56.
(b) For the general case, let V be the linear subspace of U generated by u, v
0
, . . . , v
n
. Then there is
a linear functional f
0
: V R such that f
0
(u) < 0 f
0
(v
i
) for every i. By Zorns Lemma, there is a
maximal linear subspace W U such that W V = 0. Now W + V = U (for if u / W + V , the linear
subspace W
t
generated by W u still has trivial intersection with V ), so we have an extension of f
0
to
a linear functional f : U R dened by setting f(v + w) = f
0
(v) whenever v V and w W. Now
f(u) < 0 min
in
f(v
i
), as required.
3A5E Weak topologies (a) Let U be any linear space over R and W a linear subspace of the space
U
t
of all linear functionals from U to R. Then I write T
s
(U, W) for the linear space topology dened by the
method of 2A5B from the functionals u [f(u)[ as f runs through W. (Bourbaki 87, II.6.2; Rudin 91,
3.10; Dunford & Schwartz 57, V.3.2; Taylor 64, 3.81.)
(b) I note that the weak topology of a normed space U (2A5Ia) is T
s
(U, U

), while the weak* topology

of U

(2A5Ig) is T
s
(U

, W) where W is the canonical image of U in U

. (Rudin 91, 3.14.)

(c) Let U and V be linear spaces over R and T : U V a linear operator. If W U
t
and Z V
t
are
such that gT W for every g Z, then T is continuous for T
s
(U, W) and T
s
(V, Z). (Bourbaki 87, II.6.4.)
(d) If U and V are normed spaces and T : U V is a bounded linear operator then we have an adjoint
(or conjugate, or dual) operator T
t
: V

dened by saying that T

t
g = gT for every g V

. T
t
is linear and is continuous for the weak* topologies of U

and V

. (Bourbaki 87, II.6.4; Dunford &

Schwartz 57, VI.2; Taylor 64, 4.5.)
(e) If U is a normed space and A U is convex, then the closure of A for the norm topology is the same
as the closure of A for the weak topology of U. In particular, norm-closed convex subsets (for instance,
norm-closed linear subspaces) of U are closed for the weak topology. (Rudin 91, 3.12; Lang 93, p. 88;
Dunford & Schwartz 57, V.3.13.)
3A5F Weak* topologies: Theorem If U is a normed space, the unit ball of U

is compact and
Hausdor for the weak* topology. (Rudin 91, 3.15; Lang 93, p. 71; Dunford & Schwartz 57, V.4.2;
Taylor 64, 4.61-A.)
3A5G Reexive spaces (a) A normed space U is reexive if every member of U

is of the form
f f(u) for some u U.
(b) A normed space is reexive i bounded sets are relatively weakly compact. (Dunford & Schwartz
57, V.4.8; Taylor 64, 4.61-C.)
(c) If U is a reexive space, u
n
)
nN
is a bounded sequence in U and T is an ultralter on N, then
lim
nJ
u
n
is dened in U for the weak topology. (Use (b) and the argument of 2A3Se.)
3A5Ld Normed spaces 627
3A5H (a) Uniform Boundedness Theorem Let U be a Banach space, V a normed space, and
A B(U; V ) a set such that Tu : T A is bounded in V for every u U. Then A is bounded in B(U; V ).
(Rudin 91, 2.6; Dunford & Schwartz 57, II.3.21; Taylor 64, 4.4-E.)
(b) Corollary If U is a normed space and A U is such that f[A] is bounded for every f U

, then A
is bounded. (Wilansky 64, p. 117; Taylor 64, 4.4-AS.) Consequently any relatively weakly compact set
in U is bounded. (Rudin 91, 3.18.)
3A5I Completions Let U be a normed space.
(a) Recall that U has a metric associated with the norm (2A4Bb), and that the topology dened by
is a linear space topology (2A5D, 2A5B). This topology denes a uniformity J (3A4Ad) which is also the
uniformity dened by (3A4Bd). The norm itself is a uniformly continuous function from U to R (because
[|u| |v|[ |u v| for all u, v U).
(b) Let (

U,

J) be the uniform space completion of (U, J) (3A4H). Then addition and scalar multiplication
and the norm extend uniquely to make

U a Banach space. (Schaefer 66, I.1.5; Lang 93, p. 78.)
(c) If U and V are Banach spaces with dense linear subspaces U
0
and V
0
, then any norm-preserving
isomorphism between U
0
and V
0
extends uniquely to a norm-preserving isomorphism between U and V (use
3A4G).
3A5J Normed algebras If U is a normed algebra (2A4J), its multiplication, regarded as a function
from U U to U, is continuous. (Wilansky 64, p. 259.)
3A5K Compact operators Let U and V be Banach spaces.
(a) A linear operator T : U V is compact or completely continuous if Tu : |u| 1 is relatively
compact in V for the topology dened by the norm of V .
(b) A linear operator T : U V is weakly compact if Tu : |u| 1 is relatively weakly compact in
V .
3A5L Hilbert spaces I mentioned the phrases inner product space, Hilbert space briey in 244N-
244O, without explanation, as I did not there rely on any of the abstract theory of these spaces. For the
main result of 395 we need one of their fundamental properties, so I now skim over the denitions.
(a) An inner product space is a linear space U over
R
C
together with an operator ( [ ) : U U
R
C
such that
(u
1
+u
2
[v) = (u
1
[v) + (u
2
[v), (u[v) = (u[v), (u[v) = (v[u)
(the complex conjugate of (u[v)),
(u[u) 0, u = 0 whenever (u[u) = 0
for all u, u
1
, u
2
, v U and
R
C
.
(b) If U is any inner product space, we have a norm on U dened by setting |u| =
_
(u[u) for every
u U. (Taylor 64, 3.2-B.)
(c) A Hilbert space is an inner product space which is a Banach space under the norm of (b) above,
that is, is complete in the metric dened from its norm.
(d) If U is a Hilbert space, C U is a non-empty closed convex set, and u U, then there is a unique
v C such that |u v| = inf
wC
|u w|. (Taylor 64, 4.81-A.)
628 Appendix 3A6 intro.
3A6 Group Theory
For Chapter 38 we need four denitions and two results from elementary abstract group theory.
3A6A Denition If G is a group, I will say that an element g of G is an involution if its order is 2,
that is, g
2
= e, the identity of G, but g ,= e.
3A6B Denition If G is a group, the set Aut G of automorphisms of G (that is, bijective homo-
morphisms from G to itself) is a group. For g G dene g : G G by writing g(h) = ghg
1
for every
h G; then g Aut G, and the map g g is a homomorphism from G onto a normal subgroup J of Aut G
(Rotman 84, p. 130). We call J the group of inner automorphisms of G. Members of (Aut G) J are
called outer automorphisms.
3A6C Normal subgroups For any group G, the family of normal subgroups of G, ordered by , is
a Dedekind complete lattice, with H K = HK and H K = H K. (Davey & Priestley 90, 2.8 &
2.19.)
Group Theory 629
References for Volume 3
Anderson I. [87] Combinatorics of Finite Sets. Oxford U.P., 1987. [332Xk, 3A1K.]
Balcar B., Glowczy nski W. & Jech T. [98] The sequential topology on complete Boolean algebras,
Fundamenta Math. 155 (1998) 59-78. [393Yd.]
Becker H. & Kechris A.S. [96] The descriptive set theory of Polish group actions. Cambridge U.P., 1996
(London Math. Soc. Lecture Note Series 232). [394 notes.]
Bekkali M. & Bonnet R. [89] Rigid Boolean algebras, pp. 637-678 in Monk 89. [383L.]
Bellow A. & Kolzow D. [76] Measure Theory, Oberwolfach 1975. Springer, 1976 (Lecture Notes in Math-
ematics 541).
Billingsley P. [65] Ergodic Theory and Information. Wiley, 1965. [372 notes.]
Bollobas B. [79] Graph Theory. Springer, 1979. [332Xk, 3A1K.]
Bose R.C. & Manvel B. [84] Introduction to Combinatorial Theory. Wiley, 1984. [3A1K.]
Bourbaki N. [66] General Topology. Hermann/Addison-Wesley, 1968. [393 notes, 3A3, 3A4.]
Bourbaki N. [68] Theory of Sets. Hermann/Addison-Wesley, 1968. [315 notes.]
Bourbaki N. [87] Topological Vector Spaces. Springer, 1987. [3A5.]
Bukhvalov A.V. [95] Optimization without compactness, and its applications, pp. 95-112 in Huijsmans
Kaashoek Luxemburg & de Pagter 95. [367U.]
Burke M.R. [93] Liftings for Lebesgue measure, pp. 119-150 in Judah 93. [341Lg, 345F.]
Burke M.R. [n95] Consistent liftings, privately circulated, 1995. [346Yc.]
Burke M.R. [n96] An example of S.Shelah on liftings for ccc algebras, privately circulated, 1996. [341Lh.]
Burnside W. [11] Theory of Groups of Finite Order. Cambridge U.P., 1911 (reprinted by Dover, 1955).
[383 notes.]
Chacon R.V. & Krengel U. [64] Linear modulus of a linear operator, Proc. Amer. Math. Soc. 15 (1964)
553-559. [371 notes.]
Choksi J.R. & Prasad V.S. [82] Ergodic theory of homogeneous measure algebras, pp. 367-408 of K olzow
& Maharam-Stone 82. [382I.]
Coleman A.J. & Ribenboim P. [67] (eds.) Proceedings of the Symposium in Analysis, Queens University,
June 1967. Queens University, Kingston, Ontario, 1967.
Comfort W.W. & Negrepontis S. [82] Chain Conditions in Topology. Cambridge U.P., 1982. [391L, 391
notes.]
Cziszar I. [67] Information-type measures of dierence of probability distributions and indirect observa-
tions, Studia Scientiarum Math. Hungarica 2 (1967) 299-318. [385J.]
Davey B.A. & Priestley H.A. [90] Introduction to Lattices and Order, Cambridge U.P., 1990. [3A6C.]
Dow A. & Steprans J. [93] The -linkedness of the measure algebra, Canad. Math. Bulletin 37 (1993)
42-45. [391 notes.]
Dugundji J. [66] Topology. Allyn & Bacon, 1966. [3A3, 3A4B.]
Dunford N. [36] Integration and linear operators, Trans. Amer. Math. Soc. 40 (1936) 474-494. [376
notes.]
Dunford N. & Schwartz J.T. [57] Linear Operators I. Wiley, 1957 (reprinted 1988). [356 notes, 371
notes, 376 notes, 3A5.]
Dye H.A. [59] On groups of measure preserving transformations I, Amer. J. Math. 81 (1959) 119-159.
[387 notes.]
Eigen S.J. [82] The group of measure-preserving transformations of [0, 1] has no outer automorphisms,
Math. Ann. 259 (1982) 259-270. [383D.]
Engelking R. [89] General Topology. Heldermann, 1989 (Sigma Series in Pure Mathematics 6). [393
notes, 3A3, 3A4.]
Enderton H.B. [77] Elements of Set Theory. Academic, 1977. [3A1.]
Erdos P. & Oxtoby J.C. [55] Partitions of the plane into sets having positive measure in every non-null
product set, Trans. Amer. Math. Soc. 79 (1955) 91-102. [325 notes.]
Fathi A. [78] Le groupe des transformations de [0, 1] qui preservent la mesure de Lebesgue est un groupe
simple, Israel J. Math. 29 (1978) 302-308. [381 notes, 382I.]
630 References
Fremlin D.H. [74a] Topological Riesz Spaces and Measure Theory. Cambridge U.P., 1974. [Chap. 35 intro,
354Yb, 355 notes, 356 notes, 363 notes, 371 notes, 376 notes.]
Fremlin D.H. [74b] A characterization of L-spaces, Indag. Math. 36 (1974) 270-275. [371 notes.]
Fremlin D.H. [75] Inextensible Riesz spaces, Math. Proc. Cambridge Phil. Soc. 77 (1975) 71-89. [393L.]
Fremlin D.H. [84] Consequences of Martins Axiom. Cambridge U.P., 1984. [316 notes.]
Fremlin D.H. [89] Measure algebras, pp. 876-980 in Monk 89. [341Lg, 341Zb.]
Frolk Z. [68] Fixed points of maps of extremally disconnected spaces and complete Boolean algebras,
Bull. Acad. Polon. Sci. 16 (1968) 269-275. [381 notes.]
Gaal S.A. [64] Point Set Topology. Academic, 1964. [3A3, 3A4.]
Gaifman H. [64] Concerning measures on Boolean algebras, Pacic J. Math. 14 (1964) 61-73. [391L.]
Gale D. [60] The theory of linear economic models. McGraw-Hill, 1960. [3A5D.]
Gnedenko B.V. & Kolmogorov A.N. [54] Limit Distributions for Sums of Independent Random Variables.
Addison-Wesley, 1954. [342 notes.]
Graf S. & Weizsacker H.von [76] On the existence of lower densities in non-complete measure spaces,
pp. 155-158 in Bellow & K olzow 76. [341Lc.]
Hajian A. & Ito Y. [69] Weakly wandering sets and invariant measures for a group of transformations,
J. of Math. and Mech. 18 (1969) 1203-1216. [395B.]
Hajian A., Ito Y. & Kakutani S. [75] Full groups and a theorem of Dye, Advances in Math. 17 (1975)
48-59. [387 notes.]
Halmos P.R. [60] Naive Set Theory. Van Nostrand, 1960. [3A1D.]
Huijsmans C.B., Kaashoek M.A., Luxemburg W.A.J. & de Pagter B. [95] (eds.) Operator Theory in
Function Spaces and Banach Lattices. Birkhauser, 1995.
Ionescu Tulcea C. & Ionescu Tulcea A. [69] Topics in the Theory of Lifting. Springer, 1969. [341 notes.]
James I.M. [87] Topological and Uniform Spaces. Springer, 1987. [3A3, 3A4.]
Jech T. [78] Set Theory. Academic, 1978. [316 notes, 3A1.]
Johnson R.A. [80] Strong liftings which are not Borel liftings, Proc. Amer. Math. Soc. 80 (1980) 234-236.
[345F.]
Judah H. [93] (ed.) Proceedings of the Bar-Ilan Conference on Set Theory and the Reals, 1991. Amer.
Math. Soc. (Israel Mathematical Conference Proceedings 6), 1993.
Juhasz I. [71] Cardinal Functions in Topology. North-Holland, 1975 (Math. Centre Tracts 34). [332
notes.]
Kakutani S. [41] Concrete representation of abstract L-spaces and the mean ergodic theorem, Annals of
Math. 42 (1941) 523-537. [369E.]
Kalton N.J., Peck N.T. & Roberts J.W. [84] An F-space sampler, Cambridge U.P., 1984. [375 notes.]
Kalton N.J. & Roberts J.W. [83] Uniformly exhaustive submeasures and nearly additive set functions,
Trans. Amer. Math. Soc. 278 (1983) 803-816. [392D, 392 notes.]
Kawada Y. [44]

Uber die Existenz der invarianten Integrale, Jap. J. Math. 19 (1944) 81-95. [394 notes.]
Kelley J.L. [59] Measures on Boolean algebras, Pacic J. Math. 9 (1959) 1165-1177. [391L.]
Kolmogorov A.N. [58] New metric invariants of transitive dynamical systems and automorphisms of
Lebesgue spaces, Dokl. Akad. Nauk SSSR 119 (1958) 861-864. [384P.]
Kolzow D. & Maharam-Stone D. [82] (eds.) Measure Theory Oberwolfach 1981. Springer, 1982 (Lecture
Notes in Math. 945).
Koppelberg S. [89] General Theory of Boolean Algebras, vol. 1 of Monk 89. [Chap. 31 intro, 332 notes.]
Kothe G. [69] Topological Vector Spaces I. Springer, 1969. [356 notes.]
Kranz P. & Labuda I. [93] (eds.) Proceedings of the Orlicz Memorial Conference, 1991, unpublished
manuscript, available from rst editor ([email protected]).
Krengel, U. [63]

Uber den Absolutbetrag stetiger linearer Operatoren und seine Anwendung auf ergodis-
che Zerlegungen, Math. Scand. 13 (1963) 151-187. [371Ya.]
Krieger W. [76] On ergodic ows and the isomorphism of factors, Math. Ann. 223 (1976) 19-70. [387
notes.]
Krivine J.-L. [71] Introduction to Axiomatic Set Theory. D.Reidel, 1971. [3A1.]
3A6C Group Theory 631
Kullback S. [67] A lower bound for discrimination information in terms of variation, IEEE Trans. on
Information Theory 13 (1967) 126-127. [385J.]
Kunen K. [80] Set Theory. North-Holland, 1980. [393 notes, 3A1.]
Kwapien S. [73] On the form of a linear operator on the space of all measurable functions, Bull. Acad.
Polon. Sci. 21 (1973) 951-954. [375 notes.]
Lang S. [93] Real and Functional Analysis. Springer, 1993. [3A5.]
Lindenstrauss J. & Tzafriri L. [79] Classical Banach Spaces II. Springer, 1979, reprinted in Linden-
strauss & Tzafriri 96. [354 notes, 374Xj.]
Lindenstrauss J. & Tzafriri L. [96] Classical Banach Spaces I & II. Springer, 1996.
Lipschutz S. [64] Set Theory and Related Topics. McGraw-Hill, 1964 (Schaums Outline Series). [3A1D.]
Luxemburg W.A.J. [67a] Is every integral normal?, Bull. Amer. Math. Soc. 73 (1967) 685-688. [363S.]
Luxemburg W.A.J. [67b] Rearrangement-invariant Banach function spaces, pp. 83-144 in Coleman &
Ribenboim 67. [374 notes.]
Luxemburg W.A.J. & Zaanen A.C. [71] Riesz Spaces I. North-Holland, 1971. [Chap. 35 intro.]
Macheras N.D., Musial K. & Strauss W. [p99] On products of admissible liftings and densities, J. for
Analysis and its Applications 18 (1999) 651-668. [346G.]
Macheras N.D. & Strauss W. [95] Products of lower densities, J. for Analysis and its Applications 14
(1995) 25-32. [346Xg.]
Macheras N.D. & Strauss W. [96a] On products of almost strong liftings, J. Australian Math. Soc. (A)
60 (1996) 1-23. [346Ya.]
Macheras N.D. & Strauss W. [96b] The product lifting for arbitrary products of complete probability
spaces, Atti Sem. Math. Fis. Univ. Modena 44 (1996) 485-496. [346H, 346Yb.]
Maharam D. [42] On homogeneous measure algebras, Proc. Nat. Acad. Sci. U.S.A. 28 (1942) 108-111.
[331F, 332B.]
Maharam D. [47] An algebraic characterization of measure algebras, Ann. Math. 48 (1947) 154-167.
[393J.]
Maharam D. [58] On a theorem of von Neumann, Proc. Amer. Math. Soc. 9 (1958) 987-994. [341 notes,
346 notes.]
Marczewski E. [53] On compact measures, Fund. Math. 40 (1953) 113-124. [342Ad.]
Monk J.D. [89] (ed.) Handbook of Boolean Algebra. North-Holland, 1989.
Nadkarni M.G. [90] On the existence of a nite invariant measure, Proc. Indian Acad. Sci., Math. Sci.
100 (1990) 203-220. [394 notes.]
Ornstein D.S. [74] Ergodic Theory, Randomness and Dynamical Systems. Yale U.P., 1974. [386 notes.]
Ornstein D.S. & Shields P.C. [73] An uncountable family of K-automorphisms, Advances in Math. 10
(1973) 63-88. [381 notes.]
Petersen K. [83] Ergodic Theory. Cambridge U.P., 1983. [382Ye, 384Cb, 384 notes, 385G.]
Roberts J.W. [93] Maharams problem, in Kranz & Labuda 93. [393U.]
Rotman J.J. [84] An Introduction to the Theory of Groups, Allyn & Bacon, 1984. [383 notes, 3A6B.]
Rudin W. [91] Functional Analysis. McGraw-Hill, 1991. [3A5.]
Sazonov V.V. [66] On perfect measures, A.M.S. Translations (2) 48 (1966) 229-254. [342 notes.]
Schaefer H.H. [66] Topological Vector Spaces. MacMillan, 1966; reprinted with corrections Springer, 1971.
[3A4Ad, 3A5Ib.]
Schaefer H.H. [74] Banach Lattices and Positive Operators. Springer, 1974. [Chap. 35 intro, 354 notes.]
Schubert H. [68] Topology. Allyn & Bacon, 1968. [3A3, 3A4.]
Shelah S. [Sh636] The lifting problem with the full ideal, J. Applied Analysis 4 (1998) 1-17. [341Lh.]
Sikorski R. [64] Boolean Algebras. Springer, 1964. [Chap. 31 intro.]
Sina Ya.G. [59] The notion of entropy of a dynamical system, Dokl. Akad. Nauk SSSR 125 (1959)
768-771. [384P.]
Sina Ya.G. [62] Weak isomorphism of transformations with an invariant measure, Soviet Math. 3 (1962)
1725-1729. [386E.]
632 References 3A6C
Smorodinsky M. [71] Ergodic Theory, Entropy. Springer, 1971 (Lecture Notes in Math., 214). [386
notes.]

