Elias M. Stein, Rami Shakarchi - Real Analysis - Measure Theory, Integration, and Hilbert Spaces. Vol.3.-Princeton University Press (2005)

REAL ANALYSIS
Ibookroot October 20, 2007
Princeton Lectures in Analysis
I Fourier Analysis: An Introduction
II Complex Analysis
III Real Analysis:

Measure Theory, Integration, and
Hilbert Spaces
IV Functional Analysis: Introduction

to Further Topics in Analysis
Princeton Lectures in Analysis
III
REAL ANALYSIS
Measure Theory, Integration, and
Hilbert Spaces
Elias M. Stein
&
Rami Shakarchi
PRINCETON UNIVERSITY PRESS

PRINCETON AND OXFORD
Copyright © 2005 by Princeton University Press
Published by Princeton University Press, 41 William Street,
Princeton, New Jersey 08540
In the United Kingdom: Princeton University Press,
6 Oxford Street, Woodstock, Oxfordshire OX20 1TW
All Rights Reserved
Library of Congress Control Number 2004114065

ISBN 978-0-691-11386-9
British Library Cataloging-in-Publication Data is available
The publisher would like to acknowledge the authors of this volume for
providing the camera-ready copy from which this book was printed
Printed on acid-free paper. ∞
press.princeton.edu
Printed in the United States of America
3 5 7 9 10 8 6 4
To my grandchildren
Carolyn, Alison, Jason
E.M.S.
To my parents
Mohamed & Mireille
and my brother
Karim
R.S.
Foreword
Beginning in the spring of 2000, a series of four one-semester courses

were taught at Princeton University whose purpose was to present, in
an integrated manner, the core areas of analysis. The objective was to
make plain the organic unity that exists between the various parts of the
subject, and to illustrate the wide applicability of ideas of analysis to
other fields of mathematics and science. The present series of books is
an elaboration of the lectures that were given.
While there are a number of excellent texts dealing with individual
parts of what we cover, our exposition aims at a different goal: pre-
senting the various sub-areas of analysis not as separate disciplines, but
rather as highly interconnected. It is our view that seeing these relations
and their resulting synergies will motivate the reader to attain a better
understanding of the subject as a whole. With this outcome in mind, we
have concentrated on the main ideas and theorems that have shaped the
field (sometimes sacrificing a more systematic approach), and we have
been sensitive to the historical order in which the logic of the subject
developed.
We have organized our exposition into four volumes, each reflecting
the material covered in a semester. Their contents may be broadly sum-
marized as follows:
I. Fourier series and integrals.
II. Complex analysis.

III. Measure theory, Lebesgue integration, and Hilbert spaces.
IV. A selection of further topics, including functional analysis, distri-
butions, and elements of probability theory.
However, this listing does not by itself give a complete picture of

the many interconnections that are presented, nor of the applications
to other branches that are highlighted. To give a few examples: the ele-
ments of (finite) Fourier series studied in Book I, which lead to Dirichlet
characters, and from there to the infinitude of primes in an arithmetic
progression; the X-ray and Radon transforms, which arise in a number of
viii FOREWORD
problems in Book I, and reappear in Book III to play an important role in

understanding Besicovitch-like sets in two and three dimensions; Fatou’s
theorem, which guarantees the existence of boundary values of bounded
holomorphic functions in the disc, and whose proof relies on ideas devel-
oped in each of the first three books; and the theta function, which first
occurs in Book I in the solution of the heat equation, and is then used
in Book II to find the number of ways an integer can be represented as
the sum of two or four squares, and in the analytic continuation of the
zeta function.
A few further words about the books and the courses on which they
were based. These courses where given at a rather intensive pace, with 48
lecture-hours a semester. The weekly problem sets played an indispens-
able part, and as a result exercises and problems have a similarly im-
portant role in our books. Each chapter has a series of “Exercises” that
are tied directly to the text, and while some are easy, others may require
more effort. However, the substantial number of hints that are given
should enable the reader to attack most exercises. There are also more
involved and challenging “Problems”; the ones that are most difficult, or
go beyond the scope of the text, are marked with an asterisk.
Despite the substantial connections that exist between the different
volumes, enough overlapping material has been provided so that each of
the first three books requires only minimal prerequisites: acquaintance
with elementary topics in analysis such as limits, series, differentiable
functions, and Riemann integration, together with some exposure to lin-
ear algebra. This makes these books accessible to students interested
in such diverse disciplines as mathematics, physics, engineering, and
finance, at both the undergraduate and graduate level.
It is with great pleasure that we express our appreciation to all who

have aided in this enterprise. We are particularly grateful to the stu-
dents who participated in the four courses. Their continuing interest,
enthusiasm, and dedication provided the encouragement that made this
project possible. We also wish to thank Adrian Banner and José Luis
Rodrigo for their special help in running the courses, and their efforts to
see that the students got the most from each class. In addition, Adrian
Banner also made valuable suggestions that are incorporated in the text.
FOREWORD ix
We wish also to record a note of special thanks for the following in-
dividuals: Charles Fefferman, who taught the first week (successfully
launching the whole project!); Paul Hagelstein, who in addition to read-
ing part of the manuscript taught several weeks of one of the courses, and
has since taken over the teaching of the second round of the series; and
Daniel Levine, who gave valuable help in proof-reading. Last but not
least, our thanks go to Gerree Pecht, for her consummate skill in type-
setting and for the time and energy she spent in the preparation of all
aspects of the lectures, such as transparencies, notes, and the manuscript.
We are also happy to acknowledge our indebtedness for the support
we received from the 250th Anniversary Fund of Princeton University,
and the National Science Foundation’s VIGRE program.
Elias M. Stein
Rami Shakarchi
Princeton, New Jersey

August 2002
In this third volume we establish the basic facts concerning measure

theory and integration. This allows us to reexamine and develop further
several important topics that arose in the previous volumes, as well as to
introduce a number of other subjects of substantial interest in analysis.
To aid the interested reader, we have starred sections that contain more
advanced material. These can be omitted on first reading. We also want
to take this opportunity to thank Daniel Levine for his continuing help in
proof-reading and the many suggestions he made that are incorporated
in the text.
November 2004
Contents
Foreword vii
Introduction xv
1 Fourier series: completion xvi
2 Limits of continuous functions xvi
3 Length of curves xvii
4 Differentiation and integration xviii
5 The problem of measure xviii
Chapter 1. Measure Theory 1

1 Preliminaries 1
2 The exterior measure 10
3 Measurable sets and the Lebesgue measure 16
4 Measurable functions 27
4.1 Definition and basic properties 27
4.2 Approximation by simple functions or step functions 30
4.3 Littlewood’s three principles 33
5* The Brunn-Minkowski inequality 34
6 Exercises 37
7 Problems 46
Chapter 2. Integration Theory 49

1 The Lebesgue integral: basic properties and convergence
theorems 49
2 The space L1 of integrable functions 68
3 Fubini’s theorem 75
3.1 Statement and proof of the theorem 75
3.2 Applications of Fubini’s theorem 80
4* A Fourier inversion formula 86
5 Exercises 89
6 Problems 95
Chapter 3. Differentiation and Integration 98

1 Differentiation of the integral 99
1.1 The Hardy-Littlewood maximal function 100
1.2 The Lebesgue differentiation theorem 104
2 Good kernels and approximations to the identity 108
xi
xii CONTENTS
3 Differentiability of functions 114

3.1 Functions of bounded variation 115
3.2 Absolutely continuous functions 127
3.3 Differentiability of jump functions 131
4 Rectifiable curves and the isoperimetric inequality 134
4.1* Minkowski content of a curve 136
4.2* Isoperimetric inequality 143
5 Exercises 145
6 Problems 152
Chapter 4. Hilbert Spaces: An Introduction 156

2
1 The Hilbert space L 156
2 Hilbert spaces 161
2.1 Orthogonality 164
2.2 Unitary mappings 168
2.3 Pre-Hilbert spaces 169
3 Fourier series and Fatou’s theorem 170
3.1 Fatou’s theorem 173
4 Closed subspaces and orthogonal projections 174
5 Linear transformations 180
5.1 Linear functionals and the Riesz representation the-
orem 181
5.2 Adjoints 183
5.3 Examples 185
6 Compact operators 188
7 Exercises 193
8 Problems 202
Chapter 5. Hilbert Spaces: Several Examples 207

2
1 The Fourier transform on L 207
2 The Hardy space of the upper half-plane 213
3 Constant coefficient partial differential equations 221
3.1 Weak solutions 222
3.2 The main theorem and key estimate 224
4* The Dirichlet principle 229
4.1 Harmonic functions 234
4.2 The boundary value problem and Dirichlet’s principle 243
5 Exercises 253
6 Problems 259
CONTENTS xiii
Chapter 6. Abstract Measure and Integration Theory 262

1 Abstract measure spaces 263
1.1 Exterior measures and Carathéodory’s theorem 264
1.2 Metric exterior measures 266
1.3 The extension theorem 270
2 Integration on a measure space 273
3 Examples 276
3.1 Product measures and a general Fubini theorem 276
3.2 Integration formula for polar coordinates 279
3.3 Borel measures on R and the Lebesgue-Stieltjes in-
tegral 281
4 Absolute continuity of measures 285
4.1 Signed measures 285
4.2 Absolute continuity 288
5* Ergodic theorems 292
5.1 Mean ergodic theorem 294
5.2 Maximal ergodic theorem 296
5.3 Pointwise ergodic theorem 300
5.4 Ergodic measure-preserving transformations 302
6* Appendix: the spectral theorem 306
6.1 Statement of the theorem 306
6.2 Positive operators 307
6.3 Proof of the theorem 309
6.4 Spectrum 311
7 Exercises 312
8 Problems 319
Chapter 7. Hausdorff Measure and Fractals 323

1 Hausdorff measure 324
2 Hausdorff dimension 329
2.1 Examples 330
2.2 Self-similarity 341
3 Space-filling curves 349
3.1 Quartic intervals and dyadic squares 351
3.2 Dyadic correspondence 353
3.3 Construction of the Peano mapping 355
4* Besicovitch sets and regularity 360
4.1 The Radon transform 363
4.2 Regularity of sets when d ≥ 3 370
4.3 Besicovitch sets have dimension 2 371
4.4 Construction of a Besicovitch set 374
xiv CONTENTS
5 Exercises 380
6 Problems 385
Notes and References 389
Bibliography 391
Symbol Glossary 395
Index 397
Introduction
I turn away in fright and horror from this lamentable

plague of functions that do not have derivatives.
C. Hermite, 1893
Starting in about 1870 a revolutionary change in the conceptual frame-

work of analysis began to take shape, one that ultimately led to a vast
transformation and generalization of the understanding of such basic ob-
jects as functions, and such notions as continuity, differentiability, and
integrability.
The earlier view that the relevant functions in analysis were given by
formulas or other “analytic” expressions, that these functions were by
their nature continuous (or nearly so), that by necessity such functions
had derivatives for most points, and moreover these were integrable by
the accepted methods of integration − all of these ideas began to give
way under the weight of various examples and problems that arose in
the subject, which could not be ignored and required new concepts to
be understood. Parallel with these developments came new insights that
were at once both more geometric and more abstract: a clearer under-
standing of the nature of curves, their rectifiability and their extent; also
the beginnings of the theory of sets, starting with subsets of the line, the
plane, etc., and the “measure” that could be assigned to each.
That is not to say that there was not considerable resistance to the
change of point-of-view that these advances required. Paradoxically,
some of the leading mathematicians of the time, those who should have
been best able to appreciate the new departures, were among the ones
who were most skeptical. That the new ideas ultimately won out can
be understood in terms of the many questions that could now be ad-
dressed. We shall describe here, somewhat imprecisely, several of the
most significant such problems.
xv
xvi INTRODUCTION
1 Fourier series: completion

Whenever f is a (Riemann) integrable
P function on [−π, π] we defined in
Book I its Fourier series f ∼ an einx by
Z π
1
(1) an = f (x)e−inx dx,
2π −π
and saw then that one had Parseval’s identity,

∞
X Z π
2 1
|an | = |f (x)|2 dx.
2π −π
n=−∞
However, the above relationship between functions and their Fourier

coefficients is not completely reciprocal when limited to Riemann inte-
grable functions. Thus if we consider the space R of such functions with
its square norm, and the space `2 (Z) with its norm,1 each element f in
R assigns a corresponding element {an } in `2 (Z), and the two norms are
identical. However, it is easy to construct elements in `2 (Z) that do not
correspond to functions in R. Note also that the space `2 (Z) is complete
in its norm, while R is not.2 Thus we are led to two questions:
(i) What are the putative “functions” f that arise when we complete
R? In other words: given an arbitrary sequence {an } ∈ `2 (Z) what
is the nature of the (presumed) function f corresponding to these
coefficients?
(ii) How do we integrate such functions f (and in particular verify (1))?
2 Limits of continuous functions

Suppose {fn } is a sequence of continuous functions on [0, 1]. We assume
that limn→∞ fn (x) = f (x) exists for every x, and inquire as to the nature
of the limiting function f .
If we suppose that the convergence is uniform, matters are straight-
forward and f is then everywhere continuous. However, once we drop
the assumption of uniform convergence, things may change radically and
the issues that arise can be quite subtle. An example of this is given by
the fact that one can construct a sequence of continuous functions {fn }
converging everywhere to f so that
1 We use the notation of Chapter 3 in Book I.

2 See the discussion surrounding Theorem 1.1 in Section 1, Chapter 3 of Book I.
3. Length of curves xvii
(a) 0 ≤ fn (x) ≤ 1 for all x.

(b) The sequence fn (x) is montonically decreasing as n → ∞.
(c) The limiting function f is not Riemann integrable.3
R1
However, in view of (a) and (b), the sequence 0 fn (x) dx converges to
a limit. So it is natural to ask: what method of integration can be used
to integrate f and obtain that for it
Z 1 Z 1
f (x) dx = lim fn (x) dx ?
0 n→∞ 0
It is with Lebesgue integration that we can solve both this problem

and the previous one.
3 Length of curves
The study of curves in the plane and the calculation of their lengths
are among the first issues dealt with when one learns calculus. Suppose
we consider a continuous curve Γ in the plane, given parametrically by
Γ = {(x(t), y(t))}, a ≤ t ≤ b, with x and y continuous functions of t. We
define the length of Γ in the usual way: as the supremum of the lengths
of all polygonal lines joining successively finitely many points of Γ, taken
in order of increasing t. We say that Γ is rectifiable if its length L is
finite. When x(t) and y(t) are continuously differentiable we have the
well-known formula,
Z b
¡ 0 ¢1/2
(2) L= (x (t))2 + (y 0 (t))2 dt.
a
The problems we are led to arise when we consider general curves.

More specifically, we can ask:
(i) What are the conditions on the functions x(t) and y(t) that guar-
antee the rectifiability of Γ?
(ii) When these are satisfied, does the formula (2) hold?
The first question has a complete answer in terms of the notion of func-
tions of “bounded variation.” As to the second, it turns out that if x and
y are of bounded variation, the integral (2) is always meaningful; how-
ever, the equality fails in general, but can be restored under appropriate
reparametrization of the curve Γ.
3 The limit f can be highly discontinuous. See, for instance, Exercise 10 in Chapter 1.
xviii INTRODUCTION
There are further issues that arise. Rectifiable curves, because they
are endowed with length, are genuinely one-dimensional in nature. Are
there (non-rectifiable) curves that are two-dimensional? We shall see
that, indeed, there are continuous curves in the plane that fill a square,
or more generally have any dimension between 1 and 2, if the notion of
fractional dimension is appropriately defined.
4 Differentiation and integration

The so-called “fundamental theorem of the calculus” expresses the fact
that differentiation and integration are inverse operations, and this can
be stated in two different ways, which we abbreviate as follows:
Z b
(3) F (b) − F (a) = F 0 (x) dx,
a
Z x
d
(4) f (y) dy = f (x).
dx 0
For the first assertion, the existence of continuous functions F that are
nowhere differentiable, or for which F 0 (x) exists for every x, but F 0 is
not integrable, leads to the problem of finding a general class of the F for
which (3) is valid. As for (4), the question is to formulate properly and
establish this assertion for the general class of integrable functions f that
arise in the solution of the first two problems considered above. These
questions can be answered with the help of certain “covering” arguments,
and the notion of absolute continuity.
5 The problem of measure

To put matters clearly, the fundamental issue that must be understood
in order to try to answer all the questions raised above is the problem
of measure. Stated (imprecisely) in its version in two dimensions, it
is the problem of assigning to each subset E of R2 its two-dimensional
measure m2 (E), that is, its “area,” extending the standard notion defined
for elementary sets. Let us instead state more precisely the analogous
problem in one dimension, that of constructing one-dimensional measure
m1 = m, which generalizes the notion of length in R.
We are looking for a non-negative function m defined on the family of
subsets E of R that we allow to be extended-valued, that is, to take on
the value +∞. We require:
5. The problem of measure xix
(a) m(E) = b − a if E is the interval [a, b], a ≤ b, of length b − a.

P∞ S∞
(b) m(E) = n=1 m(En ) whenever E = n=1 En and the sets En are
disjoint.
Condition (b) is the “countable additivity” of the measure m. It implies
the special case:
(b0 ) m(E1 ∪ E2 ) = m(E1 ) + m(E2 ) if E1 and E2 are disjoint.
However, to apply the many limiting arguments that arise in the theory
the general case (b) is indispensable, and (b0 ) by itself would definitely
be inadequate.
To the axioms (a) and (b) one adds the translation-invariance of m,
namely
(c) m(E + h) = m(E), for every h ∈ R.
A basic result of the theory is the existence (and uniqueness) of such
a measure, Lebesgue measure, when one limits oneself to a class of rea-
sonable sets, those which are “measurable.” This class of sets is closed
under countable unions, intersections, and complements, and contains
the open sets, the closed sets, and so forth.4
It is with the construction of this measure that we begin our study.
From it will flow the general theory of integration, and in particular the
solutions of the problems discussed above.
A chronology
We conclude this introduction by listing some of the signal events that
marked the early development of the subject.
1872 − Weierstrass’s construction of a nowhere differentiable function.
1881 − Introduction of functions of bounded variation by Jordan and
later (1887) connection with rectifiability.
1883 − Cantor’s ternary set.
1890 − Construction of a space-filling curve by Peano.
1898 − Borel’s measurable sets.
1902 − Lebesgue’s theory of measure and integration.
1905 − Construction of non-measurable sets by Vitali.
1906 − Fatou’s application of Lebesgue theory to complex analysis.
4 There is no such measure on the class of all subsets, since there exist non-measurable
sets. See the construction of such a set at the end of Section 3, Chapter 1.
1 Measure Theory
The sets whose measure we can define by virtue of the
preceding ideas we will call measurable sets; we do
this without intending to imply that it is not possible
to assign a measure to other sets.
E. Borel, 1898
This chapter is devoted to the construction of Lebesgue measure in Rd

and the study of the resulting class of measurable functions. After some
preliminaries we pass to the first important definition, that of exterior
measure for any subset E of Rd . This is given in terms of approximations
by unions of cubes that cover E. With this notion in hand we can
define measurability and thus restrict consideration to those sets that
are measurable. We then turn to the fundamental result: the collection
of measurable sets is closed under complements and countable unions,
and the measure is additive if the subsets in the union are disjoint.
The concept of measurable functions is a natural outgrowth of the
idea of measurable sets. It stands in the same relation as the concept
of continuous functions does to open (or closed) sets. But it has the
important advantage that the class of measurable functions is closed
under pointwise limits.
1 Preliminaries
We begin by discussing some elementary concepts which are basic to the
theory developed below.
The main idea in calculating the “volume” or “measure” of a subset
of Rd consists of approximating this set by unions of other sets whose
geometry is simple and whose volumes are known. It is convenient to
speak of “volume” when referring to sets in Rd ; but in reality it means
“area” in the case d = 2 and “length” in the case d = 1. In the approach
given here we shall use rectangles and cubes as the main building blocks
of the theory: in R we use intervals, while in Rd we take products of
intervals. In all dimensions rectangles are easy to manipulate and have
a standard notion of volume that is given by taking the product of the
length of all sides.
2 Chapter 1. MEASURE THEORY
Next, we prove two simple theorems that highlight the importance of

these rectangles in the geometry of open sets: in R every open set is a
countable union of disjoint open intervals, while in Rd , d ≥ 2, every open
set is “almost” the disjoint union of closed cubes, in the sense that only
the boundaries of the cubes can overlap. These two theorems motivate
the definition of exterior measure given later.
We shall use the following standard notation. A point x ∈ Rd consists
of a d-tuple of real numbers
x = (x1 , x2 , . . . , xd ), xi ∈ R, for i = 1, . . . , d.
Addition of points is componentwise, and so is multiplication by a real

scalar. The norm of x is denoted by |x| and is defined to be the standard
Euclidean norm given by
¡ ¢1/2
|x| = x21 + · · · + x2d .
The distance between two points x and y is then simply |x − y|.

The complement of a set E in Rd is denoted by E c and defined by
E c = {x ∈ Rd : x ∈
/ E}.
If E and F are two subsets of Rd , we denote the complement of F in E

by
E − F = {x ∈ Rd : x ∈ E and x ∈
/ F }.
The distance between two sets E and F is defined by
d(E, F ) = inf |x − y|,
where the infimum is taken over all x ∈ E and y ∈ F .
Open, closed, and compact sets

The open ball in Rd centered at x and of radius r is defined by
Br (x) = {y ∈ Rd : |y − x| < r}.
A subset E ⊂ Rd is open if for every x ∈ E there exists r > 0 with

Br (x) ⊂ E. By definition, a set is closed if its complement is open.
We note that any (not necessarily countable) union of open sets is
open, while in general the intersection of only finitely many open sets
1. Preliminaries 3
is open. A similar statement holds for the class of closed sets, if one
interchanges the roles of unions and intersections.
A set E is bounded if it is contained in some ball of finite radius.
A bounded set is compact if it is also closed. Compact sets enjoy the
Heine-Borel covering property:
S
• Assume E is compact, E ⊂ α Oα , and each Oα is open. Then
there are finitely many of the open sets, Oα1 , Oα2 , . . . , OαN , such
SN
that E ⊂ j=1 Oαj .
In words, any covering of a compact set by a collection of open sets

contains a finite subcovering.
A point x ∈ Rd is a limit point of the set E if for every r > 0, the ball
Br (x) contains points of E. This means that there are points in E which
are arbitrarily close to x. An isolated point of E is a point x ∈ E such
that there exists an r > 0 where Br (x) ∩ E is equal to {x}.
A point x ∈ E is an interior point of E if there exists r > 0 such
that Br (x) ⊂ E. The set of all interior points of E is called the interior
of E. Also, the closure E of the E consists of the union of E and all
its limit points. The boundary of a set E, denoted by ∂E, is the set of
points which are in the closure of E but not in the interior of E.
Note that the closure of a set is a closed set; every point in E is a
limit point of E; and a set is closed if and only if it contains all its limit
points. Finally, a closed set E is perfect if E does not have any isolated
points.
Rectangles and cubes

A (closed) rectangle R in Rd is given by the product of d one-dimensional
closed and bounded intervals
R = [a1 , b1 ] × [a2 , b2 ] × · · · × [ad , bd ],
where aj ≤ bj are real numbers, j = 1, 2, . . . , d. In other words, we have
R = {(x1 , . . . , xd ) ∈ Rd : aj ≤ xj ≤ bj for all j = 1, 2, . . . , d}.
We remark that in our definition, a rectangle is closed and has sides

parallel to the coordinate axis. In R, the rectangles are precisely the
closed and bounded intervals, while in R2 they are the usual four-sided
rectangles. In R3 they are the closed parallelepipeds.
We say that the lengths of the sides of the rectangle R are b1 −
a1 , . . . , bd − ad . The volume of the rectangle R is denoted by |R|, and
R2
R3
Figure 1. Rectangles in Rd , d = 1, 2, 3
is defined to be
|R| = (b1 − a1 ) · · · (bd − ad ).
Of course, when d = 1 the “volume” equals length, and when d = 2 it

equals area.
An open rectangle is the product of open intervals, and the interior of
the rectangle R is then
(a1 , b1 ) × (a2 , b2 ) × · · · × (ad , bd ).
Also, a cube is a rectangle for which b1 − a1 = b2 − a2 = · · · = bd − ad .

So if Q ⊂ Rd is a cube of common side length `, then |Q| = `d .
A union of rectangles is said to be almost disjoint if the interiors of
the rectangles are disjoint.
In this chapter, coverings by rectangles and cubes play a major role,
so we isolate here two important lemmas.
Lemma 1.1 If a rectangleSis the almost disjoint union of finitely many

N
other rectangles, say R = k=1 Rk , then
N
X
|R| = |Rk |.
k=1
1. Preliminaries 5
Proof. We consider the grid formed by extending indefinitely the

sides of all rectangles R1 , . . . , RN . This construction yields finitely many
rectangles R̃1 , . . . , R̃M , and a partition J1 , . . . , JN of the integers between
1 and M , such that the unions
M
[ [
R= R̃j and Rk = R̃j , for k = 1, . . . , N
j=1 j∈Jk
are almost disjoint (see the illustration in Figure 2).
RN R̃M
R1
R2 R̃1 R̃2
Figure 2. The grid formed by the rectangles Rk
PM
For the rectangle R, for example, we see that |R| = j=1 |R̃j |, since
the grid actually partitions the sides of R and each R̃j consists of taking
products of the intervals in these partitions. Thus when adding the
volumes of the R̃j we are summing the corresponding products of lengths
of the intervals that arise. Since this also holds for the other rectangles
R1 , . . . , RN , we conclude that
M
X N X
X N
X
|R| = |R̃j | = |R̃j | = |Rk |.
j=1 k=1 j∈Jk k=1
A slight modification of this argument then yields the following:

SN
Lemma 1.2 If R, R1 , . . . , RN are rectangles, and R ⊂ k=1 Rk , then
N
X
|R| ≤ |Rk |.
k=1
The main idea consists of taking the grid formed by extending all sides
of the rectangles R, R1 , . . . , RN , and noting that the sets corresponding
to the Jk (in the above proof) need not be disjoint any more.
We now proceed to give a description of the structure of open sets in
terms of cubes. We begin with the case of R.
Theorem 1.3 Every open subset O of R can be writen uniquely as a

countable union of disjoint open intervals.
Proof. For each x ∈ O, let Ix denote the largest open interval contain-
ing x and contained in O. More precisely, since O is open, x is contained
in some small (non-trivial) interval, and therefore if
ax = inf{a < x : (a, x) ⊂ O} and bx = sup{b > x : (x, b) ⊂ O}
we must have ax < x < bx (with possibly infinite values for ax and bx ).
If we now let Ix = (ax , bx ), then by construction we have x ∈ Ix as well
as Ix ⊂ O. Hence
[
O= Ix .
x∈O
Now suppose that two intervals Ix and Iy intersect. Then their union
(which is also an open interval) is contained in O and contains x. Since
Ix is maximal, we must have (Ix ∪ Iy ) ⊂ Ix , and similarly (Ix ∪ Iy ) ⊂ Iy .
This can happen only if Ix = Iy ; therefore, any two distinct intervals in
the collection I = {Ix }x∈O must be disjoint. The proof will be complete
once we have shown that there are only countably many distinct intervals
in the collection I. This, however, is easy to see, since every open interval
Ix contains a rational number. Since different intervals are disjoint, they
must contain distinct rationals, and therefore I is countable, as desired.
S∞
Naturally, if O is open and O = j=1 Ij , where the Ij ’s are disjoint
P∞
open intervals, the measure of O ought to be j=1 |Ij |. Since this rep-
resentation is unique, we could take this as a definition of measure; we
would then note that whenever O1 and O2 are open and disjoint, the mea-
sure of their union is the sum of their measures. Although this provides
1. Preliminaries 7
a natural notion of measure for an open set, it is not immediately clear

how to generalize it to other sets in R. Moreover, a similar approach in
higher dimensions already encounters complications even when defining
measures of open sets, since in this context the direct analogue of The-
orem 1.3 is not valid (see Exercise 12). There is, however, a substitute
result.
Theorem 1.4 Every open subset O of Rd , d ≥ 1, can be written as a

countable union of almost disjoint closed cubes.
Proof. We must construct a countable collection

S Q of closed cubes
whose interiors are disjoint, and so that O = Q∈Q Q.
As a first step, consider the grid in Rd formed by taking all closed cubes
of side length 1 whose vertices have integer coordinates. In other words,
we consider the natural grid of lines parallel to the axes, that is, the grid
generated by the lattice Zd . We shall also use the grids formed by cubes
of side length 2−N obtained by successively bisecting the original grid.
We either accept or reject cubes in the initial grid as part of Q accord-
ing to the following rule: if Q is entirely contained in O then we accept
Q; if Q intersects both O and Oc then we tentatively accept it; and if Q
is entirely contained in Oc then we reject it.
As a second step, we bisect the tentatively accepted cubes into 2d cubes
with side length 1/2. We then repeat our procedure, by accepting the
smaller cubes if they are completely contained in O, tentatively accepting
them if they intersect both O and Oc , and rejecting them if they are
contained in Oc . Figure 3 illustrates these steps for an open set in R2 .
O O
Step 1 Step 2
Figure 3. Decomposition of O into almost disjoint cubes

This procedure is then repeated indefinitely, and (by construction)

the resulting collection Q of all accepted cubes is countable and consists
of almost disjoint cubes. To see why their union is all of O, we note
that given x ∈ O there exists a cube of side length 2−N (obtained from
successive bisections of the original grid) that contains x and that is
entirely contained in O. Either this cube has been accepted, or it is
contained in a cube that has been previously accepted. This shows that
the union of all cubes in Q covers O.
S∞
Once again, if O = j=1 Rj where the rectangles Rj are almost dis-
P∞
joint, it is reasonable to assign to O the measure j=1 |Rj |. This is
natural since the volume of the boundary of each rectangle should be 0,
and the overlap of the rectangles should not contribute to the volume
of O. We note, however, that the above decomposition into cubes is
not unique, and it is not immediate that the sum is independent of this
decomposition. So in Rd , with d ≥ 2, the notion of volume or area, even
for open sets, is more subtle.
The general theory developed in the next section actually yields a
notion of volume that is consistent with the decompositions of open sets
of the previous two theorems, and applies to all dimensions. Before we
come to that, we discuss an important example in R.
The Cantor set

The Cantor set plays a prominent role in set theory and in analysis in
general. It and its variants provide a rich source of enlightening examples.
We begin with the closed unit interval C0 = [0, 1] and let C1 denote
the set obtained from deleting the middle third open interval from [0, 1],
that is,
C1 = [0, 1/3] ∪ [2/3, 1].
Next, we repeat this procedure for each sub-interval of C1 ; that is, we

delete the middle third open interval. At the second stage we get
C2 = [0, 1/9] ∪ [2/9, 1/3] ∪ [2/3, 7/9] ∪ [8/9, 1].
We repeat this process for each sub-interval of C2 , and so on (Figure 4).

This procedure yields a sequence Ck , k = 0, 1, 2, . . . of compact sets
with
C0 ⊃ C1 ⊃ C2 ⊃ · · · ⊃ Ck ⊃ Ck+1 ⊃ · · · .
1. Preliminaries 9
C0
0 1
C1
0 1/3 2/3 1
C2
0 1/9 2/9 1/3 2/3

7/9 8/9
1
C3
Figure 4. Construction of the Cantor set
The Cantor set C is by definition the intersection of all Ck ’s:

∞
\
C= Ck .
k=0
The set C is not empty, since all end-points of the intervals in Ck (all k)
belong to C.
Despite its simple construction, the Cantor set enjoys many interest-
ing topological and analytical properties. For instance, C is closed and
bounded, hence compact. Also, C is totally disconnected: given any
x, y ∈ C there exists z ∈/ C that lies between x and y. Finally, C is per-
fect: it has no isolated points (Exercise 1).
Next, we turn our attention to the question of determining the “size”
of C. This is a delicate problem, one that may be approached from
different angles depending on the notion of size we adopt. For instance,
in terms of cardinality the Cantor set is rather large: it is not countable.
Since it can be mapped to the interval [0, 1], the Cantor set has the
cardinality of the continuum (Exercise 2).
However, from the point of view of “length” the size of C is small.
Roughly speaking, the Cantor set has length zero, and this follows from
the following intuitive argument: the set C is covered by sets Ck whose
lengths go to zero. Indeed, Ck is a disjoint union of 2k intervals of length
3−k , making the total length of Ck equal to (2/3)k . But C ⊂ Ck for all
k, and (2/3)k → 0 as k tends to infinity. We shall define a notion of
measure and make this argument precise in the next section.
2 The exterior measure

The notion of exterior measure is the first of two important concepts
needed to develop a theory of measure. We begin with the definition and
basic properties of exterior measure. Loosely speaking, the exterior mea-
sure m∗ assigns to any subset of Rd a first notion of size; various examples
show that this notion coincides with our earlier intuition. However, the
exterior measure lacks the desirable property of additivity when taking
the union of disjoint sets. We remedy this problem in the next section,
where we discuss in detail the other key concept of measure theory, the
notion of measurable sets.
The exterior measure, as the name indicates, attempts to describe
the volume of a set E by approximating it from the outside. The set
E is covered by cubes, and if the covering gets finer, with fewer cubes
overlapping, the volume of E should be close to the sum of the volumes
of the cubes.
The precise definition is as follows: if E is any subset of Rd , the
exterior measure1 of E is
∞
X
(1) m∗ (E) = inf |Qj |,
j=1
S∞
where the infimum is taken over all countable coverings E ⊂ j=1 Qj by
closed cubes. The exterior measure is always non-negative but could be
infinite, so that in general we have 0 ≤ m∗ (E) ≤ ∞, and therefore takes
values in the extended positive numbers.
We make some preliminary remarks about the definition of the exterior
measure given by (1).
(i) It is important to note that it would not suffice to allow finite sums
in the definition of m∗ (E). The quantity that would be obtained if one
considered only coverings of E by finite unions of cubes is in general
larger than m∗ (E). (See Exercise 14.)
(ii) One can, however, replace the coverings by cubes, with coverings
by rectangles; or with coverings by balls. That the former alternative
1 Some authors use the term outer measure instead of exterior measure.
2. The exterior measure 11
yields the same exterior measure is quite direct. (See Exercise 15.) The
equivalence with the latter is more subtle. (See Exercise 26 in Chapter 3.)
We begin our investigation of this new notion by providing examples
of sets whose exterior measures can be calculated, and we check that
the latter matches our intuitive idea of volume (length in one dimension,
area in two dimensions, etc.)
Example 1. The exterior measure of a point is zero. This is clear once
we observe that a point is a cube with volume zero, and which covers
itself. Of course the exterior measure of the empty set is also zero.
Example 2. The exterior measure of a closed cube is equal to its volume.

Indeed, suppose Q is a closed cube in Rd . Since Q covers itself, we must
have m∗ (Q) ≤ |Q|. Therefore, it suffices to S
prove the reverse inequality.
∞
We consider an arbitrary covering Q ⊂ j=1 Qj by cubes, and note
that it suffices to prove that
∞
X
(2) |Q| ≤ |Qj |.
j=1
For a fixed ² > 0 we choose for each j an open cube Sj which

S∞contains Qj ,
and such that |Sj | ≤ (1 + ²)|Qj |. From the open covering j=1 Sj of the
compact set Q, we may select a finite subcovering S which, after possibly
N
renumbering the rectangles, we may write as Q ⊂ j=1 Sj . Taking the
closure of the cubes Sj , we may apply Lemma 1.2 to conclude that |Q| ≤
PN
j=1 |Sj |. Consequently,
N
X ∞
X
|Q| ≤ (1 + ²) |Qj | ≤ (1 + ²) |Qj |.
j=1 j=1
Since ² is arbitrary, we find that the inequality (2) holds; thus |Q| ≤
m∗ (Q), as desired.
Example 3. If Q is an open cube, the result m∗ (Q) = |Q| still holds.

Since Q is covered by its closure Q, and |Q| = |Q|, we immediately see
that m∗ (Q) ≤ |Q|. To prove the reverse inequality, we note that if Q0 is
a closed cube contained in Q, then m∗ (Q0 ) ≤ m∗ (Q), since any covering
of Q by a countable number of closed cubes is also a covering of Q0 (see
Observation 1 below). Hence |Q0 | ≤ m∗ (Q), and since we can choose Q0
with a volume as close as we wish to |Q|, we must have |Q| ≤ m∗ (Q).
Example 4. The exterior measure of a rectangle R is equal to its volume.

Indeed, arguing as in Example 2, we see that |R| ≤ m∗ (R). To obtain the
reverse inequality, consider a grid in Rd formed by cubes of side length
1/k. Then, if Q consists of the (finite) collection of all cubes entirely
contained in R, and Q0 the (finite) collectionS of all cubes that intersect
the complement of R, we first note that R ⊂ Q∈(Q∪Q0 ) Q. Also, a simple
argument yields
X
|Q| ≤ |R|.
Q∈Q
Moreover, there
P are O(k d−1 ) cubes2 in Q0 , and these cubes have volume
k −d , so that Q∈Q0 |Q| = O(1/k). Hence
X
|Q| ≤ |R| + O(1/k),
Q∈(Q∪Q0 )
and letting k tend to infinity yields m∗ (R) ≤ |R|, as desired.
Example 5. The exterior measure of Rd is infinite. This follows from

the fact that any covering of Rd is also a covering of any cube Q ⊂ Rd ,
hence |Q| ≤ m∗ (Rd ). Since Q can have arbitrarily large volume, we must
have m∗ (Rd ) = ∞.
Example 6. The Cantor set C has exterior measure 0. From the con-
struction of C, we know that C ⊂ Ck , where each Ck is a disjoint union
of 2k closed intervals, each of length 3−k . Consequently, m∗ (C) ≤ (2/3)k
for all k, hence m∗ (C) = 0.
Properties of the exterior measure

The previous examples and comments provide some intuition underlying
the definition of exterior measure. Here, we turn to the further study of
m∗ and prove five properties of exterior measure that are needed in what
follows.
First, we record the following remark that is immediate from the def-
inition of m∗ :
2 We remind the reader of the notation f (x) = O(g(x)), which means that |f (x)| ≤
C|g(x)| for some constant C and all x in a given range. In this particular example, there
are fewer than Ckd−1 cubes in question, as k → ∞.
S∞
• For every ² > 0, there exists a covering E ⊂ j=1 Qj with
∞
X
m∗ (Qj ) ≤ m∗ (E) + ².
j=1
The relevant properties of exterior measure are listed in a series of

observations.
Observation 1 (Monotonicity) If E1 ⊂ E2 , then m∗ (E1 ) ≤ m∗ (E2 ).
This follows once we observe that any covering of E2 by a countable

collection of cubes is also a covering of E1 .
In particular, monotonicity implies that every bounded subset of Rd
has finite exterior measure.
S∞
Observation 2 (Countable sub-additivity) If E = Ej , then
P∞ j=1
m∗ (E) ≤ j=1 m∗ (Ej ).
First, we may assume that each m∗ (Ej ) < ∞, for otherwise the in-
equality clearly holds. For any ² > 0, the
S∞definition of the exterior mea-
sure yields for each j a covering Ej ⊂ k=1 Qk,j by closed cubes with
∞
X ²
|Qk,j | ≤ m∗ (Ej ) + .
2j
k=1
S∞
Then, E ⊂ j,k=1 Qk,j is a covering of E by closed cubes, and therefore
X X ∞
∞ X
m∗ (E) ≤ |Qk,j | = |Qk,j |
j,k j=1 k=1
∞ ³
X ²´
≤ m∗ (Ej ) +
2j
j=1
∞
X
= m∗ (Ej ) + ².
j=1
Since this holds true for every ² > 0, the second observation is proved.
Observation 3 If E ⊂ Rd , then m∗ (E) = inf m∗ (O), where the infi-

mum is taken over all open sets O containing E.
By monotonicity, it is clear that the inequality m∗ (E) ≤ inf m∗ (O)

holds. ForS the reverse inequality, let ² > 0 and choose cubes Qj such
∞
that E ⊂ j=1 Qj , with
∞
X ²
|Qj | ≤ m∗ (E) + .
2
j=1
Let Q0j denote an open

S cube containing Qj , and such that |Q0j | ≤ |Qj | +
∞
²/2j+1 . Then O = j=1 Q0j is open, and by Observation 2
∞
X ∞
X
m∗ (O) ≤ m∗ (Q0j ) = |Q0j |
j=1 j=1
∞ ³
X ² ´
≤ |Qj | +
2j+1
j=1
X∞
²
≤ |Qj | +
2
j=1
≤ m∗ (E) + ².
Hence inf m∗ (O) ≤ m∗ (E), as was to be shown.
Observation 4 If E = E1 ∪ E2 , and d(E1 , E2 ) > 0, then
m∗ (E) = m∗ (E1 ) + m∗ (E2 ).
By Observation 2, we already know that m∗ (E) ≤ m∗ (E1 ) + m∗ (E2 ),

so it suffices to prove the reverse inequality. To this end, we first
S∞ select δ
such that d(E1 , E2 ) > δ > 0. Next, we choose a covering E ⊂ j=1 Qj by
P∞
closed cubes, with j=1 |Qj | ≤ m∗ (E) + ². We may, after subdividing
the cubes Qj , assume that each Qj has a diameter less than δ. In this
case, each Qj can intersect at most one of the two sets E1 or E2 . If we
denote by J1 and J2 the sets of those indices j for which Qj intersects
E1 and E2 , respectively, then J1 ∩ J2 is empty, and we have
∞
[ ∞
[
E1 ⊂ Qj as well as E2 ⊂ Qj .
j∈J1 j∈J2
Therefore,
X X
m∗ (E1 ) + m∗ (E2 ) ≤ |Qj | + |Qj |
j∈J1 j∈J2
X∞
≤ |Qj |
j=1
≤ m∗ (E) + ².
Since ² is arbitrary, the proof of Observation 4 is complete.
Observation
S∞ 5 If a set E is the countable union of almost disjoint cubes
E = j=1 Qj , then
∞
X
m∗ (E) = |Qj |.
j=1
Let Q̃j denote a cube strictly contained in Qj such that |Qj | ≤ |Q̃j | +
²/2j , where ² is arbitrary but fixed. Then, for every N , the cubes
Q̃1 , Q̃2 , . . . , Q̃N are disjoint, hence at a finite distance from one another,
and repeated applications of Observation 4 imply
Ã N
! N N
[ X X ¡ ¢
m∗ Q̃j = |Q̃j | ≥ |Qj | − ²/2j .
j=1 j=1 j=1
SN
Since j=1 Q̃j ⊂ E, we conclude that for every integer N ,
N
X
m∗ (E) ≥ |Qj | − ².
j=1
P∞
In the limit as N tends to infinity we deduce j=1 |Qj | ≤ m∗ (E) + ²
P∞
for every ² > 0, hence j=1 |Qj | ≤ m∗ (E). Therefore, combined with
Observation 2, our result proves that we have equality.
This last property shows that if a set can be decomposed into almost
disjoint cubes, its exterior measure equals the sum of the volumes of the
cubes. In particular, by Theorem 1.4 we see that the exterior measure of
an open set equals the sum of the volumes of the cubes in a decomposi-
tion, and this coincides with our initial guess. Moreover, this also yields
a proof that the sum is independent of the decomposition.
One can see from this that the volumes of simple sets that are cal-
culated by elementary calculus agree with their exterior measure. This
assertion can be proved most easily once we have developed the requisite
tools in integration theory. (See Chapter 2.) In particular, we can then
verify that the exterior measure of a ball (either open or closed) equals
its volume.
Despite observations 4 and 5, one cannot conclude in general that if
E1 ∪ E2 is a disjoint union of subsets of Rd , then
(3) m∗ (E1 ∪ E2 ) = m∗ (E1 ) + m∗ (E2 ).
In fact (3) holds when the sets in question are not highly irregular or
“pathological” but are measurable in the sense described below.
3 Measurable sets and the Lebesgue measure

The notion of measurability isolates a collection of subsets in Rd for
which the exterior measure satisfies all our desired properties, including
additivity (and in fact countable additivity) for disjoint unions of sets.
There are a number of different ways of defining measurability, but
these all turn out to be equivalent. Probably the simplest and most
intuitive is the following: A subset E of Rd is Lebesgue measurable,
or simply measurable, if for any ² > 0 there exists an open set O with
E ⊂ O and
m∗ (O − E) ≤ ².
This should be compared to Observation 3, which holds for all sets E.

If E is measurable, we define its Lebesgue measure (or measure)
m(E) by
m(E) = m∗ (E).
Clearly, the Lebesgue measure inherits all the features contained in Ob-
servations 1 - 5 of the exterior measure.
Immediately from the definition, we find:
Property 1 Every open set in Rd is measurable.
Our immediate goal now is to gather various further properties of

measurable sets. In particular, we shall prove that the collection of
measurable sets behave well under the various operations of set theory:
countable unions, countable intersections, and complements.
3. Measurable sets and the Lebesgue measure 17
Property 2 If m∗ (E) = 0, then E is measurable. In particular, if F is

a subset of a set of exterior measure 0, then F is measurable.
By Observation 3 of the exterior measure, for every ² > 0 there ex-

ists an open set O with E ⊂ O and m∗ (O) ≤ ². Since (O − E) ⊂ O,
monotonicity implies m∗ (O − E) ≤ ², as desired.
As a consequence of this property, we deduce that the Cantor set C in
Example 6 is measurable and has measure 0.
Property 3 A countable union of measurable sets is measurable.

S∞
Suppose E = j=1 Ej , where each Ej is measurable. Given ² > 0, we
may choose for each j an open setS Oj with Ej ⊂ Oj and
∞
m∗ (Oj − Ej ) ≤ ²/2j . Then the union O = j=1 Oj is open, E ⊂ O, and
S∞
(O − E) ⊂ j=1 (Oj − Ej ), so monotonicity and sub-additivity of the
exterior measure imply
∞
X
m∗ (O − E) ≤ m∗ (Oj − Ej ) ≤ ².
j=1
Property 4 Closed sets are measurable.
First, we observe that it suffices to prove that compact sets are mea-
surable. Indeed,
S∞any closed set F can be written as the union of compact
sets, say F = k=1 F ∩ Bk , where Bk denotes the closed ball of radius k
centered at the origin; then Property 3 applies.
So, suppose F is compact (so that in particular m∗ (F ) < ∞), and let
² > 0. By Observation 3 we can select an open set O with F ⊂ O and
m∗ (O) ≤ m∗ (F ) + ². Since F is closed, the difference O − F is open,
and by Theorem 1.4 we may write this difference as a countable union
of almost disjoint cubes
∞
[
O−F = Qj .
j=1
SN
For a fixed N , the finite union K = j=1 Qj is compact; therefore
d(K, F ) > 0 (we isolate this little fact in a lemma below). Since (K ∪
F ) ⊂ O, Observations 1, 4, and 5 of the exterior measure imply
m∗ (O) ≥ m∗ (F ) + m∗ (K)
N
X
= m∗ (F ) + m∗ (Qj ).
j=1
PN
Hence j=1 m∗ (Qj ) ≤ m∗ (O) − m∗ (F ) ≤ ², and this also holds in the
limit as N tends to infinity. Invoking the sub-additivity property of the
exterior measure finally yields
∞
X
m∗ (O − F ) ≤ m∗ (Qj ) ≤ ²,
j=1
as desired.
We digress briefly to complete the above argument by proving the
following.
Lemma 3.1 If F is closed, K is compact, and these sets are disjoint,
then d(F, K) > 0.
Proof. Since F is closed, S for each point x ∈ K, there exists δx > 0 so
that d(x, F ) > 3δx . Since x∈K B2δx (x) covers K, and K is compact, we
SN
may find a subcover, which we denote by j=1 B2δj (xj ). If we let δ =
min(δ1 , . . . , δN ), then we must have d(K, F ) ≥ δ > 0. Indeed, if x ∈ K
and y ∈ F , then for some j we have |xj − x| ≤ 2δj , and by construction
|y − xj | ≥ 3δj . Therefore
|y − x| ≥ |y − xj | − |xj − x| ≥ 3δj − 2δj ≥ δ,
and the lemma is proved.
Property 5 The complement of a measurable set is measurable.

If E is measurable, then for every positive integer n we may choose an
c
open set On with E ⊂ On and m∗ (On − E) ≤ 1/n. The complement S∞ Onc
is closed, hence measurable, which implies that the union S = n=1 On
is also measurable by Property 3. Now we simply note that S ⊂ E c , and
(E c − S) ⊂ (On − E),
such that m∗ (E c − S) ≤ 1/n for all n. Therefore, m∗ (E c − S) = 0, and

E c − S is measurable by Property 2. Therefore E c is measurable since
it is the union of two measurable sets, namely S and (E c − S).
Property 6 A countable intersection of measurable sets is measurable.
This follows from Properties 3 and 5, since
∞
Ã∞ !c
\ [
c
Ej = Ej .
j=1 j=1
In conclusion, we find that the family of measurable sets is closed under

the familiar operations of set theory. We point out that we have shown
more than simply closure with respect to finite unions and intersections:
we have proved that the collection of measurable sets is closed under
countable unions and intersections. This passage from finite operations
to infinite ones is crucial in the context of analysis. We emphasize, how-
ever, that the operations of uncountable unions or intersections are not
permissible when dealing with measurable sets!
Theorem
S∞ 3.2 If E1 , E2 , . . ., are disjoint measurable sets, and E =
j=1 Ej , then
∞
X
m(E) = m(Ej ).
j=1
Proof. First, we assume further that each Ej is bounded. Then, for

each j, by applying the definition of measurability to Ejc , we can choose
a closed subset Fj of Ej with m∗ (Ej − Fj ) ≤ ²/2j . For each³S fixed´N ,
N
the sets F1 , . . . , FN are compact and disjoint, so that m j=1 Fj =
PN SN
j=1 m(Fj ). Since j=1 Fj ⊂ E, we must have
N
X N
X
m(E) ≥ m(Fj ) ≥ m(Ej ) − ².
j=1 j=1
Letting N tend to infinity, since ² was arbitrary we find that

∞
X
m(E) ≥ m(Ej ).
j=1
Since the reverse inequality always holds (by sub-additivity in Observa-

tion 2), this concludes the proof when each Ej is bounded.
∞
In the general case, we select any sequence of cubes {QS k }k=1 that
d ∞
increases to R , in the sense that Qk ⊂ Qk+1 for all k ≥ 1 and k=1 Qk =
Rd . We then let S1 = Q1 and Sk = Qk − Qk−1 for k ≥ 2. If we define
measurable sets by Ej,k = Ej ∩ Sk , then
[
E= Ej,k .
j,k
The
S∞ union above is disjoint and every Ej,k is bounded. Moreover Ej =
k=1 Ej,k , and this union is also disjoint. Putting these facts together,
and using what has already been proved, we obtain

X XX X
m(E) = m(Ej,k ) = m(Ej,k ) = m(Ej ),
j,k j k j
as claimed.
With this, the countable additivity of the Lebesgue measure on mea-
surable sets has been established. This result provides the necessary
connection between the following:
• our primitive notion of volume given by the exterior measure,
• the more refined idea of measurable sets, and
• the countably infinite operations allowed on these sets.
We make two definitions to state succinctly some further consequences.
If E1 , E2 , . . . is a countable collection of subsets of S Rd that increases
∞
to E in the sense that Ek ⊂ Ek+1 for all k, and E = k=1 Ek , then we
write Ek % E.
Similarly, if ET1 , E2 , . . . decreases to E in the sense that Ek ⊃ Ek+1 for
∞
all k, and E = k=1 Ek , we write Ek & E.
Corollary 3.3 Suppose E1 , E2 , . . . are measurable subsets of Rd .

(i) If Ek % E, then m(E) = limN →∞ m(EN ).
(ii) If Ek & E and m(Ek ) < ∞ for some k, then
m(E) = lim m(EN ).

N →∞
Proof. For the first part, let G1 = E1 , G2 = E2 − E1 , and in gen-

eral Gk = Ek − Ek−1 for k ≥ 2.
S∞ By their construction, the sets Gk are
measurable, disjoint, and E = k=1 Gk . Hence
∞ N
Ã N
!
X X [
m(E) = m(Gk ) = lim m(Gk ) = lim m Gk ,
N →∞ N →∞
k=1 k=1 k=1
SN
and since k=1 Gk = EN we get the desired limit.
For the second part, we may clearly assume that m(E1 ) < ∞. Let
Gk = Ek − Ek+1 for each k, so that
∞
[
E1 = E ∪ Gk
k=1
is a disjoint union of measurable sets. As a result, we find that
N
X −1
m(E1 ) = m(E) + lim (m(Ek ) − m(Ek+1 ))
N →∞
k=1
= m(E) + m(E1 ) − lim m(EN ).
N →∞
Hence, since m(E1 ) < ∞, we see that m(E) = limN →∞ m(EN ), and the
proof is complete.
The reader should note that the second conclusion may fail without
the assumption that m(Ek ) < ∞ for some k. This is shown by the simple
example when En = (n, ∞) ⊂ R, for all n.
What follows provides an important geometric and analytic insight
into the nature of measurable sets, in terms of their relation to open and
closed sets. Its thrust is that, in effect, an arbitrary measurable set can
be well approximated by the open sets that contain it, and alternatively,
by the closed sets it contains.
Theorem 3.4 Suppose E is a measurable subset of Rd . Then, for every

² > 0:
(i) There exists an open set O with E ⊂ O and m(O − E) ≤ ².
(ii) There exists a closed set F with F ⊂ E and m(E − F ) ≤ ².
(iii) If m(E) is finite, there exists a compact set K with K ⊂ E and

m(E − K) ≤ ².
SN
(iv) If m(E) is finite, there exists a finite union F = j=1 Qj of closed
cubes such that
m(E4F ) ≤ ².
The notation E4F stands for the symmetric difference between the
sets E and F , defined by E4F = (E − F ) ∪ (F − E), which consists of
those points that belong to only one of the two sets E or F .
Proof. Part (i) is just the definition of measurability. For the second
part, we know that E c is measurable, so there exists an open set O with
E c ⊂ O and m(O − E c ) ≤ ². If we let F = Oc , then F is closed, F ⊂ E,
and E − F = O − E c . Hence m(E − F ) ≤ ² as desired.
For (iii), we first pick a closed set F so that F ⊂ E and m(E − F ) ≤
²/2. For each n, we let Bn denote the ball centered at the origin of radius
n, and define compact sets Kn = F ∩ Bn . Then E − Kn is a sequence

of measurable sets that decreases to E − F , and since m(E) < ∞, we
conclude that for all large n one has m(E − Kn ) ≤ ².
For the last part, choose a family of closed cubes {Qj }∞
j=1 so that
∞
[ ∞
X
E⊂ Qj and |Qj | ≤ m(E) + ²/2.
j=1 j=1
Since m(E) < ∞, the series converges and there exists N > 0 such that
P∞ SN
j=N +1 |Qj | < ²/2. If F = j=1 Qj , then
m(E4F ) = m(E − F ) + m(F − E)

Ã ∞ ! Ã∞ !
[ [
≤m Qj + m Qj − E
j=N +1 j=1
∞
X ∞
X
≤ |Qj | + |Qj | − m(E)
j=N +1 j=1
≤ ².
Invariance properties of Lebesgue measure

A crucial property of Lebesgue measure in Rd is its translation-invariance,
which can be stated as follows: if E is a measurable set and h ∈ Rd , then
the set Eh = E + h = {x + h : x ∈ E} is also measurable, and m(E +
h) = m(E). With the observation that this holds for the special case
when E is a cube, one passes to the exterior measure of arbitrary sets
E, and sees from the definition of m∗ given in Section 2 that m∗ (Eh ) =
m∗ (E). To prove the measurability of Eh under the assumption that E
is measurable, we note that if O is open, O ⊃ E, and m∗ (O − E) < ²,
then Oh is open, Oh ⊃ Eh , and m∗ (Oh − Eh ) < ².
In the same way one can prove the relative dilation-invariance of
Lebesgue measure. Suppose δ > 0, and denote by δE the set {δx :
x ∈ E}. We can then assert that δE is measurable whenever E is,
and m(δE) = δ d m(E). One can also easily see that Lebesgue mea-
sure is reflection-invariant. That is, whenever E is measurable, so is
−E = {−x : x ∈ E} and m(−E) = m(E).
Other invariance properties of Lebesgue measure are in Exercise 7
and 8, and Problem 4 of Chapter 2.
σ-algebras and Borel sets

A σ-algebra of sets is a collection of subsets of Rd that is closed under
countable unions, countable intersections, and complements.
The collection of all subsets of Rd is of course a σ-algebra. A more
interesting and relevant example consists of all measurable sets in Rd ,
which we have just shown also forms a σ-algebra.
Another σ-algebra, which plays a vital role in analysis, is the Borel
σ-algebra in Rd , denoted by BRd , which by definition is the smallest σ-
algebra that contains all open sets. Elements of this σ-algebra are called
Borel sets.
The definition of the Borel σ-algebra will be meaningful once we have
defined the term “smallest,” and shown that such a σ-algebra exists and
is unique. The term “smallest” means that if S is any σ-algebra that
contains all open sets in Rd , then necessarily BRd ⊂ S. Since we observe
that any intersection (not necessarily countable) of σ-algebras is again a
σ-algebra, we may define BRd as the intersection of all σ-algebras that
contain the open sets. This shows the existence and uniqueness of the
Borel σ-algebra.
Since open sets are measurable, we conclude that the Borel σ-algebra
is contained in the σ-algebra of measurable sets. Naturally, we may ask
if this inclusion is strict: do there exist Lebesgue measurable sets which
are not Borel sets? The answer is “yes.” (See Exercise 35.)
From the point of view of the Borel sets, the Lebesgue sets arise as
the completion of the σ-algebra of Borel sets, that is, by adjoining all
subsets of Borel sets of measure zero. This is an immediate consequence
of Corollary 3.5 below.
Starting with the open and closed sets, which are the simplest Borel
sets, one could try to list the Borel sets in order of their complexity. Next
in order would come countable intersections of open sets; such sets are
called Gδ sets. Alternatively, one could consider their complements, the
countable union of closed sets, called the Fσ sets.3
Corollary 3.5 A subset E of Rd is measurable
(i) if and only if E differs from a Gδ by a set of measure zero,

(ii) if and only if E differs from an Fσ by a set of measure zero.
Proof. Clearly E is measurable whenever it satisfies either (i) or (ii),

since the Fσ , Gδ , and sets of measure zero are measurable.
3 The terminology G comes from German “Gebiete” and “Durschnitt”; F comes from
δ σ
French “fermé” and “somme.”
Conversely, if E is measurable, then for each integer n ≥ 1 we may

select an open
T∞ set On that contains E, and such that m(On − E) ≤ 1/n.
Then S = n=1 On is a Gδ that contains E, and (S − E) ⊂ (On − E)
for all n. Therefore m(S − E) ≤ 1/n for all n; hence S − E has exterior
measure zero, and is therefore measurable.
For the second implication, we simply apply part (ii) of Theorem 3.4
with ² = 1/n, and take the union of the resulting closed sets.
Construction of a non-measurable set

Are all subsets of Rd measurable? In this section, we answer this question
when d = 1 by constructing a subset of R which is not measurable.4
This justifies the conclusion that a satisfactory theory of measure cannot
encompass all subsets of R.
The construction of a non-measurable set N uses the axiom of choice,
and rests on a simple equivalence relation among real numbers in [0, 1].
We write x ∼ y whenever x − y is rational, and note that this is an
equivalence relation since the following properties hold:
• x ∼ x for every x ∈ [0, 1]
• if x ∼ y, then y ∼ x
• if x ∼ y and y ∼ z, then x ∼ z.
Two equivalence classes either are disjoint or coincide, and [0, 1] is the
disjoint union of all equivalence classes, which we write as
[
[0, 1] = Eα .
α
Now we construct the set N by choosing exactly one element xα from

each Eα , and setting N = {xα }. This (seemingly obvious) step requires
further comment, which we postpone until after the proof of the following
theorem.
Theorem 3.6 The set N is not measurable.
The proof is by contradiction, so we assume that N is measurable. Let
{rk }∞
k=1 be an enumeration of all the rationals in [−1, 1], and consider
the translates
Nk = N + rk .
4 The existence of such a set in R implies the existence of corresponding non-measurable
subsets of Rd for each d, as a consequence of Proposition 3.4 in the next chapter.

We claim that the sets Nk are disjoint, and

∞
[
(4) [0, 1] ⊂ Nk ⊂ [−1, 2].
k=1
To see why these sets are disjoint, suppose that the intersection
Nk ∩ Nk0 is non-empty. Then there exist rationals rk 6= rk0 and α and
β with xα + rk = xβ + rk0 ; hence
xα − xβ = rk0 − rk .
Consequently α 6= β and xα − xβ is rational; hence xα ∼ xβ , which con-

tradicts the fact that N contains only one representative of each equiv-
alence class.
The second inclusion is straightforward since each Nk is contained in
[−1, 2] by construction. Finally, if x ∈ [0, 1], then x ∼ xα for some α, and
therefore x − xα = rk for some k. Hence x ∈ Nk , and the first inclusion
holds.
Now we may conclude the proof of the theorem. If S N were measurable,
∞
then so would be Nk for all k, and since the union k=1 Nk is disjoint,
the inclusions in (4) yield
∞
X
1≤ m(Nk ) ≤ 3.
k=1
Since Nk is a translate of N , we must have m(Nk ) = m(N ) for all k.

Consequently,
∞
X
1≤ m(N ) ≤ 3.
k=1
This is the desired contradiction, since neither m(N ) = 0 nor m(N ) > 0
is possible.
Axiom of choice
That the construction of the set N is possible is based on the following
general proposition.
• Suppose E is a set and {Eα } is a collection of non-empty subsets

of E. (The indexing set of α’s is not assumed to be countable.)
Then there is a function α 7→ xα (a “choice function”) such that
xα ∈ Eα , for all α.
In this general form this assertion is known as the axiom of choice.

This axiom occurs (at least implicitly) in many proofs in mathematics,
but because of its seeming intuitive self-evidence, its significance was
not at first understood. The initial realization of the importance of
this axiom was in its use to prove a famous assertion of Cantor, the
well-ordering principle. This proposition (sometimes referred to as
“transfinite induction”) can be formulated as follows.
A set E is linearly ordered if there is a binary relation ≤ such that:
(a) x ≤ x for all x ∈ E.
(b) If x, y ∈ E are distinct, then either x ≤ y or y ≤ x (but not both).
(c) If x ≤ y and y ≤ z, then x ≤ z.
We say that a set E can be well-ordered if it can be linearly ordered in

such a way that every non-empty subset A ⊂ E has a smallest element
in that ordering (that is, an element x0 ∈ A such that x0 ≤ x for any
other x ∈ A).
A simple example of a well-ordered set is Z+ , the positive integers with
their usual ordering. The fact that Z+ is well-ordered is an essential part
of the usual (finite) induction principle. More generally, the well-ordering
principle states:
• Any set E can be well-ordered.
It is in fact nearly obvious that the well-ordering principle implies the

axiom of choice: if we well-order E, we can choose xα to be the smallest
element in Eα , and in this way we have constructed the required choice
function. It is also true, but not as easy to show, that the converse impli-
cation holds, namely that the axiom of choice implies the well-ordering
principle. (See Problem 6 for another equivalent formulation of the Ax-
iom of Choice.)
We shall follow the common practice of assuming the axiom of choice
(and hence the validity of the well-ordering principle).5 However, we
should point out that while the axiom of choice seems self-evident the
well-ordering principle leads quickly to some baffling conclusions: one
only needs to spend a little time trying to imagine what a well-ordering
of the reals might look like!
5 It can be proved that in an appropriate formulation of the axioms of set theory, the
axiom of choice is independent of the other axioms; thus we are free to accept its validity.
4. Measurable functions 27
4 Measurable functions
With the notion of measurable sets in hand, we now turn our attention
to the objects that lie at the heart of integration theory: measurable
functions.
The starting point is the notion of a characteristic function of a set
E, which is defined by
½
1 if x ∈ E,
χE (x) =
0 if x ∈
/ E.
The next step is to pass to the functions that are the building blocks of
integration theory. For the Riemann integral it is in effect the class of
step functions, with each given as a finite sum
N
X
(5) f= ak χRk ,
k=1
where each Rk is a rectangle, and the ak are constants.

However, for the Lebesgue integral we need a more general notion, as
we shall see in the next chapter. A simple function is a finite sum
N
X
(6) f= ak χEk
k=1
where each Ek is a measurable set of finite measure, and the ak are

constants.
4.1 Definition and basic properties

We begin by considering only real-valued functions f on Rd , which we
allow to take on the infinite values +∞ and −∞, so that f (x) belongs
to the extended real numbers
−∞ ≤ f (x) ≤ ∞.
We shall say that f is finite-valued if −∞ < f (x) < ∞ for all x. In

the theory that follows, and the many applications of it, we shall almost
always find ourselves in situations where a function takes on infinite
values on at most a set of measure zero.
A function f defined on a measurable subset E of Rd is measurable,

if for all a ∈ R, the set
f −1 ([−∞, a)) = {x ∈ E : f (x) < a}
is measurable. To simplify our notation, we shall often denote the set

{x ∈ E : f (x) < a} simply by {f < a} whenever no confusion is possible.
First, we note that there are many equivalent definitions of measurable

functions. For example, we may require instead that the inverse image of
closed intervals be measurable. Indeed, to prove that f is measurable if
and only if {x : f (x) ≤ a} = {f ≤ a} is measurable for every a, we note
that in one direction, one has
∞
\
{f ≤ a} = {f < a + 1/k},
k=1
and recall that the countable intersection of measurable sets is measur-

able. For the other direction, we observe that
∞
[
{f < a} = {f ≤ a − 1/k}.
k=1
Similarly, f is measurable if and only if {f ≥ a} (or {f > a}) is measur-

able for every a. In the first case this is immediate from our definition
and the fact that {f ≥ a} is the complement of {f < a}, and in the sec-
ond case this follows from what we have just proved and the fact that
{f ≤ a} = {f > a}c . A simple consequence is that −f is measurable
whenever f is measurable.
In the same way, one can show that if f is finite-valued, then it is
measurable if and only if the sets {a < f < b} are measurable for every
a, b ∈ R. Similar conclusions hold for whichever combination of strict or
weak inequalities one chooses. For example, if f is finite-valued, then it
is measurable if and only if {a ≤ f < b} for all a, b ∈ R. By the same
arguments one sees the following:
Property 1 The finite-valued function f is measurable if and only if

f −1 (O) is measurable for every open set O, and if and only if f −1 (F ) is
measurable for every closed set F .
Note that this property also applies to extended-valued functions, if we

make the additional hypothesis that both f −1 (∞) and f −1 (−∞) are
measurable sets.
Property 2 If f is continuous on Rd , then f is measurable. If f is mea-

surable and finite-valued, and Φ is continuous, then Φ ◦ f is measurable.
In fact, Φ is continuous, so Φ−1 ((−∞, a)) is an open set O, and hence

(Φ ◦ f )−1 ((−∞, a)) = f −1 (O) is measurable.
It should be noted, however, that in general it is not true that
f ◦ Φ is measurable whenever f is measurable and Φ is continuous. See
Exercise 35.
Property 3 Suppose {fn }∞

n=1 is a sequence of measurable functions.
Then
sup fn (x), inf fn (x), lim sup, fn (x) and lim inf fn (x)
n n n→∞ n→∞
are measurable.
Proving
S that supn fn is measurable requires noting that {supn fn > a} =
n {fn > a}. This also yields the result for inf n fn (x), since this quantity
equals − supn (−fn (x)).
The result for the limsup and liminf also follows from the two obser-
vations
lim sup fn (x) = inf {sup fn } and lim inf fn (x) = sup{ inf fn }.
n→∞ k n≥k n→∞ k n≥k
Property 4 If {fn }∞
n=1 is a collection of measurable functions, and
lim fn (x) = f (x),

n→∞
then f is measurable.
Since f (x) = lim supn→∞ fn (x) = lim inf n→∞ fn (x), this property is a
consequence of property 3.
Property 5 If f and g are measurable, then

(i) The integer powers f k , k ≥ 1 are measurable.
(ii) f + g and f g are measurable if both f and g are finite-valued.
For (i) we simply note that if k is odd, then {f k > a} = {f > a1/k }, and
if k is even and a ≥ 0, then {f k > a} = {f > a1/k } ∪ {f < −a1/k }.
For (ii), we first see that f + g is measurable because
[
{f + g > a} = {f > a − r} ∩ {g > r},
r∈Q
with Q denoting the rationals.

Finally, f g is measurable because of the previous results and the fact
that
1
fg = [(f + g)2 − (f − g)2 ].
4
We shall say that two functions f and g defined on a set E are equal
almost everywhere, and write
f (x) = g(x) a.e. x ∈ E,
if the set {x ∈ E : f (x) 6= g(x)} has measure zero. We sometimes ab-

breviate this by saying that f = g a.e. More generally, a property or
statement is said to hold almost everywhere (a.e.) if it is true except on
a set of measure zero.
One sees easily that if f is measurable and f = g a.e., then g is measur-
able. This follows at once from the fact that {f < a} and {g < a} differ
by a set of measure zero. Moreover, all the properties above can be re-
laxed to conditions holding almost everywhere. For instance, if {fn }∞ n=1
is a collection of measurable functions, and
lim fn (x) = f (x) a.e.,

n→∞
then f is measurable.
Note that if f and g are defined almost everywhere on a measurable
subset E ⊂ Rd , then the functions f + g and f g can only be defined on
the intersection of the domains of f and g. Since the union of two sets of
measure zero has again measure zero, f + g is defined almost everywhere
on E. We summarize this discussion as follows.
Property 6 Suppose f is measurable, and f (x) = g(x) for a.e. x. Then

g is measurable.
In this light, Property 5 (ii) also holds when f and g are finite-valued
almost everywhere.
4.2 Approximation by simple functions or step functions

The theorems in this section are all of the same nature and provide
further insight in the structure of measurable functions. We begin by
approximating pointwise, non-negative measurable functions by simple
functions.
Theorem 4.1 Suppose f is a non-negative measurable function on Rd .

Then there exists an increasing sequence of non-negative simple functions
{ϕk }∞
k=1 that converges pointwise to f , namely,
ϕk (x) ≤ ϕk+1 (x) and lim ϕk (x) = f (x), for all x.

k→∞
Proof. We begin first with a truncation. For N ≥ 1, let QN denote

the cube centered at the origin and of side length N . Then we define

 f (x) if x ∈ QN and f (x) ≤ N ,
FN (x) = N if x ∈ QN and f (x) > N ,

0 otherwise.
Then, FN (x) → f (x) as N tends to infinity for all x. Now, we partition

the range of FN , namely [0, N ], as follows. For fixed N, M ≥ 1, we define
½ ¾
` `+1
E`,M = x ∈ QN : < FN (x) ≤ , for 0 ≤ ` < N M .
M M
Then we may form

X `
FN,M (x) = χE`,M (x).
M
`
Each FN,M is a simple function that satisfies 0 ≤ FN (x) − FN,M (x) ≤

1/M for all x. If we now choose N = M = 2k with k ≥ 1 integral, and
let ϕk = F2k ,2k , then we see that 0 ≤ FM (x) − ϕk (x) ≤ 1/2k for all x,
{ϕk } is increasing, and this sequence satisfies all the desired properties.
Note that the result holds for non-negative functions that are extended-
valued, if the limit +∞ is allowed. We now drop the assumption that f
is non-negative, and also allow the extended limit −∞.
Theorem 4.2 Suppose f is measurable on Rd . Then there exists a se-

quence of simple functions {ϕk }∞
k=1 that satisfies
|ϕk (x)| ≤ |ϕk+1 (x)| and lim ϕk (x) = f (x), for all x.
k→∞
In particular, we have |ϕk (x)| ≤ |f (x)| for all x and k.

Proof. We use the following decomposition of the function f : f (x) =
f (x) − f − (x), where
+
f + (x) = max(f (x), 0) and f − (x) = max(−f (x), 0).

Since both f + and f − are non-negative, the previous theorem yields

(1)
two increasing sequences of non-negative simple functions {ϕk (x)}∞ k=1
(2)
and {ϕk (x)}∞ k=1 which converge pointwise to f +
and f −
, respectively.
Then, if we let
(1) (2)
ϕk (x) = ϕk (x) − ϕk (x),
we see that ϕk (x) converges to f (x) for all x. Finally, the sequence {|ϕk |}
(1)
is increasing because the definition of f + , f − and the properties of ϕk
(2)
and ϕk imply that
(1) (2)
|ϕk (x)| = ϕk (x) + ϕk (x).
We may now go one step further, and approximate by step functions.

Here, in general, the convergence may hold only almost everywhere.
Theorem 4.3 Suppose f is measurable on Rd . Then there exists a se-
quence of step functions {ψk }∞
k=1 that converges pointwise to f (x) for
almost every x.
Proof. By the previous result, it suffices to show that if E is a
measurable set with finite measure, then f = χE can be approximated
by step functions. To this end, we recall part (iv) of Theorem 3.4,
which statesSN that for every ² there exist cubes Q1 , . . . , QN such that
m(E4 j=1 Qj ) ≤ ². By considering the grid formed by extending the
sides of these cubes, we SNsee that Sthere exist almost disjoint rectangles
M
R̃1 , . . . , R̃M such that j=1 Qj = j=1 R̃j . By taking rectangles Rj con-
tained in R̃j , and slightly smaller in size, we find a collection of disjoint
SM
rectangles that satisfy m(E4 j=1 Rj ) ≤ 2². Therefore
M
X
f (x) = χRj (x),
j=1
except possibly on a set of measure ≤ 2². Consequently, for every k ≥ 1,

there exists a step function ψk (x) such that if
Ek = {x : f (x) 6= ψk (x)},
S∞ T∞
then m(Ek ) ≤ 2−k . If we let FK = j=K+1 Ej and F = K=1 FK , then
m(F ) = 0 since m(FK ) ≤ 2−K , and ψk (x) → f (x) for all x in the com-
plement of F , which is the desired result.
4.3 Littlewood’s three principles

Although the notions of measurable sets and measurable functions rep-
resent new tools, we should not overlook their relation to the older con-
cepts they replaced. Littlewood aptly summarized these connections in
the form of three principles that provide a useful intuitive guide in the
initial study of the theory.
(i) Every set is nearly a finite union of intervals.