Stepanek P. & Rubin M. [89] Homogeneous Boolean algebras, pp. 679-715 in Monk 89. [381T, 381Yd,
381 notes.]
Talagrand M. [80] A simple example of a pathological submeasure, Math. Ann. 252 (1980) 97-102.
[393T.]
Talagrand M. [82a] Closed convex hull of set of measurable functions, Riemann-measurable functions
and measurability of translations, Ann. Institut Fourier (Grenoble) 32 (1982) 39-69. [346 notes.]
Talagrand M. [82b] La pathologie des rel`evements invariants, Proc. Amer. Math. Soc. 84 (1982) 379-382.
[345F.]
Talagrand M. [84] Pettis integral and measure theory. Mem. Amer. Math. Soc. 307 (1984). [346 notes.]
Taylor A.E. [64] Introduction to Functional Analysis. Wiley, 1964. [3A5.]
Vulikh B.C. [67] Introduction to the Theory of Partially Ordered Vector Spaces. Wolters-Noordho, 1967.
[364 notes.]
Wagon S. [85] The Banach-Tarski Paradox. Cambridge U.P., 1985. [394 notes.]
Wilansky A. [64] Functional Analysis. Blaisdell, 1964. [3A5.]
Zaanen A.C. [83] Riesz Spaces II. North-Holland, 1983. [Chap. 35 intro., 376K, 376 notes.]
bounded Principal topics and results 633
Index to volumes 1, 2 and 3
Principal topics and results
The general index below is intended to be comprehensive. Inevitably the entries are voluminous to the
point that they are often unhelpful. I am therefore preparing a shorter, better-annotated, index which will,
I hope, help readers to focus on particular areas. It does not mention denitions, as the bold-type entries in
the main index are supposed to lead eciently to these; and if you draw blank here you should always, of
course, try again in the main index. Entries in the form of mathematical assertions frequently omit essential
hypotheses and should be checked against the formal statements in the body of the work.
absolutely continuous real functions 225
as indenite integrals 225E
absolutely continuous additive functionals 232
characterization 232B
on measure algebras 327B
additive functionals on Boolean algebras 326, 327, 362
dominating or subordinate to a given functional 391E, 391F
Archimedean Riesz spaces 353
associate norm (on a dual function space) 369
atomless measure spaces
have elements of all possible measures 215D
automorphisms and automorphism groups
of complete Boolean algebras 381, 383
automorphisms as products of involutions 381N
normal subgroups of automorphism groups 381S, 382I
isomorphisms of automorphism groups 383D, 383E, 383J, 383M
of measure algebras 382, 384, 385, 386, 387
Banach lattices 354
order-bounded linear operators are continuous 355C
duals 356D
Banach-Ulam problem 363S
band algebra in a Riesz space 352Q
Bernoulli shifts 386
calculation of entropy 384R
as factors of given homomorphisms (Sinas theorem) 386E, 386L
specied by entropy (Ornsteins theorem) 386I, 386K
biduals of Riesz spaces 356F, 356I
Boolean algebras Chap. 31
Boolean homomorphisms 312
between measure algebras 324
induced by measurable functions 324A
continuity and uniform continuity 324F
when measure-preserving 324K
represented by measurable functions 343B, 344B, 344E
Boolean rings 311
Borel sets in R
r
111G
and Lebesgue measure 114G, 115G, 134F
bounded variation, real functions of 224
as dierences of monotonic functions 224D
integrals of their derivatives 224I
Lebesgue decomposition 226C
634 Index Cantor
Cantor set and function 134G, 134H
Caratheodorys construction of measures from outer measures 113C
Carlesons theorem (Fourier series of square-integrable functions converge a.e.) 286
Central Limit Theorem (sum of independent random variables approximately normal) 274
Lindebergs condition 274F, 274G
change of variable in the integral 235

_
J gd =
_
g d 235A, 235E, 235L
nding J 235O;
J = [ det T[ for linear operators T 263A; J = [ det
t
[ for dierentiable operators 263D
when the measures are Hausdor measures 265B, 265E
when is inverse-measure-preserving 235I

_
gd =
_
J g d 235T
characteristic function of a probability distribution 285
sequences of distributions converge in vague topology i characteristic functions converge pointwise
285L
closed subalgebra of a measure algebra 323H, 323J
classication theorem 333H, 333K
dened by a group of automorphisms 333R
compact measure space 342, 343
and representation of homomorphisms between measure algebras 343B
complete measure space 212
completion of a measure 212C et seq.
concentration of measure 264H
conditional expectation
of a function 233
as operator on L
1
() 242J
construction of measures
image measures 112E
from outer measures (Caratheodorys method) 113C
subspace measures 131A, 214A
product measures 251C, 251F, 251W, 254C
as pull-backs 132G
continued fractions 372M et seq.
Control Measure Problem 393
convergence theorems (B.Levi, Fatou, Lebesgue) 123
convergence in measure (linear space topology on L
0
)
on L
0
() 245
on L
0
(A) 367
when Hausdor/complete/metrizable 245E, 367N
and positive linear operators 367P
convex functions 233G et seq.
convolution of functions
(on R
r
) 255

_
h (f g) =
_
h(x +y)f(x)g(y)dxdy 255G
f (g h) = (f g) h 255J
convolution of measures
(on R
r
) 257

_
hd(
1

2
) =
_
h(x +y)
1
(dx)
2
(dy) 257B
of absolutely continuous measures 257F
countable sets 111F, 1A1C et seq.
countable-cocountable measure 211R
counting measure 112Bd
fundamental Principal topics and results 635
Dedekind ()-complete Boolean algebras 314
Riesz spaces 353
Dedekind completion of a Boolean algebra 314T
dierentiable functions (from R
r
to R
s
) 262, 263
direct sum of measure spaces 214K
distribution of a random variable XXX 271
as a Radon measure 271B
of (XXX) 271J
distributive laws in Boolean algebras 313B
in Riesz spaces 352E
Doobs Martingale Convergence Theorem 275G
duals
of Riesz spaces 356, 362, 369C, 369J
of Banach lattices 356D
Dyes theorem (on the full subgroup generated by a measure-preserving automorphism) 387N, 387O
entropy (of a partition or a function) 384, 385
calculation (Kolmogorov-Sina theorem) 384P; (Bernoulli shifts) 384R
Ergodic Theorem 372
(Pointwise Ergodic Theorem) 372D, 372E
(Maximal Ergodic Theorem) 372C
exhaustion, principle of 215A
extended real line 135
extension of order-continuous positive linear operators 355F, 368B, 368M
Fatous Lemma (
_
liminf liminf
_
for sequences of non-negative functions) 123B
Fejer sums (running averages of Fourier sums) converge to local averages of f 282H
uniformly if f is continuous 282G
Fourier series 282
norm-converge in L
2
282J
converge at points of dierentiability 282L
converge to midpoints of jumps, if f of bounded variation 282O
and convolutions 282Q
converge a.e. for square-integrable function 286V
Fourier transforms
on R 283, 284
formula for
_
d
c
f in terms of

f 283F
and convolutions 283M
in terms of action on test functions 284H et seq.
of square-integrable functions 284O, 286U
inversion formulae for dierentiable functions 283I; for functions of bounded variation 283L

f(y) = lim
0
1

2
_

e
iyx
e
x
2
f(x)dx a.e. 284M
free products of Boolean algebras 315
universal mapping theorem 315I
free products of measure algebras 325
universal mapping theorems 325D, 325J
Fubinis theorem 252

_
fd( ) =
__
f(x, y)dxdy 252B
when both factors -nite 252C, 252H
for characteristic functions 252D, 252F
function spaces 369, 374, 376
Fundamental Theorem of the Calculus (
d
dx
_
x
a
f = f(x) a.e.) 222; (
_
b
a
F
t
(x)dx = F(b) F(a)) 225E
636 Index Hahn
Hahn decomposition of a countably additive functional 231E
on a Boolean algebra 326I
Hahn-Banach property see Nachbin-Kelley theorem (363R)
Hardy-Littlewood Maximal Theorem 286A
Hausdor measures (on R
r
) 264
are topological measures 264E
r-dimensional Hausdor measure on R
r
a multiple of Lebesgue measure 264I
(r 1) dimensional measure on R
r
265F-265H
image measures 112E
indenite integrals
dierentiate back to original function 222E, 222H
to construct measures 234
independent random variables 272
joint distributions are product measures 272G
limit theorems 273, 274
inner regularity of measures
(with respect to compact sets) Lebesgue measure 134F
integral operators 376
represented by functions on a product space 376E
characterized by action on weakly convergent sequences 376H
into M-spaces or out of L-spaces 376M
and weak compactness 376P
disintegrated 376S
integration of real-valued functions, construction 122
as a positive linear functional 122O
acting on L
1
() 242B
by parts 225F
characterization of integrable functions 122P, 122R
over subsets 131D, 214E
functions and integrals with values in [, ] 133
invariant measures on measure algebras 394P
Jensens inequality 233I-233J
expressed in L
1
() 242K
Kakutanis theorem (on the representation of L-spaces as L
1
-spaces) 369E
Kolmogorov-Sina theorem (on the entropy of a measure-preserving homomorphism) 384P
Komlos subsequence theorem 276H
Kwapiens theorem (on maps between L
0
spaces) 375I
Lebesgues Density Theorem (in R) 223
lim
h0
1
h
_
x+h
x
f = f(x) a.e. 223A
lim
h0
1
2h
_
x+h
xh
[f(x y)[dy = 0 a.e. 223D
(in R
r
) 261C, 261E
Lebesgue measure, construction of 114, 115
further properties 134
characterized as measure space 344I
Lebesgues Dominated Convergence Theorem (
_
lim = lim
_
for dominated sequences of functions) 123C,
367J
B.Levis theorem (
_
lim = lim
_
for monotonic sequences of functions) 123A; (
_
=
_
) 226E
lifting (of a measure)
complete strictly localizable spaces have liftings 341K
translation-invariant liftings on R
r
and 0, 1
I
345
respecting product structures 346
order Principal topics and results 637
Lipschitz functions 262
dierentiable a.e. 262Q
localizable measure space
assembling partial measurable functions 213N
which is not strictly localizable 216E
Loomis-Sikorski representation of a Dedekind -complete Boolean algebra 314M
lower density (on a measure space)
strictly localizable spaces have lower densities 341H
respecting a product structure 346G
Lusins theorem (measurable functions are almost continuous) 256F
Maharams theorem
(homogeneous probability algebras of Maharam type are isomorphic) 331I
to measure algebra of 0, 1

) 331L
classication of localizable measure algebras 332J
Maharam type
(calculation of) 332S
for measure algebras of qproduct spaces 334A, 334C
relative to a subalgebra 333E
martingales 275
L
1
-bounded martingales converge a.e. 275G, 367K
when of form E(X[
n
) 275H, 275I
measurable algebra 391
characterized by weak (, )-distributivity and existence of strictly positive additive functional 391D;
by existence of uniformly exhaustive strictly positive Maharam submeasure 392J
Control Measure Problem 393
measurable envelopes
elementary properties 132E
measurable functions
(real-valued) 121
sums, products and other operations on nitely many functions 121E
limits, inma, suprema 121F
measure algebras
properties related to those of measure spaces 322B
relationships between elementary properties 322C
satisfying the countable chain condition 322G
topologies and uniformities on 323, 324; uniqueness of topology and uniformity 324H
measure-preserving Boolean homomorphisms
extension from a closed subalgebra 333C
Monotone Class Theorem 136B
monotonic functions
are dierentiable a.e. 222A
non-measurable set (for Lebesgue measure) 134B
non-paradoxical group (of automorphisms of a Boolean algebra) 394
characterizations 394D
invariant functionals 394O, 394P, 395B
order-bounded linear operators 355
order-closed ideals and order-continuous Boolean homomorphisms 313P
order-continuous linear operators 355
order*-convergence of a sequence in a lattice 367
order-dense sets in Boolean algebras 313K
in Riesz spaces 352N
638 Index Ornstein
Ornsteins theorem (on isomorphism of Bernoulli shifts with the same entropy) 386I, 386K
outer measures constructed from measures 132
elementary properties 132A
outer regularity of Lebesgue measure 134F
partially ordered linear spaces 351
perfect measure space 342K et seq.
Plancherel Theorem (on Fourier series and transforms of square-integrable functions) 282K, 284O
Poissons theorem (a count of independent rare events has an approximately Poisson distribution) 285Q
product of two measure spaces 251
basic properties of c.l.d. product measure 251I
Lebesgue measure on R
r+s
as a product measure 251M
more than two spaces 251W
Fubinis theorem 252B
Tonellis theorem 252G
and L
1
spaces 253
continuous bilinear maps on L
1
() L
1
() 253F
conditional expectation on a factor 253H
and measure algebras 325A et seq.
product of any number of probability spaces 254
basic properties of the (completed) product 254F
characterization with inverse-measure-preserving functions 254G
products of subspaces 254L
products of products 254N
determination by countable subproducts 254O
subproducts and conditional expectations 254R
and measure algebras 325I et seq.
having homogeneous measure algebras 334E
Rademachers theorem (Lipschitz functions are dierentiable a.e.) 262Q
Radon measures
on R
r
256
as completions of Borel measures 256C
indenite-integral measures 256E
image measures 256G
product measures 256K
Radon-Nikod ym theorem (truly continuous additive set-functions have densities) 232E
in terms of L
1
() 242I
on a measure algebra 327D, 365E
rearrangements (monotonic rearrangements of measurable functions) 373, 374
regular open algebras 314P
representation
of Boolean algebras see Stones theorem (311E)
of Dedekind -complete Boolean algebras 314M
of measure algebras 321J
of homomorphisms between measure algebras 343, 344
of partially ordered linear spaces 351Q
of Riesz spaces 352L, 353M, 368E et seq., 369A
of Riesz space duals 369C
of L-spaces 369E
of M-spaces 354L
Riesz spaces (vector lattices) 352
identities 352D, 352F, 352M
Weyl Principal topics and results 639
Shannon-McMillan-Breiman theorem (entropy functions of partitions generated by a measure-preserving
homomorphism converge a.e.) 385G
simple product of Boolean algebras
universal mapping theorem 315B
of measure algebras 322K
Sinas theorem (on existence of Bernoulli partitions for a measure-preserving automorphism) 386E, 386L
split interval 343J
Stone space of a Boolean ring or algebra 311E et seq.
as topological space 311I
and Boolean homomorphisms 312P, 312Q
of a Dedekind complete algebra 314S
and the countable chain condition 316B
and weak (, )-distributivity 316I
of a measure algebra 321J, 322N, 322Q
Stone-Weierstrass theorem 281
for Riesz subspaces of C
b
(X) 281A
for subalgebras of C
b
(X) 281E
for *-subalgebras of C
b
(X; C) 281G
strictly localizable measures
sucient condition for strict localizability 213O
strong law of large numbers (lim
n
1
n+1

n
i=0
(X
i
E(X
i
)) = 0 a.e.) 273
when

n=0
1
(n+1)
2
Var(X
n
) < 273D
when sup
nN
E([X
n
[
1+
) < 273H
for identically distributed X
n
273I
for martingale dierence sequences 276C, 276F
convergence of averages for | |
1
, | |
p
273N
submeasure 392
subspace measures
for measurable subspaces 131
for arbitrary subspaces 214
surface measure in R
r
265
tensor products of L
1
spaces 253, 376C
Tonellis theorem (f is integrable if
__
[f(x, y)[dxdy < ) 252G
uniformly exhaustive submeasures
and additive functionals 392F
uniformly integrable sets in L
1
246
criteria for uniform integrability 246G
in L-spaces 354P et seq.
and convergence in measure 246J
and weak compactness 247C, 356Q
in dual spaces 356O, 362E
Vitalis theorem (for coverings by intervals in R) 221A
(for coverings by balls in R
r
) 261B
weak compactness in L
1
() 247
in L-spaces 356Q
weakly (, )-distributive Boolean algebras 316
and measure algebras 322F, 391D
Riesz spaces 368Q
Weyls Equidistribution Theorem 281N
640 Index CCC
c.l.d. version 213D et seq.
L-space 354N et seq., 371
L
0
() (space of equivalence classes of measurable functions) 241
as Riesz space 241E
L
0
(A) (based on a Boolean algebra) 364, 368, 369, 375
as a quotient of a space of functions 364D
algebraic operations 364E
calculation of suprema and inma 364M
linear operators induced by Boolean homomorphisms 364R
*when A is a regular open algebra 364U
*of a compact extremally disconnected space 364W
as the domain of a linear operator 375A, 375C
positive linear operators between L
0
spaces can be derived from Riesz homomorphisms 375I
L
1
() (space of equivalence classes of integrable functions) 242
norm-completeness 242F
density of simple functions 242M
(for Lebesgue measure) density of continuous functions and step functions 242O
L
1
(A, ) (based on a measure algebra) 365
related to L
1
() 365B
as L-space 365C, 369E
linear operators induced by Boolean homomorphisms 365H, 365O, 365P
duality with L

365J
universal mapping theorems 365L, 365N
L
p
() (space of equivalence classes of pth-power-integrable functions, where 1 < p < ) 244
is a Banach lattice 244G
has dual L
q
(), where
1
p
+
1
q
= 1 244K
and conditional expectations 244M
L
p
(A) (based on a Boolean algebra) 366
related to L
p
() 366B
as Banach lattice 366C
linear operators induced by Boolean homomorphisms 366H
L

() (space of equivalence classes of bounded measurable functions) 243

duality with L
1
243F-243G
norm-completeness 243E
order-completeness 243H
L

(A) (where A is a Boolean algebra) 363

universal mapping theorems 363E
linear operators induced by Boolean homomorphisms 363F
order-completeness 363M
Nachbib-Kelley theorem (generalized Hahn-Banach theorem) 363R
L

(U; V ) (space of order-bounded linear operators) 355, 371B, 371C

L

(U; V ) (space of dierences of order-continuous linear operators) 355, 376E

M-space 354G et seq.
S(A) (space of simple functions corresponding to a Boolean ring A) 361
universal mapping theorems 361F-361I
linear operators induced by ring homomorphisms 361J
dual spaces 362
T , T
(0)
, T

(classes of operators which are norm-reducing for every | |

p
) 371F-371G, 372, 373, 374
-algebras of sets 111
generated by given families 136B, 136G
-nite measures 215B
atom General index 641
0, 1

usual measure is (Maharam) homogeneous with Maharam type 331K; is homogeneous in strict
sense 344L
General index
This index covers the full version of the work. References in bold type refer to denitions; references
in italics are passing references. Denitions marked with >>> are those in which my usage is dangerously at
variance with that of some other author.
Abels theorem 224Yi
absolute summability 226Ac
absolutely continuous additive functional 232Aa, 232B, 232D, 232F-232I, 232Xa, 232Xb, 232Xd, 232Xf,
232Xh, 234Ce, 256J, 257Xf, 327A, 327B, 327C, 362C, 362Xh, 362Xi, 363S
absolutely continuous function 225B, 225C-225G, 225K-225O, 225Xa-225Xh, 225Xn, 225Xo, 225Ya,
225Yc, 232Xb, 244Yh, 252Yj, 256Xg, 262B, 263I, 264Yp, 265Ya, 282R, 283Ci
absolutely continuous submeasure 393D, 393E, 393H, 393Yg
Absoluteness Theorem see Shoenelds Absoluteness Theorem
abstract integral operator 376D, 376E, 376F, 376H, 376J, 376K, 376M, 376P, 376Xc, 376Xk, 376Yh-
376Yj, 376Yl, 376Yn
accumulation point see cluster point (3A3La)
additive functional on an algebra of sets see nitely additive (136Xg, 231A), countably additive (231C)
additive function(al) on a Boolean algebra see nitely additive (326A, 361B), countably additive (326E),
completely additive (326J)
adherence point see cluster point (3A3La)
adjoint operator 243Fc, 243 notes, 371Xe, 371Yd, 373S-373U, 373Xu, 373Yg, 3A5Ed
algebra (over R) 361Xb, 361Xf; see also algebra of sets (231A), Baire property algebra (314Yd), Banach
algebra (2A4Jb), Boolean algebra (311A), category algebra (314Yd), normed algebra (2A4J), f-algebra
(352W), regular open algebra (314Q)
algebra of sets 113Yi, 136E, 136F, 136G, 136Xg, 136Xh, 136Xk, 136Ya, 136Yb, 231A, 231B, 231Xa,
311Bb, 311Xb, 311Xh, 312B, 315G, 315L, 362Xg, 363Yf, 381Xa; see also Boolean algebra (311A), -algebra
(111A)
Alexandro compactication see one-point compactication (3A3O)
almost continuous function 256F
almost every, almost everywhere 112Dd
almost isomorphic (inverse-measure-preserving functions) 384U, 384V, 384Xn-384Xp, 384Yf
almost surely 112De
amenable group 394 notes
analytic (complex) function 133Xc
antichain 316 notes
aperiodic Boolean homomorphism 385A, 385C, 385E, 385F, 385Xa, 385Yb, 387Ee, 387I, 387K, 387M,
387N, 387Yd
Archimedean partially ordered linear space 351R, 351Xe, 361Gb;
Riesz space 241F, 241Yb, 242Xc, 353A-353F, 353Ha, 353L-353P, 353Xa, 353Xb, 353Xd, 353Ya,
353Yd-353Yh, 354Ba, 354F, 354I-354K, 354Yi, 355Xd, 356G-356I, 361Ee, 367C, 367F, 367Xh, 367Ya, 368B,
368E-368G, 368I, 368J, 368M, 368O, 368P, 368R, 368Ya, 368Yb
area see surface measure
arrow (double arrow space, two arrows space) see split interval (343J)
associate extended Fatou norm 369H, 369I-369L, 369O, 369R, 369Xd, 374B, 374C, 374Xh, 376S
asymptotic density 273G
asymptotically equidistributed see equidistributed (281Yi)
atom (in a Boolean algebra) 316L, 316M, 316Xr-316Xt, 316Xy, 316Yq, 322Bf, 324Ye, 331H, 333Xb,
343Xe, 362Xe; see also relative atom (331A)
642 Index atom
atom (in a measure space) 211I, 211Xb, 246G, 322Bf
atom (in a Riesz space) 362Xe
atomic see purely atomic (211K, 316Lc)
atomless Boolean algebra 316Lb, 316Mb, 316Xr, 316Xs, 316Xu-316Xw, 316Yl-316Yn, 316Yp, 316Yq,
322Bg, 322Kc, 324Kf, 324Ye, 332I, 332P, 332Ya, 375B, 375Xb, 375Yb, 383E, 383F, 383L, 383Xc, 385F,
385O, 386Cd, 386C-386H, 386L; see also relatively atomless (331A)
atomless additive functional 326Ya, 326Yb-326Yd, 326Yf, 362Xf, 362Xg, 362Yi
atomless measure algebra 331C, 369Xh, 374Xl, 382G, 382H, 382J, 382Xf, 382Xg, 382Ya, 383Ld, 383M,
383O, 383P, 383Xe
atomless measure (space) 211J, 211Md, 211Q, 211Xb, 211Yd, 211Ye, 212G, 213He, 214Xe, 215D, 215Xe,
215Xf, 216A, 216Ya, 234Xb, 234Yd, 251Xq, 252Yp, 252Yr, 252Yt, 256Xd, 264Yg, 322Bg, 325Ye, 342Xc,
343Cb, 344I
atomless vector measure 326Ye
automorphism see Banach lattice automorphism, Boolean automorphism, group automorphism (3A6B),
measure space automorphism
automorphism group of a Boolean algebra 381, 382Yb, 383, 394; of a measure algebra 366Xh, 374J,
382-383, 385Xa, 394F, 394R
axiom see Banach-Ulam problem, choice (2A1J), countable choice
Baire property (for subsets of a topological space) 314Yd
Baire property algebra 314Yd, 316Yi, 341Yb, 367Yk
Baire space 314Yd, 341Yb, 364Yi, 364Yl, 367Yi, 367Yk
Baires theorem 3A3G
Baire -algebra 341Yc, 341Zb, 343Xc, 344E, 344Ya, 344Yc-344Ye
ball (in R
r
) 252Q, 252Xh; see also sphere
Banach algebra >>>2A4Jb, 363Xa
Banach function space 369, 374
Banach lattice 242G, 242Xc, 242Yc, 242Ye, 243E, 243Xb, 326Yj, 354A, 354C, 354Ee, 354L, 354Xa,
354Xb, 354Xe-354Xj, 354Xn-354Xp, 354Yd, 354Yi-354Yl, 355C, 355K, 355Xb, 355Ya, 356D, 356M, 356Xc,
356Xk, 356Yg, 356Yh, 363E, 365N, 366C, 366Xd, 367D, 367P, 369B, 369G, 369Xe, 371B-371E, 371Xa,
371Xc, 375Xb, 376C, 376M, 376Xg, 376Yj; see also extended Fatou norm, L-space (354M), L
p
, M-space
(354Gb), Orlicz space
Banach lattice automorphism 366Xh
Banach space 231Yh, 262Yi, 2A4D, 2A4E, 2A4I, 326Yk, 354Yk, 3A5H, 3A5I; see also Banach algebra
(2A4Jb), Banach lattice (242G), separable Banach space
Banach-Ulam problem 232Hc, 326 notes, 363S, 376Yg
band in a Riesz space 352 (352O), 353B, 353C, 353E, 353H, 353I, 354Bd, 354O, 355H, 355I, 356B, 356L,
362B, 362C, 362Xf-362Xi, 364Xn, 364Xo; see also complemented band, complement of a band, principal
band, projection band
band algebra (of an Archimedean Riesz space) 353B, 353D, 356Yc, 356Yd, 361Yc, 362Ya, 362Yb, 365S,
365Xm, 366Xb, 368E, 368R; see also complemented band algebra (352Q), projection band algebra (352S)
band projection 352R, 352S, 352Xl, 355Yh, 356C, 356Xe, 356Xf, 356Yb, 362B, 362D, 362Ye, 362Yi
base for a topology 3A3M
basically disconnected topological space 314Yf
Bernoulli partition 386A, 386B, 386D-386H, 386J
shift 384Q, 384R, 384S, 384Xi-384Xm, 385Xc, 386B, 386I, 386K, 386L, 386Xa, 386Xc, 386Ya
Berry-Esseen theorem 274Hc, 285 notes
bidual of a normed space (U

) 356Xh, 3A5Ga; see also order-continuous bidual

Bienaymes equality 272R
bilinear map 253 (253A), 255Xb, 363L, 376B, 376C, 376Ya-376Yc
Bochner integral 253Yf, 253Yg, 253Yi, 354Yl
Bochners theorem 285Xr
Boolean algebra Chap. 31 (311Ab), 363Xf; see also algebra of sets, complemented band algebra (352Q),
Dedekind (-)complete Boolean algebra, measure algebra (321A), projection band algebra (352S), regular
open algebra (314Q)
cellularity General index 643
Boolean automorphism 363Ye, 366Yh, 374Yd, 381C-381G, 381J, 381L, 381Xe, 381Ya, 383A, 385B, 385Xa,
387B, 387G, 387Xa, 387Yb, 387Yc, 394Ge, 394Ya, 395A; see also involution, measure-preserving Boolean
automorphism, von Neumann automorphism (387G), weakly von Neumann automorphism (387G)
Boolean homomorphism 312F, 312G-312K, 312N, 312P-312S, 312Xe, 312Xg, 312Xj, 312Ya, 313L-313N,
313P-313R, 313Xp, 313Yb, 314I, 314K, 314Xh, 316Yo, 324, 326Be, 332Yc, 343A, 352Ta, 361Xd, 363F,
363Xb, 363Xc, 384K, 385A, 385C, 385Xa see also aperiodic Boolean homomorphism (385A), Boolean
isomorphism, measure-preserving Boolean homomorphism, (sequentially) order-continuous Boolean homo-
morphism, periodic Boolean homomorphism (385A)
Boolean-independent subalgebras 315Xn
Boolean isomorphism 312L, 314I, 332Xj, 344Yb, 366Yh
Boolean ring 311 (311Aa), 361
Boolean subalgebra see subalgebra of a Boolean algebra (312A)
Boolean value (of a proposition) 364Bb
Borel algebra see Borel -algebra
Borel-Cantelli lemma 273K
Borel lifting 345F, 345Xg, 345Yc
Borel measurable function 121C, 121D, 121Eg, 121H, 121K, 121Yf, 134Fd, 134Xd, 134Yt, 135Ef, 135Xc,
135Xe, 225H, 225J, 225Yf, 233Hc, 241Be, 241I, 241Xd, 256M, 364I, 364J, 364Xg, 364Yd, 367Yp
Borel measure (on R) 211P, 216A, 341Lg, 343Xd, 345F
Borel sets in R, R
r
111G, 111Yd, 114G, 114Yd, 115G, 115Yb, 115Yd, 121Ef, 121K, 134F, 134Xc, 135C,
136D, 136Xj, 212Xc, 212Xd, 225J, 264F, 342Xi, 364H, 364Yc
Borel -algebra (of subsets of R
r
) 111Gd, 114Yg, 114Yh, 114Yi, 121J, 121Xd, 121Xe, 121Yb, 216A,
251L, 264E, 264Xb, 364G;
(of other spaces) 135C, 135Xb, 256Ye, 264Yc, 271Ya, 314Ye, 333Ya
bounded bilinear map 253Ab
bounded linear operator 253Ab, 253F, 253Gc, 253L, 253Xc, 253Yf, 253Yj, 253Yk, 2A4F, 2A4G-2A4I,
2A5If, 355C, 3A5Ed; see also (weakly) compact linear operator (3A5K), order-bounded linear operator
(355A), B(U; V ) (2A4F)
bounded set (in R
r
) 134E; (in a normed space) 2A4Bc, 3A5H; (in a linear topological space) 245Yf; see
also order-bounded (2A1Ab)
bounded support, function with see compact support
bounded variation, function of 224 (224A, 224K), 225Cb, 225M, 225Oc, 225Xh, 225Xn, 225Yc, 226Bc,
226C, 264Yp, 282M, 282O, 283L, 283Xj, 283Xk, 283Xm, 283Xn, 283Xq, 284Xk, 284Yd, 343Yc, 354Xt
bounding see

-bounding (316Yg)
Breiman see Shannon-McMillan-Breiman theorem (385G)
Cantor function 134H, 134I, 222Xd, 225N, 226Cc, 262K, 264Yn
Cantor measure 256Hc, 256Ia, 256Xk, 264Ym
Cantor set 134G, 134H, 134I, 134Xe, 256Hc, 256Xk, 264J, 264Ym, 264Yn; see also 0, 1
N
Caratheodory complete measure space see complete (211A)
Caratheodorys method (of constructing measures) 113C, 113D, 113Xa, 113Xd, 113Xg, 113Yc, 114E,
114Xa, 121Yc, 132Xc, 136Ya, 212A, 212Xg, 213C, 213Xa, 213Xf, 213Xg, 213Ya, 214H, 214Xc, 216Xb,
251C, 251Wa, 251Xd, 264C, 264K
cardinal 2A1Kb, 2A1L, 3A1B-3A1E
cardinal function (of a Boolean algebra) see cellularity (332D, Maharam type (331F)
(of a topological space) see cellularity (332Xd, density (331Yf
Carlesons theorem 282K remarks, 282 notes, 284Yg, 286U, 286V
category algebra (of a topological space) 314Yd
Cauchy distribution 285Xm, 367Xw
Cauchy lter 2A5F, 2A5G, 3A4F
Cauchys inequality 244E
Cauchy sequence 242Yc, 2A4D, 354Ed, 356Yf, 356Yg, 3A4Fe
cellularity(of a topological space) 332Xd, 332Xe; (of a Boolean algebra) 332 (332D), 365Ya; see also
magnitude (332Ga)
644 Index central
Central Limit Theorem 274G, 274I-274K, 274Xc-274Xg, 285N, 285Xn, 285Ym
Cesaro sums 273Ca, 282Ad, 282N, 282Xn
Chacon-Krengel theorem 371 notes
chain condition (in Boolean algebras or topological spaces) see ccc (316A), -linked (391L), -m-linked
(391Yh)
chain rule for Radon-Nikod ym derivatives 235Xi
change of variable in integration 235, 263A, 263D, 263F, 263G, 263I, 263Xc, 263Xe, 263Yc
characteristic function (of a set) 122Aa
(of a probability distribution) 285 (285Aa)
(of a random variable) 285 (285Ab);
charge 326A; see nitely additive functional
chargeable Boolean algebra 391D, 391J, 391N, 391X, 391Xa-391Xf, 391Yb, 391Yc, 391Yf, 392F, 394Xd
Chebyshevs inequality 271Xa
choice, axiom of 134C, 1A1G, 254Da, 2A1J, 2A1K-2A1M, 361 notes; see also countable choice
choice function 2A1J
Choquet capacity see capacity
circle group 255M, 255Yh, 345Xb, 345Xg
clopen algebra 311I, 311J, 311Xh
closed convex hull 253G, 2A5E
closed interval (in R or R
r
) 114G, 115G, 1A1A; (in general partially ordered spaces) 2A1Ab
closed set (in a topological space) 134Fb, 134Xc, 1A2E, 1A2F, 1A2G, 2A2A, 2A3A, 2A3D, 2A3Nb; (in
an ordered space) see order-closed (313D)
closed subalgebra of a measure algebra 323H-323K (323I), 323Xb, 323Xd, 323Yd, 325Xa, 327F, 331B,
331D, 332P, 332T, 333, 334Xb, 364Xe, 366J, 372G, 372J, 384D, 384G, 384Xq, 385E, 385N-385P, 385Xf,
385Yc, 385Yd, 386Be, 387L; see also order-closed subalgebra
closure of a set 2A2A, 2A2B, 2A3D, 2A3Kb
cluster point (of a lter) 3A3L; (of a sequence) 2A3O
Coarea Theorem 265 notes
conal set (in a partially ordered set) 3A1F
conality (of a partially ordered set) 3A1F
conite see nite-conite algebra (231Xa, 316Yk)
commutative (ring, algebra) 2A4Jb, 363Xf, 3A2A, 3A2Ff, 3A2Hc
commutative Banach algebra 224Yb, 243D, 255Xc, 257Ya, 363B
commutative group see abelian group
commutator in a group 381Xf, 381Xg, 381Ya
compact class of sets 342Ab, 342D
compact Hausdor space 364V, 364W, 364Yj, 364Yl, 364Ym, 3A3D, 3A3Ha, 3A3K
compact linear operator 371Ya, 371Yb, 376Yn, 3A5Ka; see also weakly compact operator (3A5Kb)
compact measure 342A, 342E-342H, 342J, 342M, 342N, 342Xd-342Xh, 342Xl, 342Xm, 342Xo, 342Ya,
342Yb, 343J, 343K, 343Xa, 343Xb, 343Ye; see also locally compact measure (342Ad)
compact set (in a topological space) 247Xc, 2A2D, 2A2E-2A2G, >>>2A3N, 2A3R, 3A3D; see also convex
compact, relatively compact (2A3N), relatively weakly compact (2A5Id), weakly compact (2A5Ic)
compact support, function with 242O, 242Pd, 242O, 242Xh, 244H, 244Yi, 256Be, 256D, 256Xh, 262Yd-
262Yg
compact support, measure with 284Yi
compact topological space >>>2A3N, 3A3Dd, 3A3J see also compact Hausdor space, locally compact
(3A3Ag)
compactication see one-point compactication (3A3O)
complement of a band (in a Riesz space) 352P, 352Xg
complementary Youngs functions 369Xc, 369Xd, 369Xr
complemented band (in a Riesz space) 352P, 352Xe, 352Xg, 353B, 353Xa; see also complemented band
algebra (352Q)
complemented band algebra 352Q, 352S, 352Xf, 353B, 356Yc
continuity General index 645
complete generation of a Boolean algebra see Maharam type (331F)
complete lattice 314Bf
complete linear topological space 245Ec, 2A5F, 2A5H, 367Nc
complete locally determined measure (space) 213C, 213D, 213H, 213J, 213L, 213O, 213Xg, 213Xi, 213Xl,
213Yd, 213Ye, 214I, 214Xc, 216D, 216E, 234F, 251Yb, 252B, 252D, 252E, 252N, 252Xb, 252Yh, 252Yj,
252Yr-252Yt, 253Yj, 253Yk, 325B, 341M, 341Pb, 342N, 364Xm, 376S, 376Xj; see also c.l.d. version (213E)
complete measure (space) 112Df, 113Xa, 122Ya, 211A, 211M, 211N, 211R, 211Xc, 211Xd, 212, 214I,
214J, 216A, 216C, 216Ya, 234A, 234D, 254F, 254G, 254J, 264Dc, 321K, 341J, 341K, 341Mb, 341Xc, 341Yd,
343B, 344C, 344I, 344Xb-344Xd, 384V; see also complete locally determined measure
complete metric space 224Ye, 323Yd, 393Xb, 3A4Fe, 3A4Hc; see also Banach space (2A4D)
complete normed space see Banach space (2A4D), Banach lattice (242G, 354Ab)
complete Riesz space see Dedekind complete (241Fc)
complete uniform space 323G, 3A4F, 3A4G, 3A4H; see also completion (3A4H)
complete (in -complete lter) 368F
complete see also Dedekind complete, Dedekind -complete, uniformly complete (354Yi)
completed indenite-integral measure see indenite-integral measure (234B)
completely additive functional on a Boolean algebra 326J, 326K-326P, 326Xe, 326Xf, 326Xh, 326Xi,
326Yp-326Ys, 327B-327E, 327Ya, 332Xo, 362Ad, 362B, 362D, 362Xe-362Xg, 363K, 363S, 365E, 394Xe; see
also M

completely additive part (of an additive functional on a Boolean algebra) 326Yq, 362Bd; (of a linear
functional on a Riesz space) 355Yh
completely continuous linear operator see compact linear operator (3A5Ka)
completely generate see -generate (331E)
completely regular topological space 353Yb, 367Yi, >>>3A3Ac, 3A3B
completion (of a Boolean algebra) see Dedekind completion (314U)
(of a linear topological space) ,
(of a measure (space)) 212 (212C), 213Fa, 213Xk, 214Xb, 214Xj, 232Xe, 234D, 235D, 235Xe,
241Xb, 242Xb, 243Xa, 244Xa, 245Xb, 251S, 251Wn, 251Xp, 252Ya, 254I, 256C, 322Da, 342Gb, 342Ib,
342Xn, 343Ac
(of a metric space) 393B
(of a normed space) 3A5Ib
(of a normed Riesz space) 354Xh, 354Yg
(of a uniform space) 325Ea, 325Yb, 3A4H
complexication of a Riesz space 354Yk, 355Yk, 356Yh, 362Yj, 363Xj, 364Yn
complex-valued function 133
component (in a Riesz space) 352Rb, 352Sb, 356Yb, 368Yj; (in a topological space) 111Ye; see also
Maharam-type- component (332Gb)
concave function 384A
conditional entropy 384D, 384E, 384G, 384N, 384Xb-384Xd, 384Xq, 384Ya, 384Yb, 385N
conditional expectation 233D, 233E, 233J, 233K, 233Xg, 233Yc, 235Yc, 242J, 246Ea, 253H, 253Le,
275Ba, 275H, 275I, 275K, 275Ne, 275Xi, 275Ya, 275Yk, 275Yl, 365R, 369Xq, 372K
conditional expectation operator 242Jf, 242K, 242L, 242Xe, 242Yk, 243J, 244M, 244Yk, 246D, 254R,
254Xp, 275Xd, 275Xe, 327Xb, 365R, 365Xl, 366J, 372G, 384B, 384D, 384E
conditionally (order)-complete see Dedekind complete (314A)
cone 3A5B, 3A5D; see also positive cone (351C)
conegligible set 112Dc, 214Cc
connected set 222Yb
conjugate operator see adjoint operator (3A5Ed)
consistent disintegration 452A, 452C, 452E, 452I, 452K, 452Xg, 452Ya, 452Yd, 454Xf, 457R, 457Xl
consistent lifting 346I, 346J, 346L, 346Xe, 346Yd
continued fractions 372M-372O; (continued-fraction approximations) 372M, 372Xj, 372Yf; (continued-
fraction coecients) 372M, 372O, 372Xh-372Xj, 372Xt, 372Ye
continuity, points of 224H, 224Ye, 225J
646 Index continuous
continuous function 121D, 121Yf, 262I, 2A2C, 2A2G, 2A3B, 2A3H, 2A3Nb, 2A3Qb, 3A3C, 3A3Eb,
3A3Ib, 3A3Nb
continuous at a point 3A3C
continuous linear functional 284Yj; see also dual linear space (2A4H)
continuous linear operator 2A4Fc; see also bounded linear operator (2A4F)
continuous outer measure see Maharam submeasure (392G)
continuous see also order-continuous (313H), semi-continuous (225H), uniformly continuous (3A4C)
continuum see c (2A1L)
control measure 393O, 393P, 393S, 393Xg, 393Yh, 393Yi
Control Measure Problem 393
convergence in mean (in L
1
() or L
1
()) 245Ib
convergence in measure (in L
0
()) 232Ya, 245 (>>>245A), 246J, 246Yc, 247Ya
(in L
0
()) 245 (>>>245A), 274Yd
- (in L
0
(A)) 367 (367M), 369M, 369Xf, 369Ye, 393M
(of sequences) 245, 246J, 246Xh, 246Xi, 253Xe, 255Yf, 273Ba, 276Yf, 367Na, 367Q, 369Yf, 376Xf
convergent almost everywhere 245C, 245K, 273Ba, 276G, 276H; see also strong law
convergent lter 2A3Q, 2A3S, 2A5Ib, 3A3Ic, 3A3Lb;
sequence 135D, 245Yi, 2A3M, 2A3Sg, 3A3Lc; see also convergent in measure (245Ad), order*-
convergent (367A)
convex function 233G, 233H-233J, 233Xb, 233Xc, 233Xd, 233Xe, 233Xh, 233Ya, 233Yc, 242K, 242Yi,
242Yj, 242Yk, 244Xm, 244Yg, 255Yk, 275Yg, 365Rb, 367Xx, 369Xb, 369Xc, 369Xd, 373Xb; see also mid-
convex (233Ya)
convex hull 2A5E; see also closed convex hull (2A5E)
convex set 233Xd, 244Yj, 262Xh, 2A5E, 326Ye, 351Ce, 3A5C, 3A5Ee, 3A5Ld
see also cone (3A5B), order-convex (351Xb)
convex structure 373Xv, 373Yf
convolution in L
0
255Fc, 255Xc, 255Yf, 255Yk
convolution of functions 255E, 255F-255K, 255O, 255Xa-255Xc, 255Xf-255Xj, 255Ya, 255Yb, 255Yd,
255Ye, 255Yi, 255Yl, 255Ym, 255Yn, 262Xj, 262Yd, 262Ye, 263Ya, 282Q, 282Xt, 283M, 283Wd, 283Wf,
283Wj, 283Xl, 284J, 284K, 284O, 284Wf, 284Wi, 284Xb, 284Xd, 373Xg, 376Xa, 376Xd
convolution of measures 257 (257A), 272S, 285R, 285Yn, 364Xf
convolution of measures and functions 257Xe, 284Xo, 284Yi
convolution of sequences 255Xe, 255Yo, 282Xq, 352Xk, 376Ym
countable (set) 111F, 114G, 115G, 1A1, 226Yc
countable antichain condition 316 notes (see ccc (316A))
countable chain condition see ccc (316A)
countable choice (axiom of) 134C
countable-cocountable algebra 211R, 211Ya, 232Hb, 316Yk, 368Xa
countable-cocountable measure 211R, 232Hb, 252L, 326Xf, 342M, 356Xc
countable sup property (in a Riesz space) 241Yd, 242Yd, 242Ye, 244Yb, 353Ye, 354Yc, 355Yi, 356Ya,
363Yb, 364Yb, 366Yb, 368Yh, 376Ye, 376Yf
countably additive functional (on a -algebra of sets) 231 (231C), 232, 246Yg, 246Yi, 327C, 362Xg,
363S; (on a Boolean algebra) >>>326E, 326F-326I, 326L, 326Xb-326Xi, 327B, 327C, 327F, 327G, 327Xa,
327Xd, 327Xe, 362Ac, 362B, 363S, 363Yh, 392Xf; see also completely additive (326J), countably subadditive
(392H), M