(ii) Every function is nearly continuous.
(iii) Every convergent sequence is nearly uniformly convergent.
The sets and functions referred to above are of course assumed to

be measurable. The catch is in the word “nearly,” which has to be
understood appropriately in each context. A precise version of the first
principle appears in part (iv) of Theorem 3.4. An exact formulation of
the third principle is given in the following important result.
Theorem 4.4 (Egorov) Suppose {fk }∞ k=1 is a sequence of measurable

functions defined on a measurable set E with m(E) < ∞, and assume
that fk → f a.e on E. Given ² > 0, we can find a closed set A² ⊂ E
such that m(E − A² ) ≤ ² and fk → f uniformly on A² .
Proof. We may assume without loss of generality that fk (x) → f (x)

for every x ∈ E. For each pair of non-negative integers n and k, let
Ekn = {x ∈ E : |fj (x) − f (x)| < 1/n, for all j > k}.
Now fix n and note that Ekn ⊂ Ek+1

n
, and Ekn % E as k tends to infinity.
By Corollary 3.3, we find that there exists kn such that m(E − Eknn ) <
1/2n . By construction, we then have
|fj (x) − f (x)| < 1/n whenever j > kn and x ∈ Eknn .

P∞
We choose N so that n=N 2−n < ²/2, and let
\
Ã² = Eknn .
n≥N
We first observe that

∞
X
m(E − Ã² ) ≤ m(E − Eknn ) < ²/2.
n=N
Next, if δ > 0, we choose n ≥ N such that 1/n < δ, and note that x ∈
Ã² implies x ∈ Eknn . We see therefore that |fj (x) − f (x)| < δ whenever
j > kn . Hence fk converges uniformly to f on Ã² .
Finally, using Theorem 3.4 choose a closed subset A² ⊂ Ã² with m(Ã² −
A² ) < ²/2. As a result, we have m(E − A² ) < ² and the theorem is
proved.
The next theorem attests to the validity of the second of Littlewood’s
principle.
Theorem 4.5 (Lusin) Suppose f is measurable and finite valued on E

with E of finite measure. Then for every ² > 0 there exists a closed set
F² , with
F² ⊂ E, and m(E − F² ) ≤ ²
and such that f |F² is continuous.
By f |F² we mean the restriction of f to the set F² . The conclusion of

the theorem states that if f is viewed as a function defined only on F² ,
then f is continuous. However, the theorem does not make the stronger
assertion that the function f defined on E is continuous at the points of
F² .
Proof. Let fn be a sequence of step functions so that fn → f a.e.
Then we may find sets En so that m(En ) < 1/2n and fn is continuous
outside En . By Egorov’s theorem, we may find a set A²/3 on which
fn → f uniformly and m(E − A²/3 ) ≤ ²/3. Then we consider
[
F 0 = A²/3 − En
n≥N
P
for N so large that n≥N 1/2n < ²/3. Now for every n ≥ N the function
fn is continuous on F 0 ; thus f (being the uniform limit of {fn }) is also
continuous on F 0 . To finish the proof, we merely need to approximate
the set F 0 by a closed set F² ⊂ F 0 such that m(F 0 − F² ) < ²/3.
5* The Brunn-Minkowski inequality

Since addition and multiplication by scalars are basic features of vector
spaces, it is not surprising that properties of these operations arise in a
fundamental way in the theory of Lebesgue measure on Rd . We have al-
ready discussed in this connection the translation-invariance and relative
5*. The Brunn-Minkowski inequality 35
dilation-invariance of Lebesgue measure. Here we come to the study of

the sum of two measurable sets A and B, defined as
A + B = {x ∈ Rd : x = x0 + x00 with x0 ∈ A and x00 ∈ B}.
This notion is of importance in a number of questions, in particular in

the theory of convex sets; we shall apply it to the isoperimetric problem
in Chapter 3.
In this regard the first (admittedly vague) question we can pose is
whether one can establish any general estimate for the measure of A + B
in terms of the measures of A and B (assuming that these three sets
are measurable). We can see easily that it is not possible to obtain an
upper bound for m(A + B) in terms of m(A) and m(B). Indeed, simple
examples show that we may have m(A) = m(B) = 0 while m(A + B) >
0. (See Exercise 20.)
In the converse direction one might ask for a general estimate of the
form
m(A + B)α ≥ cα (m(A)α + m(B)α ) ,
where α is a positive number and the constant cα is independent of A

and B. Clearly, the best one can hope for is cα = 1. The role of the
exponent α can be understood by considering convex sets. Such sets
A are defined by the property that whenever x and y are in A then
the line segment joining them, {xt + y(1 − t) : 0 ≤ t ≤ 1}, also belongs
to A. If we recall the definition λA = {λx, x ∈ A} for λ > 0, we note
that whenever A is convex, then A + λA = (1 + λ)A. However, m((1 +
λ)A) = (1 + λ)d m(A), and thus the presumed inequality can hold only
if (1 + λ)dα ≥ 1 + λdα , for all λ > 0. Now
(7) (a + b)γ ≥ aγ + bγ if γ ≥ 1 and a, b ≥ 0,
while the reverse inequality holds if 0 ≤ γ ≤ 1. (See Exercise 38.) This

yields α ≥ 1/d. Moreover, (7) shows that the inequality with the expo-
nent 1/d implies the corresponding inequality with α ≥ 1/d, and so we
are naturally led to the inequality
(8) m(A + B)1/d ≥ m(A)1/d + m(B)1/d .
Before proceeding with the proof of (8), we need to mention a technical

impediment that arises. While we may assume that A and B are mea-
surable, it does not necessarily follow that then A + B is measurable.
(See Exercise 13 in the next chapter.) However it is easily seen that this
difficulty does not occur when, for example, A and B are closed sets, or
when one of them is open. (See Exercise 19.)
With the above considerations in mind we can state the main result.
Theorem 5.1 Suppose A and B are measurable sets in Rd and their
sum A + B is also measurable. Then the inequality (8) holds.
Let us first check (8) when A and B are rectangles with side lengths
{aj }dj=1 and {bj }dj=1 , respectively. Then (8) becomes
Ã d
!1/d Ã d
!1/d Ã d
!1/d
Y Y Y
(9) (aj + bj ) ≥ aj + bj ,
j=1 j=1 j=1
which by homogeneity we can reduce to the special case where aj +

bj = 1 for each j. In fact, notice that if we replace aj , bj by λj aj , λj bj ,
with λj > 0, then both sides of (9) are multiplied by (λ1 λ2 · · · λd )1/d .
We then need only choose λj = (aj + bj )−1 . With this reduction, the
inequality (9) is an immediate consequence of the arithmetic-geometric
inequality (Exercise 39)
d
Ã d
!1/d
1X Y
xj ≥ xj , for all xj ≥ 0:
d
j=1 j=1
we add the two inequalities that result when we set xj = aj and xj = bj ,

respectively.
We next turn to the case when each A and B are the union of finitely
many rectangles whose interiors are disjoint. We shall prove (8) in this
case by induction on the total number of rectangles in A and B. We
denote this number by n. Here it is important to note that the desired
inequality is unchanged when we translate A and B independently. In
fact, replacing A by A + h and B by B + h0 replaces A + B by A + B +
h + h0 , and thus the corresponding measures remain the same. We now
choose a pair of disjoint rectangles R1 and R2 in the collection making up
A, and we note that they can be separated by a coordinate hyperplane.
Thus we may assume that for some j, after translation by an appropriate
h, R1 lies in A− = A ∩ {xj ≤ 0}, and R2 in A+ = A ∩ {0 ≤ xj }. Observe
also that both A+ and A− contain at least one less rectangle than A does,
and A = A− ∪ A+ .
We next translate B so that B− = B ∩ {xj ≤ 0} and B+ = B ∩ {xj ≥
0} satisfy
m(B± ) m(A± )
= .
m(B) m(A)
6. Exercises 37
However, A + B ⊃ (A+ + B+ ) ∪ (A− + B− ), and the union on the right

is essentially disjoint, since the two parts lie in different half-spaces.
Moreover, the total number of rectangles in either A+ and B+ , or A−
and B− is also less than n. Thus the induction hypothesis applies and
m(A + B) ≥ m(A+ + B+ ) + m(A− + B− )

¡ ¢d ¡ ¢d
≥ m(A+ )1/d + m(B+ )1/d + m(A− )1/d + m(B− )1/d
" µ ¶1/d #d " µ ¶1/d #d
m(B) m(B)
= m(A+ ) 1 + + m(A− ) 1 +
m(A) m(A)
¡ ¢d
= m(A)1/d + m(B)1/d ,
which gives the desired inequality (8) when A and B are both finite
unions of rectangles with disjoint interiors.
Next, this quickly implies the result when A and B are open sets of
finite measure. Indeed, by Theorem 1.4, for any ² > 0 we can find unions
of almost disjoint rectangles A² and B² , such that A² ⊂ A, B² ⊂ B, with
m(A) ≤ m(A² ) + ² and m(B) ≤ m(B² ) + ². Since A + B ⊃ A² + B² , the
inequality (8) for A² and B² and a passage to a limit gives the desired
result. From this, we can pass to the case where A and B are arbitrary
compact sets, by noting first that A + B is then compact, and that if
we define A² = {x : d(x, A) < ²}, then A² are open, and A² & A as ² →
0. With similar definitions for B ² and (A + B)² , we observe also that
A + B ⊂ A² + B ² ⊂ (A + B)2² . Hence, letting ² → 0, we see that (8) for
A² and B ² implies the desired result for A and B. The general case,
in which we assume that A, B, and A + B are measurable, then follows
by approximating A and B from inside by compact sets, as in (iii) of
Theorem 3.4.
6 Exercises
1. Prove that the Cantor set C constructed in the text is totally disconnected and
perfect. In other words, given two distinct points x, y ∈ C, there is a point z ∈
/C
that lies between x and y, and yet C has no isolated points.
[Hint: If x, y ∈ C and |x − y| > 1/3k , then x and y belong to two different intervals
in Ck . Also, given any x ∈ C there is an end-point yk of some interval in Ck that
satisfies x 6= yk and |x − yk | ≤ 1/3k .]
2. The Cantor set C can also be described in terms of ternary expansions.

(a) Every number in [0, 1] has a ternary expansion

∞
X
x= ak 3−k , where ak = 0, 1, or 2.
k=1
P∞
Note that this decomposition is not unique since, for example, 1/3 = k=2 2/3k .
Prove that x ∈ C if and only if x has a representation as above where every
ak is either 0 or 2.
(b) The Cantor-Lebesgue function is defined on C by
∞
X bk P∞
F (x) = if x = k=1 ak 3−k , where bk = ak /2.
2k
k=1
In this definition, we choose the expansion of x in which ak = 0 or 2.

Show that F is well defined and continuous on C, and moreover F (0) = 0 as
well as F (1) = 1.
(c) Prove that F : C → [0, 1] is surjective, that is, for every y ∈ [0, 1] there exists
x ∈ C such that F (x) = y.
(d) One can also extend F to be a continuous function on [0, 1] as follows. Note
that if (a, b) is an open interval of the complement of C, then F (a) = F (b).
Hence we may define F to have the constant value F (a) in that interval.
A geometrical construction of F is described in Chapter 3.
3. Cantor sets of constant dissection. Consider the unit interval [0, 1], and
let ξ be a fixed real number with 0 < ξ < 1 (the case ξ = 1/3 corresponds to the
Cantor set C in the text).
In stage 1 of the construction, remove the centrally situated open interval in
[0, 1] of length ξ. In stage 2, remove two central intervals each of relative length ξ,
one in each of the remaining intervals after stage 1, and so on.
Let Cξ denote the set which remains after applying the above procedure indefi-
nitely.6
(a) Prove that the complement of Cξ in [0, 1] is the union of open intervals of
total length equal to 1.
(b) Show directly that m∗ (Cξ ) = 0.
[Hint: After the kth stage, show that the remaining set has total length = (1 − ξ)k .]
4. Cantor-like sets. Construct a closed set Cˆ so that at the kth stage of the
construction one removes 2k−1 centrally situated open intervals each of length `k ,
with
`1 + 2`2 + · · · + 2k−1 `k < 1.
6 The set we call Cξ is sometimes denoted by C 1−ξ .

2
6. Exercises 39
P∞ k−1
(a) If `j are chosen small enough, then k=1 2 `k < 1. In this case, show
that m(C) ˆ = 1 − P∞ 2k−1 `k .
ˆ > 0, and in fact, m(C)
k=1
ˆ then there exists a sequence of points {xn }∞

(b) Show that if x ∈ C, n=1 such
that xn ∈ ˆ
/ C, yet xn → x and xn ∈ In , where In is a sub-interval in the
complement of Cˆ with |In | → 0.
(c) Prove as a consequence that Cˆ is perfect, and contains no open interval.
(d) Show also that Cˆ is uncountable.
5. Suppose E is a given set, and On is the open set:
On = {x : d(x, E) < 1/n}.
Show:
(a) If E is compact, then m(E) = limn→∞ m(On ).
(b) However, the conclusion in (a) may be false for E closed and unbounded; or
E open and bounded.
6. Using translations and dilations, prove the following: Let B be a ball in Rd of

radius r. Then m(B) = vd rd , where vd = m(B1 ), and B1 is the unit ball, B1 =
{x ∈ Rd : |x| < 1}.
A calculation of the constant vd is postponed until Exercise 14 in the next
chapter.
7. If δ = (δ1 , . . . , δd ) is a d-tuple of positive numbers δi > 0, and E is a subset of

Rd , we define δE by
δE = {(δ1 x1 , . . . , δd xd ) : where (x1 , . . . , xd ) ∈ E}.
Prove that δE is measurable whenever E is measurable, and
m(δE) = δ1 · · · δd m(E).
8. Suppose L is a linear transformation of Rd . Show that if E is a measurable

subset of Rd , then so is L(E), by proceeding as follows:
(a) Note that if E is compact, so is L(E). Hence if E is an Fσ set, so is L(E).
(b) Because L automatically satisfies the inequality
|L(x) − L(x0 )| ≤ M |x − x0 |
for some M , we can see that L maps√any cube of side length ` into a
cube of side length cd M `, with cd = 2 S
d. Now if P m(E) = 0, there is a
collection of cubes {Qj } such that E ⊂ j Qj , and j m(Qj ) < ². Thus
m∗ (L(E)) ≤ c0 ², and hence m(L(E)) = 0. Finally, use Corollary 3.5.
One can show that m(L(E)) = | det L| m(E); see Problem 4 in the next chapter.
9. Give an example of an open set O with the following property: the boundary
of the closure of O has positive Lebesgue measure.
[Hint: Consider the set obtained by taking the union of open intervals which are
deleted at the odd steps in the construction of a Cantor-like set.]
10. This exercise provides a construction of a decreasing sequence of positive

continuous functions on the interval [0, 1], whose pointwise limit is not Riemann
integrable.
Let Cˆ denote a Cantor-like set obtained from the construction detailed in Exer-
ˆ > 0. Let F1 denote a piecewise-linear and contin-
cise 4, so that in particular m(C)
uous function on [0, 1], with F1 = 1 in the complement of the first interval removed
in the construction of C,ˆ F1 = 0 at the center of this interval, and 0 ≤ F1 (x) ≤ 1 for
all x. Similarly, construct F2 = 1 in the complement of the intervals in stage two of
the construction of C,ˆ with F2 = 0 at the center of these intervals, and 0 ≤ F2 ≤ 1.
Continuing this way, let fn = F1 · F2 · · · Fn (see Figure 5).
F1
F2
Figure 5. Construction of {Fn } in Exercise 10
Prove the following:

(a) For all n ≥ 1 and all x ∈ [0, 1], one has 0 ≤ fn (x) ≤ 1 and fn (x) ≥ fn+1 (x).
Therefore, fn (x) converges to a limit as n → ∞ which we denote by f (x).
(b) The function f is discontinuous at every point of C. ˆ

ˆ
[Hint: Note that f (x) = 1 if x ∈ C, and find a sequence of points {xn } so
that xn → x and f (xn ) = 0.]
R R
Now fn (x) dx is decreasing, hence fn converges. However, a bounded func-
tion is Riemann integrable if and only if its set of discontinuities has measure zero.
6. Exercises 41
(The proof of this fact, which is given in the Appendix of Book I, is outlined in
Problem 4.) Since f is discontinuous on a set of positive measure, we find that f
is not Riemann integrable.
11. Let A be the subset of [0, 1] which consists of all numbers which do not have
the digit 4 appearing in their decimal expansion. Find m(A).
12. Theorem 1.3 states that every open set in R is the disjoint union of open
intervals. The analogue in Rd , d ≥ 2, is generally false. Prove the following:
(a) An open disc in R2 is not the disjoint union of open rectangles.

[Hint: What happens to the boundary of any of these rectangles?]
(b) An open connected set Ω is the disjoint union of open rectangles if and only
if Ω is itself an open rectangle.
13. The following deals with Gδ and Fσ sets.
(a) Show that a closed set is a Gδ and an open set an Fσ .

[Hint: If F is closed, consider On = {x : d(x, F ) < 1/n}.]
(b) Give an example of an Fσ which is not a Gδ .

[Hint: This is more difficult; let F be a denumerable set that is dense.]
(c) Give an example of a Borel set which is not a Gδ nor an Fσ .
14. The purpose of this exercise is to show that covering by a finite number of
intervals will not suffice in the definition of the outer measure m∗ .
The outer Jordan content J∗ (E) of a set E in R is defined by
N
X
J∗ (E) = inf |Ij |,
j=1
SN
where the inf is taken over every finite covering E ⊂ j=1 Ij , by intervals Ij .
(a) Prove that J∗ (E) = J∗ (E) for every set E (here E denotes the closure of
E).
(b) Exhibit a countable subset E ⊂ [0, 1] such that J∗ (E) = 1 while m∗ (E) = 0.
15. At the start of the theory, one might define the outer measure by taking
coverings by rectangles instead of cubes. More precisely, we define
∞
X
mR
∗ (E) = inf |Rj |,
j=1
S
where the inf is now taken over all countable coverings E ⊂ ∞ j=1 Rj by (closed)
rectangles.
Show that this approach gives rise to the same theory of measure developed in
the text, by proving that m∗ (E) = mR d
∗ (E) for every subset E of R .
[Hint: Use Lemma 1.1.]
16. The Borel-Cantelli lemma. Suppose {Ek }∞

k=1 is a countable family of
measuable subsets of Rd and that
∞
X
m(Ek ) < ∞.
k=1
Let
E = {x ∈ Rd : x ∈ Ek , for infinitely many k}

= lim sup(Ek ).
k→∞
(a) Show that E is measurable.
(b) Prove m(E) = 0.

T S
[Hint: Write E = ∞n=1 k≥n Ek .]
17. Let {fn } be a sequence of measurable functions on [0, 1] with |fn (x)| < ∞ for
a.e x. Show that there exists a sequence cn of positive real numbers such that
fn (x)
→0 a.e. x
cn
[Hint: Pick cn such that m({x : |fn (x)/cn | > 1/n}) < 2−n , and apply the Borel-
Cantelli lemma.]
18. Prove the following assertion: Every measurable function is the limit a.e. of a
sequence of continuous functions.
19. Here are some observations regarding the set operation A + B.
(a) Show that if either A and B is open, then A + B is open.
(b) Show that if A and B are closed, then A + B is measurable.
(c) Show, however, that A + B might not be closed even though A and B are
closed.
[Hint: For (b) show that A + B is an Fσ set.]
20. Show that there exist closed sets A and B with m(A) = m(B) = 0, but m(A +
B) > 0:
6. Exercises 43
(a) In R, let A = C (the Cantor set), B = C/2. Note that A + B ⊃ [0, 1].
(b) In R2 , observe that if A = I × {0} and B = {0} × I (where I = [0, 1]), then
A + B = I × I.
21. Prove that there is a continuous function that maps a Lebesgue measurable
set to a non-measurable set.
[Hint: Consider a non-measurable subset of [0, 1], and its inverse image in C by the
function F in Exercise 2.]
22. Let χ[0,1] be the characteristic function of [0, 1]. Show that there is no every-
where continuous function f on R such that
f (x) = χ[0,1] (x) almost everywhere.
23. Suppose f (x, y) is a function on R2 that is separately continuous: for each

fixed variable, f is continuous in the other variable. Prove that f is measurable
on R2 .
[Hint: Approximate f in the variable x by piecewise-linear functions fn so that
fn → f pointwise.]
24. Does there exist an enumeration {rn }∞

n=1 of the rationals, such that the
complement of
∞ „
[ «
1 1
rn − , rn +
n=1
n n
in R is non-empty?
[Hint: Find an enumeration where the only rationals outside of a fixed bounded
interval take the form rn , with n = m2 for some integer m.]
25. An alternative definition of measurability is as follows: E is measurable if for

every ² > 0 there is a closed set F contained in E with m∗ (E − F ) < ². Show that
this definition is equivalent with the one given in the text.
26. Suppose A ⊂ E ⊂ B, where A and B are measurable sets of finite measure.

Prove that if m(A) = m(B), then E is measurable.
27. Suppose E1 and E2 are a pair of compact sets in Rd with E1 ⊂ E2 , and let
a = m(E1 ) and b = m(E2 ). Prove that for any c with a < c < b, there is a compact
set E with E1 ⊂ E ⊂ E2 and m(E) = c.
[Hint: As an example, if d = 1 and E is a measurable subset of [0, 1], consider
m(E ∩ [0, t]) as a function of t.]
28. Let E be a subset of R with m∗ (E) > 0. Prove that for each 0 < α < 1, there
exists an open interval I so that
m∗ (E ∩ I) ≥ α m∗ (I).
Loosely speaking, this estimate shows that E contains almost a whole interval.
[Hint: Choose an open set O that contains E, and such that m∗ (E) ≥ α m∗ (O).
Write O as the countable union of disjoint open intervals, and show that one of
these intervals must satisfy the desired property.]
29. Suppose E is a measurable subset of R with m(E) > 0. Prove that the
difference set of E, which is defined by
{z ∈ R : z = x − y for some x, y ∈ E},
contains an open interval centered at the origin.

If E contains an interval, then the conclusion is straightforward. In general, one
may rely on Exercise 28.
[Hint: Indeed, by Exercise 28, there exists an open interval I so that m(E ∩ I) ≥
(9/10) m(I). If we denote E ∩ I by E0 , and suppose that the difference set of E0
does not contain an open interval around the origin, then for arbitrarily small a the
sets E0 , and E0 + a are disjoint. From the fact that (E0 ∪ (E0 + a)) ⊂ (I ∪ (I + a))
we get a contradiction, since the left-hand side has measure 2m(E0 ), while the
right-hand side has measure only slightly larger than m(I).]
A more general formulation of this result is as follows.
30. If E and F are measurable, and m(E) > 0, m(F ) > 0, prove that
E + F = {x + y : x ∈ E, x ∈ F }
contains an interval.
31. The result in Exercise 29 provides an alternate proof of the non-measurability

of the set N studied in the text. In fact, we may also prove the non-measurability
of a set in R that is very closely related to the set N .
Given two real numbers x and y, we shall write as before that x ∼ y whenever
the difference x − y is rational. Let N ∗ denote a set that consists of one element in
each equivalence class of ∼. Prove that N ∗ is non-measurable by using the result
in Exercise 29.
[Hint: If N ∗ is measurable, then so are its translates Nn∗ = N ∗ + rn , where {rn }∞
n=1
is an enumeration of Q. How does this imply that m(N ∗ ) > 0? Can the difference
set of N ∗ contain an open interval centered at the origin?]
32. Let N denote the non-measurable subset of I = [0, 1] constructed at the end
of Section 3.
(a) Prove that if E is a measurable subset of N , then m(E) = 0.

6. Exercises 45
(b) If G is a subset of R with m∗ (G) > 0, prove that a subset of G is non-

measurable.
[Hint: For (a) use the translates of E by the rationals.]
33. Let N denote the non-measurable set constructed in the text. Recall from the
exercise above that measurable subsets of N have measure zero.
Show that the set N c = I − N satisfies m∗ (N c ) = 1, and conclude that if E1 =
N and E2 = N c , then
m∗ (E1 ) + m∗ (E2 ) 6= m∗ (E1 ∪ E2 ),
although E1 and E2 are disjoint.

[Hint: To prove that m∗ (N c ) = 1, argue by contradiction and pick a measurable
set U such that U ⊂ I, N c ⊂ U and m∗ (U ) < 1 − ².]
34. Let C1 and C2 be any two Cantor sets (constructed in Exercise 3). Show that
there exists a function F : [0, 1] → [0, 1] with the following properties:
(i) F is continuous and bijective,
(ii) F is monotonically increasing,
(iii) F maps C1 surjectively onto C2 .

[Hint: Copy the construction of the standard Cantor-Lebesgue function.]
35. Give an example of a measurable function f and a continuous function Φ so

that f ◦ Φ is non-measurable.
[Hint: Let Φ : C1 → C2 as in Exercise 34, with m(C1 ) > 0 and m(C2 ) = 0. Let
N ⊂ C1 be non-measurable, and take f = χΦ(N ) .]
Use the construction in the hint to show that there exists a Lebesgue measurable
set that is not a Borel set.
36. This exercise provides an example of a measurable function f on [0, 1] such

that every function g equivalent to f (in the sense that f and g differ only on a
set of measure zero) is discontinuous at every point.
(a) Construct a measurable set E ⊂ [0, 1] such that for any non-empty open
sub-interval I in [0, 1], both sets E ∩ I and E c ∩ I have positive measure.
(b) Show that f = χE has the property that whenever g(x) = f (x) a.e x, then
g must be discontinuous at every point in [0, 1].
[Hint: For the first part, consider a Cantor-like set of positive measure, and add in
each of the intervals that are omitted in the first step of its construction, another
Cantor-like set. Continue this procedure indefinitely.]
37. Suppose Γ is a curve y = f (x) in R2 , where f is continuous. Show that

m(Γ) = 0.
[Hint: Cover Γ by rectangles, using the uniform continuity of f .]
38. Prove that (a + b)γ ≥ aγ + bγ whenever γ ≥ 1 and a, b ≥ 0. Also, show that

the reverse inequality holds when 0 ≤ γ ≤ 1.
[Hint: Integrate the inequality between (a + t)γ−1 and tγ−1 from 0 to b.]
39. Establish the inequality
x1 + · · · + xd
(10) ≥ (x1 · · · xd )1/d for all xj ≥ 0, j = 1, . . . , d
d
by using backward induction as follows:
(a) The inequality is true whenever d is a power of 2 (d = 2k , k ≥ 1).
(b) If (10) holds for some integer d ≥ 2, then it must hold for d − 1, that is,
one has (y1 + · · · + yd−1 )/(d − 1) ≥ (y1 · · · yd−1 )1/(d−1) for all yj ≥ 0, with
j = 1, . . . , d − 1.
[Hint: For (a), if k ≥ 2, write (x1 + · · · + x2k )/2k as (A + B)/2, where A = (x1 +
· · · + x2k−1 )/2k−1 , and apply the inequality when d = 2. For (b), apply the in-
equality to x1 = y1 , . . . , xd−1 = yd−1 and xd = (y1 + · · · + yd−1 )/(d − 1).]
7 Problems
1. Given an irrational x, one can show (using the pigeon-hole principle, for exam-
ple) that there exists infinitely many fractions p/q, with relatively prime integers
p and q such that
˛ ˛
˛ ˛
˛x − p ˛ ≤ 1 .
˛ q˛ q2
However, prove that the set of those x ∈ R such that there exist infinitely many
fractions p/q, with relatively prime integers p and q such that
˛ ˛
˛ ˛
˛x − p ˛ ≤ 1 (or ≤ 1/q 2+² ),
˛ q˛ q3
is a set of measure zero.

[Hint: Use the Borel-Cantelli lemma.]
S
2. Any open set Ω can be written as the union of closed cubes, so that Ω = Qj
with the following properties
(i) The Qj ’s have disjoint interiors.
(ii) d(Qj , Ωc ) ≈ side length of Qj . This means that there are positive constants
c and C so that c ≤ d(Qj , Ωc )/`(Qj ) ≤ C, where `(Qj ) denotes the side
length of Qj .
7. Problems 47
3. Find an example of a measurable subset C of [0, 1] such that m(C) = 0, yet the
difference set of C contains a non-trivial interval centered at the origin. Compare
with the result in Exercise 29.
[Hint: Pick the Cantor set C = C. For a fixed a ∈ [−1, 1], consider the line y =
x + a in the plane, and copy the construction of the Cantor set, but in the cube
Q = [0, 1] × [0, 1]. First, remove all but four closed cubes of side length 1/3, one at
each corner of Q; then, repeat this procedure in each of the remaining cubes (see
Figure 6). The resulting set is sometimes called a Cantor dust. Use the property
of nested compact sets to show that the line intersects this Cantor dust.]
Figure 6. Construction of the Cantor dust
4. Complete the following outline to prove that a bounded function on an interval

[a, b] is Riemann integrable if and only if its set of discontinuities has measure zero.
This argument is given in detail in the appendix to Book I.
Let f be a bounded function on a compact interval J, and let I(c, r) denote
the open interval centered at c of radius r > 0. Let osc(f, c, r) = sup |f (x) − f (y)|,
where the supremum is taken over all x, y ∈ J ∩ I(c, r), and define the oscillation
of f at c by osc(f, c) = limr→0 osc(f, c, r). Clearly, f is continuous at c ∈ J if and
only if osc(f, c) = 0.
Prove the following assertions:
(a) For every ² > 0, the set of points c in J such that osc(f, c) ≥ ² is compact.
(b) If the set of discontinuities of f has measure 0, then f is Riemann integrable.

[Hint: Given ² > 0 let A² = {c ∈ J : osc(f, c) ≥ ²}. Cover A² by a finite
number of open intervals whose total length is ≤ ². Select an appropriate
partition of J and estimate the difference between the upper and lower sums
of f over this partition.]
(c) Conversely, if f is Riemann integrable on J, then its set of discontinuities

has measure 0.
S
[Hint: The set of discontinuities of f is contained in n A1/n . Choose a
partition P such that U (f, P ) − L(f, P ) < ²/n. Show that the total length
of the intervals in P whose interior intersect A1/n is ≤ ².]
5. Suppose E is measurable with m(E) < ∞, and
E = E1 ∪ E2 , E1 ∩ E2 = ∅.
If m(E) = m∗ (E1 ) + m∗ (E2 ), then E1 and E2 are measurable.

In particular, if E ⊂ Q, where Q is a finite cube, then E is measurable if and
only if m(Q) = m∗ (E) + m∗ (Q − E).
6.∗ The fact that the axiom of choice and the well-ordering principle are equivalent
is a consequence of the following considerations.
One begins by defining a partial ordering on a set E to be a binary relation ≤
on the set E that satisfies:
(i) x ≤ x for all x ∈ E.
(ii) If x ≤ y and y ≤ x, then x = y.
(iii) If x ≤ y and y ≤ z, then x ≤ z.
If in addition x ≤ y or y ≤ x whenever x, y ∈ E, then ≤ is a linear ordering of E.
The axiom of choice and the well-ordering principle are then logically equivalent
to the Hausdorff maximal principle:
Every non-empty partially ordered set has a (non-empty) maximal
linearly ordered subset.
In other words, if E is partially ordered by ≤, then E contains a non-empty subset
F which is linearly ordered by ≤ and such that if F is contained in a set G also
linearly ordered by ≤, then F = G.
An application of the Hausdorff maximal principle to the collection of all well-
orderings of subsets of E implies the well-ordering principle for E. However, the
proof that the axiom of choice implies the Hausdorff maximal principle is more
complicated.
7.∗ Consider the curve Γ = {y = f (x)} in R2 , 0 ≤ x ≤ 1. Assume that f is twice

continuously differentiable in 0 ≤ x ≤ 1. Then show that m(Γ + Γ) > 0 if and only
if Γ + Γ contains an open set, if and only if f is not linear.
8.∗ Suppose A and B are open sets of finite positive measure. Then we have
equality in the Brunn-Minkowski inequality (8) if and only if A and B are convex
and similar, that is, there are a δ > 0 and an h ∈ Rd such that
A = δB + h.
2 Integration Theory
...amongst the many definitions that have been succes-
sively proposed for the integral of real-valued functions
of a real variable, I have retained only those which, in
my opinion, are indispensable to understand the trans-
formations undergone by the problem of integration,
and to capture the relationship between the notion of
area, so simple in appearance, and certain more com-
plicated analytical definitions of the integral.
One might ask if there is sufficient interest to oc-
cupy oneself with such complications, and if it is not
better to restrict oneself to the study of functions that
necessitate only simple definitions.... As we shall see
in this course, we would then have to renounce the
possibility of resolving many problems posed long ago,
and which have simple statements. It is to solve these
problems, and not for love of complications, that I
have introduced in this book a definition of the inte-
gral more general than that of Riemann.
H. Lebesgue, 1903
1 The Lebesgue integral: basic properties and conver-

gence theorems
The general notion of the Lebesgue integral on Rd will be defined in a
step-by-step fashion, proceeding successively to increasingly larger fam-
ilies of functions. At each stage we shall see that the integral satisfies
elementary properties such as linearity and monotonicity, and we prove
appropriate convergence theorems that amount to interchanging the in-
tegral with limits. At the end of the process we shall have achieved a
general theory of integration that will be decisive in the study of further
problems.
We proceed in four stages, by progressively integrating:
1. Simple functions
2. Bounded functions supported on a set of finite measure
3. Non-negative functions
50 Chapter 2. INTEGRATION THEORY
4. Integrable functions (the general case).
We emphasize from the onset that all functions are assumed to be mea-
surable. At the beginning we also consider only finite-valued functions
which take on real values. Later we shall also consider extended-valued
functions, and also complex-valued functions.
Stage one: simple functions

Recall from the previous chapter that a simple function ϕ is a finite sum
N
X
(1) ϕ(x) = ak χEk (x),
k=1
where the Ek are measurable sets of finite measure and the ak are con-
stants. A complication that arises from this definition is that a simple
function can be written in a multitude of ways as such finite linear com-
binations; for example, 0 = χE − χE for any measurable set E of finite
measure. Fortunately, there is an unambiguous choice for the represen-
tation of a simple function, which is natural and useful in applications.
The canonical form of ϕ is the unique decomposition as in (1), where
the numbers ak are distinct and non-zero, and the sets Ek are disjoint.
Finding the canonical form of ϕ is straightforward: since ϕ can take
only finitely many distinct and non-zero values, say c1 , . . . , cM , we may
set Fk = {x : ϕ(x) = ck }, and note that the sets Fk are disjoint. There-
PM
fore ϕ = k=1 ck χFk is the desired canonical form of ϕ.
PM
If ϕ is a simple function with canonical form ϕ(x) = k=1 ck χFk (x),
then we define the Lebesgue integral of ϕ by
Z M
X
ϕ(x) dx = ck m(Fk ).
Rd k=1
If E is a measurable subset of Rd with finite measure, then ϕ(x)χE (x)

is also a simple function, and we define
Z Z
ϕ(x) dx = ϕ(x)χE (x) dx.
E
To emphasize the choice of the Lebesgue measure m in the definition of

the integral, one sometimes writes
Z
ϕ(x) dm(x)
Rd
1. The Lebesgue integral: basic properties and convergence theorems 51
for the Lebesgue Rintegral of ϕ. In fact,

R as a matter of convenience, we
shall often write ϕ(x) dx or simply ϕ for the integral of ϕ over Rd .
Proposition 1.1 The integral of simple functions defined above satisfies

the following properties:
PN
(i) Independence of the representation. If ϕ = k=1 ak χEk is any rep-
resentation of ϕ, then
Z N
X
ϕ= ak m(Ek ).
k=1
(ii) Linearity. If ϕ and ψ are simple, and a, b ∈ R, then

Z Z Z
(aϕ + bψ) = a ϕ + b ψ.
(iii) Additivity. If E and F are disjoint subsets of Rd with finite mea-

sure, then
Z Z Z
ϕ= ϕ+ ϕ.
E∪F E F
(iv) Monotonicity. If ϕ ≤ ψ are simple, then

Z Z
ϕ ≤ ψ.
(v) Triangle inequality. If ϕ is a simple function, then so is |ϕ|, and

¯Z ¯ Z
¯ ¯
¯ ϕ¯ ≤ |ϕ|.
¯ ¯
Proof. The only conclusion that is a little tricky is the first, which
asserts that the integral of a simple function can be calculated by us-
ing any of its decompositions as a linear combination of characteristic
functions. PN
Suppose that ϕ = k=1 ak χEk , where we assume that the sets Ek are
disjoint, but we do not suppose that the numbers ak are distinct and non-
zero. For each distinct non-zero value a among the {ak } we define Ea0 =
S
Ek , where the union is taken over those indices k such
P that ak = a.
Note then that the sets Ea0 are disjoint, and m(Ea0 ) = m(Ek ), where
P
the sum is taken over the same set of k’s. Then clearly ϕ = aχEa0 ,
where the sum is over the distinct non-zero values of {ak }. Thus
Z X N
X
ϕ= am(Ea0 ) = ak m(Ek ).
k=1
PN
Next, suppose ϕ = k=1 ak χEk , where we no longer assume that the Ek
SN
are disjoint. Then we can “refine” the decomposition k=1 Ek by finding
SN Sn
sets E1∗ , E2∗ , . . . , En∗ with the property that k=1 Ek = j=1 Ej∗S ; the
sets Ej∗ (j = 1, . . . , n) are mutually disjoint; and for each k, Ek = Ej∗ ,
where the union is taken over those Ej∗ that are contained in Ek . (A proof
of thisPelementary fact can be found in Exercise 1.) For each j, let now
a∗j = ak , with the summation
Pn taken over all k such that Ek contains
Ej∗ . Then clearly ϕ = j=1 a∗j χEj∗ . However, this is a decomposition
already dealt with above because the Ej∗ are disjoint. Thus
Z X X X X
ϕ= a∗j m(Ej∗ ) = ak m(Ej∗ ) = ak m(Ek ),
Ek ⊃Ej∗
and conclusion (i) is established.

Conclusion (ii) follows by using any representation of ϕ and ψ, and
the obvious linearity of (i).
For the additivity over sets, one must note that if E and F are disjoint,
then
χE∪F = χE + χF ,
R R
and
R we may use the linearity of the integral to see that E∪F
ϕ = E
ϕ+
F
ϕ.
If η ≥ 0 is a simple function,
R then its canonical form is everywhere non-
negative, and therefore η ≥ 0 by the definition of the integral. Applying
this argument to ψ − ϕ gives the desired monotonicity property.
Finally, P
for the triangle inequality, it suffices to write ϕ in its canonical
N
form ϕ = k=1 ak χEk and observe that
N
X
|ϕ| = |ak |χEk (x).
k=1
Therefore, by the triangle inequality applied to the definition of the in-

tegral, one sees that
¯Z ¯ ¯¯X ¯
¯ X Z
¯ ¯ ¯N ¯
N
¯ ϕ¯ = ¯ a m(E )
k ¯ ≤ |a |m(E ) = |ϕ|.
¯ ¯ ¯ k
¯
k k
k=1 k=1
Incidentally, it is worthwhile to point out the following easy fact: when-

ever fR and gR are a pair of simple functions that agree almost everywhere,
then f = g. The identity of the integrals of two functions that agree
almost everywhere will continue to hold for the successive definitions of
the integral that follow.
Stage two: bounded functions supported on a set of finite

measure
The support of a measurable function f is defined to be the set of all
points where f does not vanish,
supp(f ) = {x : f (x) 6= 0}.
We shall also say that f is supported on a set E, if f (x) = 0 whenever

x∈/ E.
Since f is measurable, so is the set supp(f ). We shall next be interested
in those bounded measurable functions that have m(supp(f )) < ∞.
An important result in the previous chapter (Theorem 4.2) states the
following: if f is a function bounded by M and supported on a set E, then
there exists a sequence {ϕn } of simple functions, with each ϕn bounded
by M and supported on E, and such that
ϕn (x) → f (x) for all x.
The key lemma that follows allows us to define the integral for the class
of bounded functions supported on sets of finite measure.
Lemma 1.2 Let f be a bounded function supported on a set E of finite

measure. If {ϕn }∞
n=1 is any sequence of simple functions bounded by M ,
supported on E, and with ϕn (x) → f (x) for a.e. x, then:
Z
(i) The limit lim ϕn exists.
n→∞
Z
(ii) If f = 0 a.e., then the limit lim ϕn equals 0.
n→∞
Proof. The assertions of the lemma would be nearly obvious if we

had that ϕn converges to f uniformly on E. Instead, we recall one of
Littlewood’s principles, which states that the convergence of a sequence
of measurable functions is “nearly” uniform. The precise statement lying
behind this principle is Egorov’s theorem, which we proved in Chapter 1,
and which we apply here.
Since the measure of E is finite, given ² > 0 Egorov’s theorem guar-

antees the existence of a (closed) measurable subset A² of E such that
m(E
R − A² ) ≤ ², and ϕn → f uniformly on A² . Therefore, setting In =
ϕn we have that
Z
|In − Im | ≤ |ϕn (x) − ϕm (x)| dx
E
Z Z
= |ϕn (x) − ϕm (x)| dx + |ϕn (x) − ϕm (x)| dx
A² E−A²
Z
≤ |ϕn (x) − ϕm (x)| dx + 2M m(E − A² )
ZA²
≤ |ϕn (x) − ϕm (x)| dx + 2M ².
A²
By the uniform convergence, one has, for all x ∈ A² and all large n and
m, the estimate |ϕn (x) − ϕm (x)| < ², so we deduce that
|In − Im | ≤ m(E)² + 2M ² for all large n and m.
Since ² is arbitrary and m(E) < ∞, this proves that {In } is a Cauchy
sequence and hence converges, as desired.
For the second part, we note that if f = 0, we may repeat the argument
above to find that |In | ≤ m(E)² + M ², which yields limn→∞ In = 0, as
was to be shown.
Using Lemma 1.2 we can now turn to the integration of bounded func-
tions that are supported on sets of finite measure. For such a function f
we define its Lebesgue integral by
Z Z
f (x) dx = lim ϕn (x) dx,
n→∞
where {ϕn } is any sequence of simple functions satisfying: |ϕn | ≤ M ,

each ϕn is supported on the support of f , and ϕn (x) → f (x) for a.e. x
as n tends to infinity. By the previous lemma, we know that this limit
exists.
R
Next, we must first show that f is independent of the limiting se-
quence {ϕn } used, in order for the integral to be well-defined. There-
fore, suppose that {ψn } is another sequence of simple functions that is
bounded by M , supported on supp(f ), and such that ψn (x) → f (x) for
a.e. x as n tends to infinity. Then, if ηn = ϕn − ψn , the sequence {ηn }
consists of simple functions bounded by 2M , supported on a set of fi-
nite measure, and such that ηn → 0 a.e. as n tends to infinity. We may
R
therefore conclude, by the second part of the lemma, that ηn → 0 as n
tends to infinity. Consequently, the two limits
Z Z
lim ϕn (x) dx and lim ψn (x) dx
n→∞ n→∞
(which exist by the lemma) are indeed equal.

If E is a subset of Rd with finite measure, and f is bounded with
m(supp(f )) < ∞, then it is natural to define
Z Z
f (x) dx = f (x)χE (x) dx.
E
R
Clearly, if f is itself simple, then f as defined above coincides with
the integral of simple functions studied earlier. This extension of the def-
inition of integration also satisfies all the basic properties of the integral
of simple functions.
Proposition 1.3 Suppose f and g are bounded functions supported on

sets of finite measure. Then the following properties hold.
(i) Linearity. If a, b ∈ R, then

Z Z Z
(af + bg) = a f +b g.
(ii) Additivity. If E and F are disjoint subsets of Rd , then

Z Z Z
f= f+ f.
E∪F E F
(iii) Monotonicity. If f ≤ g, then

Z Z
f≤ g.
(iv) Triangle inequality. |f | is also bounded, supported on a set of finite

measure, and
¯Z ¯ Z
¯ ¯
¯ f ¯ ≤ |f |.
¯ ¯
All these properties follow by using approximations by simple functions,

and the properties of the integral of simple functions given in Proposi-
tion 1.1.
We are now in a position to prove the first important convergence
theorem.
Theorem 1.4 (Bounded convergence theorem) Suppose that {fn }

is a sequence of measurable functions that are all bounded by M , are
supported on a set E of finite measure, and fn (x) → f (x) a.e. x as n →
∞. Then f is measurable, bounded, supported on E for a.e. x, and
Z
|fn − f | → 0 as n → ∞.
Consequently,
Z Z
fn → f as n → ∞.
Proof. From the assumptions one sees at once that f is bounded by M

almost everywhere and vanishes outside E, except possibly on a set of
measure zero. Clearly, the triangle
R inequality for the integral implies
that it suffices to prove that |fn − f | → 0 as n tends to infinity.
The proof is a reprise of the argument in Lemma 1.2. Given ² > 0, we
may find, by Egorov’s theorem, a measurable subset A² of E such that
m(E − A² ) ≤ ² and fn → f uniformly on A² . Then, we know that for
all sufficiently large n we have |fn (x) − f (x)| ≤ ² for all x ∈ A² . Putting
these facts together yields
Z Z Z
|fn (x) − f (x)| dx ≤ |fn (x) − f (x)| dx + |fn (x) − f (x)| dx
A² E−A²
≤ ²m(E) + 2M m(E − A² )
for all large n. Since ² is arbitrary, the proof of the theorem is complete.
We note that the above convergence theorem is a statement about the

interchange of an integral and a limit, since its conclusion simply says
Z Z
lim fn = lim fn .
n→∞ n→∞
A useful observation that we can make at this point is the following:

R if
f ≥ 0 is bounded and supported on a set of finite measure E and f = 0,
then f = 0 almost everywhere. Indeed, if for each integer k ≥ 1 we set

Ek = {x ∈ E : f (x) ≥ 1/k}, then the fact that k −1 χEk (x) ≤ f (x) implies
Z
−1
k m(Ek ) ≤ f,
by monotonicity Sof the integral. Thus m(Ek ) = 0 for all k, and since
∞
{x : f (x) > 0} = k=1 Ek , we see that f = 0 almost everywhere.
Return to Riemann integrable functions

We shall now show that Riemann integrable functions are also Lebesgue
integrable. When we combine this with the bounded convergence theo-
rem we have just proved, we see that Lebesgue integration resolves the
second problem in the Introduction.
Theorem 1.5 Suppose f is Riemann integrable on the closed interval
[a, b]. Then f is measurable, and
Z R Z L
f (x) dx = f (x) dx,
[a,b] [a,b]
where the integral on the left-hand side is the standard Riemann integral,
and that on the right-hand side is the Lebesgue integral.
Proof. By definition, a Riemann integrable function is bounded, say
|f (x)| ≤ M , so we need to prove that f is measurable, and then establish
the equality of integrals.
Again, by definition of Riemann integrability,1 we may construct two
sequences of step functions {ϕk } and {ψk } that satisfy the following
properties: |ϕk (x)| ≤ M and |ψk (x)| ≤ M for all x ∈ [a, b] and k ≥ 1,
ϕ1 (x) ≤ ϕ2 (x) ≤ · · · ≤ f ≤ · · · ≤ ψ2 (x) ≤ ψ1 (x),
and
Z R Z R Z R
(2) lim ϕk (x) dx = lim ψk (x)dx = f (x) dx.
k→∞ [a,b] k→∞ [a,b] [a,b]
Several observations are in order. First, it follows immediately from their

definition that for step functions the Riemann and Lebesgue integrals
agree; therefore
(3)
Z R Z L Z R Z L
ϕk (x) dx = ϕk (x) dx and ψk (x) dx = ψk (x) dx
[a,b] [a,b] [a,b] [a,b]
1 See also Section 1 of the Appendix in Book I.

for all k ≥ 1. Next, if we let
ϕ̃(x) = lim ϕk (x) and ψ̃(x) = lim ψk (x),

k→∞ k→∞
we have ϕ̃ ≤ f ≤ ψ̃. Moreover, both ϕ̃ and ψ̃ are measurable (being the

limit of step functions), and the bounded convergence theorem yields
Z L Z L
lim ϕk (x) dx = ϕ̃(x) dx
k→∞ [a,b] [a,b]
and
Z L Z L
lim ψk (x) dx = ψ̃(x) dx.
k→∞ [a,b] [a,b]
This together with (2) and (3) yields

Z L
(ψ̃(x) − ϕ̃(x)) dx = 0,
[a,b]
and since ψk − ϕk ≥ 0, we must have ψ̃ − ϕ̃ ≥ 0. By the observation

following the proof of the bounded convergence theorem, we conclude
that ψ̃ − ϕ̃ = 0 a.e., and therefore ϕ̃ = ψ̃ = f a.e., which proves that f
is measurable. Finally, since ϕk → f almost everywhere, we have (by
definition)
Z L Z L
lim ϕk (x) dx = f (x) dx,
k→∞ [a,b] [a,b]
RR RL
and by (2) and (3) we see that [a,b]
f (x) dx = [a,b]
f (x) dx, as desired.
Stage three: non-negative functions

We proceed with the integrals of functions that are measurable and non-
negative but not necessarily bounded. It will be important to allow
these functions to be extended-valued, that is, these functions may take
on the value +∞ (on a measurable set). We recall in this connection the
convention that one defines the supremum of a set of positive numbers
to be +∞ if the set is unbounded.
In the case of such a function f we define its (extended) Lebesgue
integral by
Z Z
f (x) dx = sup g(x) dx,
g
where the supremum is taken over all measurable functions g such that
0 ≤ g ≤ f , and where g is bounded and supported on a set of finite
measure.
With the above definition of the integral, there are only two possible
cases;
R the supremum is either finite, or infinite. In the first case, when
f (x) dx < ∞, we shall say that f is Lebesgue integrable or simply
integrable.
Clearly, if E is any measurable subset of Rd , and f ≥ 0, then f χE is
also positive, and we define
Z Z
f (x) dx = f (x)χE (x) dx.
E
Simple examples of functions on Rd that are integrable (or non-integrable)

are given by
½
|x|−a if |x| ≤ 1,
fa (x) =
0 if |x| > 1.
1
Fa (x) = , all x ∈ Rd .
1 + |x|a
Then fa is integrable exactly when a < d, while Fa is integrable exactly

when a > d. See the discussion following Corollary 1.10 and also Exer-
cise 10.
Proposition 1.6 The integral of non-negative measurable functions en-

joys the following properties:
(i) Linearity. If f, g ≥ 0, and a, b are positive real numbers, then
Z Z Z
(af + bg) = a f + b g.
(ii) Additivity. If E and F are disjoint subsets of Rd , and f ≥ 0, then

Z Z Z
f= f+ f.
E∪F E F
(iii) Monotonicity. If 0 ≤ f ≤ g, then

Z Z
f ≤ g.
(iv) If g is integrable and 0 ≤ f ≤ g, then f is integrable.

(v) If f is integrable, then f (x) < ∞ for almost every x.
R
(vi) If f = 0, then f (x) = 0 for almost every x.
Proof. Of the first four assertions, only (i) is not an immediate

consequence of the definitions, and to prove it we argue as follows. We
take a = b = 1 and note that if ϕ ≤ f and ψ ≤ g, where both ϕ and ψ are
bounded and supported on sets of finite measure, then ϕ + ψ ≤ f + g,
and ϕ + ψ is also bounded and supported on a set of finite measure.
Consequently
Z Z Z
f + g ≤ (f + g).
To prove the reverse inequality, suppose η is bounded and supported on a

set of finite measure, and η ≤ f + g. If we define η1 (x) = min(f (x), η(x))
and η2 = η − η1 , we note that
η1 ≤ f and η2 ≤ g.
Moreover both η1 , η2 are bounded and supported on sets of finite mea-

sure. Hence
Z Z Z Z Z Z
η = (η1 + η2 ) = η1 + η2 ≤ f + g.
Taking the supremum over η yields the required inequality.

To prove the conclusion (v) we argue as follows. Suppose Ek = {x :
f (x) ≥ k}, and E∞ = {x : f (x) = ∞}. Then
Z Z
f ≥ χEk f ≥ km(Ek ),
hence m(Ek ) → 0 as k → ∞. Since Ek & E∞ , Corollary 3.3 in the pre-

vious chapter implies that m(E∞ ) = 0.
The proof of (vi) is the same as the observation following Theorem 1.4.
We now turn our attention to some important convergence theorems

for the class of non-negative measurable functions. To motivate the re-
sults that follow, we ask the following question: R SupposeR fn ≥ 0 and
fn (x) → f (x) for almost every x. Is it true that fn dx → f dx ? Un-
fortunately, the example that follows provides a negative answer to this,
and shows that we must change our formulation of the question to obtain
a positive convergence result.
Let
½
n if 0 < x < 1/n,
fn (x) =
0 otherwise.
R
Then fn (x) → 0 for all x, yet fn (x) dx = 1 for all n. In this particular
example, the limit of the integrals is greater than the integral of the limit
function. This turns out to be the case in general, as we shall see now.
Lemma 1.7 (Fatou) Suppose {fn } is a sequence of measurable func-
tions with fn ≥ 0. If limn→∞ fn (x) = f (x) for a.e. x, then
Z Z
f ≤ lim inf fn .
n→∞
Proof. Suppose 0 ≤ g ≤ f , where g is bounded and supported on a

set E of finite measure. If we set gn (x) = min(g(x), fn (x)), then gn is
measurable, supported on E, and gn (x) → g(x) a.e., so by the bounded
convergence theorem
Z Z
gn → g.
R R
By construction, we also have gn ≤ fn , so that gn ≤ fn , and therefore
Z Z
g ≤ lim inf fn .
n→∞
Taking the supremum over all g yields the desired inequality.

R
In particular, we do not exclude the cases f = ∞, or lim inf n→∞ fn =
∞.
We can now immediately deduce the following series of corollaries.
Corollary 1.8 Suppose f is a non-negative measurable function, and
{fn } a sequence of non-negative measurable functions with fn (x) ≤ f (x)
and fn (x) → f (x) for almost every x. Then
Z Z
lim fn = f.
n→∞
R R
Proof. Since fn (x) ≤ f (x) a.e x, we necessarily have fn ≤ f for
all n; hence
Z Z
lim sup fn ≤ f.
n→∞
This inequality combined with Fatou’s lemma proves the desired limit.
In particular, we can now obtain a basic convergence theorem for the
class of non-negative measurable functions. Its statement requires the
following notation.
In analogy with the symbols % and & used to describe increasing and
decreasing sequences of sets, we shall write
fn % f
whenever {fn }∞
n=1 is a sequence of measurable functions that satisfies
fn (x) ≤ fn+1 (x) a.e x, all n ≥ 1 and lim fn (x) = f (x) a.e x.
n→∞
Similarly, we write fn & f whenever
fn (x) ≥ fn+1 (x) a.e x, all n ≥ 1 and lim fn (x) = f (x) a.e x.
n→∞
Corollary 1.9 (Monotone convergence theorem) Suppose {fn } is

a sequence of non-negative measurable functions with fn % f . Then
Z Z
lim fn = f.
n→∞
The monotone convergence theorem has the following useful conse-

quence:
P∞
Corollary 1.10 Consider a series k=1 ak (x), where ak (x) ≥ 0 is mea-
surable for every k ≥ 1. Then
Z X
∞ ∞ Z
X
ak (x) dx = ak (x) dx.
k=1 k=1
P∞ R P∞
If k=1 ak (x) dx is finite, then the series k=1 ak (x) converges for
a.e. x.
Pn P∞
Proof. Let fn (x) = k=1 ak (x) and f (x) = k=1 ak (x). The func-
tions fn are measurable, fn (x) ≤ fn+1 (x), and fn (x) → f (x) as n tends
to infinity. Since
Z n Z
X
fn = ak (x) dx,
k=1
the monotone convergence theorem implies

∞ Z
X Z X
∞
ak (x) dx = ak (x) dx.
k=1 k=1
PR P∞
If ak < ∞, then the above implies that ak (x) is integrable,
k=1P
∞
and by our earlier observation, we conclude that k=1 ak (x) is finite
almost everywhere.
We give two nice illustrations of this last corollary.
The first consists of another proof of the Borel-Cantelli lemma (see
Exercise 16, Chapter 1), which
P says that if E1 , E2 , . . . is a collection
of measurable subsets with m(Ek ) < ∞, then the set of points that
belong to infinitely many sets Ek has measure zero. To prove this fact,
we let
ak (x) = χEk (x),
andPnote that a point x belongs to infinitely

P many sets Ek if and only
∞
if
P∞ k=1 R a k (x) = ∞. Our assumption on m(Ek ) Psays precisely that
∞
k=1 a k (x) dx < ∞, and the corollary implies that k=1 ak (x) is finite
except possibly on a set of measure zero, and thus the Borel-Cantelli
lemma is proved.
The second illustration will be useful in our discussion of approxima-
tions to the identity in Chapter 3. Consider the function
½ 1
|x|d+1
if x 6= 0,
f (x) =
0 otherwise.
We prove that f is integrable outside any ball, |x| ≥ ², and moreover

Z
C
f (x) dx ≤ , for some constant C > 0.
|x|≥² ²
Indeed, if we let Ak = {x ∈ Rd : 2k ² < |x| ≤ 2k+1 ²}, and define
∞
X 1
g(x) = ak (x) where ak (x) = χAk (x),
(2k ²)d+1
k=0
R R
then we must have f (x) ≤ g(x), and hence f ≤ g. Since the set Ak
is obtained from A = {1 < |x| < 2} by a dilation of factor 2k ², we have
by the relative dilation-invariance properties of the Lebesgue measure,

that m(Ak ) = (2k ²)d m(A). Also by Corollary 1.10, we see that
Z ∞
X ∞
X
m(Ak ) (2k ²)d C
g= k d+1
= m(A) k d+1
= ,
(2 ²) (2 ²) ²
k=0 k=0
where C = 2m(A). Note that the same dilation-invariance property in

fact shows that
Z Z
dx 1 dx
d+1
= d+1
.
|x|≥² |x| ² |x|≥1 |x|
See also the identity (7) below.
Stage four: general case

If f is any real-valued measurable function on Rd , we say that f is
Lebesgue integrable (or just integrable) if the non-negative measur-
able function |f | is integrable in the sense of the previous section.
If f is Lebesgue integrable, we give a meaning to its integral as follows.
First, we may define
f + (x) = max(f (x), 0) and f − (x) = max(−f (x), 0),
so that both f + and f − are non-negative and f + − f − = f . Since f ± ≤

|f |, both functions f + and f − are integrable whenever f is, and we then
define the Lebesgue integral of f by
Z Z Z
f= f+ − f −.
In practice one encounters many decompositions f = f1 − f2 , where

f1 , f2 are both non-negative integrable functions, and one would expect
that regardless of the decomposition of f , we always have
Z Z Z
f= f1 − f2 .
In other words, the definition of the integral should be independent of the

decomposition f = f1 − f2 . To see why this is so, suppose f = g1 − g2
is another decomposition where both g1 and g2 are non-negative and
integrable. Since f1 − f2 = g1 − g2 we have f1 + g2 = g1 + f2 ; but both
sides of this last identity consist of positive measurable functions, so the

linearity of the integral in this case yields
Z Z Z Z
f1 + g2 = g1 + f2 .
Since all integrals involved are finite, we find the desired result
Z Z Z Z
f1 − f2 = g1 − g2 .
In considering the above definitions it is useful to keep in mind the

following small observations. Both the integrability of f , and the value
of its integral are unchanged if we modify f arbitrarily on a set of measure
zero. It is therefore useful to adopt the convention that in the context
of integration we allow our functions to be undefined on sets of measure
zero. Moreover, if f is integrable, then by (v) of Proposition 1.6, it is
finite-valued almost everywhere. Thus, availing ourselves of the above
convention, we can always add two integrable functions f and g, since
the ambiguity of f + g, due to the extended values of each, resides in a
set of measure zero. Moreover, we note that when speaking of a function
f , we are, in effect, also speaking about the collection of all functions
that equal f almost everywhere.
Simple applications of the definition and the properties proved previ-
ously yield all the elementary properties of the integral:
Proposition 1.11 The integral of Lebesgue integrable functions is lin-

ear, additive, monotonic, and satisfies the triangle inequality.
We now gather two results which, although instructive in their own

right, are also needed in the proof of the next theorem.
Proposition 1.12 Suppose f is integrable on Rd . Then for every ² > 0:
(i) There exists a set of finite measure B (a ball, for example) such
that
Z
|f | < ².
Bc
(ii) There is a δ > 0 such that

Z
|f | < ² whenever m(E) < δ.
E
The last condition is known as absolute continuity.

Proof. By replacing f with |f | we may assume without loss of gener-
ality that f ≥ 0.
For the first part, let BN denote the ball of radius N centered at the
origin, and note that if fN (x) = f (x)χBN (x), then fN ≥ 0 is measur-
able, fN (x) ≤ fN +1 (x), and limN →∞ fN (x) = f (x). By the monotone
convergence theorem, we must have
Z Z
lim fN = f.
N →∞
In particular, for some large N ,

Z Z
0 ≤ f − f χBN < ²,
R
and since 1 − χBN = χBN c , this implies c f < ², as we set out to prove.
BN
For the second part, assuming again that f ≥ 0, we let fN (x) = f (x)χEN
where
EN = {x : f (x) ≤ N }.
Once again, fN ≥ 0 is measurable, fN (x) ≤ fN +1 (x), and given ² > 0

there exists (by the monotone convergence theorem) an integer N > 0
such that
Z
²
(f − fN ) < .
2
We now pick δ > 0 so that N δ < ²/2. If m(E) < δ, then

Z Z Z
f= (f − fN ) + fN
E E E
Z Z
≤ (f − fN ) + fN
E
Z
≤ (f − fN ) + N m(E)
² ²
≤ + = ².
2 2
This concludes the proof of the proposition.
Intuitively, integrable functions should in some sense vanish at infinity
since their integrals are finite, and the first part of the proposition at-
taches a precise meaning to this intuition. One should observe, however,
that integrability need not guarantee the more naive pointwise vanishing
as |x| becomes large. See Exercise 6.
We are now ready to prove a cornerstone of the theory of Lebesgue
integration, the dominated convergence theorem. It can be viewed as a
culmination of our efforts, and is a general statement about the interplay
between limits and integrals.
Theorem 1.13 Suppose {fn } is a sequence of measurable functions such

that fn (x) → f (x) a.e. x, as n tends to infinity. If |fn (x)| ≤ g(x), where
g is integrable, then
Z
|fn − f | → 0 as n → ∞,
and consequently
Z Z
fn → f as n → ∞.
Proof. For each N ≥ 0 let EN = {x : |x| ≤ N, g(x) ≤ N }. Given

² > 0, we may argue as in the
R first part of the previous lemma, to see
that there exists N so that E c g < ². Then the functions fn χEN are
N
bounded (by N ) and supported on a set of finite measure, so that by the
bounded convergence theorem, we have
Z
|fn − f | < ², for all large n.
EN
Hence, we obtain the estimate

Z Z Z
|fn − f | = |fn − f | + |fn − f |
EN c
EN
Z Z
≤ |fn − f | + 2 g
EN c
EN
≤ ² + 2² = 3²
for all large n. This proves the theorem.
Complex-valued functions
If f is a complex-valued function on Rd , we may write it as
f (x) = u(x) + iv(x),

where u and v are real-valued functions called the real and imaginary
parts of f , respectively. The function f is measurable if and only if both u
and v are measurable. We then say that f is Lebesgue integrable if the
function |f (x)| = (u(x)2 + v(x)2 )1/2 (which is non-negative) is Lebesgue
integrable in the sense defined previously.
It is clear that
|u(x)| ≤ |f (x)| and |v(x)| ≤ |f (x)|.
Also, if a, b ≥ 0, one has (a + b)1/2 ≤ a1/2 + b1/2 , so that
|f (x)| ≤ |u(x)| + |v(x)|.
As a result of these simple inequalities, we deduce that a complex-valued

function is integrable if and only if both its real and imaginary parts are
integrable. Then, the Lebesgue integral of f is defined by
Z Z Z
f (x) dx = u(x) dx + i v(x) dx.
Finally, if E is a measurable subset of Rd , and f is a complex-valued

measurable function on E, we say that Rf is Lebesgue
R integrable on E if
f χE is integrable on Rd , and we define E f = f χE .
The collection of all complex-valued integrable functions on a mea-

surable subset E ⊂ Rd forms a vector space over C. Indeed, if f and g
are integrable, then so is f + g, since the triangle inequality gives |(f +
g)(x)| ≤ |f (x)| + |g(x)|, and monotonicity of the integral then yields
Z Z Z
|f + g| ≤ |f | + |g| < ∞.
E E E
Also, it is clear that if a ∈ C and if f is integrable, then so is af . Finally,

the integral continues to be linear over C.
2 The space L1 of integrable functions

The fact that the integrable functions form a vector space is an impor-
tant observation about the algebraic properties of such functions. A
fundamental analytic fact is that this vector space is complete in the
appropriate norm.
2. The space L1 of integrable functions 69
For any integrable function f on Rd we define the norm2 of f ,

Z
kf k = kf kL1 = kf kL1 (Rd ) = |f (x)| dx.
Rd
The collection of all integrable functions with the above norm gives a
(somewhat imprecise) definition of the space L1 (Rd ). We also note that
kf k = 0 if and only if f = 0 almost everywhere (see Proposition 1.6),
and this simple property of the norm reflects the practice we have al-
ready adopted not to distinguish two functions that agree almost every-
where. With this in mind, we take the precise definition of L1 (Rd ) to be
the space of equivalence classes of integrable functions, where we define
two functions to be equivalent if they agree almost everywhere. Often,
however, it is convenient to retain the (imprecise) terminology that an
element f ∈ L1 (Rd ) is an integrable function, even though it is only an
equivalence class of such functions. Note that by the above, the norm
kf k of an element f ∈ L1 (Rd ) is well-defined by the choice of any inte-
grable function in its equivalence class. Moreover, L1 (Rd ) inherits the
property that it is a vector space. This and other straightforward facts
are summarized in the following proposition.
Proposition 2.1 Suppose f and g are two functions in L1 (Rd ).
(i) kaf kL1 (Rd ) = |a| kf kL1 (Rd ) for all a ∈ C.
(ii) kf + gkL1 (Rd ) ≤ kf kL1 (Rd ) + kgkL1 (Rd ) .
(iii) kf kL1 (Rd ) = 0 if and only if f = 0 a.e.
(iv) d(f, g) = kf − gkL1 (Rd ) defines a metric on L1 (Rd ).
In (iv), we mean that d satisfies the following conditions. First, d(f, g) ≥
0 for all integrable functions f and g, and d(f, g) = 0 if and only if f = g
a.e. Also, d(f, g) = d(g, f ), and finally, d satisfies the triangle inequality
d(f, g) ≤ d(f, h) + d(h, g), for all f, g, h ∈ L1 (Rd ).
A space V with a metric d is said to be complete if for every Cauchy
sequence {xk } in V (that is, d(xk , x` ) → 0 as k, ` → ∞) there exists
x ∈ V such that limk→∞ xk = x in the sense that
d(xk , x) → 0, as k → ∞.
Our main goal of completing the space of Riemann integrable functions
will be attained once we have established the next important theorem.
2 In this chapter the only norm we consider is the L1 -norm, so we often write kf k for
kf kL1 . Later, we shall have occasion to consider other norms, and then we shall modify
our notation accordingly.
Theorem 2.2 (Riesz-Fischer) The vector space L1 is complete in its

metric.
Proof. Suppose {fn } is a Cauchy sequence in the norm, so that kfn −

fm k → 0 as n, m → ∞. The plan of the proof is to extract a subsequence
of {fn } that converges to f , both pointwise almost everywhere and in
the norm.
Under ideal circumstances we would have that the sequence {fn } con-
verges almost everywhere to a limit f , and we would then prove that the
sequence converges to f also in the norm. Unfortunately, almost every-
where convergence does not hold for general Cauchy sequences (see Exer-
cise 12). The main point, however, is that if the convergence in the norm
is rapid enough, then almost everywhere convergence is a consequence,
and this can be achieved by dealing with an appropriate subsequence of
the original sequence.
Indeed, consider a subsequence {fnk }∞ k=1 of {fn } with the following
property:
kfnk+1 − fnk k ≤ 2−k , for all k ≥ 1.
The existence of such a subsequence is guaranteed by the fact that kfn −

fm k ≤ ² whenever n, m ≥ N (²), so that it suffices to take nk = N (2−k ).
We now consider the series whose convergence will be seen below,
∞
X
f (x) = fn1 (x) + (fnk+1 (x) − fnk (x))
k=1
and
∞
X
g(x) = |fn1 (x)| + |fnk+1 (x) − fnk (x)|,
k=1
and note that

Z ∞ Z
X Z ∞
X
|fn1 | + |fnk+1 − fnk | ≤ |fn1 | + 2−k < ∞.
k=1 k=1
So the monotone convergence theorem implies that g is integrable, and

since |f | ≤ g, hence so is f . In particular, the series defining f converges
almost everywhere, and since the partial sums of this series are precisely
the fnk (by construction of the telescopic series), we find that
fnk (x) → f (x) a.e. x.

To prove that fnk → f in L1 as well, we simply observe that |f − fnk | ≤ g

for all k, and apply the dominated convergence theorem to get kfnk −
f kL1 → 0 as k tends to infinity.
Finally, the last step of the proof consists in recalling that {fn } is
Cauchy. Given ², there exists N such that for all n, m > N we have
kfn − fm k < ²/2. If nk is chosen so that nk > N , and kfnk − f k < ²/2,
then the triangle inequality implies
kfn − f k ≤ kfn − fnk k + kfnk − f k < ²
whenever n > N . Thus {fn } has the limit f in L1 , and the proof of the
theorem is complete.
Since every sequence that converges in the norm is a Cauchy sequence
in that norm, the argument in the proof of the theorem yields the fol-
lowing.
Corollary 2.3 If {fn }∞ 1

n=1 converges to f in L , then there exists a sub-
sequence {fnk }∞
k=1 such that
fnk (x) → f (x) a.e. x.
We say that a family G of integrable functions is dense in L1 if for any

f ∈ L1 and ² > 0, there exists g ∈ G so that kf − gkL1 < ². Fortunately
we are familiar with many families that are dense in L1 , and we describe
some in the theorem that follows. These are useful when one is faced
with the problem of proving some fact or identity involving integrable
functions. In this situation a general principle applies: the result is often
easier to prove for a more restrictive class of functions (like the ones in
the theorem below), and then a density (or limiting) argument yields the
result in general.
Theorem 2.4 The following families of functions are dense in L1 (Rd ):
(i) The simple functions.