countably additive part (of an additive functional on a Boolean algebra) 326Yn, 362Bc, 362Ye, 393Yg
countably full group (of automorphisms of a Boolean algebra) 381Xd
countably generated -algebra 381Yb, 381Yc, 382Yf
countably separated (measure space, -algebra of sets) 343D, 343E-343H, 343K, 343L, 343Xe, 343Xh,
343Yb, 343Ye, 343Yf, 344B, 344C, 344I, 344Xb-344Xd, 384V, 387C, 387D
countably subadditive functional 392H
counting measure 112Bd, 122Xd, 122 notes, 211N, 211Xa, 212Ya, 226A, 241Xa, 242Xa, 243Xl, 244Xi,
244Xj, 244Xn, 245Xa, 246Xc, 251Xb, 251Xh, 252K, 255Yo, 264Db, 324Xe
determined General index 647
cover see measurable envelope (132D)
covering theorem 221A, 261B, 261F, 261Xc, 261Ya, 261Yi, 261Yk
cozero set 3A3Pa
cycle notation (for automorphisms of Boolean algebras) 381G, 381H, 381I, 385B; see also pseudo-cycle
(387Xc)
cyclic automorphism 381G; see also exchanging involution (381G)
cylinder (in measurable cylinder) 254Aa, 254F, 254G, 254Q, 254Xa
Davies R.O. 325 notes
decimal expansions 273Xf
decomposable measure (space) see strictly localizable (211E)
decomposition (of a measure space) 211E, 211Ye, 213O, 213Xh, 214Ia, 214K, 214M, 214Xi, 322L; see
also Hahn decomposition of an additive functional (231F), Jordan decomposition of an additive functional
(231F), Lebesgue decomposition of a countably additive functional (232I), Lebesgue decomposition of a
function of bounded variation (226)
decreasing rearrangement (of a function) 252Yp; (of an element of M
0,
(A)) 373 (373C), 374B-374D,
374J, 374L, 374Xa, 374Xj, 374Ya-374Yd
Dedekind complete Boolean algebra 314B, 314Ea, 314Fa, 314G, 314Ha, 314I, 314Ja, 314K, 314P, 314S,
314T, 314Va, 314Xa, 314Xb, 314Xf, 314Xg, 314Xh, 314Yd, 314Yg, 314Yh, 315F, 315Yf, 316Fa, 316K, 316Xt,
316Yi, 316Yn, 331Yd, 332Xa, 333Bc, 352Q, 363Mb, 363P-363S, 364O, 385B, 385C, 368A-368D, 368Xe,
375C, 375D, 375Xa, 375Xc, 375Ya, 381, 383D, 383J, 393A-393C, 393J, 393Xc-393Xf, 393Ya, 393Yc-393Yf
Dedekind complete partially ordered set 135Ba, 314Aa, 314B, 314Ya, 315De, 343Yc, 3A6Csee also
Dedekind complete Boolean algebra, Dedekind complete Riesz space
Dedekind complete Riesz space 241Fc, 241G, 241Xf, 242H, 242Yc, 243H, 243Xj, 244L, 353G, 353I,
353Jb, 353K, 353Yb, 354Ee, 355, 356B, 356K, 356Xj, 356Ye, 361H, 363Mb, 363P, 363S, 364O, 366C,
366G, 368H-368J, 371B-371D, 371Xd, 371Xe, 375J, 375Ya
Dedekind completion of a Boolean algebra 314T, 314U, 314Xe, 314Yh, 315Yd, 322O, 361Yc, 365Xm,
366Xk, 368Xb, 383Lf, 383Xb, 383Xd, 391Xc; see also localization (322P)
Dedekind completion of a Riesz space 368I, 368J, 368Xb, 368Yb, 369Xp
Dedekind -complete Boolean algebra 314C-314F, 314Hb, 314Jb, 314M, 314N, 314Xa, 314Xc, 314Ye,
314Yf, 315O, 316C, 316Fa, 316Hb, 316J, 316Xg, 316Yi, 316Yk, 321A, 324Yb, 324Yc, 326Fg, 326I, 326O,
326P, 326Xc, 326Xi, 326Yf, 326Yo, 344Yb, 362Xa, 363Ma, 363P, 363Xh, 363Yh, 364, 367Xk, 368Yd, 375Yb,
381B, 381C, 381T, 381Ya, 392H, 392Xd, 392Xf, 393A, 393E, 393M, 393O-393Q, 393S, 393Xb, 393Yg-393Yi
Dedekind -complete partially ordered set 314Ab, 315De, 367Bf; see also Dedekind -complete Boolean
algebra, Dedekind -complete Riesz space
Dedekind -complete Riesz space 241Fb, 241G, 241Xe, 241Yb, 241Yh, 242Yg, 243H, 243Xb, 353G,
353H, 353J, 353Xb, 353Yc, 353Yd, 354Xn, 354Yi, 354Yl, 356Xc, 356Xd, 363M-363P, 364C, 364E
delta function see Diracs delta function (284R)
delta system see -system
De Morgans laws 311Xf; see also distributive laws
dense set in a topological space 254Yc, 2A3U, 313Xj, 323Dc, 3A3E, 3A3G, 3A3Ie, 3A4Ff; see also
nowhere dense (3A3Fa), order-dense (313J, 352Na), quasi-order-dense (352Na)
density (of a topological space) 331Ye, 331Yf, 365Ya, 393Yf
density function (of a random variable) 271H, 271I-271K, 271Xc-271Xe, 272T, 272Xd, 272Xj; see also
Radon-Nikod ym derivative (232If, 234B)
density point 223B, 223Xi, 223Yb
density topology 223Yb, 223Yc, 223Yd, 261Yf
density see also asymptotic density (273G), Lebesgues Density Theorem (223A)
derivative of a function (of one variable) 222C, 222E, 222F, 222G, 222H, 222I, 222Yd, 225J, 225L, 225Of,
225Xc, 226Be, 282R; (of many variables) 262F, 262G, 262P; see also gradient, partial derivative
determinant of a matrix 2A6A
determined see locally determined measure space (211H), locally determined negligible sets (213I)
determined by coordinates (in W is determined by coordinates in J) 254M, 254O, 254R, 254S, 254T,
254Xp, 254Xr, 311Xh, 325N, 325Xh
648 Index Devil
Devils Staircase see Cantor function (134H)
dierentiability, points of 222H, 225J
dierentiable function 123D, 222A, 224I, 224Kg, 224Yc, 225L, 225Of, 225Xc, 225Xn, 233Xc, 252Yj,
255Xg, 255Xh, 262 (262F), 263D, 263Xc, 263Yc, 265E, 265Xd, 265Ya, 274E, 282L, 282Xs, 282Xk, 283I-
283K, 283Xm, 284Xk; see also derivative
dierentiable relative to its domain 262F
diused measure see atomless measure (211J)
dilation 286C
dimension see Hausdor dimension (264 notes))
Diracs delta function 257Xa, 284R, 284Xn, 284Xo, 285H, 285Xp
direct image (of a set under a function or relation) 1A1B
direct sum of measure spaces 214K, 214L, 214Xi-214Xl, 241Xg, 242Xd, 243Xe, 244Xg, 245Yh, 251Xh,
251Xi, 322Kb, 332C, 342Gd, 342Ic, 342Xn, 343Yb
directed set see downwards-directed (2A1Ab), upwards-directed (2A1Ab)
Dirichlet kernel 282D; see also modied Dirichlet kernel (282Xc)
disconnected see extremally disconnected (3A3Ae), basically disconnected
discrete topology (3A3Ah)
disjoint family (of sets) 112Bb; (in a Boolean ring) 311Gb; (in a Riesz space) 352C
disjoint sequence theorem 246G, 246Ha, 246Yd, 246Ye, 246Yf, 246Yj, 354Rb, 356O, 356Xm
disjoint set (in a Boolean ring) 311Gb; (in a Riesz space) 352C
disjoint union topology 315Xe
distribution see Schwartzian distribution, tempered distribution
distribution of a random variable 241Xc, 271B, 271C, 271D-271G, 272G, 272S, 272Yf, 285Ab, 285Mb,
364Xd, 364Xf, 367Xv, 373Xh; see also empirical distribution (273 notes)
distribution function of a random variable >>>271G, 271Xb, 271Yb, 271Yc, 272Xe, 272Yc, 273Xg, 273Xh,
285P
distributive lattice 367Yb, 381Xi, 3A1Ic; see also (2, )-distributive (367Yd)
distributive laws in Boolean algebras 313B; see also De Morgans laws, weakly -distributive (316Yg),
weakly (, )-distributive (316G)
distributive laws in Riesz spaces 352E; see also weakly -distributive, weakly (, )-distributive (368N)
Dominated Convergence Theorem see Lebesgues Dominated Convergence Theorem (123C)
Doobs Martingale Convergence Theorem 275G
double arrow space see split interval (343J)
doubly stochastic matrix 373Xe
downwards-directed partially ordered set 2A1Ab
dual linear operator see adjoint operator (3A5Ed)
dual linear space (of a normed space) 243G, 244K, 2A4H, 356D, 356N, 356O, 356P, 356Xg, 356Yh, 365I,
365J, 366C, 366D, 369K, 3A5A, 3A5C, 3A5E-3A5H
dual
see also algebraic dual, bidual, order-bounded dual (356A), order-continuous dual (356A), sequen-
tially order-continuous dual (356A)
Dunfords theorem 376N
dyadic cycle system (for a Boolean automorphism) 387G, 387K, 387M
Dyes theorem 387O, 387 notes
Dynkin class 136A, 136B, 136Xb
Eberleins theorem 2A5J, 356 notes
Egorovs theorem 131Ya, 215Yb
Egorov property (of a Boolean algebra) 316Yg, 316Yh
empirical distribution 273Xh, 273 notes
entropy 384; (of a partition of unity) 384C, 385G, 385K, 385N, 385Q, 386B-386D, 386F-386H; (of a
measure-preserving Boolean homomorphism) 384M, 385N, 385Xe, 385Yc, 385Yd, 386C-386E, 386I-386L,
387Yd; see also conditional entropy (384D), entropy metric (384Xq)
entropy metric 384Xq, 384Yg
nitely General index 649
enumerate, enumeration 3A1B
envelope see measurable envelope (132D), upr(a, C) (314V)
equidecomposable (in G-equidecomposable) 395Ya; (in G--equidecomposable) 394 (394A)
equidistributed sequence (in a topological probability space) 281N, 281Yi, 281Yj, 281Yk
Equidistribution Theorem see Weyls Equidistribution Theorem (281N)
equivalent norms 355Xb, 369Xd
equiveridical 121B, 212B, 312B
ergodic automorphism (of a Boolean algebra) 387Xg, 394Ge
ergodic group of automorphisms (of a Boolean algebra) 394Ge, 394Q, 394R, 394Xa, 394Xg, 394Yd
ergodic inverse-measure-preserving function 372P, 372Q, 372Xo, 372Xs, 372Yi, 372Yj, 387Yf
ergodic measure-preserving Boolean homomorphism (of a measure algebra) 372P, 372Q, 372Xk, 372Xl,
384Se, 385F, 385H, 386C-386E, 386J, 386Xb, 387Xg, 387Ye, 395Xc
Ergodic Theorem 372D, 372E, 372Ya, 385Xc; see also Maximal Ergodic Theorem (372C), Mean Ergodic
Theorem (372Xa), Wieners Dominated Ergodic Theorem (372Yb)
essential supremum of a family of measurable sets 211G, 213K, 215B, 215C; of a real-valued function
243D, 243I, 376S, 376Xn
essentially bounded function 243A
Etemadis lemma 272U
Euclidean metric (on R
r
) 2A3Fb
Euclidean topology 1A2, 2A2, 2A3Ff, 2A3Tc
even function 255Xb, 283Yb, 283Yc
exchangeable sequence of random variables 276Xe
exchanging involution 381G, 381H, 381K, 381Yb
exhaustion, principle of 215A, 215C, 215Xa, 215Xb, 232E, 246Hc, 342B, 365 notes
exhaustive submeasure 392Bb, 392C, 392Hc, 392Xb-392Xd, 393A-393C, 393H; see also uniformly ex-
haustive (392Bc)
expectation of a random variable 271Ab, 271E, 271F, 271I, 271Xa, 272Q, 272Xb, 272Xi, 285Ga, 285Xo;
see also conditional expectation (233D)
extended Fatou norm 369 (369F), 376S, 376Xn, 376Yl; see also T -invariant extended Fatou norm
(374Ab), rearrangement-invariant extended Fatou norm (374Eb)
extended real line 121C, 135
extension of nitely additive functionals 391G
extension of measures 132Yd, 212Xk, 327Xf; see also completion (212C), c.l.d. version (213E)
extremally disconnected topological space 314S, 353Yb, 363Yc, 364W, 364Yk-364Ym, 368G, 368Yc,
3A3Ae, 3A3Bd
extreme point of a convex set 353Yf
factor (of an automorphism of a Boolean algebra) 386Ac, 386L, 386Xb
fair-coin probability 254J
Fatous Lemma 123B, 133K, 135G, 135Hb, 365Xc, 365Xd, 367Xf
Fatou norm on a Riesz space 244Yf, 354Da, 354Eb, 354J, 354Xk, 354Xn, 354Xo, 354Ya, 354Ye, 356Da,
367Xg, 369G, 369Xe, 371B-371D, 371Xc, 371Xd; see also extended Fatou norm (369F)
Fejer integral 283Xf, 283Xh-283Xj
Fejer kernel 282D
Fejer sums 282Ad, 282B-282D, 282G-282I, 282Yc
Feller, W. Chap. 27 intro.
eld (of sets) see algebra (136E)
lter 254Ta, 2A1I, 2A1N, 2A1O, 2A5F; see also convergent lter (2A3Q), ultralter (2A1N)
nite-conite algebra 231Xa, 231Xc, 316Yk, 392Xc
nite rank operators see abstract integral operators (376D)
nitely additive functional on an algebra of sets 136Xg, 136Ya, 136Yb, 231A, 231B, 231Xb-231Xe,
231Ya-231Yh, 327C; see also countably additive, completely additive
650 Index nitely
nitely additive function(al) on a Boolean algebra 326A, 326B-326D, 326Fa, 326G, 326Kg, 326Xa-326Xc,
326Xe, 326Xg, 326Xi, 326Yt-326Yd, 326Yg-326Yj, 326Yk, 326Yl-326Yn, 326Yq, 326Yr, 327B, 327C, 331B,
361Xa-361Xc, 391D-391G, 391I, 391J, 391N, 393Be, 393C, 394N, 394O; see also chargeable Boolean algebra
(391X), completely additive functional (326J), countably additive functional (231C, 326E)
nitely additive function(al) on a Boolean ring 361B, 361C, 361E-361I, 365Eb
nitely additive measure 326A
rst category see meager (3A3Fb)
xed-point subalgebra (in a Boolean algebra) 394G, 394H, 394I, 394K-394N, 394P, 394Xb-394Xf, 394Yb,
394Yc
Fourier coecients 282Aa, 282B, 282Cb, 282F, 282I, 282J, 282M, 282Q, 282R, 282Xa, 282Xg, 282Xq,
282Xt, 282Ya, 283Xu, 284Ya, 284Yg
Fouriers integral formula 283Xm
Fourier series 121G, 282 (282Ac)
Fourier sums 282Ab, 282B-282D, 282J, 282L, 282O, 282P, 282Xi-282Xk, 282Xp, 282Xt, 282Yd, 286V,
286Xb
Fourier transform 133Xd, 133Yc, 283 (283A, 283Wa), 284 (284H, 284Wd), 285Ba, 285D, 285Xd,
285Ya, 286U, 286Ya
Fourier-Stieltjes transform see characteristic function (285A)
Frechet lter 2A3Sg
free product of Boolean algebras 315 (315H), 316Yl, 325, 326Q, 326Yt, 326Yu, 361Yf, 383Le, 391Xb,
391Ye; see also localizable measure algebra free product (325Ea), probability algebra free product (325K)
Frolks theorem 381J, 381Yb
Fubinis theorem 252B, 252C, 252H, 252R, 393Td
full local semigroup (of partial automorphisms of a Boolean algebra) 394A, 394B, 394C, 394Gc, 394Xc,
394Ya, 394Yd
full outer measure 132F, 132G, 132Yd, 133Yf, 134D, 134Yt, 214F, 322Jb, 324Cb, 343Xa
full subgroup of Aut A 381M, 381N, 381O, 381R, 381S, 381Xb, 381Xc, 381Xi, 381Xj, 381Xk, 382C, 387;
see also countably full (381Xd)
subsemigroup of Aut A 387Ya
fully non-paradoxical group of automorphisms of a Boolean algebra 394E, 394F, 394H, 394I, 394K-394R,
394Xd-394Xf, 394Xh, 394Ya-394Yc, 395B, 395Ya
fully symmetric extended Fatou norm see T -invariant (374Ab)
function 1A1B
function space see Banach function space (369)
functionally closed see zero set (3A3Pa)
functionally open see cozero set (3A3Pa)
Fundamental Theorem of Calculus 222E, 222H, 222I, 225E
Fundamental Theorem of Statistics 273Xh, 273 notes
Gaifmans example 391N, 391Yf-391Yh,
game see innite game, Banach-Mazur game
Gamma function see -function (225Xj)
Gauss C.F. 372Yg
Gaussian random variable, distribution see normal random variable (274Ad)
generated topology 3A3Ma
generated (-)algebra of sets 111Gb, 111Xe, 111Xf, 121J, 121Xd, 136B, 136C, 136G, 136Xc, 136X,
136Xl, 136Yb
generated ()-subalgebra of a Boolean algebra 313F, 313M, 314G, 314Ye, 331E
generating Bernoulli partition 386Ab, 386Bg
generating set in a Boolean algebra 331E; see also -generating (331E), -generating (331E)
Glivenko-Cantelli theorem 273 notes
group 255Yn, 255Yo; see also amenable group, automorphism group, circle group, ergodic group, simple
group
group automorphism 3A6B; see also inner automorphism (3A6B), outer automorphism (3A6B)
independent General index 651
Hahn-Banach theorem 363R, 373Yd, 3A5A, 3A5C
Hahn decomposition of an additive functional 231F, 326I, 326O
see also Vitali-Hahn-Saks theorem (246Yg)
Hajian-Ito theorem 395B
half-open interval (in R or R
r
) 114Aa, 114G, 114Xe, 114Yj, 115Ab, 115Xa, 115Xc, 115Yd
Halls Marriage Lemma 3A1K
Halmos-Rokhlin-Kakutani lemma 385E
Hardy-Littlewood Maximal Theorem 286A
Hausdor dimension 264 notes
Hausdor measure 264 (264C, 264Db, 264K, 264Yo), 343Ye, 345Xb, 345Xg; see also normalized
Hausdor measure (265A)
Hausdor metric (on a space of closed subsets) 246Yb
Hausdor outer measure 264 (264A, 264K, 264Yo)
Hausdor topology 2A3E, 2A3L, 2A3Mb, 2A3S, 3A3Aa, 3A3D, 3A3Id, 3A3K, 3A4Ac, 3A4Fd
Hausdor uniformity 3A4Ac
Hilbert space 244N, 244Yj, 3A5L; see also inner product space (3A5L)
Hilbert transform 376Ym
Holders inequality 244Eb
homeomorphism 3A3Ce
homogeneous Boolean algebra 331M, 331N, 331Xj, 331Yg, 331Yh, 331Yi, 331Yj, 381Q, 381T, 381Yd,
383E, 383F, 383G, 383La, 383Xc; see also Maharam-type-homogeneous (331Fc), relatively Maharam-type-
homogeneous (333Ac)
measure algebra 331N, 332M, 373Yb, 374H, 374Yc, 375Kb, 382E, 382F, 382I, 394R; see also quasi-
homogeneous (374G)
measure space
probability algebra 333P, 333Yc, 334E, 372Xk
homomorphic image (of a Boolean algebra) 314M
homomorphism see Boolean homomorphism (312F), group homomorphism, lattice homomorphism
(3A1I), ring homomorphism (3A2D)
hull see convex hull (2A5E), closed convex hull (2A5E)
Humpty Dumptys privilege 372 notes
ideal in an algebra of sets 232Xc, 312C, 363Yf;see also -ideal (112Db)
in a Boolean ring or algebra 311D, 312C, 312K, 312O, 312Xd, 312Xi, 312Xk, 381S, 392Xd; see also
principal ideal (312D), -ideal (313E)
in a Riesz space see solid linear subspace (352J)
in a ring 352Xl, 3A2E, 3A2F, 3A2G
idempotent 363Xf, 363Xg
identically distributed random variables 273I, 274 notes, 276Yg, 285Xn, 285Yc, 372Xg; see also exchange-
able sequence (276Xe)
image lter 2A1Ib, 2A3Qb, 2A3S
image measure 112E, 112F, 112Xd, 123Ya, 132G, 132Yb, 132Yf, 211Xd, 212Bd, 212Xg, 235L, 254O,
256G, 342Xh, 342Xj, 343Xb, 343Yf, 386Xd
image measure catastrophe 235J, 343 notes
indecomposable transformation see ergodic inverse-measure-preserving function (372Pb)
indenite integral 131Xa, 222D-222F, 222H, 222I, 222Xa-222Xc, 222Yc, 224Xg, 225E, 225Od, 225Xh,
232D, 232E, 232Yf, 232Yi; see also indenite-integral measure
indenite-integral measure 234 (234A), 235M, 235P, 235Xi, 253I, 256E, 256J, 256L, 256Xe, 256Yd, 257F,
257Xe, 263Ya, 275Yi, 275Yj, 285Dd, 285Xe, 285Ya, 322Xf, 342Xd, 342Xn, 342Yd; see also uncompleted
indenite-integral measure
independence 272 (272A)
independent elements of L
0
(A) 364Xe, 364Xf, 367Xw
independent family in a probability algebra 325Xe, 325Yf, 371Yc
652 Index independent
independent random variables 272Ac), 272D-272I, 272L, 272M, 272P-272U, 272Xb, 272Xd, 272Xh-272Xj,
272Ya-272Yd, 272Yf, 272Yg, 273B, 273D, 273E, 273H, 273I, 273L-273N, 273Xh, 273Xi, 273Xk, 274B-274D,
274F-274K, 274Xc, 274Xd, 274Xg, 275B, 275Yh, 276Af, 285I, 285Xf, 285Xg, 285Xm-285Xo, 285Yc, 285Yk,
285Yl
independent sets 272Aa, 272Bb, 272F, 272N, 273F, 273K
independent subalgebras of a probability algebra 325L, 325Xe, 325Xf, 325Xg, 327Xe, 364Xe, 384Q, 384Sf,
384Xd-384Xf, 386B; see also Boolean-independent (315Xn)
independent -algebras 272Ab, 272B, 272D, 272F, 272J, 272K, 272M, 272O, 275Ym
Individual Ergodic Theorem see Ergodic Theorem (372D, 372E)
induced automorphism (on a principal ideal of a Boolean algebra) 387E, 387I, 387J, 387Xb
induced topology see subspace topology (2A3C)
inductive denitions 2A1B
inmum in a Boolean algebra 313A-313C
innity 112B, 133A, 135
information 384Cb
initial ordinal 2A1E, 2A1F, 2A1K
injective ring homomorphism 3A2Da, 3A2G
inner automorphism of a group 3A6B
inner measure 113Yh, 212Yc, 213Xe, 213Yc, 324Cb
inner product space 244N, 253Xe, 3A5L; see also Hilbert space (3A5L)
inner regular measure 256A, 256B, 342Aa, 342C, 342Xa
with respect to closed sets 256Ya, 342F
with respect to compact sets 342Xh, 343Yc
integrable function 122 (122M), 123Ya, 133B, 133Db, 133Dc, 133F, 133J, 133Xa, 135Fa, 212B, 212F,
213B, 213G; see also Bochner integrable function (253Yf), L
1
() (242A)
integral 122 (122E, 122K, 122M), 363L; (on an L-space) 356P; see also integrable function, Lebesgue