(ii) The step functions.
(iii) The continuous functions of compact support.
Proof. Let f be an integrable function on Rd . First, we may assume

that f is real-valued, because we may approximate its real and imaginary
parts independently. If this is the case, we may then write f = f + − f − ,
where f + , f − ≥ 0, and it now suffices to prove the theorem when f ≥ 0.
For (i), Theorem 4.1 in Chapter 1 guarantees the existence of a se-

quence {ϕk } of non-negative simple functions that increase to f point-
wise. By the dominated convergence theorem (or even simply the mono-
tone convergence theorem) we then have
kf − ϕk kL1 → 0 as k → ∞.
Thus there are simple functions that are arbitrarily close to f in the L1
norm.
For (ii), we first note that by (i) it suffices to approximate simple
functions by step functions. Then, we recall that a simple function is
a finite linear combination of characteristic functions of sets of finite
measure, so it suffices to show that if E is such a set, then there is a
step function ψ so that kχE − ψkL1 is small. However, we now recall
that this argument was already carried out in the proof of Theorem 4.3,
Chapter 1. Indeed, there it is shownSthat there is an almost disjoint
M
family of rectangles {Rj } with m(E4 j=1 Rj ) ≤ 2². Thus χE and ψ =
P
j χRj differ at most on a set of measure 2², and as a result we find
that kχE − ψkL1 < 2².
By (ii), it suffices to establish (iii) when f is the characteristic function
of a rectangle. In the one-dimensional case, where f is the characteristic
function of an interval [a, b], we may choose a continuous piecewise linear
function g defined by
½
1 if a ≤ x ≤ b,
g(x) =
0 if x ≤ a − ² or x ≥ b + ²,
and with g linear on the intervals [a − ², a] and [b, b + ²]. Then kf −

gkL1 < 2². In d dimensions, it suffices to note that the characteristic
function of a rectangle is the product of characteristic functions of inter-
vals. Then, the desired continuous function of compact support is simply
the product of functions like g defined above.
The results above for L1 (Rd ) lead immediately to an extension in which
Rd can be replaced by any fixed subset E of positive measure. In fact
if E is such a subset, we can define L1 (E) and carry out the arguments
that are analogous to L1 (Rd ). Better yet, we can proceed by extending
any function f on E by setting f˜ = f on E and f˜ = 0 on E c , and defining
kf kL1 (E) = kf˜kL1 (Rd ) . The analogues of Proposition 2.1 and Theorem 2.2
then hold for the space L1 (E).
Invariance Properties
If f is a function defined on Rd , the translation of f by a vector h ∈ Rd
is the function fh , defined by fh (x) = f (x − h). Here we want to examine
some basic aspects of translations of integrable functions.
First, there is the translation-invariance of the integral. One way to
state this is as follows: if f is an integrable function, then so is fh and
Z Z
(4) f (x − h) dx = f (x) dx.
Rd Rd
We check this assertion first when f = χE , the characteristic function

of a measurable set E. Then obviously fh = χEh , where Eh = {x + h :
x ∈ E}, and thus the assertion follows because m(Eh ) = m(E) (see Sec-
tion 3 in Chapter 1). As a result of linearity, the identity (4) holds for
all simple functions. Now if f is non-negative and {ϕn } is a sequence of
simple functions that increase pointwise a.e to f (such a sequence exists
by Theorem 4.1 in the previous chapter), then {(ϕn )h } is a sequence of
simple functions that increase to fh pointwise a.e, and the monotone con-
vergence theorem implies (4) in thisR special case. Thus,R if f is complex-
valued and integrable we see that Rd |f (x − h)| dx = Rd |f (x)| dx, which
shows that fh ∈ L1 (Rd ) and also kfh k = kf k. From the definitions, we
then conclude that (4) holds whenever f ∈ L1 .
Incidentally, using the relative invariance of Lebesgue measure under
dilations and reflections (Section 3, Chapter 1) one can prove in the same
way that if f (x) is integrable, so is f (δx), δ > 0, and f (−x), and
(5) Z Z Z Z
δd f (δx) dx = f (x) dx, while f (−x) dx = f (x) dx.
Rd Rd Rd Rd
We digress to record for later use two useful consequences of the above
invariance properties:
(i) Suppose that f and g are a pair of measurable functions on Rd so
that for some fixed x ∈ Rd the function y 7→ f (x − y)g(y) is integrable.
As a consequence, the function y 7→ f (y)g(x − y) is then also integrable
and we have
Z Z
(6) f (x − y)g(y) dy = f (y)g(x − y) dy.
Rd Rd
This follows from (4) and (5) on making the change of variables which
replaces y by x − y, and noting that this change is a combination of a
translation and a reflection.
The integral on the left-hand side is denoted by (f ∗ g)(x) and is de-

fined as the convolution of f and g. Thus (6) asserts the commutativity
of the convolution product.
(ii) Using (5) one has that for all ² > 0

Z Z
dx −a+d dx
(7) a
=² a
whenever a > d,
|x|≥² |x| |x|≥1 |x|
and
Z Z
dx dx
(8) = ²−a+d whenever a < d.
|x|≤²|x|a |x|≤1 |x| a
R dx
R dx
It can also be seen that the integrals |x|≥1 |x| a and |x|≤1 |x|a
(respec-
tively, when a > d and a < d) are finite by the argument that appears
after Corollary 1.10.
Translations and continuity

We shall next examine how continuity properties of f are related to the
way the translations fh vary with h. Note that for any given x ∈ Rd , the
statement that fh (x) → f (x) as h → 0 is the same as the continuity of
f at the point x.
However, a general f which is integrable may be discontinuous at ev-
ery x, even when corrected on a set of measure zero; see Exercise 15.
Nevertheless, there is an overall continuity that an arbitrary f ∈ L1 (Rd )
enjoys, one that holds in the norm.
Proposition 2.5 Suppose f ∈ L1 (Rd ). Then
kfh − f kL1 → 0 as h → 0.
The proof is a simple consequence of the approximation of integrable
functions by continuous functions of compact support as given in The-
orem 2.4. In fact for any ² > 0, we can find such a function g so that
kf − gk < ². Now
fh − f = (gh − g) + (fh − gh ) − (f − g).
However, kfh − gh k = kf − gk < ², while since g is continuous and has
compact support we have that clearly
Z
kgh − gk = |g(x − h) − g(x)| dx → 0 as h → 0.
Rd
So if |h| < δ, where δ is sufficiently small, then kgh − gk < ², and as a

result kfh − f k < 3², whenever |h| < δ.
3. Fubini’s theorem 75
3 Fubini’s theorem
In elementary calculus integrals of continuous functions of several vari-
ables are often calculated by iterating one-dimensional integrals. We
shall now examine this important analytic device from the general point
of view of Lebesgue integration in Rd , and we shall see that a number of
interesting issues arise.
In general, we may write Rd as a product
Rd = Rd1 × Rd2 where d = d1 + d2 , and d1 , d2 ≥ 1.
A point in Rd then takes the form (x, y), where x ∈ Rd1 and y ∈ Rd2 .
With such a decomposition of Rd in mind, the general notion of a slice,
formed by fixing one variable, becomes natural. If f is a function in
Rd1 × Rd2 , the slice of f corresponding to y ∈ Rd2 is the function f y of
the x ∈ Rd1 variable, given by
f y (x) = f (x, y).
Similarly, the slice of f for a fixed x ∈ Rd1 is fx (y) = f (x, y).
In the case of a set E ⊂ Rd1 × Rd2 we define its slices by
E y = {x ∈ Rd1 : (x, y) ∈ E} and Ex = {y ∈ Rd2 : (x, y) ∈ E}.
See Figure 1 for an illustration.
Rd2
Ey
y
Ex
x Rd1
Figure 1. Slices E y and Ex (for fixed x and y) of a set E
3.1 Statement and proof of the theorem

That the theorem that follows is not entirely straightforward is clear
from the first difficulty that arises in its formulation, involving the mea-
surability of the functions and sets in question. In fact, even with the
assumption that f is measurable on Rd , it is not necessarily true that

the slice f y is measurable on Rd1 for each y; nor does the corresponding
assertion necessarily hold for a measurable set: the slice E y may not
be measurable for each y. An easy example arises in R2 by placing a
one-dimensional non-measurable set on the x-axis; the set E in R2 has
measure zero, but E y is not measurable for y = 0. What saves us is that,
nevertheless, measurability holds for almost all slices.
The main theorem is as follows. We recall that by definition all inte-
grable functions are measurable.
Theorem 3.1 Suppose f (x, y) is integrable on Rd1 × Rd2 . Then for al-
most every y ∈ Rd2 :
(i) The slice f y is integrable on Rd1 .

R
(ii) The function defined by Rd1 f y (x) dx is integrable on Rd2 .
Moreover:
Z µZ ¶ Z
(iii) f (x, y) dx dy = f.
R d2 R d1 Rd
Clearly, the theorem is symmetric in x and y so that we also may

R conclude
that the slice fx is integrable on Rd2 for a.e. x. Moreover, Rd2 fx (y) dy
is integrable, and
Z µZ ¶ Z
f (x, y) dy dx = f.
R d1 R d2 Rd
In particular, Fubini’s theorem states that the integral of f on Rd can

be computed by iterating lower-dimensional integrals, and that the iter-
ations can be taken in any order
Z µZ ¶ Z µZ ¶ Z
f (x, y) dx dy = f (x, y) dy dx = f.
R d2 R d1 R d1 R d2 Rd
We first note that we may assume that f is real-valued, since the

theorem then applies to the real and imaginary parts of a complex-valued
function. The proof of Fubini’s theorem which we give next consists of a
sequence of six steps. We begin by letting F denote the set of integrable
functions on Rd which satisfy all three conclusions in the theorem, and
set out to prove that L1 (Rd ) ⊂ F .
We proceed by first showing that F is closed under operations such
as linear combinations (Step 1) and limits (Step 2). Then we begin to
construct families of functions in F. Since any integrable function is the

“limit” of simple functions, and simple functions are themselves linear
combinations of sets of finite measure, the goal quickly becomes to prove
that χE belongs to F whenever E is a measurable subset of Rd with
finite measure. To achieve this goal, we begin with rectangles and work
our way up to sets of type Gδ (Step 3), and sets of measure zero (Step 4).
Finally, a limiting argument shows that all integrable functions are in F.
This will complete the proof of Fubini’s theorem.
Step 1. Any finite linear combination of functions in F also belongs

to F.
Indeed, let {fk }N
k=1 ⊂ F. For each k there exists a set Ak ⊂ R
d2
of
y d1
measure 0 so that fk is integrable on R whenever y ∈ / Ak . Then, if
SN
A = k=1 Ak , the set A has measure 0, and in the complement of A,
the y-slice corresponding to any finite linear combination of the fk is
measurable, and also integrable. By linearity of the integral, we then
conclude that any linear combination of the fk ’s belongs to F.
Step 2. Suppose {fk } is a sequence of measurable functions in F so
that fk % f or fk & f , where f is integrable (on Rd ). Then f ∈ F.
By taking −fk instead of fk if necessary, we note that it suffices to
consider the case of an increasing sequence. Also, we may replace fk
by fk − f1 and assume that the fk ’s are non-negative. Now, we observe
that an application of the monotone convergence theorem (Corollary 1.9)
yields
Z Z
(9) lim fk (x, y) dx dy = f (x, y) dx dy.
k→∞ Rd Rd
d2 y
By assumption, for each k there exists a set S∞Ak ⊂ R , so that fk is
d1
integrable on R whenever y ∈ / Ak . If A = k=1 Ak , then m(A) = 0 in
/ A, then fky is integrable on Rd1 for all k, and, by the
Rd2 , and if y ∈
monotone convergence theorem, we find that
Z Z
gk (y) = fky (x) dx increases to a limit g(y) = f y (x) dx
R d1 R d1
as k tends to infinity. By assumption, each gk (y) is integrable, so that

another application of the monotone convergence theorem yields
Z Z
(10) gk (y) dy → g(y) dy as k → ∞.
Rd2 R d2
By the assumption that fk ∈ F we have

Z Z
gk (y) dy = fk (x, y) dx dy,
Rd2 Rd
and combining this fact with (9) and (10), we conclude that
Z Z
g(y) dy = f (x, y) dx dy.
Rd2 Rd
Since f is integrable, the right-hand integral is finite, and this proves that
g is integrable. Consequently g(y) < ∞ a.e. y, hence f y is integrable for
a.e. y, and
Z µZ ¶ Z
f (x, y) dx dy = f (x, y) dx dy.
Rd2 R d1 Rd
This proves that f ∈ F as desired.

Step 3. Any characteristic function of a set E that is a Gδ and of finite
measure belongs to F.
We proceed in stages of increasing order of generality.
(a) First suppose E is a bounded open cube in Rd , such that E = Q1 ×
Q2 , where Q1 and Q2 are open cubes in Rd1 and Rd2 , respectively. Then,
for each y the function χE (x, y) is measurable in x, and integrable with
Z ½
|Q1 | if y ∈ Q2 ,
g(y) = χE (x, y) dx
Rd1 0 otherwise.
Consequently, g = |Q1 |χQ2 is also measurable and integrable, with

Z
g(y) dy = |Q1 | |Q2 |.
Rd2
R
Since we initially have Rd
χE (x, y) dx dy = |E| = |Q1 | |Q2 |, we deduce
that χE ∈ F.
(b) Now suppose E is a subset of the boundary of some closed cube.
Then,
R since the boundary of a cube has measure 0 in Rd , we have
χ (x, y) dx dy = 0.
Rd E
Next, we note, after an investigation of the various possibilities, that
y d1
for almost
R every y, the slice E has measure 0 in R , and therefore if
g(y)
R = Rd1 χE (x, y) dx we have g(y) = 0 for a.e. y. As a consequence,
d
R 2
g(y) dy = 0, and therefore χE ∈ F.
(c) Suppose now E is a finite union of closed cubes whose interiors are
SK
disjoint, E = k=1 Qk . Then, if Q̃k denotes the interior of Qk , we may
write χE as a linear combination of the χQ̃k and χAk where Ak is a
subset of the boundary of Qk for k = 1, . . . , K. By our previous analysis,
we know that χQk and χAk belong to F for all k, and since Step 1
guarantees that F is closed under finite linear combinations, we conclude
that χE ∈ F, as desired.
(d) Next, we prove that if E is open and of finite measure, then χE ∈
F. This follows from taking a limit in the previous case. Indeed, by
Theorem 1.4 in Chapter 1, we may write E as a countable union of
almost disjoint closed cubes
∞
[
E= Qj .
j=1
Pk
Consequently, if we let fk = j=1 χQj , then we note that the functions
fk increase to f = χE , which is integrable since m(E) is finite. Therefore,
we may conclude by Step 2 that f ∈ F.
(e) Finally, if E is a Gδ of finite measure, then χE ∈ F. Indeed, by
definition, there exist open sets Õ1 , Õ2 , . . ., such that
∞
\
E= Õk .
k=1
Since E has finite measure, there exists an open set Õ0 of finite measure
with E ⊂ Õ0 . If we let
k
\
Ok = O0 ∩ Õj ,
j=1
then we note that we have a decreasing sequence of open sets of finite

measure O1 ⊃ O2 ⊃ · · · with
∞
\
E= Ok .
k=1
Therefore, the sequence of functions fk = χOk decreases to f = χE , and

since χOk ∈ F for all k by (d) above, we conclude by Step 2 that χE
belongs to F.
Step 4. If E has measure 0, then χE belongs to F.
Indeed, since E is measurable, we may choose a set G of type Gδ with

E ⊂ G and m(G) = 0 (Corollary 3.5, Chapter 1). Since χG ∈ F (by the
previous step) we find that
Z µZ ¶ Z
χG (x, y) dx dy = χG = 0.
R d2 Rd1 Rd
Therefore
Z
χG (x, y) dx = 0 for a.e. y.
Rd1
Consequently, the slice Gy has measure 0 for a.e. y. The simple obser-
vation
R that E y ⊂ Gy then shows that E y has measure 0 for a.e. y, and
χ (x, y) dx = 0 for a.e. y. Therefore,
Rd1 E
Z µZ ¶ Z
χE (x, y) dx dy = 0 = χE ,
Rd2 Rd1 Rd
and thus χE ∈ F, as was to be shown.

Step 5. If E is any measurable subset of Rd with finite measure, then
χE belongs to F.
To prove this, recall first that there exists a set of finite measure G of
type Gδ , with E ⊂ G and m(G − E) = 0. Since
χE = χG − χG−E ,
and F is closed under linear combinations, we find that χE ∈ F, as

desired.
Step 6. This is the final step, which consists of proving that if f is
integrable, then f ∈ F.
We note first that f has the decomposition f = f + − f − , where both f +
and f − are non-negative and integrable, so by Step 1 we may assume
that f is itself non-negative. By Theorem 4.1 in the previous chapter,
there exists a sequence {ϕk } of simple functions that increase to f . Since
each ϕk is a finite linear combination of characteristic functions of sets
with finite measure, we have ϕk ∈ F by Steps 5 and 1, hence f ∈ F by
Step 2.
3.2 Applications of Fubini’s theorem

Theorem 3.2 Suppose f (x, y) is a non-negative measurable function on
Rd1 × Rd2 . Then for almost every y ∈ Rd2 :
(i) The slice f y is measurable on Rd1 .

R
(ii) The function defined by Rd1 f y (x) dx is measurable on Rd2 .
Moreover:
Z µZ ¶ Z
(iii) f (x, y) dx dy = f (x, y) dx dy in the extended sense.
R d2 Rd1 Rd
In practice, this theorem is often used in conjunction with Fubini’s

theorem.3 Indeed, supposeR we are given a measurable function f on Rd
and asked to compute Rd f . To justify the use of iterated integration, we
first apply the present theorem to |f |. Using it, we may freely compute
(or estimate) the iterated integrals of the non-negative function |f |. If
these
R are finite, Theorem 3.2 guarantees that f is integrable, that is,
|f | < ∞. Then the hypothesis in Fubini’s theorem is verified, and we
may use that theorem in the calculation of the integral of f .
Proof of Theorem 3.2. Consider the truncations
½
f (x, y) if |(x, y)| < k and f (x, y) < k,
fk (x, y) =
0 otherwise.
Each fk is integrable, and by part (i) in Fubini’s theorem there exists a

y
set Ek ⊂ Rd2 of measure 0 such S that the slice fk (x)y is measurable for all
c
y ∈ Ek . Then, if we set E = k Ek , we find that f (x) is measurable for
all y ∈ E c and all k. Moreover, m(E) = 0. Since fky % f y , the monotone
convergence theorem implies that if y ∈ / E, then
Z Z
fk (x, y) dx % f (x, y) dx as k → ∞.
Rd1 R d1
R
Again by Fubini’s
R theorem, Rd1 fk (x, y) dx is measurable for all y ∈ E c ,
hence so is Rd1 f (x, y) dx. Another application of the monotone conver-
gence theorem then gives
Z µZ ¶ Z µZ ¶
(11) fk (x, y) dx dy → f (x, y) dx dy.
Rd2 Rd1 Rd2 Rd1
By part (iii) in Fubini’s theorem we know that

Z µZ ¶ Z
(12) fk (x, y) dx dy = fk .
Rd2 R d1 Rd
3 Theorem 3.2 was formulated by Tonelli. We will, however, use the short-hand of
referring to it, as well as Theorem 3.1 and Corollary 3.3, as Fubini’s theorem.
A final application of the monotone convergence theorem directly to fk

also gives
Z Z
(13) fk → f.
Rd Rd
Combining (11), (12), and (13) completes the proof of Theorem 3.2.
Corollary 3.3 If E is a measurable set in Rd1 × Rd2 , then for almost

every y ∈ Rd2 the slice
E y = {x ∈ Rd1 : (x, y) ∈ E}
is a measurable subset of Rd1 . Moreover m(E y ) is a measurable function

of y and
Z
m(E) = m(E y ) dy.
Rd2
This is an immediate consequence of the first part of Theorem 3.2 applied

to the function χE . Clearly a symmetric result holds for the x-slices in
Rd2 .
We have thus established the basic fact that if E is measurable on
Rd1 × Rd2 , then for almost every y ∈ Rd2 the slice E y is measurable in
Rd1 (and also the symmetric statement with the roles of x and y inter-
changed). One might be tempted to think that the converse assertion
holds. To see that this is not the case, note that if we let N denote a
non-measurable subset of R, and then define
E = [0, 1] × N ⊂ R × R,
we see that
½
y [0, 1] if y ∈ N ,
E =
∅ if y ∈
/ N.
Thus E y is measurable for every y. However, if E were measurable, then

the corollary would imply that Ex = {y ∈ R : (x, y) ∈ E} is measurable
for almost every x ∈ R, which is not true since Ex is equal to N for all
x ∈ [0, 1].
A more striking example is that of a set E in the unit square [0, 1] ×
[0, 1] that is not measurable, and yet the slices E y and Ex are measurable
with m(E y ) = 0 and m(Ex ) = 1 for each x, y ∈ [0, 1]. The construction
of E is based on the existence of a highly paradoxical ordering ≺ of
the reals, with the property that {x : x ≺ y} is a countable set for each
y ∈ R. (The construction of this ordering is discussed in Problem 5.)
Given this ordering we let
E = {(x, y) ∈ [0, 1] × [0, 1], with x ≺ y}.
Note that for each y ∈ [0, 1], E y = {x : x ≺ y}; thus E y is countable and
m(E y ) = 0. Similarly m(Ex ) = 1, because Ex is the complement of a
denumerable set in [0, 1]. If E were measurable, it would contradict the
formula in Corollary 3.3.
In relating a set E to its slices Ex and E y , matters are straightforward
for the basic sets which arise when we consider Rd as the product Rd1 ×
Rd2 . These are the product sets E = E1 × E2 , where Ej ⊂ Rdj .
Proposition 3.4 If E = E1 × E2 is a measurable subset of Rd , and

m∗ (E2 ) > 0, then E1 is measurable.
Proof. By Corollary 3.3, we know that for a.e. y ∈ Rd2 , the slice
function
(χE1 ×E2 )y (x) = χE1 (x)χE2 (y)
is measurable as a function of x. In fact, we claim that there is some

y ∈ E2 such that the above slice function is measurable in x; for such a
y we would have χE1 ×E2 (x, y) = χE1 (x), and this would imply that E1
is measurable.
To prove the existence of such a y, we use the assumption that m∗ (E2 ) >
0. Indeed, let F denote the set of y ∈ Rd2 such that the slice E y is
measurable. Then m(F c ) = 0 (by the previous corollary). However,
E2 ∩ F is not empty
S because m∗ (E2 ∩ F ) > 0. To see this, note that
E2 = (E2 ∩ F ) (E2 ∩ F c ), hence
0 < m∗ (E2 ) ≤ m∗ (E2 ∩ F ) + m∗ (E2 ∩ F c ) = m∗ (E2 ∩ F ),
because E2 ∩ F c is a subset of a set of measure zero.

To deal with a converse of the above result, we need the following
lemma.
Lemma 3.5 If E1 ⊂ Rd1 and E2 ⊂ Rd2 , then
m∗ (E1 × E2 ) ≤ m∗ (E1 ) m∗ (E2 ),
with the understanding that if one of the sets Ej has exterior measure
zero, then m∗ (E1 × E2 ) = 0.
Proof. Let ² > 0. By definition, we can find cubes {Qk }∞

k=1 in R
d1
0 ∞ d2
and {Q` }`=1 in R such that
∞
[ ∞
[
E1 ⊂ Qk , and E2 ⊂ Q0`
k=1 `=1
and
∞
X ∞
X
|Qk | ≤ m∗ (E1 ) + ² and |Q0` | ≤ m∗ (E2 ) + ².
k=1 `=1
S∞
Since E1 × E2 ⊂ k,`=1 Qk × Q0` , the sub-additivity of the exterior mea-
sure yields
∞
X
m∗ (E1 × E2 ) ≤ |Qk × Q0` |
k,`=1
Ã∞ !Ã ∞
!
X X
= |Qk | |Q0` |
k=1 `=1
≤ (m∗ (E1 ) + ²)(m∗ (E2 ) + ²).
If neither E1 nor E2 has exterior measure 0, then from the above we find
m∗ (E1 × E2 ) ≤ m∗ (E1 ) m∗ (E2 ) + O(²),
and since ² is arbitrary, we must have m∗ (E1 × E2 ) ≤ m∗ (E1 ) m∗ (E2 ).

If for instance m∗ (E1 ) = 0, consider for each positive integer j the
set E2j = E2 ∩ {y ∈ Rd2 : |y| ≤ j}. Then, by the above argument, we
find that m∗ (E1 × E2j ) = 0. Since (E1 × E2j ) % (E1 × E2 ) as j → ∞, we
conclude that m∗ (E1 × E2 ) = 0.
Proposition 3.6 Suppose E1 and E2 are measurable subsets of Rd1 and

Rd2 , respectively. Then E = E1 × E2 is a measurable subset of Rd . More-
over,
m(E) = m(E1 ) m(E2 ),
with the understanding that if one of the sets Ej has measure zero, then
m(E) = 0.
Proof. It suffices to prove that E is measurable, because then the

assertion about m(E) follows from Corollary 3.3. Since each set Ej is
measurable, there exist sets Gj ⊂ Rdj of type Gδ , with Gj ⊃ Ej and

m∗ (Gj − Ej ) = 0 for each j = 1, 2. (See Corollary 3.5 in Chapter 1.)
Clearly, G = G1 × G2 is measurable in Rd1 × Rd2 and
(G1 × G2 ) − (E1 × E2 ) ⊂ ((G1 − E1 ) × G2 ) ∪ (G1 × (G2 − E2 )) .
By the lemma we conclude that m∗ (G − E) = 0, hence E is measurable.
As a consequence of this proposition we have the following.
Corollary 3.7 Suppose f is a measurable function on Rd1 . Then the

function f˜ defined by f˜(x, y) = f (x) is measurable on Rd1 × Rd2 .
Proof. To see this, we may assume that f is real-valued, and recall

first that if a ∈ R and E1 = {x ∈ Rd1 : f (x) < a}, then E1 is measurable
by definition. Since
{(x, y) ∈ Rd1 × Rd2 : f˜(x, y) < a} = E1 × Rd2 ,
the previous proposition shows that {f˜(x, y) < a} is measurable for each
a ∈ R. Thus f˜(x, y) is a measurable function on Rd1 × Rd2 , as desired.
Finally, we return to an interpretation of the integral
R that arose first in
the calculus. We have in mind the notion that f describes the “area”
under the graph of f . Here we relate this to the Lebesgue integral and
show how it extends to our more general context.
Corollary 3.8 Suppose f (x) is a non-negative function on Rd , and let
A = {(x, y) ∈ Rd × R : 0 ≤ y ≤ f (x)}.
Then:
(i) f is measurable on Rd if and only if A is measurable in Rd+1 .
(ii) If the conditions in (i) hold, then
Z
f (x) dx = m(A).
Rd
Proof. If f is measurable on Rd , then the previous proposition guar-

antees that the function
F (x, y) = y − f (x)
is measurable on Rd+1 , so we immediately see that A = {y ≥ 0} ∩ {F ≤

0} is measurable.
Conversely, suppose that A is measurable. We note that for each
x ∈ Rd1 the slice Ax = {y ∈ R : (x, y) ∈ A} is a closed segment, namely
Ax = [0, f (x)]. Consequently Corollary 3.3 (with the roles of x and y
interchanged) yields the measurability of m(Ax ) = f (x). Moreover
Z Z Z
m(A) = χA (x, y) dx dy = m(Ax ) dx = f (x) dx,
R d1 Rd1
as was to be shown.
We conclude this section with a useful result.
Proposition 3.9 If f is a measurable function on Rd , then the function

f˜(x, y) = f (x − y) is measurable on Rd × Rd .
By picking E = {z ∈ Rd : f (z) < a}, we see that it suffices to prove

that whenever E is a measurable subset of Rd , then Ẽ = {(x, y) : x − y ∈
E} is a measurable subset of Rd × Rd .
Note first that if O is an open set, then Õ is also open. Taking count-
able intersections shows that if E is a Gδ set, then so is Ẽ. Assume
now that m(Ẽk ) = 0 for each k, where Ẽk = Ẽ ∩ Bk and Bk = {|y| < k}.
Again, take O to be open in Rd , and let us calculate m(Õ ∩ Bk ). We
have that χÕ∩Bk = χO (x − y)χBk (y). Hence
Z
m(Õ ∩ Bk ) = χO (x − y)χBk (y) dy dx
Z µZ ¶
= χO (x − y) dx χBk (y) dy
= m(O) m(Bk ),
by the translation-invariance of the measure. Now if m(E) = 0, there is

a sequence of open sets On such that E ⊂ On and m(On ) → 0. It follows
from the above that Ẽk ⊂ Õn ∩ Bk and m(Õn ∩ Bk ) → 0 in n for each
fixed k. This shows m(Ẽk ) = 0, and hence m(Ẽ) = 0. The proof of the
proposition is concluded once we recall that any measurable set E can
be written as the difference of a Gδ and a set of measure zero.
4* A Fourier inversion formula

The question of the inversion of the Fourier transform encompasses in
effect the problem at the origin of Fourier analysis. This issue involves
4*. A Fourier inversion formula 87
establishing the validity of the inversion formula for a function f in terms

of its Fourier transform fˆ, that is,
Z
(14) ˆ
f (ξ) = f (x)e−2πix·ξ dx,
Rd
Z
(15) f (x) = fˆ(ξ)e2πix·ξ dξ.
Rd
We have already encountered this problem in Book I in the rudimen-

tary case when in fact both f and fˆ were continuous and had rapid (or
moderate) decrease at infinity. In Book II we also considered the ques-
tion in the one-dimensional setting, seen from the viewpoint of complex
analysis. The most elegant and useful formulations of Fourier inversion
are in terms of the L2 theory, or in its greatest generality stated in the
language of distributions. We shall take up these matters systematically
later.4 It will, nevertheless, be enlightening to digress here to see what
our knowledge at this stage teaches us about this problem. We intend to
do this by presenting a variant of the inversion formula appropriate for
L1 , one that is both simple and adequate in many circumstances.
To begin with, we need to have an idea of what can be said about the
Fourier transform of an arbitrary function in L1 (Rd ).
Proposition 4.1 Suppose f ∈ L1 (Rd ). Then fˆ defined by (14) is con-
tinuous and bounded on Rd .
In fact, since |f (x)e−2πix·ξ | = |f (x)|,R the integral representing fˆ con-
verges for each ξ and supξ∈Rd |fˆ(ξ)| ≤ Rd |f (x)| dx = kf k. To verify the
continuity, note that for every x, f (x)e−2πix·ξ → f (x)e−2πix·ξ0 as ξ → ξ0 ,
where ξ0 is any point in Rd ; hence fˆ(ξ) → fˆ(ξ0 ) by the dominated con-
vergence theorem.
One can assert a little more than the boundedness of fˆ; namely, one
has fˆ(ξ) → 0 as |ξ| → ∞, but not much more can be said about the
decrease at infinity of fˆ. (See Exercises 22 and 25.) As a consequence,
for general f ∈ L1 (Rd ) the function fˆ is not in L1 (Rd ), and the presumed
formula (15) becomes problematical. The following theorem evades this
difficulty and yet is useful in a number of situations.
Theorem 4.2 Suppose f ∈ L1 (Rd ) and assume also that fˆ ∈ L1 (Rd ).
Then the inversion formula (15) holds for almost every x.
An immediate corollary is the uniqueness of the Fourier transform
on L1 .
4 The L2 theory will be dealt with in Chapter 5, and distributions will be studied in
Book IV.
Corollary 4.3 Suppose fˆ(ξ) = 0 for all ξ. Then f = 0 a.e.
The proof of the theorem requires only that we adapt the earlier argu-
ments carried out for Schwartz functions in Chapter 5 of Book I to the
present context. We begin with the “multiplication formula.”
Lemma 4.4 Suppose f and g belong to L1 (Rd ). Then

Z Z
ˆ
f (ξ)g(ξ) dξ = f (y)ĝ(y) dy.
Rd Rd
Note that both integrals converge in view of the proposition above. Con-
sider the function F (ξ, y) = g(ξ)f (y)e−2πiξ·y defined for (ξ, y) ∈ Rd ×
Rd = R2d . It is measurable as a function on R2d in view of Corollary 3.7.
We now apply Fubini’s theorem to observe first that
Z Z Z Z
|F (ξ, y)| dξ dy = |g(ξ)| dξ |f (y)| dy < ∞.
Rd Rd Rd Rd
R R R ¡R ¢
Next, if we evaluate Rd Rd F (ξ, y) dξ dy by writing it as Rd Rd F (ξ, y) dξ dy
we get the left-hand side of the desired equality. Evaluating the double
integral in the reverse order gives as the right-hand side, proving the
lemma.
2
Next we consider the modulated Gaussian, g(ξ) = e−πδ|ξ| e2πix·ξ , where
for the moment δ and x are fixed, with δ > 0 and x ∈ Rd . An elementary
calculation gives5
Z
2 2
ĝ(y) = e−πδ|ξ| e2πi(x−y)·ξ dξ = δ −d/2 e−π|x−y| /δ ,
Rd
which we will abbreviate as Kδ (x − y). We recognize Kδ as a “good

kernel” that satisfies:
Z
(i) Kδ (y) dy = 1.
Rd
Z
(ii) For each η > 0, Kδ (y) dy → 0 as δ → 0.
|y|>η
Applying the lemma gives

Z Z
ˆ −πδ|ξ|2 2πix·ξ
(16) f (ξ)e e dξ = f (y)Kδ (x − y) dy.
Rd Rd
5 See for example Chapter 6 in Book I.

5. Exercises 89
Note that since fˆ ∈ L1 (Rd ), the dominated R convergence theorem shows

that the left-hand side of (16) converges to Rd fˆ(ξ)e2πix·ξ dξ as δ → 0, for
each x. As for the right-hand side, we make two successive change of vari-
ables y → y + x (a translation), and y → −y (a reflection), and take into
account the corresponding invariance of the integrals
R (see equations (4)
and (5)). Thus the right-hand side becomes Rd f (x − y)Kδ (y) dy, and
we will prove that this function converges in the L1 -norm to f as δ → 0.
Indeed, we can write the difference as
Z Z
∆δ (x) = f (x − y)Kδ (y) dy − f (x) = (f (x − y) − f (x))Kδ (y) dy,
Rd Rd
because of property (i) above. Thus

Z
|∆δ (x)| ≤ |f (x − y) − f (x)|Kδ (y) dy.
Rd
We can now apply Fubini’s theorem, recalling that the measurability

of f (x) and f (x − y) on Rd × Rd are established in Corollary 3.7 and
Proposition 3.9. The result is
Z
k∆δ k ≤ kfy − f kKδ (y) dy, where fy (x) = f (x − y).
Rd
Now, for given ² > 0 we can find (by Proposition 2.5) η > 0 so small such
that kfy − f k < ² when |y| < η. Thus
Z Z
k∆δ k ≤ ² + kfy − f kKδ (y) dy ≤ ² + 2kf k Kδ (y) dy.
|y|>η |y|>η
The first inequality follows by using (i) again; the second holds because
kfy − f k ≤ kfy k + kf k = 2kf k. Therefore, with the use of (ii), the com-
bination above is ≤ 2² if δ is sufficiently small. To summarize: the right-
hand side of (16) converges to f in the L1 -norm as δ → 0, and thus
by Corollary 2.3 there is a subsequence that converges to f (x) almost
everywhere, and the theorem is proved.
Note that an immediate consequence of the theorem and the proposi-
tion is that if fˆ were in L1 , then f could be modified on a set of measure
zero to become continuous everywhere. This is of course impossible for
the general f ∈ L1 (Rd ).
5 Exercises
∗ ∗ ∗
1. Given a collection of sets
SF 1 , F2 , . . . , Fn , construct another collection F1 , F2 , . . . , FN ,
SN
n n ∗ ∗
with N = 2 − 1, so that k=1 Fk = j=1 Fj ; the collection {Fj } is disjoint; also
S
Fk = Fj∗ ⊂Fk Fj∗ , for every k.
[Hint: Consider the 2n sets F10 ∩ F20 ∩ · · · ∩ Fn0 where each Fk0 is either Fk or Fkc .]
2. In analogy to Proposition 2.5, prove that if f is integrable on Rd and δ > 0,

then f (δx) converges to f (x) in the L1 -norm as δ → 1.
3. Suppose f is integrable on (−π, π] and extended to R by making it periodic of

period 2π. Show that
Z π Z
f (x) dx = f (x) dx,
−π I
where I is any interval in R of length 2π.

[Hint: I is contained in two consecutive intervals of the form (kπ, (k + 2)π).]
4. Suppose f is integrable on [0, b], and

Z b
f (t)
g(x) = dt for 0 < x ≤ b.
x t
Prove that g is integrable on [0, b] and

Z b Z b
g(x) dx = f (t) dt.
0 0
5. Suppose F is a closed set in R, whose complement has finite measure, and let
δ(x) denote the distance from x to F , that is,
δ(x) = d(x, F ) = inf{|x − y| : y ∈ F }.
Consider
Z
δ(y)
I(x) = dy.
R |x − y|2
(a) Prove that δ is continuous, by showing that it satisfies the Lipschitz condi-
tion
|δ(x) − δ(y)| ≤ |x − y|.
(b) Show that I(x) = ∞ for each x ∈

/ F.
(c) Show that I(x) < ∞ for a.e. x ∈ F . This may be surprising in view of the
fact that the Lispshitz condition cancels only one power of |x − y| in the
integrand of I.
5. Exercises 91
R
[Hint: For the last part, investigate F
I(x) dx.]
6. Integrability of f on R does not necessarily imply the convergence of f (x) to 0

as x → ∞.
(a) There exists a positive continuous function f on R so that f is integrable
on R, but yet lim supx→∞ f (x) = ∞.
(b) However, if we assume that f is uniformly continuous on R and integrable,

then lim|x|→∞ f (x) = 0.
[Hint: For (a), construct a continuous version of the function equal to n on the
segment [n, n + 1/n3 ), n ≥ 1.]
7. Let Γ ⊂ Rd × R, Γ = {(x, y) ∈ Rd × R : y = f (x)}, and assume f is measurable

on Rd . Show that Γ is a measurable subset of Rd+1 , and m(Γ) = 0.
Rx
8. If f is integrable on R, show that F (x) = −∞
f (t) dt is uniformly continuous.
9. Tchebychev inequality. Suppose f ≥ 0, and f is integrable. If α > 0 and

Eα = {x : f (x) > α}, prove that
Z
1
m(Eα ) ≤ f.
α
10. Suppose f ≥ 0, and let E2k = {x : f (x) > 2k } and Fk = {x : 2k < f (x) ≤
2k+1 }. If f is finite almost everywhere, then
∞
[
Fk = {f (x) > 0},
k=−∞
and the sets Fk are disjoint.

Prove that f is integrable if and only if
∞
X ∞
X
2k m(Fk ) < ∞, if and only if 2k m(E2k ) < ∞.
k=−∞ k=−∞
Use this result to verify the following assertions. Let

 
|x|−a if |x| ≤ 1, |x|−b if |x| > 1,
f (x) = and g(x) =
0 otherwise, 0 otherwise.
Then f is integrable on Rd if and only if a < d; also g is integrable on Rd if and

only if b > d.
R
11. Prove that if f is integrable on Rd , real-valued, and
R E f (x) dx ≥ 0 for ev-
ery measurable E, then f (x) ≥ 0 a.e. x. As a result, if E f (x) dx = 0 for every
measurable E, then f (x) = 0 a.e.
12. Show that there are f ∈ L1 (Rd ) and a sequence {fn } with fn ∈ L1 (Rd ) such
that
kf − fn kL1 → 0,
but fn (x) → f (x) for no x.

[Hint: In R, let fn = χIn , where In is an appropriately chosen sequence of intervals
with m(In ) → 0.]
13. Give an example of two measurable sets A and B such that A + B is not
measurable.
[Hint: In R2 take A = {0} × [0, 1] and B = N × {0}.]
14. In Exercise 6 of the previous chapter we saw that m(B) = vd rd , whenever B

is a ball of radius r in Rd and vd = m(B1 ), with B1 the unit ball. Here we evaluate
the constant vd .
(a) For d = 2, prove using Corollary 3.8 that

Z 1
v2 = 2 (1 − x2 )1/2 dx,
−1
and hence by elementary calculus, that v2 = π.
(b) By similar methods, show that

Z 1
vd = 2vd−1 (1 − x2 )(d−1)/2 dx.
0
(c) The result is
π d/2
vd = .
Γ(d/2 + 1)
Another derivation is in Exercise 5 in Chapter 6 below. Relevant facts about the

gamma and beta functions can be found in Chapter 6 of Book II.
15. Consider the function defined over R by


x−1/2 if 0 < x < 1,
f (x) =
0 otherwise.
For a fixed enumeration {rn }∞

n=1 of the rationals Q, let
∞
X
F (x) = 2−n f (x − rn ).
n=1
5. Exercises 93
Prove that F is integrable, hence the series defining F converges for almost every
x ∈ R. However, observe that this series is unbounded on every interval, and in
fact, any function F̃ that agrees with F a.e is unbounded in any interval.
16. Suppose f is integrable on Rd . If δ = (δ1 , . . . , δd ) is a d-tuple of non-zero real

numbers, and
f δ (x) = f (δx) = f (δ1 x1 , . . . , δd xd ),
show that f δ is integrable with

Z Z
f δ (x) dx = |δ1 |−1 · · · |δd |−1 f (x) dx.
Rd Rd
17. Suppose f is defined on R2 as follows: f (x, y) = an if n ≤ x < n + 1 and n ≤

y < n + 1, (n ≥ 0); f (x, y) = −an if n ≤ xP
< n + 1 and n + 1 ≤ y < n + 2, (n ≥ 0);
while f (x,P
y) = 0 elsewhere. Here an = k≤n bk , with {bk } a positive sequence
such that ∞ k=0 bk = s < ∞.
y
R
(a) Verify thatReach
`R slice f and´ fx is integrable. Also for all x, fx (y) dy = 0,
and hence f (x, y) dy dx = 0.
R y R
(b) However, f (x) dx = a0 if 0 ≤ y R< 1, and f y (x) dx = an − an−1 if n ≤
y < n + 1 with n ≥ 1. Hence y 7→ f y (x) dx is integrable on (0, ∞) and
Z „Z «
f (x, y) dx dy = s.
R
(c) Note that R×R
|f (x, y)| dx dy = ∞.
18. Let f be a measurable finite-valued function on [0, 1], and suppose that |f (x) −
f (y)| is integrable on [0, 1] × [0, 1]. Show that f (x) is integrable on [0, 1].
19. Suppose f is integrable on Rd . For each α > 0, let Eα = {x : |f (x)| > α}.
Prove that
Z Z ∞
|f (x)| dx = m(Eα ) dα.
Rd 0
20. The problem (highlighted in the discussion preceding Fubini’s theorem) that
certain slices of measurable sets can be non-measurable may be avoided by re-
stricting attention to Borel measurable functions and Borel sets. In fact, prove the
following:
Suppose E is a Borel set in R2 . Then for every y, the slice E y is a Borel set in
R.
[Hint: Consider the collection C of subsets E of R2 with the property that each
slice E y is a Borel set in R. Verify that C is a σ-algebra that contains the open
sets.]
21. Suppose that f and g are measurable functions on Rd .
(a) Prove that f (x − y)g(y) is measurable on R2d .
(b) Show that if f and g are integrable on Rd , then f (x − y)g(y) is integrable

on R2d .
(c) Recall the definition of the convolution of f and g given by

Z
(f ∗ g)(x) = f (x − y)g(y) dy.
Rd
Show that f ∗ g is well defined for a.e. x (that is, f (x − y)g(y) is integrable
on Rd for a.e. x).
(d) Show that f ∗ g is integrable whenever f and g are integrable, and that
kf ∗ gkL1 (Rd ) ≤ kf kL1 (Rd ) kgkL1 (Rd ) ,
with equality if f and g are non-negative.
(e) The Fourier transform of an integrable function f is defined by

Z
fˆ(ξ) = f (x)e−2πix·ξ dx.
Rd
Check that fˆ is bounded and is a continuous function of ξ. Prove that for

each ξ one has
\
(f ∗ g)(ξ) = fˆ(ξ)ĝ(ξ).
22. Prove that if f ∈ L1 (Rd ) and

Z
fˆ(ξ) = f (x)e−2πixξ dx,
Rd
then fˆ(ξ) → 0 as |ξ| → ∞. (This is the Riemann-Lebesgue lemma.)

R
[Hint: Write fˆ(ξ) = 1 d [f (x) − f (x − ξ 0 )]e−2πixξ dx, where ξ 0 = 1
2 R
ξ
2 |ξ|2
, and use
Proposition 2.5.]
23. As an application of the Fourier transform, show that there does not exist a
function I ∈ L1 (Rd ) such that
f ∗I =f for all f ∈ L1 (Rd ).

6. Problems 95
24. Consider the convolution

Z
(f ∗ g)(x) = f (x − y)g(y) dy.
Rd
(a) Show that f ∗ g is uniformly continuous when f is integrable and g bounded.
(b) If in addition g is integrable, prove that (f ∗ g)(x) → 0 as |x| → ∞.
1
25. Show that for each ² > 0 the function F (ξ) = (1+|ξ|2 )²
is the Fourier transform
of an L1 function.
2 R∞
[Hint: With Kδ (x) = e−π|x| /δ δ −d/2 consider f (x) = 0
Kδ (x)e−πδ δ ²−1 dδ. Use
Fubini’s theorem to prove f ∈ L1 (Rd ), and
Z ∞ 2
fˆ(ξ) = e−πδ|ξ| e−πδ δ ²−1 dδ,
0
and evaluate the last integral as π −² Γ(²) (1+|ξ|

1
2 )² . Here Γ(s) is the gamma function
R ∞ −t s−1
defined by Γ(s) = 0 e t dt.]
6 Problems
R 2π
1. If f is integrable on [0, 2π], then 0 f (x)e−inx dx → 0 as |n| → ∞.
Show as a consequence that if E is a measurable subset of [0, 2π], then
Z
m(E)
cos2 (nx + un ) dx → , as n → ∞
E 2
for any sequence {un }.

[Hint: See Exercise 22.]
2. Prove the Cantor-Lebesgue theorem: if
∞
X ∞
X
An (x) = (an cos nx + bn sin nx)
n=0 n=0
converges for x in a set of positive measure (or in particular for all x), then an → 0
and bn → 0 as n → ∞.
[Hint: Note that An (x) → 0 uniformly on a set E of positive measure.]
3. A sequence {fk } of measurable functions on Rd is Cauchy in measure if for

every ² > 0,
m({x : |fk (x) − f` (x)| > ²}) → 0 as k, ` → ∞.

We say that {fk } converges in measure to a (measurable) function f if for every

²>0
m({x : |fk (x) − f (x)| > ²}) → 0 as k → ∞.
This notion coincides with the “convergence in probability” of probability theory.

Prove that if a sequence {fk } of integrable functions converges to f in L1 , then
{fk } converges to f in measure. Is the converse true?
We remark that this mode of convergence appears naturally in the proof of
Egorov’s theorem.
4. We have already seen (in Exercise 8, Chapter 1) that if E is a measurable set

in Rd , and L is a linear transformation of Rd to Rd , then L(E) is also measurable,
and if E has measure 0, then so has L(E). The quantitative statement is
m(L(E)) = | det(L)| m(E).
As a special case, note that the Lebesgue measure is invariant under rotations.
(For this special case see also Exercise 26 in the next chapter.)
The above identity can be proved using Fubini’s theorem as follows.
(a) Consider first the case d = 2, and L a “strictly” upper triangular transfor-
mation x0 = x + ay, y 0 = y. Then
χL(E) (x, y) = χE (L−1 (x, y)) = χE (x − ay, y).
Hence
Z „Z «
m(L(E)) = χE (x − ay, y) dy
R×R
Z „Z «
= χE (x, y) dx dy
R×R
= m(E),
by the translation-invariance of the measure.

(b) Similarly m(L(E)) = m(E) if L is strictly lower triangular. In general, one
can write L = L1 ∆L2 , where Lj are strictly (upper and lower) triangular
and ∆ is diagonal. Thus m(L(E)) = | det(L)|m(E), if one uses Exercise 7
in Chapter 1.
5. There is an ordering ≺ of R with the property that for each y ∈ R the set
{x ∈ R : x ≺ y} is at most countable.
The existence of this ordering depends on the continuum hypothesis, which
asserts: whenever S is an infinite subset of R, then either S is countable, or S has
the cardinality of R (that is, can be mapped bijectively to R).6
6 This assertion, formulated by Cantor, is like the well-ordering principle independent
of the other axioms of set theory, and so we are also free to accept its validity.
6. Problems 97
[Hint: Let ≺ denote a well-ordering of R, and define the set X by X = {y ∈

R : the set {x : x ≺ y} is not countable}. If X is empty we are done. Otherwise,
consider the smallest element y in X, and use the continuum hypothesis.]
3 Differentiation and Integration
The Maximal Problem:
The problem is most easily grasped when stated
in the language of cricket, or any other game in which
a player compiles a series of scores of which an average
is recorded.
G. H. Hardy and J. E. Littlewood, 1930
That differentiation and integration are inverse operations was already

understood early in the study of the calculus. Here we want to reexamine
this basic idea in the framework of the general theory studied in the
previous chapters. Our objective is the formulation and proof of the
fundamental theorem of the calculus in this setting, and the development
of some of the concepts that occur. We shall try to achieve this by
answering two questions, each expressing one of the ways of representing
the reciprocity between differentiation and integration.
The first problem involved may be stated as follows.
• SupposeR f is integrable on [a, b] and F is its indefinite integral

x
F (x) = a f (y) dy. Does this imply that F is differentiable (at
least for almost every x), and that F 0 = f ?
We shall see that the affirmative answer to this question depends

on ideas that have broad application and are not limited to the one-
dimensional situation.
For the second question we reverse the order of differentiation and
integration.
• What conditions on a function F on [a, b] guarantee that F 0 (x) ex-

ists (for a.e. x), that this function is integrable, and that moreover
Z b
F (b) − F (a) = F 0 (x) dx ?
a
While this problem will be examined from a narrower perspective than

the first, the issues it raises are deep and the consequences entailed are
1. Differentiation of the integral 99
far-reaching. In particular, we shall find that this question is connected

to the problem of rectifiability of curves, and as an illustration of this
link, we shall establish the general isoperimetric inequality in the plane.
1 Differentiation of the integral

We begin with the first problem, that is, the study of differentiation of
the integral. If f is given on [a, b] and integrable on that interval, we let
Z x
F (x) = f (y) dy, a ≤ x ≤ b.
a
To deal with F 0 (x), we recall the definition of the derivative as the limit
of the quotient
F (x + h) − F (x)
when h tends to 0.
h
We note that this quotient takes the form (say in the case h > 0)
Z x+h Z
1 1
f (y) dy = f (y) dy,
h x |I| I
where we use the notation I = (x, x + h) and |I| for the length of this
interval. At this point, we pause to observe that the above expression
is the “average” value of f over I, and that in the limit as |I| → 0,
we might expect that these averages tend to f (x). Reformulating the
question slightly, we may ask whether
Z
1
lim f (y) dy = f (x)
|I| → 0
x ∈ I
|I| I
holds for suitable points x. In higher dimensions we can pose a similar

question, where the averages of f are taken over appropriate sets that
generalize the intervals in one dimension. Initially we shall study this
problem where the sets involved are the balls B containing x, with their
volume m(B) replacing the length |I| of I. Later we shall see that as a
consequence of this special case similar results will hold for more general
collections of sets, those that have bounded “eccentricity.”
With this in mind we restate our first problem in the context of Rd ,
for all d ≥ 1.
100 Chapter 3. DIFFERENTIATION AND INTEGRATION
Suppose f is integrable on Rd . Is it true that

Z
1
lim f (y) dy = f (x), for a.e. x?
m(B) → 0
x ∈ B
m(B) B
The limit is taken as the volume of open balls B containing

x goes to 0.
We shall refer to this question as the averaging problem. We remark
that if B is any ball of radius r in Rd , then m(B) = vd rd , where vd is
the measure of the unit ball. (See Exercise 14 in the previous chapter.)
Note of course that in the special case when f is continuous at x , the
limit does converge to f (x). Indeed, given ² > 0, there exists δ > 0 such
that |f (x) − f (y)| < ² whenever |x − y| < δ. Since
Z Z
1 1
f (x) − f (y) dy = (f (x) − f (y)) dy,
m(B) B m(B) B
we find that whenever B is a ball of radius < δ/2 that contains x, then
¯ Z ¯ Z
¯ 1 ¯
¯f (x) − f (y) dy ¯≤ 1 |f (x) − f (y)| dy < ²,
¯ m(B) ¯ m(B)
B B
as desired.
The averaging problem has an affirmative answer, but to establish that
fact, which is qualitative in nature, we need to make some quantitative
estimates bearing on the overall behavior of the averages of f . This will
be done in terms of the maximal averages of |f |, to which we now turn.
1.1 The Hardy-Littlewood maximal function

The maximal function that we consider below arose first in the one-
dimensional situation treated by Hardy and Littlewood. It seems that
they were led to the study of this function by toying with the question
of how a batsman’s score in cricket may best be distributed to maximize
his satisfaction. As it turns out, the concepts involved have a universal
significance in analysis. The relevant definition is as follows.
If f is integrable on Rd , we define its maximal function f ∗ by
Z
1
f ∗ (x) = sup |f (y)| dy, x ∈ Rd ,
x∈B m(B) B
where the supremum is taken over all balls containing the point x. In
other words, we replace the limit in the statement of the averaging prob-
lem by a supremum, and f by its absolute value.
The main properties of f ∗ we shall need are summarized in a theorem.

Theorem 1.1 Suppose f is integrable on Rd . Then:
(i) f ∗ is measurable.
(ii) f ∗ (x) < ∞ for a.e. x.
(iii) f ∗ satisfies
A
(1) m({x ∈ Rd : f ∗ (x) > α}) ≤ kf kL1 (Rd )
α
R
for all α > 0, where A = 3d , and kf kL1 (Rd ) = Rd |f (x)| dx.
Before we come to the proof we want to clarify the nature of the main
conclusion (iii). As we shall observe, one has that f ∗ (x) ≥ |f (x)| for a.e.
x; the effect of (iii) is that, broadly speaking, f ∗ is not much larger than
|f |. From this point of view, we would have liked to conclude that f ∗ is
integrable, as a result of the assumed integrability of f . However, this
is not the case, and (iii) is the best substitute available (see Exercises 4
and 5).
An inequality of the type (1) is called a weak-type inequality be-
cause it is weaker than the corresponding inequality for the L1 -norms.
Indeed, this can be seen from the Tchebychev inequality (Exercise 9 in
Chapter 2), which states that for an arbitrary integrable function g,
1
m({x : |g(x)| > α}) ≤ kgkL1 (Rd ) , for all α > 0.
α
We should add that the exact value of A in the inequality (1) is unim-
portant for us. What matters is that this constant be independent of α
and f .
The only simple assertion in the theorem is that f ∗ is a measurable
function. Indeed, the set Eα = {x ∈ Rd : f ∗ (x) > α} is open, because if
x ∈ Eα , there exists a ball B such that x ∈ B and
Z
1
|f (y)| dy > α.
m(B) B
Now any point x close enough to x will also belong to B; hence x ∈ Eα
as well.
The two other properties of f ∗ in the theorem are deeper, with (ii)
being a consequence of (iii). This follows at once if we observe that
{x : f ∗ (x) = ∞} ⊂ {x : f ∗ (x) > α}

for all α. Taking the limit as α tends to infinity, the third property yields
m({x : f ∗ (x) = ∞}) = 0.
The proof of inequality (1) relies on an elementary version of a Vitali
covering argument.1
Lemma 1.2 Suppose B = {B1 , B2 , . . . , BN } is a finite collection of open

balls in Rd . Then there exists a disjoint sub-collection Bi1 , Bi2 , . . . , Bik
of B that satisfies
Ã N
! k
[ X
m B` ≤ 3d m(Bij ).
`=1 j=1
Loosely speaking, we may always find a disjoint sub-collection of balls

that covers a fraction of the region covered by the original collection of
balls.
Proof. The argument we give is constructive and relies on the fol-
lowing simple observation: Suppose B and B 0 are a pair of balls that
intersect, with the radius of B 0 being not greater than that of B. Then
B 0 is contained in the ball B̃ that is concentric with B but with 3 times
its radius.
As a first step, we pick a ball Bi1 in B with maximal (that is, largest)
radius, and then delete from B the ball Bi1 as well as any balls that
intersect Bi1 . Thus all the balls that are deleted are contained in the
ball B̃i1 concentric with Bi1 , but with 3 times its radius.
The remaining balls yield a new collection B 0 , for which we repeat the
procedure. We pick Bi2 with largest radius in B 0 , and then delete from
B 0 the ball Bi2 and any ball that intersects Bi2 . Continuing this way we
find, after at most N steps, a collection of disjoint balls Bi1 , Bi2 , . . . , Bik .
Finally, to prove that this disjoint collection of balls satisfies the in-
equality in the lemma, we use the observation made at the beginning of
the proof. We let B̃ij denote the ball concentric with Bij , but with 3
times its radius. Since any ball B in B must intersect a ball Bij and have
equal or smaller radius than Bij , we must have B ⊂ B̃ij , thus
Ã N
! Ã k
! k k
[ [ X X
m B` ≤m B̃ij ≤ m(B̃ij ) = 3d m(Bij ).
`=1 j=1 j=1 j=1
1 We note that the lemma that follows is the first of a series of covering arguments that
occur below in the theory of differentiation; see also Lemma 3.9 and its corollary, as well
as Lemma 3.5, where the covering assertion is more implicit.
B̃
B0
Figure 1. The balls B and B̃
In the last step we have used the fact that in Rd a dilation of a set by
δ > 0 results in the multiplication by δ d of the Lebesgue measure of this
set.
The proof of (iii) in Theorem 1.1 is now in reach. If we let Eα = {x :
f ∗ (x) > α}, then for each x ∈ Eα there exists a ball Bx that contains x,
and such that
Z
1
|f (y)| dy > α.
m(Bx ) Bx
Therefore, for each ball Bx we have

Z
1
(2) m(Bx ) < |f (y)| dy.
α Bx
S
Fix a compact subset K of Eα . Since K is covered by x∈Eα Bx , we
SN
may select a finite subcover of K, say K ⊂ `=1 B` . The covering lemma
guarantees the existence of a sub-collection Bi1 , . . . , Bik of disjoint balls
with
Ã N
! k
[ X
(3) m B` ≤ 3d m(Bij ).
`=1 j=1
Since the balls Bi1 , . . . , Bik are disjoint and satisfy (2) as well as (3), we
find that
ÃN ! k k Z
[ X 3d X
d
m(K) ≤ m B` ≤ 3 m(Bij ) ≤ |f (y)| dy
α
`=1 j=1 j=1 Bij
Z
3d
= |f (y)| dy
α Skj=1 Bi
j
Z
3d
≤ |f (y)| dy.
α Rd
Since this inequality is true for all compact subsets K of Eα , the proof
of the weak type inequality for the maximal operator is complete.
1.2 The Lebesgue differentiation theorem

The estimate obtained for the maximal function now leads to a solution
of the averaging problem.
Theorem 1.3 If f is integrable on Rd , then
Z
1
(4) lim f (y) dy = f (x) for a.e. x.
m(B) → 0
x ∈ B
m(B) B
Proof. It suffices to show that for each α > 0 the set

 
 ¯ Z ¯ 
¯ 1 ¯
Eα = x : lim sup ¯ ¯ ¯
f (y) dy − f (x)¯ > 2α
 m(B) → 0 m(B) B 
x ∈ B
has measure
S∞ zero, because this assertion then guarantees that the set
E = n=1 E1/n has measure zero, and the limit in (4) holds at all points
of E c .
We fix α, and recall Theorem 2.4 in Chapter 2, which states that for
each ² > 0 we may select a continuous function g of compact support with
kf − gkL1 (Rd ) < ². As we remarked earlier, the continuity of g implies
that
Z
1
lim g(y) dy = g(x), for all x.
m(B) → 0
x ∈ B
m(B) B
1
R
Since we may write the difference m(B) B
f (y) dy − f (x) as
Z Z
1 1
(f (y) − g(y)) dy + g(y) dy − g(x) + g(x) − f (x)
m(B) B m(B) B
we find that
¯ Z ¯
¯ 1 ¯
lim sup ¯¯ f (y) dy − f (x)¯¯ ≤ (f − g)∗ (x) + |g(x) − f (x)|,
m(B) → 0 m(B) B
x ∈ B
where the symbol ∗ indicates the maximal function. Consequently, if
Fα = {x : (f − g)∗ (x) > α} and Gα = {x : |f (x) − g(x)| > α}
then Eα ⊂ (Fα ∪ Gα ), because if u1 and u2 are positive, then u1 + u2 >

2α only if ui > α for at least one ui . On the one hand, Tchebychev’s
inequality yields
1
m(Gα ) ≤ kf − gkL1 (Rd ) ,
α
and on the other hand, the weak type estimate for the maximal function
gives
A
m(Fα ) ≤ kf − gkL1 (Rd ) .
α
The function g was selected so that kf − gkL1 (Rd ) < ². Hence we get
A 1
m(Eα ) ≤ ² + ².
α α
Since ² is arbitrary, we must have m(Eα ) = 0, and the proof of the the-
orem is complete.
Note that as an immediate consequence of the theorem applied to |f |,
we see that f ∗ (x) ≥ |f (x)| for a.e. x, with f ∗ the maximal function.
We have worked so far under the assumption that f is integrable. This
“global” assumption is slightly out of place in the context of a “local”
notion like differentiability. Indeed, the limit in Lebesgue’s theorem is
taken over balls that shrink to the point x, so the behavior of f far from
x is irrelevant. Thus, we expect the result to remain valid if we simply
assume integrability of f on every ball.
To make this precise, we say that a measurable function f on Rd
is locally integrable, if for every ball B the function f (x)χB (x) is
integrable. We shall denote by L1loc (Rd ) the space of all locally integrable
functions. Loosely speaking, the behavior at infinity does not affect the
local integrability of a function. For example, the functions e|x| and
|x|−1/2 are both locally integrable, but not integrable on Rd .
Clearly, the conclusion of the last theorem holds under the weaker
assumption that f is locally integrable.
Theorem 1.4 If f ∈ L1loc (Rd ), then

Z
1
lim f (y) dy = f (x), for a.e. x.
m(B) → 0
x ∈ B
m(B) B
Our first application of this theorem yields an interesting insight into

the nature of measurable sets. If E is a measurable set and x ∈ Rd , we
say that x is a point of Lebesgue density of E if
m(B ∩ E)
lim = 1.
m(B) → 0
x ∈ B
m(B)
Loosely speaking, this condition says that small balls around x are almost
entirely covered by E. More precisely, for every α < 1 close to 1, and
every ball of sufficiently small radius containing x, we have
m(B ∩ E) ≥ αm(B).
Thus E covers at least a proportion α of B.

An application of Theorem 1.4 to the characteristic function of E im-
mediately yields the following:
Corollary 1.5 Suppose E is a measurable subset of Rd . Then:
(i) Almost every x ∈ E is a point of density of E.
(ii) Almost every x ∈
/ E is not a point of density of E.
We next consider a notion that for integrable functions serves as a useful
substitute for pointwise continuity.
If f is locally integrable on Rd , the Lebesgue set of f consists of all
points x ∈ Rd for which f (x) is finite and
Z
1
lim |f (y) − f (x)| dy = 0.
m(B) → 0
x ∈ B
m(B) B
At this stage, two simple observations about this definition are in order.
First, x belongs to the Lebesgue set of f whenever f is continuous at x.
Second, if x is in the Lebesgue set of f , then
Z
1
lim f (y) dy = f (x).
m(B) → 0
x ∈ B
m(B) B
Corollary 1.6 If f is locally integrable on Rd , then almost every point

belongs to the Lebesgue set of f .
Proof. An application of Theorem 1.4 to the function |f (y) − r| shows

that for each rational r, there exists a set Er of measure zero, such that
Z
1
lim |f (y) − r| dy = |f (x) − r| whenever x ∈ / Er .
m(B) → 0
x ∈ B
m(B) B
S
If E = r∈Q Er , then m(E) = 0. Now suppose that x ∈ / E and f (x) is
finite. Given ² > 0, there exists a rational r such that |f (x) − r| < ².
Since
Z Z
1 1
|f (y) − f (x)| dy ≤ |f (y) − r| dy + |f (x) − r|,
m(B) B m(B) B
we must have
Z
1
lim sup |f (y) − f (x)| dy ≤ 2²,
m(B) → 0 m(B) B
x ∈ B
and thus x is in the Lebesgue set of f . The corollary is therefore proved.
Remark. Recall from the definition in Section 2 of Chapter 2 that

elements of L1 (Rd ) are actually equivalence classes, with two functions
being equivalent if they differ on a set of measure zero. It is interesting
to observe that the set of points where the averages (4) converge to a
limit is independent of the representation of f chosen, because
Z Z
f (y) dy = g(y) dy
B B
whenever f and g are equivalent. Nevertheless, the Lebesgue set of f

depends on the particular representative of f that we consider.
We shall see that the Lebesgue set of a function enjoys a universal
property in that at its points the function can be recovered by a wide
variety of averages. We will prove this both for averages over sets that
generalize balls, and in the setting of approximations to the identity.
Note that the theory of differentiation developed so far uses averages
over balls, but as we mentioned earlier, one could ask whether similar
conclusions hold for other families of sets, such as cubes or rectangles.
The answer depends in a fundamental way on the geometric properties
of the family in question. For example, we now show that in the case of
cubes (and more generally families of sets with bounded “eccentricity”)
the above results carry over. However, in the case of the family of all
rectangles the existence of the limit almost everywhere and the weak
type inequality fail (see Problem 8).
A collection of sets {Uα } is said to shrink regularly to x (or has
bounded eccentricity at x) if there is a constant c > 0 such that for
each Uα there is a ball B with
x ∈ B, Uα ⊂ B, and m(Uα ) ≥ c m(B).
Thus Uα is contained in B, but its measure is comparable to the measure

of B. For example, the set of all open cubes containing x shrink regularly
to x. However, in Rd with d ≥ 2 the collection of all open rectangles
containing x does not shrink regularly to x. This can be seen if we
consider very thin rectangles.
Corollary 1.7 Suppose f is locally integrable on Rd . If {Uα } shrinks
regularly to x, then
Z
1
lim f (y) dy = f (x)
m(Uα ) → 0
x ∈ U
m(Uα ) Uα
α
for every point x in the Lebesgue set of f .

The proof is immediate once we observe that if x ∈ B with Uα ⊂ B
and m(Uα ) ≥ cm(B), then
Z Z
1 1
|f (y) − f (x)| dy ≤ |f (y) − f (x)| dy.
m(Uα ) Uα cm(B) B
2 Good kernels and approximations to the identity

We shall now turn to averages of functions given as convolutions,2 which
can be written as
Z
(f ∗ Kδ )(x) = f (x − y)Kδ (y) dy.
Rd
Here f is a general integrable function, which we keep fixed, while the Kδ

vary over a specific family of functions, referred to as kernels. Expressions
of this kind arise in many questions (for instance, in the Fourier inversion
theorem of the previous chapter), and were already discussed in Book I.
In our initial consideration we called these functions “good kernels” if
they are integrable and satisfy the following conditions for δ > 0:
2 Some basic properties of convolutions are described in Exercise 21 of the previous
chapter.
2. Good kernels and approximations to the identity 109
Z
(i) Kδ (x) dx = 1.
Rd
Z
(ii) |Kδ (x)| dx ≤ A.
Rd
(iii) For every η > 0,

Z
|Kδ (x)| dx → 0 as δ → 0.
|x|≥η
Here A is a constant independent of δ.

The main use of these kernels was that whenever f is bounded, then
(f ∗ Kδ )(x) → f (x) as δ → 0, at every point of continuity of f . To obtain
a similar conclusion, one also valid at all points of the Lebesgue set
of f , we need to strengthen somewhat our assumptions on the kernels
Kδ . To reflect this situation we adopt a different terminology and refer
to the resulting narrower class of kernels as approximations to the
identity. The assumptions are again that the Kδ are integrable and
satisfy conditions (i) but, instead of (ii) and (iii), we assume:
(ii0 ) |Kδ (x)| ≤ Aδ −d for all δ > 0.

(iii0 ) |Kδ (x)| ≤ Aδ/|x|d+1 for all δ > 0 and x ∈ Rd .3
We observe that these requirements are stronger and imply the conditions
in the definition of good kernels. Indeed, we first prove (ii). For that, we
use the second illustration of Corollary 1.10 in Chapter 2, which gives
Z
dx C
(5) d+1
≤ for some C > 0 and all ² > 0.
|x|≥² |x| ²
Then, using the estimates (ii0 ) and (iii0 ) when |x| < δ and |x| ≥ δ, re-
spectively, yields
Z Z Z
|Kδ (x)| dx = |Kδ (x)| dx + |Kδ (x)| dx
Rd |x|<δ |x|≥δ
Z Z
dx 1
≤A + Aδ dx
|x|<δ δd |x|≥δ |x|d+1
0 00
≤ A + A < ∞.
3 Sometimes the condition (iii0 ) is replaced by the requirement |K (x)| ≤ Aδ ² /|x|d+²

δ
for some fixed ² > 0. However, the special case ² = 1 suffices in most circumstances.
Finally, the last condition of a good kernel is also verified, since another
application of (5) gives
Z Z
dx
|Kδ (x)| dx ≤ Aδ
|x|≥η |x|≥η |x|d+1
0
Aδ
≤ ,
η
and this last expression tends to 0 as δ → 0.

The term “approximation to the identity” originates in the fact that
the mapping f 7→ f ∗ Kδ converges to the identity mapping f 7→ f , as
δ → 0, in various senses, as we shall see below. It is also connected with
the following heuristics. Figure 2 pictures a typical approximation to the
identity: for each δ > 0, the kernel is supported on the set |x| < δ and
has height 1/2δ. As δ tends to 0, this family of kernels converges to the
1/2δ
−δ 0 δ
Figure 2. An approximation to the identity
so-called unit mass at the origin or Dirac delta “function.” The latter
is heuristically defined by
½ Z
∞ if x = 0
D(x) = and D(x) dx = 1.
0 if x 6= 0
Since each Kδ integrates to 1, we may say loosely that

Kδ → D as δ → 0.
R
If we think of the convolution f ∗ D as f (x − y)D(y) dy, the product
f (x − y)D(y) is 0 except when y = 0, and the mass of D is concentrated
at y = 0, so we may intuitively expect that
(f ∗ D)(x) = f (x).
Thus f ∗ D = f , and D plays the role of the identity for convolutions.
We should mention that this discussion can be formalized and D given
a precise definition either in terms of Lebesgue-Stieltjes measures, which
we take up in Chapter 6, or in terms of “generalized functions” (that is,
distributions), which we defer to Book IV.
We now turn to a series of examples of approximations to the identity.
Example 1. Suppose ϕ is a non-negative bounded function in Rd that
is supported on the unit ball |x| ≤ 1, and such that
Z
ϕ = 1.
Rd
Then, if we set Kδ (x) = δ −d ϕ(δ −1 x), the family {Kδ }δ>0 is an approx-
imation to the identity. The simple verification is left to the reader.
Important special cases are in the next two examples.
Example 2. The Poisson kernel for the upper half-plane is given by

1 y
Py (x) = , x ∈ R,
π x2 + y 2
where the parameter is now δ = y > 0.
Example 3. The heat kernel in Rd is defined by

1 2
Ht (x) = d/2
e−|x| /4t .
(4πt)
Here t > 0 and we have δ = t1/2 . Alternatively, we could set δ = 4πt to
make the notation consistent with the specific usage in Chapter 2.
Example 4. The Poisson kernel for the disc is


 1 1 − r2
1 if |x| ≤ π,
Pr (x) = 2π 1 − 2r cos x + r2
2π  0 if |x| > π.
Here we have 0 < r < 1 and δ = 1 − r.
Example 5. The Fejér kernel is defined by


1
 1 sin2 (N x/2)
if |x| ≤ π,
FN (x) = 2πN sin2 (x/2)
2π 
0 if |x| > π,
where δ = 1/N .
We note that Examples 2 through 5 have already appeared in Book I.
We now turn to a general result about approximations to the identity
that highlights the role of the Lebesgue set.
Theorem 2.1 If {Kδ }δ>0 is an approximation to the identity and f is
integrable on Rd , then
(f ∗ Kδ )(x) → f (x) as δ → 0
for every x in the Lebesgue set of f . In particular, the limit holds for
a.e. x.
Since the integral of each kernel Kδ is equal to 1, we may write
Z
(f ∗ Kδ )(x) − f (x) = [f (x − y) − f (x)] Kδ (y) dy.
Consequently,
Z
|(f ∗ Kδ )(x) − f (x)| ≤ |f (x − y) − f (x)| |Kδ (y)| dy,
and it now suffices to prove that the right-hand side tends to 0 as δ goes
to 0. The argument we give depends on a simple result that we isolate
in the next lemma.
Lemma 2.2 Suppose that f is integrable on Rd , and that x is a point of
the Lebesgue set of f . Let
Z
1
A(r) = d |f (x − y) − f (x)| dy, whenever r > 0.
r |y|≤r
Then A(r) is a continuous function of r > 0, and
A(r) → 0 as r → 0.
Moreover, A(r) is bounded, that is, A(r) ≤ M for some M > 0 and all
r > 0.
Proof. The continuity of A(r) follows by invoking the absolute conti-

nuity in Proposition 1.12 of Chapter 2.
The fact that A(r) tends to 0 as r tends to 0 follows since x belongs
to the Lebesgue set of f , and the measure of a ball of radius r is vd rd .
This and the continuity of A(r) for 0 < r ≤ 1 show that A(r) is bounded
when 0 < r ≤ 1. To prove that A(r) is bounded for r > 1, note that
Z Z
1 1
A(r) ≤ d |f (x − y)| dy + d |f (x)| dy
r |y|≤r r |y|≤r
≤ r−d kf kL1 (Rd ) + vd |f (x)|,
and this concludes the proof of the lemma.

We now return to the proof of the theorem. The key consists in writing
the integral over Rd as a sum of integrals over annuli as follows:
Z Z ∞ Z
X
|f (x − y) − f (x)| |Kδ (y)| dy = + .
|y|≤δ k=0 2k δ<|y|≤2k+1 δ
By using the property (ii0 ) of the approximation to the identity, the first
term is estimated by
Z Z
c
|f (x − y) − f (x)| |Kδ (y)| dy ≤ d |f (x − y) − f (x)| dy
|y|≤δ δ |y|≤δ
≤ c A(δ).
Each term in the sum is estimated similarly, but this time by using
property (iii0 ) of approximations to the identity:
Z
|f (x − y) − f (x)| |Kδ (y)| dy
2k δ<|y|≤2k+1 δ
Z
cδ
≤ |f (x − y) − f (x)| dy
(2k δ)d+1 |y|≤2k+1 δ
Z
c0
≤ |f (x − y) − f (x)| dy
2k (2k+1 δ)d |y|≤2k+1 δ
≤ c0 2−k A(2k+1 δ).
Putting these estimates together, we find that

∞
X
|(f ∗ Kδ )(x) − f (x)| ≤ c A(δ) + c0 2−k A(2k+1 δ).
k=0
P
Given ² > 0, we first choose N so large that k≥N 2−k < ². Then, by
making δ sufficiently small, we have by the lemma
A(2k δ) < ²/N, whenever k = 0, 1, . . . , N − 1.
Hence, recalling that A(r) is bounded, we find
|(f ∗ Kδ )(x) − f (x)| ≤ C²
for all sufficiently small δ, and the theorem is proved.

In addition to this pointwise result, convolutions with approximations
to the identity also provide convergence in the L1 -norm.
Theorem 2.3 Suppose that f is integrable on Rd and that {Kδ }δ>0 is
an approximation to the identity. Then, for each δ > 0, the convolution
Z
(f ∗ Kδ )(x) = f (x − y)Kδ (y) dy
Rd
is integrable, and
k(f ∗ Kδ ) − f kL1 (Rd ) → 0, as δ → 0.
The proof is merely a repetition in a more general context of the argument

2
in the special case where Kδ (x) = δ −d/2 e−π|x| /δ given in Section 4*,
Chapter 2, and so will not be repeated.
3 Differentiability of functions
We now take up the second question raised at the beginning of this
chapter, that of finding a broad condition on functions F that guarantees
the identity
Z b
(6) F (b) − F (a) = F 0 (x) dx.
a
There are two phenomena that make a general formulation of this identity
problematic. First, because of the existence of non-differentiable func-
tions,4 the right-hand side of (6) might not be meaningful if we merely
assumed F was continuous. Second, even if F 0 (x) existed for every x,
the function F 0 would not necessarily be (Lebesgue) integrable. (See
Exercise 12.)
4 In particular, there are continuous nowhere differentiable functions. See Chapter 4 in
Book I, or also Chapter 7 below.

3. Differentiability of functions 115
How do we deal with these difficulties? One way is by limiting ourselves

to those functions F that arise as indefinite integrals (of integrable func-
tions). This raises the issue of how to characterize such functions, and
we approach that problem via the study of a wider class, the functions
of bounded variation. These functions are closely related to the question
of rectifiability of curves, and we start by considering this connection.
3.1 Functions of bounded variation

Let γ be a parametrized curve in the plane given by z(t) = (x(t), y(t)),
where a ≤ t ≤ b. Here x(t) and y(t) are continuous real-valued functions
on [a, b]. The curve γ is rectifiable if there exists M < ∞ such that, for
any partition a = t0 < t1 < · · · < tN = b of [a, b],
N
X
(7) |z(tj ) − z(tj−1 )| ≤ M.
j=1
By definition, the length L(γ) of the curve is the supremum over all
partitions of the sum on the left-hand side, that is,
N
X
L(γ) = sup |z(tj ) − z(tj−1 )|.
a=t0 <t1 <···<tN =b
j=1
Alternatively, L(γ) is the infimum of all M that satisfy (7). Geomet-

rically, the quantity L(γ) is obtained by approximating the curve by
polygonal lines and taking the limit of the length of these polygonal
lines as the interval [a, b] is partitioned more finely (see the illustration
in Figure 3).
Naturally, we may now ask the following questions: What analytic
condition on x(t) and y(t) guarantees rectifiability of the curve γ? In
particular, must the derivatives of x(t) and y(t) exist? If so, does one
have the desired formula
Z b
L(γ) = (x0 (t)2 + y 0 (t)2 )1/2 dt?
a
The answer to the first question leads directly to the class of functions
of bounded variation, a class that plays a key role in the theory of dif-
ferentiation.
Suppose F (t) is a complex-valued function defined on [a, b], and a =
t0 < t1 < · · · < tN = b is a partition of this interval. The variation of F
Figure 3. Approximation of a rectifiable curve by polygonal lines
on this partition is defined by

N
X
|F (tj ) − F (tj−1 )|.
j=1
The function F is said to be of bounded variation if the variations of

F over all partitions are bounded, that is, there exists M < ∞ so that
N
X
|F (tj ) − F (tj−1 )| ≤ M
j=1
for all partitions a = t0 < t1 < · · · < tN = b. In this definition we do not

assume that F is continuous; however, when applying it to the case of
curves, we will suppose that F (t) = z(t) = x(t) + iy(t) is continuous.
We observe that if a partition P̃ given by a = t̃0 < t̃1 < · · · < t̃M = b is
a refinement5 of a partition P given by a = t0 < t1 < · · · < tN = b, then
the variation of F on P̃ is greater than or equal to the variation of F on
P.
Theorem 3.1 A curve parametrized by (x(t), y(t)), a ≤ t ≤ b, is rectifi-
able if and only if both x(t) and y(t) are of bounded variation.
The proof is immediate once we observe that if F (t) = x(t) + iy(t), then
F (tj ) − F (tj−1 ) = (x(tj ) − x(tj−1 )) + i (y(tj ) − y(tj−1 )) ,
5 We say that a partition P̃ of [a, b] is a refinement of a partition P of [a, b] if every
point in P also belongs to P̃.

and if a and b are real, then |a + ib| ≤ |a| + |b| ≤ 2|a + ib|.
Intuitively, a function of bounded variation cannot oscillate too often
with amplitudes that are too large. Some examples should help clarify
this assertion.
We first fix some terminology. A real-valued function F defined on
[a, b] is increasing if F (t1 ) ≤ F (t2 ) whenever a ≤ t1 ≤ t2 ≤ b. If the
inequality is strict, we say that F is strictly increasing.
Example 1. If F is real-valued, monotonic, and bounded, then F is of
bounded variation. Indeed, if for example F is increasing and bounded
by M , we see that
N
X N
X
|F (tj ) − F (tj−1 )| = F (tj ) − F (tj−1 )
j=1 j=1
= F (b) − F (a) ≤ 2M.
Example 2. If F is differentiable at every point, and F 0 is bounded,

then F is of bounded variation. Indeed, if |F 0 | ≤ M , the mean value
theorem implies
|F (x) − F (y)| ≤ M |x − y|, for all x, y ∈ [a, b],

PN
hence j=1 |F (tj ) − F (tj−1 )| ≤ M (b − a). (See also Exercise 23.)
Example 3. Let
½
xa sin(x−b ) for 0 < x ≤ 1,
F (x) =
0 if x = 0.
Then F is of bounded variation on [0, 1] if and only if a > b (Exercise 11).

Figure 4 illustrates the three cases a > b, a = b, and a < b.
The next result shows that in some sense the first example above ex-
hausts all functions of bounded variation. For its proof, we need the fol-
lowing definitions. The total variation of f on [a, x] (where a ≤ x ≤ b)
is defined by
N
X
TF (a, x) = sup |F (tj ) − F (tj−1 )|,
j=1
a = 2, b = 1
a = 1, b = 1
a = 1/2, b = 1
Figure 4. Graphs of xa sin(x−b ) for different values of a and b
where the sup is over all partitions of [a, x]. The preceding definition
makes sense if F is complex-valued. The succeeding ones require that
F is real-valued. In the spirit of the first definition, we say that the
positive variation of F on [a, x] is
X
PF (a, x) = sup F (tj ) − F (tj−1 ),
(+)
where the sum is over all j such that F (tj ) ≥ F (tj−1 ), and the supremum
is over all partitions of [a, x]. Finally, the negative variation of F on
[a, x] is defined by
X
NF (a, x) = sup −[F (tj ) − F (tj−1 )],
(−)
where the sum is over all j such that F (tj ) ≤ F (tj−1 ), and the supremum
is over all partitions of [a, x].
Lemma 3.2 Suppose F is real-valued and of bounded variation on [a, b].

Then for all a ≤ x ≤ b one has
F (x) − F (a) = PF (a, x) − NF (a, x),
and
TF (a, x) = PF (a, x) + NF (a, x).
Proof. Given ² > 0 there exists a partition a = t0 < · · · < tN = x of

[a, x], such that
¯ ¯ ¯ ¯
¯ X ¯ ¯ X ¯
¯ ¯ ¯ ¯
¯PF − ¯ ¯
F (tj ) − F (tj−1 )¯ < ² and ¯NF − −[F (tj ) − F (tj−1 )]¯¯ < ².
¯
¯ (+) ¯ ¯ (−) ¯
(To see this, it suffices to use the definition to obtain similar estimates
for PF and NF with possibly different partitions, and then to consider a
common refinement of these two partitions.) Since we also note that
X X
F (x) − F (a) = F (tj ) − F (tj−1 ) − −[F (tj ) − F (tj−1 )],
(+) (−)
we find that |F (x) − F (a) − [PF − NF ]| < 2², which proves the first iden-
tity.
For the second identity, we also note that for any partition of a = t0 <
· · · < tN = x of [a, x] we have
N
X X X
|F (tj ) − F (tj−1 )| = F (tj ) − F (tj−1 ) + −[F (tj ) − F (tj−1 )],
j=1 (+) (−)
hence TF ≤ PF + NF . Also, the above implies

X X
F (tj ) − F (tj−1 ) + −[F (tj ) − F (tj−1 )] ≤ TF .
(+) (−)
Once again, one can argue using common refinements of partitions in the
definitions of PF and NF to deduce the inequality PF + NF ≤ TF , and
the lemma is proved.
Theorem 3.3 A real-valued function F on [a, b] is of bounded variation

if and only if F is the difference of two increasing bounded functions.
Proof. Clearly, if F = F1 − F2 , where each Fj is bounded and in-

creasing, then F is of bounded variation.
Conversely, suppose F is of bounded variation. Then, we let F1 (x) =
PF (a, x) + F (a) and F2 (x) = NF (a, x). Clearly, both F1 and F2 are in-
creasing, of bounded variation, and by the lemma F (x) = F1 (x) − F2 (x).
Observe that as a consequence, a complex-valued function of bounded

variation is a (complex) linear combination of four increasing functions.
Returning to the curve γ parametrized by a continuous function z(t) =
x(t) + iy(t), we want to make some comment about its associated length
function. Assuming that the curve is rectifiable, we define L(A, B) as the
length of the segment of γ that arises as the image of those t for which
A ≤ t ≤ B, with a ≤ A ≤ B ≤ b. Note that L(A, B) = TF (A, B), where
F (t) = z(t). We see that
(8) L(A, C) + L(C, B) = L(A, B) if A ≤ C ≤ B.
We also observe that L(A, B) is a continuous function of B (and of

A). Since it is an increasing function, to prove its continuity in B from
the left, it suffices to see that for each B and ² > 0, we can find B1 < B
such that L(A, B1 ) ≥ L(A, B) − ². We do this by first finding a partition
A = t0 < t1 < · · · < tN = B such that the length of the corresponding
polygonal line is ≥ L(A, B) − ²/2. By continuity of the function z(t),
we can find a B1 , with tN −1 < B1 < B, such that |z(B) − z(B1 )| < ²/2.
Now for the refined partition t0 < t1 < · · · < tN −1 < B1 < B, the length
of the polygonal line is still ≥ L(A, B) − ²/2. Therefore, the length
for the partition t0 < t1 < · · · < tN −1 = B1 is ≥ L(A, B) − ², and thus
L(A, B1 ) ≥ L(A, B) − ².
To prove continuity from the right at B, let ² > 0, pick any C > B,
and choose a partition B = t0 < t1 < · · · < tN = C such that L(B, C) −
PN −1
²/2 < j=0 |z(tj+1 ) − z(tj )|. By considering a refinement of this par-
tition if necessary, we may assume since z is continuous that |z(t1 ) −
z(t0 )| < ²/2. If we denote B1 = z(t1 ), then we get
L(B, C) − ²/2 < ²/2 + L(B1 , C).
Since L(B, B1 ) + L(B1 , C) = L(B, C) we have L(B, B1 ) < ², and there-

fore L(A, B1 ) − L(A, B) < ².
Note that what we have observed can be re-stated as follows: if a
function of bounded variation is continuous, then so is its total variation.
The next result lies at the heart of the theory of differentiation.
Theorem 3.4 If F is of bounded variation on [a, b], then F is differen-

tiable almost everywhere.
In other words, the quotient
F (x + h) − F (x)
lim
h→0 h
exists for almost every x ∈ [a, b]. By the previous result, it suffices to
consider the case when F is increasing. In fact, we shall first also assume
that F is continuous. This makes the argument simpler. As for the
general case, we leave that till later. (See Section 3.3.) It will then
be instructive to examine the nature of the possible discontinuities of a
function of bounded variation, and reduce matters to the case of “jump
functions.”
We begin with a nice technical lemma of F. Riesz, which has the effect
of a covering argument.
Lemma 3.5 Suppose G is real-valued and continuous on R. Let E be

the set of points x such that
G(x + h) > G(x) for some h = hx > 0.
If E is non-empty, then it must be open, and hence

S can be written as a
countable disjoint union of open intervals E = (ak , bk ). If (ak , bk ) is a
finite interval in this union, then
G(bk ) − G(ak ) = 0.
Proof. Since G is continuous, it is clear that E is open whenever it is

non-empty and can therefore be written as a disjoint union of countably
many open intervals (Theorem 1.3 in Chapter 1). If (ak , bk ) denotes a
finite interval in this decomposition, then ak ∈/ E; therefore we cannot
have G(bk ) > G(ak ). We now suppose that G(bk ) < G(ak ). By continu-
ity, there exists ak < c < bk so that
G(ak ) + G(bk )
G(c) = ,
2
and in fact we may choose c farthest to the right in the interval (ak , bk ).
Since c ∈ E, there exists d > c such that G(d) > G(c). Since bk ∈ / E, we
must have G(x) ≤ G(bk ) for all x ≥ bk ; therefore d < bk . Since G(d) >
G(c), there exists (by continuity) c0 > d with c0 < bk and G(c0 ) = G(c),
which contradicts the fact that c was chosen farthest to the right in
(ak , bk ). This shows that we must have G(ak ) = G(bk ), and the lemma
is proved.
Note. This result sometimes carries the name “rising sun lemma” for
the following reason. If one thinks of the sun rising from the east (at
the right) with the rays of light parallel to the x-axis, then the points
(x, G(x)) on the graph of G, with x ∈ E, are precisely the points which
are in the shade; these points appear in bold in Figure 5.
Figure 5. Rising sun lemma
A slight modification of the proof of Lemma 3.5 gives:
Corollary 3.6 Suppose G is real-valued and continuous on a closed in-

terval [a, b]. If E denotes the set of points x in (a, b) so that G(x + h) >
G(x) for some h > 0, then E is either empty or open. In the latter
case, it is a disjoint union of countably many intervals (ak , bk ), and
G(ak ) = G(bk ), except possibly when a = ak , in which case we only have
G(ak ) ≤ G(bk ).
For the proof of the theorem, we define the quantity
F (x + h) − F (x)
4h (F )(x) = .
h
We also consider the four Dini numbers at x defined by
D+ (F )(x) = lim sup 4h (F )(x)

h → 0
h > 0
D+ (F )(x) = lim inf 4h (F )(x)

h → 0
h > 0
D− (F )(x) = lim sup 4h (F )(x)

h → 0
h < 0
D− (F )(x) = lim inf 4h (F )(x).

h → 0
h < 0
Clearly, one has D+ ≤ D+ and D− ≤ D− . To prove the theorem it

suffices to show that
(i) D+ (F )(x) < ∞ for a.e. x, and;
(ii) D+ (F )(x) ≤ D− (F )(x) for a.e. x.
Indeed, if these results hold, then by applying (ii) to −F (−x) instead of
F (x) we obtain D− (F )(x) ≤ D+ (F )(x) for a.e. x. Therefore
D+ ≤ D− ≤ D− ≤ D+ ≤ D+ < ∞ for a.e. x.
Thus all four Dini numbers are finite and equal almost everywhere, hence
F 0 (x) exists for almost every point x.
We recall that we assume that F is increasing, bounded, and continu-
ous on [a, b]. For a fixed γ > 0, let
Eγ = {x : D+ (F )(x) > γ}.
First, we assert that Eγ is measurable. (The proof of this simple fact is

outlined in Exercise 14.) Next, we apply Corollary 3.6S to the function
G(x) = F (x) − γx, and note that we then have Eγ ⊂ k (ak , bk ), where
F (bk ) − F (ak ) ≥ γ(bk − ak ). Consequently,
X
m(Eγ ) ≤ m((ak , bk ))
k
1X
≤ F (bk ) − F (ak )
γ
k
1
≤ (F (b) − F (a)).
γ
Therefore m(Eγ ) → 0 as γ tends to infinity, and since {D+ F (x) < ∞} ⊂

Eγ for all γ, this proves that D+ F (x) < ∞ almost everywhere.
Having fixed real numbers r and R such that R > r, we let
E = {x ∈ [a, b] : D+ (F )(x) > R and r > D− (F )(x)}.
We will have shown D+ (F )(x) ≤ D− (F )(x) almost everywhere once we

prove that m(E) = 0, since it then suffices to let R and r vary over the
rationals with R > r.
To prove that m(E) = 0 we may assume that m(E) > 0 and arrive at
a contradiction. Because R/r > 1 we can find an open set O such that
E ⊂ O ⊂ (a, b), yet m(O) < m(E) S · R/r.
Now O can be written as In , with In disjoint open intervals. Fix
n and apply Corollary 3.6 to the function G(x) = −F (−x) + rx on the
interval −In . Reflecting through the origin again yields an open set
S
k (ak , bk ) contained in In , where the intervals (ak , bk ) are disjoint, with
F (bk ) − F (ak ) ≤ r(bk − ak ).
However, on each interval (ak , bk ) we apply CorollaryS3.6, this time to

G(x) = F (x) − Rx. We thus obtain an open set On = k,j (ak,j , bk,j ) of
disjoint open intervals (ak,j , bk,j ) with (ak,j , bk,j ) ⊂ (ak , bk ) for every j,
and
F (bk,j ) − F (ak,j ) ≥ R(bk,j − ak,j ).
Then using the fact that F is increasing we find that

X 1 X
m(On ) = (bk,j − ak,j ) ≤ F (bk,j ) − F (ak,j )
R
k,j k,j
1 X r X
≤ F (bk ) − F (ak ) ≤ (bk − ak )
R R
k k
r
≤ m(In ).
R
Note that On ⊃ E ∩ In , since D+ F (x) > R and r > D− F (x) for each
x ∈ E; of course, In ⊃ On . We now sum in n. Therefore
X X r X r
m(E) = m(E ∩ In ) ≤ m(On ) ≤ m(In ) = m(O) < m(E).
R R
n n
The strict inequality gives a contradiction and Theorem 3.4 is proved, at

least when F is continuous.
Let us see how far we have come regarding (6) if F is a monotonic
function.
Corollary 3.7 If F is increasing and continuous, then F 0 exists almost

everywhere. Moreover F 0 is measurable, non-negative, and
Z b
F 0 (x) dx ≤ F (b) − F (a).
a
In particular, if F is bounded on R, then F 0 is integrable on R.

Proof. For n ≥ 1, we consider the quotient
F (x + 1/n) − F (x)
Gn (x) = .
1/n
By the previous theorem, we have that Gn (x) → F 0 (x) for a.e. x, which
shows in particular that F 0 is measurable and non-negative.
We now extend F as a continuous function on all of R. By Fatou’s
lemma (Lemma 1.7 in Chapter 2) we know that
Z b Z b
F 0 (x) dx ≤ lim inf Gn (x) dx.
a n→∞ a
To complete the proof, it suffices to note that

Z b Z b Z b
1 1
Gn (x) dx = F (x + 1/n) dx − F (x) dx
a 1/n a 1/n a
Z b+1/n Z b
1 1
= F (y) dy − F (x) dx
1/n a+1/n 1/n a
Z b+1/n Z a+1/n
1 1
= F (x) dx − F (x) dx.
1/n b 1/n a
Since F is continuous, the first and second terms converge to F (b) and
F (a), respectively, as n goes to infinity, so the proof of the corollary is
complete.
We cannot go any farther than the inequality in the corollary if we
allow all continuous increasing functions, as is shown by the following
important example.
The Cantor-Lebesgue function

The following simple construction yields a continuous function F : [0, 1] →
[0, 1] that is increasing with F (0) = 0 and F (1) = 1, but F 0 (x) = 0 al-
most everywhere! Hence F is of bounded variation, but
Z b
F 0 (x) dx 6= F (b) − F (a).
a
Consider the standard triadic Cantor set C ⊂ [0, 1] described at the

end of Section 1 in Chapter 1, and recall that
∞
\
C= Ck ,
k=0
where each Ck is a disjoint union of 2k closed intervals. For example,

C1 = [0, 1/3] ∪ [2/3, 1]. Let F1 (x) be the continuous increasing function
on [0, 1] that satisfies F1 (0) = 0, F1 (x) = 1/2 if 1/3 ≤ x ≤ 2/3, F1 (1) = 1,
and F1 is linear on C1 . Similarly, let F2 (x) be continuous and increasing,
and such that


 0 if x = 0,


 1/4 if 1/9 ≤ x ≤ 2/9,
F2 (x) = 1/2 if 1/3 ≤ x ≤ 2/3,



 3/4 if 7/9 ≤ x ≤ 8/9,

1 if x = 1,
and F2 is linear on C2 . See Figure 6.
3/4
1/2
1/4
0 1/9 2/9 1/3 2/3 7/9 8/9 1
Figure 6. Construction of F2
This process yields a sequence of continuous increasing functions

{Fn }∞
n=1 such that clearly
|Fn+1 (x) − Fn (x)| ≤ 2−n−1 .

Hence {Fn }∞ n=1 converges uniformly to a continuous limit F called the
Cantor-Lebesgue function (Figure 7).6 By construction, F is increas-
ing, F (0) = 0, F (1) = 1, and we see that F is constant on each interval
of the complement of the Cantor set. Since m(C) = 0, we find that
F 0 (x) = 0 almost everywhere, as desired.
6 The reader may check that indeed this function agrees with the one given in Exercise 2
of Chapter 1.
0 1
Figure 7. The Cantor-Lebesgue function
The considerations in this section, as well as this last example, show

that the assumption of bounded variation guarantees the existence of a
derivative almost everywhere, but not the validity of the formula
Z b
F 0 (x) dx = F (b) − F (a).
a
In the next section, we shall present a condition on a function that will

completely settle the problem of establishing the above identity.
3.2 Absolutely continuous functions

A function F defined on [a, b] is absolutely continuous if for any ² > 0
there exists δ > 0 so that
N
X N
X
|F (bk ) − F (ak )| < ² whenever (bk − ak ) < δ,
k=1 k=1
and the intervals (ak , bk ), k = 1, . . . , N are disjoint. Some general re-

marks are in order.
• From the definition, it is clear that absolutely continuous functions

are continuous, and in fact uniformly continuous.
• If F is absolutely continuous on a bounded interval, then it is also of
bounded variation on the same interval. Moreover, as is easily seen,
its total variation is continuous (in fact absolutely continuous). As
a consequence the decomposition of such a function F into two
monotonic functions given in Section 3.1 shows that each of these

functions is continuous.
Rx
• If F (x) = a f (y) dy where f is integrable, then F is absolutely
continuous. This follows at once from (ii) in Proposition 1.12,
Chapter 2.
In fact, this last remark shows that absolute continuity is a necessary

Rb
condition to impose on F if we hope to prove a F 0 (x) dx = F (b) − F (a).
Theorem 3.8 If F is absolutely continuous on [a, b], then F 0 (x) exists

almost everywhere. Moreover, if F 0 (x) = 0 for a.e. x, then F is constant.
Since an absolutely continuous function is the difference of two continu-

ous monotonic functions, as we have seen above, the existence of F 0 (x)
for a.e. x follows from what we have already proved. To prove that
F 0 (x) = 0 a.e. implies F is constant requires a more elaborate version of
the covering argument in Lemma 1.2. For the moment we revert to the
generality of d dimensions to describe this.
A collection B of balls {B} is said to be a Vitali covering of a set E
if for every x ∈ E and any η > 0 there is a ball B ∈ B, such that x ∈ B
and m(B) < η. Thus every point is covered by balls of arbitrarily small
measure.
Lemma 3.9 Suppose E is a set of finite measure and B is a Vitali cov-

ering of E. For any δ > 0 we can find finitely many balls B1 , . . . , BN in
B that are disjoint and so that
N
X
m(Bi ) ≥ m(E) − δ.
i=1
Proof. We apply the elementary Lemma 1.2 iteratively, with the

aim of exhausting the set E. It suffices to take δ sufficiently small, say
δ < m(E), and using the just cited covering lemma, we can find an initial
PN1
collection of disjoint balls B1 , B2 , . . . , BN1 in B such that i=1 m(Bi ) ≥
γδ. (For simplicity of notation, we have written γ = 3−d .) Indeed, first
we have m(E 0 ) ≥ δ for an appropriate compact subset E 0 of E. Because
of the compactness of E 0 , we can cover it by finitely many balls from B,
and then the previous lemma allows us toPselect a disjoint sub-collection
N1
of these balls B1 , B2 , . . . , BN1 such that i=1 m(Bi ) ≥ γm(E 0 ) ≥ γδ.
With B1 , . . . , BN1 as our initial sequence of balls, we consider two
PN1
possibilities: either i=1 m(Bi ) ≥ m(E) − δ and we are done with N =
PN1
N1 ; or, contrariwise, i=1 m(Bi ) < m(E) − δ. In the second case, with
SN1
E2 = E − i=1 Bi , we have m(E2 ) > δ (recall that m(Bi ) = m(Bi )). We
then repeat the previous argument, by choosing a compact subset E20 of
E2 with m(E20 ) ≥ δ, and by noting that the balls in B that are disjoint
SN1
from i=1 Bi still cover E2 and in fact give a Vitali covering for E2 , and
hence for E20 . Thus we can chooseP a finite disjoint collection of these
balls Bi , N1 < i ≤ N2 , so that N1 <i≤N2 m(Bi ) ≥ γδ. Therefore, now
PN2
i=1 m(Bi ) ≥ 2γδ, and the balls Bi , 1 ≤ i ≤ N2 , are disjoint.
PN2
We again consider two alternatives, whether or not i=1 m(Bi ) ≥
m(E) − δ. In the first case, we are done with N2 = N , and in the second
case, we proceed as before. If, continuing this way, we had reached the
k th stage and not stopped before then, we would have selected a collection
of disjoint balls with the sum of their measures ≥ kγδ. In any case, our
process achieves the desired goal by the k th stage if k ≥ (m(E) − δ)/γδ,
PNk
since in this case i=1 m(Bi ) ≥ m(E) − δ.
A simple consequence is the following.
Corollary 3.10 We can arrange the choice of the balls so that
N
[
m(E − Bi ) < 2δ.
i=1
In fact, let O be an open set, with O ⊃ E and m(O − E) < δ. Since

we are dealing with a Vitali covering of E, we can restrictSall of our
N
choices above to balls contained in O. If we do this, then (E − i=1 Bi ) ∪
SN
i=1 Bi ⊂ O, where the union on the left-hand side is a disjoint union.
Hence
N
[ N
[
m(E − Bi ) ≤ m(O) − m( Bi ) ≤ m(E) + δ − (m(E) − δ) = 2δ.
i=1 i=1
We now return to the situation on the real line. To complete the proof
of the theorem it suffices to show that under its hypotheses we have
F (b) = F (a), since if that is proved, we can replace the interval [a, b] by
any sub-interval. Now let E be the set of those x ∈ (a, b) where F 0 (x)
exists and is zero. By our assumption m(E) = b − a. Next, momentarily
fix ² > 0. Since for each x ∈ E we have
¯ ¯
¯ F (x + h) − F (x) ¯
lim ¯ ¯ = 0,
h→0 ¯ h ¯
then for each η > 0 we have an open interval I = (ax , bx ) ⊂ [a, b] con-
taining x, with
|F (bx ) − F (ax )| ≤ ²(bx − ax ) and bx − ax < η.
The collection of these intervals forms a Vitali covering of E, and

hence by the lemma, for δ > 0, we can select finitely many Ii , 1 ≤ i ≤ N ,
Ii = (ai , bi ), which are disjoint and such that
N
X
(9) m(Ii ) ≥ m(E) − δ = (b − a) − δ.
i=1
However, |F (bi ) − F (ai )| ≤ ²(bi − ai ), and upon adding these inequalities

we get
N
X
|F (bi ) − F (ai )| ≤ ²(b − a),
i=1
since the intervals Ii are disjoint and lie in [a, b]. Next consider the
SN
complement of j=1 Ij in [a, b]. It consists of finitely many closed in-
SM
tervals k=1 [αk , βk ] with total length ≤ δ because of (9). Thus by the
absolute continuity of F (if δ is chosen appropriately in terms of ²),
PM
k=1 |F (βk ) − F (αk )| ≤ ². Altogether, then,
N
X M
X
|F (b) − F (a)| ≤ |F (bi ) − F (ai )| + |F (βk ) − F (αk )| ≤ ²(b − a) + ².
i=1 k=1
Since ² was positive but otherwise arbitrary, we conclude that F (b) −

F (a) = 0, which we set out to show.
The culmination of all our efforts is contained in the next theorem. In
particular, it resolves our second problem of establishing the reciprocity
between differentiation and integration.
Theorem 3.11 Suppose F is absolutely continuous on [a, b]. Then F 0
exists almost everywhere and is integrable. Moreover,
Z x
F (x) − F (a) = F 0 (y) dy, for all a ≤ x ≤ b.
a
Rb
By selecting x = b we get F (b) − F (a) = a F 0 (y) dy.
Conversely, if f is integrable on [a, b], then there exists an absolutely
continuous function F suchR that F 0 (x) = f (x) almost everywhere, and in
x
fact, we may take F (x) = a f (y) dy.
Proof. Since we know that a real-valued absolutely continuous

function is the difference of two continuous increasing functions,
R x Corol-
lary 3.7 shows that F 0 is integrable on [a, b]. Now let G(x) = a F 0 (y) dy.
Then G is absolutely continuous; hence so is the difference G(x) − F (x).
By the Lebesgue differentiation theorem (Theorem 1.4), we know that
G0 (x) = F 0 (x) for a.e. x; hence the difference F − G has derivative 0 al-
most everywhere. By the previous theorem we conclude that F − G is
constant, and evaluating this expression at x = a gives the desired result.
The converseR x is a consequence of the observation we made earlier,
namely that a f (y) dy is absolutely continuous, and the Lebesgue dif-
ferentiation theorem, which gives F 0 (x) = f (x) almost everywhere.
3.3 Differentiability of jump functions

We now examine monotonic functions that are not assumed to be con-
tinuous. The resulting analysis will allow us to remove the continuity
assumption made earlier in the proof of Theorem 3.4.
As before, we may assume that F is increasing and bounded. In par-
ticular, these two conditions guarantee that the limits
F (x− ) = lim F (y) and F (x+ ) = lim F (y)

y → x y → x
y < x y > x
exist. Then of course F (x− ) ≤ F (x) ≤ F (x+ ), and the function F is

continuous at x if F (x− ) = F (x+ ); otherwise, we say that it has a jump
discontinuity. Fortunately, dealing with these discontinuities is manage-
able, since there can only be countably many of them.
Lemma 3.12 A bounded increasing function F on [a, b] has at most
countably many discontinuities.
Proof. If F is discontinuous at x, we may choose a rational number
rx so that F (x− ) < rx < F (x+ ). If f is discontinuous at x and z with
x < z, we must have F (x+ ) ≤ F (z − ), hence rx < rz . Consequently, to
each rational number corresponds at most one discontinuity of F , hence
F can have at most a countable number of discontinuities.
Now let {xn }∞
n=1 denote the points where F is discontinuous, and let
αn denote the jump of F at xn , that is, αn = F (x+ −
n ) − F (xn ). Then
F (x+ −
n ) = F (xn ) + αn
and
F (xn ) = F (x−
n ) + θn αn , for some θn , with 0 ≤ θn ≤ 1.
If we let

 0 if x < xn ,
jn (x) = θn if x = xn ,

1 if x > xn ,
then we define the jump function associated to F by

∞
X
JF (x) = αn jn (x).
n=1
For simplicity, and when no confusion is possible, we shall write J instead

of JF .
Our first observation is that if F is bounded, then we must have
∞
X
αn ≤ F (b) − F (a) < ∞,
n=1
and hence the series defining J converges absolutely and uniformly.
Lemma 3.13 If F is increasing and bounded on [a, b], then:
(i) J(x) is discontinuous precisely at the points {xn } and has a jump
at xn equal to that of F .
(ii) The difference F (x) − J(x) is increasing and continuous.
Proof. If x 6= xn for all n, each jn is continuous at x, and since the

series converges uniformly, J must be continuous at x. If x = xN for
some N , then we write
N
X ∞
X
J(x) = αn jn (x) + αn jn (x).
n=1 n=N +1
By the same argument as above, the series on the right-hand side is

continuous at x. Clearly, the finite sum has a jump discontinuity at xN
of size αN .
For (ii), we note that (i) implies at once that F − J is continuous.
Finally, if y > x we have
X
J(y) − J(x) ≤ αn ≤ F (y) − F (x),
x<xn ≤y
where the last inequality follows since F is increasing. Hence
F (x) − J(x) ≤ F (y) − J(y),
and the difference F − J is increasing, as desired.

Since we may write F (x) = [F (x) − J(x)] + J(x), our final task is to
prove that J is differentiable almost everywhere.
Theorem 3.14 If J is the jump function considered above, then J 0 (x)
exists and vanishes almost everywhere.
Proof. Given any ² > 0, we note that the set E of those x where
J(x + h) − J(x)
(10) lim sup >²
h→0 h
is a measurable set. (The proof of this little fact is outlined in Exercise 14

below.) Suppose δ = Pm(E). We need to show that δ = 0. Now observe
that since the series αn arising in the definition of J converges,
P then for
any η, to be chosen later, we can find an N so large that n>N αn < η.
We then write
X
J0 (x) = αn jn (x),
n>N
and because of our choice of N we have
(11) J0 (b) − J0 (a) < η.
However, J − J0 is a finite sum of terms αn jn (x), and therefore the set

of points where (10) holds, with J replaced by J0 , differs from E by
at most a finite set, the points {x1 , x2 , . . . , xN }. Thus we can find a
compact set K, with m(K) ≥ δ/2, so that lim suph→0 J0 (x+h)−J h
0 (x)
>²
for each x ∈ K. Hence there are intervals (ax , bx ) containing x, x ∈ K, so
that J0 (bx ) − J0 (ax ) > ²(bx − ax ). We can first choose a finite collection
of these intervals that covers K, and then apply LemmaP1.2 to select
n
intervals I1 , I2 , . . . , In which are disjoint, and for which j=1 m(Ij ) ≥
m(K)/3. The intervals Ij = (aj , bj ) of course satisfy
J0 (bj ) − J0 (aj ) > ²(bj − aj ).
Now,
N
X X ² ²
J0 (b) − J0 (a) ≥ J0 (bj ) − J0 (aj ) > ² (bj − aj ) ≥ m(K) ≥ δ.
3 6
j=1
Thus by (11), ²δ/6 < η, and since we are free to choose η, it follows that
δ = 0 and the theorem is proved.
4 Rectifiable curves and the isoperimetric inequality

We turn to the further study of rectifiable curves and take up first the
validity of the formula
Z b
(12) L= (x0 (t)2 + y 0 (t)2 )1/2 dt,
a
for the length L of the curve parametrized by (x(t), y(t)).

We have already seen that rectifiable curves are precisely the curves
where, besides the assumed continuity of x(t) and y(t), these functions
are of bounded variation. However a simple example shows that for-
mula (12) does not always hold in this context. Indeed, let x(t) = F (t)
and y(t) = F (t), where F is the Cantor-Lebesgue function and 0 ≤ t ≤ 1.
Then this parametrized√ curve traces out the straight line from (0, 0) to
(1, 1) and has length 2, yet x0 (t) = y 0 (t) = 0 for a.e. t.
The integral formula expressing the length of L is in fact valid if we
assume in addition that the coordinate functions of the parametrization
are absolutely continuous.
Theorem 4.1 Suppose (x(t), y(t)) is a curve defined for a ≤ t ≤ b. If

both x(t) and y(t) are absolutely continuous, then the curve is rectifiable,
and if L denotes its length, we have
Z b
L= (x0 (t)2 + y 0 (t)2 )1/2 dt.
a
Note that if F (t) = x(t) + iy(t) is absolutely continuous then it is auto-

matically of bounded variation, and hence the curve is rectifiable. The
identity (12) is an immediate consequence of the proposition below, which
can be viewed as a more precise version of Corollary 3.7 for absolutely
continuous functions.
Proposition 4.2 Suppose F is complex-valued and absolutely continu-

ous on [a, b]. Then
Z b
TF (a, b) = |F 0 (t)| dt.
a
4. Rectifiable curves and the isoperimetric inequality 135
In fact, because of Theorem 3.11, for any partition a = t0 < t1 < · · · <
tN = b of [a, b], we have
¯ ¯
N
X N ¯ Z tj
X ¯
¯ ¯
|F (tj ) − F (tj−1 )| = ¯ F 0 (t) dt¯
¯ tj−1 ¯
j=1 j=1
N Z
X tj
≤ |F 0 (t)| dt
j=1 tj−1
Z b
= |F 0 (t)| dt.
a
So this proves
Z b
(13) TF (a, b) ≤ |F 0 (t)| dt.
a
To prove the reverse inequality, fix ² > 0, and using Theorem 2.4 in
Chapter 2 find a step function g on [a, b], such that F 0 = g + h with
Rb Rx Rx
a
|h(t)| dt < ². Set G(x) = a g(t) dt, and H(x) = a h(t) dt. Then F =
G + H, and as is easily seen
TF (a, b) ≥ TG (a, b) − TH (a, b).
However, by (13) TH (a, b) < ², so that
TF (a, b) ≥ TG (a, b) − ².
Now partition the interval [a, b], as a = t0 < · · · < tN = b, so that the step
function g is constant on each of the intervals (tj−1 , tj ), j = 1, 2, . . . , N .
Then
N
X
TG (a, b) ≥ |G(tj ) − G(tj−1 )|
j=1
¯
N ¯ Z tj
¯
X ¯
¯ ¯
= ¯ g(t) dt¯
¯ tj −1 ¯
j=1
X tjZ
= |g(t)| dt
tj−1
Z b
= |g(t)| dt.
a
Rb Rb
Since a
|g(t)| dt ≥ a
|F 0 (t)| dt − ², we obtain as a consequence that
Z b
TF (a, b) ≥ |F 0 (t)| dt − 2²,
a
and letting ² → 0 we establish the assertion and also the theorem.

Now, any curve (viewed as the image of a mapping t 7→ z(t)) can in
fact be realized by many different parametrizations. A rectifiable curve,
however, has associated to it a unique natural parametrization, the arc-
length parametrization. Indeed, let L(A, B) denote the length function
(considered in Section 3.1), and for the variable t in [a, b] set s = s(t) =
L(a, t). Then s(t), the arc-length, is a continuous increasing function
which maps [a, b] to [0, L], where L is the length of the curve. The arc-
length parametrization of the curve is now given by the pair z̃(s) =
x̃(s) + iỹ(s), where z̃(s) = z(t), for s = s(t). Notice that in this way the
function z̃(s) is well defined on [0, L], since if s(t1 ) = s(t2 ), t1 < t2 , then
in fact z(t) does not vary in the interval [t1 , t2 ] and thus z(t1 ) = z(t2 ).
Moreover |z̃(s1 ) − z̃(s2 )| ≤ |s1 − s2 |, for all pairs s1 , s2 ∈ [0, L], since the
left-hand side of the inequality is the distance between two points on the
curve, while the right-hand side is the length of the portion of the curve
joining these two points. Also, as s varies from 0 to L, z̃(s) traces out
the same points (in the same order) that z(t) does as t varies from a to b.
Theorem 4.3 Suppose (x(t), y(t)), a ≤ t ≤ b, is a rectifiable curve that

has length L. Consider the arc-length parametrization z̃(s) = (x̃(s), ỹ(s))
described above. Then x̃ and ỹ are absolutely continuous, |z̃ 0 (s)| = 1 for
almost every s ∈ [0, L], and
Z L
L= (x̃0 (s)2 + ỹ 0 (s)2 )1/2 ds.
0
Proof. We noted that |z̃(s1 ) − z̃(s2 )| ≤ |s1 − s2 |, so it follows im-

mediately that z̃(s) is absolutely continuous, hence differentiable almost
everywhere. Moreover, this inequality also proves that |z̃ 0 (s)| ≤ 1, for
almost every s. By definition the total variation of z̃ equals L, and by
RL
the previous theorem we must have L = 0 |z̃ 0 (s)| ds. Finally, we note
that this identity is possible only when |z̃ 0 (s)| = 1 almost everywhere.
4.1* Minkowski content of a curve

The proof we give below of the isoperimetric inequality depends in a key
way on the concept of the Minkowski content. While the idea of this
content has an interest on its own right, it is particularly relevant for us

here. This is because the rectifiability of a curve is tantamount to having
(finite) Minkowski content, with that quantity the same as the length of
the curve.
We begin our discussion of these matters with several definitions. A

curve parametrized by z(t) = (x(t), y(t)), a ≤ t ≤ b, is said to be simple
if the mapping t 7→ z(t) is injective for t ∈ [a, b]. It is a closed simple
curve if the mapping t 7→ z(t) is injective for t in [a, b), and z(a) = z(b).
More generally, a curve is quasi-simple if the mapping is injective for t
in the complement of finitely many points in [a, b].
Figure 8. A quasi-simple curve
We shall find it convenient to designate by Γ the pointset traced out by

the curve z(t) as t varies in [a, b], that is, Γ = {z(t) : a ≤ t ≤ b}. For any
compact set K ⊂ R2 (we take K = Γ below), we denote by K δ the open
set that consists of all points at distance (strictly) less than δ from K,
K δ = {x ∈ R2 : d(x, K) < δ}.
Γδ
Figure 9. The curve Γ and the set Γδ

We then say that the set K has Minkowski content7 if the limit
m(K δ )
lim
δ→0 2δ
exists. When this limit exists, we denote it by M(K).
Theorem 4.4 Suppose Γ = {z(t), a ≤ t ≤ b} is a quasi-simple curve. The

Minkowski content of Γ exists if and only if Γ is rectifiable. When this
is the case and L is the length of the curve, then M(Γ) = L.
To prove the theorem, we also consider for any compact set K
m(K δ ) m(K δ )
M∗ (K) = lim sup and M∗ (K) = lim inf
δ→0 2δ δ→0 2δ
(both taken as extended positive numbers). Of course M∗ (K) ≤ M∗ (K).

To say that the Minkowski content exists is the same as saying that
M∗ (K) < ∞ and M∗ (K) = M∗ (K). Their common value is then M(K).
The theorem just stated is the consequence of two propositions con-
cerning M∗ (K) and M∗ (K). The first is as follows.
Proposition 4.5 Suppose Γ = {z(t), a ≤ t ≤ b} is a quasi-simple curve.

If M∗ (Γ) < ∞, then the curve is rectifiable, and if L denotes its length,
then
L ≤ M∗ (Γ).
The proof depends on the following simple observation.
Lemma 4.6 If Γ = {z(t), a ≤ t ≤ b} is any curve, and ∆ = |z(b) − z(a)|

is the distance between its end-points, then m(Γδ ) ≥ 2δ∆.
Proof. Since the distance function and the Lebesgue measure are
invariant under translations and rotations (see Section 3 in Chapter 1
and Problem 4 in Chapter 2) we may transform the situation by an
appropriate composition of these motions. Therefore we may assume
that the end-points of the curve have been placed on the x-axis, and
thus we may suppose that z(a) = (A, 0), z(b) = (B, 0) with A < B, and
∆ = B − A (in the case A = B the conclusion is automatically verified).
By the continuity of the function x(t), there is for each x in [A, B] a
value t in [a, b], such that x = x(t). Since Q = (x(t), y(t)) ∈ Γ, the set
7 This is one-dimensional Minkowski content; variants are in Exercise 28 and also in
Chapter 7 below.
Γδ contains a segment parallel to the y-axis, of length 2δ centered at Q

lying above x (see Figure 10). In other words the slice (Γδ )x contains
the interval (y(t) − δ, y(t) + δ), and hence m1 ((Γδ )x ) ≥ 2δ (where m1 is
the one-dimensional Lebesgue measure). However by Fubini’s theorem
Z Z B
m(Γδ ) = m1 ((Γδ )x ) dx ≥ m1 ((Γδ )x ) dx ≥ 2δ(B − A) = 2δ∆,
R A
and the lemma is proved.
A x = x(t) B
Figure 10. The situation in Lemma 4.6
We now pass to the proof of the proposition. Let us assume first that
the curve is simple. Let P be any partition a = t0 < t1 < · · · < tN = b
of the interval [a, b], and let LP denote the length of the corresponding
polygonal line, that is,
N
X
LP = |z(tj ) − z(tj−1 )|.
j=1
For each ² > 0, the continuity of t 7→ z(t) guarantees the existence of N

proper closed sub-intervals Ij = [aj , bj ] of (tj−1 , tj ), so that
N
X
|z(bj ) − z(aj )| ≥ LP − ².
j=1
Let Γj denote the segment of the curve given by Γj = {z(t); t ∈ Ij }. Since

the closed intervals I1 , . . . , IN are disjoint, it follows by the simplicity of
the curve
SN that the compact SN sets Γ1 , Γ2 , . . . , ΓN are disjoint. However,
Γ ⊃ j=1 Γj and Γδ ⊃ j=1 (Γj )δ . Moreover, the disjointness of the Γj
implies that the sets (Γj )δ are also disjoint for sufficiently small δ. Hence
for those δ, the previous lemma applied to each Γj gives
N
X X
δ
m(Γ ) ≥ m((Γj )δ ) ≥ 2δ |z(bj ) − z(aj )|.
j=1
As a result, m(Γδ )/(2δ) ≥ LP − ², and a passage to the limit gives

M∗ (Γ) ≥ LP − ². Since this inequality is true for all partitions P and
all ² > 0, it implies that the curve is rectifiable and its length does not
exceed M∗ (Γ).
The proof when the curve is merely quasi-simple is similar, except
the partitions P considered must be refined so as to include as partition
points those (finitely many) points in whose complement (in [a, b]) the
mapping t 7→ z(t) is injective. The details may be left to the reader.
The second proposition is in the reverse direction.
Proposition 4.7 Suppose Γ = {z(t), a ≤ t ≤ b} is a rectifiable curve with

length L. Then
M∗ (Γ) ≤ L.
The quantities M∗ (Γ) and L are of course independent of the parametriza-

tion used; since the curve is rectifiable, it will be convenient to use the arc-
length parametrization. Thus we write the curve as z(s) = (x(s), y(s)),
with 0 ≤ s ≤ L, and recall that then z(s) is absolutely continuous and
|z 0 (s)| = 1 for a.e. s ∈ [0, L].
We first fix any 0 < ² < 1, and find a measurable set E² ⊂ R and a
positive number r² such that m(E² ) < ² and
¯ ¯
¯ z(s + h) − z(s) ¯
(14) sup ¯ ¯ − z (s)¯¯ < ² for all s ∈ [0, L] − E² .
0
0<|h|<r² h
Indeed, for each integer n, let

¯ ¯
¯ z(s + h) − z(s) ¯
Fn (s) = sup ¯ − z 0
(s)¯
0<|h|<1/n
¯ h ¯
(where z(s) has been extended outside [0, L], so that z(s) = z(0), when
s < 0, and z(s) = z(L) when s > L). Because z(s) is continuous the
supremum of h in the definition of Fn (s) can be replaced by a supremum
of countably many measurable functions, and hence each Fn is measur-
able. However, Fn (s) → 0, as n → ∞ for a.e s ∈ [a, b]. Thus by Egorov’s
theorem the convergence is uniform outside a set E² with m(E² ) < ²,
and so we merely need to choose r² = 1/n for sufficiently large n to es-

tablish (14). It will be convenient in what follows to assume, as we may,
that z 0 (s) exists and |z 0 (s)| = 1 for every s ∈
/ E² .
Now for any 0 < ρ < r² (with ρ < 1), we partition the interval [0, L]
into consecutive closed intervals, each of length ρ, (except that the last
interval may have length ≤ ρ). Then there is a total of N ≤ L/ρ + 1 such
intervals that arise. We call these intervals I1 , I2 , . . . , IN , and divide them
into two classes. The first class, those intervals Ij we call “good,” are the
ones that enjoy the property that Ij 6⊂ E² . The second class, S those which
are “bad,” have the property that Ij ⊂ E² . As a result, Ij bad Ij ⊂ E² ,
hence the union has measure < ². S
N
We have of course that [0, L] ⊂ j=1 Ij , and if we denote by Γj the
SN
segment of Γ given by {z(s) : s ∈ Ij }, then Γ = j=1 Γj , and as a result
SN PN
Γδ = j=1 (Γj )δ and m(Γδ ) ≤ j=1 m((Γj )δ ).
We consider first the contribution of m((Γj )δ ) when Ij is a good in-
terval. Recall that for such Ij = [aj , bj ] there is an s0 ∈ Ij which is not
in E² , and therefore (14) holds for s = s0 . Let us now visualize Γj by in-
troducing a coordinate system such that z(s0 ) = 0 and z 0 (s0 ) = 1 (which
we may assume after a suitable translation and rotation). We maintain
the notations z(s) and Γj for the so transformed segment of the curve.
aj s0 bj
Γj
aj − s0 − ²ρ aj − s0 0 = z(s0 ) bj − s0 bj − s0 + ²ρ
Figure 11. Estimate of m((Γj )δ ) for a good interval Ij
Note that as h varies over the interval [aj − s0 , bj − s0 ], s0 + h varies

over Ij = [aj , bj ]. Therefore Γj is contained in the rectangle
[aj − s0 − ²ρ, bj − s0 + ²ρ] × [−²ρ, ²ρ],

since |h| ≤ ρ < r² by construction, and |z(s0 + h) − h| < ²|h| by (14). See
Figure 11. Thus (Γj )δ is contained in the rectangle
[aj − s0 − ²ρ − δ, bj − s0 + ²ρ + δ] × [−²ρ − δ, ²ρ + δ],
which has measure ≤ (ρ + 2²ρ + 2δ)(2²ρ + 2δ). Therefore, since ² ≤ 1,

we have
(15) m((Γj )δ ) ≤ 2δρ + O(²δρ + δ 2 + ²ρ2 ),
where the bound arising in O is independent of ², δ, and ρ. This is our

desired estimate for the good intervals.
To pass to the remaining intervals we use the fact that |z(s) − z(s0 )| ≤
|s − s0 | for all s and s0 . Thus in every case Γj is contained in a ball
(disc) of radius ρ, and hence (Γj )δ is contained in a ball of radius ρ + δ.
Therefore we have the crude estimate
(16) m((Γj )δ ) = O(δ 2 + ρ2 ).
We now sum (15) over the good intervals (of which there are at most
L/ρ + 1), and (16) over the bad intervals. There are at most ²/ρ + 1
of the latter kind, since their union is included in E² and this set has
measure < ². Altogether, then,
¡ ¢
m(Γδ ) ≤ 2δL + 2δρ + O(²δ + δ 2 /ρ + ²ρ) + O (²/ρ + 1)(δ 2 + ρ2 ) ,
which simplifies to the inequalities

µ ¶
m(Γδ ) δ ²ρ ²δ ρ2
≤L+O ρ+²+ + + +δ+
2δ ρ δ ρ δ
µ ¶
δ ²ρ ρ2
≤L+O ρ+²+ + + ,
ρ δ δ
where in the last line we have used the fact that ² < 1 and ρ < 1. In
order to obtain a favorable estimate from this as δ → 0, we need to
choose ρ (the length of the sub-intervals) very roughly of the same size
as δ. An effective choice is ρ = δ/²1/2 . If we fix this choice and restrict
our attention to δ for which 0 < δ < ²1/2 r² , then automatically ρ < r² ,
as required by (14). Inserting ρ = δ/²1/2 in the above inequality gives
µ ¶
m(Γδ ) δ δ
≤L+O + ² + ²1/2 + ,
2δ ²1/2 ²
and thus
m(Γδ )
lim sup ≤ L + O(² + ²1/2 ).
δ→0 2δ
Now we can let ² → 0 to obtain the desired conclusion M∗ (Γ) ≤ L, and

the proofs of the proposition and theorem are complete.
4.2* Isoperimetric inequality

The isoperimetric inequality in the plane states, in effect, that among all
curves of a given length it is the circle that encloses the maximum area.
A simple form of this theorem already appeared in Book I. While the
proof given there had the virtue of being brief and elegant, it did suffer
several shortcomings. Among them the “area” in the statement was
defined indirectly via a technical artifice, and the scope of the conclusion
was limited because only relatively smooth curves were considered. Here
we want to remedy those defects and deal with a general version of the
result.
We suppose that Ω is a bounded open subset of R2 , and that its bound-
ary Ω − Ω, is a rectifiable curve Γ, with length `(Γ). We do not require
that Γ be a simple closed curve. The isoperimetric theorem then asserts
the following.
Theorem 4.8 4π m(Ω) ≤ `(Γ)2 .
Proof. For each δ > 0 we consider the outer set
Ω+ (δ) = {x ∈ R2 : d(x, Ω) < δ},
and the inner set
Ω− (δ) = {x ∈ R2 : d(x, Ωc ) ≥ δ}.
Thus Ω− (δ) ⊂ Ω ⊂ Ω+ (δ).

We notice that for Γδ = {x : d(x, Γ) < δ} we have
(17) Ω+ (δ) = Ω− (δ) ∪ Γδ ,
and that this union is disjoint. Moreover, if D(δ) is the open ball (disc)
of radius δ centered at the origin, D(δ) = {x ∈ R2 , |x| < δ}, then clearly
½
Ω+ (δ) ⊃ Ω + D(δ),
(18)
Ω ⊃ Ω− (δ) + D(δ).
Ω+ (δ)
Ω− (δ)
Figure 12. The sets Ω, Ω− (δ) and Ω+ (δ)
We now apply the Brunn-Minkowski inequality (Theorem 5.1 in Chap-

ter 1) to the first inclusion, and obtain
m(Ω+ (δ)) ≥ (m(Ω)1/2 + m(D(δ))1/2 )2 .
Since m(D(δ)) = πδ 2 (this standard formula is established in Exercise 14

in the previous chapter), and (A + B)2 ≥ A2 + 2AB whenever A and B
are positive, we find that
m(Ω+ (δ)) ≥ m(Ω) + 2π 1/2 δ m(Ω)1/2 .
Similarly, m(Ω) ≥ m(Ω− (δ)) + 2π 1/2 δ m(Ω− (δ))1/2 using the second in-
clusion in (18), which implies
−m(Ω− (δ)) ≥ −m(Ω) + 2π 1/2 δ m(Ω− (δ))1/2 .
Now by (17)
m(Γδ ) = m(Ω+ (δ)) − m(Ω− (δ)),
and by the inequalities above, we have
m(Γδ ) ≥ 2π 1/2 δ(m(Ω)1/2 + m(Ω− (δ))1/2 ).
We now divide both sides by 2δ and take the limsup as δ → 0. This

yields
M∗ (Γ) ≥ π 1/2 (2m(Ω)1/2 ),

5. Exercises 145
since Ω− (δ) % Ω as δ → 0. However, by Proposition 4.7, `(Γ) ≥ M∗ (Γ),

so
`(Γ) ≥ 2π 1/2 m(Ω)1/2 ,
which proves the theorem.

Remark. A similar result holds even without the assumption that the
boundary is a (rectifiable) curve. In fact the proof shows that for any
bounded open set Ω whose boundary is Γ we have
4π m(Ω) ≤ M∗ (Γ)2 .
5 Exercises
R
1. Suppose ϕ is an integrable function on Rd with Rd
ϕ(x) dx = 1. Set Kδ (x) =
δ −d ϕ(x/δ), δ > 0.
(a) Prove that {Kδ }δ>0 is a family of good kernels.
(b) Assume in addition that ϕ is bounded and supported in a bounded set.

Verify that {Kδ }δ>0 is an approximation to the identity.
(c) Show that Theorem 2.3 (convergence in the L1 -norm) holds for good kernels
as well.
2. Suppose {Kδ } is a family of kernels that satisfies:
(i) |Kδ (x)| ≤ Aδ −d for all δ > 0.
(ii) |Kδ (x)| ≤ Aδ/|x|d+1 for all δ > 0.

R∞
(iii) −∞ Kδ (x) dx = 0 for all δ > 0.
Thus Kδ satisfies conditions (i) and (ii) of approximations to the identity, but the
average value of Kδ is 0 instead of 1. Show that if f is integrable on Rd , then
(f ∗ Kδ )(x) → 0 for a.e. x, as δ → 0.
3. Suppose 0 is a point of (Lebesgue) density of the set E ⊂ R. Show that for each
of the individual conditions below there is an infinite sequence of points xn ∈ E,
with xn 6= 0, and xn → 0 as n → ∞.
(a) The sequence also satisfies −xn ∈ E for all n.
(b) In addition, 2xn belongs to E for all n.

Generalize.
4. Prove that if f is integrable on Rd , and f is not identically zero, then
c
f ∗ (x) ≥ , for some c > 0 and all |x| ≥ 1.
|x|d
Conclude that f ∗ is not integrable on Rd . Then, show that the weak type estimate
m({x : f ∗ (x) > α}) ≤ c/α

R
for all α > 0 whenever |f | = 1,
R is best possible in the following sense: if f is
supported in the unit ball with |f | = 1, then
m({x : f ∗ (x) > α}) ≥ c0 /α
for some c0 > 0 and all sufficiently small α.

R
[Hint: For the first part, use the fact that B |f | > 0 for some ball B.]
5. Consider the function on R defined by

8
< 1
if |x| ≤ 1/2,
f (x) = |x|(log 1/|x|)2
: 0 otherwise.
(a) Verify that f is integrable.
(b) Establish the inequality
c
f ∗ (x) ≥ for some c > 0 and all |x| ≤ 1/2,
|x|(log 1/|x|)
to conclude that the maximal function f ∗ is not locally integrable.
6. In one dimension there is a version of the basic inequality (1) for the maximal
function in the form of an identity. We define the “one-sided” maximal function
Z x+h
∗ 1
f+ (x) = sup |f (y)| dy.
h>0 h x
If Eα+ = {x ∈ R : f+
∗
(x) > α}, then
Z
1
m(Eα+ ) = |f (y)| dy.
α +
Eα
Rx
[Hint: Apply Lemma 3.5 to F (x) = 0 |f (y)| dy − αx. Then Eα+ is the union of
R bk
disjoint intervals (ak , bk ) with a |f (y)| dy = α(ak − bk ).]
k
5. Exercises 147
7. Using Corollary 1.5, prove that if a measurable subset E of [0, 1] satisfies

m(E ∩ I) ≥ α m(I) for some α > 0 and all intervals I in [0, 1], then E has measure
1. See also Exercise 28 in Chapter 1.
8. Suppose A is a Lebesgue measurable set in R with

S∞ m(A) > 0. Does there exist
a sequence {sn }∞
n=1 such that the complement of n=1 (A + sn ) in R has measure
zero?
[Hint: For every ² > 0,Sfind an interval I² of length `² such that m(A ∩ I² ) ≥
(1 − ²)m(I² ). Consider ∞k=−∞ (A + tk ), with tk = k`² . Then vary ².]
9. Let F be a closed subset in R, and δ(x) the distance from x to F , that is,
δ(x) = d(x, F ) = inf{|x − y| : y ∈ F }.
Clearly, δ(x + y) ≤ |y| whenever x ∈ F . Prove the more refined estimate
δ(x + y) = o(|y|) for a.e. x ∈ F ,
that is, δ(x + y)/|y| → 0 for a.e. x ∈ F .

[Hint: Assume that x is a point of density of F .]
10. Construct an increasing function on R whose set of discontinuities is pre-

cisely Q.
11. If a, b > 0, let


xa sin(x−b ) for 0 < x ≤ 1,
f (x) =
0 if x = 0.
Prove that f is of bounded variation in [0, 1] if and only if a > b. Then, by tak-
ing a = b, construct (for each 0 < α < 1) a function that satisfies the Lipschitz
condition of exponent α
|f (x) − f (y)| ≤ A|x − y|α
but which is not of bounded variation.

[Hint: Note that if h > 0, the difference |f (x + h) − f (x)| can be estimated by
C(x + h)a , or C 0 h/x by the mean value theorem. Then, consider two cases,
whether xa+1 ≥ h or xa+1 < h. What is the relationship between α and a?]
12. Consider the function F (x) = x2 sin(1/x2 ), x 6= 0, with F (0) = 0. Show that
F 0 (x) exists for every x, but F 0 is not integrable on [−1, 1].
13. Show directly from the definition that the Cantor-Lebesgue function is not
absolutely continuous.
14. The following measurability issues arose in the discussion of differentiability

of functions.
(a) Suppose F is continuous on [a, b]. Show that
F (x + h) − F (x)
D+ (F )(x) = lim sup
h → 0 h
h > 0
is measurable.
P∞
(b) Suppose J(x) = n=1 αn jn (x) is a jump function as in Section 3.3. Show
that
J(x + h) − J(x)
lim sup
h→0 h
is measurable.
[Hint: For (a), the continuity of F allows one to restrict to countably

˛ many h in tak-˛
˛ N (x) ˛
ing the limsup. For (b), given k > m, let Fk,m = sup1/k≤|h|≤1/m ˛ JN (x+h)−J
N
h ˛,
PN N
where JN (x) = n=1 αn jn (x). Note that each Fk,m is measurable. Then, succes-
sively, let N → ∞, k → ∞, and finally m → ∞.]
15. Suppose F is of bounded variation and continuous. Prove that F = F1 − F2 ,

where both F1 and F2 are monotonic and continuous.
16. Show that if F is of bounded variation in [a, b], then:

Rb
(a) a
|F 0 (x)| dx ≤ TF (a, b).
Rb
(b) a
|F 0 (x)| dx = TF (a, b) if and only if F is absolutely continuous.
Rb
As a result of (b), the formula L = a |z 0 (t)| dt for the length of a rectifiable curve
parametrized by z holds if and only if z is absolutely continuous.
17. Prove that if {K² }²>0 is a family of approximations to the identity, then
sup |(f ∗ K² )(x)| ≤ cf ∗ (x)

²>0
for some constant c > 0 and all integrable f .
18. Verify the agreement between the two definitions given for the Cantor-Lebesgue
function in Exercise 2, Chapter 1 and in Section 3.1 of this chapter.
19. Show that if f : R → R is absolutely continuous, then
(a) f maps sets of measure zero to sets of measure zero.
(b) f maps measurable sets to measurable sets.

5. Exercises 149
20. This exercise deals with functions F that are absolutely continuous on [a, b]
and are increasing. Let A = F (a) and B = F (b).
(a) There exists such an F that is in addition strictly increasing, but such that
F 0 (x) = 0 on a set of positive measure.
(b) The F in (a) can be chosen so that there is a measurable subset E ⊂ [A, B],
m(E) = 0, so that F −1 (E) is not measurable.
(c) Prove, however, that for any increasing absolutely continuous F , and E a
measurable subset of [A, B], the set F −1 (E) ∩ {F 0 (x) > 0} is measurable.
Rx
[Hint: (a) Let F (x) = a χK (x) dx, where K is the complement of a Cantor-like
set C of positive measure. For (b),
R note that F (C) is a set of measure zero. Finally,
for (c) prove first that m(O) = F −1 (O) F 0 (x) dx for any open set O.]
21. Let F be absolutely continuous and increasing on [a, b] with F (a) = A and
F (b) = B. Suppose f is any measurable function on [A, B].
(a) Show that f (F (x))F 0 (x) is measurable on [a, b]. Note: f (F (x)) need not be
measurable by Exercise 20 (b).
(b) Prove the change of variable formula: If f is integrable on [A, B], then so is
f (F (x))F 0 (x), and
Z B Z b
f (y) dy = f (F (x))F 0 (x) dx.
A a
R
[Hint: Start with the identity m(O) = F −1 (O)
F 0 (x) dx used in (c) of Exercise 20
above.]
22. Suppose that F and G are absolutely continuous on [a, b]. Show that their
product F G is also absolutely continuous. This has the following consequences.
(a) Whenever F and G are absolutely continuous in [a, b],

Z b Z b
F 0 (x)G(x) dx = − F (x)G0 (x) dx + [F (x)G(x)]ba .
a a
(b) Let F be absolutely continuous in [−π, π] with F (π) = F (−π). Show that
if
Z π
1
an = F (x)e−inx dx,
2π −π
P
such that F (x) ∼ an einx , then
X
F 0 (x) ∼ inan einx .
(c) What happens if F (−π) 6= F (π)? [Hint: Consider F (x) = x.]
23. Let F be continuous on [a, b]. Show the following.
(a) Suppose (D+ F )(x) ≥ 0 for every x ∈ [a, b]. Then F is increasing on [a, b].
(b) If F 0 (x) exists for every x ∈ (a, b) and |F 0 (x)| ≤ M , then |F (x) − F (y)| ≤
M |x − y| and F is absolutely continuous.
[Hint: For (a) it suffices to show that F (b) − F (a) ≥ 0. Assume otherwise. Hence
with G² (x) = F (x) − F (a) + ²(x − a), for sufficiently small ² > 0 we have G² (a) =
0, but G² (b) < 0. Now let x0 ∈ [a, b) be the greatest value of x0 such that G² (x0 ) ≥
0. However, (D+ G² )(x0 ) > 0.]
24. Suppose F is an increasing function on [a, b].
(a) Prove that we can write
F = F A + FC + FJ ,
where each of the functions FA , FC , and FJ is increasing and:
(i) FA is absolutely continuous.

(ii) FC is continuous, but FC0 (x) = 0 for a.e. x.
(iii) FJ is a jump function.
(b) Moreover, each component FA , FC , FJ is uniquely determined up to an

additive constant.
The above is the Lebesgue decomposition of F . There is a corresponding

decomposition for any F of bounded variation.
25. The following shows the necessity of allowing for general exceptional sets of
measure zero in the differentiation Theorems 1.4, 3.4, and 3.11. Let E be any set
of measure zero in Rd . Show that:
(a) There exists a non-negative integrable f in Rd , so that

Z
1
lim inf f (y) dy = ∞ for each x ∈ E.
m(B) → 0 m(B) B
x ∈ B
(b) When d = 1 this may be restated as follows. There is an increasing abso-

lutely continuous function F so that
D+ (F )(x) = D− (F )(x) = ∞, for each x ∈ E.

5. Exercises 151
P∞
[Hint: Find open sets On ⊃ E, with m(On ) < 2−n , and let f (x) = n=1 χOn (x).]
26. An alternative way of defining the exterior measure m∗ (E) of an arbitrary set
E, as given in Section 2 of Chapter 1, is to replace the coverings ofPE by cubes
∞
with coverings by balls. That is, suppose we define S∞ mB∗ (E) as inf j=1 m(Bj ),
where the infimum is taken over all coverings E ⊂ j=1 Bj by open balls. Then
m∗ (E) = mB ∗ (E). (Observe that this result leads to an alternate proof that the
Lebesgue measure is invariant under rotations.)
Clearly m∗ (E) ≤ mB ∗ (E). Prove the reverse inequality by showingSthe follow-
ing.
P For any ² > 0, there is a collection of balls {Bj } such that E ⊂ j Bj while
j m(Bj ) ≤ m∗ (E) + ². Note also that for any preassigned δ, we can choose the
balls to have diameter < δ.
[Hint: Assume first that E is measurable, and pick O open so that O ⊃ E and
0
m(O
PN − E) < ² . Next, using Corollary
S 3.10, find balls B1 , . . . , BN such
SNthat
j=1 m(B j ) ≤ m(E) + 2²0 and m(E − N 0
j=1 Bj ) ≤ 3² . Finally, cover E − j=1 Bj
by a union of cubes, the sum of whose measures is ≤ 4²0 , and replace these cubes
by balls that contain them. For the general E, begin by applying the above when
E is a cube.]
27. A rectifiable curve has a tangent line at almost all points of the curve. Make
this statement precise.
28. A curve in Rd is a continuous map t 7→ z(t) of an interval [a, b] into Rd .
(a) State and prove the analogues of the conditions dealing with the rectifiability
of curves and their length that are given in Theorems 3.1, 4.1, and 4.3.
(b) Define the (one-dimensional) Minkowski content M(K) of a compact set in

Rd as the limit (if it exists) of
m(K δ )
as δ → 0,
md−1 (B(δ))
where md−1 (B(δ)) is the measure (in Rd−1 ) of the ball defined by B(δ) =
{x ∈ Rd−1 , |x| < δ}. State and prove analogues of Propositions 4.5 and 4.7
for curves in Rd .
29. Let Γ = {z(t), a ≤ t ≤ b} be a curve, and suppose it satisfies a Lipschitz

condition with exponent α, 1/2 ≤ α ≤ 1, that is,
|z(t) − z(t0 )| ≤ A|t − t0 |α for all t, t0 ∈ [a, b].
Show that m(Γδ ) = O(δ 2−1/α ) for 0 < δ ≤ 1.
30. A bounded function F is said to be of bounded variation on R if F is of

bounded variation on any finite sub-interval [a, b], and supa,b TF (a, b) < ∞.
Prove that such an F enjoys the following two properties:
R
(a) R
|F (x + h) − F (x)| dx ≤ A|h|, for some constant A and all h ∈ R.
R
(b) | R F (x)ϕ0 (x) dx| ≤ A, where ϕ ranges over all C 1 functions of bounded
support with supx∈R |ϕ(x)| ≤ 1.
For the converse, and analogues in Rd , see Problem 6∗ below.

[Hint: For (a), write F = F1 − F2 , where Fj are monotonic and bounded. For (b),
deduce this from (a).]
31. Let F be the Cantor-Lebesgue function described in Section 3.1. Consider the
curve that is the graph of F , that is, the curve given by x(t) = t and y(t) = F (t)
with 0 ≤ t ≤ 1. Prove that the length L(x) of the segment 0 ≤ t ≤ x of the curve
is given by L(x) = x + F (x). Hence the total length of the curve is 2.
32. Let f : R → R. Prove that f satisfies the Lipschitz condition
|f (x) − f (y)| ≤ M |x − y|
for some M and all x, y ∈ R, if and only if f satisfies the following two properties:
(i) f is absolutely continuous.
(ii) |f 0 (x)| ≤ M for a.e. x.
6 Problems
1. Prove the following variant of the Vitali covering lemma: If E is covered in

the Vitali sense by a family B of balls, and 0 < m∗ (E) < ∞, then for every η > 0
there exists a disjoint collection of balls {Bj }∞
j=1 in B such that
∞
! ∞
[ X
m∗ E/ Bj =0 and |Bj | ≤ (1 + η)m∗ (E).
j=1 j=1
2. The following simple one-dimensional covering lemma can be used in a number

of different situations.
Suppose I1 , I2 , . . . , IN is a given finite collection of open intervals in R. Then
there are two finite sub-collections I10 , I20 , . . . , IK
0
, and I100 , I200 , . . . , IL00 , so that each
sub-collection consists of mutually disjoint intervals and
N
[ K
[ L
[
Ij = Ik0 ∪ I`00 .
j=1 k=1 `=1
Note that, in contrast with Lemma 1.2, the full union is covered and not merely a
part.
6. Problems 153
[Hint: Choose I10 to be an interval whose left end-point is as far left as possible.
Discard all intervals contained in I10 . If the remaining intervals are disjoint from
I10 , select again an interval as far to the left as possible, and call it I20 . Otherwise
choose an interval that intersects I10 , but reaches out to the right as far as possible,
and call this interval I100 . Repeat this procedure.]
3.∗ There is no direct analogue of Problem 2 in higher dimensions. However, a full

covering is afforded by the Besicovitch covering lemma. A version of this lemma
states that there is an integer N (dependent only on the dimension d) with the
following property. Suppose E is any bounded set in Rd that is covered by a
collection B of balls in the (strong) sense that for each x ∈ E, there is a B ∈ B
whose center is x. Then, there are N sub-collections B1 , B2 , . . . , BN of the original
collection B, such that each Bj is a collection of disjoint balls, and moreover,
[
E⊂ B, where B0 = B1 ∪ B2 ∪ · · · ∪ BN .
B∈B0
4. A real-valued function ϕ defined on an interval (a, b) is convex if the region

lying above its graph {(x, y) ∈ R2 : y > ϕ(x), a ≤ x ≤ b} is a convex set, as defined
in Section 5*, Chapter 1. Equivalently, ϕ is convex if
ϕ(θx1 + (1 − θ)x2 ) ≤ θϕ(x1 ) + (1 − θ)ϕ(x2 )
for every x1 , x2 ∈ (a, b) and 0 ≤ θ ≤ 1. One can also observe as a consequence that
we have the following inequality of the slopes:
ϕ(x + h) − ϕ(x) ϕ(y) − ϕ(x) ϕ(y) − ϕ(y − h)
≤ ≤ ,
h y−x h
whenever x < y, h > 0, and x + h < y.
The following can then be proved.
(a) ϕ is continuous on (a, b).
(b) ϕ satisfies a Lipschitz condition of order 1 in any proper closed sub-interval
[a0 , b0 ] of (a, b). Hence ϕ is absolutely continuous in each sub-interval.
(c) ϕ0 exists at all but an at most denumerable number of points, and ϕ0 = D+ ϕ
is an increasing function with
Z y
ϕ(y) − ϕ(x) = ϕ0 (t) dt.
x
Rx
(d) Conversely, if ψ is any increasing function on (a, b), then ϕ(x) = c
ψ(t) dt
is a convex function in (a, b) (for c ∈ (a, b)).
5. Suppose that F is continuous on [a, b], F 0 (x) exists for every x ∈ (a, b), and
F 0 (x) is integrable. Then F is absolutely continuous and
Z b
F (b) − F (a) = F 0 (x) dx.
a
x x+h y−h y
Figure 13. A convex function
[Hint: Assume F 0 (x) ≥ 0 for a.e. x. We want to conclude that F (b) ≥ F (a). Let
E be the set of measure 0 of those x such that F 0 (x) < 0. Then according to
Exercise 25, there is a function Φ which is increasing, absolutely continuous, and for
which D+ Φ(x) = ∞, x ∈ E. Consider F + δΦ, for each δ and apply the result (a)
in Exercise 23.]
6.∗ The following converse to Exercise 30 characterizes functions of bounded vari-

ation.
Suppose F is a bounded measurable function on R. If F satisfies either of
conditions (a) or (b) in that exercise, then F can be modified on a set of measure
zero so as to become a function of bounded variation on R.
Moreover, on Rd we have the following assertion. Suppose F is a bounded
measurable function on Rd . Then the following two conditions on F are equivalent:
R
(a0 ) Rd |F (x + h) − F (x)| dx ≤ A|h|, for all h ∈ Rd .
R ∂ϕ
(b0 ) | Rd
F (x) ∂xj
dx| ≤ A, for all j = 1, . . . , d,
for all ϕ ∈ C 1 that have bounded support, and for which supx∈Rd |ϕ(x)| ≤ 1.
The class of functions that satisfy either (a0 ) or (b0 ) is the extension to Rd of
the class of functions of bounded variation.
7. Consider the function

∞
X n
f1 (x) = 2−n e2πi2 x
.
n=0
(a) Prove that f1 satisfies |f1 (x) − f1 (y)| ≤ Aα |x − y|α for each 0 < α < 1.
(b)∗ However, f1 is nowhere differentiable, hence not of bounded variation.