integral (122Nb), lower integral (133I), Riemann integral (134K), upper integral (133I)
integral operator see kernel operator, abstract integral operator (376)
integration by parts 225F, 225Oe, 252Xi
integration by substitution see change of variable in integration
interior of a set 2A3D, 2A3Ka
interpolation see Riesz Convexity Theorem
intersection number (of a family in a Boolean algebra) 391H, 391I-391K, 391Xh, 391Xi, 392E
interval see half-open interval (114Aa, 115Ab), open interval (111Xb, split interval (343J)
invariant function (in G-invariant function) 394A, 394N-394P, 394Xd, 394Xe, 394Z, 395B, 395Xd
ideal 381S
lower density 346Xd
see also rearrangement-invariant (374E), T -invariant (374A), translation-invariant
inverse (of an element in an f-algebra) 353Pc, 353Yh
inverse Fourier transform 283Ab, 283B, 283Wa, 283Xb, 284I; see also Fourier transform
inverse image (of a set under a function or relation) 1A1B
inverse-measure-preserving function 132G, 134Yl, 134Ym, 134Yn, 235G, 235H, 235I, 235Xe, 241Xh,
242Xf, 243Xn, 244Xo, 246Xf, 254G, 254H, 254K, 254O, 254Xc-254Xf, 254Xh, 254Yb, 324M, 324N, 341P,
341Xd, 341Yc, 341Yd, 343Ad, 343C, 343J, 343Xd, 343Xh, 343Yd, 365Xi, 372H-372L, 372Xe, 372Xf, 372Yh,
382Yd, 384S-384V, 384V; isomorphic inverse-measure-preserving functions 384Tb; see also almost isomor-
phic (384U); see also ergodic inverse-measure-preserving function (372P), image measure (112E), mixing
inverse-measure-preserving function (372P)
Inversion Theorem (for Fourier series and transforms) 282G-282I, 282L, 282O, 282P, 283I, 283L, 284C,
284M; see also Carlesons theorem
invertible element(in an f-algebra) 353Pc, 353Yh
involution in a group 3A6A, 381Xl
involution in Aut A 381K, 381N, 381O, 381Xd, 381Xe, 381Xf, 381Xh, 381Xk, 381Ya, 381Yc, 381Z, 382D,
382G, 382Yf, 384Xl; see also exchanging involution (381G), many involutions (381P)
Lebesgue General index 653
irreducible continuous function 313Ye
isodiametric inequality 264H, 264 notes
isolated point 316M, 316Yj, 316Yq, 3A3Pb
isometry 324Yh, 3A4Ff, 3A4G
isomorphic see almost isomorphic (384U)
isomorphism see Boolean isomorphism, measure space isomorphism, order-isomorphism
Jacobian 263Ea
Jensens inequality 233I, 242Yi, 365Rb
joint distribution see distribution (271C)
Jordan decomposition of an additive functional 231F, 231Ya, 232C, 326D, 326H, 326M
Kakutanis theorem 369E, 369Yb; see also Halmos-Rokhlin-Kakutani lemma (385E)
Kalton-Roberts theorem 392F, 392J
Kawadas theorem 394Q
Kechris A.S. see Becker-Kechris theorem, Nadkarni-Becker-Kechris theorem
Kelleys criterion 391L; see also Nachbin-Kelley theorem
kernel (of a ring homomorphism) 3A2D, 3A2Eb, 3A2G; see also Dirichlet kernel (282D), Fejer kernel
(282D), modied Dirichlet kernel (282Xc)
kernel operator 373Xf, 376
Kirzbrauns theorem 262C
Kolmogorovs Strong Law of Large Numbers 273I, 275Yn, 372Xg
Kolmogorov-Sina theorem 384P
Komlos theorem 276H, 276Yh, 367Xn
Kroneckers lemma 273Cb
Kwapiens theorem 375I
Lacey-Thiele Lemma 286M
Laplaces central limit theorem 274Xe
Laplace transform 123Xc, 123Yb, 133Xc, 225Xe
laterally complete Riesz space 368L, 368M
lattice 231Yc, 2A1Ad, 311L, 311Xi, 311Yb, 315Xb, 315Yg, 367A, 367B, 367Xa, 367Ye, 3A1I; see also
Riesz space (=vector lattice) (352A), Banach lattice (242G, 354Ab)
homomorphism 352Nd, 3A1Ia
norm see Riesz norm (242Xg, 354Aa)
of normal subgroups 381Xi, 382H, 382J, 382Xf, 382Xg, 382Ya, 3A6C
law of a random variable see distribution (271C)
law of large numbers see strong law (273)
Lebesgue, H. Vol. 1, intro., Chap. 27, intro.
Lebesgue Covering Lemma 2A2Ed
Lebesgue decomposition of a countably additive functional 232I, 232Yb, 232Yg, 256Ib
Lebesgue decomposition of a function of bounded variation 226C, 226Dc, 226Ya, 232Yb
Lebesgue density see lower Lebesgue density (341E)
Lebesgues Density Theorem 223, 261C, 275Xg
Lebesgues Dominated Convergence Theorem 123C, 133G, 363Yh, 367J, 367Xf
Lebesgue extension see completion (212C)
Lebesgue integrable function 122Nb, 122Yb, 122Ye, 122Yf
Lebesgue integral 122Nb
Lebesgue measurable function 121C, 121D, 134Xd, 225H, 262K, 262P, 262Yc
Lebesgue measurable set 114E, 114F, 114G, 114Xe, 114Ye, 115E, 115F, 115G, 115Yc
Lebesgue measure (on R) 114 (114E), 131Xb, 133Xc, 133Xd, 134G-134L, 212Xc, 216A, Chap. 22, 242Xi,
246Yd, 246Ye, 252N, 252O, 252Xf, 252Xg, 252Yj, 252Yp, 255, 342Xo, 343H, 344K, 345Xc, 346Xe, 372Xd,
382Xe, 383Q
(on R
r
) 115 (115E), 132C, 132Ee, 133Yc, 134, 211M, 212Xd, 245Yj, 251M, 251Wi, 252Q,
252Xh, 252Yu, 254Xk, 255A, 255K, 255L, 255Xd, 255Yc, 255Yd, 256Ha, 256J-256L, 264H, 264I, 342Ja,
344K, 345B, 345D, 345Xf
654 Index Lebesgue
(on [0, 1], [0, 1[) 211Q, 216A, 252Yq, 254K, 254Xh, 254Xj, 254Xk, 254Xl, 324Yg, 343Cb, 343J,
343Xd, 344K, 384Xj, 384Xk, 384Yd
(on other subsets of R
r
) 242O, 244Hb, 244I, 244Yh, 246Yf, 246Yl, 251Q, 252Ym, 255M, 255N,
255O, 255Ye, 255Yh, 343Cc, 343H, 343M, 344J, 345Xd, 346Xb, 372Yj, 372Yk, 387Yf
algebra 331Xd, 331Xe, 373C, 374C, 374D, 374Xa, 374Xf, 374Xh, 374Ya, 374Yd, 375Xe, 381Yd
Lebesgue negligible set 114E, 115E, 134Yk
Lebesgue outer measure 114C, 114D, 114Xc, 114Yd, 115C, 115D, 115Xb, 115Xd, 115Yb, 132C, 134A,
134D, 134Fa
Lebesgue set of a function 223D, 223Xf, 223Xg, 223Xh, 223Yg, 261E, 261Ye
Lebesgue-Stieltjes measure 114Xa, 114Xb, 114Yb, 114Yc, 114Yf, 131Xc, 132Xh, 134Xc, 211Xb, 212Xd,
212Xi, 225Xf, 232Xb, 232Yb, 235Xb, 235Xg, 235Xh, 252Xi, 256Xg, 271Xb, 224Yh, 342Xe
left-translation-invariant see translation-invariant
length of a curve 264Yl, 265Xd, 265Ya
length of an interval 114Ab
B.Levis theorem 123A, 123Xa, 133K, 135G, 135Hb, 226E, 242E, 365Df, 365Dh
Levi property of a normed Riesz space 242Yb, 244Ye, 354Db, 354J, 354N, 354Xi, 354Xn, 354Xo, 354Yi,
356Da, 356L, 356Xk, 365C, 366D, 369G, 369Xe, 369Yg, 371C, 371D, 371Xc, 371Xd
Levys martingale convergence theorem 275I
Levys metric 274Ya, 285Yd
lexicographic ordering 351Xa
Liapounos central limit theorem 274Xg
Liapounos vector measure theorem 326Ye
lifting of a measure (algebra) Chap. 34 (341A, 341Ya), 363Xe, 363Yg; see also linear lifting (363Xe),
translation-invariant lifting (345A)
Lifting Theorem 341K, 363Yg
limit of a lter 2A3Q, 2A3R, 2A3S
limit of a sequence 2A3M, 2A3Sg
limit ordinal 2A1Dd
Lindebergs central limit theorem 274F-274H, 285Ym
Lindebergs condition 274H
linear functional 3A5D; see also positive linear functional
linear lifting 341Xe, 341Ye, 345Yb, 363Xe, 363Yg
linear operator 262Gc, 263A, 265B, 265C, 2A6, 355D, 3A5E; see also bounded linear operator (2A4F),
(weakly) compact linear operator (3A5K), continuous linear operator, order-bounded linear operator
(355A), positive linear operator (351F)
linear order see totally ordered set (2A1Ac)
linear space topology see linear topological space (2A5A), weak topology (2A5I), weak* topology
(2A5I), T
s
(U, V ) (3A5E)
linear subspace (of a normed space) 2A4C; (of a partially ordered linear space) 351E, 351Rc
linear topological space 245D, 284Ye, 2A5A, 2A5B, 2A5C, 2A5Eb, 2A5F, 2A5G, 2A5H, 2A5I, 367N,
367Yn, 367Yo, 393L, 393M, 393O-393Q, 393Yh, 393Yi, 3A4Ad, 3A4Bd, 3A4Fg
linked set(in a Boolean algebra) 391L; (in -linked set, <-linked set) 391Yh; see also -linked (391L),
-m-linked (391Yh)
Lipschitz constant 262A, 262C, 264Yj
Lipschitz function 225Yc, 262 (262A), 263F, 264Yj, 282Yb
local convergence in measure see convergence in measure (245A)
local semigroup see full local semigroup (394A)
localizable measure algebra 322Ae, 322Be, 322C, 322K-322O, 323Gc, 323H, 323J, 323K, 323Xc, 323Yb,
323Z, 324Yf, 325C, 325D, 332B, 332C, 332J, 332O, 332S, 332T, 332Xo, 332Ya, 332Yc, 333H, 333Ia, 364Xk,
365J, 365K, 365S, 365T, 366Xb, 366Xf, 367Nc, 367V, 369A, 369B, 369D, 369E, 369R, 373L-373N, 373R,
373T, 373U, 375Yd, 382B-382E, 382Ga, 382Xa, 382Xb, 382Xd, 382Xg, 383M, 383N, 383Pa, 383Ya-383Yd,
391Cb, 394Xg, 394Xh, 394Yd, 395A; see also localization (322P)
Marczewski General index 655
localizable measure algebra free product 325D, 325Ea, 325H, 325Xa, 325Xb, 325Xc, 327Ya, 333E-333H,
333K, 333Q, 333R, 334B, 334Xb, 376B, 376E, 376F, 376Xb, 376Yb; see also probability algebra free product
(325K)
localizable measure (space) 211G, 211L, 211Ya, 211Yb, 212G, 213Hb, 213L-213N, 213Xl, 213Xm, 214Id,
214J, 214Xa, 214Xd, 214Xf, 216C, 216E, 216Ya, 216Yb, 234F, 234G, 234Ye, 241G, 241Ya, 243G, 243H,
245Ec, 245Yf, 252Yq, 252Yr, 252Yt, 254U, 322Be, 322N, 325B, 342N, 342Yb, 363S, 364Xm; see also strictly
localizable (211E)
localization of a semi-nite measure algebra 322P, 322Yb, 332Yb, 365Xm, 366Xe, 366Xf, 369Xp
locally compact measure (space) 342Ad, 342H, 342I, 342L-342N, 342Xc, 342Xd, 343B, 343Xa, 343Xg,
343Ya, 344H
locally compact topological space 3A3Ag, 3A3B, 3A3G, 3A3Hb
locally convex linear topological space 393Yi
locally determined measure (space) 211H, 211L, 211Ya, 216Xb, 216Ya, 216Yb, 251Xc, 252Ya, 322N; see
also complete locally determined measure
locally determined negligible sets 213I, 213J-213L, 213Xk, 213Xl, 214Ib, 214Xg, 214Xh, 216Yb, 234Yb,
252Yb, 376Yk
locally nite measure 256A, 256C, 256G, 256Xa, 256Ya
locally integrable function 242Xi, 255Xh, 255Xi, 256E, 261Xa, 262Yg
Lo`eve, M. Chap. 27 intro.
logistic map see quadratic map (372Xn)
Loomis-Sikorski theorem 314M
Lorentz norm 374Xj, 374Yb
lower density on a measure space 341 (341C, 341Ya), 343Yc, 345, 346F, 346G, 346Xa, 346Xc, 346Xd,
346Ye, 346Yf, 346Zb, 363Xe; see also lower Lebesgue density (341E)
lower integral 133I, 133J, 133Xe, 135H, 364Xm
lower Lebesgue density 223Yf, 341E
lower Riemann integral 134Ka
lower semi-continuous function 225H, 225I, 225Xl, 225Xm, 225Yd, 225Ye, 323Cb, 367Xx
Lusins theorem 134Yc, 256F
Luxemburg, W.A.J. 363 notes
magnitude of (an element in) a measure algebra 332Ga, 332J, 332O, 382Xc, 382Xd, 382Xg
magnitude of a measure space 332Ga, 343Yb, 344I
Maharam algebra see localizable measure algebra (322Ae)
Maharam measure (space) see localizable (211G)
Maharam submeasure 392G, 392H-392J, 392Xc, 392Xd, 392Xf, 393A-393C, 393E, 393M, 393Xc-393Xf,
393Ya, 393Yf, 393Yg
Maharams theorem 331I, 332B
Maharam type of a Boolean algebra 331F, 331H, 331Xc, 331Xg, 331Xh, 331Xj, 331Yd, 332H, 365Ya,
373Yb, 393Xe, 393Xf, 393Ya, 393Yf; see also relative Maharam type (333Aa)
of a measure algebra 331I-331K, 331Xd-331Xf, 331Xi, 332M, 332N, 332R-332T, 333D, 334,
369Xg, 375Kb, 386K, 386L, 386Xc, 387N, 387O
of a measure (space) 331Fc, 365Xp, 367Xr
Maharam-type-homogeneous Boolean algebra 331Fb, 331H, 331Xh, 331Xj, 332A, 332H, 332Xa; see also
homogeneous (331M)
measure algebra 331I, 331K, 331L, 331N, 331Xd, 331Xe, 331Yj, 344Xe; see also relatively
Maharam-type-homogeneous (333Ac)
measure (space) 331Fc, 334Xe, 334Xf, 334Ya, 341Yc, 341Yd, 346E
Maharam-type- component in a Boolean algebra 332Gb, 332H, 332J, 332O, 332P, 332Xj, 332Ya, 383N,
394Xh
many involutions (group with many involutions) 381P, 381Q-381S, 381Xi-381Xk, 382G, 383A, 383C,
383D, 383I, 394Xc
Marczewski functional 343E
656 Index Markov
Markov time see stopping time (275L)
Marriage Lemma see Halls Marriage Lemma (3A1K)
martingale 275 (275A, 275Cc, 275Cd, 275Ce); see also reverse martingale
martingale convergence theorems 275G-275I, 275K, 275Xf, 367K, 369Xq
martingale dierence sequence 276A, 276B, 276C, 276E, 276Xd, 276Ya, 276Yb, 276Ye, 276Yg
martingale inequalities 275D, 275F, 275Xb, 275Yc-275Ye, 276Xb, 372 notes
Max-ow Min-cut Theorem 332Xk
maximal element in a partially ordered set 2A1Ab
Maximal Ergodic Theorem 372C
maximal theorems 275D, 275Yc, 275Yd, 276Xb, 286A, 286T, 372C, 372Yb
Mazur S. see Banach-Mazur game
McMillan see Shannon-McMillan-Breiman theorem (385G)
meager set 314L, 314M, 314Yd, 316I, 316Yf, 316Yi, 341Yb, 3A3F, 3A3Ha
mean (of a random variable) see expectation (271Ab)
Mean Ergodic Theorem 372Xa
see also convergence in mean (245Ib)
measurable algebra 391B, 391C, 391D, 391K, 391Xf, 391Z, 392J, 393A, 393J, 393Ye, 394P, 394Q, 394Xf,
395B, 395Xd
measurable cover see measurable envelope (132D)
measurable envelope 132D, 132E, 132F, 132Xf, 132Xg, 132Xh, 134Fc, 134Xc, 213K-213M, 214G, 216Yc,
322I, 322J; see also full outer measure (132F)
measurable envelope property 213Xl, 214Xl, 364Xm
measurable function (taking values in R) 121 (121C), 122Ya, 212B, 212F, 213Yd, 214La, 214Ma, 235C,
235K, 252O, 252P, 256F, 256Yb, 256Yc, 316Yi, 322Yf; (taking values in R
r
) 121Yf), 256G; (taking values
in other spaces) 133Da, 133E, 133Yb, 135E, 135Xd, 135Yf; (, T)-measurable function 121Yb, 251Ya;
see also Borel measurable, Lebesgue measurable
measurable set 112A; -measurable set 212Cd; see also relatively measurable (121A)
measurable space 111Bc
measurable transformation 235, 365H; see also inverse-measure-preserving function
measure 112A; see also control measure (393O), vector measure (393O)
(in measures E, E is measured by ) >>>112Be
measure algebra 211Yb, 211Yc, vol. 3 (321A); isomorphic measure algebras 331I, 331L, 332B, 332C,
332J, 332K, 332P, 332Q, 332Ya, 332Yb
measure algebra of a measure space 321H-321K (321I), 322B, 322N, 327C; see also usual topology on a
measure algebra (323A)
measure-algebra topology 323 (323A), 324F, 324G, 324H, 324K, 324Xb, 324Xc, 327B, 327C, 365Ea,
367R, 367Xs, 367Yp
measure-algebra uniformity 323A, 323B, 323D, 323G, 324F, 324H, 327B
measure metric (on a measure algebra) 323A, 323Ca, 323Xg, 324Yh
measure-preserving Boolean automorphism 332L, 333Gb, 333P-333R, 333Yc, 343Jc, 344C, 344E-344G,
344Xf, 372Xk, 372Xl, 372Yi, 373Yb, 374E, 374G, 374L, 374Yc, 382D, 382Ye, 384P, 384Sb, 384Sf, 384Xg,
384Yd, 385P, 386B-386H, 386J, 386L, 387C, 387E, 387F, 387H-387K, 387M-387O, 387Xa-387Xe, 387Yf,
395Xc; see also automorphism group of a measure algebra, two-sided Bernoulli shift (384Qb)
measure-preserving Boolean homomorphism 324I, 324J-324P, 324Xd, 324Xe, 324Yf, 324Yh, 325C, 325D,
325I, 325J, 325Xd, 327Xa, 331D, 332L-332O, 332Q, 332Xm, 332Xn, 333C, 333D, 333F, 333Gb, 333Xc,
333Yd, 343B, 366Xf, 369Xm, 369Xn, 372G, 372H, 372Xc, 373Bd, 373U, 373Xi, 382Ye, 384L-384Q, 384T,
384V, 384Xf-384Xi, 384Xm, 384Xp, 384Yb-384Ye, 385D-385H, 385N, 385P, 385Xe, 385Xf, 385Yb-385Yd,
386A, 386B, 386F; isomorphic measure-preserving Boolean homomorphisms 384Ta; see also automorphism
of a measure algebra, Bernoulli shift (384Q), measure-preserving ring homomorphism
measure-preserving function see inverse-measure-preserving function (235G), measure space automor-
phism, measure space isomorphism
measure-preserving ring automorphism 366L, 366Xh
measure-preserving ring homomorphism 361Ad, 365O, 365Q, 366H, 366K, 366L, 366Xg, 366Xh, 366Yg,
369Xl, 372F; see also measure-preserving Boolean homomorphism
one-point General index 657
measure space 112 (112A), 113C, 113Yi
measure space automorphism 255A, 255Ca, 255N, 255Ya, 255 notes, 344C, 344E-344G, 344Xf, 345Ab,
345Xa, 346Xd, 384Sb
measure space isomorphism 254K, 254Xj, 254Xk, 254Xl, 255Ca, 255Mb, 344I, 344J, 344K, 344L, 344Xa-
344Xc, 344Xe-344Xg, 344Yc
metric 2A3F, 2A4Fb; see also entropy metric (384Xq), Euclidean metric (2A3Fb), Hausdor metric
(246Yb), Levys metric (274Ya), measure metric (323Ad), pseudometric (2A3F)
metric outer measure 264Xb, 264Yc
metric space 224Ye, 261Yi, 316Yj, 3A4Fe, 3A4Ff, 3A4Hc; see also complete metric space, metrizable
space (2A3Ff)
metrically transitive transformation see ergodic inverse-measure-preserving function (372Pb)
metrizable (topological) space 2A3Ff, 2A3L; see also metric space, separable metrizable space
uniformity 3A4Bc
mid-convex function 233Ya
minimal element in a partially ordered set 2A1Ab
mixing inverse-measure-preserving function 372P, 372Q, 372R, 372Xk, 372Xm-372Xo, 372Xq, 372Xr,
372Xu, 372Xv, 372Yh
mixing measure-preserving homomorphism 333P, 333Yc, 372P, 372Q, 372R, 372Xk, 372Xl, 372Xp, 384Se,
386Xb, 387Ye
modied Dirichlet kernel 282Xc
modulation 286C
Monotone Class Theorem 136B, 312Xc, 313G, 313Xd
Monotone Convergence Theorem see B.Levis theorem (133A)
monotonic function 121D, 222A, 222C, 222Yb, 224D, 313Xh
Monte Carlo integration 273J, 273Ya
multilinear map 253Xc
multiplicative identity (in an f-algebra) 353P, 353Yg
multiplicative Riesz homomorphism 352Xl, 353Pd, 361J, 363F, 364R
Nachbin-Kelley theorem 363R, 363Yj
natural extension (of an inverse-measure-preserving function) 382Ye
negligible set 112D, 131Ca, 214Cb, 344H, 344Yf
Neumann, J. von 341 notes
non-decreasing sequence of sets 112Ce
non-increasing sequence of sets 112Cf
non-measurable set 134B, 134D, 134Xg
norm 2A4B; (of a linear operator) 2A4F, 2A4G, 2A4I; (of a matrix) 262H, 262Ya; (norm topology)
242Xg, 2A4Bb; see also Fatou norm (354Da), Riesz norm (354Aa)
normal density function 274A, 283N, 283We, 283Wf
normal distribution function 274Aa, 274F-274K, 274M, 274Xe, 274Xg
normal random variable 274A, 274B, 285E, 285Xm, 285Xn
normal subgroup 381S, 381Xi, 381Xk, 382I, 382Xd, 382Yb; see also lattice of normal subgroups
normal subspace (of a Riesz space) see band (352O)
normalized Hausdor measure 264 notes, 265 (265A)
normed algebra 2A4J, 3A5J
normed space 224Yf, 2A4 (2A4Ba); see also Banach space (2A4D)
Novikovs Separation Theorem see Second Separation Theorem
nowhere dense set 313R, 313Yb, 315Yb, 316I, 316Yc, 316Yf, 316Yh, 3A3F
nowhere rigid Boolean algebra 383H, 383I-383L, 383Xa-383Xd
null set see negligible (112Da)
odd function 255Xb, 283Yd
Ogasawaras representation theorem for Riesz spaces 368 notes
one-point compactication 311Ya, 3A3O
658 Index open
open interval 111Xb, 114G, 115G, 1A1A, 2A2I
open map 313R
open set 111Gc, 111Yc, 114Yd, 115G, 115Yb, 133Xb, 134Fa, 134Yj, 135Xa, 1A2A, 1A2B, 1A2D,
256Ye, 2A3A, 2A3G; (in R) 111Ye, 114G, 134Xc, 2A2I; see also topology (2A3A)
operator topology see very weak operator topology (373K)
optional time see stopping time (275L)
orbit of a bijection 387C, 387Xa, 387Xb, 387Yb, 387Yc
order (in a Boolean ring) 311H
order-bounded dual Riesz space (U