6. Problems 155
8.∗ Let R denote the set of all rectangles in R2 that contain the origin, and with
sides parallel to the coordinate axis. Consider the maximal operator associated to
this family, namely
Z
∗ 1
fR (x) = sup |f (x − y)| dy.
R∈R m(R) R
∗
(a) Then, f 7→ fR does not satisfy the weak type inequality
∗ A
m({x : fR (x) > α}) ≤ kf kL1
α
for all α > 0, all integrable f , and some A > 0.
(b) Using this, one can show that there exists f ∈ L1 (R) so that for R ∈ R
Z
1
lim sup f (x − y) dy = ∞ for almost every x.
diam(R)→0 m(R) R
Here diam(R) = supx ,y∈R |x − y| equals the diameter of the rectangle.
[Hint: For part (a), let B be the unit ball, and consider the function ϕ(x) =
χB (x)/m(B). For δ > 0, let ϕδ (x) = δ −2 ϕ(x/δ). Then
1
(ϕδ )∗R (x) → as δ → 0,
|x1 | |x2 |
for every (x1 , x2 ), with x1 x2 6= 0. If the weak type inequality held, then we would
have
A
m({|x| ≤ 1 : |x1 x2 |−1 > α}) ≤ .
α
This is a contradiction since the left-hand side is of the order of (log α)/α as α
tends to infinity.]
4 Hilbert Spaces: An
Introduction
Born barely 10 years ago, the theory of integral equa-

tions has attracted wide attention as much as for its
inherent interest as for the importance of its applica-
tions. Several of its results are already classic, and no
one doubts that in a few years every course in analysis
will devote a chapter to it.
M. Plancherel, 1912
There are two reasons that account for the importance of Hilbert
spaces. First, they arise as the natural infinite-dimensional generaliza-
tions of Euclidean spaces, and as such, they enjoy the familiar properties
of orthogonality, complemented by the important feature of complete-
ness. Second, the theory of Hilbert spaces serves both as a conceptual
framework and as a language that formulates some basic arguments in
analysis in a more abstract setting.
For us the immediate link with integration theory occurs because of
the example of the Lebesgue space L2 (Rd ). The related example of
L2 ([−π, π]) is what connects Hilbert spaces with Fourier series. The
latter Hilbert space can also be used in an elegant way to analyze the
boundary behavior of bounded holomorphic functions in the unit disc.
A basic aspect of the theory of Hilbert spaces, as in the familiar finite-
dimensional case, is the study of their linear transformations. Given the
introductory nature of this chapter, we limit ourselves to rather brief
discussions of several classes of such operators: unitary mappings, pro-
jections, linear functionals, and compact operators.
1 The Hilbert space L2

A prime example of a Hilbert space is the collection of square inte-
grable functions on Rd , which is denoted by L2 (Rd ), and consists of
all complex-valued measurable functions f that satisfy
Z
|f (x)|2 dx < ∞.
Rd
1. The Hilbert space L2 157
The resulting L2 (Rd )-norm of f is defined by

µZ ¶1/2
2
kf kL2 (Rd ) = |f (x)| dx .
Rd
The reader should compare those definitions with these for the space
L1 (Rd ) of integrable functions and its norm that were described in Sec-
tion 2, Chapter 2. A crucial difference is that L2 has an inner product,
which L1 does not. Some relative inclusion relations between those spaces
are taken up in Exercise 5.
The space L2 (Rd ) is naturally equipped with the following inner prod-
uct:
Z
(f, g) = f (x)g(x) dx, whenever f, g ∈ L2 (Rd ),
Rd
which is intimately related to the L2 -norm since
(f, f )1/2 = kf kL2 (Rd ) .
As in the case of integrable functions, the condition kf kL2 (Rd ) = 0 only

implies f (x) = 0 almost everywhere. Therefore, we in fact identify func-
tions that are equal almost everywhere, and define L2 (Rd ) as the space
of equivalence classes under this identification. However, in practice it is
often convenient to think of elements in L2 (Rd ) as functions, and not as
equivalence classes of functions.
For the definition of the inner product (f, g) to be meaningful we need
to know that f g is integrable on Rd whenever f and g belong to L2 (Rd ).
This and other basic properties of the space of square integrable functions
are gathered in the next proposition.
In the rest of this chapter we shall denote the L2 -norm by k · k (drop-
ping the subscript L2 (Rd )) unless stated otherwise.
Proposition 1.1 The space L2 (Rd ) has the following properties:
(i) L2 (Rd ) is a vector space.
(ii) f (x)g(x) is integrable whenever f, g ∈ L2 (Rd ), and the Cauchy-

Schwarz inequality holds: |(f, g)| ≤ kf k kgk.
(iii) If g ∈ L2 (Rd ) is fixed, the map f 7→ (f, g) is linear in f , and also

(f, g) = (g, f ).
(iv) The triangle inequality holds: kf + gk ≤ kf k + kgk.

158 Chapter 4. HILBERT SPACES: AN INTRODUCTION
Proof. If f, g ∈ L2 (Rd ), then since |f (x) + g(x)| ≤ 2 max(|f (x)|, |g(x)|),

we have
|f (x) + g(x)|2 ≤ 4(|f (x)|2 + |g(x)|2 ),
therefore
Z Z Z
|f + g|2 ≤ 4 |f |2 + 4 |g|2 < ∞,
hence f + g ∈ L2 (Rd ). Also, if λ ∈ C we clearly have λf ∈ L2 (Rd ), and

part (i) is proved.
To see why f g is integrable whenever f and g are in L2 (Rd ), it suffices
to recall that for all A, B ≥ 0, one has 2AB ≤ A2 + B 2 , so that
Z
1£ ¤
(1) |f g| ≤ kf k2 + kgk2 .
2
To prove the Cauchy-Schwarz inequality, we first observe that if either

kf k = 0 or kgk = 0, then f g = 0 is zero almost everywhere, hence (f, g) =
0 and the inequality is obvious. Next, if we assume that kf k = kgk = 1,
then we get theR desired inequality |(f, g)| ≤ 1. This follows from the fact
that |(f, g)| ≤ |f g|, and inequality (1). Finally, in the case when both
kf k and kgk are non-zero, we normalize f and g by setting
f˜ = f /kf k and g̃ = g/kgk,
so that kf˜k = kg̃k = 1. By our previous observation we then find
|(f˜, g̃)| ≤ 1.
Multiplying both sides of the above by kf k kgk yields the Cauchy-Schwarz

inequality.
Part (iii) follows from the linearity of the integral.
Finally, to prove the triangle inequality, we use the Cauchy-Schwarz
inequality as follows:
kf + gk2 = (f + g, f + g)
= kf k2 + (f, g) + (g, f ) + kgk2
≤ kf k2 + 2 |(f, g)| + kgk2
≤ kf k2 + 2 kf k kgk + kgk2
= (kf k + kgk)2 ,
and taking square roots completes the argument.

1. The Hilbert space L2 159
We turn our attention to the notion of a limit in the space L2 (Rd ).

The norm on L2 induces a metric d as follows: if f, g ∈ L2 (Rd ), then
d(f, g) = kf − gkL2 (Rd ) .
A sequence {fn } ⊂ L2 (Rd ) is said to be Cauchy if d(fn , fm ) → 0 as

n, m → ∞. Moreover, this sequence converges to f ∈ L2 (Rd ) if d(fn , f ) →
0 as n → ∞.
Theorem 1.2 The space L2 (Rd ) is complete in its metric.
In other words, every Cauchy sequence in L2 (Rd ) converges to a function

in L2 (Rd ). This theorem, which is in sharp contrast with the situation for
Riemann integrable functions, is a graphic illustration of the usefulness
of Lebesgue’s theory of integration. We elaborate on this point and its
relation to Fourier series in Section 3 below.
Proof. The argument given here follows closely the proof in Chapter 2
that L1 is complete. Let {fn }∞ 2
n=1 be a Cauchy sequence in L , and
∞
consider a subsequence {fnk }k=1 of {fn } with the following property:
kfnk+1 − fnk k ≤ 2−k , for all k ≥ 1.
If we now consider the series whose convergence will be seen below,

∞
X
f (x) = fn1 (x) + (fnk+1 (x) − fnk (x))
k=1
and
∞
X
g(x) = |fn1 (x)| + |(fnk+1 (x) − fnk (x))|,
k=1
together the partial sums
K
X
SK (f )(x) = fn1 (x) + (fnk+1 (x) − fnk (x))
k=1
and
K
X
SK (g)(x) = |fn1 (x)| + |fnk+1 (x) − fnk (x)|,
k=1

K
X
kSK (g)k ≤ kfn1 k + kfnk+1 − fnk k
k=1
K
X
≤ kfn1 k + 2−k .
k=1
Letting K tend to R infinity, and applying the monotone convergence theo-

rem proves that |g|2 < ∞, and since |f | ≤ g, we must have f ∈ L2 (Rd ).
In particular, the series defining f converges almost everywhere, and
since (by construction of the telescopic series) the (K − 1)th partial sum
of this series is precisely fnK , we find that
fnk (x) → f (x) a.e. x.
To prove that fnk → f in L2 (Rd ) as well, we simply observe that |f −

SK (f )|2 ≤ (2g)2 for all K, and apply the dominated convergence theorem
to get kfnk − f k → 0 as k tends to infinity.
Finally, the last step of the proof consists of recalling that {fn } is
Cauchy. Given ², there exists N such that for all n, m > N we have
kfn − fm k < ²/2. If nk is chosen so that nk > N , and kfnk − f k < ²/2,
kfn − f k ≤ kfn − fnk k + kfnk − f k < ²
whenever n > N . This concludes the proof of the theorem.

An additional useful property of L2 (Rd ) is contained in the following
theorem.
Theorem 1.3 The space L2 (Rd ) is separable, in the sense that there
exists a countable collection {fk } of elements in L2 (Rd ) such that their
linear combinations are dense in L2 (Rd ).
Proof. Consider the family of functions of the form rχR (x), where r
is a complex number with rational real and imaginary parts, and R is
a rectangle in Rd with rational coordinates. We claim that finite linear
combinations of these type of functions are dense in L2 (Rd ).
Suppose f ∈ L2 (Rd ) and let ² > 0. Consider for each n ≥ 1 the func-
tion gn defined by
½
f (x) if |x| ≤ n and |f (x)| ≤ n,
gn (x) =
0 otherwise.
2. Hilbert spaces 161
Then |f − gn |2 ≤ 4|f |2 and gn (x) → f (x) almost everywhere.1 The dom-

inated convergence theorem implies that kf − gn k2L2 (Rd ) → 0 as n tends
to infinity; therefore we have
kf − gN kL2 (Rd ) < ²/2 for some N .
Let g = gN , and note that g is a bounded function supported on a

bounded set; thusRg ∈ L1 (Rd ). We may now find a step function ϕ so
that |ϕ| ≤ N and |g − ϕ| < ²2 /16N (Theorem 2.4, Chapter 2). By re-
placing the coefficients and rectangles that appear in the canonical form
of ϕ by complex numbers with rational real and imaginary parts, and
rectangles
R with rational coordinates, we may find a ψ with |ψ| ≤ N and
|g − ψ| < ²2 /8N . Finally, we note that
Z Z
|g − ψ|2 ≤ 2N |g − ψ| < ²2 /4.
Consequently kg − ψk < ²/2, therefore kf − ψk < ², and the proof is

complete.
The example L2 (Rd ) possesses all the characteristic properties of a
Hilbert space, and motivates the definition of the abstract version of this
concept.
2 Hilbert spaces
A set H is a Hilbert space if it satisfies the following:
(i) H is a vector space over C (or R).2
(ii) H is equipped with an inner product (·, ·), so that
• f 7→ (f, g) is linear on H for every fixed g ∈ H,

• (f, g) = (g, f ),
• (f, f ) ≥ 0 for all f ∈ H.
We let kf k = (f, f )1/2 .
(iii) kf k = 0 if and only if f = 0.
1 Bydefinition f ∈ L2 (Rd ) implies that |f |2 is integrable, hence f (x) is finite for a.e x.
2 Atthis stage we consider both cases, where the scalar field can be either C or R.
However, in many applications, such as in the context of Fourier analysis, one deals
primarily with Hilbert spaces over C.
(iv) The Cauchy-Schwarz and triangle inequalities hold
|(f, g)| ≤ kf k kgk and kf + gk ≤ kf k + kgk
for all f, g ∈ H.
(v) H is complete in the metric d(f, g) = kf − gk.
(vi) H is separable.
We make two comments about the definition of a Hilbert space. First,

the Cauchy-Schwarz and triangle inequalities in (iv) are in fact easy
consequences of assumptions (i) and (ii). (See Exercise 1.) Second, we
make the requirement that H be separable because that is the case in
most applications encountered. That is not to say that there are no
interesting non-separable examples; one such example is described in
Problem 2.
Also, we remark that in the context of a Hilbert space we shall of-
ten write limn→∞ fn = f or fn → f to mean that limn→∞ kfn − f k = 0,
which is the same as d(fn , f ) → 0.
We give some examples of Hilbert spaces.
Example 1. If E is a measurable subset of Rd with m(E) > 0, we let
L2 (E) denote the space of square integrable functions that are supported
on E,
½ Z ¾
2 2
L (E) = f supported on E, so that |f (x)| dx < ∞ .
E
The inner product and norm on L2 (E) are then

Z µZ ¶1/2
2
(f, g) = f (x)g(x) dx and kf k = |f (x)| dx .
E E
Once again, we consider two elements of L2 (E) to be equivalent if they

differ only on a set of measure zero; this guarantees that kf k = 0 implies
f = 0. The properties (i) through (vi) follow from these of L2 (Rd ) proved
above.
Example 2. A simple example is the finite-dimensional complex Eu-

clidean space. Indeed,
CN = {(a1 , . . . , aN ) : ak ∈ C}
becomes a Hilbert space when equipped with the inner product

N
X
a k bk ,
k=1
where a = (a1 , . . . , aN ) and b = (b1 , . . . , bN ) are in CN . The norm is then

Ã N
!1/2
X
kak = |ak |2 .
k=1
One can formulate in the same way the real Hilbert space RN .
Example 3. An infinite-dimensional analogue of the above example is

the space `2 (Z). By definition
( ∞
)
X
`2 (Z) = (. . . , a−2 , a−1 , a0 , a1 , . . .) : ai ∈ C, |an |2 < ∞ .
n=−∞
If we denote infinite sequences by a and b, the inner product and norm

on `2 (Z) are
∞
Ã ∞
!1/2
X X
(a, b) = a k bk and kak = |ak |2 .
k=−∞ k=−∞
We leave the proof that `2 (Z) is a Hilbert space as Exercise 4.

While this example is very simple, it will turn out that all infinite-
dimensional (separable) Hilbert spaces are `2 (Z) in disguise.
Also, a slight variant of this space is `2 (N), where we take only one-
sided sequences, that is,
( ∞
)
X
2 2
` (N) = (a1 , a2 , . . .) : ai ∈ C, |an | < ∞ .
n=1
The inner product and norm are then defined in the same way with the
sums extending from n = 1 to ∞.
A characteristic feature of a Hilbert space is the notion of orthogo-
nality. This aspect, with its rich geometric and analytic consequences,
distinguishes Hilbert spaces from other normed vector spaces. We now
describe some of these properties.
2.1 Orthogonality
Two elements f and g in a Hilbert space H with inner product (·, ·) are
orthogonal or perpendicular if
(f, g) = 0, and we then write f ⊥ g.
The first simple observation is that the usual theorem of Pythagoras

holds in the setting of abstract Hilbert spaces:
Proposition 2.1 If f ⊥ g, then kf + gk2 = kf k2 + kgk2 .
Proof. It suffices to note that (f, g) = 0 implies (g, f ) = 0, and there-
fore
kf + gk2 = (f + g, f + g) = kf k2 + (f, g) + (g, f ) + kgk2

= kf k2 + kgk2 .
A finite or countably infinite subset {e1 , e2 , . . .} of a Hilbert space H

is orthonormal if
½
1 when k = `,
(ek , e` ) =
0 when k 6= `.
In other words, each ek has unit norm and is orthogonal to e` whenever

` 6= k.
P
Proposition 2.2 If {ek }∞k=1 is orthonormal, and f = ak ek ∈ H where
the sum is finite, then
X
kf k2 = |ak |2 .
The proof is a simple application of the Pythagorean theorem.

Given an orthonormal subset {e1 , e2 , . . .} = {ek }∞
k=1 of H, a natural
problem is to determine whether this subset spans all of H, that is,
whether finite linear combinations of elements in {e1 , e2 , . . .} are dense
in H. If this is the case, we say that {ek }∞
k=1 is an orthonormal basis
for H. If we are in the presence of an orthonormal basis, we might expect
that any f ∈ H takes the form
∞
X
f= ak ek ,
k=1
for some constants ak ∈ C. In fact, taking the inner product of both

sides with ej , and recalling that {ek } is orthonormal yields (formally)
(f, ej ) = aj .
This question is motivated by Fourier series. In fact, a good insight

into the theorem below is afforded by consideringRπ the case where H
1
is L2 ([−π, π]) with inner product (f, g) = 2π −π
f (x)g(x) dx, and the
orthonormal set {ek }∞ k=1 is merely a relabeling of the exponentials
{einx }∞
n=−∞ . P∞
Adapting the notation used in Fourier series, we write f ∼ k=1 ak ek ,
where aj = (f, ej ) for all j.
In the next theorem, we provide four equivalent characterizations that
{ek } is an orthonormal basis for H.
Theorem 2.3 The following properties of an orthonormal set {ek }∞

k=1
are equivalent.
(i) Finite linear combinations of elements in {ek } are dense in H.

(ii) If f ∈ H and (f, ej ) = 0 for all j, then f = 0.
PN
(iii) If f ∈ H, and SN (f ) = k=1 ak ek , where ak = (f, ek ), then SN (f ) →
f as N → ∞ in the norm.
P∞
(iv) If ak = (f, ek ), then kf k2 = k=1 |ak |2 .
Proof. We prove that each property implies the next, with the last
one implying the first.
We begin by assuming (i). Given f ∈ H with (f, ej ) = 0 for all j, we
wish to prove that f = 0. By assumption, there exists a sequence {gn }
of elements in H that are finite linear combinations of elements in {ek },
and such that kf − gn k tends to 0 as n goes to infinity. Since (f, ej ) = 0
for all j, we must have (f, gn ) = 0 for all n; therefore an application of
the Cauchy-Schwarz inequality gives
kf k2 = (f, f ) = (f, f − gn ) ≤ kf k kf − gn k for all n.
Letting n → ∞ proves that kf k2 = 0; hence f = 0, and (i) implies (ii).

Now suppose that (ii) is verified. For f ∈ H we define
N
X
SN (f ) = ak ek , where ak = (f, ek ),
k=1
and prove first that SN (f ) converges to some element g ∈ H. Indeed,

one notices that the definition of ak implies (f − SN (f )) ⊥ SN (f ), so
the Pythagorean theorem and Proposition 2.2 give
N
X
2 2 2 2
(2) kf k = kf − SN (f )k + kSN (f )k = kf − SN (f )k + |ak |2 .
k=1
PN
Hence kf k2 ≥ k=1 |ak |2 , and letting N tend to infinity we obtain Bessel’s
inequality
∞
X
|ak |2 ≤ kf k2 ,
k=1
P∞
which implies that the series k=1 |ak |2 converges. Therefore, {SN (f )}∞
N =1
forms a Cauchy sequence in H since
N
X
kSN (f ) − SM (f )k2 = |ak |2 whenever N > M .
k=M +1
Since H is complete, there exists g ∈ H such that SN (f ) → g as N tends

to infinity.
Fix j, and note that for all sufficiently large N , (f − SN (f ), ej ) =
aj − aj = 0. Since SN (f ) tends to g, we conclude that
(f − g, ej ) = 0 for all j.
P∞
Hence f = g by assumption (ii), and we have proved that f = k=1 ak ek .
Now assume that (iii) holds. Observe from (2) that we immediately
get in the limit as N goes to infinity
∞
X
2
kf k = |ak |2 .
k=1
Finally, if (iv) holds, then again from (2) we see that kf − SN (f )k

converges to 0. Since each SN (f ) is a finite linear combination of elements
in {ek }, we have completed the circle of implications, and the theorem
is proved.
In particular, a closer look at the proof shows that Bessel’s inequality
holds for any orthonormal family {ek }. In contrast, the identity
∞
X
kf k2 = |ak |2 , where ak = (f, ek ),
k=1
which is called Parseval’s identity, holds if and only if {ek }∞

k=1 is also
an orthonormal basis.
Now we turn our attention to the existence of a basis.
Theorem 2.4 Any Hilbert space has an orthonormal basis.
The first step in the proof of this fact is to recall that (by definition)
a Hilbert space H is separable. Hence, we may choose a countable col-
lection of elements F = {hk } in H so that finite linear combinations of
elements in F are dense in H.
We start by recalling a definition already used in the case of finite-
dimensional vector spaces. Finitely many elements g1 , . . . , gN are said to
be linearly independent if whenever
a 1 g1 + · · · + a N gN = 0 for some complex numbers ai ,
then a1 = a2 = · · · = aN = 0. In other words, no element gj is a lin-

ear combination of the others. In particular, we note that none of the
gj can be 0. We say that a countable family of elements is linearly
independent if all finite subsets of this family are linearly independent.
If we next successively disregard the elements hk that are linearly
dependent on the previous elements h1 , h2 , . . . , hk−1 , then the result-
ing collection h1 = f1 , f2 , . . . , fk , . . . consists of linearly independent ele-
ments, whose finite linear combinations are the same as those given by
h1 , h2 , . . . , hk , . . ., and hence these linear combinations are also dense in
H.
The proof of the theorem now follows from an application of a familiar
construction called the Gram-Schmidt process. Given a finite family
of elements {f1 , . . . , fk } we call the span of this family the set of all
elements which are finite linear combinations of the elements {f1 , . . . , fk }.
We denote the span of {f1 , . . . , fk } by Span({f1 , . . . , fk }).
We now construct a sequence of orthonormal vectors e1 , e2 , . . . such
that Span({e1 , . . . , en }) = Span({f1 , . . . , fn }) for all n ≥ 1. We do this
by induction.
By the linear independence hypothesis, f1 6= 0, so we may take e1 =
f1 /kf1 k. Next, assume that orthonormal vectors e1 , . . . , ek have been
found such that Span({e1 , . . . , ek }) = Span({f1 , . . . , fk }) for a given k.
Pk
We then try e0k+1 as fk+1 + j=1 aj ej . To have (e0k+1 , ej ) = 0 requires
that aj = −(fk+1 , ej ), and this choice of aj for 1 ≤ j ≤ k assures that
e0k+1 is orthogonal to e1 , . . . , ek . Moreover our linear independence hy-
pothesis assures that e0k+1 6= 0; hence we need only “renormalize” and
take ek+1 = e0k+1 /ke0k+1 k to complete the inductive step. With this we
have found an orthonormal basis for H
Note that we have implicitly assumed that the number of linearly in-
dependent elements f1 , f2 , . . . is infinite. In the case where there are only
N linearly independent vectors f1 , . . . , fN , then e1 , . . . , eN constructed
in the same way also provide an orthonormal basis for H. These two
cases are differentiated in the following definition. If H is a Hilbert space
with an orthonormal basis consisting of finitely many elements, then we
say that H is finite-dimensional. Otherwise H is said to be infinite-
dimensional.
2.2 Unitary mappings

A correspondence between two Hilbert spaces that preserves their struc-
ture is a unitary transformation. More precisely, suppose we are given
two Hilbert spaces H and H0 with respective inner products (·, ·)H and
(·, ·)H0 , and the corresponding norms k · kH and k · kH0 . A mapping
U : H → H0 between these space is called unitary if:
(i) U is linear, that is, U (αf + βg) = αU (f ) + βU (g).
(ii) U is a bijection.
(iii) kU f kH0 = kf kH for all f ∈ H.
Some observations are in order. First, since U is bijective it must
have an inverse U −1 : H0 → H that is also unitary. Part (iii) above also
implies that if U is unitary, then
(U f, U g)H0 = (f, g)H for all f, g ∈ H.
To see this, it suffices to “polarize,” that is, to note that for any vector
space (say over C) with inner product (·, ·) and norm k · k, we have
· µ ¶¸
1 2 2 F 2 F 2
(F, G) = kF + Gk − kF − Gk + i k + Gk − k − Gk
4 i i
whenever F and G are elements of the space.

The above leads us to say that the two Hilbert spaces H and H0 are
unitarily equivalent or unitarily isomorphic if there exists a unitary
mapping U : H → H0 . Clearly, unitary isomorphism of Hilbert spaces is
an equivalence relation.
With this definition we are now in a position to give precise meaning
to the statement we made earlier that all infinite-dimensional Hilbert
spaces are the same and in that sense `2 (Z) in disguise.
Corollary 2.5 Any two infinite-dimensional Hilbert spaces are unitarily

equivalent.
Proof. If H and H0 are two infinite-dimensional Hilbert spaces, we

may select for each an orthonormal basis, say
{e1 , e2 , . . .} ⊂ H and {e01 , e02 , . . .} ⊂ H0 .

P∞
Then, consider the mapping defined as follows: if f = k=1 ak ek , then
∞
X
U (f ) = g, where g= ak e0k .
k=1
Clearly, the mapping U is both linear and invertible. Moreover, by Par-

seval’s identity, we must have
∞
X
kU f k2H0 = kgk2H0 = |ak |2 = kf k2H ,
k=1
and the corollary is proved.

Consequently, all infinite-dimensional Hilbert spaces are unitarily equiv-
alent to `2 (N), and thus, by relabeling, to `2 (Z). By similar reasoning
we also have the following:
Corollary 2.6 Any two finite-dimensional Hilbert spaces are unitarily

equivalent if and only if they have the same dimension.
Thus every finite-dimensional Hilbert space over C (or over R) is equiv-

alent with Cd (or Rd ), for some d.
2.3 Pre-Hilbert spaces

Although Hilbert spaces arise naturally, one often starts with a pre-
Hilbert space instead, that is, a space H0 that satisfies all the defining
properties of a Hilbert space except (v); in other words H0 is not assumed
to be complete. A prime example arose implicitly early in the study of
Fourier series with the space H0 = R of Riemann integrable functions
on [−π, π] with the usual inner product; we return to this below. Other
examples appear in the next chapter in the study of the solutions of
partial differential equations.
Fortunately, every pre-Hilbert space H0 can be completed.
Proposition 2.7 Suppose we are given a pre-Hilbert space H0 with in-

ner product (·, ·)0 . Then we can find a Hilbert space H with inner product
(·, ·) such that
(i) H0 ⊂ H.
(ii) (f, g)0 = (f, g) whenever f, g ∈ H0 .
(iii) H0 is dense in H.
A Hilbert space satisfying properties like H in the above proposition is
called a completion of H0 . We shall only sketch the construction of
H, since it follows closely Cantor’s familiar method of obtaining the real
numbers as the completion of the rationals in terms of Cauchy sequences
of rationals.
Indeed, consider the collection of all Cauchy sequences {fn } with fn ∈
H0 , 1 ≤ n < ∞. One defines an equivalence relation in this collection
by saying that {fn } is equivalent to {fn0 } if fn − fn0 converges to 0 as
n → ∞. The collection of equivalence classes is then taken to be H. One
then easily verifies that H inherits the structure of a vector space, with
an inner product (f, g) defined as limn→∞ (fn , gn ), where {fn } and {gn }
are Cauchy sequences in H0 , representing, respectively, the elements f
and g in H. Next, if f ∈ H0 we take the sequence {fn }, with fn = f for
all n, to represent f as an element of H, giving H0 ⊂ H. To see that
H is complete, let {F k }∞ k=1 be a Cauchy sequence in H, with each F
k
represented by {fnk }∞ k
n=1 , fn ∈ H0 . If we define F ∈ H as represented by
n n n
the sequence {fn } with fn = fN (n) , where N (n) is so that |fN (n) − fj | ≤
1/n for j ≥ N (n), then we note that F k → F in H.
One can also observe that the completion H of H0 is unique up to
isomorphism. (See Exercise 14.)
3 Fourier series and Fatou’s theorem

We have already seen an interesting relation between Hilbert spaces and
some elementary facts about Fourier series. Here we want to pursue this
idea and also connect it with complex analysis.
When considering Fourier series, it is natural to begin by turning to
the broader class of all integrable functions on [−π, π]. Indeed, note that
L2 ([−π, π]) ⊂ L1 ([−π, π]), by the Cauchy-Schwarz inequality, since the
interval [−π, π] has finite measure. Thus, if f ∈ L1 ([−π, π]) and n ∈ Z,
we define the nth Fourier coefficient of f by
Z π
1
an = f (x)e−inx dx.
2π −π
3. Fourier series and Fatou’s theorem 171
P∞
The Fourier series of f is then formally n=−∞ an einx , and we write
∞
X
f (x) ∼ an einx
n=−∞
to indicate that the sum on the right is the Fourier series of the func-
tion on the left. The theory developed thus far provides the natural
generalization of some earlier results obtained in Book I.
Theorem 3.1 Suppose f is integrable on [−π, π].
(i) If an = 0 for all n, then f (x) = 0 for a.e. x.
P∞ |n| inx
(ii) n=−∞ an r e tends to f (x) for a.e. x, as r → 1, r < 1.
The second conclusion is the almost everywhereR“Abel summability” to
1 π
f of its Fourier series. Note that since |an | ≤ 2π −π
|f (x)| dx, the series
P
an r|n| einx converges absolutely and uniformly for each r, 0 ≤ r < 1.
Proof. The first conclusion is an immediate consequence of the second.
To prove the latter we recall the identity
∞
X 1 − r2
r|n| einy = Pr (y) =
1 − 2r cos y + r2
n=−∞
for the Poisson kernel; see Book I, Chapter 2. Starting with our given
f ∈ L1 ([−π, π]) we extend it as a function on R by making it periodic of
period 2π.3 We then claim that for every x
X∞ Z π
|n| inx 1
(3) an r e = f (x − y)Pr (y) dy.
2π −π
n=−∞
Indeed, by the dominated convergence theorem the right-hand side equals

X Z π
1
r|n| f (x − y)einy dy.
2π −π
Moreover, for each x and n
Z π Z π+x
f (x − y)einy dy = f (y)ein(x−y) dy
−π −π+x
Z π
inx
=e f (y)e−iny dy = einx 2πan .
−π
3 Note that we may without loss of generality assume that f (π) = f (−π) so as to make
the periodic extension unambiguous.

The first equality follows byRtranslation invariance

R (see Section 3, Chap-
π
ter 2), and the second since −π F (y) dy = I F (y) dy whenever F is peri-
odic of period 2π and I is an interval of length 2π (Exercise 3, Chapter 2).
With these observations, the identity (3) is established. We can now in-
voke the facts about approximations to the identity (Theorem 2.1 and
Example 4, Chapter 3) to conclude that the left-hand side of (3) tends to
f (x) at every point of the Lebesgue set of f , hence almost everywhere.
(To be correct, the hypotheses of the theorem require that f be integrable
on all of R. We can achieve this for our periodic function by setting f
equal to zero outside [−2π, 2π], and then (3) still holds for this modified
f , whenever x ∈ [−π, π].)
We return to the more restrictive setting of L2 . We express the essen-
tial conclusions of Theorem 2.3 in the context
R π of Fourier series. With
1
f ∈ L2 ([−π, π]), we write as before an = 2π −π
f (x)e−inx
dx.
Theorem 3.2 Suppose f ∈ L2 ([−π, π]). Then:
(i) We have Parseval’s relation

∞
X Z π
1
|an |2 = |f (x)|2 dx.
2π −π
n=−∞
(ii) The mapping f 7→ {an } is a unitary correspondence between

L2 ([−π, π]) and `2 (Z).
(iii) The Fourier series of f converges to f in the L2 -norm, that is,
Z π
1
|f (x) − SN (f )(x)|2 dx → 0 as N → ∞,
2π −π
P
where SN (f ) = |n|≤N an einx .
To apply the Rprevious results, we let H = L2 ([−π, π]) with inner prod-
1 π
uct (f, g) = 2π −π
f (x)g(x) dx, and take the orthonormal set {ek }∞ k=1
to be the exponentials {einx }∞ n=−∞ , with k = 1 when n = 0, k = 2n for
n > 0, and k = 2|n| − 1 for n < 0.
By the previous result, assertion (ii) of Theorem 2.3 holds and thus
all the other conclusions hold. We therefore havePParseval’s relation,
and from (iv) we conclude that kf − SN (f )k2 = |n|>N |an |2 → 0 as
N → ∞. Similarly, if {an } ∈ `2 (Z) is given, then kSN (f ) − SM (f )k2 →
0, as N, M → ∞. Hence the completeness of L2 guarantees that there is
an f ∈ L2 such that kf − SN (f )k → 0, and one verifies directly that f
3. Fourier series and Fatou’s theorem 173
has {an } as its Fourier coefficients. Thus we deduce that the mapping
f 7→ {an } is onto and hence unitary. This is a key conclusion that holds
in the setting on L2 and was not valid in an earlier context of Riemann
integrable functions. In fact the space R of such functions on [−π, π] is
not complete in the norm, containing as it does the continuous functions,
but R is itself restricted to bounded functions.
3.1 Fatou’s theorem

Fatou’s theorem is a remarkable result in complex analysis. Its proof
combines elements of Hilbert spaces, Fourier series, and deeper ideas of
differentiation theory, and yet none of these notions appear in its state-
ment. The question that Fatou’s theorem answers may be put simply as
follows.
Suppose F (z) is holomorphic in the unit disc D = {z ∈ C :

|z| < 1}. What are conditions on F that guarantee that F (z)
will converge, in an appropriate sense, to boundary values
F (eiθ ) on the unit circle?
In general a holomorphic function in the unit disc can behave quite

erratically near the boundary. It turns out, however, that imposing a
simple boundedness condition is enough to obtain a strong conclusion.
If F is a function defined in the unit disc D, we say that F has a radial
limit at the point −π ≤ θ ≤ π on the circle, if the limit
lim F (reiθ )
r → 1
r < 1
exists.
Theorem 3.3 A bounded holomorphic function F (reiθ ) on the unit disc

has radial limits at almost every θ.
P∞
Proof. We know that F (z) has a power series expansion n=0 an z n in
D that converges absolutely andP∞uniformly whenever z = reiθ and r < 1.
n inθ
In fact, for r < 1 the series n=0 an r e is the Fourier series of the
function F (reiθ ), that is,
Z π
n 1
an r = F (reiθ )e−inθ dθ when n ≥ 0,
2π −π
and the integral vanishes when n < 0. (See also Chapter 3, Section 7 in
Book II).
We pick M so that |F (z)| ≤ M , for all z ∈ D. By Parseval’s identity

∞
X Z π
2 2n 1
|an | r = |F (reiθ )|2 dθ for each 0 ≤ r < 1.
2π −π
n=0
P
Letting r → 1 one sees that |an |2 converges (and is ≤ M 2 ). We now let
iθ 2
F (e ) be the L -function whose Fourier coefficients are an when n ≥ 0,
and 0 when n < 0. Hence by conclusion (ii) in Theorem 3.1
∞
X
an rn einθ → F (eiθ ), for a.e θ,
n=0
concluding the proof of the theorem.

If we examine the argument given above we see that the same conclu-
sion holds for a larger class of functions. In this connection, we define
the Hardy space H 2 (D) to consist of all holomorphic functions F on
the unit disc D that satisfy
Z π
1
sup |F (reiθ )|2 dθ < ∞.
0≤r<1 2π −π
We also define the “norm” for functions F in this class, kF kH 2 (D) , to be

the square root of the above quantity.
One notes that if F is bounded, then F ∈ H 2 (D), and moreover the
conclusion of the existence of radial limits almost everywhere holds for
any F ∈ H 2 (D), by the same argument given for the bounded P∞ case.
4
Fi-
2 n
nally,
P∞ one notes that F ∈ H P∞(D) if and only if F (z) = n=0 a n z with
2 2 2
n=0 |an | < ∞; moreover, n=0 |an | = kF kH 2 (D) . This states in par-
ticular that H 2 (D) is in fact a Hilbert space that can be viewed as the
“subspace” `2 (Z+ ) of `2 (Z), consisting of all {an } ∈ `2 (Z), with an = 0
when n < 0.
Some general considerations of subspaces and their concomitant or-
thogonal projections will be taken up next.
4 Closed subspaces and orthogonal projections

A linear subspace S (or simply subspace) of H is a subset of H that
satisfies αf + βg ∈ S whenever f, g ∈ S and α, β are scalars. In other
words, S is also a vector space. For example in R3 , lines passing through
4 An even more general statement is given in Problem 5∗ .

4. Closed subspaces and orthogonal pro jections 175
the origin and planes passing through the origin are the one-dimensional
and two-dimensional subspaces, respectively.
The subspace S is closed if whenever {fn } ⊂ S converges to some
f ∈ H, then f also belongs to S. In the case of finite-dimensional Hilbert
spaces, every subspace is closed. This is, however, not true in the gen-
eral case of infinite-dimensional Hilbert spaces. For instance, as we
have already indicated, the subspace of Riemann integrable functions
in L2 ([−π, π]) is not closed, nor is the subspace obtained by fixing a ba-
sis and taking all vectors that are finite linear combinations of these basis
elements. It is useful to note that every closed subspace S of H is itself a
Hilbert space, with the inner product on S that which is inherited from
H. (For the separability of S, see Exercise 11.)
Next, we show that a closed subspace enjoys an important character-
istic property of Euclidean geometry.
Lemma 4.1 Suppose S is a closed subspace of H and f ∈ H. Then:
(i) There exists a (unique) element g0 ∈ S which is closest to f , in the

sense that
kf − g0 k = inf kf − gk.
g∈S
(ii) The element f − g0 is perpendicular to S, that is,
(f − g0 , g) = 0 for all g ∈ S.
The situation in the lemma can be visualized as in Figure 1.
g0
Figure 1. Nearest element to f in S

Proof. If f ∈ S, then we choose f = g0 , and there is nothing left

to prove. Otherwise, we let d = inf g∈S kf − gk, and note that we must
/ S and S is closed. Consider a sequence {gn }∞
have d > 0 since f ∈ n=1 in
S such that
kf − gn k → d as n → ∞.
We claim that {gn } is a Cauchy sequence whose limit will be the desired
element g0 . In fact, it would suffice to show that a subsequence of {gn }
converges, and this is immediate in the finite-dimensional case because
a closed ball is compact. However, in general this compactness fails, as
we shall see in Section 6, and so a more intricate argument is needed at
this point.
To prove our claim, we use the parallelogram law, which states that
in a Hilbert space H
£ ¤
(4) kA + Bk2 + kA − Bk2 = 2 kAk2 + kBk2 for all A, B ∈ H.
The simple verification of this equality, which consists of writing each

norm in terms of the inner product, is left to the reader. Putting A =
f − gn and B = f − gm in the parallelogram law, we find
£ ¤
k2f − (gn + gm )k2 + kgm − gn k2 = 2 kf − gn k2 + kf − gm k2 .
However S is a subspace, so the quantity 12 (gn + gm ) belongs to S, hence
1
k2f − (gn + gm )k = 2kf − (gn + gm )k ≥ 2d.
2
Therefore
£ ¤
kgm − gn k2 = 2 kf − gn k2 + kf − gm k2 − k2f − (gn + gm )k2
£ ¤
≤ 2 kf − gn k2 + kf − gm k2 − 4d2 .
By construction, we know that kf − gn k → d and kf − gm k → d as n, m →

∞, so the above inequality implies that {gn } is a Cauchy sequence. Since
H is complete and S closed, the sequence {gn } must have a limit g0 in
S, and then it satisfies d = kf − g0 k.
We prove that if g ∈ S, then g ⊥ (f − g0 ). For each ² (positive or neg-
ative), consider the perturbation of g0 defined by g0 − ²g. This element
belongs to S, hence
kf − (g0 − ²g)k2 ≥ kf − g0 k2 .
Since kf − (g0 − ²g)k2 = kf − g0 k2 + ²2 kgk2 + 2² Re(f − g0 , g), we find

that
(5) 2² Re(f − g0 , g) + ²2 kgk2 ≥ 0.
If Re(f − g0 , g) < 0, then taking ² small and positive contradicts (5).

If Re(f − g0 , g) > 0, a contradiction also follows by taking ² small and
negative. Thus Re(f − g0 , g) = 0. By considering the perturbation g0 −
i²g, a similar argument gives Im(f − g0 , g) = 0, and hence (f − g0 , g) =
0.
Finally, the uniqueness of g0 follows from the above observation about
orthogonality. Suppose g̃0 is another point in S that minimizes the
distance to f . By taking g = g0 − g˜0 in our last argument we find
(f − g0 ) ⊥ (g0 − g˜0 ), and the Pythagorean theorem gives
kf − g̃0 k2 = kf − g0 k2 + kg0 − g̃0 k2 .
Since by assumption kf − g̃0 k2 = kf − g0 k2 , we conclude that kg0 − g̃0 k =

0, as desired.
Using the lemma, we may now introduce a useful concept that is an-
other expression of the notion of orthogonality. If S is a subspace of a
Hilbert space H, we define the orthogonal complement of S by
S ⊥ = {f ∈ H : (f, g) = 0 for all g ∈ S}.
Clearly, S ⊥ is also a subspace of H, and moreover S ∩ S ⊥ = {0}. To see

this, note that if f ∈ S ∩ S ⊥ , then f must be orthogonal to itself; thus
0 = (f, f ) = kf k, and therefore f = 0. Moreover, S ⊥ is itself a closed
subspace. Indeed, if fn → f , then (fn , g) → (f, g) for every g, by the
Cauchy-Schwarz inequality. Hence if (fn , g) = 0 for all g ∈ S and all n,
then (f, g) = 0 for all those g.
Proposition 4.2 If S is a closed subspace of a Hilbert space H, then
H = S ⊕ S ⊥.
The notation in the proposition means that every f ∈ H can be written

uniquely as f = g + h, where g ∈ S and h ∈ S ⊥ ; we say that H is the
direct sum of S and S ⊥ . This is equivalent to saying that any f in H
is the sum of two elements, one in S, the other in S ⊥ , and that S ∩ S ⊥
contains only 0.
The proof of the proposition relies on the previous lemma giving the
closest element of f in S. In fact, for any f ∈ H, we choose g0 as in the
lemma and write
f = g0 + (f − g0 ).
By construction g0 ∈ S, and the lemma implies f − g0 ∈ S ⊥ , and this

shows that f is the sum of an element in S and one in S ⊥ . To prove that
this decomposition is unique, suppose that
f = g + h = g̃ + h̃ where g, g̃ ∈ S and h, h̃ ∈ S ⊥ .
Then, we must have g − g̃ = h̃ − h. Since the left-hand side belongs to

S while the right-hand side belongs to S ⊥ the fact that S ∩ S ⊥ = {0}
implies g − g̃ = 0 and h̃ − h = 0. Therefore g = g̃ and h = h̃ and the
uniqueness is established.
With the decomposition H = S ⊕ S ⊥ one has the natural projection
onto S defined by
PS (f ) = g, where f = g + h and g ∈ S, h ∈ S ⊥ .
The mapping PS is called the orthogonal projection onto S and sat-

isfies the following simple properties:
(i) f 7→ PS (f ) is linear,
(ii) PS (f ) = f whenever f ∈ S,
(iii) PS (f ) = 0 whenever f ∈ S ⊥ ,
(iv) kPS (f )k ≤ kf k for all f ∈ H.
Property (i) means that PS (αf1 + βf2 ) = αPS (f1 ) + βPS (f2 ), whenever
f1 , f2 ∈ H and α and β are scalars.
It will be useful to observe the following. Suppose {ek } is a (finite
or infinite) collection of orthonormal vectors in H. Then the orthogonal
projection
P P in the closure of the subspace spanned by {ek } is given by
P (f ) = k (f, ek )ek . In case the collection is infinite, the sum converges
in the norm of H.
We illustrate this with two examples that arise in Fourier analysis.
P∞
Example 1. On L2 ([−π, π]), recall that if f (θ) ∼ n=−∞ an einθ then
the partial sums of the Fourier series are
N
X
SN (f )(θ) = an einθ .
n=−N
Therefore, the partial sum operator SN consists of the projection onto

the closed subspace spanned by {e−N , . . . , eN }.
The sum SN can be realized as a convolution
Z π
1
SN (f )(θ) = DN (θ − ϕ)f (ϕ) dϕ,
2π −π
where DN (θ) = sin((N + 1/2)θ)/ sin(θ/2) is the Dirichlet kernel.
Example 2. Once again, consider L2 ([−π, π]) and let S denote the
subspace that consists of all F ∈ L2 ([−π, π]) with
∞
X
F (θ) ∼ an einθ .
n=0
In other words, S is the space of square integrable functions whose

Fourier coefficients an vanish for n < 0. From the proof of Fatou’s theo-
rem, this implies that S can be identified with the Hardy space H 2 (D),
where D is the unit disc, and so is a closed subspace unitarily isomorphic
to `2 (Z+ ). Therefore, using this identification, if P denotes the orthogo-
nal projection from L2 ([−π, π]) to S, we may also write P (f )(z) for the
element corresponding to H 2 (D), that is,
∞
X
P (f )(z) = an z n .
n=0
Given f ∈ L2 ([−π, π]), we define the Cauchy integral of f by
Z
1 f (ζ)
C(f )(z) = dζ,
2πi γ ζ −z
where γ denotes the unit circle and z belongs to the unit disc. Then we
have the identity
P (f )(z) = C(f )(z), for all z ∈ D.
Indeed, since f ∈ L2 it follows by the Cauchy-Schwarz inequality that

f ∈ L1 ([−π, π]), and therefore we may interchange the sum and integral
in the following calculation (recall |z| < 1):

∞
X X∞ µ Z π ¶
n 1 iθ −inθ
P (f )(z) = an z = f (e )e dθ z n
2π −π
n=0 n=0
Z π X∞
1
= f (eiθ ) (e−iθ z)n dθ
2π −π
n=0
Z π iθ
1 f (e )
= dθ
2π −π 1 − e−iθ z
Z π
1 f (eiθ ) iθ
= ie dθ
2πi −π eiθ − z
= C(f )(z).
5 Linear transformations
The focus of analysis in Hilbert spaces is largely the study of their lin-
ear transformations. We have already encountered two classes of such
transformations, the unitary mappings and the orthogonal projections.
There are two other important classes we shall deal with in this chapter
in some detail: the “linear functionals” and the “compact operators,”
and in particular those that are symmetric.
Suppose H1 and H2 are two Hilbert spaces. A mapping T : H1 → H2
is a linear transformation (also called linear operator or operator)
if
T (af + bg) = aT (f ) + bT (g) for all scalars a, b and f, g ∈ H1 .
Clearly, linear operators satisfy T (0) = 0.

We shall say that a linear operator T : H1 → H2 is bounded if there
exists M > 0 so that
(6) kT (f )kH2 ≤ M kf kH1 .
The norm of T is denoted by kT kH1 →H2 or simply kT k and defined by
kT k = inf M,
where the infimum is taken over all M so that (6) holds. A trivial example
is given by the identity operator I, with I(f ) = f . It is of course a
unitary operator and a projection, with kIk = 1.
5. Linear transformations 181
In what follows we shall generally drop the subscripts attached to the

norms of elements of a Hilbert space, when this causes no confusion.
Lemma 5.1 kT k = sup{|(T f, g)| : kf k ≤ 1, kgk ≤ 1}, where of course

f ∈ H1 and g ∈ H2 .
Proof. If kT k ≤ M , the Cauchy-Schwarz inequality gives
|(T f, g)| ≤ M whenever kf k ≤ 1 and kgk ≤ 1;
thus sup{|(T f, g)| : kf k ≤ 1, kgk ≤ 1} ≤ kT k.

Conversely, if sup{|(T f, g)| : kf k ≤ 1, kgk ≤ 1} ≤ M , we claim that
kT f k ≤ M kf k for all f . If f or T f is zero, there is nothing to prove.
Otherwise, f 0 = f /kf k and g 0 = T f /kT f k have norm 1, so by assump-
tion
|(T f 0 , g 0 )| ≤ M.
But since |(T f 0 , g 0 )| = kT f k/kf k this gives kT f k ≤ M kf k, and the

lemma is proved.
A linear transformation T is continuous if T (fn ) → T (f ) whenever
fn → f . Clearly, linearity implies that T is continuous on all of H1 if
and only if it is continuous at the origin. In fact, the conditions of being
bounded or continuous are equivalent.
Proposition 5.2 A linear operator T : H1 → H2 is bounded if and only

if it is continuous.
Proof. If T is bounded, then kT (f ) − T (fn )kH2 ≤ M kf − fn kH1 ,

hence T is continuous. Conversely, suppose that T is continuous but
not bounded. Then for each n there exists fn 6= 0 such that kT (fn )k ≥
nkfn k. The element gn = fn /(nkfn k) has norm 1/n, hence gn → 0.
Since T is continuous at 0, we must have T (gn ) → 0, which contradicts
the fact that kT (gn )k ≥ 1. This proves the proposition.
In the rest of this chapter we shall assume that all linear operators are
bounded, hence continuous. It is noteworthy to recall that any linear
operator between finite-dimensional Hilbert spaces is necessarily contin-
uous.
5.1 Linear functionals and the Riesz representation theorem

A linear functional ` is a linear transformation from a Hilbert space
H to the underlying field of scalars, which we may assume to be the
complex numbers,
` : H → C.
Of course, we view C as a Hilbert space equipped with its standard norm,

the absolute value.
A natural example of a linear functional is provided by the inner prod-
uct on H. Indeed, for fixed g ∈ H, the map
`(f ) = (f, g)
is linear, and also bounded by the Cauchy-Schwarz inequality. Indeed,
|(f, g)| ≤ M kf k, where M = kgk.
Moreover, `(g) = M kgk so we have k`k = kgk. The remarkable fact is

that this example is exhaustive, in the sense that every continuous linear
functional on a Hilbert space arises as an inner product. This is the so-
called Riesz representation theorem.
Theorem 5.3 Let ` be a continuous linear functional on a Hilbert space

H. Then, there exists a unique g ∈ H such that
`(f ) = (f, g) for all f ∈ H.
Moreover, k`k = kgk.
Proof. Consider the subspace of H defined by
S = {f ∈ H : `(f ) = 0}.
Since ` is continuous the subspace S, which is called the null-space of `,

is closed. If S = H, then ` = 0 and we take g = 0. Otherwise S ⊥ is non-
trivial and we may pick any h ∈ S ⊥ with khk = 1. With this choice of h
we determine g by setting g = `(h)h. Thus if we let u = `(f )h − `(h)f ,
then u ∈ S, and therefore (u, h) = 0. Hence
0 = (`(f )h − `(h)f, h) = `(f )(h, h) − (f, `(h)h).
Since (h, h) = 1, we find that `(f ) = (f, g) as desired.

At this stage we record the following remark for later use. Let H0
be a pre-Hilbert space whose completion is H. Suppose `0 is a linear
functional on H0 which is bounded, that is, |`0 (f )| ≤ M kf k for all f ∈
H0 . Then `0 has an extension ` to a bounded linear functional on H,

with |`(f )| ≤ M kf k for all f ∈ H. This extension is also unique. To see
this, one merely notes that {`0 (fn )} is a Cauchy sequence whenever the
vectors {fn } belong to H0 , and fn → f in H, as n → ∞. Thus we may
define `(f ) as limn→∞ `0 (fn ). The verification of the asserted properties
of ` is then immediate. (This result is a special case of the extension
Lemma 1.3 in the next chapter.)
5.2 Adjoints
The first application of the Riesz representation theorem is to determine
the existence of the “adjoint” of a linear transformation.
Proposition 5.4 Let T : H → H be a bounded linear transformation.
There exists a unique bounded linear transformation T ∗ on H so that:
(i) (T f, g) = (f, T ∗ g),
(ii) kT k = kT ∗ k,
(iii) (T ∗ )∗ = T .
The linear operator T ∗ : H → H satisfying the above conditions is called
the adjoint of T .
To prove the existence of an operator satisfying (i) above, we observe
that for each fixed g ∈ H, the linear functional ` = `g , defined by
`(f ) = (T f, g),
is bounded. Indeed, since T is bounded one has kT f k ≤ M kf k; hence

the Cauchy-Schwarz inequality implies that
|`(f )| ≤ kT f k kgk ≤ Bkf k,
where B = M kgk. Consequently, the Riesz representation theorem guar-

antees the existence of a unique h ∈ H, h = hg , such that
`(f ) = (f, h).
Then we define T ∗ g = h, and note that the association T ∗ : g 7→ h is

linear and satisfies (i).
The fact that kT k = kT ∗ k follows at once from (i) and Lemma 5.1:
kT k = sup{|(T f, g)| : kf k ≤ 1, kgk ≤ 1}

= sup{|(f, T ∗ g)| : kf k ≤ 1, kgk ≤ 1}
= kT ∗ k.
To prove (iii), note that (T f, g) = (f, T ∗ g) for all f and g if and only
if (T ∗ f, g) = (f, T g) for all f and g, as one can see by taking complex
conjugates and reversing the roles of f and g.
We record here a few additional remarks.
(a) In the special case when T = T ∗ (we say that T is symmetric), then
(7) kT k = sup{|(T f, f )| : kf k = 1}.
This should be compared to Lemma 5.1, which holds for any linear oper-
ator. To establish (7), let M = sup{|(T f, f )| : kf k = 1}. By Lemma 5.1
it is clear that M ≤ kT k. Conversely, if f and g belong on H, then one
has the following “polarization” identity which is easy to verify
1
(T f, g) = [(T (f + g), f + g) − (T (f − g), f − g)
4
+ i (T (f + ig), f + ig) − i (T (f − ig), f − ig)].
For any h ∈ H, the quantity (T h, h) is real, because T = T ∗ , hence

(T h, h) = (h, T ∗ h) = (h, T h) = (T h, h). Consequently
1
Re(T f, g) = [(T (f + g), f + g) − (T (f − g), f − g)] .
4
£ ¤
Now |(T h, h)| ≤ M khk2 , so |Re(T f, g)| ≤ M 2 2
4 kf + gk + kf − gk , and
an application of the parallelogram law (4) then implies
M
|Re(T f, g)| ≤ [kf k2 + kgk2 ].
2
So if kf k ≤ 1 and kgk ≤ 1, then |Re(T f, g)| ≤ M . In general, we may
replace g by eiθ g in the last inequality to find that whenever kf k ≤ 1 and
kgk ≤ 1, then |(T f, g)| ≤ M , and invoking Lemma 5.1 once again gives
the result, kT k ≤ M .
(b) Let us note that if T and S are bounded linear transformations of H to
itself, then so is their product T S, defined by (T S)(f ) = T (S(f )). More-
over we have automatically (T S)∗ = S ∗ T ∗ ; in fact, (T Sf, g) = (Sf, T ∗ g) =
(f, S ∗ T ∗ g).
(c) One can also exhibit a natural connection between linear transforma-
tions on a Hilbert space and their associated bilinear forms. Suppose first
that T is a bounded operator in H. Define the corresponding bilinear
form B by
(8) B(f, g) = (T f, g).

Note that B is linear in f and conjugate linear in g. Also by the Cauchy-

Schwarz inequality |B(f, g)| ≤ M kf k kgk, where M = kT k. Conversely if
B is linear in f , conjugate linear in g and satisfies |B(f, g)| ≤ M kf k kgk,
there is a unique linear transformation so that (8) holds with M = kT k.
This can be proved by the argument of Proposition 5.4; the details are
left to the reader.
5.3 Examples
Having presented the elementary facts about Hilbert spaces, we now
digress to describe briefly the background of some of the early develop-
ments of the theory. A motivating problem of considerable interest was
that of the study of the “eigenfunction expansion” of a differential oper-
ator L. A particular case, that of a Sturm-Liouville operator, arises on
an interval [a, b] of R with L defined by
d2
L= − q(x),
dx2
where q is a given real-valued function. The question is then that of
expanding an “arbitrary” function in terms of the eigenfunctions ϕ, that
is those functions that satisfy L(ϕ) = µϕ for some µ ∈ R. The classi-
cal example of this is that of Fourier series, where L = d2 /dx2 on the
interval [−π, π] with each exponential einx an eigenfunction of L with
eigenvalue µ = −n2 .
When made precise in the “regular” case, the problem for L can be
resolved by considering an associated “integral operator” T defined on
L2 ([a, b]) by
Z b
T (f )(x) = K(x, y)f (y) dy,
a
with the property that for suitable f ,
LT (f ) = f.
It turns out that a key feature that makes the study of T tractable is
a certain compactness it enjoys. We now pass to the definitions and
elaboration of some of these ideas, and begin by giving two relevant
illustrations of classes of operators on Hilbert spaces.
Infinite diagonal matrix

Suppose {ϕk }∞
k=1 is an orthonormal basis of H. Then, a linear transfor-
mation T : H → H is said to be diagonalized with respect to the basis
{ϕk } if
T (ϕk ) = λk ϕk , where λk ∈ C for all k.
In general, a non-zero element ϕ is called an eigenvector of T with

eigenvalue λ if T ϕ = λϕ. So the ϕk above are eigenvectors of T , and
the numbers λk are the corresponding eigenvalues.
So if
∞
X ∞
X
f∼ ak ϕk then Tf ∼ ak λk ϕk .
k=1 k=1
The sequence {λk } is called the multiplier sequence corresponding to

T.
In this case, one can easily verify the following facts:
• kT k = supk |λk |.
• T ∗ corresponds to the sequence {λk }; hence T = T ∗ if and only if

the λk are real.
• T is unitary if and only if |λk | = 1 for all k.
• T is an orthogonal projection if and only if λk = 0 or 1 for all k.
As a particular example, consider H = L2 ([−π, π]), and assume that

every f ∈ L2 ([−π, π]) is extended to R by periodicity, so that f (x +
2π) = f (x) for all x ∈ R. Let ϕk (x) = eikx for k ∈ Z. For a fixed h ∈ R
the operator Uh defined by
Uh (f )(x) = f (x + h)
is unitary with λk = eikh . Hence

∞
X ∞
X
Uh (f ) ∼ ak λk eikx if f∼ ak eikx .
k=−∞ k=−∞
Integral operators, and in particular, Hilbert-Schmidt

operators
Let H = L2 (Rd ). If we can define an operator T : H → H by the formula
Z
T (f )(x) = K(x, y)f (y) dy whenever f ∈ L2 (Rd ),
Rd
we say that the operator T is an integral operator and K is its asso-

ciated kernel.
In fact, it was the problem of invertibility related to such operators,
and more precisely the question of solvability of the equation f − T f = g
for given g, that initiated the study of Hilbert spaces. These equations
were then called “integral equations.”
In general a bounded linear transformation cannot be expressed as an
(absolutely convergent) integral operator. However, there is an inter-
esting class for which this is possible and which has a number of other
worthwhile properties: Hilbert-Schmidt operators, those with a ker-
nel K that belongs to L2 (Rd × Rd ).
Proposition 5.5 Let T be a Hilbert-Schmidt operator on L2 (Rd ) with
kernel K.
(i) If f ∈ L2 (Rd ), then for almost every x the function y 7→ K(x, y)f (y)
is integrable.
(ii) The operator T is bounded from L2 (Rd ) to itself, and
kT k ≤ kKkL2 (Rd ×Rd ) ,
where kKkL2 (Rd ×Rd ) is the L2 -norm of K on Rd × Rd = R2d .
(iii) The adjoint T ∗ has kernel K(y, x).

Proof. By Fubini’s theorem we know that for almost every x, the
function y 7→ |K(x, y)|2 is integrable. Then, part (i) follows directly from
an application of the Cauchy-Schwarz inequality.
For (ii), we make use again of the Cauchy-Schwarz inequality as follows
¯Z ¯ Z
¯ ¯
¯ K(x, y)f (y) dy ¯ ≤ |K(x, y)||f (y)| dy
¯ ¯
µZ ¶1/2 µZ ¶1/2
2 2
≤ |K(x, y)| dy |f (y)| dy .
Therefore, squaring this and integrating in x yields

Z µZ Z ¶
2 2 2
kT f kL2 (Rd ) ≤ |K(x, y)| dy |f (y)| dy dx
= kKk2L2 (Rd ×Rd ) kf k2L2 (Rd ) .
Finally, part (iii) follows by writing out (T f, g) in terms of a double

integral, and then interchanging the order of integration, as is permissible
by Fubini’s theorem.
Hilbert-Schmidt operators can be defined analogously for the Hilbert

space L2 (E), where E is a measurable subset of Rd . We leave it to the
reader to formulate an prove the analogue of Proposition 5.5 that holds
in this case.
Hilbert-Schmidt operators enjoy another important property: they are
compact. We will now discuss this feature in more detail.
6 Compact operators
We shall use the notion of sequential compactness in a Hilbert space H:
a set X ⊂ H is compact if for every sequence {fn } in X, there exists a
subsequence {fnk } that converges in the norm to an element in X.
Let H denote a Hilbert space, and B the closed unit ball in H,
B = {f ∈ H : kf k ≤ 1}.
A well-known result in elementary real analysis says that in a finite-

dimensional Euclidean space, a closed and bounded set is compact. How-
ever, this does not carry over to the infinite-dimensional case. The fact
is that in this case the unit ball, while closed and bounded, is not com-
pact. To see this, consider the sequence {fn } = {en }, where the en are
orthonormal. By the Pythagorean theorem, ken − em k2 = 2 if n 6= m, so
no subsequence of the {en } can converge.
In the infinite-dimensional case we say that a linear operator T : H →
H is compact if the closure of
T (B) = {g ∈ H : g = T (f ) for some f ∈ B}
is a compact set. Equivalently, an operator T is compact if, whenever

{fk } is a bounded sequence in H, there exists a subsequence {fnk } so
that T fnk converges. Note that a compact operator is automatically
bounded.
Note that by what has been said, a linear transformation is in general
not compact (take for instance the identity operator!). However, if T is
of finite rank, which means that its range is finite-dimensional, then
it is automatically compact. It turns out that dealing with compact
operators provides us with the closest analogy to the usual theorems of
(finite-dimensional) linear algebra. Some relevant analytic properties of
compact operators are given by the proposition below.
Proposition 6.1 Suppose T is a bounded linear operator on H.

6. Compact operators 189
(i) If S is compact on H, then ST and T S are also compact.
(ii) If {Tn } is a family of compact linear operators with kTn − T k → 0

as n tends to infinity, then T is compact.
(iii) Conversely, if T is compact, there is a sequence {Tn } of operators

of finite rank such that kTn − T k → 0.
(iv) T is compact if and only if T ∗ is compact.
Proof. Part (i) is immediate. For part (ii) we use a diagonalization

argument. Suppose {fk } is a bounded sequence in H. Since T1 is com-
pact, we may extract a subsequence {f1,k }∞ k=1 of {fk } such that T1 (f1,k )
converges. From {f1,k } we may find a subsequence {f2,k }∞ k=1 such that
T2 (f2,k ) converges, and so on. If we let gk = fk,k , then we claim {T (gk )}
is a Cauchy sequence. We have
kT (gk ) − T (g` )k ≤ kT (gk ) − Tm (gk )k + kTm (gk ) − Tm (g` )k+

+ kTm (g` ) − T (g` )k.
Since kT − Tm k → 0 and {gk } is bounded, we can make the first and

last term each < ²/3 for some large m independent of k and `. With this
fixed m, we note that by construction kTm (gk ) − Tm (g` )k < ²/3 for all
large k and `. This proves our claim; hence {T (gk )} converges in H.
To prove (iii) let {ek }∞
k=1 be a basis of H and let Qn be the orthogonal
projectionPon the subspace spannedPby the ek with k > n. Then clearly
∞
Qn (g) ∼ k>n ak ek whenever g ∼ k=1 ak ek , and kQn gk2 is a decreas-
ing sequence that tends to 0 as n → ∞ for any g ∈ H. We claim that
kQn T k → 0 as n → ∞. If not, there is a c > 0 so that kQn T k ≥ c, and
hence for each n we can find fn , with kfn k = 1 so that kQn T fn k ≥ c.
Now by compactness of T , choosing an appropriate subsequence {fnk },
we have T fnk → g for some g. But Qnk (g) = Qnk T fnk − Qnk (T fnk − g),
and hence we conclude that kQnk (g)k ≥ c/2, for large k. This contradic-
tion shows that kQn T k → 0. So if Pn is the complementary projection
on the finite-dimensional space spanned by e1 , . . . , en , I = Pn + Qn , then
kQn T k → 0 means that kPn T − T k → 0. Since each Pn T is of finite rank,
assertion (iii) is established.
Finally, if T is compact the fact that kPn T − T k → 0 implies kT ∗ Pn −
T ∗ k → 0, and clearly T ∗ Pn is again of finite rank. Thus we need only
appeal to the second conclusion to prove the last.
We now state two further observations about compact operators.
• If T can be diagonalized with respect to some basis {ϕk } of eigen-

vectors and corresponding eigenvalues {λk }, then T is compact if
and only if |λk | → 0. See Exercise 25.
• Every Hilbert-Schmidt operator is compact.
To prove the second point, recall that a Hilbert-Schmidt operator is

given on L2 (Rd ) by
Z
T (f )(x) = K(x, y)f (y) dy, where K ∈ L2 (Rd × Rd ).
Rd
If {ϕk }∞ 2 d
k=1 denotes an orthonormal basis for L (R ), then the collection
{ϕk (x)ϕ` (y)}k,`≥1 is an orthonormal basis for L2 (Rd × Rd ); the proof of
this simple fact is outlined in Exercise 7. As a result
∞
X P
K(x, y) ∼ ak` ϕk (x)ϕ` (y), with k,` |ak` |2 < ∞.
k,`=1
We define an operator
Z
Pn
Tn f (x) = Kn (x, y)f (y)dy, where Kn (x, y) = k,`=1 ak` ϕk (x)ϕ` (y).
Rd
Then, each Tn has finite-dimensional range, hence is compact. Moreover,

X
kK − Kn k2L2 (Rd ×Rd ) = |ak` |2 → 0 as n → ∞.
k ≥ n or ` ≥ n
By Proposition 5.5, kT − Tn k ≤ kK − Kn kL2 (Rd ×Rd ) , so we can conclude

the proof that T is compact by appealing to Proposition 6.1.
The climax of our efforts regarding compact operators is the infinite-
dimensional version of the familiar diagonalization theorem in linear al-
gebra for symmetric matrices. Using a similar terminology, we say that
a bounded linear operator T is symmetric if T ∗ = T . (These operators
are also called “self-adjoint” or “Hermitian.”)
Theorem 6.2 (Spectral theorem) Suppose T is a compact symmet-

ric operator on a Hilbert space H. Then there exists an (orthonormal)
basis {ϕk }∞
k=1 of H that consists of eigenvectors of T . Moreover, if
T ϕk = λk ϕk ,
then λk ∈ R and λk → 0 as k → ∞.
6. Compact operators 191
Conversely, every operator of the above form is compact and symmetric.

The collection {λk } is called the spectrum of T .
Lemma 6.3 Suppose T is a bounded symmetric linear operator on a

Hilbert space H.
(i) If λ is an eigenvalue of T , then λ is real.
(ii) If f1 and f2 are eigenvectors corresponding to two distinct eigen-

values, then f1 and f2 are orthogonal.
Proof. To prove (i), we first choose a non-zero eigenvector f such

that T (f ) = λf . Since T is symmetric (that is, T = T ∗ ), we find that
λ(f, f ) = (T f, f ) = (f, T f ) = (f, λf ) = λ(f, f ),
where we have used in the last equality the fact that the inner product is
conjugate linear in the second variable. Since f 6= 0, we must have λ = λ
and hence λ ∈ R.
For (ii), suppose f1 and f2 have eigenvalues λ1 and λ2 , respectively.
By the previous argument both λ1 and λ2 are real, and we note that
λ1 (f1 , f2 ) = (λ1 f1 , f2 )
= (T f1 , f2 )
= (f1 , T f2 )
= (f1 , λ2 f2 )
= λ2 (f1 , f2 ).
Since by assumption λ1 6= λ2 we must have (f1 , f2 ) = 0 as desired.

For the next lemma note that every non-zero element of the null-space
of T − λI is an eigenvector with eigenvalue λ.
Lemma 6.4 Suppose T is compact, and λ 6= 0. Then the dimension of

the null space of T − λI is finite. Moreover, the eigenvalues of T form
at most a denumerable set λ1 , . . . , λk , . . ., with λk → 0 as k → ∞. More
specifically, for each µ > 0, the linear space spanned by the eigenvectors
corresponding to the eigenvalues λk with |λk | > µ is finite-dimensional.
Proof. Let Vλ denote the null-space of T − λI, that is, the eigenspace
of T corresponding to λ. If Vλ is not finite-dimensional, there exists
a countable sequence of orthonormal vectors {ϕk } in Vλ . Since T is
compact, there exists a subsequence {ϕnk } such that T (ϕnk ) converges.
But since T (ϕnk ) = λϕnk and λ 6= 0, we conclude that ϕnk converges,

which is a contradiction since kϕnk − ϕnk0 k2 = 2 if k 6= k 0 .
The rest of the lemma follows if we can show that for each µ > 0, there
are only finitely many eigenvalues whose absolute values are greater than
µ. We argue again by contradiction. Suppose there are infinitely many
distinct eigenvalues whose absolute values are greater than µ, and let
{ϕk } be a corresponding sequence of eigenvectors. Since the eigenvalues
are distinct, we know from the previous lemma that {ϕk } is orthogonal,
and after normalization, we may assume that this set of eigenvectors is
orthonormal. One again, since T is compact, we may find a subsequence
so that T (ϕnk ) converges, and since
T (ϕnk ) = λnk ϕnk
the fact that |λnk | > µ leads to a contradiction, since {ϕk } is an or-
thonormal set and thus kλnk ϕnk − λnj ϕnj k2 = λ2nk + λ2nj ≥ 2µ2 .
Lemma 6.5 Suppose T 6= 0 is compact and symmetric. Then either kT k

or −kT k is an eigenvalue of T .
Proof. By the observation (7) made earlier, either
kT k = sup{(T f, f ) : kf k = 1} or − kT k = inf{(T f, f ) : kf k = 1}.
We assume the first case, that is,
λ = kT k = sup{(T f, f ) : kf k = 1},
and prove that λ is an eigenvalue of T . (The proof of the other case is

similar.)
We pick a sequence {fn } ⊂ H such that kfn k = 1 and (T fn , fn ) → λ.
Since T is compact, we may assume also (by passing to a subsequence of
{fn } if necessary) that {T fn } converges to a limit g ∈ H. We claim that
g is an eigenvector of T with eigenvalue λ. To see this, we first observe
that T fn − λfn → 0 because
kT fn − λfn k2 = kT fn k2 − 2λ(T fn , fn ) + λ2 kfn k2

≤ kT k2 kfn k2 − 2λ(T fn , fn ) + λ2 kfn k2
≤ 2λ2 − 2λ(T fn , fn ) → 0.
Since T fn → g, we must have λfn → g, and since T is continuous, this

implies that λT fn → T g. This proves that λg = T g. Finally, we must
7. Exercises 193
have g 6= 0, for otherwise kTn fn k → 0, hence (T fn , fn ) → 0, and λ =

kT k = 0, which is a contradiction.
We are now equipped with the necessary tools to prove the spectral
theorem. Let S denote the closure of the linear space spanned by all
eigenvectors of T . By Lemma 6.5, the space S is non-empty. The goal
is to prove that S = H. If not, then since
(9) S ⊕ S ⊥ = H,
S ⊥ would be non-empty. We will have reached a contradiction once

we show that S ⊥ contains an eigenvector of T . First, we note that T
respects the decomposition (9). In other words, if f ∈ S then T f ∈ S,
which follows from the definitions. Also, if g ∈ S ⊥ then T g ∈ S ⊥ . This
is because T is symmetric and maps S to itself, and hence
(T g, f ) = (g, T f ) = 0 whenever g ∈ S ⊥ and f ∈ S.
Now consider the operator T1 , which by definition is the restriction of

T to the subspace S ⊥ . The closed subspace S ⊥ inherits its Hilbert space
structure from H. We see immediately that T1 is also a compact and
symmetric operator on this Hilbert space. Moreover, if S ⊥ is non-empty,
the lemma implies that T1 has a non-zero eigenvector in S ⊥ . This eigen-
vector is clearly also an eigenvector of T , and therefore a contradiction
is obtained. This concludes the proof of the spectral theorem.
Some comments about Theorem 6.2 are in order. If in its statement we
drop either of the two assumptions (the compactness or symmetry of T ),
then T may have no eigenvectors. (See Exercises 32 and 33.) However,
when T is a general bounded linear transformation which is symmetric,
there is an appropriate extension of the spectral theorem that holds for
it. Its formulation and proof require further ideas that are deferred to
Chapter 6.
7 Exercises
1. Show that properties (i) and (ii) in the definition of a Hilbert space (Section 2)
imply property (iii): the Cauchy-Schwarz inequality |(f, g)| ≤ kf k · kgk and the
triangle inequality kf + gk ≤ kf k + kgk.
[Hint: For the first inequality, consider (f + λg, f + λg) as a positive quadratic
function of λ. For the second, write kf + gk2 as (f + g, f + g).]
2. In the case of equality in the Cauchy-Schwarz inequality we have the following.

If |(f, g)| = kf k kgk and g 6= 0, then f = cg for some scalar c.
[Hint: Assume kf k = kgk = 1 and (f, g) = 1. Then f − g and g are orthogonal,

while f = f − g + g. Thus kf k2 = kf − gk2 + kgk2 .]
3. Note that kf + gk2 = kf k2 + kgk2 + 2Re(f, g) for any pair of elements in a

Hilbert space H. As a result, verify the identity kf + gk2 + kf − gk2 = 2(kf k2 +
kgk2 ).
4. Prove from the definition that `2 (Z) is complete and separable.
5. Establish the following relations between L2 (Rd ) and L1 (Rd ):

(a) Neither the inclusion L2 (Rd ) ⊂ L1 (Rd ) nor the inclusion L1 (Rd ) ⊂ L2 (Rd )
is valid.
(b) Note, however, that if f is supported on a set E of finite measure and if f ∈
L2 (Rd ), applying the Cauchy-Schwarz inequality to f χE gives f ∈ L1 (Rd ),
and
kf kL1 (Rd ) ≤ m(E)1/2 kf kL2 (Rd ) .
(c) If f is bounded (|f (x)| ≤ M ), and f ∈ L1 (Rd ), then f ∈ L2 (Rd ) with

1/2
kf kL2 (Rd ) ≤ M 1/2 kf kL1 (Rd ) .
[Hint: For (a) consider f (x) = |x|−α , when |x| ≤ 1 or when |x| > 1.]
6. Prove that the following are dense subspaces of L2 (Rd ).

(a) The simple functions.
(b) The continuous functions of compact support.
7. Suppose {ϕk }∞ 2 d
k=1 is an orthonormal basis for L (R ). Prove that the collection
{ϕk,j }1≤k,j<∞ with ϕk,j (x, y) = ϕk (x)ϕj (y) is an orthonormal basis of L2 (Rd ×
Rd ).
[Hint: First verify that the R{ϕk,j } are orthonormal, by Fubini’s theorem. Next,
for each j consider Fj (x) = Rd F (x, y)ϕj (y) dy. If one assumes that (F, ϕk,j ) = 0
R
for all j, then Fj (x)ϕk (x) dx = 0.]
8. Let η(t) be a fixed strictly positive continuous function on [a, b]. Define Hη =
L2 ([a, b], η) to be the space of all measurable functions f on [a, b] such that
Z b
|f (t)|2 η(t) dt < ∞.
a
Define the inner product on Hη by

Z b
(f, g)η = f (t)g(t)η(t) dt.
a
7. Exercises 195
(a) Show that Hη is a Hilbert space, and that the mapping U : f 7→ η 1/2 f gives
a unitary correspondence between Hη and the usual space L2 ([a, b]).
(b) Generalize this to the case when η is not necessarily continuous.
9. Let H1 = L2 ([−π, π]) be the Hilbert

Rπ space of functions F (eiθ ) on the unit circle
with inner product (F, G) = 2π −π F (eiθ )G(eiθ ) dθ. Let H2 be the space L2 (R).
1
Using the mapping

i−x
x 7→
i+x
of R to the unit circle, show that:
(a) The correspondence U : F → f , with

„ «
1 i−x
f (x) = F
π 1/2 (i + x) i+x
gives a unitary mapping of H1 to H2 .
(b) As a result,
 „ «n ff∞
1 i−x 1
π 1/2 i+x i+x n=−∞
is an orthonormal basis of L2 (R).
10. Let S denote a subspace of a Hilbert space H. Prove that (S ⊥ )⊥ is the

smallest closed subspace of H that contains S.
11. Let P be the orthogonal projection associated with a closed subspace S in a

Hilbert space H, that is,
P (f ) = f if f ∈ S and P (f ) = 0 if f ∈ S ⊥ .
(a) Show that P 2 = P and P ∗ = P .
(b) Conversely, if P is any bounded operator satisfying P 2 = P and P ∗ = P ,

prove that P is the orthogonal projection for some closed subspace of H.
(c) Using P , prove that if S is a closed subspace of a separable Hilbert space,

then S is also a separable Hilbert space.
12. Let E be a measurable subset of Rd , and suppose S is the subspace of L2 (Rd )

of functions that vanish for a.e. x ∈
/ E. Show that the orthogonal projection P on
S is given by P (f ) = χE · f , where χE is the characteristic function of E.
13. Suppose P1 and P2 are a pair of orthogonal projections on S1 and S2 , respec-

tively. Then P1 P2 is an orthogonal projection if and only if P1 and P2 commute,
that is, P1 P2 = P2 P1 . In this case, P1 P2 projects onto S1 ∩ S2 .
14. Suppose H and H0 are two completions of a pre-Hilbert space H0 . Show that
there is a unitary mapping from H to H0 that is the identity on H0 .
[Hint: If f ∈ H, pick a Cauchy sequence {fn } in H0 that converges to f in H. This
sequence will also converge to an element f 0 in H0 . The mapping f 7→ f 0 gives the
required unitary mapping.]
15. Let T be any linear transformation from H1 to H2 . If we suppose that H1 is

finite-dimensional, then T is automatically bounded. (If H1 is not assumed to be
finite-dimensional this may fail; see Problem 1 below.)
16. Let F0 (z) = 1/(1 − z)i .
(a) Verify that |F0 (z)| ≤ eπ/2 in the unit disc, but that limr→1 F0 (r) does not
exist.
[Hint: Note that |F0 (r)| = 1 and F0 (r) oscillates between ±1 infinitely often
as r → 1.]
(b) Let {αn }∞

n=1 be an enumeration of the rationals, and let
∞
X
F (z) = δ j F0 (ze−iαj ),
j=1
where δ is sufficiently small. Show that limr→1 F (reiθ ) fails to exist when-
ever θ = αj , and hence F fails to have a radial limit for a dense set of points
on the unit circle.
17. Fatou’s theorem can be generalized by allowing a point to approach the

boundary in larger regions, as follows.
For each 0 < s < 1 and point z on the unit circle, consider the region Γs (z)
defined as the smallest closed convex set that contains z and the closed disc Ds (0).
In other words, Γs (z) consists of all lines joining z with points in Ds (0). Near the
point z, the region Γs (z) looks like a triangle. See Figure 2.
We say that a function F defined in the open unit disc has a non-tangential
limit at a point z on the circle, if for every 0 < s < 1, the limit
lim F (w)
w → z
w ∈ Γs (z)
exists.
Prove that if F is holomorphic and bounded on the open unit disc, then F has
a non-tangential limit for almost every point on the unit circle.
[Hint: Show that the Poisson integral of a function f has non-tangential limits at
every point of the Lebesgue set of f .]
7. Exercises 197
Γs (z)
Figure 2. The region Γs (z)
18. Let H denote a Hilbert space, and L(H) the vector space of all bounded linear
operators on H. Given T ∈ L(H), we define the operator norm
kT k = inf{B : kT vk ≤ Bkvk, for all v ∈ H}.
(a) Show that kT1 + T2 k ≤ kT1 k + kT2 k whenever T1 , T2 ∈ L(H).
(b) Prove that
d(T1 , T2 ) = kT1 − T2 k
defines a metric on L(H).
(c) Show that L(H) is complete in the metric d.
19. If T is a bounded linear operator on a Hilbert space, prove that
kT T ∗ k = kT ∗ T k = kT k2 = kT ∗ k2 .
20. Suppose H is an infinite-dimensional Hilbert space. We have seen an example

of a sequence {fn } in H with kfn k = 1 for all n, but for which no subsequence
of {fn } converges in H. However, show that for any sequence {fn } in H with
kfn k = 1 for all n, there exist f ∈ H and a subsequence {fnk } such that for all
g ∈ H, one has
lim (fnk , g) = (f, g).

k→∞
One says that {fnk } converges weakly to f .

[Hint: Let g run through a basis for H, and use a diagonalization argument. One
can then define f by giving its series expansion with respect to the chosen basis.]
21. There are several senses in which a sequence of bounded operators {Tn } can
converge to a bounded operator T (in a Hilbert space H). First, there is con-
vergence in the norm, that is, kTn − T k → 0, as n → ∞. Next, there is a weaker
convergence, which happens to be called strong convergence, that requires that
Tn f → T f , as n → ∞, for every vector f ∈ H. Finally, there is weak conver-
gence (see also Exercise 20) that requires (Tn f, g) → (T f, g) for every pair of
vectors f, g ∈ H.
(a) Show by examples that weak convergence does not imply strong convergence,
nor does strong convergence imply convergence in the norm.
(b) Show that for any bounded operator T there is a sequence {Tn } of bounded
operators of finite rank so that Tn → T strongly as n → ∞.
22. An operator T is an isometry if kT f k = kf k for all f ∈ H.

(a) Show that if T is an isometry, then (T f, T g) = (f, g) for every f, g ∈ H.
Prove as a result that T ∗ T = I.
(b) If T is an isometry and T is surjective, then T is unitary and T T ∗ = I.
(c) Give an example of an isometry that is not unitary.
(d) Show that if T ∗ T is unitary then T is an isometry.

[Hint: Use the fact that (T f, T f ) = (f, f ) for f replaced by f ± g and f ± ig.]
23. Suppose {Tk } is a collection of bounded operators on a Hilbert space H, with

kTk k ≤ 1 for all k. Suppose also that
Tk Tj∗ = Tk∗ Tj = 0 for all k 6= j.

P
Let SN = N k=−N Tk .
Show that SN (f ) converges as N → ∞, for every f ∈ H. If T (f ) denotes the
limit, prove that kT k ≤ 1.
A generalization is given in Problem 8∗ below.
[Hint: Consider first the case when only finitely many of the Tk are non-zero, and
note that the ranges of the Tk are mutually orthogonal.]
24. Let {ek }∞

k=1 denote an orthonormal set Pin a Hilbert space H. If {ck }∞
k=1 is a
sequence of positive real numbers such that c2k < ∞, then the set
∞
X
A={ ak ek : |ak | ≤ ck }
k=1
is compact in H.
25. Suppose T is a bounded operator that is diagonal with respect to a basis {ϕk },
with T ϕk = λk ϕk . Then T is compact if and only if λk → 0.
7. Exercises 199
[Hint: If λk → 0, then note that kPn T − T k → 0, where Pn is the orthogonal

projection on the subspace spanned by ϕ1 , ϕ2 , . . . , ϕn . ]
26. Suppose w is a measurable function on Rd with 0 < w(x) < ∞ for a.e. x, and
K is a measurable function on R2d that satisfies:
Z
(i) |K(x, y)|w(y) dy ≤ Aw(x) for almost every x ∈ Rd , and
Rd
Z
(ii) |K(x, y)|w(x) dx ≤ Aw(y) for almost every y ∈ Rd .
Rd
Prove that the integral operator defined by

Z
T f (x) = K(x, y)f (y) dy, x ∈ Rd
Rd
is bounded on L2 (Rd ) with kT kR≤ A. R

Note as a special case that if |K(x, y)| dy ≤ A for all x, and |K(x, y)| dx ≤ A
for all y, then kT k ≤ A.
[Hint: Show that if f ∈ L2 (Rd ), then
Z »Z –1/2
1/2 1/2 2 −1
|K(x, y)| |f (y)| dy ≤ A w(x) |K(x, y)| |f (y)| w(y) dy .]
27. Prove that the operator

Z ∞
1 f (y)
T f (x) = dy
π 0 x+y
2
is bounded on L (0, ∞) with norm kT k ≤ 1.
[Hint: Use Exercise 26 with an appropriate w.]
28. Suppose H = L2 (B), where B is the unit ball in Rd . Let K(x, y) be a mea-
surable function on B × B that satisfies |K(x, y)| ≤ A|x − y|−d+α for some α > 0,
whenever x, y ∈ B. Define
Z
T f (x) = K(x, y)f (y)dy.
B
(a) Prove that T is a bounded operator on H.

(b) Prove that T is compact.
(c) Note that T is a Hilbert-Schmidt operator if and only if α > d/2.
[Hint: For (b), consider the operators Tn associated with the truncated kernels
Kn (x, y) = K(x, y) if |x − y| ≥ 1/n and 0 otherwise. Show that each Tn is compact,
and that kTn − T k → 0 as n → ∞.]
29. Let T be a compact operator on a Hilbert space H, and assume λ 6= 0.

(a) Show that the range of λI − T defined by
{g ∈ H : g = (λI − T )f, for some f ∈ H}
is closed. [Hint: Suppose gj → g, where gj = (λI − T )fj . Let Vλ denote

the eigenspace of T corresponding to λ, that is, the kernel of λI − T . Why
can one assume that fj ∈ Vλ⊥ ? Under this assumption prove that {fj } is a
bounded sequence.]
(b) Show by example that this may fail when λ = 0.
(c) Show that the range of λI − T is all of H if and only if the null-space of
λI − T ∗ is trivial.
30. Let H = L2 ([−π, π]) with [−π, π] identified as the unit circle. Fix a bounded
sequence {λn }∞
n=−∞ of complex numbers, and define an operator T f by
∞
X ∞
X
T f (x) ∼ λn an einx whenever f (x) ∼ an einx .
n=−∞ n=−∞
Such an operator is called a Fourier multiplier operator, and the sequence

{λn } is called the multiplier sequence.
(a) Show that T is a bounded operator on H and kT k = supn |λn |.
(b) Verify that T commutes with translations, that is, if we define τh (x) =
f (x − h) then
T ◦ τh = τh ◦ T for every h ∈ R.
(c) Conversely, prove that if T is any bounded operator on H that commutes

with translations, then T is a Fourier multiplier operator. [Hint: Consider
T (einx ).]
31. Consider a version of the sawtooth function defined on [−π, π) by5
K(x) = i(sgn(x)π − x),
and extended to R with period 2π. Suppose f ∈ L1 ([−π, π]) is extended to R with
period 2π, and define
Z π
1
T f (x) = K(x − y)f (y) dy
2π −π
Z π
1
= K(y)f (x − y) dy.
2π −π
5 The symbol sgn(x) denotes the sign function: it equals 1 or −1 if x is positive or
negative respectively, and 0 if x = 0.

7. Exercises 201
Rπ
(a) Show that F (x) = T f (x) is absolutely continuous, and if −π
f (y)dy = 0,
then F 0 (x) = if (x) a.e. x.
(b) Show that the mapping f 7→ T f is compact and symmetric on L2 ([−π, π]).
(c) Prove that ϕ(x) ∈ L2 ([−π, π]) is an eigenfunction for T if and only if ϕ(x)
is (up to a constant multiple) equal to einx for some integer n 6= 0 with
eigenvalue 1/n, or ϕ(x) = 1 with eigenvalue 0.
(d) Show as a result that {einx }n∈Z is an orthonormal basis of L2 ([−π, π]).
Note that in Book I, Chapter 2, Exercise 8, it is shown that the Fourier series
of K is
X einx
K(x) ∼ .
n
n6=0
32. Consider the operator T : L2 ([0, 1]) → L2 ([0, 1]) defined by
T (f )(t) = tf (t).
(a) Prove that T is a bounded linear operator with T = T ∗ , but that T is not
compact.
(b) However, show that T has no eigenvectors.
33. Let H be a Hilbert space with basis {ϕk }∞

k=1 . Verify that the operator T
defined by
1
T (ϕk ) = ϕk+1
k
is compact, but has no eigenvectors.
34. Let K be a Hilbert-Schmidt kernel which is real and symmetric. Then, as we

saw, the operator T whose kernel is K is compact and symmetric. Let {ϕk (x)} be
the eigenvectors (with eigenvalues λk ) that diagonalize T . Then:
P
(a) |λk |2 < ∞.
k
P
(b) K(x, y) ∼ λk ϕk (x)ϕk (y) is the expansion of K in the basis {ϕk (x)ϕk (y)}.
(c) Suppose T is a compact operator

P which is symmetric. Then T is of Hilbert-
Schmidt type if and only if n |λn |2 < ∞, where {λn } are the eigenvalues
of T counted according to their multiplicities.
35. Let H be a Hilbert space. Prove the following variants of the spectral theorem.
(a) If T1 and T2 are two linear symmetric and compact operators on H that
commute (that is, T1 T2 = T2 T1 ), show that they can be diagonalized simul-
taneously. In other words, there exists an orthonormal basis for H which
consists of eigenvectors for both T1 and T2 .
(b) A linear operator on H is normal if T T ∗ = T ∗ T . Prove that if T is normal

and compact, then T can be diagonalized.
[Hint: Write T = T1 + iT2 where T1 and T2 are symmetric, compact and
commute.]
(c) If U is unitary, and U = λI − T , where T is compact, then U can be diago-

nalized.
8 Problems
1. Let H be an infinite-dimensional Hilbert space. There exists a linear functional

` defined on H that is not bounded (and hence not continuous).
[Hint: Using the axiom of choice (or one of its equivalent forms), construct an
algebraic basis of H, {eα }; it has the property that every element of H is uniquely
a finite linear combination of the {eα }. Select a denumerable collection {en }∞n=1 ,
and define ` to satisfy the requirement that `(en ) = nken k for all n ∈ N.]
2.∗ The following is an example of a non-separable Hilbert space. We consider

the collection of exponentials {eiλx } on R, where λ ranges over the real numbers.
Let H0 denote the space of finite linear combinations of these exponentials. For
f, g ∈ H0 , we define the inner product as
Z T
1
(f, g) = lim f (x)g(x) dx.
T →∞ 2T −T
(a) Show that this limit exists, and
N
X
(f, g) = aλk bλk
k=1
PN PN
if f (x) = k=1 aλk eiλk x and g(x) = k=1 bλk e
iλk x
.
(b) With this inner product H0 is a pre-Hilbert space. Notice that kf k ≤

supx |f (x)|, if f ∈ H0 , where kf k denotes the norm hf, f i1/2 . Let H be
0
the completion of H0 . Then H is not separable because eiλx and eiλ x are
orthonormal if λ 6= λ0 .
A continuous function F defined on R is called almost periodic if it is the
uniform limit (on R) of elements in H0 . Such functions can be identified
with (certain) elements in the completion H: We have H0 ⊂ AP ⊂ H, where
AP denotes the almost periodic functions.
8. Problems 203
(c) A continuous function F is in AP if for every ² > 0 we can find a length

L = L² such that any interval I ⊂ R of length L contains an “almost period”
τ satisfying
sup |F (x + τ ) − F (x)| < ².

x
(d) An equivalent characterization is that F is in AP if and only if every se-

quence F (x + hn ) of translates of F contains a subsequence that converges
uniformly.
3. The following is a direct generalization of Fatou’s theorem: if u(reiθ ) is harmonic

in the unit disc and bounded there, then limr→1 u(reiθ ) exists for a.e. θ.
R 2π 2
[Hint: Let an (r) = 2π1
0
u(reiθ )e−inθ dθ. Then a00n (r) + r1 a0n (r) − nr2 an (r) = 0,
P
hence an (r) = An rn + Bn r−n , n 6= 0, and as a result6 u(reiθ ) = ∞ −∞ an r
|n| inθ
e .
From this one can proceed as in the proof of Theorem 3.3.]
4.∗ This problem provides some examples of functions that fail to have radial limits
almost everywhere.
P 2n
(a) At almost every point of the boundary unit circle, the function ∞ n=0 z
fails to have a radial limit.
P 2n P
(b) More generally, suppose F (z) = ∞ n=0 an z . Then, if |an |2 = ∞ the
functionP F fails to have radial limits at almost every boundary point. How-
ever, if |an |2 < ∞, then F ∈ H 2 (D), and we know by the proof of Theo-
rem 3.3 that F does have radial limits almost everywhere.
5.∗ Suppose F is holomorphic in the unit disc, and

Z π
1
sup log+ |F (reiθ )| dθ < ∞,
0≤r<1 2π −π
where log+ u = log u if u ≥ 1, and log+ u = 0 if u < 1.

Then limr→1 F (reiθ ) exists for almost every θ.
The above condition is satisfied whenever (say)
Z π
1
sup |F (reiθ )|p dθ < ∞, for some p > 0,
0≤r<1 2π −π
(since epu ≥ pu, u ≥ 0).

Functions that satisfy the latter condition are said to belong to the Hardy
space H p (D).
6.∗ If T is compact, and λ 6= 0, show that
6 See also Section 5, Chapter 2 in Book I.

(a) λI − T is injective if and only if λI − T ∗ is injective.
(b) λI − T is injective if and only if λI − T is surjective.

This result, known as the Fredholm alternative, is often combined with that in
Exercise 29.
7. Show that the identity operator on L2 (Rd ) cannot be given as an (absolutely)

convergent integral operator. More precisely, if K(x, y) is a measurable function
d d 2 d
Ron R × R with the property that for each f ∈ L (R ), the integral T (f )(x) =
Rd
K(x, y)f (y) dy converges for almost every x, then T (f ) 6= f for some f .
[Hint: Prove that otherwise for any pair of disjoint balls B1 and B2 in Rd , we
would have that K(x, y) = 0 for a.e. (x, y) ∈ B1 × B2 .]
8.∗ Suppose {Tk } is a collection of bounded opeartors on a Hilbert space H. As-

sume that
kTk Tj∗ k ≤ ak−j and kTk∗ Tj k ≤ a∗k−j ,

P∞
for positive constants {an } with the property that −∞ an = A < ∞. Then
P
SN (f ) converges as N → ∞, for every f ∈ H, with SN = N −N T k . Moreover,
T = limN →∞ SN satisfies kT k ≤ A.
9. A discussion of a class of regular Sturm-Liouville operators follows. Other

special examples are given in the problems below.
Suppose [a, b] is a bounded interval, and L is defined on functions f that are
twice continuously differentiable in [a, b] (we write, f ∈ C 2 ([a, b])) by
d2 f
L(f )(x) = − q(x)f (x).
dx2
Here the function q is continuous and real-valued on [a, b], and we assume for
simplicity that q is non-negative. We say that ϕ ∈ C 2 ([a, b]) is an eigenfunction
of L with eigenvalue µ if L(ϕ) = µϕ, under the assumption that ϕ satisfies the
boundary conditions ϕ(a) = ϕ(b) = 0. Then one can show:
(a) The eigenvalues µ are strictly negative, and the eigenspace corresponding
to each eigenvalue is one-dimensional.
(b) Eigenvectors corresponding to distinct eigenvalues are orthogonal in L2 ([a, b]).
(c) Let K(x, y) be the “Green’s kernel” defined as follows. Choose ϕ− (x) to be
a solution of L(ϕ− ) = 0, with ϕ− (a) = 0 but ϕ0− (a) 6= 0. Similarly, choose
ϕ+ (x) to be a solution of L(ϕ+ ) = 0 with ϕ+ (b) = 0, but ϕ0+ (b) 6= 0. Let
w = ϕ0+ (x)ϕ− (x) − ϕ0− (x)ϕ+ (x), be the “Wronskian” of these solutions, and
note that w is a non-zero constant.
Set
(
ϕ− (x)ϕ+ (y)
if a ≤ x ≤ y ≤ b,
K(x, y) = w
ϕ+ (x)ϕ− (y)
w
if a ≤ y ≤ x ≤ b.
8. Problems 205
Then the operator T defined by

Z b
T (f )(x) = K(x, y)f (y) dy
a
is a Hilbert-Schmidt operator, and hence compact. It is also symmetric.

Moreover, whenever f is continuous on [a, b], T f is of class C 2 ([a, b]) and
L(T f ) = f.
(d) As a result, each eigenvector of T (with eigenvalue λ) is an eigenvector of L

(with eigenvalue µ = 1/λ). Hence Theorem 6.2 proves the completeness of
the orthonormal set arising from normalizing the eigenvectors of L.
10.∗ Let L be defined on C 2 ([−1, 1]) by
d2 f df
L(f )(x) = (1 − x2 ) − 2x .
dx2 dx
If ϕn is the nth Legendre polynomial, given by

„ «n
d
ϕn (x) = (1 − x2 )n , n = 0, 1, 2, . . . ,
dx
then Lϕn = −n(n + 1)ϕn .

When normalized the ϕn form an orthonormal basis of L2 ([−1, 1]) (see also
Problem 2, Chapter 3 in Book I, where ϕn is denoted by Ln .)
11.∗ The Hermite functions hk (x) are defined by the generating identity
∞
X tk 2 2
hk (x) = e−(x /2−2tx+t ) .
k!
k=0
` d
´
(a) They satisfy the “creation” and “annihilation” identities x − dx hk (x) =
` d
´
hk+1 (x) and x + dx hk (x) = hk−1 (x) for k ≥ 0 where h−1 (x) = 0. Note
2 2
that h0 (x) = e−x /2 , h1 (x) = 2xe−x /2 , and more generally hk (x) =
2
Pk (x)e−x /2 , where Pk is a polynomial of degree k.
(b) Using (a) one sees that the hk are eigenvectors of the operator L = −d2 /dx2 +
x2 , with L(hk ) = λk hk , where λk = 2k + 1. One observes that these func-
tions are mutually orthogonal. Since
Z
[hk (x)]2 dx = π 1/2 2k k! = ck ,
R
we can normalize them obtaining a orthonormal sequence {Hk }, with Hk =

−1/2 R
ck hk . This sequence is complete in L2 (Rd ) since R f Hk dx = 0 for all k
R∞ x2
implies −∞ f (x)e− 2 +2tx dx = 0 for all t ∈ C.
P Hk (x)Hk (y)
(c) Suppose that K(x, y) = ∞ k=0 , and also F (x) = T (f )(x) =
R λk
R
K(x,
P y)f (y) dy. Then T is a symmetric
P∞ ak Hilbert-Schmidt operator, and
if f ∼ ∞ k=0 a k H k , then F ∼ k=0 λk H k .
One can show on the basis of (a) and (b) that whenever f ∈ L2 (R), not only is
F ∈ L2 (R), but also x2 F (x) ∈ L2 (R). Moreover, F can be corrected on a set of
measure zero, so it is continuously differentiable, F 0 is absolutely continuous, and
F 00 ∈ L2 (R). Finally, the operator T is the inverse of L in the sense that
LT (f ) = LF = −F 00 + x2 F = f for every f ∈ L2 (R).
(See also Problem 7* in Chapter 5 of Book I.)

5 Hilbert
Examples
Spaces: Several
What is the difference between a mathematician and

a physicist? It is this: To a mathematician all Hilbert
spaces are the same; for a physicist, however, it is their
different realizations that really matter.
Attributed to E. Wigner, ca. 1960
Hilbert spaces arise in a large number of different contexts in analysis.

Although it is a truism that all (infinite-dimensional) Hilbert spaces are
the same, it is in fact their varied and distinct realizations and separate
applications that make them of such interest in mathematics. We shall
illustrate this via several examples.
To begin with, we consider the Plancherel formula and the resulting
unitary character of the Fourier transform. The relevance of these ideas
to complex analysis is then highlighted by the study of holomorphic func-
tions in a half-space that belong to the Hardy space H 2 . That function
space itself is another interesting realization of a Hilbert space. The con-
siderations here are analogous to the ideas that led us to Fatou’s theorem
for the unit disc, but are of a more involved character.
We next see how complex analysis and the Fourier transform com-
bine to guarantee the existence of solutions to linear partial differential
equations with constant coefficients. The proof relies on a basic L2 es-
timate, which once established can be exploited by simple Hilbert space
techniques.
Our final example is Dirichlet’s principle and its applications to the
boundary value problem for harmonic functions. Here the Hilbert space
that arises is given by Dirichlet’s integral, and the solution is expressed
by aid of an appropriate orthogonal projection operator.
1 The Fourier transform on L2

The Fourier transform of a function f on Rd is defined by
Z
(1) fˆ(ξ) = f (x) e−2πix·ξ dx,
Rd
208 Chapter 5. HILBERT SPACES: SEVERAL EXAMPLES
and its attached inversion is given by

Z
(2) f (x) = fˆ(ξ) e2πix·ξ dξ.
Rd
These formulas have already appeared in several different contexts.

We considered first (in Book I) the properties of the Fourier transform
in the elementary setting by restricting to functions in the Schwartz class
S(Rd ). The class S consists of functions f that are smooth (indefinitely
differentiable) and such that for each multi-index α and β, the function
∂ β
xα ( ∂x ) f is bounded on Rd .1 We saw that on this class the Fourier trans-
form is a bijection, that the inversion formula (2) holds, and moreover
we have the Plancherel identity
Z Z
(3) ˆ 2
|f (ξ)| dξ = |f (x)|2 dx.
Rd Rd
Turning now to more general (in particular, non-continuous) functions,

we note that the largest class for which the integral defining fˆ(ξ) con-
verges (absolutely) is the space L1 (Rd ). For it, we saw in Chapter 2 that
a (relative) inversion formula is valid.
Beyond these particular facts, what we would like here is to reestablish
in the general context the symmetry between f and fˆ that holds for S.
This is where the special role of the Hilbert space L2 (Rd ) enters.
We shall define the Fourier transform on L2 (Rd ) as an extension of its
definition on S. For this purpose, we temporarily adopt the notational
device of denoting by F0 and F the Fourier transform on S and its
extension to L2 , respectively.
The main results we prove are the following.
Theorem 1.1 The Fourier transform F0 , initially defined on S(Rd ),

has a (unique) extension F to a unitary mapping of L2 (Rd ) to itself. In
particular,
kF(f )kL2 (Rd ) = kf kL2 (Rd )
for all f ∈ L2 (Rd ).
The extension F will be given by a limiting process: if {fn } is a sequence

in the Schwartz space that converges to f in L2 (Rd ), then {F0 (fn )} will
α
1 Recall that xα = xα 1 α2
1 x2 · · · xd
d ∂ β
and ( ∂x ∂ β1
) = ( ∂x ∂
) · · · ( ∂x )βd , where α =
1 d
(α1 , . . . , αd ) and β = (β1 , . . . , βd ), with αj and βj positive integers. The order of α is
denoted by |α| and defined to be α1 + · · · + αd .
1. The Fourier transform on L2 209
converge to an element in L2 (Rd ) which we will define as the Fourier

transform of f . To implement this approach we have to see that every L2
function can be approximated by elements in the Schwartz space.
Lemma 1.2 The space S(Rd ) is dense in L2 (Rd ). In other words, given
any f ∈ L2 (Rd ), there exists a sequence {fn } ⊂ S(Rd ) such that
kf − fn kL2 (Rd ) → 0 as n → ∞.
For the proof of the lemma, we fix f ∈ L2 (Rd ) and ² > 0. Then, for
each M > 0, we define
½
f (x) if |x| ≤ M and |f (x)| ≤ M ,
gM (x) =
0 otherwise.
Then, |f (x) − gM (x)| ≤ 2|f (x)|, hence |f (x) − gM (x)|2 ≤ 4|f (x)|2 , and
since gM (x) → f (x) as M → ∞ for almost every x, the dominated con-
vergence theorem guarantees that for some M , we have
kf − gM kL2 (Rd ) < ².
We write g = gM , note that this function is bounded and supported on

a bounded set, and observe that it now suffices to approximate g by
functions in the Schwartz space. To achieve this goal, we use a method
called regularization, which consists of “smoothing” g by convolving it
with an approximation of the identity. Consider a function ϕ(x) on Rd
with the following properties:
(a) ϕ is smooth (indefinitely differentiable).
(b) ϕ is supported in the unit ball.
(c) ϕ ≥ 0.
Z
(d) ϕ(x) dx = 1.
Rd
For instance, one can take

( 1
− 1−|x|
ϕ(x) = ce 2
if |x| < 1,
0 if |x| ≥ 1,
where the constant c is chosen so that (d) holds.

Next, we consider the approximation to the identity defined by
Kδ (x) = δ −d ϕ(x/δ).
The key observation is that g ∗ Kδ belongs to S(Rd ), with this convolu-

tion in fact bounded and supported on a fixed bounded set, uniformly in
δ (assuming for example that δ ≤ 1). Indeed, we may write
Z Z
(g ∗ Kδ )(x) = g(y)Kδ (x − y) dy = g(x − y)Kδ (y) dy,
in view of the identity (6) in Chapter 2. We note that since g is supported

on some bounded set and Kδ vanishes outside the ball of radius δ, the
function g ∗ Kδ is supported in some fixed bounded set independent of δ.
Also, the function g is bounded by construction, hence
Z
|(g ∗ Kδ )(x)| ≤ |g(x − y)|Kδ (y) dy
Z
≤ sup |g(z)| Kδ (y) dy = sup |g(z)|,
z∈Rd z∈Rd
which shows that g ∗ Kδ is also uniformly bounded in δ. Moreover, from

the first integral expression for g ∗ Kδ above, one may differentiate under
the integral sign to see that g ∗ Kδ is smooth and all of its derivatives
have support in some fixed bounded set.
The proof of the lemma will be complete if we can show that g ∗ Kδ
converges to g in L2 (Rd ). Now Theorem 2.1 in Chapter 3 guarantees
that for almost every x, the quantity |(g ∗ Kδ )(x) − g(x)|2 converges to 0
as δ tends to 0. An application of the bounded convergence theorem
(Theorem 1.4 in Chapter 2) yields
k(g ∗ Kδ ) − gk2L2 (Rd ) → 0 as δ → 0.
In particular, k(g ∗ Kδ ) − gkL2 (Rd ) < ² for an appropriate δ and hence

kf − g ∗ Kδ kL2 (Rd ) < 2², and choosing a sequence of ² tending to zero
gives the construction of the desired sequence {fn }.
For later purposes it is useful to observe that the proof of the above
lemma establishes the following assertion: if f belongs to both L1 (Rd )
and L2 (Rd ), then there is a sequence {fn }, fn ∈ S(Rd ), that converges
to f in both the L1 -norm and the L2 -norm.
Our definition of the Fourier transform on L2 (Rd ) combines the above

density of S with a general “extension principle.”
Lemma 1.3 Let H1 and H2 denote Hilbert spaces with norms k · k1 and
k · k2 , respectively. Suppose S is a dense subspace of H1 and T0 : S → H2
a linear transformation that satisfies kT0 (f )k2 ≤ ckf k1 whenever f ∈ S.
1. The Fourier transform on L2 211
Then T0 extends to a (unique) linear transformation T : H1 → H2 that

satisfies kT (f )k2 ≤ ckf k1 for all f ∈ H1 .
Proof. Given f ∈ H1 , let {fn } be a sequence in S that converges to f ,

and define
T (f ) = lim T0 (fn ),
n→∞
where the limit is taken in H2 . To see that T is well-defined we must

verify that the limit exists, and that it is independent of the sequence
{fn } used to approximate f . Indeed, for the first point, we note that
{T (fn )} is a Cauchy sequence in H2 because by construction {fn } is
Cauchy in H1 , and the inequality verified by T0 yields
kT0 (fn ) − T0 (fm )k2 ≤ ckfn − fm k1 → 0 as n, m → ∞;
thus {T0 (fn )} is Cauchy, hence converges in H2 .

Second, to justify that the limit is independent of the approximating
sequence, let {gn } be another sequence in S that converges to f in H1 .
Then
kT0 (fn ) − T0 (gn )k2 ≤ ckfn − gn k1 ,
and since kfn − gn k1 ≤ kfn − f k1 + kf − gn k1 , we conclude that {T0 (gn )}

converges to a limit in H2 that equals the limit of {T0 (fn )}.
Finally, we recall that if fn → f and T0 (fn ) → T (f ), then kfn k1 →
kf k1 and kT0 (fn )k2 → kT (f )k2 , so in the limit as n → ∞, the inequality
kT (f )k2 ≤ ckf k1 holds for all f ∈ H1 .
In the present case of the Fourier transform, we apply this lemma with
H1 = H2 = L2 (Rd ) (equipped with the L2 -norm), S = S(Rd ), and T0 =
F0 the Fourier transform defined on the Schwartz space. The Fourier
transform on L2 (Rd ) is by definition the unique (bounded) extension of
F0 to L2 guaranteed by Lemma 1.3. Thus if f ∈ L2 (Rd ) and {fn } is any
sequence in S(Rd ) that converges to f (that is, kf − fn kL2 (Rd ) → 0 as
n → ∞), we define the Fourier transform of f by
(4) F(f ) = lim F0 (fn ),

n→∞
where the limit is taken in the L2 sense. Clearly, the argument in the
proof of the lemma shows that in our special case the extension F con-
tinues to satisfy the identity (3):
kF(f )kL2 (Rd ) = kf kL2 (Rd ) whenever f ∈ L2 (Rd ).

The fact that F is invertible on L2 (and thus F is a unitary mapping)

is also a consequence of the analogous property on S(Rd ). Recall that
on the Schwartz space, F0−1 is given by formula (2), that is,
Z
−1
F0 (g)(x) = g(ξ)e2πix·ξ dξ,
Rd
and satisfies again the identity kF0−1 (g)kL2 = kgkL2 . Therefore, arguing
in the same fashion as above, we can extend F0−1 to L2 (Rd ) by a limiting
argument. Then, given f ∈ L2 (Rd ), we choose a sequence {fn } in the
Schwartz space so that kf − fn kL2 → 0. We have
fn = F0−1 F0 (fn ) = F0 F0−1 (fn ),
and taking the limit as n tends to infinity, we see that
f = F −1 F(f ) = FF −1 (f ),
and hence F is invertible. This concludes the proof of Theorem 1.1.

Some remarks are in order.
(i) Suppose f belongs to both L1 (Rd ) and L2 (Rd ). Are the two definitions
of the Fourier transform the same? That is, do we have F(f ) = fˆ, with
F(f ) defined by the limiting process in Theorem 1.1 and fˆ defined by the
convergent integral (1)? To prove that this is indeed the case we recall
that we can approximate f by a sequence {fn } in S so that fn → f both
in the L1 -norm and the L2 -norm. Since F0 (fn ) = fˆn , a passage to the
limit gives the desired conclusion. In fact, F0 (fn ) converges to F(f ) in
the L2 -norm, so a subsequence converges to F(f ) almost everywhere; see
the analogous statement for L1 in Corollary 2.3, Chapter 2. Moreover,
sup |fˆn (ξ) − fˆ(ξ)| ≤ kfn − f kL1 (Rd ) ,

ξ∈Rd
hence fˆn converges to fˆ everywhere, and the assertion is established.

(ii) The theorem gives a rather abstract definition of the Fourier trans-
form on L2 . In view of what we have just said, we can also define the
Fourier transform more concretely as follows. If f ∈ L2 (Rd ), then
Z
ˆ
f (ξ) = lim f (x)e−2πix·ξ dx,
R→∞ |x|≤R
where the limit is taken in the L2 -norm. Note in fact that if χR denotes
the characteristic function of the ball {x ∈ Rd : |x| ≤ R}, then for each
R the function f χR is in both L1 and L2 , and f χR → f in the L2 -norm.
2. The Hardy space of the upper half-plane 213
(iii) The identity of the various definitions of the Fourier transform dis-
cussed above allows us to choose fˆ as the preferred notation for the
Fourier transform. We adopt this practice in what follows.
2 The Hardy space of the upper half-plane

We will apply the L2 theory of the Fourier transform to holomorphic
functions in the upper half-plane. This leads us to consider the relevant
analogues of the Hardy space and Fatou’s theorem discussed in the previ-
ous chapter.2 It incidentally provides an answer to the following natural
question: What are the functions f ∈ L2 (R) whose Fourier transforms
are supported on the half-line (0, ∞)?
Let R2+ = {z = x + iy, x ∈ R, y > 0} be the upper half-plane. We
define the Hardy space H 2 (R2+ ) to consist of all functions F analytic
in R2+ with the property that
Z
(5) sup |F (x + iy)|2 dx < ∞.
y>0 R
We define the corresponding norm, kF kH 2 (R2+ ) , to be the square root of

the quantity (5).
Let us first describe a (typical) example of a function F in H 2 (R2+ ).
We start with a function F̂0 that belongs to L2 (0, ∞), and write
Z ∞
(6) F (x + iy) = F̂0 (ξ)e2πiξz dξ, z = x + iy, y > 0.
0
(The choice of the particular notation F̂0 will become clearer below.)
We claim that for any δ > 0 the integral (6) converges absolutely and
uniformly as long as y ≥ δ. Indeed, |F̂0 (ξ)e2πiξz | = |F̂0 (ξ)|e−2πξy , hence
by the Cauchy-Schwarz inequality
Z ∞ µZ ∞ ¶1/2 µZ ∞ ¶1/2
2πiξz 2 −4πξδ
|F̂0 (ξ)e | dξ ≤ |F̂0 (ξ)| dξ e dξ ,
0 0 0
from which the asserted convergence is established. From the uniform

convergence it follows that F (z) is holomorphic in the upper half-plane.
Moreover, by Plancherel’s theorem
Z Z ∞
2
|F (x + iy)| dx = |F̂0 (ξ)|2 e−4πξy dξ ≤ kF̂0 k2L2 (0,∞) ,
R 0
2 Further motivation and some elementary background material may be found in The-
orem 3.5 in Chapter 4 of Book II.

and in fact, by the monotone convergence theorem,

Z
sup |F (x + iy)|2 dx = kF̂0 k2L2 (0,∞) .
y>0 R
In particular, F belongs to H 2 (R2+ ). The main result we prove next is

the converse, that is, every element of the space H 2 (R2+ ) is in fact of the
form (6).
Theorem 2.1 The elements F in H 2 (R2+ ) are exactly the functions

given by (6), with F̂0 ∈ L2 (0, ∞). Moreover
kF kH 2 (R2+ ) = kF̂0 kL2 (0,∞) .
This shows incidentally that H 2 (R2+ ) is a Hilbert space that is isomorphic

to L2 (0, ∞) via the correspondence (6).
The crucial point in the proof of the theorem is the following fact. For
any fixed strictly positive y, we let F̂y (ξ) denote the Fourier transform
of the L2 function F (x + iy), x ∈ R. Then for any pair of choices of y,
y1 and y2 , we have that
(7) F̂y1 (ξ)e2πy1 ξ = F̂y2 (ξ)e2πy2 ξ for a.e. ξ.
To establish this assertion we rely on a useful technical observation.
Lemma 2.2 If F belongs to H 2 (R2+ ), then F is bounded in any proper

half-plane {z = x + iy, y ≥ δ}, where δ > 0.
To prove this we exploit the mean-value property of holomorphic func-

tions. This property may be stated in two alternative ways. First, in
terms of averages over circles,
Z 2π
1
(8) F (ζ) = F (ζ + reiθ ) dθ if 0 < r ≤ δ.
2π 0
(Note that if ζ lies in the upper half-plane, Im(ζ) > δ, then the disc
centered at ζ of radius r belongs to R2+ .) Alternatively, integrating over
r, we have the mean-value property in terms of discs,
Z
1
(9) F (ζ) = 2 F (ζ + z) dx dy, z = x + iy.
πδ |z|<δ
These assertions actually hold for harmonic functions in R2 (see Corol-

lary 7.2, Chapter 3 in Book II for the result about holomorphic functions,
and Lemma 2.8, Chapter 5 in Book I for the case of harmonic functions);
later in this chapter we in fact prove the extension of (9) to Rd .
From (9) we see from the Cauchy-Schwarz inequality that
Z
2 1
|F (ζ)| ≤ 2 |F (ζ + z)|2 dx dy.
πδ |z|<δ
Writing z = x + iy and ζ = ξ + iη, with η > δ, we see that the disc

Bδ (ζ) of center ζ and radius δ is contained in the strip {z + ζ : z =
x + iy, −δ < y < δ}, and moreover this strip lies in the half-plane R2+ .
See Figure 1.
Bδ (ζ)
Figure 1. Disc contained in a strip
This gives the following majorization:

Z Z Z
|F (ζ + z)|2 dx dy ≤ |F (ζ + x + iy)|2 dx dy
|z|<δ |y|<δ R
Z
≤ 2δ sup |F (x + i(η + y))|2 dx.
−δ<y<δ R
Recalling that η > δ, we see that the last expression is in fact majorized
by
Z
2δ sup |F (x + iy)|2 dx = 2δ kF k2H 2 (R2 ) .
+
y>0 R
2
In all |F (ζ)|2 ≤ 2
πδ kF kH 2 in the half-plane Im(ζ) > 0, which proves the
lemma.
We now turn to the proof of the identity (7). Starting with F in
H 2 (R2+ ), we improve it by replacing it with the function F ² defined by
1
F ² (z) = F (z) , with ² > 0.
(1 − i²z)2
Observe that |F ² (z)| ≤ |F (z)| when Im(z) > 0; also F ² (z) → F (z) for
each such z, as ² → 0. This shows that for each y > 0, F ² (x + iy) →
F (x + iy) in the L2 -norm. Moreover, the lemma guarantees that each

F ² satisfies the decay estimate
µ ¶
² 1
F (z) = O whenever Im(z) > δ, for some δ > 0.
1 + x2
We assert first that (7) holds with F replaced by F ² . This is a simple

consequence of contour integration applied to the function
G(z) = F ² (z)e−2πizξ .
In fact we integrate G(z) over the rectangle with vertices −R + iy1 , R +

iy1 , R + iy2 , −R + iy2 , and let R → ∞. If we take into account that
G(z) = O(1/(1 + x2 )) in this rectangle, then we find that
Z Z
G(z) dz = G(z) dz,
L1 L2
where Lj is the line {x + iyj : x ∈ R}, j = 1, 2. Since

Z Z
G(z) dz = F ² (x + iyj )e−2πi(x+iyj )ξ dx,
Lj R
This means that
F̂y²1 (ξ)e2πy1 ξ = F̂y²2 (ξ)e2πy2 ξ .
Since F ² (x + iyj ) → F (x + iyj ) in the L2 -norm as ² → 0, we then ob-

tain (7).
The identity we have just proved states that Fˆy (ξ)e2πyξ is independent
of y, y > 0, and thus there is a function F̂0 (ξ) so that Fˆy (ξ)e2πξy = F̂0 (ξ);
as a result
Fˆy (ξ) = F̂0 (ξ)e−2πξy for all y > 0.
Therefore by Plancherel’s identity

Z Z
2
|F (x + iy)| dx = |F̂0 (ξ)|2 e−4πξy dξ,
R R
and hence
Z
sup |F̂0 (ξ)|2 e−4πξy dξ = kF k2H 2 (R2 ) < ∞.
+
y>0 R
Finally this in turn implies that F̂0 (ξ) = 0 for almost every ξ ∈ (−∞, 0).
For if this were not the case, then for appropriate positive numbers a, b,
and c we could have that |F̂0R(ξ)| ≥ a for ξ in a set E in (−∞, −b), with
m(E) ≥ c. This would give |F̂0 (ξ)|2 e−4πξy dξ ≥ a2 ce4πby , which grows
indefinitely as y → ∞. The contradiction thus obtained shows that F̂0 (ξ)
vanishes almost everywhere when ξ ∈ (−∞, 0).
To summarize, for each y > 0 the function F̂y (ξ) equals F̂0 (ξ)e−2πξy ,
with F̂0 ∈ L2 (0, ∞). The Fourier inversion formula then yields the repre-
sentation (6) for an arbitrary element of H 2 , and the proof of the theorem
is concluded.
The second result we deal with may be viewed as the half-plane ana-
logue of Fatou’s theorem in the previous chapter.
Theorem 2.3 Suppose F belongs to H 2 (R2+ ). Then limy→0 F (x + iy) =
F0 (x) exists in the following two senses:
(i) As a limit in the L2 (R)-norm.
(ii) As a limit for almost every x.
Thus F has boundary values (denoted by F0 ) in either of the two senses
above. The function F0 is sometimes referred to as the boundary-value
function of f . The proof of (i) is immediate from what we already know.
Indeed, if F0 is the L2 function whose Fourier transform is F̂0 , then
Z ∞
kF (x + iy) − F0 (x)k2L2 (R) = |F̂0 (ξ)|2 |e−2πξy − 1|2 dy,
0
and this tends to zero as y → 0 by the dominated convergence theorem.

To prove the almost everywhere convergence, we establish the Poisson
integral representation
Z Z
(10) fˆ(ξ)e−2π|ξ|y e2πixξ dξ = f (x − t)Py (t) dt,
R R
with
1 y
Py (x) =
π y 2 + x2
the Poisson kernel.3 This identity holds for every (x, y) ∈ R2+ and any
function f in L2 (R). To see this, we begin by noting the following ele-
mentary integration formulas:
Z ∞
i
(11) e2πiξz dξ = if Im(z) > 0,
0 2πz
3 This is the analogue in R of the identity (3) for the circle, given in Chapter 4.
and
Z
1 y
(12) e−2π|ξ|y e2πiξx dξ = if y > 0.
R π y + x2
2
The first is an immediate consequence of the fact that

Z N
1
e2πiξz dξ = [e2πiN z − 1]
0 2πiz
if we let N → ∞. To prove the second formula, we write the integral as

Z ∞ Z ∞
e−2πξy e2πiξx dξ + e−2πξy e−2πiξx dξ,
0 0
which equals
· ¸
i 1 1 1 y
+ =
2π x + iy −x + iy π y 2 + x2
by (11).
Next we establish (10) when f belongs to (say) the space S. Indeed, for
fixed (x, y) ∈ R2+ consider the function Φ(t, ξ) = f (t)e−2πiξt e−2π|ξ|y e2πiξx
on R2 = {(ξ, t)}. Since |Φ(t, ξ)| = |f (t)|e−2π|ξ|y , then (because f is rapidly
decreasing) Φ is integrable over R2 . Applying Fubini’s theorem yields
Z µZ ¶ Z µZ ¶
Φ(t, ξ) dξ dt = Φ(t, ξ) dt dξ.
R R R R
R
The right-hand side Robviously gives R fˆ(ξ)e−2π|ξ|y e2πixξ dξ, while the
left-hand side yields R f (t)Py (x − y) dt in view of (12) above. However,
if we use the relation (6) in Chapter 2 we see that
Z Z
f (t)Py (x − y) dt = f (x − t)Py (t) dt.
R R
Thus the Poisson integral representation (10) holds for every f ∈ S. For
a general f ∈ L2 (R) we consider a sequence {fn } of elements in S, so
that fn → f (and also fˆn → fˆ) in the L2 -norm. A passage to the limit
then yields the formula for f from the corresponding formula for each
fn . Indeed, by the Cauchy-Schwarz inequality we have
¯Z ¯ µZ ¶1/2
¯ ¯
¯ [fˆ(ξ) − fˆn (ξ)]e−2π|ξ|y
e2πixξ
dξ ¯ ≤ kfˆ − fˆn kL2 e−4π|ξ|y
dξ ,
¯ ¯
R R
and also
¯Z ¯ µZ ¶1/2
¯ ¯
¯ [f (x − t) − fn (x − t)]Py (t) dt¯ ≤ kf − fn kL2 2
|Py (t)| dt ,
¯ ¯
R R
and the right-hand sides tend to 0 because for each fixed (x, y) ∈ R2+ the
functions e−2π|ξ|y , ξ ∈ R, and Py (t), t ∈ R, belong to L2 (R).
Having established the Poisson integral representation (10), we return
to our given element F ∈ H 2 (R2+ ). We know that there is an L2 function
F̂0 (ξ) (which vanishes when ξ < 0) such that (6) holds. With F0 the
L2 (R) function whose Fourier transform is F̂0 (ξ), we see from (10), with
f = F0 , that
Z
F (x + iy) = F0 (x − t)Py (t) dt.
R
From this we deduce the fact that F (x + iy) → F0 (x) a.e in x as y → 0,

since the family {Py } is an approximation of the identity for which The-
orem 2.1 in Chapter 3 applies. There is, however, one small obstacle that
has to be overcome: the theorem as stated applied to L1 functions and
not to functions in L2 . Nevertheless, given the nature of the approxima-
tion to the identity, a simple “localization” argument will succeed. We
proceed as follows.
It will suffice to see that for any large N , which is fixed, F (x + iy) →
F0 (x), for a.e x with |x| < N . To do this, decompose F0 as G + H, where
G(x) = F0 (x) when |x| > 2N , G(x) = 0 when |x| ≥ 2N ; thus H(x) = 0
if |x| ≤ 2N but |H(x)| ≤ |F0 (x)|. Note that now G ∈ L1 and
Z Z Z
F0 (x − t)Py (t) dt = G(x − t)Py (t) dt + H(x − t)Py (t) dt.
R R R
Therefore, by the above mentioned theorem in Chapter 3, the first in-

tegral on the right-hand side converges for a.e x to G(x) = F0 (x) when
|x| < N . While when |x| < N the integrand of the second integral van-
ishes when |t| < N (since then |x − t| < 2N ). That integral is therefore
majorized by
µZ ¶1/2 µZ ¶1/2
2 2
|H(x − t)| dt |Py (t)| dt .
R |t|≥N
¡R ¢1/2
However
R R
|H(x − t)|2 dt ≤ kF0 kL2 , while (as is easily seen)
2
|t|≥N
|Py (t)| dt → 0 as y → 0. Hence F (x + iy) → F0 (x) for a.e x with
|x| < N , as y → 0, and since N is arbitrary, the proof of Theorem 2.3 is

now complete.
The following comments may help clarify the thrust of the above the-
orems.
(i) Let S be the subspace of L2 (R) consisting of all functions F0 arising in
Theorem 2.3. Then, since the functions F0 are exactly those functions in
L2 whose Fourier transform is supported on the half-line (0, ∞), we see
that S is a closed subspace. We might be tempted to say that S consists
of those functions in L2 that arise as boundary values of holomorphic
functions in the upper half-plane; but this heuristic assertion is not exact
if we do not add a quantitative restriction such as in the definition (5)
of the Hardy space. See Exercise 4.
(ii) Suppose we defined P to be the orthogonal projection on the subspace
[
S of L2 . Then, as is easily seen, (P f )(ξ) = χ(ξ)fˆ(ξ) for any f ∈ L2 (R);
here χ is the characteristic function of (0, ∞). The operator P is also
closely related to the Cauchy integral. Indeed, if F is the (unique)
element in H 2 (R2+ ) whose boundary function (according to Theorem 2.3)
is P (f ), then
Z
1 f (t)
F (z) = dt, z ∈ R2+ .
2πi R t − z
To prove this it suffices to verify that for any f ∈ L2 (R) and any fixed
z = x + iy ∈ R2+ , we have
Z ∞ Z
1 f (t)
fˆ(ξ)e2πiξz dξ = dt.
0 2πi R t−z
This is proved in the same way as the Poisson integral representation (10)
except here we use the identity (11) instead of (12). The details may be
left to the interested reader. Also, the reader might note the close analogy
between this version of the Cauchy integral for the upper-half plane, and
a corresponding version for the unit disc, as given in Example 2, Section 4
of Chapter 4.
(iii) In analogy with the periodic case discussed in Exercise 30 of Chap-
ter 4, we define a Fourier multiplier operator T on R to be a linear
operator on L2 (R) determined by a bounded function m (the multi-
plier), such that T is defined by the formula (T [ f )(ξ) = m(ξ)fˆ(ξ) for
2
any f ∈ L (R). The orthogonal projection P above is such an operator
and its multiplier is the characteristic function χ(ξ). Another closely
related operator of this type is the Hilbert transform H defined by
3. Constant coefficient partial differential equations 221
P = I+iH
2 . Then H is a Fourier multiplier operator corresponding to the
multiplier 1i sign(ξ). Among the many important properties of H is its
connection to conjugate harmonic functions. Indeed, for f a real-valued
function in L2 (R), f and H(f ) are, respectively, the real and imaginary
parts of the boundary values of a function in the Hardy space. More
about the Hilbert transform can be found in Exercises 9 and 10 and
Problem 5 below.
3 Constant coefficient partial differential equations

We turn our attention to solving the linear partial differential equation
(13) L(u) = f,
where the operator L takes the form
X µ ¶α
∂
L= aα
∂x
|α|≤n
with aα ∈ C constants.
In the study of the classical examples of L, such as the wave equation,
the heat equation, and Laplace’s equation, one already sees the Fourier
transform entering in an important way.4 For general L, this key role
is further indicated by the following simple observation. If, for example,
we try to solve this equation with both u and f elements in S, then this
is equivalent to the algebraic equation
P (ξ)û(ξ) = fˆ(ξ),
where P (ξ) is the characteristic polynomial of f defined by
X
P (ξ) = aα (2πiξ)α .
|α|≤n
This is because one has the Fourier transform identity

µ\ ¶
∂αf
(ξ) = (2πiξ)α fˆ(ξ).
∂xα
Thus a solution u in the space S (if it exists) would be uniquely deter-
mined by
fˆ(ξ)
û(ξ) = .
P (ξ)
4 See for example Chapters 5 and 6 in Book I.

In a more general setting, matters are not so easy: aside from the ques-
tion of defining (13), the Fourier transform is not directly applicable;
also, solutions that we prove to exist (but are not unique!) have to be
understood in a wider sense.
3.1 Weak solutions

As the reader may have guessed, it will not suffice to restrict our attention
to those functions for which L(u) is defined in the usual way, but instead
a broader notion is needed, one involving the idea of “weak solutions.”
To describe this concept, we start with a given open set Ω in Rd and
consider the space C0∞ (Ω), which consists of the indefinitely differentiable
functions5 having compact support in Ω.6 We have the following fact.
Lemma 3.1 The space C0∞ (Ω) is dense in L2 (Ω) in the norm k · kL2 (Ω) .
The proof is essentially a repetition of that of Lemma 1.2. We take the

precaution of modifying the definition of gM given there to be: gM (x) =
f (x) if |x| ≤ M , d(x, Ωc ) ≥ 1/M and |f (x)| ≤ M , and gM (x) = 0 oth-
erwise. Also, when we regularize gM , we replace it with gM ∗ ϕδ , with
δ < 1/2M . Then the support of gM ∗ ϕδ is still compact and at a distance
≥ 1/2M from Ωc .
We next consider the adjoint operator of L defined by
X µ ¶α
∗ |α| ∂
L = (−1) aα .
∂x
|α|≤n
The operator L∗ is called the adjoint of L because, in analogy with

the definition of the adjoint of a bounded linear transformation given in
Section 5.2 of the previous chapter, we have
(14) (Lϕ, ψ) = (ϕ, L∗ ψ) whenever ϕ, ψ ∈ C0∞ (Ω),
where (·, ·) denotes the inner product on L2 (Ω) (which is the restriction
of the usual inner product on L2 (Rd )). The identity (14) is proved by
successive integration by parts. Indeed, consider first the special case
when L = ∂/∂xj , and then L∗ = −∂/∂xj . If we use Fubini’s theorem,
integrating first in the xj variable, then in this case (14) reduces to the
5 Indefinitely differentiable functions are also referred to as C ∞ functions, or smooth
functions.
6 This means that the closure of the support of f , as defined in Section 1 of Chapter 2,
is compact and contained in Ω.

familiar one-dimensional formula

Z ∞µ ¶ Z ∞ µ ¶
dϕ dψ
ψ dx = − ϕ dx,
−∞ dx −∞ dx
with the integrated boundary terms vanishing because of the assumed

support properties of ψ (or ϕ). Once established for L = ∂/∂xj , 1 ≤ j ≤
α
n, then (14) follows for L = (∂/∂x) by iteration, and hence for general
L by linearity.
At this point we digress momentarily to consider besides C0∞ (Ω) some
other spaces of differentiable functions on Ω that will be useful later.
The space C n (Ω) consists of all functions f on Ω that have continuous
partial derivatives of order ≤ n. Also, the space C n (Ω) consists of those
functions on Ω that can be extended to functions in Rd that belong to
C n (Rd ). Thus, in an obvious sense, we have the inclusion relation
C0∞ (Ω) ⊂ C n (Ω) ⊂ C n (Ω), for each positive integer n.
Returning to our partial differential operator L, it is useful to observe

that the formula
(Lu, ψ) = (u, L∗ ψ)
continues to hold (with the same proof) if we merely assume that u ∈

C n (Ω) without assuming it has compact support, while still supposing
ψ ∈ C0∞ (Ω).
In particular, if we have L(u) = f in the ordinary sense (sometimes
called the “strong” sense), which requires the assumption that u ∈ C n (Ω)
in order to define the partial derivatives entering in Lu, then we would
also have
(15) (f, ψ) = (u, L∗ ψ) for all ψ ∈ C0∞ (Ω).
This leads to the following important definition: if f ∈ L2 (Ω), a function

u ∈ L2 (Ω) is a weak solution of the equation Lu = f in Ω if (15) holds.
Of course an ordinary solution is always a weak solution.
Significant instances of weak solutions that are not ordinary solutions
already arise in elementary situations such as in the study of the one-
dimensional wave equation. Here L(u) = (∂ 2 u/∂x2 ) − (∂ 2 u/∂t2 ), so the
underlying space is R2 = {(x1 , x2 ) : with x1 = x, x2 = t}. Suppose, for
example, we consider the case of the “plucked string.”7 We are then
7 See Chapter 1 in Book I.

looking at the solution of L(u) = 0 subject to the boundary conditions

u(x, 0) = f (x) and (∂u/∂t)(x, 0) = 0 for 0 ≤ x ≤ π, where the graph of
f is piecewise linear and is illustrated in Figure 2.
0 p π
Figure 2. Initial position of a plucked string
If one extends f to [−π, π] by making it odd, and then to all of R

by periodicity (of period 2π), then the solution is given by d’Alembert’s
formula
f (x + t) + f (x − t)
u(x, t) = .
2
In the present case u is not twice continuously differentiable, and it is
therefore not an ordinary solution. Nevertheless it is a weak solution.
To see this, approximate f by a sequence of functions fn that are C ∞
and such that fn → f uniformly on every compact subset of R.8 If we
define un (x, t) as [fn (x + t) + fn (x − t)]/2, we can check directly that
L(un ) = 0 and hence (un , L∗ ψ) = 0 for all ψ ∈ C0∞ (R2 ), and thus by
uniform convergence we obtain that (u, L∗ ψ) = 0 as desired.
A different example illustrating the nature of weak solutions arises for
the operator L = d/dx on R. If we suppose Ω = (0, 1), then with u and
f in L2 (Ω), we have that Lu = f in the weak sense if and only if there is
an absolutely continuous function F on [0, 1] such that F (x) = u(x) and
F 0 (x) = f (x) almost everywhere. For more about this, see Exercise 14.
3.2 The main theorem and key estimate

We now turn to the general theorem guaranteeing the existence of solu-
tions of partial differential equations with constant coefficients
Theorem 3.2 Suppose Ω is a bounded open subset of Rd . Given a linear
partial differential operator L with constant coefficients, there exists a
8 One may write, for example, fn = f ∗ ϕ1/n , where {ϕ² } is the approximation to the
identity, as in the proof of Lemma 1.2.
bounded linear operator K on L2 (Ω) such that whenever f ∈ L2 (Ω), then
L(Kf ) = f in the weak sense.
In other words, u = K(f ) is a weak solution to L(u) = f .
The heart of the matter lies in an inequality that we state next, but
whose proof (which uses the Fourier transform) is postponed until the
next section.
Lemma 3.3 There exists a constant c such that
kψkL2 (Ω) ≤ ckL∗ ψkL2 (Ω) whenever ψ ∈ C0∞ (Ω).
The usefulness of this lemma comes about for the following reason.
If L is a finite-dimensional linear transformation, the solvability of L
(the fact that it is surjective) is of course equivalent with the fact that
its adjoint L∗ is injective. In effect, the lemma provides the analytic
substitute for this reasoning in an infinite-dimensional setting.
We first prove the theorem assuming the validity of the inequality in
the lemma.
Consider the pre-Hilbert space H0 = C0∞ (Ω) equipped with the inner
product and norm
hϕ, ψi = (L∗ ϕ, L∗ ψ), kψk20 = kL∗ ψkL2 (Ω) .
Following the results in Section 2.3 of Chapter 4, we let H denote the

completion of H0 . By Lemma 3.3, a Cauchy sequence in the k · k0 -norm
is also Cauchy in the L2 (Ω)-norm; hence we may identify H with a
subspace of L2 (Ω). Also, L∗ , initially defined as a bounded operator
from H0 to L2 (Ω), extends to a bounded operator L∗ from H to L2 (Ω)
(by Lemma 1.3). For a fixed f ∈ L2 (Ω), consider the linear map `0 :
C0∞ (Ω) → C defined by
`0 (ψ) = (ψ, f ) for ψ ∈ C0∞ (Ω).
The Cauchy-Schwarz inequality together with another application of

Lemma 3.3 yields
|`0 (ψ)| = |(ψ, f )| ≤ kψkL2 (Ω) kf kL2 (Ω)

≤ ckL∗ ψkL2 (Ω) kf kL2 (Ω)
≤ c0 kψk0 ,
with c0 = ckf kL2 (Ω) . Hence `0 is bounded on the pre-Hilbert space H0 .

Therefore, ` extends to a bounded linear functional on H (see Section 5.1,
Chapter 4), and the above inequalities show that k`k ≤ ckf kL2 (Ω) . By
the Riesz representation theorem applied to ` on the Hilbert space H
(Theorem 5.3 in Chapter 4), there exists U ∈ H such that
`(ψ) = hψ, U i = (L∗ ψ, L∗ U ) for all ψ ∈ C0∞ (Ω).
Here h·, ·i denotes the extension to H of the initial inner product on H0 ,

and L∗ also denotes the extension of L∗ originally given on H0 .
If we let u = L∗ U , then u ∈ L2 (Ω), and we find that
`(ψ) = (ψ, f ) = (L∗ ψ, u) for all ψ ∈ C0∞ (Rd ).
Hence
(f, ψ) = (u, L∗ ψ) for all ψ ∈ C0∞ (Rd ),
and by definition, u is a weak solution to the equation Lu = f in Ω. If

we let Kf = u, we see that once f is given, Kf is uniquely determined
by the above steps. Since kU k0 = k`k ≤ ckf kL2 (Ω) we see that
kKf kL2 (Ω) = kukL2 (Ω) = kL∗ U kL2 (Ω) = kU k0 ≤ ckf kL2 (Ω) ,
whence K : L2 (Ω) → L2 (Ω) is bounded.
Proof of the main estimate

To complete the proof of the theorem, we must still prove the estimate
in Lemma 3.3, that is,
kψkL2 (Ω) ≤ ckL∗ ψkL2 (Ω) whenever ψ ∈ C0∞ (Ω).
The reasoning below relies on an important fact: if f has compact

support in R, then fˆ(ξ) initially defined for ξ ∈ R extends to an entire
function for ζ = ξ + iη ∈ C. This observation reduces the problem to an
inequality about holomorphic functions and polynomials.
Lemma 3.4 Suppose P (z) = z m + · · · + a1 z + a0 is a polynonial of de-

gree m with leading coefficient 1. If F is a holomorphic function on C,
then
Z 2π
1
|F (0)|2 ≤ |P (eiθ )F (eiθ )|2 dθ.
2π 0
Proof. The lemma is a consequence of the special case when P = 1

Z 2π Z 2π
2 1
(16) |F (0)| ≤ |F (eiθ )|2 dθ.
2π 0 0
This assertion follows directly from the mean-value identity (8) in Sec-
tion 2 with ζ = 0 and r = 1, via the Cauchy-Schwarz inequality. With it
we begin by factoring P :
Y Y
P (z) = (z − α) (z − β) = P1 (z)P2 (z),
|α|≥1 |β|<1
where each product is finite and taken over the roots of P whose absolute
values are ≥ 1 and < 1,Qrespectively.
Note that |P1 (0)| = |α|≥1 |α| ≥ 1.
For P2 we write
(z − β) = −(1 − βz)ψβ (z),

β−z
where ψβ (z) = 1−βz are the “Blaschke factors” that have the obvious
property that they are holomorphic in a region containing the closed
iθ
unit disc
Q and |ψβ (e )| = 1; see also Chapter 8 in Book II. We iθwrite
P̃2 = |β|<1 (1 − βz) and P̃ = P1 P̃2 . Thus |P̃ (0)| ≥ 1, while |P̃ (e )| =
|P (eiθ )| for every θ. We now apply (16) to the function P̃ F in place of
F and find that
Z 2π
2 2 1
|F (0)| ≤ |P̃ (0)F (0)| ≤ |P̃ (eiθ )F (eiθ )|2 dθ
2π 0
Z 2π
1
= |P (eiθ )F (eiθ )|2 dθ,
2π 0
which gives the desired conclusion.

We turn to the proof of the inequality kψk ≤ ckL∗ ψk for all ψ ∈ C0∞ (Ω)
in the special case of one dimension, that is, Ω ⊂ R.
Suppose f is an L2 function supported on the interval [−M, M ]. Then
Z M
fˆ(ξ) = f (x)e−2πixξ dx
−M
whenever ξ ∈ R. In fact, the above integral converges whenever ξ is re-

placed by ζ = ξ + iη ∈ C, and we may extend fˆ to a holomorphic func-
tion of ζ in the whole complex plane. An application of the Plancherel
formula (for fixed η) yields

Z ∞ Z ∞
ˆ 2
|f (ξ + iη)| dξ ≤ e4πM |η|
|f (x)|2 dx.
−∞ −∞
We use this observation in the following context. We may assume (upon

multiplying L by a suitable constant) that
X µ ¶k
∂
L∗ = (−1)k ak ,
∂x
0≤k≤n
P
where an = (2πi)−n . If we let Q(ξ) = 0≤k≤n (−1)k ak (2πiξ)k be its
characteristic polynomial, then we note that
d
L∗ ψ(ξ) = Q(ξ)ψ̂(ξ) whenever ψ ∈ C0∞ (R).
If M is chosen so large that Ω ⊂ [−M, M ], then our previous observation

gives
Z ∞ Z ∞
2 4πM |η|
(17) |Q(ξ + iη)ψ̂(ξ + iη)| dξ ≤ e |L∗ ψ(x)|2 dx.
−∞ −∞
Picking η = i sin θ, and making a translation by cos θ yields

Z ∞
|Q(ξ + cos θ + i sin θ)ψ̂(ξ + cos θ+i sin θ)|2 dξ ≤
−∞
Z ∞
≤ e4πM |L∗ ψ(x)|2 dx.
−∞
An application of Lemma 3.4 with F (z) = ψ̂(ξ + z) and Q(ξ + z) in place

of P (z) then gives
Z 2π
2 1
|ψ̂(ξ)| ≤ |Q(ξ + cos θ + i sin θ)ψ̂(ξ + cos θ + i sin θ)|2 dθ.
2π 0
We now integrate in ξ over R, and on the right-hand side interchange
the order of the ξ and θ integrations; also by translation invariance we
replace the integration in the ξ variable by that in the variable ξ + cos θ.
Using (17) the result is
Z 2π Z
2 1
kψ̂kL2 (R) ≤ |Q(ξ + i sin θ)ψ̂(ξ + i sin θ)|2 dξ dθ
2π 0
Z R
4πM
≤e |L∗ ψ(x)|2 dx,
R
4*. The Dirichlet principle 229
which by Plancherel’s identity proves the main lemma in the one-dimensional

case.
The higher
P dimensional case is a modification of the argument above.
Let Q = |α|≤n (−1) aα (2πiξ)α be the characteristic polynomial of L∗ .
α
Then we can choose a new set of orthogonal axes (whose coordinates we

denote by (ξ1 , . . . , ξd )) so that if ξ = (ξ1 , ξ 0 ) with ξ 0 = (ξ2 , . . . , ξd ), then
after multiplying by a suitable constant
n−1
X
(18) Q(ξ) = (2πi)−n ξ1n + ξ1j qj (ξ 0 ),
j=0
where qj (ξ 0 ) are polynomials of ξ 0 (of degrees ≤ n − j).