) 356 (356Aa), 362A, 363K, it 365I, 365J, 366D, 366Ya, 368P,

371Xe, 375Yb
order-bounded set (in a partially ordered space) 2A1Ab, 354Xd, 368Xe
order-bounded linear operator (between Riesz spaces) 355 (355A); see also order-bounded dual (356A),
L

(355A)
order-closed set in a partially ordered space 313Da, 313Fa, 313Id, 313Xb, 313Xc, 314Ya, 316Fb, 323D,
352Oa, 354Xp; (ideal in a Boolean algebra) 313Eb, 313Pa, 313Qa, 313Xo, 314Xb, 314Yh, 316Xh, 316Xi,
316Xn, 316Xu; (Riesz subspace of a Riesz space) 353J, 354Xp, 375C; (subalgebra of a Boolean algebra)
313E-313G, 314Xg, 314Xh, 314Yc, 315Xm, 322Md, 322Xh, 323H-323J, 323Yc, 331E, 331G, 331Yb, 331Yc,
333Bc, 393Xd; see also sequentially order-closed (313Db)
order-complete see Dedekind complete (241Ec, 314Aa)
order-continuous bidual Riesz space (U

) 356I, 356J, 369D

order-continuous Boolean homomorphism 313L-313N, 313Xl-313Xn, 313Xp, 313Yf, 314Fa, 314G, 314R,
314T, 315Jc, 315P, 315Ya, 316Fd, 316Xy, 322Yd, 324E-324G, 324K, 324Ya, 324Ye, 324Yf, 325Aa, 325C,
325D, 325H, 326Kf, 331H, 331J, 332Xm, 332Xn, 334Xb, 343B, 344A, 344C, 344E, 344F, 352Xh, 363Ff,
364Rc, 364Yh, 366Xf, 369Xm, 369Xn, 373Bd, 373U, 375Ya, 385C, 395Xa
order-continuous dual Riesz space (U

) 356 (356Ac), 362A, 363K, 363S, 365I, 365J, 366D, 366Ya,

366Yc, 367Xf, 368Pc, 368Yj, 369A, 369C, 369D, 369K, 369Q, 369Xa, 369Xe, 369Yh, 371Xe, 375B, 375J,
375Yd
order-continuous norm (on a Riesz space) 242Yc, 242Ye, 244Yd, 313Yd, 326Yj, 354Dc, 354E, 354N,
354Xi, 354Xj, 354Xl, 354Xm, 354Xo, 354Xp, 354Yc-354Yh, 355K, 355Yg, 356D, 356M, 356Yg, 365C, 366D,
367E, 367Xu, 367Yn, 367Yr, 369B, 369Xf, 369Xg, 369Yc-369Yf, 371Xa-371Xc, 371Ya, 374Xe, 374Xg, 374Yb,
376L, 376M
order-continuous order-preserving function 313Ha, 313I, 313Xi, 313Yc, 313Yd, 315D, 315Yg, 316Fc,
326Kc, 361Cf, 361Gb, 363Eb, 363Ff, 366Yd, 367Xb, 367Yc, 384Ya, 385Yc, 385Yd, 392Xb, 394N; see also
sequentially order-continuous
order-continuous positive linear operator 351Ga, 355G, 355H, 355K, 364Rc, 366H, 368I, 375Xa, 375Xb,
375Yd; see also order-continuous dual (356A), order-continuous Riesz homomorphism, L

(355G)
order-continuous Riesz homomorphism 327Yb, 351Xc, 352N, 352Oe, 352Rb, 352Ub, 352Xd, 353O, 353Xc,
356I, 361Je, 355F, 366H, 367Xh, 368B, 375C, 375Xc, 375Ya
order-continuous ring homomorphism (between Boolean rings) 361A, 361Je, 365P
order*-convergent sequence in a partially ordered set 245Xc, 356Xd, 367 (>>>367A), 368Yg, 368Yh,
369Xq, 376G, 376H, 376Xe, 376Xg, 376Ye, 376Yh, 393Xa, 393Yd, 393Ye; (in C(X)) 367L, 367Yi, 367Yj,
367Yk; (in L
0
()) 245C, 245K, 245L, 245Xc, 245Xd, 376J; (in L
0
(A)) 367, 368Yi, 372D-372G, 372Xb,
385G, 385H
order-convex set in a partially ordered set 351Xb
order-dense set in a Boolean algebra 313J, 313K, 313Xm, 314Yg, 332A
order-dense Riesz subspace of a Riesz space 352N, 353A, 353D, 354Ef, 354I, 354Ya, 355F, 355J, 356I,
363C, 363Xd, 364L, 365F, 365G, 367Yf, 368B-368E, 368G-368I, 368M, 368P, 368S, 368Ya, 368Ye, 369A,
369B, 369C, 369D, 369G, 375D, 375Xc; see also quasi-order-dense (352Na)
order-dense subalgebra of a Boolean algebra 313O, 313Xj, 313Xn, 313Ye, 314I, 314T, 314Xf, 316Xf,
316Xo, 316Xr, 316Yk, 316Yn, 323Dc, 363Xd, 391Xj
order-isomorphism 312L
order*-limit see order*-convergent (367A)
order-preserving function 313H, 313I, 313La, 315D, 324Yg, 326Bf, 361Ce
order topology (on a partially ordered set) 313Xb, 313Xj, 313Yc, 326Yr, 367Yl, 393Xa, 393Yd
porous General index 659
order unit (in a Riesz space) 243C, 353L, 353M, 363N, 368Ya; see also order-unit norm (354Ga), standard
order unit (354Gc), weak order unit (353L)
order-unit norm 354F, 354G (354Ga), 354I-354K, 354Yi, 353Yf, 355Xc, 356N, 356O; see also M-space
(354Gb)
ordered set see partially ordered set (2A1Aa), totally ordered set (2A1Ac), well-ordered set (2A1Ae)
ordinal 2A1C, 2A1D-2A1F, 2A1K
ordinate set 252N, 252Yg, 252Yh
Orlicz norm 369Xd, 369Xi, 369Xj, 369Xn, 369Yc, 369Yd, 369Ye, 369Yg, 374Xc
Orlicz space 369Xd, 369Xi, 369Xj, 373Xm
Ornsteins theorem 386I, 386K
orthogonal matrix 2A6B, 2A6C
orthogonal projection in Hilbert space 244Nb, 244Yj, 244Yk, 366K
orthonormal vectors 2A6B
outer automorphism of a group 383G, 383O, 383Pb, 383Q, 383Yc, 383Yd, 3A6B
outer measure 113 (113A), 114Xd, 132B, 132Xh, 136Ya, 212Ea, 212Xa, 212Xb, 212Xg, 213C, 213Xa,
213Xg, 213Xk, 213Ya, 251B, 251Wa, 251Xd, 254B, 254G, 254L, 254S, 254Xq, 254Yd, 264B, 264Xa, 264Ya,
264Yo, 393T, 393Xh, 393Yh; see also Lebesgue outer measure (114C, 115C), metric outer measure
(264Yc), regular outer measure (132Xa), submeasure (392A)
dened from a measure 113Ya, 132 (132B), 213C, 213F, 213Xa, 213Xg-213Xi, 213Xk, 213Yd,
214Cd, 215Yc, 251O, 251R, 251Wk, 251Wm, 251Xm, 251Xo, 252Yh, 254Xb, 264F, 264Yd
outer regular measure 256Xi
Parsevals identity 284Qd
partial derivative 123D, 252Yj, 262I, 262J, 262Xh, 262Yb, 262Yc
partial lower density on a measure space 341Dc, 341N
partial order see partially ordered set (2A1Aa)
partially ordered linear space 231Yc, 241E, 241Yg, 326C, 351 (351A), 355Xa, 361C, 361G, 362Aa; see
also Riesz space (352A)
partially ordered set 2A1Aa, 313D, 313F, 313H, 313I, 313Xb, 313Xg, 313Xh, 313Yc, 315C, 315D, 315Xb,
315Yg
partition of unity in a Boolean ring or algebra 311G, 313K, 313L, 313Xk, 315E, 315F, 315Xk, 315Xq,
316Xp, 322E, 332E, 332I, 332Xi, 332Yb, 352T, 375H, 375I, 381B, 381C, 381J, 382F, 384C, 384D, 384G-
384P, 384R, 384Xa-384Xe, 384Xi, 384Xq, 384Ya-384Yc, 384Yg, 385B, 385C, 385G, 385K, 385L, 385N, 385O,
385Q, 385Xd, 385Xe, 387Lb; see also Bernoulli partition (386A)
pathological submeasure 393Tc
Peano curve 134Yl, 134Ym, 134Yn, 134Yo
perfect measure (space) 342K, 342L, 342M, 342Xh-342Xo, 343K-343M, 343Xg, 343Xh, 344C, 344I,
344Xb-344Xd, 344Yf, 384V
perfect Riesz space 356J, 356K-356M, 356P, 356Xg, 356Xi-356Xk, 356Ye, 356Yg, 365C, 365K, 366D,
369D, 369K, 369Q, 369Ya, 374M, 374Xk, 374Xl, 374Ya, 376P, 376Xl
periodic Boolean homomorphism of a Boolean algebra 385A, 385B, 385C, 385Xa
periodic extension of a function on ], ] 282Ae
permutation group see symmetric group
Plancherel Theorem (on Fourier series and transforms of square-integrable functions) 282K, 284O, 284Qd
pointwise convergence (topology on a space of functions) 281Yf
pointwise convergent see order*-convergent (245Cb, 367A)
pointwise topology see pointwise convergence
Poisson distribution 285Q, 285Xo
Poissons theorem 285Q
polar coordinates 263G, 263Xf
Polyas urn scheme 275Xc
polynomial (on R
r
) 252Yu
porous set 223Ye, 261Yg, 262L
660 Index positive
positive cone 253G, 253Xi, 253Yd, 351C
positive denite function 283Xt, 285Xr
positive linear operator (between partially ordered linear spaces) 351F, 355B, 355E, 355K, 355Xa, 361G,
367P, 375I-375K, 375Xa, 375Xe, 375Xg, 376Cc; see also (sequentially) order-continuous positive linear
operator, Riesz homomorphism (351H)
predictable sequence 276Ec
presque partout 112De
primitive product measure 251C, 251E, 251F, 251H, 251K, 251Wa, 251Xa-251Xc, 251Xe, 251Xf, 251Xj,
251Xl-251Xp, 252Yc, 252Yd, 252Yg, 253Ya-253Yc, 253Yg, 325Ya
principal band (in a Riesz space) 352V, 353C, 353H, 353Xb, 362Yd
principal ideal in a Boolean ring or algebra 312D, 312E, 312J, 312S, 312Xf, 312Yf, 314E, 314Xb, 315E,
315Xl, 316Xb, 316Xk, 316Xv, 321Xa, 322H-322J, 322Xf, 323Xe, 325Xb, 331Fb, 331H, 331M, 332A, 332L,
332P, 332Xh, 352Sc, 352Xf, 364Xo, 383Lb, 391Xe, 391Xj
principal projection property (of a Riesz space) 353Xb, 353Yd, 354Yi
principal ultralter 2A1N
probability algebra 322Aa, 322Ba, 322Ca, 322G, 322Mb, 384, 385G, 385J-385L, 385N-385Q, 385Xe,
385Yc, 385Yd, 386, 391B
probability algebra free product 325F, 325G, 325I-325M (325K) 325Xd, 334D, 334Xc, 384Xm, 387Ye
probability density function see density (271H)
probability space 211B, 211L, 211Q, 211Xb, 211Xc, 211Xd, 212G, 213Ha, 215B, 243Xi, 253H, 253Xh,
254, Chap. 27, 322Ba
product Boolean algebra see simple product (315A), free product (315H)
product f-algebra 352Wc, 364S
product measure Chap. 25; see also c.l.d. product measure (251F, 251W), primitive product measure
(251C), product probability measure (254C)
product partial order 315C, 315D, 315Xb
product partially ordered linear space 351L, 351Rd, 351Xc, 351Xd; see also product Riesz space
product probability measure 254 (254C), 272G, 272J, 272M, 275J, 275Yi, 275Yj, 281Yk, 325I, 334C,
334E, 334Xf, 334Xg, 334Ya, 342Gf, 342Xn, 343H, 346, 372Xf, 384S
product Riesz space 352K, 352T, 354Xb
product ring 3A2H
product topology 281Yc, 2A3T, 315Xf, 315Yb, 315Yd, 323L, 391Yc, 3A3I, 3A3J, 3A3K
product uniformity 3A4E
see also inner product space
projection (on a Riesz space) see band projection (352Rb)
projection band (in a Riesz space) 352R (352Ra), 352S, 352T, 352Xg, 353E, 353H, 353I, 353Xb, 361Xh,
361Ye, 362B
projection band algebra 352S, 352T, 361K, 363J, 363N, 363O, 364Q
pseudo-cycle (for a Boolean automorphism) 387Xc
pseudometric 2A3F, 2A3G, 2A3H, 2A3I, 2A3J, 2A3K, 2A3L, 2A3Mc, 2A3S, 2A3T, 2A3Ub, 2A5B, 323A,
3A3Be, 3A4B, 3A4D-3A4F
pseudo-simple function 122Ye, 133Ye
pull-back measures 132G
purely atomic Boolean algebra 316Lc, 316M, 316Xr-316Xx, 316Yk, 316Yl, 316Yq, 322Bh, 322Kc, 324Kg,
331Yd, 332Xb, 332Xc, 362Xf, 384J, 384Xq, 394Xf
purely atomic measure (space) 211K, 211N, 211R, 211Xb, 211Xc, 211Xd, 212G, 213He, 214Xe, 234Xb,
251Xr, 322Bh, 342Xn, 375Ya
purely innite measurable set 213 notes
quadratic map 372Xn
quasi-homogeneous measure algebra 374G, 374H-374M, 374Xl, 374Ye, 394Xg
quasi-order-dense Riesz subspace (of a Riesz space) 352N, 353A, 353K, 353Ya
quasi-Radon measure (space) 256Ya, 263Ya
representation General index 661
quasi-simple function 122Yd, 133Yd
quasi-Stonian topological space 314Yf
quotient Boolean algebra (or ring) 312K, 312Xk, 313Q, 313Xo, 314C, 314D, 314M, 314N, 314Yd, 316C,
316D, 316Xh, 316Xi, 316Xn, 316Xu, 316Ye, 316Yp, 361M, 363Gb, 391Xe, 391Yb, 392Xd
quotient partially ordered linear space 241Yg, 351J, 351K, 351Xb; see also quotient Riesz space
quotient Riesz space 241Yg, 241Yh, 242Yg, 352Jb, 352U, 354Yi, 361M, 364C, 364U
quotient ring 3A2F, 3A2G
quotient topology 245Ba
Rademachers theorem 262Q
Radon measure(on R or R
r
) 256 (256A), 284R, 284Yi, 342Jb, 342Xj; (on ], ]) 257Yb
Radon-Nikod ym derivative 232Hf, 232Yj, 234B, 234Ca, 234Yd, 234Yf, 235Xi, 256J, 257F, 257Xe, 257Xf,
272Xe, 272Yc, 275Ya, 275Yi, 285Dd, 285Xe, 285Ya, 363S
Radon-Nikod ym theorem 232E, 232F, 232G, 234G, 235Xk, 242I, 244Yk, 327D, 365E, 365Xf
Radon probability measure (on R or R
r
) 271B, 271C, 271Xb, 285Aa, 285M, 364Xd, 372Xf; (on other
spaces) 256Ye, 271Ya
Radon product measure (of nitely many spaces) 256K
random variable 271Aa
rapidly decreasing test function 284 (284A, 284Wa), 285Dc, 285Xd, 285Ya
rare see nowhere dense (3A3Fa)
rearrangement 373Ya; see also decreasing rearrangement (252Yp, 373C)
rearrangement-invariant extended Fatou norm 374Eb, 374F, 374K
rearrangement-invariant set 374Ea, 374F, 374M, 374Xk, 374Xl, 374Ye
Recurrence Theorem (for a measure-preserving Boolean homomorphism) 382Yc, 382Yd, 385D
recursion 2A1B
reduced power of R 351M, 351Q, 351Yd, 352L, 352M, 368F
rene 311Gd
reexive Banach space 372A, 3A5G
regular embedding of Boolean algebras 313N
regular measure see inner regular (256Ac)
regular open algebra (of a topological space) 314P-314S (314Q), 314Yd, 315Xe, 315Yd, 316Xe, 316Yb,
316Yd, 316Yf, 316Yj, 316Yq, 332Xe, 363Yc, 364U-364W, 364Yh-364Yl, 368Ya, 391Yc, 391Yf; (of R) 316K,
316Yb, 316Yn, 316Yo, 331Xh, 331Yi, 367Yq, 375Xe, 375Yf, 375Z, 376Yi, 391Xd
regular open set 314O, 314P, 314Q
regular operator (between Riesz spaces) 355 notes
regular outer measure 132C, 132Xa, 214Hb, 251Xm, 254Xb, 264Fb
regular topological space 2A5J, 316Yf, >>>3A3Ab, 3A3Ba, 3A3De; see also completely regular (3A3Ac)
regularly embedded (subalgebra of a Boolean algebra) 313N, 313O, 313P, 316Xm, 316Xs, 326Kf; (Riesz
subspace) 352Ne, 352Xd, 354Xk, 354Xm, 367F, 368Pa, 368S
relation 1A1B
relative atom in a Boolean algebra 331A
relative Maharam type of a Boolean algebra over a subalgebra 333A, 333B, 333C, 333E, 333F, 333Yb
relatively atomless (Boolean algebra) 331A, 385E, 387L, 394Xc
relatively compact set 2A3Na, 2A3Ob, 3A3De
relatively Maharam-type-homogeneous 333Ac, 333Bb
relatively measurable set 121A
relatively von Neumann transformation 387Ga, 387N, 387Xg, 387Yd, 387Ye
relatively weakly compact set (in a normed space) 247C, 2A5I, 356Q, 356Xl, 3A5Gb, 3A5Hb, 3A5Kb;
(in other linear spaces) 376O, 376P, 376Xl
repeated integral 252 (252A); see also Fubinis theorem, Tonellis theorem
representation of homomorphisms (between Boolean algebras) 344Ya, 344Yd, 364Xr; (between measure
algebras) 324A, 324B, 343A, 343B, 343G, 343J, 343M, 343Xc, 343Xf, 343Yd, 344A-344C, 344E-344G, 344Xf,
344Yc
662 Index respects
respects coordinates (said of a lifting) 346A, 346C, 346E, 346Xg, 346Yc, 346Za
reverse martingale 275K
Riemann integrable function 134K, 134L, 281Yh, 281Yi
Riemann integral 134K, 242 notes, 363Yi
Riemann-Lebesgue lemma 282E
Riesz Convexity Theorem 244 notes
Riesz homomorphism (between partially ordered linear spaces) 351H, 351J, 351L, 351Q, 351Xc, 351Ya,
352G; (between Riesz spaces) 352G-352J, 352W, 352Xb, 352Xd, 353Pd, 353Yf, 355Xe, 356Xh, 361Gc, 361J,
361Xg, 362Xe, 363Ec, 363F, 363Xb, 363Xc, 364R, 365N, 375I, 375J, 375Xg, 375Ya, 376Cc; see also order-
continuous Riesz homomorphism
Riesz norm 242Xg, 354A, 354B, 354D, 354F, 354M, 354Xc-354Xf, 354Xh, 354Yb, 354Yf, 354Yk, 355Xc,
356D, 356Xg, 356Xh; see also Fatou norm (354Da), order-continuous norm (354Dc), order-unit norm
(354Ga)
Riesz space (= vector lattice) 241Ed, 241F, 241Yc, 241Yg, chap. 35 (352A), 361Gc, 367C, 367E, 367Xc,
367Xg, 367Yo; see also Archimedean Riesz space (352), Banach lattice (354Ab), Riesz norm (354A)
Riesz subspace (of a partially ordered linear space) 352I; (of a Riesz space) 352I, 352J, 352L, 352M, 353A,
354O, 354Rc; see also band (352O), order-dense Riesz subspace (352N), solid linear subspace (351I)
rigid Boolean algebra 383Ha, 383L; see also nowhere rigid (383Hb)
ring 3A2 (3A2A); see also Boolean ring (311Aa)
ring homomorphism 3A2D, 3A2F-3A2H
ring homomorphism between Boolean rings 311D, 312Xf, 312Xg, 312Xh, 312Yc, 312Yd, 312Ye, 312Yf,
361A, 361Cc, 361J, 361Xe, 361Xg, 375H
Rokhlin see Halmos-Rokhlin-Kakutani lemma (385E)
root algebra (of a Bernoulli shift) 384Q, 384R, 384S, 386B, 386Ya
Saks see Vitali-Hahn-Saks theorem (246Yg)
saltus function 226B, 226Db, 226Xa
Schroder-Bernstein theorem 2A1G, 332 notes, 344D, 344Xa
Schwartz function see rapidly decreasing test function (284A)
Schwartzian distribution 284R, 284 notes; see also tempered distribution
self-supporting set (in a topological measure space) 256Xf
semi-continuous function see lower semi-continuous (225H)
semi-nite measure algebra 322 (322Ad), 323Dd, 323Ga, 323Xa, 324K, 324Xb, 325Ae, 325D, 327B,
331C, 332E, 332F, 332I, 332R, 332Xi, 332Yb, 364L, 365E, 365G, 365J, 365P, 365S, 366E, 366Xe, 366Xf,
366Xk, 367Nb, 368S, 369H, 369Xa, 371Xc, 373R, 375I, 382E, 383Ld, 383P, 383Xf, 391Ca
semi-nite measure (space) 211F, 211L, 211Ya, 212G, 213A, 213B, 213Hc, 213Xc, 213Xd, 213Xj, 213Xl,
213Xm, 213Ya-213Yc 214Xe, 214Xh, 215B, 216Xa, 216Yb, 234F, 235O, 235Xd, 235Xe, 241G, 241Ya, 241Yd,
243G, 245Ea, 245J, 245Xd, 245Xj, 245Xl, 246J, 246Xh, 251J, 251Xc, 252P, 252Yf, 253Xf, 253Xg, 322Bd,
322Yd, 327C, 327D, 342L, 342Xa, 342Xc, 342Xn, 343B, 344H, 365Xp, 367Xr
semi-nite version of a measure 213Xc, 213Xd, 322Xb
semigroup see topological semigroup
semi-martingale see submartingale (275Yf)
seminorm 2A5D
semi-ring of sets 115Ye
separable (topological) space 2A3Ud, 316Xd, 316Yb, 316Yj, 367Xr, 391Yc, 3A3E
separable Banach space 244I, 254Yc, 365Xp, 366Xc, 369Xg
separable metrizable space 245Yj, 264Yb, 284Ye
sequentially order-closed set in a partially ordered space 313Db, 313Xg, 313Yc, 316Fb, 353Ja, 367Yc;
see also -ideal (313Ec), -subalgebra (313Ec)
sequentially order-continuous additive function (on a Boolean algebra) 326Gc, 363Eb
Boolean homomorphism 313Lc, 313Pb, 313Qb, 313Yb, 314Fb, 314Hb, 314Xc, 314Ye, 315Ya,
316Fd, 324A, 324B, 324Kd, 324Xa, 324Xe, 324Yc, 326Ff, 343Ab, 363Ff, 364G, 364H, 364R, 364Xp, 364Yc,
364Yf, 364Yn, 365H, 365Xg, 375Ya, 375Ye
strictly General index 663
dual (of a Riesz space)(U