To see that such a choice is possible, write Q = Qn + Q0 , where Qn is
homogeneous of degree n and Q0 has degree < n. Then since we may
assume Qn 6= 0 there is (after multiplying Q by a suitable constant),
a unit vector γ so that Qn (γ) = (2πi)−n . Then Qn (ξ) = (2πi)−n rn if
ξ = γr, r ∈ R. We can then take the ξ1 -axis to lie along γ, and the
ξ2 , . . . , ξd -axes to be in mutually orthogonal directions, from which the
form (18) is clear.
Proceeding now as before we obtain
Z 2π
0 2 1
|ψ̂(ξ1 , ξ )| ≤ |Q(ξ1 + eiθ , ξ 0 )ψ̂(ξ1 + eiθ , ξ 0 )|2 dθ
2π 0
for each (ξ1 , ξ 0 ) ∈ Rd . An integration9 then gives

Z 2π Z
1
kψ̂k2L2 (Rd ) ≤ |Q(ξ1 + i sin θ, ξ 0 )ψ̂(ξ1 + i sin θ, ξ 0 )|2 dξ dθ.
2π 0 Rd
If we suppose that the projection of the (bounded) set Ω on the x1 -axis

is contained in [−M, M
R ], we see as before that the right-hand side above
is majorized by e4πM Rd |L∗ ψ(x)|2 dx, finishing the proof of Lemma 3.3
and hence that of the theorem.
4* The Dirichlet principle

Dirichlet’s principle arose in the study of the boundary-value problem
for Laplace’s equation. Stated in the case of two dimensions it refers to
the classical problem of finding the steady-state temperature of a plate
9 We note that by the rotational invariance of Lebesgue measure (Problem 4 in Chap-
ter 2 and Exercise 26 in Chapter 3), integration in ξ can be carried out in the new
coordinates as well.
whose boundary is exposed to a given temperature distribution. The

issue raised is the following question, called the Dirichlet problem:
If Ω is a bounded open set in R2 and f a continuous function on the
boundary ∂Ω, we wish to find a function u(x1 , x2 ) such that
½
4u = 0 in Ω,
(19)
u = f on ∂Ω.
Thus we need to determine a function that is C 2 (twice continuously

differentiable) in Ω, whose Laplacian10 is zero, and which is continuous
on the closure of Ω, with u|∂Ω = f .
With either Ω or f satisfying special symmetry conditions, the solution
to this problem can sometimes be written out explicitly. For instance, if
Ω is the unit disc, then
Z π
iθ 1
u(re ) = f (ϕ)Pr (θ − ϕ) dϕ,
2π −π
where Pr is the Poisson kernel (for the disc). We also obtained (in Books I
and II) explicit formulas for the solution of the Dirichlet problem for some
unbounded domains. For example, when Ω is the upper half-plane the
solution is
Z
u(x, y) = Py (x − t)f (t) dt,
R
where Py (x) is the analogous Poisson kernel for the upper half-plane. A
somewhat similar convolution formula was obtained when Ω is a strip.
Also, the Dirichlet problem can be solved explicitly for certain Ω by using
conformal mappings.11
In general, however, there are no explicit solutions, and other methods
must be found. An idea that was used intially was based on an approach
of wide utility in mathematics and physics: to find the equilibrium state
of a system one seeks to minimize an appropriate “energy” or “action.”
In the present case the role of this energy is played by the Dirichlet
integral, which is defined for appropriate functions U by
Z Z ¯ ¯ ¯ ¯
¯ ∂U ¯2 ¯ ∂U ¯2
D(U ) = |∇U |2 = ¯ ¯ +¯ ¯
¯ ¯ ¯ ∂x2 ¯ dx1 dx2 .
Ω Ω ∂x1
(Note the similarity with the expression of the “potential energy” in the
case of the vibrating string in Chapters 3 and 6 of Book I.) In fact,
P
10 The Laplacian of a function u in Rd is defined by 4u = dk=1 ∂ 2 u/∂x2k .
11 The close relation between conformal maps and the Dirichlet problem is discussed in
the last part of Section 1 of Chapter 8, in Book II.
that approach underlies the proof Riemann proposed for his well-known
mapping theorem. About this early history R. Courant has written:
Already some years before the rise of Riemann’s genius,

C.F. Gauss and W. Thompson had observed that the bound-
ary value problem of the harmonic differential equation 4u =
uxx + uyy = 0 for a domain G in the x, y-plane can be re-
duced to the problem of minimizing the integral D[φ] for the
domain G, under the condition that the functions φ admitted
to competition have the prescribed boundary values. Because
of the positive character of D[φ] the existence of a solution
for the latter problem was considered obvious and hence the
existence for the former assured. As a student in Dirichlet’s
lectures, Riemann had been fascinated by this convincing ar-
gument: soon afterwards he used it, under the name “Dirich-
let’s Principle,” in a more varied and spectacular manner as
the very foundation of his new geometric function theory.
The application of Dirichlet’s principle was thought to have been jus-

tified by the following simple observation:
Proposition 4.1 Suppose there exists a function u ∈ C 2 (Ω) that mini-

mizes D(U ) among all U ∈ C 2 (Ω) with U |∂Ω = f . Then u is harmonic
in Ω.
Proof. For functions F and G in C 2 (Ω) define the following inner-

product
Z µ ¶
∂F ∂G ∂F ∂G
hF, Gi = + dx1 dx2 .
Ω ∂x1 ∂x1 ∂x2 ∂x2
We then note that D(u) = hu, ui. If v is any function in C 2 (Ω) with
v|∂Ω = 0, then for all ² we have
D(u + ²v) ≥ D(u),
since u + ²v and u have the same boundary values, and u minimizes the
Dirichlet integral. We note, however, that
D(u + ²v) = D(u) + ²2 D(v) + ²hu, vi + ²hv, ui.
Hence
²2 D(v) + ²hu, vi + ²hv, ui ≥ 0,

and since ² can be both positive or negative, this can happen only if
Rehu, vi = 0. Similarly, considering the perturbation u + i²v, we find
Imhu, vi = 0, and therefore hu, vi = 0. An integration by parts then pro-
vides
Z
0 = hu, vi = − (4u)v
Ω
for all v ∈ C 2 (Ω) with v|∂Ω = 0. This implies that 4u = 0 in Ω, and of

course u equals f on the boundary.
Nevertheless, several serious objections were later raised to Dirich-

let’s principle. The first was by Weierstrass, who pointed out that it
was not clear (and had not been proved) that a minimizing function for
the Dirichlet integral exists, so there might simply be no winner to the
implied competition in Proposition 4.1. He argued by analogy with a
simpler one-dimensional problem: that of minimizing the integral
Z 1
D(ϕ) = |xϕ0 (x)|2 dx
−1
among all C 1 functions on [−1, 1] that satisfy ϕ(−1) = −1 and ϕ(1) = 1.

The minimum value achieved by this integral is zero. To verify this, let
ψ be a smooth non-decreasing function on R that satisfies ψ(x) = 1 for
x ≥ 1, and ψ(x) = −1 if x ≤ −1. For each 0 < ² < 1, we consider the
function

 1 if ² ≤ x,
ϕ² (x) = ψ(x/²) if −² < x < ²,

−1 if x ≤ −².
Then ϕ² satisfies the desired constraints, and if M denotes a bound for

the derivative of ψ, we find
Z ²
D(ϕ² ) = |x|2 |²−1 ψ 0 (x/²)|2 dx
−²
Z ²
≤ |ψ 0 (x/²)|2 dx
−²
≤ 2²M 2 .
In the limit as ² tends to 0, we find that the minimum value of the integral
D(ϕ) is zero. This minimum value cannot be reached by a C 1 function
satisfying the boundary conditions, since D(ϕ) = 0 implies ϕ0 (x) = 0 and
thus ϕ is constant.
A further objection was raised by Hadamard, who remarked that D(u)

may be infinite even for a solution u of the boundary value problem:
thus, in effect, there may simply be no competitors who qualify for the
competition!
To illustrate this point, we return to the disc, and consider the function
∞
X n
f (θ) = fα (θ) = 2−nα ei2 θ
n=0
for α > 0. This function first appeared in Chapter 4 of Book I, where it

is shown that fα is continuous but nowhere differentiable if α ≤ 1. The
solution of the Dirichlet problem on the unit disc with boundary value
fα is given by the Poisson integral
∞
X n n
u(r, θ) = r2 2−nα ei2 θ .
n=0
However, the use of polar coordinates gives

¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯2
¯ ∂u ¯2 ¯ ∂u ¯2 ¯ ∂u ¯2 1 ¯ ∂u ¯
¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯
¯ ∂x1 ¯ + ¯ ∂x2 ¯ = ¯ ∂r ¯ + r2 ¯ ∂θ ¯ .
Thus
ZZ Ã¯ ¯ ¯ ¯ ! Z ρZ Ã¯ ¯ 2 ¯ ¯2 !
¯ ∂u ¯2 ¯ ∂u ¯2 2π ¯ ∂u ¯ ¯ ¯
¯ ¯ +¯ ¯ dx1 dx2 = ¯ ¯ + ¯ ∂u ¯ dθrdr
1
¯ ∂x1 ¯ ¯ ∂x2 ¯ ¯ ∂r ¯ r ¯ ∂θ ¯
2
Dρ 0 0
where Dρ is the disc of radius 0 < ρ < 1 centered at the origin. Since
∂u X n −nα 2n −1 i2n θ ∂u X 2n −nα n i2n θ

∼ 2 2 r e and ∼ r 2 i2 e ,
∂r ∂θ
applications of Parseval’s identity lead to
Z Z Ã¯ ¯ ¯ ¯ ! Z ∞
ρX
¯ ∂u ¯2 ¯ ∂u ¯2 n+1
¯ ¯ +¯ ¯ dx1 dx2 ≈ 22n+1 2−2nα r2 −1
dr
¯ ∂x1 ¯ ¯ ∂x2 ¯
Dρ 0 n=0
∞
X n+1
2 n −2nα
= ρ 2 2 ,
n=0
which tends to infinity as ρ → 1 if α ≤ 1/2.

One can formulate this objection in a more precise way by appealing
to the result in Exercise 20.
Despite these significant difficulties, Dirichlet’s principle can indeed be

validated, if applied in the appropriate way. A key insight is that the
space of competing functions arising in the proof of the above proposition
is itself a pre-Hilbert space, with inner product h·, ·i given there. The
desired solution lies in the completion of this pre-Hilbert space, and this
requires the L2 theory for its analysis. These ideas were clearly not
available at the time Dirichlet’s principle was first formulated and used.
In what follows we shall describe how these additional concepts can
be exploited. We will begin our presentation in the more general d-
dimensional setting, but conclude with the application of these tech-
niques to the solution of the two-dimensional problem (19). As an impor-
tant preliminary matter we start with the study of some basic properties
of harmonic functions.
4.1 Harmonic functions

Throughout this section Ω will denote an open subset of Rd . A function u
is harmonic in Ω if it is twice continuously differentiable12 and u solves
d
X ∂2u
4u = = 0.
∂x2j
j=1
We shall see that harmonic functions can be characterized by a number

of equivalent properties.13 Adapting the terminology used in Section 3,
we say that u is weakly harmonic in Ω if
(20) (u, 4ψ) = 0 for every ψ ∈ C0∞ (Ω).
Note that the left-hand side of (20) is well-defined for any u that is inte-
grable on compact subsets of Ω. Thus, in particular, a weakly harmonic
function needs to be defined only almost everywhere. Clearly, however,
any harmonic function is weakly harmonic.
Another notion is the mean-value property generalizing the iden-
tity (9) in Section 2 for holomorphic functions. A continuous function u
defined in Ω satisfies this property if
Z
1
(21) u(x0 ) = u(x) dx
m(B) B
for each ball B whose center is x0 and whose closure B is contained in Ω.
12 In other words, u is in C 2 (Ω) in the notation of Section 3.1.

13 Note that in the case of one dimension, harmonic functions are linear and so their
theory is essentially trivial.
The following two theorems give alternative characterizations of har-

monic functions. Their proofs are closely intertwined.
Theorem 4.2 If u is harmonic in Ω, then u satisfies the mean-value
property (21). Conversely, a continuous function satisfying the mean-
value property is harmonic.
Theorem 4.3 Any weakly harmonic function u in Ω can be corrected
on a set of measure zero so that the resulting function is harmonic in Ω.
The above statement says that for a given weakly harmonic function u
there exists a harmonic function ũ, so that ũ(x) = u(x) for a.e. x ∈ Ω.
Notice since ũ is necessarily continuous it is uniquely determined by u.
Before we prove the theorems, we deduce a noteworthy corollary. It is
a version of the maximum principle.
Corollary 4.4 Suppose Ω is a bounded open set, and let ∂Ω = Ω − Ω
denote its boundary. Assume that u is continuous in Ω and is harmonic
in Ω. Then
max |u(x)| = max |u(x)|.
x∈Ω x∈∂Ω
Proof. Since the sets Ω and ∂Ω are compact and u is continuous, the
two maxima above are clearly attained. We suppose that maxx∈Ω |u(x)|
is attained at an interior point x0 ∈ Ω, for otherwise there is nothing to
prove. R
1
Now by the mean-value property, |u(x0 )| ≤ m(B) B
|u(x)| dx. If for
0 0
some point x ∈ B we had |u(x )| < |u(x0 )|, then a similar inequality
would hold in a small neighborhood of x0 ,R and since |u(x)| ≤ |u(x0 )|
1
throughout B, the result would be that m(B) B
|u(x)| dx < |u(x0 )|, which
is a contradiction. Hence |u(x)| = |u(x0 )| for each x ∈ B. Now this is
true for each ball Br of radius r, centered at x0 , such that Br ⊂ Ω. Let
r0 be the least upper bound of such r; then B r0 intersects the boundary
Ω at some point x̃. Since |u(x)| = |u(x0 )| for all x ∈ B r , r < r0 , it follows
by continuity that |u(x̃)| = |u(x0 )|, proving the corollary.
Turning to the proofs of the theorems, we first establish a variant
of Green’s formula (for the unit ball) that does not explicitly involve
boundary terms.14 Here u, v, and η are assumed to be twice continuously
differentiable functions in a neighborhood of the closure of B, but η is
also supposed to be supported in a compact subset of B.
14 The more usual version requires integration over the (boundary) sphere, a topic
deferred to the next chapter. See also Exercises 6 and 7 in that chapter.
Lemma 4.5 We have the identity

Z Z
(v4u − u4v)η dx = u(∇v · ∇η) − v(∇u · ∇η) dx.
B B
³ ´
∂u ∂u ∂u
Here ∇u is the gradient of u, that is, ∇u = ∂x1 , ∂x2 , . . . , ∂xd and
d
X ∂v ∂η
∇v · ∇η = ,
∂xj ∂xj
j=1
with ∇u · ∇η defined similarly.

In fact, by integrating by parts as in the proof of (14) we have
Z Z Z
∂u ∂v ∂η
vη dx = − u η dx − uv dx.
B ∂x j B ∂x j B ∂x j
We then repeat this with u replaced by ∂u/∂xj , and sum in j to obtain

Z Z Z
(4u)vη dx = − (∇u · ∇v)η dx − (∇u · ∇η)v dx.
B B B
This yields the lemma if we subtract from this the symmetric formula
with u and v interchanged.
We shall apply the lemma when u is a given harmonic function, while
v is one of the three following “test” functions: first, v(x) = 1; second,
v(x) = |x|2 ; and third, v(x) = |x|−d+2 if d ≥ 3, while v(x) = log |x| if
d = 2. The relevance of these choices arises because 4v = 0 in the first
case, while 4v is a non-zero constant in the second case; also v in the
third case is a constant multiple of a “fundamental solution,” and in
particular v(x) is harmonic for x 6= 0.
When v(x) = 1, we take η = η²+ , where η²+ (x) = 1 for |x| ≤ 1 − ²,
η²+ (x)
= 0 for 1, and |∇η²+ (x)| ≤ c/². We accomplish this by setting
³ |x| ≥ ´
|x|−1+²
η²+ (x) =χ ² for 1 − ² ≤ |x| ≤ 1, where χ is a fixed C 2 function
on [0, 1] that equals 1 in [0, 1/4] and equals 0 in [3/4, 1]. A picture of η²+
is given in Figure 3.
Since u is harmonic, we see that with v = 1, Lemma 4.5 implies
Z
(22) ∇u · ∇η²+ dx = 0.
B
Next we take v(x) = |x| ; then clearly 4v = 2d, and with η = η²+ the
2
lemma yields:
Z Z Z
2d uη²+ dx = |x|2 (∇u · ∇η²+ ) dx − 2 u(x · ∇η²+ ) dx.
B B B
1−² 1 |x|
Figure 3. The function η²+
However, since ∇η²+ is supported in the spherical shell S²+ = {x : 1 − ² ≤

|x| ≤ 1}, we see that
Z Z
|x|2 (∇u · ∇η²+ ) dx = (∇u · ∇η²+ ) dx + O(²),
B B
and hence by (22) we get

Z Z
(23) d u dx = − lim u(x · ∇η²+ ) dx.
B ²→0 B
We finally turn to v(x) = |x|−d+2 , when d ≥ 3, and calculate (4v)(x)

for x 6= 0 to see that it vanishes there. In fact, since ∂|x|/∂xj = xj /|x|,
we note that
∂|x|a ∂ 2 |x|a
= axj |x|a−2 and = a|x|a−2 + a(a − 2)x2j |x|a−4 .
∂xj ∂x2j
Upon adding in j, we obtain that 4(|x|a ) = [da + a(a − 2)]|x|a−2 , and

this is zero if a = −d + 2 (or a = 0). A similar argument shows that
4(log |x|) = 0 when d = 2 and x 6= 0.
We now apply the lemma with this v and η = η² defined as follows:
η² (x) = 1 − χ(|x|/²) for |x| ≤ ²,

η² (x) = 1 ³ ´ for ² ≤ |x| ≤ 1 − ²,
+ |x|−1+²
η² (x) = η² (x) = χ ² for 1 − ² ≤ |x| ≤ 1.
The picture for η² is as follows (Figure 4):

0 ² 1−² 1 |x|
Figure 4. The function η²
We note that |∇η² | is O(1/²) throughout. Now both u and v are

harmonic in the support of η² , and in this case ∇η² is supported only
near the unit sphere (in the shell S²+ ) or near the origin (in the ball
B² = {|x| < ²}). Thus the right-hand side of the identity of the lemma
gives two contributions, one over S²+ and the other over B² . We consider
the first contribution (when d ≥ 3); it is
Z Z
u∇(|x|−d+2 ) · ∇η² dx − |x|−d+2 (∇u · ∇η²+ ) dx.
S²+ S²+
R
Now the first integral is (−d + 2) S²+ u|x|−d (x · ∇η²+ ) dx, which by (23)
R
tends to c B u dx as ² → 0, where c is the constant (2 − d)d, since |x|−d −
1 = O(²) over S²+ . The second term tends to zero as ² → 0 because of (22)
and the fact that the integrand there is supported in the shell S²+ . A
similar argument for d = 2, with v(x) = log |x|, yields the result with
c = 1.
To consider the contribution near the origin, that is, over B² , we tem-
porarily make the additional assumption that u(0) = 0. Then because
of the differentiability assumption satisfied by a harmonic function, we
have u(x)
R = O(|x|) as |x| → 0. Now over B² we have two terms, the first
being B² u∇(|x|−d+2 )∇η² dx, which is majorized by
Z µZ ¶
O(²)|x|−d+1 O(1/²) dx ≤ O |x|−d+1 dx ≤ O(²),
B² |x|≤²
because of (8) in Section 2 of Chapter 2. This term tends to 0 with ².

R
The second term is B²
|x|−d+2 (∇u · ∇η² ) dx, which is majorized by
Z
c1
|x|−d+2 = c2 ²,
² |x|≤²
using the result just cited. We have used the fact that ∇u is bounded
and ∇η² is O(1/²) throughout B. Letting ² → 0 we see that this term
tends to zero also. A similar argument works when d = 2.
Thus we have proved that if u is harmonic Rin a neighborhood of the
closure of the unit ball B, and u(0) = 0, then B u dx = 0. We can drop
the assumption u(0) = 0 by applying the conclusion we have just reached
to u(x) − u(0) in place of u(x). Therefore we have achieved the mean-
value property (21) for the unit ball.
Now suppose Br (x0 ) = {x : |x − x0 | < r} is the ball of radius r cen-
tered at x0 , and consider U (x) = u(x0 + rx). If we suppose that u is har-
monic in Br (x0 ), then clearly U is harmonic in the unit ball (indeed, the
property of being harmonic is unchanged under translations x → x + x0
and dilations x → rx, as is easily verified). Thus if u were supported
R in Ω,
1
and Br (x0 ) ⊂ Ω, then by the result just proved U (0) = m(B) B
U (x) dx,
which means that
Z Z
1 1
u(x0 ) = u(x0 + rx) dx = d u(x0 + x) dx
m(B) |x|≤1 r m(B) |x|≤r
Z
1
= u(x) dx,
m(Br (x0 )) Br
by the relative invariance of Lebesgue measure under dilations and trans-

lations. This establishes (21) in general.
The converse property

To prove this, we first show that the mean-value property allows a useful
extension of itself. For this purpose, we fix a function ϕ(y) that is contin-
uous in the closed unit ball {|y| ≤ 1} and is radial (that is, ϕ(y) = Φ(|y|)
for an appropriateR Φ), and extend ϕ to be zero when |y| > 1. Suppose
in addition that ϕ(y) dy = 1. We then claim the following:
Lemma 4.6 Whenever u satisfies the mean-value property (21) in Ω,
and the closure of the ball {x : |x − x0 | < r} lies in Ω, then
(24) Z Z
u(x0 ) = u(x0 − ry)ϕ(y) dy = u(x0 − y)ϕr (y) dy = (u ∗ ϕr )(x0 ),
Rd Rd
where ϕr (y) = r−d ϕ(y/r).

That the second of the two identities holds is an immediate consequence

of the change of variables y → y/r; the rightmost equality is merely the
definition of u ∗ ϕr .
We can prove (24) as a consequence of a simple observation about
integration. Let ψ(y) be another function on the ball {|y| ≤ 1}, which
we assume is bounded. For each N , a large positive integer, denote by
B(j) the ball {|y| ≤ j/N }. Recall that ϕ(y) = Φ(|y|). Then
Z N
X µ ¶Z
j
(25) ϕ(y)ψ(y) dy = lim Φ ψ(y) dy.
N →∞ N B(j)−B(j−1)
j=1
To verify this, note that the left-hand side of (25) equals

N Z
X
ϕ(y)ψ(y) dy.
j=1 B(j)−B(j−1)
However, sup1≤j≤N supy∈B(j)−B(j−1) |ϕ(y) − Φ(j/N )| = ²N , which tends

to zero as N → ∞, since ϕ is radial, continuous, and ϕ(y) = Φ(|y|). Thus
PN R
the left-hand side of (25) differs from j=1 Φ(j/N ) B(j)−B(j−1) ψ(y) dy
R
by at most ²N |y|≤1 |ψ(y)| dy, proving (25).
We now use this in the case where ψ(y) = u(x0 − ry) and ϕ is as before.
Then
Z XN µ ¶Z
j
u(x0 − ry)ϕ(y) dy = lim Φ u(x0 − ry) dy.
N →∞ N B(j)−B(j−1)
j=1
However, it follows from the mean-value property assumed for u that

Z
u(x0 − ry) dy = u(x0 )[m(B(j)) − m(B(j − 1))].
B(j)−B(j−1)
Therefore, the right-hand side above equals

N
X µ ¶Z
j
u(x0 ) lim Φ dy,
N →∞ N B(j)−B(j−1)
j=1
and this
R is u(x0 ) if we use (25) again, this time with ψ = 1, and recall
that ϕ(y) dy = 1. We have therefore proved the lemma.
We see from this that every continuous function which satisfies the
mean-value property is its own regularization! To be precise, we have
(26) u(x) = (u ∗ ϕr )(x)

whenever x ∈ Ω and the distance from x to the boundary of Ω is larger

than r. If we now require in addition that ϕ ∈ C0∞ {|y| < 1}, then by the
discussion in Section 1 we conclude that u is smooth throughout Ω.
Let us now establish that such functions are harmonic. Indeed, by
Taylor’s theorem, for every x0 ∈ Ω
d
X d
1 X
(27) u(x0 + x) − u(x0 ) = aj x j + ajk xj xk + ²(x),
2
j=1 j,k=1
R
where ²(x) = O(|x|3 ) as |x| → 0. We note next that |x|≤r xj dx = 0 and
R
x x dx = 0 for all j and k with k 6= j. This follows by carrying
|x|≤r j k
out the integrations first in the xj variable and noting that the integral
vanishes
R because
R xj is2 an odd function. Also by an obvious symmetry
2
x dx = |x|≤r xk dx, and by the relative dilation-invariance (see
|x|≤r j R
Section 3, Chapter 1) these are equal to r2 |x|≤r (x1 /r)2 dx =
R
rd+2 |x|≤1 x21 dx = crd+2 , with c > 0. We now integrate both sides of (27)
over the ball {|x| ≤ r}, divide by rd , and use the mean-value property.
The result is that
d µ Z ¶
c 2X cr2 1
r ajj = (4u)(x0 ) = O |²(x)| dx = O(r3 ).
2 2 rd |x|≤r
j=1
Letting r → 0 then gives 4u(x0 ) = 0. Since x0 was an arbitrary point

of Ω, the proof of Theorem 4.2 is concluded.
Theorem 4.3 and some corollaries

We come now to the proof of Theorem 4.3. Let us assume that u is
weakly harmonic in Ω. For each ² > 0 we define Ω² to be the set of
points in Ω that are at a distance greater than ² from its boundary:
Ω² = {x ∈ Ω : d(x, ∂Ω) > ²}.
Notice that Ω² is open, and that every point of Ω belongs to Ω² if ²

is small enough. Then the regularization u ∗ ϕr = ur considered in the
previous theorem is defined in Ω² , for r < ², and as we have noted is a
smooth function there. We next observe that it is weakly harmonic in
Ω² . In fact, for ψ ∈ C0∞ (Ω² ) we have

Z µZ ¶
(ur , 4ψ) = u(x − ry)ϕ(y) dy (4ψ)(x) dx
Rd Rd
Z µZ ¶
= ϕ(y) u(x − ry)(4ψ)(x) dx dy,
Rd Rd
by Fubini’s theorem, and the inner integral vanishes for y, |y| ≤ 1, be-
cause it equals (u, 4ψr ), with ψr = ψ(x + ry). Thus we have
(u ∗ ϕr , 4ψ) = 0,
and hence u ∗ ϕr is weakly harmonic. Next, since this regularization is

automatically smooth it is then also harmonic. Moreover, we claim that
(28) (u ∗ ϕr1 )(x) = (u ∗ ϕr2 )(x)
whenever x ∈ Ω² and r1 + r2 < ². Indeed, (u ∗ ϕr1 ) ∗ ϕr2 = u ∗ ϕr1 as

we have shown in (26) above. However convolutions are commutative
(see Remark (6) in Chapter 2); thus (u ∗ ϕr1 ) ∗ ϕr2 = (u ∗ ϕr2 ) ∗ ϕr1 =
u ∗ ϕr2 , and (28) is proved.
Now we can let r1 tend to zero, while keeping r2 fixed. We know by the
properties of approximations to the identity that u ∗ ϕr1 (x) → u(x) for
almost every x in Ω² ; hence u(x) equals ur2 (x) for almost every x ∈ Ω² .
Thus u can be corrected on Ω² (setting it equal to ur2 ), so that it becomes
harmonic there. Now since ² can be taken arbitrarily small, the proof of
the theorem is complete.
We state several further corollaries arising out of the above theorems.

Corollary 4.7 Every harmonic function is indefinitely differentiable.
Corollary 4.8 Suppose {un } is a sequence of harmonic functions in Ω
that converges to a function u uniformly on compact subsets of Ω as
n → ∞. Then u is also harmonic.
The first of these corollaries was already proved as a consequence
of (26). For the second, we use the fact that each un satisfies the mean-
value property
Z
1
un (x0 ) = un (x) dx
m(B) B
whenever B is a ball with center at x0 , and B ⊂ Ω. Thus by the uniform

convergence it follows that u also satisfies this property, and hence u is
harmonic.
We should point out that these properties of harmonic functions on

Rd are reminiscent of similar properties of holomorphic functions. But
this should not be surprising, given the close connection between these
two classes of functions in the special case d = 2.
4.2 The boundary value problem and Dirichlet’s principle

The d-dimensional Dirichlet boundary value problem we are concerned
with may be stated as follows. Let Ω be an open bounded set in Rd .
Given a continuous function f defined on the boundary ∂Ω, we wish to
find a function u that is continuous in Ω, harmonic in Ω, and such that
u = f on ∂Ω.
An important preliminary observation is that the solution to the prob-
lem, if it exists, is unique. Indeed, if u1 and u2 are two solutions
then u1 − u2 is harmonic in Ω and vanishes on the boundary. Thus by
the maximum principle (Corollary 4.4) we have u1 − u2 = 0, and hence
u1 = u2 .
Turning to the existence of a solution, we shall now pursue the ap-
proach of Dirichlet’s principle outlined earlier.
We consider the class of functions C 1 (Ω), and equip this space with
the inner product
Z
hu, vi = (∇u · ∇v) dx,
Ω
where of course
d
X ∂u ∂v
∇u · ∇v = .
∂xj ∂xj
j=1
With this inner product, we have a corresponding norm given by

kuk2 = hu, ui. We note that kuk = 0 is the same as ∇u = 0 through-
out Ω, which means that u is constant on each connected component of
Ω. Thus we are led to consider equivalence classes in C 1 (Ω) of elements
modulo functions that are constant on components of Ω. These then
form a pre-Hilbert space with inner product and norm given as above.
We call this pre-Hilbert space H0 .
In studying the completion H of H0 and its applications to the bound-
ary value problem, the following lemma is needed.
Lemma 4.9 Let Ω be an open bounded set in Rd . Suppose v belongs to
C 1 (Ω) and v vanishes on ∂Ω. Then
Z Z
(29) |v(x)|2 dx ≤ cΩ |∇v(x)|2 dx.
Ω Ω
Proof. This conclusion could in fact be deduced from the considera-

tions given in Lemma 3.3. We prefer to prove this easy version separately
to highlight a simple idea that we shall also use later. It should be noted
that the argument yields the estimate cΩ ≤ d(Ω)2 , where d(Ω) is the
diameter of Ω.
We proceed on the basis of the following observation. Suppose f is a
function in C 1 (I), where I = (a, b) is an interval in R. Assume that f
vanishes at one of the end-points of I. Then
Z Z
2 2
(30) |f (t)| dt ≤ |I| |f 0 (t)|2 dt,
I I
where |I| denotes the length of I. Rs

Indeed, suppose f (a) = 0. Then f (s) = a f 0 (t) dt, and by the Cauchy-
Schwarz inequality
Z s Z
2
|f (s)| ≤ |I| 0 2
|f (t)| dt ≤ |I| |f 0 (t)|2 dt.
a I
Integrating this in s over I then yields (30).

To prove (29), write x = (x1 , x0 ) with x1 ∈ R and x0 ∈ Rd−1 and ap-
ply (30) to f defined by f (x1 ) = v(x1 , x0 ), with x0 fixed. Let J(x0 )
be the open set in R that is the corresponding slice of Ω given by
{x1 ∈ R : (x1 , x0 ) ∈ Ω}. The set J(x0 ) can be written as a disjoint union
of open intervals Ij . (Note that in fact f (x1 ) vanishes at both end-points
of each Ij .) For each j, on applying (30) we obtain
Z Z
0 2 2
|v(x1 , x )| dx1 ≤ |Ij | |∇v(x1 , x0 )|2 dx1 .
Ij Ij
Now since |Ij | ≤ d(Ω), summing over the disjoint intervals Ij gives
Z Z
|v(x1 , x0 )|2 dx1 ≤ d(Ω)2 |∇v(x1 , x0 )|2 dx1 ,
J(x0 ) J(x0 )
and an integration over x0 ∈ Rd then leads to (29).

Now let S0 denote the linear subspace of C 1 (Ω) consisting of functions
that vanish on the boundary of Ω. We note that distinct elements of S0
remain distinct under the equivalence relation defining H0 (since con-
stants on each component that vanish on the boundary are zero), and so
S0 may be identified with a subspace of H0 . Denote by S the closure in
H of this subspace, and let PS be the orthogonal projection of H onto S.
With these preliminaries out of the way, we first try to solve the bound-
ary value problem with f given on ∂Ω under the additional assumption
that f is the restriction to ∂Ω of a function F in C 1 (Ω). (How this
additional hypothesis can be removed will be explained below.) Fol-
lowing the prescription of Dirichlet’s principle, we seek a sequence {un }
with un ∈ C 1 (Ω) and un |∂Ω = F |∂Ω , such that the Dirichlet integrals
kun k2 converge to a minimum value. This means that un = F − vn ,
with vn ∈ S0 , and that limn→∞ kun k minimizes the distance from F to
S0 . Since S = S0 , this sequence also minimizes the distance from F to
S in H.
Now what do the elementary facts about orthogonal projections teach
us? According to the proof of Lemma 4.1 in the previous chapter, we
conclude that the sequence {vn }, and hence also the sequence {un },
both converge in the norm of H, the former having a limit PS (F ). Now
applying Lemma 4.9 to vn − vm we deduce that {vn } and {un } are also
Cauchy in the L2 (Ω)-norm, and thus converge also in the L2 -norm. Let
u = limn→∞ un . Then
(31) u = F − PS (F ).
We see that u is weakly harmonic. Indeed, whenever ψ ∈ C0∞ (Ω), then

ψ ∈ S, and hence by (31) hu, ψi = 0. Therefore hun , ψi → 0, but by
integration by parts, as we have seen,
Z Z
hun , ψi = (∇un · ∇ψ) dx = − un 4ψ dx = −(un , 4ψ).
Ω Ω
As a result, (u, 4ψ) = 0, and so u is weakly harmonic and thus can be

corrected on a set of measure zero to become harmonic.
This is the purported solution to our problem. However, two issues
still remain to be resolved.
The first is that while u is the limit of a sequence {un } of continuous
functions in Ω and un |∂Ω = f , for each n, it is not clear that u itself is
continuous in Ω and u|∂Ω = f .
The second issue is that we restricted our argument above to those
f defined on the boundary of Ω that arise as restrictions of functions
in C 1 (Ω).
The second obstacle is the easier of the two to overcome, and this can
be done by the use of the following lemma, applied to the set Γ = ∂Ω.
Lemma 4.10 Suppose Γ is a compact set in Rd , and f is a continuous
function on Γ. Then there exists a sequence {Fn } of smooth functions
on Rd so that Fn → f uniformly on Γ.
In fact, supposing we can deal with the first issue raised, then with the
lemma we proceed as follows. We find the functions Un that are har-
monic in Ω, continuous on Ω, and such that Un |∂Ω = Fn |∂Ω . Now since
the {Fn } converges uniformly (to f ) on ∂Ω, it follows by the maximum
principle that the sequence {Un } converges uniformly to a function u
that is continuous on Ω, has the property that u|∂Ω = f , and which is
moreover harmonic (by Corollary 4.8 above). This achieves our goal.
The proof of Lemma 4.10 is based on the following extension principle.
Lemma 4.11 Let f be a continuous function on a compact subset Γ of
Rd . Then there exists a function G on Rd that is continuous, and so that
G|∂Γ = f .
Proof. We begin with the observation that if K0 and K1 are two
disjoint compact sets, there exists a continuous function 0 ≤ g(x) ≤ 1 on
Rd which takes the value 0 on K0 and 1 on K1 . Indeed, if d(x, Ω) denotes
the distance from x to Ω, we see that
d(x, K0 )
g(x) =
d(x, K0 ) + d(x, K1 )
has the required properties.
Now, we may assume without loss of generality that f is non-negative
and bounded by 1 on Γ. Let
K0 = {x ∈ Γ : 2/3 ≤ f (x) ≤ 1} and K1 = {x ∈ Γ : 0 ≤ f (x) ≤ 1/3},
so that K0 and K1 are disjoint. Clearly, the observation before the

lemma guarantees that there exists a function 0 ≤ G1 (x) ≤ 1/3 on Rd
which takes the value 1/3 on K0 and 0 on K1 . Then we see that
2
0 ≤ f (x) − G1 (x) ≤ for all x ∈ Γ.
3
We now repeat the argument with f replaced by f − G1 . In the first
step, we have gone from 0 ≤ f ≤ 1 to 0 ≤ f − G1 ≤ 2/3. Consequently,
we may find a continuous function G2 on Rd so that
µ ¶2
2
0 ≤ f (x) − G1 (x) − G2 (x) ≤ on Γ,
3
and 0 ≤ G2 ≤ 13 23 . Repeating this process, we find continuous functions

Gn on Rd such that
µ ¶N
2
0 ≤ f (x) − G1 (x) − · · · − GN (x) ≤ on Γ,
3
1
¡ 2 ¢N −1
and 0 ≤ GN ≤ 3 3 on Rd . If we define
∞
X
G= Gn ,
n=1
then G is continuous and equals f on Γ.

To complete the proof of Lemma 4.10, we argue as follows. We regu-
larize the function G obtained in Lemma 4.11 by defining
Z Z
−d
F² (x) = ² G(x − y)ϕ(y/²) dy = G(y)ϕ² (x − y) dy,
Rd Rd
with ϕ² (y) = ²−d ϕ(y/²), where ∞

R ϕ is a non-negative C0 function sup-
ported in the unit ball with ϕ(y) dy = 1. Then each F² is a C ∞ func-
tion. However,
Z
F² (x) − G(x) = (G(y) − G(x))ϕ² (x − y) dy.
Since the integration above is restricted to |x − y| ≤ ², then if x ∈ Γ, we

see that
Z
|F² (x) − G(x)| ≤ sup |G(x) − G(y)| ϕ² (x − y) dy
|x−y|≤²
≤ sup |G(x) − G(y)|.

|x−y|≤²
The last quantity tends to zero with ² by the uniform continuity of G

near Γ, and if we choose ² = 1/n we obtain our desired sequence.
The two-dimensional theorem

We now take up the problem of whether the proposed solution u takes
on the desired boundary values. Here we limit our discussion to the case
of two dimensions for the reason that in the higher dimensional situation
the problems that arise involve a number of questions that would take
us beyond the scope of this book. In contrast, in two dimensions, while
the proof of the result below is a little tricky, it is within the reach of the
Hilbert space methods we have been illustrating.
The Dirichlet problem can be solved (in two dimensions as well as
in higher dimensions) only if certain restrictions are made concerning
the nature of the domain Ω. The regularity we shall assume, while not
optimal,15 is broad enough to encompass many applications, and yet

has a simple geometric form. It can be described as follows. We fix an
initial triangle T0 in R2 . To be precise, we assume that T0 is an isosceles
triangle whose two equal sides have length `, and make an angle α at
their common vertex. The exact values of ` and α are unimportant;
they may both be taken as small as one wishes, but must be kept fixed
throughout our discussion. With the shape of T0 thus determined, we
say that T is a special triangle if it is congruent to T0 , that is, T arises
from T0 by a translation and rotation. The vertex of T is defined to be
the intersection of its two equal sides.
The regularity property of Ω we assume, the outside-triangle con-
dition, is as follows: with ` and α fixed, for each x in the boundary of
Ω, there is a special triangle with vertex x whose interior lies outside Ω.
(See Figure 5.)
Ω
α
x
`
T0
T
∂Ω
Figure 5. The triangle T0 and the special triangle T
Theorem 4.12 Let Ω be an open bounded set in R2 that satisfies the

outside-triangle condition. If f is a continuous function on ∂Ω, then the
boundary value problem 4u = 0 with u continuous in Ω and u|∂Ω = f is
always uniquely solvable.
Some comments are in order.
(1) If Ω is bounded by a polygonal curve, it satisfies the conditions of
the theorem.
(2) More generally, if Ω is appropriately bounded by finitely many Lips-
chitz curves, or in particular C 1 curves, the conditions are also satisfied.
(3) There are simple examples where the problem is not solvable: for
instance, if Ω is the punctured disc. This example of course does not
satisfy the outside-triangle condition.
15 The optimal conditions involve the notion of capacity of sets.

(4) The conditions on Ω in this theorem are not optimal: one can con-
struct examples of Ω when the problem is solvable for which the above
regularity fails.
For more details on the above, see Exercise 19 and Problem 4.
We turn to the proof of the theorem. It is based on the following
proposition, which may be viewed as a refined version of Lemma 4.9
above.
Proposition 4.13 For any bounded open set Ω in R2 that satisfies the
outside-triangle condition there are two constants c1 < 1 and c2 > 1 such
that the following holds. Suppose z is a point in Ω whose distance from
∂Ω is δ. Then whenever v belongs to C 1 (Ω) and v|∂Ω = 0, we have
Z Z
2 2
(32) |v(x)| dx ≤ Cδ |∇v(x)|2 dx.
Bc1 δ (z) Bc2 δ (z)∩Ω
The bound C can be chosen to depend only on the diameter of Ω and the
parameters ` and α which determine the triangles T .
Bc2 δ (z) Ω
Bc1 δ (z)
Figure 6. The situation in Proposition 4.13
Let us see how the proposition proves the theorem. We have already
shown that it suffices to assume that f is the restriction to ∂Ω of an
F that belongs to C 1 (Ω). We recall we had the minimizing sequence
un = F − vn , with vn ∈ C 1 (Ω) and vn |∂Ω = 0. Moreover, this sequence
converges in the norm of H and L2 (Ω) to a limit v, such that u = F − v
is harmonic in Ω. Then since (32) holds for each vn , it also holds for
v = F − u; that is,
Z Z
2 2
(33) |(F − u)(x)| dx ≤ Cδ |∇(F − u)(x)|2 dx.
Bc1 δ (z) Bc2 δ (z)∩Ω
To prove the theorem it suffices, in view of the continuity of u in Ω, to

show that if y is any fixed point in ∂Ω, and z is a variable point in Ω,
then u(z) → f (y) as z → y. Let δ = δ(z) denote the distance of z from
the boundary. Then δ(z) ≤ |z − y| and therefore δ(z) → 0 as z → y.
We now consider the averages of F and u taken over the discs centered
at z of radius c1 δ(z) (recall that c1 < 1). We denote these averages
by Av(F )(z) and Av(u)(z), respectively. Then by the Cauchy-Schwarz
inequality, we have
Z
2 1
|Av(F )(z) − Av(u)(z)| ≤ |F − u|2 dx,
π(c1 δ)2 Bc δ (z)∩Ω
1
which by (33) is then majorized by

Z
0
C |∇(F − u)|2 dx.
Bc2 δ (z)∩Ω
The absolute continuity of the integral guarantees that the last integral
tends to zero with δ, since m(Bc2 δ ) → 0. However, by the mean-value
property, Av(u)(z) = u(z), while by the continuity of F in Ω,
Z
1
Av(F )(z) = F (x) dx → f (y),
m(Bc1 δ (z)) Bc δ (z)
1
because F |∂Ω = f and z → y. Altogether this gives u(z) → f (y), and the
theorem is proved, once the proposition is established.
To prove the proposition, we construct for each z ∈ Ω whose distance
from ∂Ω is δ, and for δ sufficiently small, a rectangle R with the following
properties:
(1) R has side lengths 2c1 δ and M δ (with c1 ≤ 1/2, M ≤ 4).
(2) Bc1 δ (z) ⊂ R.
(3) Each segment in R, that is parallel to and of length equal to the

length of the long side, intersects the boundary of Ω.
To obtain R we let y be a point in ∂Ω so that δ = |z − y|, and we apply

the outside-triangle condition at y. As a result, the line joining z with
y and one of the sides of the special triangle whose vertex is at y must
make an angle β < π. (In fact β ≤ π − α/2, as is easily seen.) Now after
a suitable rotation and translation we may assume that y = 0 and that
the angle going from the x2 -axis to the line joining z to 0 is equal to the
γ
∂Ω
γ
Side of triangle
x2
Figure 7. Placement of the rectangle R
angle of the side of the triangle to the x2 -axis. This angle can be taken
to be γ, with γ > α/4. (See Figure 7.)
There is an alternate possibility that occurs with this figure reflected
through the x2 -axis.
With this picture in mind we construct the rectangle R as indicated
in Figure 8.
It has its long side parallel to the x2 -axis, contains the disc Bc1 δ (z),
and every segment R parallel to the x2 -axis intersects the (extension) of
the side of the triangle.
Note that the coordinates of z are (−δ sin γ, δ cos γ). We choose c1 <
sin γ, then Bc1 δ (z) lies in the same (left) half-plane as z.
We next focus our attention on two points: P1 , which lies on the x1 -
axis at the intersection of this axis with the far side of the rectangle; and
P2 , which is at the corner of that side of the rectangle, that is, at the
intersection of the (continuation) of the side of the outside triangle and
the further side of the rectangle. The coordinates of P1 are (−a, 0), where
a = δc1 + δ sin γ. The coordinates of P2 are (−a, −a cos γ
sin γ ). Note that the
distance of P2 from the origin is a/ sin γ, which is δ + c1 δ/ sin γ ≤ 2δ,
since c1 < sin γ.
Now we observe that the length of the larger side of the rectangle is
the sum of the part that lies above the x1 -axis and the part that lies
below. The upper part has length the sum of the radius of the disc plus
the height of z, and this is c1 δ + δ cos γ ≤ 2δ. The lower part has length
equal to a/ tan γ, which is δ cos γ + δc1 cos γ
sin γ ≤ 2δ, since c1 < sin γ. Thus
Bc1 δ (z)
γ
∂Ω
P1
x1
x2
P2
Figure 8. The disc Bc1 δ (z) and the rectangle R containing it
we find that the length of the side is ≤ 4δ.

Now it is clear from the construction that each vertical segment in R
starting from the disc Bc1 δ (z) when continued downward and parallel to
the x2 -axis intersects the line joining 0 to P2 , (which is a continuation
of the side of the triangle). Moreover, if the length ` of this side of the
triangle exceeds the distance of P2 from the origin, then the segment in-
tersects the triangle. When this intersection occurs the segment starting
from Bc2 δ (z) must also intersect the boundary of Ω, since the triangle
lies outside Ω. Therefore if ` ≥ 2δ the desired intersection occurs, and
each of the conclusions (1), (2), and (3) are verified. (We shall lift the
restriction δ ≤ `/2 momentarily.)
Now we integrate over each line segment parallel to the x2 -axis in R,
including its portion in Bc1 δ (z), which is continued downward until it
meets ∂Ω. Call such a segment I(x1 ). Then, using (30) we see that
Z Z ¯ ¯2
¯ ∂v ¯
2 2 2
|v(x1 , x2 )| dx2 ≤ M δ ¯ ¯
¯ ∂x2 (x1 , x2 )¯ dx2 ,
I(x1 ) I(x1 )
and an integration in x1 gives

Z Z
|v(x)|2 dx ≤ M δ 2 |∇v(x)|2 dx.
R∩Ω R∩Ω
5. Exercises 253
However, we note that Bc1 δ (z) ⊂ R, and Bc2 δ (z) ⊃ R when c2 ≥ 2. Thus
the desired inequality (32) is established, still under the assumption that
δ is small, that is, δ ≤ `/2. When δ > `/2 it suffices merely to use the
crude estimate (29) and the proposition is then proved. The proof of the
theorem is therefore complete.
5 Exercises
1. Suppose f ∈ L2 (Rd ) and k ∈ L1 (Rd ).

R
(a) Show that (f ∗ k)(x) = f (x − y)k(y) dy converges for a.e. x.
(b) Prove that kf ∗ kkL2 (Rd ) ≤ kf kL2 (Rd ) kkkL1 (Rd ) .
\
(c) Establish (f ∗ k)(ξ) = k̂(ξ)fˆ(ξ) for a.e. ξ.
(d) The operator T f = f ∗ k is a Fourier multiplier operator with multiplier

m(ξ) = k̂(ξ).
[Hint: See Exercise 21 in Chapter 2.]
2. Consider the Mellin transform defined initially for continuous functions f of

compact support in R+ = {t ∈ R : t > 0} and x ∈ R by
Z ∞
Mf (x) = f (t)tix−1 dt.
0
Prove that (2π)−1/2 M extends to a unitary operator from L2 (R+ , dt/t) to L2 (R).
The Mellin transform serves on R+ , with its multiplicative structure, the same
purpose as the Fourier transform on R, with its additive structure.
3. Let F (z) be a bounded holomorphic function in the half-plane. Show in two

ways that limy→0 F (x + iy) exists for a.e. x.
(a) By using the fact that F (z)/(z + i) is in H 2 (R2+ ).
“ ”
(b) By noting that G(z) = F i 1−z
1+z
is a bounded holomorphic function in the
unit disc, and using Exercise 17 in the previous chapter.
4. Consider F (z) = ei/z /(z + i) in the upper half-plane. Note that F (x + iy) ∈
L2 (R), for each y > 0 and y = 0. Observe also that F (z) → 0 as |z| → 0. However,
F ∈/ H 2 (R2+ ). Why?
5. For a < b, let Sa,b denote the strip {z = x + iy, a < y < b}. Define H 2 (Sa,b )
to consist of the holomorphic functions F in Sa,b so that
Z
kF k2H 2 (Sa,b ) = sup |F (x + iy)|2 dx < ∞.
a<y<b R2
Define H 2 (Sa,∞ ) and H 2 (S−∞,b ) to be the obvious variants of the Hardy spaces
for the half-planes {z = x + iy, y > a} and {z = x + iy, y < b}, respectively.
(a) Show that F ∈ H 2 (Sa,b ) if and only if F can be written as
Z
F (z) = f (ξ)e−2πizξ dξ,
R
R
with R
|f (ξ)|2 (e4πaξ + e4πbξ ) dξ < ∞.
(b) Prove that every F ∈ H 2 (Sa,b ) can be decomposed as F = G1 + G2 , where

G∈ H 2 (Sa,∞ ) and G2 ∈ H 2 (S−∞,b ).
(c) Show that lima<y<b,y→a F (x + iy) = Fa (x) exists in the L2 -norm and also
almost everywhere, with a similar result for lima<y<b,y→b F (x + iy).
6. Suppose Ω is an open set in C = R2 , and let H be the subspace of L2 (Ω)

consisting of holomorphic functions on Ω. Show that H is a closed subspace of
L2 (Ω), and hence is a Hilbert space with inner product
Z
(f, g) = f (z)g(z) dx dy, where z = x + iy.
Ω
c
[Hint: Prove that for f ∈ H, we have |f (z)| ≤ d(z,Ωc )
kf k for z ∈ Ω, where c =
−1/2
π , using the mean-value property (9). Thus if {fn } is a Cauchy sequence in
H, it converges uniformly on compact subsets of Ω.]
7. Following up on the previous exercise, prove:
(a) If {ϕn }∞
n=0 is an orthonormal basis of H, then
∞
X c2
|ϕn (z)|2 ≤ for z ∈ Ω.
n=0
d(z, Ωc )
(b) The sum

∞
X
B(z, w) = ϕn (z)ϕn (w)
n=0
converges absolutely for (z, w) ∈ Ω × Ω, and is independent of the choice of

the orthonormal basis {ϕn } of H.
(c) To prove (b) it is useful to characterize the function B(z, w), called the
Bergman kernel, by the following property. Let T be the linear transfor-
mation on L2 (Ω) defined by
Z
T f (z) = B(z, w)f (w) du dv, w = u + iv.
Ω
Then T is the orthogonal projection of L2 (Ω) to H.

5. Exercises 255
P∞
(d) Suppose that Ω is the unit disc. Then f ∈ H exactly when f (z) = n=0 an z n ,
with
∞
X
|an |2 (n + 1)−1 < ∞.
n=0
n
Also, the sequence { z (n+1) ∞
π 1/2
}n=0 is an orthonormal basis of H. Moreover,
in this case
1
B(z, w) = .
π(1 − zw)2
8. Continuing with Exercise 6, suppose Ω is the upper half-plane R2+ . Then every
f ∈ H has a representation
√ Z ∞
(34) f (z) = 4π fˆ0 (ξ)e2πiξz dξ, z ∈ R2+ ,
0
R∞
where 0 |fˆ0 (ξ)|2 dξ
ξ
< ∞. Moreover, the mapping fˆ0 → f given by (34) is a uni-
tary mapping from L2 ((0, ∞), dξ
ξ
) to H.
9. Let H be the Hilbert transform. Verify that

(a) H ∗ = −H, H 2 = −I, and H is unitary.
(b) If τh denotes the translation operator, τh (f )(x) = f (x − h), then H com-

mutes with τh , τh H = Hτh .
(c) If δa denotes the dilation operator, δa (f )(x) = f (ax) with a > 0, then H
commutes with δa , δa H = Hδa .
A converse is given in Problem 5 below.
10. Let f ∈ L2 (R) and let u(x, y) be the Poisson integral of f , that is u = (f ∗
Py )(x), as given in (10) above. Let v(x, y) = (Hf ∗ Py )(x), the Poisson integral of
the Hilbert transform of f . Prove that:
(a) F (x + iy) = u(x, y) + iv(x, y) is analytic in the half-plane R2+ , so that u and
v are conjugate harmonic functions. We also have f = limy→0 u(x, y) and
Hf = limy→0 v(x, y).
1
R dt
(b) F (z) = πi R
f (t) t−z .
1 x
(c) v(x, y) = f ∗ Qy , where Qy (x) = π x2 +y 2
is the conjugate Poisson kernel.
i
[Hint: Note that πz
= Py (x) + iQy (x), z = x + iy.]
11. Show that

 „ «n ff∞
1 i−z
π 1/2 (i + z) i+z n=0
is an orthonormal basis of H 2 (R2+ ).

n “ ”n o ∞
Note that π1/21(i+x) i−x
i+x
is an orthonormal basis of L2 (R); see Exer-
n=0
cise 9 in the previous chapter.
[Hint: It suffices to show that if F ∈ H 2 (R2+ ) and
Z ∞
(x + i)n
F (x) dx = 0 for n = 0, 1, 2, . . .,
−∞ (x − i)n+1
then F = 0. Use the Cauchy integral formula to prove that

„ «n
d
(F (z)(z + i)n )|z=i = 0,
dz
and thus F (n) (i) = 0 for n = 0, 1, 2, . . ..]
12. We consider whether the inequality
kukL2 (Ω) ≤ ckL(u)kL2 (Ω)
can hold for open sets Ω that are unbounded.
(a) Assume d ≥ 2. Show that for each constant coefficient partial differential
operator L, there are unbounded connected open sets Ω for which the above
holds for all u ∈ C0∞ (Ω).
(b) Show that kukL2 (Rd ) ≤ ckL(u)kL2 (Rd ) for all u ∈ C0∞ (Rd ) if and only if
|P (ξ)| ≥ c > 0 all ξ, where P is the characteristic polynomial of L.
[Hint: For (a) consider first L = (∂/∂x1 )n and a strip {x : −1 < x1 < 1}.]
13. Suppose L is a linear partial differential operator with constant coefficients.

Show that when d ≥ 2, the linear space of solutions u of L(u) = 0 with u ∈ C ∞ (Rd )
is not finite-dimensional.
[Hint: Consider the zeroes ζ of P (ζ), ζ ∈ Cd , where P is the characteristic poly-
nomial of L.]
14. Suppose F and G are two integrable functions on a bounded interval [a, b].
Show that G is the weak derivative of F if and only if F can be corrected on a set
of measure 0, such that F is absolutely continuous and F 0 (x) = G(x) for almost
every x.
[Hint: If G is the weak derivative of F , use an approximation to show that
Z b Z b
G(x)ϕ(x)dx = − F (x)ϕ0 (x)dx
a a
holds for the function ϕ illustrated in Figure 9.]

5. Exercises 257
0 α−h α β β+h
Figure 9. The function ϕ in Exercise 14
15. Suppose f ∈ L2 (Rd ). Prove that there exists g ∈ L2 (Rd ) such that
„ «α
∂
f (x) = g(x)
∂x
in the weak sense, if and only if
(2πiξ)α fˆ(ξ) = ĝ(ξ) ∈ L2 (Rd ).
16. Sobolev embedding theorem. Suppose n is the smallest integer > d/2. If
„ «α
∂
f ∈ L2 (Rd ) and f ∈ L2 (Rd )
∂x
in the weak sense, for all 1 ≤ |α| ≤ n, then f can be modified on a set of measure
zero so that f is continuous and bounded.
[Hint: Express f in terms of fˆ, and show that fˆ ∈ L1 (Rd ) by the Cauchy-Schwarz
inequality.]
17. The conclusion of the Sobolev embedding theorem fails when n = d/2. Con-
sider the case d = 2, and let f (x) = (log 1/|x|)α η(x), where η is a smooth cut-
off function with η = 1 for x near the origin, but η(x) = 0 if |x| ≥ 1/2. Let
0 < α < 1/2.
(a) Verify that ∂f /∂x1 and ∂f /∂x2 are in L2 in the weak sense.
(b) Show that f cannot be corrected on a set of measure zero such that the
resulting function is continuous at the origin.
18. Consider the linear partial differential operator

X „ «α
∂
L= aα .
∂x
|α|≤n
Then
X
P (ξ) = aα (2πiξ)α
|α|≤n
is called the characteristic polynomial of L. The differential operator L is said

to be elliptic if
|P (ξ)| ≥ c|ξ|n for some c > 0 and all ξ sufficiently large.
P
(a) Check that L is elliptic if and only if |α|=n aα (2πξ)α vanishes only when
ξ = 0.
(b) If L is elliptic, prove that for some c > 0 the inequality

‚ „ «α ‚
‚ ∂ ‚ ` ´
‚ ϕ‚ ≤ c kLϕkL2 (Rd ) + kϕkL2 (Rd )
‚ ∂x ‚
L2 (Rd )
holds for all ϕ ∈ C0∞ (Ω) and |α| ≤ n.
(c) Conversely, if (b) holds then L is elliptic.
19. Suppose u is harmonic in the punctured unit disc D∗ = {z ∈ C : 0 < |z| < 1}.
(a) Show that if u is also continuous at the origin, then u is harmonic throughout
the unit disc.
[Hint: Show that u is weakly harmonic.]
(b) Prove that the Dirichlet problem for the punctured unit disc is in general
not solvable.
20. Let F be a continuous function onR the closure D of the unit disc. Assume that
F is in C 1 on the (open) disc D, and D |∇F |2 < ∞.
P∞Let f (eiθ ) denote the restriction
P∞ of F to the unit circle, and write f (eiθ ) ∼
inθ
n=−∞ an e . Prove that n=−∞ |n| |an |2 < ∞.
P R
[Hint: Write F (reiθ ) ∼ ∞ n=−∞ Fn (r)e
inθ
, with Fn (1) = an . Express D |∇F |2 in
polar coordinates, and use the fact that
Z 1 Z 1
1
|F (1)|2 ≤ L−1 |F 0 (r)|2 dr + L |F (r)|2 dr,
2 1/2 1/2
for L ≥ 2; apply this to F = Fn , L = |n|.]

6. Problems 259
6 Problems
1. Suppose F0 (x) ∈ L2 (R). Then a necessary and sufficient condition that there
exists an entire analytic function F , such that |F (z)| ≤ Aea|z| for all z ∈ C, and
F0 (x) = F (x) a.e. x ∈ R, is that F̂0 (ξ) = 0 whenever |ξ| > a/2π.
R∞
[Hint: Consider the regularization F ² (z) = −∞ F (z − t)ϕ² (t) dt and apply to it
the considerations in Theorem 3.3 of Chapter 4 in Book II.]
2. Suppose Ω is an open bounded subset of R2 . A boundary Lipschitz arc γ is

a portion of ∂Ω which after a rotation of the axes is represented as
γ = {(x1 , x2 ) : x2 = η(x1 ), a ≤ x1 ≤ b},
where a < b and γ ⊂ ∂Ω. It is also supposed that
(35) |η(x1 ) − η(x01 )| ≤ M |x1 − x01 |, whenever x1 , x01 ∈ [a, b],
and moreover if γδ = {(x1 , x2 ) : x2 − δ ≤ η(x1 ) ≤ x2 }, then γδ ∩ Ω = ∅ for some

δ > 0. (Note that the condition (35) is satisfied if η ∈ C 1 ([a, b]).)
Suppose Ω satisfies the following condition.
S There are finitely many open discs
D1 , D2 , . . . , DN with the property that j Dj contains ∂Ω and for each j, ∂Ω ∩ Dj
is a boundary Lipschitz arc (see Figure 10). Then Ω verifies the outside-triangle
condition of Theorem 4.12, guaranteeing the solvability of the boundary value
problem.
Dj
Figure 10. A domain with boundary Lipschitz arcs
3.∗ Suppose the bounded domain Ω has as its boundary a closed simple continuous
curve. Then the boundary value problem is solvable for Ω. This is because there
exists a conformal map Φ of the unit disc D to Ω that extends to a continuous

bijection from D to Ω. (See Section 1.3 and Problem 6∗ in Chapter 8 of Book II.)
4. Consider the two domains Ω in R2 given by Figure 11.
Domain I Domain II
Figure 11. Domains with a cusp
The set I has as its boundary a smooth curve, with the exception of an (inside)
cusp. The set II is similar, except it has an outside cusp. Both I and II fall
within the scope of the result of Problem 3, and hence the boundary value problem
is solvable in each case. However, II satisfies the outside-triangle condition while
I does not.
5. Let T be a Fourier multiplier operator on L2 (Rd ). That is, suppose there

is a bounded function m such that (T [ f )(ξ) = m(ξ)fˆ(ξ), all f ∈ L2 (Rd ). Then T
commutes with translations, τh T = T τh , where τh (f )(x) = f (x − h), for all h ∈ Rd .
Conversely any bounded operator on L2 (Rd ) that commutes with translations
is a Fourier multiplier operator.
[Hint: It suffices to prove that if a bounded operator T̂ commutes with multiplica-
tion by exponentials e2πiξ·h , h ∈ Rd , then there is an m so that T̂ g(ξ) = m(ξ)g(ξ)
for all g ∈ L2 (Rd ). To do this, show first that
T̂ (Φg) = ΦT̂ (g), all g ∈ L2 (Rd ), whenever Φ ∈ C0∞ (Rd ).
Next, for large N , choose Φ so that it equals 1 in the ball |ξ| ≤ N . Then m(ξ) =
T̂ (Φ)(ξ) for |ξ| ≤ N .]
As a consequence of this theorem show that if T is a bounded operator on L2 (R)
that commutes with translations and dilations (as in Exercise 9 above), then
(a) If (T f )(−x) = T (f (−x)) it follows T = cI, where c is an appropriate con-
stant and I the identity operator.
(b) If (T f )(−x) = −T (f (−x)), then T = cH, where c is an appropriate constant
and H the Hilbert transform.
6. This problem provides an example of the contrast between analysis on L1 (Rd )

and L2 (Rd ).
6. Problems 261
Recall that if f is locally integrable on Rd , the maximal function f ∗ is defined

by
Z
1
f ∗ (x) = sup |f (y)| dy,
x∈B m(B) B
where the supremum is taken over all balls containing the point x.
Complete the following outline to prove that there exists a constant C so that
kf ∗ kL2 (Rd ) ≤ Ckf kL2 (Rd ) .
In other words, the map that takes f to f ∗ (although not linear) is bounded
on L2 (Rd ). This differs notably from the situation in L1 (Rd ), as we observed in
Chapter 3.
(a) For each α > 0, prove that if f ∈ L2 (Rd ), then

Z
2A
m({x : f ∗ (x) > α}) ≤ |f (x)| dx.
α |f |>α/2
Here, A = 3d will do.

[Hint: Consider f1 (x) = f (x) if |f (x)| ≥ α/2 and 0 otherwise. Check that
f1 ∈ L1 (Rd ), and
{x : f ∗ (x) > α} ⊂ {x : f1∗ (x) > α/2}.]
(b) Show that

Z Z ∞
|f ∗ (x)|2 dx = 2 αm(Eα )dα,
Rd 0
where Eα = {x : f ∗ (x) > α}.
(c) Prove that kf ∗ kL2 (Rd ) ≤ Ckf kL2 (Rd ) .

6 Abstract Measure and
Integration Theory
What immediately suggest itself, then, is that these

characteristic properties themselves be treated as the
main object of investigation, by defining and dealing
with abstract objects which need satisfy no other con-
ditions than those required by the very theory to be
developed.
This procedure has been made use of − more or
less consciously − by mathematicians of every era.
The geometry of Euclid and the literal algebra of the
sixteenth and seventeenth centuries arose in this way.
But only in more recent times has this method, called
the axiomatic method, been consistently developed
and carried through to its logical conclusion.
It is our intention to treat the theories of measure
and integration by means of the axiomatic method just
described.
C. Carathéodory, 1918
In much of mathematics integration plays a significant role. It is used,

in one form or another, when dealing with questions that arise in analysis
on a variety of different spaces. While in some situations it suffices to
integrate continuous or other simple functions on these spaces, the deeper
study of a number of other problems requires integration based on the
more refined ideas of measure theory. The development of these ideas,
going beyond the setting of the Euclidean space Rd , is the goal of this
chapter.
The starting point is a fruitful insight of Carathéodory and the re-
sulting theorems that lead to construction of measures in very general
circumstances. Once this has been achieved, the deduction of the fun-
damental facts about integration in the general context then follows a
familiar path.
We apply the abstract theory to obtain several useful results: the
theory of product measures; the polar coordinate integration formula,
which is a consequence of this; the construction of the Lebesgue-Stieltjes
integral and its corresponding Borel measure on the real line; and the
1. Abstract measure spaces 263
general notion of absolute continuity. Finally, we treat some of the basic

limit theorems of ergodic theory. This not only gives an illustration of
the abstract framework we have established, but also provides a link with
the differentiation theorems studied in Chapter 3.
1 Abstract measure spaces

A measure space consists of a set X equipped with two fundamental
objects:
(I) A σ-algebra M of “measurable” sets, which is a non-empty col-
lection of subsets of X closed under complements and countable
unions and intersections.
(II) A measure µ : M → [0, ∞] with the following defining property:
if E1 , E2 , . . . is a countable family of disjoint sets in M, then
Ã∞ ! ∞
[ X
µ En = µ(En ).
n=1 n=1
A measure space is therefore often denoted by the triple (X, M, µ) to em-

phasize its three main components. Sometimes, however, when there is
no ambiguity we will abbreviate this notation by referring to the measure
space as (X, µ), or simply X.
A feature that a measure space often enjoys is the property of being
σ-finite. This means that X can be written as the union of countably
many measurable sets of finite measure.
At this early stage we give only two simple examples of measure spaces:
(i) The first is the discrete example with X a countable set, X =
{xn }∞
n=1 , M the collection of all subsets of X, and the measure
µ determined by µ(xn ) = µn , with {µn }∞ n=1 a given sequence
P of
(extended) non-negative numbers. Note that µ(E) = xn ∈E µn .
When µn = 1 for all n, we call µ the counting measure, and also
denote it by #. In this case integration will amount to nothing but
the summation of (absolutely) convergent series.
d
(ii) Here X =R R , M is the collection of Lebesgue measurable sets, and
µ(E) = E f dx, where f is a given non-negative measurable func-
tion on Rd . The case f = 1 corresponds to the Lebesgue measure.
The countable additivity of µ follows from the usual additivity and
limiting properties of integrals of non-negative functions proved in
Chapter 2.
264 Chapter 6. ABSTRACT MEASURE AND INTEGRATION THEORY
The construction of measure spaces relevant for most applications require

further ideas, and to these we now turn.
1.1 Exterior measures and Carathéodory’s theorem

To begin the construction of a measure and its corresponding measurable
sets in the general setting requires, as in the special case of Lebesgue mea-
sure considered in Chapter 1, a prerequisite notion of “exterior” measure.
This is defined as follows.
Let X be a set. An exterior measure (or outer measure) µ∗ on
X is a function µ∗ from the collection of all subsets of X to [0, ∞] that
satisfies the following properties:
(i) µ∗ (∅) = 0.
(ii) If E1 ⊂ E2 , then µ∗ (E1 ) ≤ µ∗ (E2 ).
(iii) If E1 , E2 , . . . is a countable family of sets, then
Ã∞ ! ∞
[ X
µ∗ Ej ≤ µ∗ (Ej ).
j=1 j=1
For instance, the exterior Lebesgue measure m∗ in Rd defined in Chap-

ter 1 enjoys all these properties. In fact, this example belongs to a
large class of exterior measures that can be obtained using “coverings”
by a family of special sets whose measures are taken as known. This
idea is systematized by the notion of a “premeasure” taken up below in
Section 1.3. A different type of example is the exterior α-dimensional
Hausdorff measure m∗α defined in Chapter 7.
Given an exterior measure µ∗ , the problem that one faces is how to de-
fine the corresponding notion of measurable sets. In the case of Lebesgue
measure in Rd such sets were characterized by their difference from open
(or closed) sets, when considered in terms of µ∗ . For the general case,
Carathéodory found an ingenious substitute condition. It is as follows.
A set E in X is Carathéodory measurable or simply measurable
if one has
(1) µ∗ (A) = µ∗ (E ∩ A) + µ∗ (E c ∩ A) for every A ⊂ X.
In other words, E separates any set A in two parts that behave well
in regard to the exterior measure µ∗ . For this reason, (1) is sometimes
referred to as the separation condition. One can show that in Rd with the
Lebesgue exterior measure the notion of measurability (1) is equivalent
to the definition of Lebesgue measurability given in Chapter 1. (See

Exercise 3.)
A first observation we make is that to prove a set E is measurable, it
suffices to verify
µ∗ (A) ≥ µ∗ (E ∩ A) + µ∗ (E c ∩ A) for all A ⊂ X,
since the reverse inequality is automatically verified by the sub-additivity
property (iii) of the exterior measure. We see immediately from the
definition that sets of exterior measure zero are necessarily measurable.
The remarkable fact about the definition (1) is summarized in the next
theorem.
Theorem 1.1 Given an exterior measure µ∗ on a set X, the collection
M of Carathéodory measurable sets forms a σ-algebra. Moreover, µ∗
restricted to M is a measure.
Proof. Clearly, ∅ and X belong to M and the symmetry inherent
in condition (1) shows that E c ∈ M whenever E ∈ M. Thus M is non-
empty and closed under complements.
Next, we prove that M is closed under finite unions of disjoint sets,
and µ∗ is finitely additive on M. Indeed, if E1 , E2 ∈ M, and A is any
subset of X, then
µ∗ (A) = µ∗ (E2 ∩ A) + µ∗ (E2c ∩ A)
= µ∗ (E1 ∩ E2 ∩ A) + µ∗ (E1c ∩ E2 ∩ A)+
+ µ∗ (E1 ∩ E2c ∩ A) + µ∗ (E1c ∩ E2c ∩ A)
≥ µ∗ ((E1 ∪ E2 ) ∩ A) + µ∗ ((E1 ∪ E2 )c ∩ A),
where in the first two lines we have used the measurability condition
on E2 and then E1 , and where the last inequality was obtained using
the sub-additivity of µ∗ and the fact that E1 ∪ E2 = (E1 ∩ E2 ) ∪ (E1c ∩
E2 ) ∪ (E1 ∩ E2c ). Therefore, we have E1 ∪ E2 ∈ M, and if E1 and E2 are
disjoint, we find
µ∗ (E1 ∪ E2 ) = µ∗ (E1 ∩ (E1 ∪ E2 )) + µ∗ (E1c ∩ (E1 ∪ E2 ))
= µ∗ (E1 ) + µ∗ (E2 ).
Finally, it suffices to show that M is closed under countable unions of
disjoint sets, and that µ∗ is countably additive on M. Let E1 , E2 , . . .
denote a countable collection of disjoint sets in M, and define
n
[ ∞
[
Gn = Ej and G= Ej .
j=1 j=1
For each n, the set Gn is a finite union of sets in M, hence Gn ∈ M.

Moreover, for any A ⊂ X we have
µ∗ (Gn ∩ A) = µ∗ (En ∩ (Gn ∩ A)) + µ∗ (Enc ∩ (Gn ∩ A))

= µ∗ (En ∩ A) + µ∗ (Gn−1 ∩ A)
Xn
= µ∗ (Ej ∩ A),
j=1
where the last equality is obtained by induction. Since we know that

Gn ∈ M, and Gc ⊂ Gcn , we find that
n
X
µ∗ (A) = µ∗ (Gn ∩ A) + µ∗ (Gcn ∩ A) ≥ µ∗ (Ej ∩ A) + µ∗ (Gc ∩ A).
j=1
Letting n tend to infinity, we obtain

∞
X
µ∗ (A) ≥ µ∗ (Ej ∩ A) + µ∗ (Gc ∩ A) ≥ µ∗ (G ∩ A) + µ∗ (Gc ∩ A)
j=1
≥ µ∗ (A).
Therefore all the inequalities above are equalities, and we conclude that
G ∈ M, as desired. Moreover, by taking A = G in the above, we find
that µ∗ is countably additive on M, and the proof of the theorem is
complete.
Our previous observation that sets of exterior measure 0 are Carathéodory
measurable shows that the measure space (X, M, µ) in the theorem
is complete: whenever F ∈ M satisfies µ(F ) = 0 and E ⊂ F , then
E ∈ M.
1.2 Metric exterior measures

If the underlying set X is endowed with a “distance function” or “met-
ric,” there is a particular class of exterior measures that is of interest in
practice. The importance of these exterior measures is that they induce
measures on the natural σ-algebra generated by the open sets in X.
A metric space is a set X equipped with a function d : X × X →
[0, ∞) that satisfies:
(i) d(x, y) = 0 if and only if x = y.
(ii) d(x, y) = d(y, x) for all x, y ∈ X.
(iii) d(x, z) ≤ d(x, y) + d(y, z), for all x, y, z ∈ X.
The last property is of course called the triangle inequality, and a func-
tion d that satisfies all these conditions is called a metric on X. For
example, the set Rd with d(x, y) = |x − y| is a metric space. Another
example is provided by the space of continuous functions on a compact
set K with d(f, g) = supx∈K |f (x) − g(x)|.
A metric space (X, d) is naturally equipped with a family of open balls.
Here
Br (x) = {y ∈ X : d(x, y) < r}
defines the open ball of radius r centered at x. Together with this, we say
that a set O ⊂ X is open if for any x ∈ O there exists r > 0 so that the
open ball Br (x) is contained in O. A set is closed if its complement is
open. With these definitions, one checks easily that an (arbitrary) union
of open sets is open, and a similar intersection of closed sets is closed.
Finally, on a metric space X we can define, as in Section 3 of Chapter 1,
the Borel σ-algebra, BX , that is the smallest σ-algebra of sets in X
that contains the open sets of X. In other words BX is the intersection
of all σ-algebras that contain the open sets. Elements in BX are called
Borel sets.
We now turn our attention to those exterior measures on X with the
special property of being additive on sets that are “well separated.” We
show that this property guarantees that this exterior measure defines a
measure on the Borel σ-algebra. This is achieved by proving that all
Borel sets are Carathéodory measurable.
Given two sets A and B in a metric space (X, d), the distance between
A and B is defined by
d(A, B) = inf{d(x, y) : x ∈ A and y ∈ B}.
Then an exterior measure µ∗ on X is a metric exterior measure if it

satisfies
µ∗ (A ∪ B) = µ∗ (A) + µ∗ (B) whenever d(A, B) > 0.
This property played a key role in the case of exterior Lebesgue measure.
Theorem 1.2 If µ∗ is a metric exterior measure on a metric space X,

then the Borel sets in X are measurable. Hence µ∗ restricted to BX is a
measure.
Proof. By the definition of BX it suffices to prove that closed sets

in X are Carathéodory measurable. Therefore, let F denote a closed set
and A a subset of X with µ∗ (A) < ∞. For each n > 0, let
An = {x ∈ F c ∩ A : d(x, F ) ≥ 1/n}.
S∞
Then An ⊂ An+1 , and since F is closed we have F c ∩ A = n=1 An .
Also, the distance between F ∩ A and An is ≥ 1/n, and since µ∗ is a
metric exterior measure, we have
(2) µ∗ (A) ≥ µ∗ ((F ∩ A) ∪ An ) = µ∗ (F ∩ A) + µ∗ (An ).
Next, we claim that
(3) lim µ∗ (An ) = µ∗ (F c ∩ A).

n→∞
To see this, let Bn = An+1 ∩ Acn and note that
1
d(Bn+1 , An ) ≥ .
n(n + 1)
Indeed, if x ∈ Bn+1 and d(x, y) < 1/n(n + 1) the triangle inequality shows
that d(y, F ) < 1/n, hence y ∈
/ An . Therefore
µ∗ (A2k+1 ) ≥ µ∗ (B2k ∪ A2k−1 ) = µ∗ (B2k ) + µ∗ (A2k−1 ),
and this implies that
k
X
µ∗ (A2k+1 ) ≥ µ∗ (B2j ).
j=1
A similar argument also gives
k
X
µ∗ (A2k ) ≥ µ∗ (B2j−1 ).
j=1
P P
Since µ∗ (A) is finite, we find that both series µ∗ (B2j ) and µ∗ (B2j−1 )
are convergent. Finally, we note that
∞
X
µ∗ (An ) ≤ µ∗ (F c ∩ A) ≤ µ∗ (An ) + µ∗ (Bj ),
j=n+1
and this proves the limit (3). Letting n tend to infinity in the inequal-
ity (2) we find that µ∗ (A) ≥ µ∗ (F ∩ A) + µ∗ (F c ∩ A), and hence F is
measurable, as was to be shown.
Given a metric space X, a measure µ defined on the Borel sets of X
will be referred to as a Borel measure. Borel measures that assign a
finite measure to all balls (of finite radius) also satisfy a useful regularity
property. The requirement that µ(B) < ∞ for all balls B is satisfied in
many (but not in all) circumstances that arise in practice.1 When it does
hold, we get the following proposition.
Proposition 1.3 Suppose the Borel measure µ is finite on all balls in

X of finite radius. Then for any Borel set E and any ² > 0, there are
an open set O and a closed set F such that E ⊂ O and µ(O − E) < ²,
while F ⊂ E and µ(E − F ) < ².
Proof.
S∞ We need the following preliminary observation. Suppose
∗
F = k=1 Fk , where the Fk are closed sets. Then for any ² > 0, we can
find a closed set F ⊂ F ∗ such that µ(F ∗ − F ) < ². To prove this we can
assume that the sets {Fk } are increasing. Fix a point x0 ∈SX, and let Bn
∞
denote the ball {x : d(x, x0 ) < n}, with B0 = {∅}. Since n=1 Bn = X,
we have that
[
F∗ = F ∗ ∩ (B n − Bn−1 ).
Now for each n, F ∗ ∩ (B n − Bn−1 ) is the limit as k → ∞ of the increasing

sequence of closed sets Fk ∩ (B n − Bn−1 ), so (recalling that Bn has finite
measure) we can find an N = N (n) so that (F ∗ − FN (n) ) ∩ (B n − Bn−1 )
has measure less than ²/2n . If we now let
∞
[ ¡ ¢
F = FN (n) ∩ (B n − Bn−1 ) ,
n=1
P∞
it follows that the measure of F ∗ − F is less that n
n=1 ²/2 = ². We
also see that F ∩ B k is closed since it is the finite union of closed sets.
Thus F itself is closed because, as is easily seen, any set F is closed
whenever the sets F ∩ B k are closed for all k.
Having established the observation, we call C the collection of all sets
that satisfy the conclusions of the proposition. Notice first that if E
belongs to C then automatically so does its complement.
1 This restriction is not always valid for the Hausdorff measures that are considered in
the next chapter.