c
) 356Ab, 356B, 356D, 356L, 356Xa, 356Xb, 356Xc, 356Xd, 356Xf,
356Ya, 362Ac, 363K, 363S
function 313Hb, 313Ic, 313Xg, 313Yc, 315D, 316Fc, 361Cf, 361Gb, 367Xb, 367Yc, 375Xd, 392H
positive linear operator or functional 351Gb, 355G, 355I, 361Gb, 363Eb, 363Ff, 364R, 375A; see
also sequentially order-continuous dual (356A), L

c
(355G)
Riesz homomorphism 361Jf, 375Ya
ring homomorphism between Boolean rings 361Ac, 361Jf, 375Ka
Shannon-McMillan-Breiman theorem 385G, 385Xc
shift operators (on function spaces based on topological groups) 286C; see also Bernoulli shift (384Q)
Shoenelds Absoluteness Theorem 393 notes
Sierpi nski Class Theorem see Monotone Class Theorem (136B)
signed measure see countably additive functional (231C)
simple function 122 (>>>122A), 242M, 361D
simple group 381T, 382I, 382Xa, 382Xb, 382Xc
simple product of Boolean algebras 315 (315A), 316Xc, 316Xl, 316Xw, 332Xa, 332Xg, 364S, 381Yd,
391Xb, 391Xj
simple product of measure algebras 322K, 323L, 325Xc, 332B, 332Xm, 332Xn, 333H, 333Ia, 333K, 333R,
366Xi, 383Lc
Sinas theorem 386E, 386L
singular additive functional 232Ac, 232I, 232Yg
smooth function (on R or R
r
) 242Xi, 255Xi, 262Yd, 262Ye, 262Yf, 262Yg, 284A, 284Wa
smoothing by convolution 261Ye
solid hull (of a subset of a Riesz space) 247Xa, 352Ja
solid set (in a partially ordered linear space) 351I; (in a Riesz space) 354Xg
linear subspace (of a partially ordered linear space) 351J, 351K, 351Yb; (of a Riesz space) 352J, 353J,
353K, 353N, 355F, 355J, 355Yj, 368Ye, 382J; see also band (352O)
Souslin property see ccc (316Ab)
space-lling curve 134Yl
spectrum (of an M-space) 354L
sphere, surface measure on 265F-265H, 265Xa-265Xc, 265Xe
spherical polar coordinates 263Xf, 265F
split interval (= double arrow space) 343J, 343Xf, 343Yc, 344Xf
square-integrable function 244Na; see also L
2
standard extension of a countably additive functional 327F, 327G, 327Xa, 327Xd, 327Xe, 327Yc
standard normal distribution, standard normal random variable 274A
standard order unit (in an M-space) 354Gc, 354H, 354L, 356N, 356P, 363Ba, 363Ye
Steiner symmetrization 264H, 264 notes
step-function 226Xa
Stieltjes measure see Lebesgue-Stieltjes measure (114Xa)
Stirlings formula 252Yn
stochastic see doubly stochastic matrix
stochastically independent see independent (272A, 325L, 325Xe)
Stone representation of a Boolean ring or algebra 311E, 352 notes; see also Stone space
Stone space of a Boolean ring or algebra 311E, 311F, 311I-311K, 311Xg, 311Ya, 312O-312S, 312Xi-
312Xk, 312Yc-312Yf, 313C, 313R, 313Xp, 313Yb, 314M, 314S, 314T, 314Yd, 315H, 315Xc, 316B, 316I,
316Yb, 316Yc, 316Yi, 363A, 363Yf, 387Yb
Stone space of a measure algebra 321J, 321K, 322N, 322Q, 322Xi, 322Yb, 322Yf, 341O, 341P, 342Jc,
343B, 344A, 344Xe, 346K, 346L, 346Xf
Stone-Weierstrass theorem 281A, 281E, 281G, 281Ya, 281Yg
stopping time 275L, 275M-275O, 275Xi, 275Xj
strictly localizable measure (space) 211E, 211L, 211N, 211Ye, 212G, 213Ha, 213J, 213L, 213O, 213Xh,
213Xn, 213Ye, 214Ia, 214J, 215Xf, 216E, 234F, 235P, 251N, 251P, 251Xn, 252B, 252D, 252E, 252Ys, 252Yt,
322Kd, 322N, 322Qb, 322Xi, 325He, 341H, 341K, 342Hb, 343B, 344C, 344I, 344Xb, 344Xc, 346Xd, 384V
664 Index strictly
strictly positive additive functional (on a Boolean algebra) 391D, 391J, 391N, 395Xd see also chargeable
Boolean algebra (391X)
submeasure (on a Boolean algebra) 392B, 392F, 392I, 392J, 392Xc, 393Xa, 393Xb, 393Yd
strong law of large numbers 273D, 273H, 273I, 273Xh, 275Yn, 276C, 276F, 276Ye, 276Yg
strongly mixing see mixing (372P)
subalgebra of a Boolean algebra 312A, 312B, 312M, 312N, 312Xb, 312Xc, 312Xj, 313Fc, 313G, 313Xd,
313Xe, 315Xn, 315Xp, 316Xa, 331E, 331G, 332Xf, 363G, 391Xb, 391Xf, 391Xj; see also Boolean-independent
subalgebras (315Xn), closed subalgebra (323I), xed-point subalgebra (394Ga), independent subalge-
bras (325L), order-closed subalgebra, order-dense subalgebra, regularly embedded subalgebra (313N), -
subalgebra (233A, 313E)
subhomomorphism see -subhomomorphism (375E)
sublattice 3A1Ib
submartingale 275Yf, 275Yg
submeasure (on a Boolean algebra) 392 (392A), 393B, 393T, 393U, 393Yb; see also Maharam submea-
sure (392G), outer measure (113A), pathological submeasure (393Tc)
subring 3A2C
of a Boolean ring 311Xd, 312Xa, 312Xg
subspace measure 113Yb, 214A, 214B, 214C, 214H, 214I, 214Xb-214Xh, 216Xa, 216Xb, 241Ye, 242Yf,
243Ya, 244Yc, 245Yb, 251P, 251Q, 251Wl, 251Xn, 251Yb, 254L, 254Ye, 264Yf, 322I, 322J, 322Xg, 322Yd,
343H, 343M, 343Xa; (on a measurable subset) 131A, 131B, 131C, 132Xb, 214J, 214K, 214Xa, 214Xi, 241Yf,
247A, 342Ga, 342Ia, 342Xn, 343L, 344J, 344Xa, 344Xe, 344Xf; (integration with respect to a subspace
measure) 131D, 131E-131H, 131Xa-131Xc, 133Dc, 133Xa, 214D, 214E-214G, 214M
subspace of a normed space 2A4C
subspace topology 2A3C, 2A3J, 3A4Db
subspace uniformity 3A4D
subspace -algebra 121A, 214Ce
substitution see integration by substitution
successor cardinal 2A1Fc
ordinal 2A1Dd
sum over arbitrary index set 112Bd, 226A
sum of measures 112Xe, 112Ya, 212Xe, 212Xh, 212Xi, 212Xj, 212Yd, 212Ye, 334Xb, 334Xd
summable family of real numbers 226A, 226Xf
support of an additive functional on a Boolean algebra 326Xi
support of an automorphism of a Boolean algebra 381L, 381R, 381S, 381Yb, 383B
support of a measure (on a topological space) 256Xf, 257Xd
support of a submeasure on a Boolean algebra 392Xe
support see also bounded support, compact support
supporting (element in a Boolean algebra supporting an automorphism) 381D, 381E, 383B; see also
self-supporting set (256Xf), support
supremum 2A1Ab
in a Boolean algebra 313A-313C
surface measure see normalized Hausdor measure (265A)
symmetric distribution 272Ye
symmetric dierence (in a Boolean algebra) 311G
symmetrization see Steiner symmetrization
Tarskis theorem 394 notes
tempered distribution 284 notes
tempered function 284 (284D), 286D
tempered measure 284Yi
tensor product of linear spaces 253 notes, 376Ya-376Yc
tent map 372Xm, 384Xk
test function 242Xi, 284 notes; see also rapidly decreasing test function (284A)
upper General index 665
thick set see full outer measure (132F)
tight family of measures see uniformly tight (285Xj)
Tonellis theorem 252G, 252H, 252R
topological measure (space) 256A
topological space 2A3 (2A3A), 3A3
topological vector space see linear topological space (2A5A)
topology 2A2, 2A3 (2A3A); see also convergence in measure (245A, 367M), linear space topology
(2A5A), measure-algebra topology (323Ab)
total order see totally ordered set (2A1Ac)
total variation (of an additive functional) 231Yh; (of a function) see variation (224A)
totally nite measure algebra 322Ab, 322Bb, 322C, 322Mb, 322Qc, 323Ad, 323Ca, 324Kb, 324P, 327B,
331B, 331D, 332M, 332P, 332Q, 333C-333G, 333J-333N, 333Q, 333R, 366Yh, 373Xj, 373Xn, 373Yb, 375Fb,
375H, 382F, 382I, 382J, 382Xf, 382Ya, 382Yc, 383O, 383Xe, 385D-385F, 385M, 385Xd, 385Xf, 391B, 394R,
395Xa
totally nite measure (space) 211C, 211L, 211Xb, 211Xc, 211Xd, 212G, 213Ha, 214Ia, 214Ja, 215Yc,
232B, 232G, 243I, 243Xk, 245Fd, 245Xe, 245Ye, 246Xi, 246Ya, 322Bb, 382Yd
totally ordered set 135Ba, 2A1Ac, 3A1F
trace (of a -algebra) see subspace -algebra (121A)
transnite recursion 2A1B
translation-invariant inverse-measure-preserving function 386Xd
translation-invariant lifting 345A, 345B-345D, 345Ya-345Yc, 346C
translation-invariant lower density 345Xf, 345Ya
translation-invariant measure 115Xd, 134A, 134Ye, 134Yf, 255A, 255Ba, 255Yn, 345A, 345Xb, 345Ya
truly continuous additive functional 232Ab, 232B-232E, 232H, 232I, 232Xa, 232Xb, 232Xf, 232Xh,
232Ya, 232Ye, 234Ce, 327C, 362Xh, 363S
Tychono space see completely regular (3A3Ac)
Tychonos theorem 3A3J
type see Maharam type (331F)
Ulam S. see Banach-Ulam problem
ultralter 254Yd, 2A1N, 2A1O, 2A3R, 2A3Se, 351Yd, 3A3De, 3A3Lc; see also principal ultralter
(2A1N)
Ultralter Theorem 2A1O
uncompleted indenite-integral measure 234Cc
uniferent homomorphism 312F
Uniform Boundedness Theorem 3A5H
uniformity 3A4 (3A4A), 3A5I; see also measure-algebra uniformity (323Ab), uniform space (3A4A)
uniformly complete Riesz space 354Yi, 354Yk
uniformly continuous function 224Xa, 255K, 3A4C, 3A4Ec, 3A4G
uniformly convergent (sequence of functions) 3A3N
uniformly distributed sequence see equidistributed (281Yi)
uniformly exhaustive submeasure 392Bc, 392C, 392E, 392F, 392J, 392Xc, 393A
uniformly integrable set (in L
1
) 246 (>>>246A), 252Yp, 272Yd, 273Na, 274J, 275H, 275Xi, 275Yl, 276Xd,
276Yb; (in L
1
()) 246 (246A), 247C, 247D, 247Xe, 253Xd, 354Q; (in an L-space) 354P, 354Q, 354R,
356O, 356Q, 356Xm, 362E, 362Yf-362Yh, 371Xf; (in L
1
(A, )) 367Xm, 373Xj, 373Xn
uniformly tight (set of measures) 285Xj, 285Xk, 285Ye, 285Yf
unit ball in R
r
252Q
universal mapping theorems 253F, 254G, 315B, 315I, 315 notes, 325C, 325H, 325J, 369Xk
universally complete Riesz space see laterally complete (368L)
up-crossing 275E, 275F
upper envelope see upr(a, C) (314V)
upper integral 133I, 133J, 133K, 133Xe, 133Yf, 135H, 252Ye, 252Yh, 252Yi, 253J, 253K
upper Riemann integral 134Ka
666 Index upwards
upwards-directed partially ordered set 2A1Ab
usual measure on 0, 1
I
254J; see under 0, 1
I
usual measure on TX 254J; see under TX
vague topology (on a space of measures) 274Ld, 274Xh, 274Ya-274Yd, 285K, 285L, 285S, 285U, 285Xk,
285Xq, 285Yd, 285Yg-285Yi, 285Yn, 367Xv
variance of a random variable 271Ac, 271Xa, 272R, 272Xf, 285Gb, 285Xo
variation of a function 224 (224A, 224K, 224Yd, 224Ye), 226B, 226Db, 226Xc, 226Xd, 226Yb; see
also bounded variation (224A)
of a measure see total variation (231Yh)
vector integration see Bochner integral (253Yf)
vector lattice see Riesz space (241E, 352A)
vector measure 326Ye, 361Gb, 393O, 393P, 393Q, 393S, 393Xg, 393Yh, 393Yi
very weak operator topology 373K, 373L, 373Xp, 373Xv, 373Yg
virtually measurable function 122Q, 122Xe, 122Xf, 212Bb, 212F, 241A, 252E
Vitali cover 261Ya
Vitalis theorem 221A, 221Ya, 221Yc, 221Yd, 261B, 261Yk
Vitali-Hahn-Saks theorem 246Yg, 362Yh
volume 115Ac
of a ball in R
r
252Q, 252Xh
von Neumann transformation 387G, 387H, 387Xd, 387Xd, 387Xg see also relatively von Neumann
(387G), weakly von Neumann (387G)
Walds equation 272Xh
wandering see weakly wandering (395C)
weak operator topology see very weak operator topology (373K)
weak order unit (in a Riesz space) 353L, 353P, 353Yg, 368Yd
weak topology of a normed space 247Ya, 2A5I, 356Yf, 356Yg, 3A5Eb
see also very weak operator topology (373K), (relatively) weakly compact, weakly compactly gener-
ated, weakly convergent, weakly K-countably determined
weak* topology on a dual space 253Yd, 285Yg, 2A5Ig, 3A5E, 3A5F; see also vague topology (274Ld)
weakly compact linear operator 371Xb, 376Q, 3A5Kb; see also compact operator (3A5Ka)
weakly compact set (in a linear topological space) 247C, 247Xa, 247Xc, 247Xd, 2A5I, 376Yj; see also
relatively weakly compact (2A5Id)
weakly convergent sequence in a normed space 247Yb
weakly von Neumann transformation 387G, 387I, 387K, 387Xf, 387Ye
weakly wandering element 395C
weakly -renable see hereditarily weakly -renable
weakly -distributive Boolean algebra 316Yg, 316Yh, 362Ye, 375Yc, 391Ya
weakly (, )-distributive Boolean algebra 316G, 316H-316K, 316Xk-316Xq, 316Xx, 316Yf, 316Yg,
316Yi, 316Yj, 316Yp, 322F, 325Yd, 362D, 367Yl, 368Q, 368R, 368Xe, 368Yg, 368Yi, 376Yf, 391D, 391K,
391Ya, 391Yg, 391Z, 392I, 393Yd, 394Yc, 395Ya
weakly (, )-distributive Riesz space 368N, 368O-368S, 368Xd, 368Xe, 368Ye, 368Yf, 368Yh, 368Yj
376H
Weierstrass approximation theorem 281F; see also Stone-Weierstrass theorem
well-ordered set 2A1Ae, 2A1B, 2A1Dg, 2A1Ka; see also ordinal (2A1C)
Well-ordering Theorem 2A1Ka
Weyls Equidistribution Theorem 281M, 281N, 372Xo
Wieners Dominated Ergodic Theorem 372Yb
Youngs function >>>369Xc, 369Xd, 369Xr, 369Yc, 369Yd, 373Xm
Youngs inequality 255Ym
Zermelos Well-ordering Theorem 2A1Ka
Zermelo-Fraenkel set theory 3A1A
III General index 667
zero-dimensional topological space 311I-311K, 315Xf, 316Xo, 316Yd, 353Yc, 3A3Ad, 3A3Bd
zero-one law 254S, 272O, 272Xf, 272Xg, 325Xg
zero set in a topological space 313Yb, 316Yh, 324Yb, 3A3Pa
Zorns lemma 2A1M, 3A1G
a.e. (almost everywhere) 112Dd
a.s. (almost surely) 112De
Aut (in Aut A) see automorphism group of a Boolean algebra (381A); (in Aut

A) see automorphism
group of a measure algebra (382A)
A
L
(Lebesgue measure algebra) 373C
B (in B(x, ), closed ball) 261A, 2A2B
B (in B(U; V ), space of bounded linear operators) 253Xb, 253Yj, 253Yk, 2A4F, 2A4G, 2A4H, 371B-371D,
371G, 371Xd, 371Yc, 376M, 3A5H; (B(U; U)) 395Xb
c (in c(A), where A is a Boolean algebra) see cellularity (332D)
c (in c(X), where X is a topological space) see cellularity (332Xd)
c (the cardinal of R or TN) 2A1H, 2A1L, 343I, 343Yb, 344H, 344Yf, 382Xc, 391Yd, 391Ye
C (in C(X), where X is a topological space) 243Xo, 281Yc, 281Ye, 281Yf, 352Xj, 353M, 353Xd, 354L,
354Yf, 353Yc, 363A, 367L, 367Yi, 367Yj, 367Yk, 368Ya
C([0, 1]) 242 notes, 352Xg, 356Xb, 368Yf
C
b
(in C
b
(X), where X is a topological space) 281A, 281E, 281G, 281Ya, 281Yd, 281Yg, 285Yg, 352Xj,
354Hb
C

(in C

(X), for extremally disconnected X) 364W, 364Yl, 368G

ccc 354Xq, 354Xs, 355Ye
ccc
0
354Xa, 354Xd, 354Xi, 371Yc
cac (countable antichain condition) 316 notes
ccc Boolean algebra 316Aa, 316C-316F, 316Xa-316Xj, 316Yc-316Ye, 316Yg, 316Yp, 322G, 324Yd,
325Yd, 326L, 326Xi, 331Ge, 332D, 332H, 363Yb, 364Yb, 367Yl, 368Yg, 368Yi, 391M, 391Xa, 392Ca, 392I,
393J, 393Yd, 393Ye, 394Xf, 394Yb
ccc topological space 316Ab, 316B, 316Xd, 316Xe, 316Ya, 316Yd
cf (in cf P) see conality (3A1Fb)
c.l.d. product measure 251-253 (251F, 251W), 254Db, 254U, 254Ye, 256K, 256L, 325A, 325B, 325C,
325H, 334A, 334Xa, 342Ge, 342Id, 342Xn, 343H, 354Yl, 376J, 376R, 376S, 376Yc
c.l.d. version of a measure (space) 213E, 213F-213H, 213M, 213Xb-213Xe, 213Xg, 213Xk, 213Xn, 213Yb,
214Xf, 214Xj, 232Ye, 234Yf, 241Ya, 242Yh, 244Ya, 245Yc, 251Ic, 251S, 251Wn, 251Xd, 251Xj, 251Xk,
252Ya, 322D, 322Qb, 322Xc, 322Xi, 322Yb, 324Xc, 324Xe, 342Gb, 342Ib, 342Xn, 343H, 343Ye
CM

(formulations of the Control Measure Problem) 393A, 393H, 393J, 393L, 393P, 393Xe, 393Xg
d (in d(X)) see density (331Yf)
D (in D
n
(A, ), where A is a subset of a Boolean algebra, and is a homomorphism) 384K, 384L, 384M
diam (in diamA) = diameter
dom (in domf): the domain of a function f
ess sup see essential supremum (243Da)
E (in E(X), expectation of a random variable) 271Ab
f
(in A
f
) 361Ad
T (in T(B), T(B)) 323D
f-algebra 241H, 241 notes, 352W, 352Xj-352Xm, 353O, 353P, 353Xd, 353Yg, 353Yh, 361Eh, 363B,
364C-364E, 367Yg
G

set 264Xe
h (in h()) see entropy (384M); (in h(, A)) 384M, 384N-384P, 384Xq, 384Yb, 386C
H (in H(A)) see entropy of a partition (384C); (in H(A[B)) see conditional entropy (384D)
I
|
see split interval (343J)
668 Index LLL

1
(in
1
(X)) 242Xa, 243Xl, 246Xd, 247Xc, 247Xd, 354Xa, 356Xc

1
(=
1
(N)) 246Xc, 354M, 354Xd, 356Xl

2
244Xn, 282K, 282Xg, 355Yb, 371Ye, 376Yh, 376Ym, 376Yn

p
(in
p
(X)) 244Xn, 354Xa

(in

(X)) 243Xl, 281B, 281D, 354Ha, 354Xa, 361D, 361L

(=

(N)) 243Xl, 354Xj, 356Xa, 371Yd, 382J

-complemented subspace 363Yd

L-space 354M, 354N-354P, 354R, 354Xt, 354Yj, 356N, 356P, 356Q, 356Xm, 356Yf, 362A, 362B, 362Yj,
365C, 365Xc, 365Xd, 367Xn, 369E, 371A-371E, 371Xa, 371Xb, 371Xf, 371Ya, 376M, 376P, 376Yi, 376Yj
L
0
(in L
0
()) 121Xb, 121Ye, 241 (241A), 245, 253C, 253Ya; (in L
0
()) 345Yb, 364C, 364D, 364E,
364J, 364Yi; see also L
0
(241C), L
0
strict
(241Yh), L
0
C
(241J)
L
0
strict
241Yh
L
0
C
(in L
0
C
()) 241J, 253L
L
0
(in L
0
()) 241 (241A), 242B, 242J, 243A, 243B, 243D, 243Xe, 243Xj, 245, 253Xe, 253Xf, 253Xg,
271De, 272H, 323Xf, 345Yb, 352Xj, 364Jc, 376Yc; (in L
0
C
()) 241J; (in L
0
(A)) 364 (364A), 368A-368E,
368H, 368K, 368M, 368Qb, 368R, 368S, 368Xa, 368Xe, 368Ya, 368Yd, 368Yi, 369, 372H, 375, 376B,
376Yb, 393M, 394I; (in L
0
C
(A)) 364Yn; see also L
0
(241A, 364C)
L
1
(in L
1
()) 122Xc, 242A, 242Da, 242Pa, 242Xb; (in L
1
strict
() 242Yg, 341Ye; (in L
1
C
() 242P, 255Yn;
(in L
1
V
()) 253Yf; see also L
1
, | |
1
L
1
(in L
1
()) 242 (242A), 243De, 243F, 243G, 243J, 243Xf, 243Xg, 243Xh, 245H, 245J, 245Xh, 245Xi,
246, 247, 253, 254R, 254Xp, 254Ya, 254Yc, 255Xc, 257Ya, 282Bd, 327D, 341Ye, 354M, 354Q, 354Xa,
365B, 376N, 376Q, 376S, 376Yl; (in L
1
V
()) 253Yf, 253Yi, 354Yl; (in L
1
(A, ) or L
1

) 365 (365A), 366Yc,
367J, 367U, 367Yt, 369E, 369N, 369O, 369P, 371Xc, 371Yb, 371Yc, 371Yd, 372B, 372C, 372E, 372G, 372Xc,
376C, 385G, 385H, 385J; see also L
1
, L
1
C
, | |
1
L
1
C
() 242P, 243K, 246K, 246Yl, 247E, 255Xc; see also convolution of functions
L
2
(in L
2
()) 244Ob, 253Yj, 286; (in L
2
C
()) 284N, 284O, 284Wh, 284Wi, 284Xi, 284Xk-284Xm, 284Yg;
see also L
2
, L
p
, | |
2
L
2
(in L
2
()) 244N, 244Yk, 247Xe, 253Xe, 355Ye, 372 notes; (in L
2
C
()) 282K, 282Xg, 284P; (in L
2
(A, ))
366K, 366L, 366Xh, 395A, 395Xb; see also L
2
, L
p
, | |
2
L
p
(in L
p
()) 244 (244A), 246Xg, 252Ym, 253Xh, 255K, 255Og, 255Yc, 255Yd, 255Yl, 255Ym, 261Xa,
263Xa, 273M, 273Nb, 281Xd, 282Yc, 284Xj, 286A; see also L
p
, L
2
, | |
p
L
p
(in L
p
(), 1 < p < ) 244 (244A), 245G, 245Xj, 245Xk, 245Yg, 246Xh, 247Ya, 253Xe, 253Xi,
253Yk, 255Yf, 354Xa, 354Yk, 366B, 376N; (in L
p
(A, ) = L
p

, 1 < p < ) 366 (366A), 369L, 371Gd,
372Xp, 372Xq, 372Yb, 373Bb, 373F, 376Xb;(in L
p
C
(), 1 < p < ) 354Yk;(in L
p
(A, ), 0 < p < 1) 366Ya,
366Yg;see also L
p
, | |
p
L

(in L

()) 243A, 243D, 243I, 243Xa, 243Xl, 243Xn; (in L

()) 341Xe, 363H;see also L

C
243K
L

strict
243Xb
L

(in L

()) 243 (243A), 253Yd, 341Xe, 352Xj, 354Hc, 354Xa, 363I, 376Xn; (in L

(A)) 363 (363A),

364K, 364Xh, 365I, 365J, 365K, 365Xk, 368Q, 394N; see also L

, L

C
, | |

C
243K, 243Xm
L

(where is an extended Fatou norm) 369G, 369J, 369K, 369M, 369O, 369R, 369Xi, 374Xd, 374Xi;
see also Orlicz space (369Xd), L
p
, M
1,
(369N), M
,1
(369N)
L (in L(U; V ), space of linear operators) 253A, 253Xa, 351F, 351Xd, 351Xe
L

(in L

(U; V ), space of order-bounded linear operators) 355 (355A), 356Xi, 361H, 361Xc, 361Yb,
363Q, 365N, 371B-371E, 371Gb, 371Xb-371Xe, 371Ya, 371Yc-371Ye, 375Kb, 376J, 376Xe, 376Ym; see also
order-bounded dual (356A)
L

c
(in L

c
(U; V )) 355G, 355I, 355Yi, 376Yf; see also sequentially order-continuous dual (356A)
L

(in L

(U; V )) 355G, 355H, 355J, 355K, 355Yg, 355Yi, 355Yj, 371B-371D, 371Gb, 376D, 376E, 376H,
376K, 376Xj, 376Yf; see also order-continuous dual (356A)
lim (in limT) 2A3S; (in lim
xJ
) 2A3S
liminf (in liminf
n
) 1A3 (1A3Aa), 2A3Sg; (in liminf
0
) 2A2H; (in liminf
xJ
) 2A3S
limsup (in limsup
n
) 1A3 (1A3Aa), 2A3Sg; (in limsup
0
) 2A2H, 2A3Sg; (in limsup
xJ
f(x))
2A3S
TTT General index 669
ln
+
275Yd
M (in M(A), space of bounded nitely additive functionals) 362B, 362E, 363K
M-space 354Gb, 354H, 354L, 354Xq, 354Xr, 356P, 356Xj, 363B, 363O, 371Xd, 376M; see also order-unit
norm (354Ga)
M
0
(in M
0
(A, ) = M
0

) 366F, 366G, 366H, 366Yb, 366Yd, 366Yg, 373D, 373P, 373Xl
M
0,
252Yp
M
0,
(in M
0,
(A, ) = M
0,

) 373C, 373D, 373E, 373F, 373I, 373Q, 373Xo, 374B, 374J, 374L
M
1,0
(in M
1,0
(A, ) = M
1,0

) 366F, 366G, 366H, 366Ye, 369P, 369Q, 369Yh, 371F, 371G, 372D, 372Ya,
373G, 373H, 373J, 373S, 373Xp, 373Xr, 374Xe
M
1,
(in M
1,
()) 234Yd, 244Xl, 244Xm, 244Xo, 244Yc; (in M
1,
(A, ) = M
1,

) 369N, 369O-369Q,
369Xi-369Xk, 369Xm, 369Xq, 373, 374A, 374B, 374M
M
,0
(in M
,0
(A, )) 366Xd, 366Yc
M
,1
(in M
,1
(A, ) = M
,1

) 369N, 369O, 369P, 369Q, 369Xi, 369Xj, 369Xk, 369Xl, 369Yh, 373K,
373M, 374B, 374M, 374Xa, 374Ya
M

(in M

(A), space of countably additive functionals) 362B, 362Xd, 362Xh, 362Xi, 362Ya, 362Yb, 363K
M

(in M

(A), space of completely additive functionals) 326Yp, 327D, 362B, 362D, 362Xd, 362Xg, 362Xi,
362Ya, 362Yb, 363K
N 3A1H
N N 111Fb
N
N
372Xi
see also TN
^ see ideal of negligibles
On (the class of ordinals) 3A1E
p (in p(t)) 385I, 385J
T (in TX) 311Ba, 311Xe, 312B, 312C, 313Ec, 313Xf, 363S, 381Xe, 382Yb; (usual measure on TX) 254J,
254Xf, 254Xq, 254Yd
TN 1A1Hb, 2A1Ha, 2A1Lb, 315O, 316Yo, 324Yg, 326Yg, 374Xk; (usual measure on) 273G, 273Xd, 273Xe
T(N
N
) 316Yg
p.p. (presque partout) 112De
Pr(X > a), Pr(XXX E) etc. 271Ad
Q (the set of rational numbers) 111Eb, 1A1Ef, 364Yg
q (in q(t)) 384A, 385O
R (the set of real numbers) 111Fe, 1A1Ha, 2A1Ha, 2A1Lb, 352M
R
X
245Xa, 256Ye, 352Xj, 375Ya, 3A3K; see also Euclidean metric, Euclidean topology
R
X
[T see reduced power (351M)
R
C
2A4A
R see extended real line (135)
RO (in RO(X)) see regular open algebra (314Q)
S (in S(A)) 243I, 361 (361D), 363C, 363Xg, 364K, 364Xh, 365F, 368Q, 369O; (in S
f
= S(A
f
)) 242M,
244H, 365F, 365G, 369O, 369P; (in S(A)

) 362A; (in S(A)

c
) 362Ac; (in S(A)

) 362Ad; (in S
C
(A)) 361Xj,
361Yd
S see rapidly decreasing test function (284A)
S
1
(the unit circle, as topological group) see circle group
S
6
(the group of permutations of six elements) 383 notes
sf (in
sf
) see semi-nite version of a measure (213Xc); (in

sf
) 213Xf, 213Xg, 213Xk
T
2
topology see Hausdor (2A3E, 3A3Aa)
T
(0)
(in T
(0)
,
) 371F, 371G, 371H, 372D, 372Xb, 372Yb, 372Yc, 373B, 373G, 373J, 373R, 373S, 373Xp,
373Xq, 373Xr, 373Xu, 373Xv
T (in T
,
) 244Xm, 244Xo, 244Yc, 246Yc, 373 (373A); see also T -invariant (374A)
670 Index TTT
T

(in T

,
) 373 (373Ab), 376Xa, 376Xh
T -invariant extended Fatou norm 374Ab, 374B-374D, 374Fa, 374Xb, 374Xd-374Xj, 374Yb
T -invariant set 374Aa, 374M, 374Xa, 374Xi, 374Xk, 374Xl, 374Ya, 374Ye
T
m
see convergence in measure (245A)
T
s
(in T
s
(U, V )) 373M, 373Xq, 376O, 3A5E; see also weak topology (2A5I), weak* topology (2A5Ig)
U (in U(x, )) 1A2A
upr (in upr(a, C)) 314V, 314Xg, 333Xa, 365Rc, 394G, 394I, 394K-349N,
Var (in Var(X)) see variance (271Ac); (in Var
D
f, Var f) see variation (224A)
w

-topology see weak* topology 2A5Ig

Z (the set of integers) 111Eb, 1A1Ee; (as topological group) 255Xe
Z
2
(the group 0, 1) 311Bc, 311D
ZFC see Zermelo-Fraenkel set theory

r
(volume of unit ball in R
r
) 252Q, 252Xh, 265F, 265H, 265Xa, 265Xb, 265Xe
-function 225Xj, 225Xk, 252Xh, 252Yk, 252Yn, 255Xj
-system 2A1Pa
-renable see hereditarily weakly -renable

L
(in 373) 373C

XXX
see distribution of a random variable (271C)
- Theorem see Monotone Class Theorem (136B)
-additive see countably additive (231C, 326E)
-algebra of sets 111 (111A), 136Xb, 136Xi, 212Xk, 314D, 314M, 314N, 314Yd, 316D, 322Ya, 326Ys,
343D, 344D, 362Xg, 363H; see also Borel -algebra (111G)
-algebra dened by a random variable 272C, 272D
-complete see Dedekind -complete (241Fb, 314Ab)
-eld see -algebra (111A)
-nite measure algebra 322Ac, 322Bc, 322C, 322G, 322M, 323Gb, 323Ya, 324K, 325Eb, 327B, 331N,
362Xd, 367Nd, 367Q, 367Xq, 367Xs, 369Xg, 382E
-nite measure (space) 211D, 211L, 211M, 211Xe, 212G, 213Ha, 213Ma, 214Ia, 214Ja, 215B, 215C,
215Ya, 215Xe, 215Yb, 216A, 232B, 232F, 234F, 235O, 235R, 235Xe, 235Xk, 241Yd, 243Xi, 245Eb, 245K,
245L, 245Xe, 251K, 251Wg, 252B-252E, 252H, 252P, 252R, 252Xc, 252Yb, 252Yl, 322Bc, 342Xi, 362Xh,
365Xp, 367Xr, 376I, 376J, 376N, 376S
-generating set in a Boolean algebra 331E
-ideal (in a Boolean algebra) 313E, 313Pb, 313Qb, 314C, 314D, 314L, 314N, 314Yd, 316C, 316D, 316Xi,
316Ye, 321Ya, 322Ya, 392Xd
(of sets) 112Db, 211Xc, 212Xf, 212Xk, 313Ec, 322Ya, 363H
-linked Boolean algebra 391L, 391M, 391N, 391Xj, 391Yb, 391Yd, 391Ye, 391Z, 393Xf; -m-linked
Boolean algebra 391Yh, 393Ya
-order complete see Dedekind -complete (314Ab)
-order-continuous see sequentially order-continuous (313H)
-subalgebra of a Boolean algebra 313E, 313F, 313G, 313Xd, 313Xe, 314Eb, 314Fb, 314Hb, 314Jb,
314Xc, 315Yc, 321G, 321Xb, 322M, 323Z, 324Xb, 326Fg, 331E, 331G, 364Xc, 366I; see also order-closed
subalgebra
-subalgebra of sets 233 (233A), 321Xb, 323Xb
-subhomomorphism between Boolean algebras 375E, 375F-375H, 375Xd, 375Yc, 375Ye
(, )-distributive see weakly (, )-distributive (316G)

iI
a
i
112Bd, 222Ba, 226A
special symbols General index 671
(in (A)) see Maharam type (331Fa); (in
C
(A)) see relative Maharam type (333Aa)
-additive functional on a Boolean algebra see completely additive (326J)
-additive measure 256M, 256Xb, 256Xc
-generating set in a Boolean algebra 313Fb, 313M, 331E, 331F, 331G, 331Yb, 331Yc
see normal distribution function (274Aa)
(in A, where A is a set) 122Aa; (in a, where a belongs to a Boolean ring) 361D, 361Ef, 361L,
361M, 364K; (the function : A L
0
(A)) 364Kc, 367R
(the rst innite ordinal) 2A1Fa, 3A1H; (in [X]
<
) 3A1Cd, 3A1J

-bounding Boolean algebra see weakly -distributive (316Yg)

1
(the rst uncountable ordinal) 2A1Fc

1
-saturated ideal in a Boolean algebra 316C, 316D, 341Lh, 344Yd, 344Ye
see also T
1

2
2A1Fc

(the th uncountable initial ordinal) 3A1E

see closure (2A2A, 2A3Db)
(in f g, u v, , f, f ) see convolution (255E, 255O, 255Xe, 255Yn)
* (in weak*) see weak* topology (2A5Ig);(in U

= B(U; R), linear topological space dual) see dual

(2A4H)(in u

) see decreasing rearrangement (373C);(in

) see outer measure dened by a measure (132B)

(in

) see inner measure dened by a measure (113Yh)

(in U

) see order-bounded dual (356A)

c
(in U

c
) see sequentially order-continuous dual (356A)

(in U

) see order-continuous dual (356A); (in U

) see order-continuous bidual

(in E F, set dierence) 111C
_
(in
_
f,
_
fd,
_
f(x)(dx)) 122E, 122K, 122M, 122Nb, >>>363L; (in
_
A
f) 131D, 214D, 235Xf;
see also subspace measure; (in
_
u) 242Ab, 242B, 242D, 363L, 365D, 365Xa; (in
_
A
u) 242Ac; (in
_
a
u)
365D, 365Xb; see also upper integral, lower integral (133I)
_
see upper integral (133I)
_
see lower integral (133I)
R
_
see Riemann integral (134K)
(in fA, the restriction of a function to a set) 121Eh
[ [ (in a Riesz space) 241Ee, 242G, 352 (352C), 354Aa, 354Bb
| |
e
see order-unit norm (354Ga)
| |
1
(on L
1
()) 242 (242D), 246F, 253E, 275Xd, 282Ye; (on L
1
()) 242D, 242Yg, 273Na, 273Xi; (on
L
1
(A, )) 365A, 365B, 365C, 385G, 385H; (on the
1
-sum of Banach lattices) 354Xb, 354Xo
| |
2
244Da, 273Xj, 366Yh; see also L
2
, | |
p
| |
p
(for 1 < p < ) 244 (244Da), 246Xb, 246Xh, 246Xi, 252Ym, 252Yp, 253Xe, 253Xh, 273M, 273Nb,
275Xe, 275Xf, 275Xh, 276Ya, 366A, 366C, 366D, 366H, 366J, 366Xa, 366Xi, 366Yf, 367Xp, 369Oe, 372Xb,
372Yb, 374Xb; see also L
p
, L
p
, | |
p,q
| |
p,q
(the Lorentz norm) 374Yb
| |

243D, 243Xb, 243Xo, 244Xh, 273Xk, 281B, 354Xb, 354Xo, 356Xc, 361D, 361Ee, 361I, 361J,
361L, 361M, 363A, 364Xh; see also essential supremum (243D), L

, L

| |
1,
369O, 369P, 369Xh-369Xj, 371Gc, 372D, 372F, 373F, 373Xk; see also M
1,
, M
1,0
| |
,1
369N, 369O, 369Xi, 369Xj, 369Xl; see also M
,1
(in f g) 253B, 253C, 253J, 253L, 253Ya, 253Yb; (in u v) 253 (253E); (in A B, a b) see free
product (315M)

(in

iI
A
i
) see free product (315H)

(in

T) 251D, 251K, 251L, 251Xk, 251Ya, 252P, 252Xd, 252Xe, 253C

(in

iI

i
) 251Wb, 251Wf, 254E, 254F, 254Mc, 254Xc, 254Xi, 343Xb

(in

iI

i
) 254F; (in

iI
X
i
) 254Aa
# (in #(X), the cardinal of X) 2A1Kb
672 Index special symbols

(in

nN
E
n
) 111C; (in

/) 1A1F

(in

nN
E
n
) 111C; (in

c) 1A2F
(in EF, symmetric dierence) 111C, 311Ba
, (in a Boolean ring or algebra) 311Ga, 313Xi, 323B
\ , . (in a Boolean ring or algebra) 311Ga, 323B
, (in a Boolean ring or algebra) 311H, 323Xa
(in I R) see ideal (3A2Ea)
(

) (in (

b), (

c) etc.) see cycle notation (381G), cyclic automorphism, exchanging involution

(381G)
+
(in
+
, successor cardinal) 2A1Fc; (in f
+
, where f is a function) 121Xa, 241Ef; (in u
+
, where u
belongs to a Riesz space) 241Ef, 352C; (in U
+
, where U is a partially ordered linear space) 351C; (in
F(x
+
), where F is a real function) 226Bb

(in f

, where f is a function) 121Xa, 241Ef; (in u

, in a Riesz space) 241Ef, 352C; (in F(x

),
where F is a real function) 226Bb
2 (in 2

) 3A1D
, (in a lattice) 121Xa, 2A1Ad; (in A B, where A, B are partitions of unity in a Boolean algebra)
384F
(in A

, in a Boolean algebra) 313Xo; (in A

, in a Riesz space) 352O, 352P, 352Q, 352R, 352Xg; (in

V

, in a Hilbert space) see orthogonal complement; see also complement of a band

t
(in U
t
) see algebraic dual; (in T
t
) see adjoint operator

(in z

i) 3A1H

,

(in

f,

f) see Fourier transform, inverse Fourier transform (283A)
(in

h(u), where h is a Borel function and u L
0
) 241I, 241Xd, 241Xi, 245Dd, 364I, 364J, 364Xg,
364Xq, 364Yd, 364Ye, 367I, 367S, 367Xl, 367Ys
0, 1
I
(usual measure on) 254J, 254Xe, 254Yc, 272N, 273Xb, 331J-331L, 332B, 332C, 332N, 332Xm,
332Xn, 341Yc, 341Yd, 341Zb, 342Jd, 343Ca, 343I, 343Xc, 343Yd, 344G, 344L, 344Xg, 345Ab, 345C-345E,
345Xa, 346C, 382Xb; (when I = N) 254K, 254Xd, 254Xj, 256Xk, 261Yd, 341Xb, 343Cb, 343H, 343M,
345Yc, 346Zb, 387H; see also TX
(usual topology of) 311Xh, 3A3K; (when I = N) 314Ye
(open-and-closed algebra of) 311Xh, 315Xh, 316Xq, 316Yj, 316Ym, 331Yg, 391Xd, 393A, 393F
(regular open algebra of) 316Yj
(2, )-distributive lattice 367Yd
(in ) see absolutely continuous (232A)

G
(in a

G
b) 394A, 394G, 394I, 394K, 394Ma, 394Xb
see innity
[ ] (in [a, b]) see closed interval (115G, 1A1A, 2A1Ab); (in f[A], f
1
[B], R[A], R
1
[B]) 1A1B; (in
[X]

, [X]
<
, [X]

) 3A1J; (in [X]

<
) 3A1Cd, 3A1J
[[ ]] (in f[[T]]) see image lter (2A1Ib)
[[ ]] (in [[u > ]], [[u ]], [[u E]] etc.) 363Xh, 364A, 364B, 364H, 364Xa, 364Xc, 364Yc; (in [[ > ]])
326O, 326P
[ [ (in [a, b[) see half-open interval (115Ab, 1A1A)
] ] (in ]a, b]) see half-open interval (1A1A)
] [ (in ]a, b[) see open interval (115G, 1A1A)
| (in b : a|) 394I, 394J, 394K-394M, 394Xa
| (in b : a|) 394I, 394J, 394K-394M, 394Xa
(in E) 234E, 235Xf