S∞
Suppose now that E = k=1 Ek , with each Ek ∈ C. Then there are
open
S∞ sets Ok , Ok ⊃ EkS , with µ(Ok − Ek ) < ²/2k . However, P∞if O =
∞ k
O
k=1 k , then O − E ⊂ k=1 (O k − Ek ), and so µ(O − E) ≤ k=1 ²/2 =
².
k
Next,
S∞there are closed sets Fk ⊂ Ek with µ(E k − Fk ) < ²/2 . Thus if
∗
F = k=1 Fk , we see as before that µ(E − F ) < ². However, F ∗ is not
∗
necessarily closed, so we can use our preliminary observation to find a

closed set F ⊂ F ∗ with µ(F ∗
S∞− F ) < ². Thus µ(E − F ) < 2². Since ² is
arbitrary, this proves that k=1 Ek belongs to C.
Let us finally note that any open set O is in C. The property regarding
containment by open sets is immediate. To find a closed F ⊂ O, so
that µ(O − F ) < ², let Fk = {x ∈SB k : d(x, Oc ) ≥ 1/k}. Then it is clear
∞
that each Fk is closed and O = k=1 Fk . We then need only apply the
observation again to find the required set F . Thus we have shown that C
is a σ-algebra that contains the open sets, and hence all Borel sets. The
proposition is therefore proved.
1.3 The extension theorem

As we have seen, a class of measurable sets on X can be constructed
once we start with a given exterior measure. However, the definition of
an exterior measure usually depends on a more primitive idea of measure
defined on a simpler class of sets. This is the role of a premeasure defined
below. As we will show, any premeasure can be extended to a measure
on X. We begin with several definitions.
Let X be a set. An algebra in X is a non-empty collection of subsets
of X that is closed under complements, finite unions, and finite intersec-
tions. Let A be an algebra in X. A premeasure on an algebra A is a
function µ0 : A → [0, ∞] that satisfies:
(i) µ0 (∅) = 0.
(ii) If
S∞ E1 , E2 , . . . is a countable collection of disjoint sets in A with
k=1 Ek ∈ A, then
Ã ∞
! ∞
[ X
µ0 Ek = µ0 (Ek ).
k=1 k=1
In particular, µ0 is finitely additive on A.
Premeasures give rise to exterior measures in a natural way.

Lemma 1.4 If µ0 is a premeasure on an algebra A, define µ∗ on any

subset E of X by
(∞ ∞
)
X [
µ∗ (E) = inf µ0 (Ej ) : E ⊂ Ej , where Ej ∈ A for all j .
j=1 j=1
Then, µ∗ is an exterior measure on X that satisfies:
(i) µ∗ (E) = µ0 (E) for all E ∈ A.
(ii) All sets in A are measurable in the sense of (1).
Proof. Proving that µ∗ is an exterior measure presents no difficulty.

To see why the restriction of µ∗ to A coincides with µ0 , suppose that
E ∈ A. Clearly, one always has µ∗ (E)S≤ µ0 (E) since E covers itself. To
∞
prove the reverse inequality let E ⊂ j=1 Ej , where Ej ∈ A for all j.
Then, if we set
Ã k−1
!
[
Ek0 = E ∩ Ek − Ej ,
j=1
S∞
the sets Ek0 are disjoint elements of A, Ek0 ⊂ Ek and E = k=1 Ek0 . By
(ii) in the definition of a premeasure, we have
∞
X ∞
X
µ0 (E) = µ0 (Ek0 ) ≤ µ0 (Ek ).
k=1 k=1
Therefore, we find that µ0 (E) ≤ µ∗ (E), as desired.

Finally, we must prove that sets in A are measurable for µ∗ . Let A
be any subset of X, E ∈ A, and ² > 0. By definition, there S∞ exists a
countable collection E1 , E2 , . . . of sets in A such that A ⊂ j=1 Ej and
∞
X
µ0 (Ej ) ≤ µ∗ (A) + ².
j=1
Since µ0 is a premeasure, it is finitely additive on A and therefore

∞
X ∞
X ∞
X
µ0 (Ej ) = µ0 (E ∩ Ej ) + µ0 (E c ∩ Ej )
j=1 j=1 j=1
≥ µ∗ (E ∩ A) + µ∗ (E c ∩ A).
Since ² is arbitrary, we conclude that µ∗ (A) ≥ µ∗ (E ∩ A) + µ∗ (E c ∩ A),

as desired.
The σ-algebra generated by an algebra A is by definition the smallest
σ-algebra that contains A. The above lemma then provides the necessary
step for extending µ0 on A to a measure on the σ-algebra generated by
A.
Theorem 1.5 Suppose that A is an algebra of sets in X, µ0 a premea-

sure on A, and M the σ-algebra generated by A. Then there exists a
measure µ on M that extends µ0 .
One notes below that µ is the only such extension of µ0 under the as-
sumption that µ is σ-finite.
Proof. The exterior measure µ∗ induced by µ0 defines a measure µ on
the σ-algebra of Carathéodory measurable sets. Therefore, by the result
in the previous lemma, µ is also a measure on M that extends µ0 . (We
should observe that in general the class M is not as large as the class of
all sets that are measurable in the sense of (1).)
To prove that this extension is unique whenever µ is σ-finite, we argue
as follows. Suppose that ν is another measure on M that coincides with
µ0 on A, and supposeSthat F ∈ M has finite measure. We claim that
µ(F ) = ν(F ). If F ⊂ Ej , where Ej ∈ A, then
∞
X ∞
X
ν(F ) ≤ ν(Ej ) = µ0 (Ej ),
j=1 j=1
so
S that ν(F ) ≤ µ(F ). To prove the reverse inequality, note that if E =
Ej , then the fact that ν and µ are two measures that agree on A gives
n
[ n
[
ν(E) = lim ν( Ej ) = lim µ( Ej ) = µ(E).
n→∞ n→∞
j=1 j=1
If the sets Ej are chosen so that µ(E) ≤ µ(F ) + ², then the fact that
µ(F ) < ∞ implies µ(E − F ) ≤ ², and therefore
µ(F ) ≤ µ(E) = ν(E) = ν(F ) + ν(E − F ) ≤ ν(F ) + µ(E − F )

≤ µ(F ) + ².
Since ² is arbitrary, we find that µ(F ) ≤ ν(F ), as desired.

Finally, we use this last resultSto prove that if µ is σ-finite, then µ =
ν. Indeed, we may write X = Ej , where E1 , E2 , . . . is a countable
2. Integration on a measure space 273
collection of disjoint sets in A with µ(Ej ) < ∞. Then for any F ∈ M we

have
X X
µ(F ) = µ(F ∩ Ej ) = ν(F ∩ Ej ) = ν(F ),
and the uniqueness is proved.

For later use we record the following observation about the premeasure
µ0 on the algebra A and the resulting measure µ∗ that is implicit in the
argument given above. The details of the proof may be left to the reader.
We define Aσ as the collection of sets that are countable unions of sets
in A, and Aσδ as the sets that arise as countable intersections of sets in
Aσ .
Proposition 1.6 For any set E and any ² > 0, there are sets E1 ∈
Aσ and E2 ∈ Aσδ , such that E ⊂ E1 , E ⊂ E2 , and µ∗ (E1 ) ≤ µ∗ (E) + ²,
while µ∗ (E2 ) = µ∗ (E).
2 Integration on a measure space

Once we have established the basic properties of a measure space X, the
fundamental facts about measurable functions and integration of such
functions on X can be deduced as in the case of the Lebesgue measure
on Rd . Indeed, the results in Section 4 of Chapter 1 and all of Chapter 2
go over to the general case, with proofs remaining almost word-for-word
the same. For this reason we shall not repeat these arguments but limit
ourselves to the bare statement of the main points. The reader should
have no difficulty in filling in the missing details.
To avoid unnecessary complications we will assume throughout that
the measure space (X, M, µ) under consideration is σ-finite.
Measurable functions
A function f on X with values in the extended real numbers is measur-
able if
f −1 ([−∞, a)) = {x ∈ X : f (x) < a} ∈ M for all a ∈ R.
With this definition, the basic properties of measurable functions ob-

tained in the case of Rd with the Lebesgue measure continue to hold.
(See Properties 3 through 6 for measurable functions in Chapter 1.) For
instance, the collection of measurable functions is closed under the ba-
sic algebraic manipulations. Also, the pointwise limits of measurable
functions are measurable.
The notion of “almost everywhere” that we use now is with respect to

the measure µ. For instance, if f and g are measurable functions on X,
we write f = g a.e. to say that
µ ({x ∈ X : f (x) 6= g(x)}) = 0.
A simple function on X takes the form
N
X
ak χ Ek ,
k=1
where Ek are measurable sets of finite measure and ak are real numbers.
Approximations by simple functions played an important role in the defi-
nition of the Lebesgue integral. Fortunately, this result continues to hold
in our abstract setting.
• Suppose f is a non-negative measurable function on a measure

space (X, M, µ). Then there exists a sequence of simple functions
{ϕk }∞
k=1 that satisfies
ϕk (x) ≤ ϕk+1 (x) and lim ϕk (x) = f (x) for all x.

k→∞
In general, if f is only measurable, there exists a sequence of simple

functions {ϕk }∞k=1 that satisfies
|ϕk (x)| ≤ |ϕk+1 (x)| and lim ϕk (x) = f (x) for all x.
k→∞
The proof of this result can be obtained with some obvious minor
modifications of the proofs of Theorems 4.1 and 4.2 in Chapter 1. Here,
one makes use of the technical S condition imposed on X, that of being σ-
finite. Indeed, if we write X = Fk , where Fk ∈ M are of finite measure,
then the sets Fk play the role of the cubes Qk in the proof of Theorem 4.1,
Chapter 1.
Another important result that generalizes immediately is Egorov’s the-
orem.
• Suppose {fk }∞
k=1 is a sequence of measurable functions defined on
a measurable set E ⊂ X with µ(E) < ∞, and fk → f a.e. Then
for each ² > 0 there is a set A² with A² ⊂ E, µ(E − A² ) ≤ ², and
such that fk → f uniformly on A² .
2. Integration on a measure space 275
Definition and main properties of the integral

The four-step approach to the construction of the Lebesgue integral that
begins with its definition on simple functions given in Chapter 2 carries
over to the situation of a σ-finite measure space (X, M, µ). This leads
to the notion of the integral, with respect to the measure µ, of a non-
negative measurable function f on X. This integral is denoted by
Z
f (x) dµ(x),
X
R R R
which we sometimes simplify as X f dµ, f dµ or f , when no con-
fusion is possible. Finally, we say that a measurable function f is inte-
grable if
Z
|f (x)| dµ(x) < ∞.
X
The elementary properties of the integral, such as linearity and mono-

tonicity, continue to hold in this general setting, as well as the following
basic limit theorems.
(i) Fatou’s lemma. If {fn } is a sequence of non-negative measurable

functions on X, then
Z Z
lim inf fn dµ ≤ lim inf fn dµ.
n→∞ n→∞
(ii) Monotone convergence. If {fn } is a sequence of non-negative mea-

surable functions with fn % f , then
Z Z
lim fn = f.
n→∞
(iii) Dominated convergence. If {fn } is a sequence of measurable func-

tions with fn → f a.e., and such that |fn | ≤ g for some integrable
g, then
Z
|fn − f | dµ → 0 as n → ∞,
and consequently
Z Z
fn dµ → f dµ as n → ∞.
The spaces L1 (X, µ) and L2 (X, µ)

The equivalence classes (modulo functions that vanish almost every-
where) of integrable functions on (X, M, µ) form a vector space equipped
with a norm. This space is denoted by L1 (X, µ) and its norm is
Z
(4) kf kL1 (X,µ) = |f (x)| dµ(x).
X
2
Similarly we can define
R L (X,2µ) to be the equivalence class of measurable
functions for which X |f (x)| dµ(x) < ∞. The norm is then
µZ ¶1/2
2
(5) kf kL2 (X,µ) = |f (x)| dµ(x) .
X
There is also an inner product on this space given by

Z
(f, g) = f (x)g(x) dµ(x).
X
The proofs of Proposition 2.1 and Theorem 2.2 in Chapter 2, as well as

the results in Section 1 of Chapter 4, extend to this general case and
give:
• The space L1 (X, µ) is a complete normed vector space.

• The space L2 (X, µ) is a (possibly non-separable) Hilbert space.
3 Examples
We now discuss some useful examples of the general theory.
3.1 Product measures and a general Fubini theorem

Our first example concerns the construction of product measures, and
leads to a general form of the theorem that expresses a multiple integral
as a repeated integral, extending the case of Euclidean space considered
in Section 3 of Chapter 2.
Suppose (X1 , M1 , µ1 ) and (X2 , M2 , µ2 ) are a pair of measure spaces.
We want to describe the product measure µ1 × µ2 on the space X =
X1 × X2 = {(x1 , x2 ) : x1 ∈ X1 , x2 ∈ X2 }.
We will assume here that the two measure spaces are each complete
and σ-finite.
We begin by considering measurable rectangles: these are subsets
of X of the form A × B, with A and B measurable sets, that is, A ∈ M1
3. Examples 277
and B ∈ M2 . We then let A denote the collection of all sets in X that are
finite unions of disjoint measurable rectangles. It is easy to check that A
is an algebra of subsets of X. (Indeed, the complement of a measurable
rectangle is the union of three disjoint such rectangles, while the union
of two measurable rectangles is the disjoint union of at most six such
rectangles.) From now on we abbreviate our terminology by referring to
measurable rectangles simply as “rectangles.”
On the rectangles we define the function µ0 by µ0 (A × B) = µ1 (A)µ2 (B).
Now the fact that µ0 has a unique extension to the algebra A for which
µ0 becomes a premeasure is a consequence of the following fact: when-
ever a rectangle A × B is the disjoint
S∞ union of a countable collection of
rectangles {Aj × Bj }, A × B = j=1 Aj × Bj , then
∞
X
(6) µ0 (A × B) = µ0 (Aj × Bj ).
j=1
To prove this, observe that if x1 ∈ A, then for each x2 ∈ B the point

(x1 , x2 ) belongs to exactly one Aj × Bj . Therefore we see that B is the
disjoint union of the Bj for which x1 ∈ Aj . By the countable additivity
property of the measure µ2 this has as an immediate consequence the
fact that
∞
X
χA (x1 )µ2 (B) = χAj (x1 )µ2 (Bj ).
j=1
Hence integrating in
Px∞1 and using the monotone convergence theorem we
get µ1 (A)µ2 (B) = j=1 µ1 (Aj )µ2 (Bj ), which is (6).
Now that we know that µ0 is a premeasure on A, we obtain from The-
orem 1.5 a measure (which we denote by µ = µ1 × µ2 ) on the σ-algebra
M of sets generated by the algebra A of measurable rectangles. In this
way, we have defined the product measure space (X1 × X2 , M, µ1 × µ2 ).
Given a set E in M we shall now consider slices
Ex1 = {x2 ∈ X2 : (x1 , x2 ) ∈ E} and E x2 = {x1 ∈ X1 : (x1 , x2 ) ∈ E}.
We recall the definitions according to which Aσ denotes the collection

of sets that are countable unions of elements of A, and Aσδ the sets
that arise as countable intersections of sets from Aσ . We then have the
following key fact.
Proposition 3.1 If E belongs to Aσδ , then E x2 is µ1 -measurable for

every x2 ; moreover, µ1 (E x2 ) is a µ2 -measurable function. In addition
Z
(7) µ1 (E x2 ) dµ2 = (µ1 × µ2 )(E).
X2
Proof. One notes first that all the assertions hold immediately when
E is a (measurable) rectangle. Next suppose E is a set in Aσ . Then we
can decompose it as a countable union of disjoint rectangles S Ej . (If the
Ej are not already disjoint we only need to replace the Ej by k≤j Ek −
S S∞
k≤j−1 Ek .) Then for each x2 we have E
x2
= j=1 Ejx2 , and we observe
that {Ejx2 } are disjoint sets. Thus by (7) applied to each rectangle Ej
and the monotone convergence theorem we get our conclusion for each
set E ∈ Aσ .
Next assume E ∈ Aσδ and that (µ1 × µ2 )(E) < ∞. ThenTthere is
∞
a sequence {Ej } of sets with Ej ∈ Aσ , Ej+1 ⊂ Ej , and E = j=1 Ej .
x2
We let fj (x2 ) = µ1 (Ej ) and f (x2 ) = µ1 (E x2 ). To see that E x2 is µ1 -
measurable and f (x2 ) is well-defined, note that E x2 is the decreasing
limit of the sets Ejx2 , which we have seen by the above are measur-
able. Moreover, since E1 ∈ Aσ and (µ1 × µ2 )(E1 ) < ∞, we see that
fj (x2 ) → f (x2 ), as j → ∞ for each x2 . Thus f (x2 ) is measurable. How-
ever, {fj (x2 )} is a decreasing sequence of non-negative functions, hence
Z Z
f (x2 ) dµ2 (x) = lim fj (x2 ) dµ2 (x),
X2 j→∞ X2
and therefore (7) is proved in the case when (µ1 × µ2 )(E) < ∞. Now
since we assumed both µ1 and µ2 are σ-finite, we can find sequences F1 ⊂
F
S2∞⊂ · · · ⊂ Fj ⊂S· ∞
· · ⊂ X1 and G1 ⊂ G2 ⊂ · · · ⊂ Gj ⊂ · · · ⊂ X2 , with
F
j=1 j = X1 , j=1 Gj = X2 , µ1 (Fj ) < ∞, and µ2 (Gj ) < ∞ for all j.
Then we merely need to replace E by Ej = E ∩ (Fj × Gj ), and let j → ∞
to obtain the general result.
We now extend the result in the above proposition to an arbitrary
measurable set E in X1 × X2 , that is, E ∈ M, the σ-algebra generated
by the measurable rectangles.
Proposition 3.2 If E is an arbitrary measurable set in X, then the

conclusion of Proposition 3.1 are still valid except that we only assert that
E x2 is µ1 -measurable and µ1 (E x2 ) is defined for almost every x2 ∈ X2 .
Proof. Consider first the case when E is a set of measure zero.

Then we know by Proposition 1.6 that there is a set F ∈ Aσδ such that
3. Examples 279
E ⊂ F and (µ1 × µ2 )(F ) = 0. Since E x2 ⊂ F x2 for every x2 and F x2 has

µ1 -measure zero for almost every x2 by (7) applied to F , the assumed
completeness of the measure µ2 shows that E x2 is measurable and has
measure zero for those x2 . Thus the desired conclusion holds when E
has measure zero.
If we drop this assumption on E, we can invoke Proposition 1.6 again
to find an F ∈ Aσδ , F ⊃ E, such that F − E = Z has measure zero.
Since F x2 − E x2 = Z x2 we can apply the case we have just proved, and
find that for almost all x2 the set E x2 is measurable and µ1 (E x2 ) =
µ1 (F x2 ) − µ1 (Z x2 ). From this the proposition follows.
We now obtain the main result, generalizing Fubini’s theorem in Chap-
ter 2.
Theorem 3.3 In the setting above, suppose f (x1 , x2 ) is an integrable
function on (X1 × X2 , µ1 × µ2 ).
(i) For almost every x2 ∈ X2 , the slice f x2 (x1 ) = f (x1 , x2 ) is inte-
grable on (X1 , µ1 ).
R
(ii) X1 f (x1 , x2 ) dµ1 is an integrable function on X2 .
R ³R ´ R
(iii) X2 X1 f (x1 , x2 ) dµ1 dµ2 = X1 ×X2 f dµ1 × µ2 .
Proof. Note that if the desired conclusions hold for finitely many
functions, they also hold for their linear combinations. In particular it
suffices to assume that f is non-negative. When f = χE , where E is a set
of finite measure, what we wish to prove is contained in Proposition 3.2.
Hence the desired result also holds for simple functions. Therefore by
the monotone convergence theorem it is established for all non-negative
functions, and the theorem is proved.
We remark that in general the product space (X, M, µ) constructed
above is not complete. However, if we define the completed space (X, M, µ)
as in Exercise 2, the theorem continues to hold in this completed space.
The proof requires only a simple modification of the argument in Propo-
sition 3.2.
3.2 Integration formula for polar coordinates

The polar coordinates of a point x ∈ Rd − {0} are the pair (r, γ), where
0 < r < ∞ and γ belongs to the unit sphere S d−1 = {x ∈ Rd , |x| = 1}.
These are determined by
x
(8) r = |x|, γ = , and reciprocally by x = rγ.
|x|
Our intention here is to deal with the formula that, with appropriate
definitions and under suitable hypotheses, states:
Z Z µZ ∞ ¶
d−1
(9) f (x) dx = f (rγ)r dr dσ(γ).
Rd S d−1 0
For this we consider the following pair of measure spaces. First,

(X1 , M1 , µ1 ), where X1 = (0, ∞), M1 is the collection of Lebesgue mea-
d−1
surable
R d−1 sets in (0, ∞), and dµ1 (r) = r drd−1 in the sense that µ1 (E) =
E
r dr. Next, X2 is the unit sphere S , and the measure µ2 is
the one in effect determined by (9) with µ2 = σ. Indeed given any set
E ⊂ S d−1 we let Ẽ = {x ∈ Rd : x/|x| ∈ E, 0 < |x| < 1} be the “sector”
in the unit ball whose “end-points” are in E. We shall say E ∈ M2
exactly when Ẽ is a Lebesgue measurable subset of Rd , and define
µ2 (E) = σ(E) = d · m(Ẽ), where m is Lebesgue measure in Rd .
With this it is clear that both (X1 , M1 , µ1 ) and (X2 , M2 , µ2 ) satisfy
all the properties of complete and σ-finite measure spaces. We note also
that the sphere S d−1 has a metric on it given by d(γ, γ 0 ) = |γ − γ 0 |, for
γ, γ 0 ∈ S d−1 . If E is an open set (with respect to this metric) in S d−1 ,
then Ẽ is open in Rd , and hence E is a measurable set in S d−1 .
Theorem 3.4 Suppose f is an integrable function on Rd . Then for al-

most every γ ∈ S d−1 the slice f γ defined by f γ (r) = f (rγ)
R∞is an integrable
function with respect to the measure rd−1 dr. Moreover, 0 f γ (r)rd−1 dr
is integrable on S d−1 and the identity (9) holds.
There is a corresponding result with the order of integration of r and

γ reversed.
Proof. We consider the product measure µ = µ1 × µ2 on X1 × X2
given by Theorem 3.3. Since the space X1 × X2 = {(r, γ) : 0 < r <
∞ and γ ∈ S d−1 } can be identified with Rd − {0}, we can think of µ
as a measure of the latter space, and our main task is to identify it with
the (restriction of) Lebesgue measure on that space. We claim first that
(10) m(E) = µ(E)
whenever E is a measurable rectangle E = E1 × E2 , and in this case

µ(E) = µ1 (E1 )µ2 (E2 ). In fact this holds for E2 an arbitrary measurable
subset of S d−1 and E1 = (0, 1), because then E = E1 × E2 is the sector
E˜2 , while µ1 (E1 ) = 1/d.
Because of the relative dilation-invariance of Lebesgue measure, (10)
also holds when E = (0, b) × E2 , b > 0. A simple limiting argument then
proves the result for sets E1 = (0, a], and by subtraction to all open
3. Examples 281
intervals E1 = (a, b), and thus for all open sets. Thus we have m(E1 ×
E2 ) = µ1 (E1 )µ2 (E2 ) for all open sets E1 , and hence for all closed sets,
and therefore for all Lebesgue measurable sets. (In fact, we can find
sets F1 ⊂ E1 ⊂ O1 with F1 closed and O1 open, such that m1 (O1 ) − ² ≤
m1 (E1 ) ≤ m1 (F1 ) + ², and apply the above to F1 × E2 and O1 × E2 .)
So we have established the identity (10) for all measurable rectangles
and as a result for all finite unions of measurable rectangles. This is
the algebra A that occurs in the proof of Theorem 3.3, and hence by
the uniqueness in Theorem 1.5, the identity extends to the σ-algebra
generated by A, which is the σ-algebra M on which the measure µ is
defined. To summarize, whenever E ∈ M, the assertion (9) holds for
f = χE .
d
To go further we note that any openS∞ set in R − {0} can be written
as a countable union of rectangles j=1 Aj × Bj , where Aj and Bj are
open in (0, ∞) and S d−1 , respectively. (This small technical point is
taken up in Exercise 12.) It follows that any open set is in M, and
therefore so is any Borel set. Thus (9) is valid for χE whenever E is
any Borel set in Rd − {0}. The result then goes over to any Lebesgue
set E 0 ⊂ Rd − {0}, since such a set can be written as a disjoint union
E 0 = E ∪ Z, where E is a Borel set and Z ⊂ F , with F a Borel set
of measure zero. To finish the proof we follow the familiar steps of
deducing (9) for simple functions, and then by monotonic convergence
for non-negative integrable functions, and from that for the general case.
3.3 Borel measures on R and the Lebesgue-Stieltjes integral

The Stieltjes integral was introduced to provide a generalization of the
Rb
Riemann integral a f (x) dx, where the increments dx were replaced by
the increments dF (x) for a given increasing function F on [a, b]. We wish
to pursue this idea from the general point of view taken in this chapter.
The question that is then raised is that of characterizing the measures
on R that arise in this way, and in particular measures defined on the
Borel sets on the real line.
To have a unique correspondence between measures and increasing
functions as we shall have below, we need first to normalize these func-
tions appropriately. Recall that an increasing function F can have at
most a countable number of discontinuities. If x0 is such a discontinuity,
then
lim F (x) = F (x−

0) and lim F (x) = F (x+
0)
x < x0 x > x0
x → x0 x → x0
both exist, while F (x− + −

0 ) < F (x0 ) and F (x0 ) is some value between F (x0 )
+
and F (x0 ). We shall now modify F at x0 , if necessary, by setting
F (x0 ) = F (x+
0 ), and we do this for every point of discontinuity. The
function F so obtained is now still increasing, yet right-continuous at ev-
ery point, and we say such functions are normalized. The main result
is then as follows.
Theorem 3.5 Let F be an increasing function on R that is normalized.

Then there is a unique measure µ (also denoted by dF ) on the Borel
sets B on R such that µ((a, b]) = F (b) − F (a) if a < b. Conversely, if
µ is a measure on B that is finite on bounded intervals, then F defined
by F (x) = µ((0, x]), x > 0, F (0) = 0 and, F (x) = −µ((−x, 0]), x < 0, is
increasing and normalized.
Before we come to the proof, we remark that the condition that µ be

finite on bounded intervals is crucial. In fact, the Hausdorff measures
that will be considered in the next chapter provide examples of Borel
measures on R of a very different character from those treated in the
theorem.
Proof. We define a function µ∗ on all subsets of R by
∞
X
µ∗ (E) = inf (F (bj ) − F (aj )),
j=1
S∞
where the infimum is taken over all coverings of E of the form j=1 (aj , bj ].
It is easy to verify that µ∗ is an exterior measure on R. We observe
next that µ∗ ((a, b]) = (F (b) − F (a)), if a < b. Clearly µ∗ ((a, b])
S∞≤ F (b) −
F (a), since (a, b], then covers itself. Next, suppose that j=1 (aj , bj ]
covers (a, b]; then it covers [a0 , b] for any a < a0 < b. However, by the
right-continuity of F , if ² > 0 is given, we can always choose S b0j > bj such
∞
that F (bj ) ≤ F (bj ) + ²/2 . Now the union of open intervals j=1 (aj , b0j )
0 j
SN
covers [a0 , b]. By the compactness of this interval, j=1 (aj , b0j ) covers
[a0 , b] for some N . Thus since F is increasing we have
N
X N
X
F (b) − F (a0 ) ≤ F (b0j ) − F (aj ) ≤ (F (bj ) − F (aj ) + ²/2j )
j=1 j=1
≤ µ∗ ((a, b]) + ².
Thus letting a0 → a, and using the right-continuity of F again, we see

that F (b) − F (a) ≤ µ∗ ((a, b]) + ². Since ² was arbitrary this then proves
F (b) − F (a) = µ∗ ((a, b]).
3. Examples 283
Next we show that µ∗ is a metric exterior measure (for the usual

metric d(x, x0 ) = |x − x0 | on the real line). Since µ∗ is an exterior measure
we have µ∗ (E1 ∪ E2 ) ≤ µ∗ (E1 ) + µ∗ (E2 ); thus it suffices to see that the
reverse inequality holds whenever d(E1 , E2 ) ≥ δ, for some S∞δ > 0.
Suppose that we are given a positive ², and that j=1 (aj , bj ] is a
covering of E1 ∪ E2 such that
∞
X
F (bj ) − F (aj ) ≤ µ∗ (E1 ∪ E2 ) + ².
j=1
We may assume, after subdividing the intervals (aj , bj ] into smaller half-
open intervals, that each interval in the covering has length less than δ.
When this is so each interval can intersect at most one of the two sets E1
or E2 . If we denote by J1 and J2 the sets of those indices for which (aj , bj ]
intersects ES1 and E2 , respectively, then S
J1 ∩ J2 is empty; moreover, we
have E1 ⊂ j∈J1 (aj , bj ] as well as E2 ⊂ j∈J2 (aj , bj ]. Therefore
X X
µ∗ (E1 ) + µ∗ (E2 ) ≤ F (bj ) − F (aj ) + F (bj ) − F (aj )
j∈J1 j∈J2
∞
X
≤ F (bj ) − F (aj ) ≤ µ∗ (E1 ∪ E2 ) + ².
j=1
Since ² was arbitrary, we see that µ∗ (E1 ) + µ∗ (E2 ) ≤ µ∗ (E1 ∪ E2 ), as we

intended to show.
We can now invoke Theorem 1.5. This guarantees the existence of a
measure µ for which the Borel sets are measurable; moreover, we have
µ((a, b]) = F (b) − F (a), since clearly (a, b]) is a Borel set and we have
previously seen that µ∗ ((a, b]) = F (b) − F (a).
To prove that µ is the unique Borel measure on R for which µ((a, b]) =
F (b) − F (a), let us suppose that ν is another Borel measure with this
property. It now suffices to show that ν = µ on all Borel sets. S∞
We can write any open interval as a disjoint union (a, b) = j=1 (aj , bj ],
by choosing {bj }∞ j=1 to be a strictly increasing sequence with a < bj < b,
bj → b as j → ∞, and taking a1 = a, aj+1 = bj . Since ν and µ agree on
each (aj , bj ], it follows that ν and µ agree on (a, b), and hence on all
open intervals, and therefore on all open sets. Moreover, clearly ν and µ
are finite on all bounded intervals; thus the regularity in Proposition 1.3
allows one to conclude that µ = ν on all Borel sets.
Conversely, if we start with a Borel measure µ on R that is finite on
bounded intervals, we can define the function F as in the statement of the
theorem. Then clearly F is increasing. To see that it is right-continuous,
note that if, for instance, x0 > 0, the sets En = (0, x0 + 1/n] decrease
to E = (0, x0 ] as n → ∞, hence µ(En ) → µ(E), since µ(E1 ) < ∞. This
means that F (x0 + 1/n) → F (x0 ). Since F is increasing, this implies
that F is right-continuous at x0 . The argument for any x0 ≤ 0 is similar,
and thus the theorem is proved.
Remarks. Several comments about the theorem are in order.
(i) Two increasing functions F and G give the same measure if F −

G is constant. The converse if also true because F (b) − F (a) =
G(b) − G(a) for all a < b exactly when F − G is constant.
(ii) The measure µ constructed in the proof of the theorem is defined
on a larger σ-algebra than the Borel sets, and is actually complete.
However, in applications, its restriction to the Borel sets often suf-
fices.
(iii) If F is an increasing normalized function given on a closed interval
[a, b], we can extend it to R by setting F (x) = F (a) for x < a, and
F (x) = F (b) for x > b. For the resulting measure µ, the intervals
(−∞, a] and (b, ∞) have measure zero. One then often writes
Z Z b
f (x) dµ(x) = f (x) dF (x),
R a
for every f that is integrable with respect to µ. If F arises from an

increasing function F0 defined on R, one may wish to account for
the possible jump of F0 at a. In this case it is sometimes useful to
define
Z b Z b
f (x) dF (x) as f (x) dµ0 (x),
a− a
where µ0 is the measure on R corresponding to F0 .

(iv) Note that the above definition of the Lebesgue-Stieltjes integral
extends to the case when F is of bounded variation. Indeed
P4suppose
F is a complex-valued function on [a, b] such that F = j=1 ²j Fj ,
where each Fj is increasing and normalized, and ²j are ±1 or ±i.
Rb P4 Rb
Then we can define a f (x) dF (x) as j=1 ²j a f (x) dFj (x); here
we require
P4 that f be integrable with respect to the Borel measure
µ = j=1 µj , where µj is the measure corresponding to Fj .
(v) The value of these integrals can be calculated more directly in the
following cases.
4. Absolute continuity of measures 285
(a) If F is an absolutely continuous function on [a, b], then

Z b Z b
f (x) dF (x) = f (x)F 0 (x) dx
a a
for every Borel measurable function f that is integrable with

respect to µ = dF .
(b) Suppose F is a pure jump function as in Section 3.3, Chap-
ter 3, with jumps {αn }∞ ∞
n=1 at the points {xn }n=1 . Then when-
ever f is, say, continuous and vanishes outside some finite
interval we have
Z b X∞
f (x) dF (x) = f (xn )αn .
a n=1
In particular, for the measure µ we have µ({xn }) = αn and

µ(E) = 0 for all sets that do not contain any of the xn .
(c) A special instance arises when F = H, the Heaviside function
defined by H(x) = 1 for x ≥ 0, and H(x) = 0 for x < 0. Then
Z ∞
f (x) dH(x) = f (0),
−∞
which is another expression for the Dirac delta function arising

in Section 2 of Chapter 3.
Further details about (v) can be found in Exercise 11.
4 Absolute continuity of measures

The generalization of the notion of absolute continuity considered in
Chapter 3 requires that we extend the ideas of a measure to encompass
set functions that may be positive or negative. We describe this notion
first.
4.1 Signed measures

Loosely speaking, a signed measure possesses all the properties of a mea-
sure, except that it may take positive or negative values. More precisely,
a signed measure ν on a σ-algebra M is a mapping that satisfies:
(i) The set function ν is extended-valued in the sense that −∞ <

ν(E) ≤ ∞ for all E ∈ M.
(ii) If {Ej }∞
j=1 are disjoint subsets of M, then
Ã∞ ! ∞
[ X
ν Ej = ν(Ej ).
j=1 j=1
P
Note that for this to hold the sum ν(Ej ) must be independent of
S∞
the rearrangements of terms, so that if ν( j=1 Ej ) is finite, it implies
that the sum converges absolutely.
Examples of signed measures arise naturally if we drop the assumption
that f be non-negative in the expression
Z
ν(E) = f dµ,
E
where (X, M, µ) is a measure space and f is µ-measurable. In fact,

to ensure that ν satisfies (i) and (ii) the function f is required
R − to be
“integrable” with
R respect to µ in the extended sense that f dµ must
be finite, while f + dµ may be infinite.
Given a signed measure ν on (X, M) it is always possible to find a
(positive) measure µ that dominates ν, in the sense that
ν(E) ≤ µ(E) for all E,
and that in addition is the “smallest” µ that has this property.

The construction is in effect an abstract version of the decomposition
of a function of bounded variation as the difference of two increasing
functions, as carried out in Chapter 3. We proceed as follows. We define
a function |ν| on M, called the total variation of ν, by
∞
X
|ν|(E) = sup |ν(Ej )|,
j=1
where the supremum S is taken over all partitions of E, that is, over all
∞
countable unions E = j=1 Ej , where the sets Ej are disjoint and belong
to M.
The fact that |ν| is actually additive is not obvious, and is given in the
proof below.
Proposition 4.1 The total variation |ν| of a signed measure ν is itself

a (positive) measure that satisfies ν ≤ |ν|.
∞
Proof. SupposeS {Ej }j=1 is a countable collection of disjoints sets in
M, and let E = Ej . It suffices to prove:
X X
(11) |ν|(Ej ) ≤ |ν|(E) and |ν|(E) ≤ |ν|(Ej ).
Let αj be a real number that

S satisfies αj < |ν|(Ej ). By definition, each
Ej can be written as Ej = i Fi,j , where the Fi,j are disjoint, belong to
M, and
∞
X
αj ≤ |ν(Fi,j )|.
i=1
S
Since E = i,j Fi,j , we have
X X
αj ≤ |ν(Fi,j )| ≤ |ν|(E).
j,i
Consequently, taking the supremum over the numbers αj gives the first
inequality in (11).
For the reverse inequality, let Fk be any other partition of E. For a
fixed k, {Fk ∩ Ej }j is a partition of Fk , so
¯ ¯
X X ¯¯X ¯
¯
|ν(Fk )| = ¯ ν(Fk ∩ Ej )¯ ,
¯ ¯
k k j
since ν is a signed measure. An application of the triangle inequality and

the fact that {Fk ∩ Ej }k is a partition of Ej gives
X XX
|ν(Fk )| ≤ |ν(Fk ∩ Ej )|
k k j
XX
= |ν(Fk ∩ Ej )|
j k
X
≤ |ν|(Ej ).
j
Since {Fk } was an arbitrary partition of E, we obtain the second in-

equality in (11) and the proof is complete.
It is now possible to write ν as the difference of two (positive) measures.
To see this, we define the positive variation and negative variation
of ν by
1 1
ν+ = (|ν| + ν) and ν− = (|ν| − ν).
2 2
By the proposition we see that ν + and ν − are measures, and they clearly
satisfy
ν = ν+ − ν− and |ν| = ν + + ν − .
In the above if ν(E) = ∞ for a set E, then |ν|(E) = ∞, and ν − (E) is

defined to be zero.
We also make the following definition: we say that the signed measure
ν is σ-finite if the measure |ν| is σ-finite. Since ν ≤ |ν| and | − ν| = |ν|,
we find that
−|ν| ≤ ν ≤ |ν|.
As a result, if ν is σ-finite, then so are ν + and ν − .
4.2 Absolute continuity

Given two measures defined on a common σ-algebra we describe here the
relationships that can exist between them. More concretely, consider two
measures ν and µ defined on the σ-algebra M; two extreme scenarios
are
(a) ν and µ are “supported” on separate parts of M.

(b) The support of ν is an essential part of the support of µ.
Here we adopt the terminology that the measure ν is supported on a

set A, if ν(E) = ν(E ∩ A) for all E ∈ M.
The Lebesgue-Radon-Nikodym theorem below states that in a precise
sense the relationship between any two measures ν and µ is a combination
of the above two possibilities.
Mutually singular and absolutely continuous measures

Two signed measures ν and µ on (X, M) are mutually singular if there
are disjoint subsets A and B in M so that
ν(E) = ν(A ∩ E) and µ(E) = µ(B ∩ E) for all E ∈ M.
Thus ν and µ are supported on disjoint subsets. We use the symbol

ν ⊥ µ to denote the fact that the measures are mutually singular.
In contrast, if ν is a signed measure and µ a (positive) measure on M,
we say that ν is absolutely continuous with respect to µ if
(12) ν(E) = 0 whenever E ∈ M and µ(E) = 0.

Thus if ν is supported in a set A, then A must be an essential part of the

support of µ in the sense that µ(A) > 0. We use the symbol ν ¿ µ to
indicate that ν is absolutely continuous with respect to µ. Note that if
ν and µ are mutually singular, and ν is also absolutely continuous with
respect to µ, then ν vanishes identically.
An important example is given by integration with respect to µ. In-
deed, if Rf ∈ L1 (X, µ), or if f is merely
R + integrable in the extended sense
−
(where f < ∞, but possibly f = ∞), then the signed measure ν
defined by
Z
(13) ν(E) = f dµ
E
is absolutely continuous with respect to µ. We shall use the shorthand

dν = f dµ to indicate that ν is defined by (13).
This is a variant of the notion of absolute continuity that arose in
Chapter 3 in the special case of R (with M the Lebesgue measurable
sets and dµ = dx the Lebesgue measure). In fact, with ν defined by (13)
and f an integrable function, we saw that in place of (12) we had the
following stronger assertion:
(14)
For each ² > 0, there is a δ > 0 such that µ(E) < δ implies |ν(E)| < ².
In the general situation the relation between the two conditions (12)
and (14) is clarified by the following observation.
Proposition 4.2 The assertion (14) implies (12). Conversely, if |ν| is
a finite measure, then (12) implies (14).
That (12) is a consequence of (14) is obvious because µ(E) = 0 gives
|ν(E)| < ² for every ² > 0. To prove the converse, it suffices to consider
the case when ν is positive, upon replacing ν by |ν|. We then assume
that (14) does not hold. This means that it fails for some fixed ² >
−n
0. Hence for each n, there is a measurable set E Tn∞with∗ µ(En ) < 2∗
∗
while ν(En ) ≥ ². Now let E =P
S lim supn→∞ En = n=1 En , where En =
∗ k n−1
k≥n Ek . Then since µ(En ) ≤ k≥n 1/2 = 1/2 , and the decreasing
sets {Ek } are contained in a set of finite measure (E1∗ ), we get µ(E ∗ ) = 0.
∗
However ν(En∗ ) ≥ ν(En ) ≥ ², and the ν measure is assumed finite. So

ν(E ∗ ) = limn→∞ ν(En∗ ) ≥ ², which gives a contradiction.
After these preliminaries we can come to the main result. It guarantees
among other things a converse to the representation (13); it was proved
in the case of R by Lebesgue, and in the general case by Radon and
Nikodym.
Theorem 4.3 Suppose µ is a σ-finite positive measure on the measure

space (X, M) and ν a σ-finite signed measure on M. Then there exist
unique signed measures νa and νs on M such that νa ¿ µ, νs ⊥ µ and
ν = νa + νs . In addition, the measure νa takes the form dνa = f dµ; that
is,
Z
νa (E) = f (x) dµ(x)
E
for some extended µ-integrable function f .

Note the following consequence. If ν is absolutely continuous with respect
to µ, then dν = f dµ, and this assertion can be viewed as a generalization
of Theorem 3.11 in Chapter 3.
There are several known proofs of the above theorem. The argument
given below, due to von Neumann, has the virtue that it exploits elegantly
the application of a simple Hilbert space idea.
We start with the case when both ν and µ are positive and finite. Let
ρ = ν + µ, and consider the transformation on L2 (X, ρ) defined by
Z
`(ψ) = ψ(x) dν(x).
X
The mapping ` defines a bounded linear functional on L2 (X, ρ) since

Z Z
|`(ψ)| ≤ |ψ(x)| dν(x) ≤ |ψ(x)| dρ(x)
X X
µZ ¶1/2
1/2 2
≤ (ρ(X)) |ψ(x)| dρ(x) ,
X
where the last inequality follows by the Cauchy-Schwarz inequality. But

L2 (X, ρ) is a Hilbert space, so the Riesz representation theorem (in Chap-
ter 4) guarantees the existence of g ∈ L2 (X, ρ) such that
Z Z
(15) ψ(x) dν(x) = ψ(x)g(x) dρ(x) for all ψ ∈ L2 (X, ρ).
X X
If E ∈ M with ρ(E) > 0, when we set ψ = χE in (15) and recall that

ν ≤ ρ, we find
Z
1
0≤ g(x) dρ(x) ≤ 1,
ρ(E) E
from which we conclude that

R 0 ≤ g(x) ≤ 1 for a.e. x (with respect to the
measure ρ). In fact, 0 ≤ E g(x) dρ(x) for all sets E ∈ M implies that
R
g(x) ≥ 0 almost everywhere. In the same way, 0 ≤ E (1 − g(x)) dρ(x)
for all E ∈ M guarantees that g(x) ≤ 1 almost everywhere. Therefore
we may clearly assume 0 ≤ g(x) ≤ 1 for all x without disturbing the
identity (15), which we rewrite as
Z Z
(16) ψ(1 − g) dν = ψg dµ.
Consider now the two sets
A = {x ∈ X : 0 ≤ g(x) < 1} and B = {x ∈ X : g(x) = 1},
and define two measures νa and νs on M by
νa (E) = ν(A ∩ E) and νs (E) = ν(B ∩ E).
To see why νs ⊥ µ, it suffices to note that setting ψ = χB in (16) gives

Z
0 = χB dµ = µ(B).
Finally, we set ψ = χE (1 + g + · · · + g n ) in (16) :

Z Z
n+1
(17) (1 − g ) dν = g(1 + · · · + g n ) dµ.
E E
Since (1 − g n+1 )(x) = 0 if x ∈ B, and (1 − g n+1 )(x) → 1 if x ∈ A, the

dominated convergence theorem implies that the left-hand side of (17)
1
converges to ν(A ∩ E) = νa (E). Also, 1 + g + · · · + g n converges to 1−g ,
so we find in the limit that
Z
g
νa (E) = f dµ, where f = 1−g .
E
1
Note that f ∈ L (X, µ), since νa (X) ≤ ν(X) < ∞. If µ and ν are
S σ-finite
and positive we may clearly find sets Ej ∈ M such that X = Ej and
µ(Ej ) < ∞, ν(Ej ) < ∞ for all j.
We may define positive and finite measures on M by
µj (E) = µ(E ∩ Ej ) and νj (E) = ν(E ∩ Ej ),
and then we can write for each j, νj = νj,a + νj,s where νj,s ⊥ µj and
νj,a = fj dµj . Then it suffices to set
X X X
f= fj , νs = νj,s , and νa = νj,a .
Finally, if ν is signed, then we apply the argument separately to the

positive and negative variations of ν.
To prove the uniqueness of the decomposition, suppose we also have
ν = νa0 + νs0 , where νa0 ¿ µ and νs0 ⊥ µ. Then
νa − νa0 = νs0 − νs .
The left-hand side is absolutely continuous with respect to µ, and the

right-hand side is singular with respect to µ. Thus both sides are zero
and the theorem is proved.
5* Ergodic theorems
Ergodic theory had its beginnings in certain problems in statistical me-
chanics studied in the late nineteenth century. Since then it has grown
rapidly and has gained wide influence in a number of mathematical disci-
plines, in particular those related to dynamical systems and probability
theory. It is not our purpose to try to give an account of this broad
and fascinating theory. Rather, we restrict our presentation to some of
the basic limit theorems that lie at its foundation. These theorems are
most naturally formulated in the general context of abstract measure
spaces, and thus for us they serve as excellent illustrations of the general
framework developed in this chapter.
The setting for the theory is a σ-finite measure space (X, M, µ) en-
dowed with a mapping τ : X → X such that whenever E is a measurable
subset of X, then so is τ −1 (E), and µ(τ −1 (E)) = µ(E). Here τ −1 (E) is
the pre-image of E under τ ; that is, τ −1 (E) = {x ∈ X : τ (x) ∈ E}. A
mapping τ with these properties is called a measure-preserving trans-
formation. If in addition for such a τ we have the feature that it is a
bijection and τ −1 is also a measure-preserving transformation, then τ is
referred to as a measure-preserving isomorphism.
Let us note that if τ is a measure-preserving transformation, then
f (τ (x)) is measurable if f (x) is measurable, and is integrable if f is
integrable; moreover, then
Z Z
(18) f (τ (x)) dµ(x) = f (x) dµ(x).
X X
Indeed, if χE is the characteristic function of the set E, we note that

χE (τ (x)) = χτ −1 (E) (x), and so the assertion holds for characteristic func-
tions of measurable sets and thus for simple functions, and hence by the
usual limiting arguments for all non-negative measurable functions, and
5*. Ergodic theorems 293
then integrable functions. For later purposes we record here an equiva-

lent statement: whenever f is a real-valued measurable function and α
is any real number, then
µ({x : f (x) > α}) = µ({x : f (τ (x)) > α}).
Before we proceed further, we describe several examples of measure-

preserving transformations:
(i) Here X = Z, the integers, with µ its counting measure; that is,
µ(E) = #(E) = the number of integers in E, for any E ⊂ Z. We
define τ to be the unit translation, τ : n 7→ n + 1. Note that τ gives
a measure-preserving isomorphism of Z.
(ii) Another easy example is X = Rd with Lebesgue measure, and τ a

translation, τ : x 7→ x + h for some fixed h ∈ Rd . This is of course
a measure-preserving isomorphism. (See the section on invariance
properties of the Lebesgue measure in Chapter 1.)
(iii) Here X is the unit circle, given as R/Z, with the measure induced
from Lebesgue measure on R. That is, we may realize X as the unit
interval (0, 1], and take µ to be the Lebesgue measure restricted
to this interval. For any real number α, the translation x 7→ x +
α, taken modulo Z, is well defined on X = R/Z, and is measure-
preserving. (See the related Exercise 3 in Chapter 2.) It can be
interpreted as a rotation of the circle by angle 2πα.
(iv) In this example X is again (0, 1] with Lebesgue measure µ, but τ

is the doubling map τ (x) = 2x mod 1. It is easy to verify that
τ is a measure-preserving transformation. Indeed, any set E ⊂
(0, 1] has two pre-images E1 and E2 , the first in (0, 1/2] and the
second in (1/2, 1], both of measure µ(E)/2, if E is measurable.
(See Figure 1.) However, τ is not an isomorphism, since τ is not
injective.
(v) A trickier example is given by the transformation that is key in

the theory of continued fractions. Here X = [0, 1) and τ is defined
by τ (x) = h1/xi, the fractional part of 1/x; when x = 0 we set
1
τ (0) = 0. Gauss observed, in effect, that the measure dµ = 1+x dx
is preserved by the transformation τ . Note that each x ∈ (0, 1) has
infinitely many pre-images under τ ; that is, the sequence {1/(x +
k)}∞k=1 . More about this example can be found in Problems 8
through 10 below.
E1 E2 E
0 1/2 1 0 1/2 1
Figure 1. Pre-images E1 and E2 under the doubling map
Having pointed out these examples, we can now return to the general
theory. The notions described above are of interest, in part, because they
abstract the idea of a dynamical system, one whose totality of states is
represented by the space X, with each point x ∈ X giving a particular
state of the system. The mapping τ : X → X then describes the trans-
formation of the system after a unit of time has elapsed. For such a
system there is often associated a notion of “volume” or “mass” that is
unchanged by the evolution, and this is the role of the invariant measure
µ. The iterates, τ n = τ ◦ τ ◦ · · · ◦ τ (n times) describe the evolution of
the system after n units of time, and a principal concern is the average
behaviour, as n → ∞, of various quantities associated with the system.
Thus one is led to study averages
n−1
1X
(19) An (f )(x) = f (τ k (x)),
n
k=0
and their limits as n → ∞. To this we now turn.
5.1 Mean ergodic theorem

The first theorem dealing with the averages (19) that we consider is
purely Hilbert-space in character. Historically it preceded both Theo-
rems 5.3 and 5.4 which will be proved below.
For the specific application of the theorem below, one takes the Hilbert
space H to be L2 (X, M, µ). Given the measure-preserving transforma-
tion τ on X, we define the linear operator T on H by
(20) T (f )(x) = f (τ (x)).
Then T is an isometry; that is,
(21) kT f k = kf k,
where k · k denotes the Hilbert space (that is, the L2 ) norm. This is clear
from (18) with f replaced by |f |2 . Observe that if τ were also supposed
to be a measure-preserving isomorphism, then T would be invertible and
hence unitary; but we do not assume this.
Now with T as above, consider the subspace S of invariant vec-
tors, S = {f ∈ H : T (f ) = f }. Clearly, because of (21), the subspace
S is closed. Let P denote the orthogonal projection on this subspace.
The theorem that follows deals with the “mean” convergence, meaning
convergence in the norm.
Theorem 5.1 Suppose T is an isometry of the Hilbert space H, and let

P be the orthogonal projection on the subspace of the invariant vectors of
T . Let An = n1 (I + T + T 2 + · · · + T n−1 ). Then for each f ∈ H, An (f )
converges to P (f ) in norm, as n → ∞.
Together with the subspace S defined above we consider the subspaces

S∗ = {f ∈ H : T ∗ (f ) = f } and S1 = {f ∈ H : f = g − T g, g ∈ H}; here
T ∗ denotes the adjoint of T . Then S∗ , like S, is closed, but S1 is not
necessarily closed. We denote its closure by S1 . The proof of the theorem
is based on the following lemma.
Lemma 5.2 The following relations hold among the subspaces S, S∗ ,

and S1 .
(i) S = S∗ .
(ii) The orthogonal complement of S1 is S.
Proof. First, since T is an isometry, we have that (T f, T g) = (f, g)

for all f, g ∈ H, and thus T ∗ T = I. (See Exercise 22 in Chapter 4.) So
if T f = f then T ∗ T f = T ∗ f , which means that f = T ∗ f . To prove the
converse inclusion, assume T ∗ f = f . As a consequence (f, T ∗ f − f ) = 0,
and thus (f, T ∗ f ) − (f, f ) = 0; that is, (T f, f ) = kf k2 . However, kT f k =
kf k, so we have in the above an instance of equality for the Cauchy-
Schwarz inequality. As a result of Exercise 2 in Chapter 4 we get T f =
cf , which by the above gives T f = f . Thus part (i) is proved.
Next we observe that f is in the orthogonal complement of S1 ex-
actly when (f, g − T g) = 0, for all g ∈ H. However, this means that
(f − T ∗ f, g) = 0 for all g, and hence f = T ∗ f , which by part (i) means
f ∈ S.
Having established the lemma we can finish the proof of the theorem.
Given any f ∈ H, we write f = f0 + f1 , where f0 ∈ S and f1 ∈ S1 (since
S and S1 are orthogonal complements). We also fix ² > 0 and pick f10 ∈
S1 such that kf1 − f10 k < ². We then write
(22) An (f ) = An (f0 ) + An (f10 ) + An (f1 − f10 ),
and consider each term separately.

For the first term, we recall that P is the orthogonal projection on S,
so P (f ) = f0 , and since T f0 = f0 we deduce
n−1
1X k
An (f0 ) = T (f0 ) = f0 = P (f ) for every n ≥ 1.
n
k=0
For the second term, we recall the definition of S1 and pick a g ∈ H

with f10 = g − T g. Thus
n−1 n−1
1X k 1X k
An (f10 ) = T (1 − T )(g) = T (g) − T k+1 (g)
n n
k=0 k=0
1
= (g − T n (g)).
n
Since T is an isometry, the above identity shows that An (f10 ) converges
to 0 in the norm as n → ∞.
For the last term, we use once again the fact that each T k is an isometry
to obtain
n−1
1X k
kAn (f1 − f10 )k ≤ kT (f1 − f10 )k ≤ kf1 − f10 k < ².
n
k=0
Finally, from (22) and the above three observations, we deduce that
lim supn→∞ kAn (f ) − P (f )k ≤ ², and this concludes the proof of the the-
orem.
5.2 Maximal ergodic theorem

We now turn to the question of almost everywhere convergence of the
averages (19). As in the case of the averages that occur in the differ-
entiation theorems of Chapter 3, the key to dealing with such pointwise
limits lies in estimates for their corresponding maximal functions. In the
present case this function is defined by
m−1
1 X
(23) f ∗ (x) = sup |f (τ k (x))|.
1≤m<∞ m
k=0
Theorem 5.3 Whenever f ∈ L1 (X, µ), the maximal function f ∗ (x) is

finite for almost every x. Moreover, there is a universal constant A so
that
A
(24) µ({x : f ∗ (x) > α}) ≤ kf kL1 (X,µ) for all α > 0.
α
There are several proofs of this theorem. The one we choose emphasizes
the close connection to the maximal function given in Section 1.1 of
Chapter 3, and we shall in fact deduce the present theorem from the
one-dimensional case of that chapter. This argument gives the value
A = 6 for the constant in (24). By a different argument one can obtain
A = 1, but this improvement is not relevant in what follows.
Before beginning the proof, we make some preliminary remarks. Note
that in the present case the function f ∗ is automatically measurable,
since it is the supremum of a countable number of measurable functions.
Also, we may assume that our function f is non-negative, since otherwise
we may replace it by |f |.
Step 1. The case when X = Z and τ : n 7→ n + 1.

For each function f on Z, we consider its extension f˜ to R defined by
f˜(x) = f (n) for n ≤ x < n + 1, n ∈ Z. (See Figure 2.)
f (n) f˜(x)
−1 n=0 1 2 −1 x=0 1 2
Figure 2. Extension of f to R
S
Similarly, if E ⊂ Z, denote by Ẽ the set in R given by Ẽ = n∈E [n, n +
1).
R Note thatP as a result of these definitions we have m(Ẽ) = #(E) and
R
f (x) dx = n∈Z f (n), and thus kf˜kL1 (R) = kf kL1 (Z) . Here m is the
˜
Lebesgue measure on R, and # is the counting measure on Z. Note also
that
m−1
X Z m
f (n + k) = f˜(n + t) dt.
k=0 0
Rm Rm
However, because 0
f˜(n + t) dt ≤ −1
f˜(x + t) dt whenever x ∈ [n, n +
1), we see that
m−1 µ ¶ Z m
1 X m+1 1
f (n + k) ≤ f˜(x + t) dt if x ∈ [n, n + 1).
m m m + 1 −1
k=0
Taking the supremum over all m ≥ 1 in the above and noting that (m +
1)/m ≤ 2, we obtain
(25) f ∗ (n) ≤ 2(f˜)∗ (x) whenever x ∈ [n, n + 1).
To be clear about the notation here: f ∗ (n) denotes the maximal function
of f on Z defined by (23), with f (τ k (n)) = f (n + k), while (f˜)∗ is the
maximal function as defined in Chapter 3, of the extended function f˜
on R.
By (25)
#({n : f ∗ (n) > α}) ≤ m({x ∈ R : (f˜)∗ (x) > α/2}),

R
and thus the latter is majorized by A0 /(α/2) f˜(x) dx = 2A0 /αkf˜kL1 (R) ,
according to the maximal theorem for R. The constant A0 that occurs in
that theorem (there denoted by A) can be taken to be 3. Hence we have
6
(26) #({n : f ∗ (n) > α}) ≤ kf kL1 (Z) ,
α
since kf˜kL1 (R) = kf kL1 (Z) . This disposes of the special case when X = Z.
Step 2. The general case.
By a sleight-of-hand we shall “transfer” the result for Z just proved to
the general case. We proceed as follows.
For every positive integer N , we consider the truncated maximal func-
∗
tion fN defined as
m−1
∗ 1 X
fN (x) = sup f (τ k (x)).
1≤m≤N m
k=0
∗ ∗
Since {fN (x)} forms an increasing sequence with N , and limN →∞ fN (x) =
∗
f (x) for every x, it suffices to show that
∗ A
(27) µ{x : fN (x) > α} ≤ kf kL1 (X,µ) ,
α
with constant A independent of N . Letting N → ∞ will then give the
desired result.
So in place of f ∗ we estimate fN
∗
, and to simplify our notation we write
∗
the latter as f , dropping the N subscript. Our argument will compare
the maximal function f ∗ with the special case arising for Z. To clarify
the formula below we temporarily adopt the expedient of denoting the
second maximal function by M(f ). Thus for a positive function f on Z
we set
m−1
1 X
M(f )(n) = sup f (n + k).
1≤m m
k=0
Now starting with a function f on X that is integrable, we define the

function F on X × Z by
½
f (τ n (x)) if n ≥ 0,
F (x, n) =
0 if n < 0.
Then
m−1 m−1
1 X 1 X
Am (f )(x) = f (τ k (x)) = F (x, k).
m m
k=0 k=0
In the above we replace x by τ n (x); then since τ k (τ n (x)) = τ n+k (x), we

have
m−1
1 X
Am (f )(τ n (x)) = F (x, n + k).
m
k=0
Now we fix a large positive a and set b = a + N . We also write Fb for

the truncated function on X × Z defined by Fb (x, n) = F (x, n) if n < b,
Fb (x, n) = 0 otherwise. We then have
m−1
1 X
Am (f )(τ n (x)) = Fb (x, n + k) if m ≤ N and n < a.
m
k=0
Thus
(28) f ∗ (τ n (x)) ≤ M(Fb )(x, n) if n < a.

(Recall that f ∗ is actually fN

∗
!) This is the comparison of the two maxi-
mal functions we wished to obtain. Now set Eα = {x : f ∗ (x) > α}. Then
by the measure-preserving character of τ , µ({x : f ∗ (τ n (x)) > α}) =
µ(Eα ). Hence on the product space X × Z the product measure µ × #
of the set {(x, n) ∈ X × Z : f ∗ (τ n (x)) > α, 0 ≤ n < a} equals aµ(Eα ).
However, because of (28) the µ × # measure of this set is no more than
Z
#({n ∈ Z : M(Fb )(x, n) > α}) dµ.
X
Because of the maximal estimate (26) for Z, we see that the integrand
above is no more than
b−1
A A X
kFb (x, n)kL1 (Z) = f (τ n (x)),
α α
n=0
with of course A = 6. R n
R Hence, integrating this over X and recalling that X f (τ (x)) dµ =
X
f (x) dµ gives us
A A
aµ(Eα ) ≤ b kf kL1 (X) = (a + N ) kf kL1 (X) .
α α
¡ ¢
Thus µ(Eα ) ≤ A N
α 1 + a kf kL (X) , and letting a → ∞ yields estimate (27).
1
As we have seen, a final limit as N → ∞ then completes the proof.
5.3 Pointwise ergodic theorem

The last of the series of limit theorems we will study is the pointwise
(or individual) ergodic theorem, which combines ideas of the first two
theorems. At this stage it will be convenient to assume that the measure
space (X, µ) is finite; we can then normalize the measure and suppose
µ(X) = 1.
Theorem 5.4 Suppose f is integrable

Pm−1 overk X. Then for almost every
1
x ∈ X the averages Am (f ) = m k=0 f (τ (x)) converge to a limit as
m → ∞.
Corollary 5.5 If we denote this limit by P 0 (f ), we have that

Z Z
0
|P (f )(x)| dµ(x) ≤ |f (x)| dµ(x).
X X
Moreover P 0 (f ) = P (f ) whenever f ∈ L2 (X, µ).

The idea of the proof is as follows. We first show that Am (f ) converges

to a limit almost everywhere for a set of functions f that is dense in
L1 (X, µ). We then use the maximal theorem to show that this implies
the conclusion for all integrable functions.
We remark to begin with that because the total measure of X is 1, we
have L2 (X, µ) ⊂ L1 (X, µ) and kf kL1 ≤ kf kL2 , and moreover L2 (X, µ) is
dense in L1 (X, µ). In fact, if f belongs to L1 , consider the sequence
{fn } defined by fn (x) = f (x) if |f (x)| ≤ n, fn (x) = 0 otherwise. Then
each fn is clearly in L2 , while by the dominated convergence theorem
kf − fn kL1 → 0.
Now starting with an integrable f and any ² > 0 we shall see that we
can write f = F + H, where kHkL1 < ², and F = F0 + (1 − T )G, where
both F0 and G belong to L2 , and T (F0 ) = F0 , with T (F0 ) = F0 (τ (x)). To
obtain this decomposition of f , we first write f = f 0 + h0 , where f 0 ∈ L2
and kh0 kL1 < ²/2, which we can do in view of the density of L2 in L1
as seen above. Next, since the subspaces S and S1 of Lemma 5.2 are
orthogonal complements in L2 , we can find F0 ∈ S, F1 ∈ S1 , such that
f 0 = F0 + F1 + h with khkL2 < ²/2. Because F1 ∈ S1 is automatically
of the form F1 = (1 − T )G, we obtain f = F + H, with F = F0 + (1 −
T )G and H = h + h0 . Thus kHkL1 ≤ khkL1 + kh0 kL1 and since khkL1 ≤
khkL2 < ²/2 we have achieved our desired decomposition of f .
1
Now Am (F ) = Am (F0 ) + Am ((1 − T )G) = F0 + m (1 − T m (G)), as we
1 m
have already seen in the proof of Theorem 5.1. Note that m T (G) =
1 m
m G(τ (x)) converges
P∞ to zero as m → ∞ for almost every x ∈ X. In-
deed, the series m=1 m12 (G(τ m (x)))2 converges almost everywhere by
the monotone convergence theorem, since its integral over X is
∞
X ∞
X
1 m 2 2 1
kT GkL2 = kGkL2 ,
m2 m2
m=1 m=1
which is finite.
As a result, Am (F )(x) converges for almost every x ∈ X. Finally,
to prove the corresponding convergence for Am (f )(x), we argue as in
Theorem 1.3 in Chapter 3 and set
Eα = {x : lim sup |An (f )(x) − Am (f )(x)| > α}.

N →∞ n,m≥N
Then it suffices to see that µ(Eα ) = 0 for all α > 0. However, since
An (f ) − Am (f ) = An (F ) − Am (F ) + An (H) − Am (H), and Am (F )(x) con-
verges almost everywhere as m → ∞, it follows that almost every point
in the set Eα is contained in Eα0 , where
Eα0 = {x : sup |An (H)(x) − Am (H)(x)| > α},

n,m≥N
and thus µ(Eα ) ≤ µ(Eα0 ) ≤ µ({x : 2 supm |Am (H)(x)| > α}). The last
quantity is majorized by A/(α/2)kHkL1 ≤ 2²A/α by Theorem 5.3. Since
² was arbitrary we see that µ(Eα ) = 0, and hence Am (f )(x) is a Cauchy
sequence for almost every x, and the theorem is proved.
To establish the corollary, observe that if f ∈ L2 (X), we know by
Theorem 5.1 that Am (f ) converges to P (f ) in the L2 -norm, and hence
a subsequence converges almost everywhere to that limit, showing that
P (f ) = P 0 (f ) in that case. Next, for any f that is merely integrable, we
have
Z m−1 Z Z
1 X
|Am (f )| dx ≤ |f (τ k (x))| dµ(x) = |f (x)| dµ(x),
X m X X
k=0
0
and thus sinceR Am0 (f ) → P (f ) almost
R everywhere, we get by Fatou’s
lemma that X |P (f )(x)| dµ(x) ≤ X |f (x)| dµ(x). With this the corol-
lary is also proved.
It can be shown that the conclusions of the theorem and corollary are
still valid if we drop the assumption that the space X has finite measure.
The modifications of the argument needed to obtain this more general
conclusion are outlined in Exercise 26.
5.4 Ergodic measure-preserving transformations

The adjective “ergodic” is commonly applied to the three limit theorems
proved above. It also has a related but separate usage describing an
important class of transformations of the space X.
We say that a measure-preserving transformation τ of X is ergodic
if whenever E is a measurable set that is “invariant,” that is, E and
τ −1 (E) differ by sets of measure zero, then either E or E c has measure
zero.
There is a useful rephrasing of this condition of ergodicity. Expanding
the definition used in Section 5.1 we say that a measurable function f
is invariant if f (x) = f (τ (x)) for a.e. x ∈ X. Then τ is ergodic exactly
when the only invariant functions are equivalent to constants. In fact,
let τ be an ergodic transformation, and assume that f is a real-valued in-
variant function. Then each of the sets Ea = {x : f (x) > a} is invariant,
hence µ(Ea ) = 0 or µ(Eac ) = 0 for each a. However, if f is not equivalent
to a constant, then both µ(Ea ) and µ(Eac ) must have strictly positive
measure for some a. In the converse direction we merely need to note
that if all characteristic functions of measurable sets that are invariant
must be constants, then τ is ergodic.
The following result subsumes the conclusion of Theorem 5.4 for er-
godic transformations. We keep to the assumption of that theorem that
the underlying space X has measure equal to 1.
Corollary 5.6 Suppose τ is an ergodic measure-preserving transforma-

tion. For any integrable function f we have
m−1 Z
1 X
f (τ k (x)) converges to f dµ for a.e. x ∈ X as m → ∞.
m X
k=0
The result has the interpretation that the “time average” of f equals its
“space average.”
Proof. By Theorem 5.1 we know that the averages Am (f ) converge
to P (f ), whenever f ∈ L2 , where P is the orthogonal projection on the
subspace of invariant vectors. Since in this case the invariant vectors
form a one-dimensional space Rspanned by the constant functions, we
observe that P (f ) = 1(f, 1) = X f dµ, where 1 designates the function
identically equal to 1 on X. To verify this, note that P is the identity on
constants and annihilates all functions orthogonal to constants. Next we
write any f ∈ L1 as g + h, where g ∈ L2 and khkL1 < ². Then P 0 (f ) =
P 0 (g) + P 0 (h). However, we also know that P 0 (g) = P (g), and kP 0 (h)k ≤
khkL1 < ² by the corollary to Theorem 5.4. Thus
Z Z
0
P (f ) − f dµ = (g − f ) dµ + P 0 (h)
X X
R
yields that kP 0 (f ) − RX f dµkL1 ≤ kg − f kL1 + ² < 2². This shows that
P 0 (f ) is the constant X f dµ and the assertion is proved.
We shall now elaborate on the nature of ergodicity and illustrate its
thrust in terms of several examples.
a) Rotations of the circle

Here we take up the example described in (iii) at the beginning of
Section 5*. On the unit circle R/Z with the induced Lebesgue measure,
we consider the action τ given by x 7→ x + α mod 1. The result is
• The mapping τ is ergodic if and only if α is irrational.

To begin with, if α is irrational we know by the equidistribution theorem

that
n−1 Z 1
1X
(29) f (x + kα) → f (x) dx as n → ∞
n 0
k=0
for every x if f is continuous on [0, 1] and periodic (f (0) = f (1)). The

argument used to prove this goes as follows.2 First we verify that (29)
holds whenever f (x) = e2πinx , n ∈ Z, by considering the cases n = 0 and
n 6= 0 separately. It then follows that (29) is valid for any trigonometric
polynomial (a finite linear combination of these exponentials). Finally,
any continuous and periodic function can be uniformly approximated by
trigonometric polynomials, so (29) goes over to the general case.
Now if P is the projection on invariant L2 -functions, then Theorem 5.1
and (29) show that P projects onto the constants, when restricted to the
continuous periodic functions. Since this subspace is dense in L2 , we
see that P still projects all of L2 on constants; hence the invariant L2 -
functions are constants and thus τ is ergodic.
On the other hand, suppose α = p/q. Choose any set E0 ⊂ S (0, 1/q), so
q−1
that 0 < m(E0 ) < 1/q, and let E denote the disjoint union r=0 (E0 +
r/q). Then clearly E is invariant under τ : x 7→ x + p/q, and 0 < m(E) =
qm(E0 ) < 1; thus τ is not ergodic.
The property (29) we used, which involves the existence of the limit
at all points, is actually stronger than ergodicity: it implies that the
measure dµ = dx is uniquely ergodic for this mapping τ . That means
that if ν is any measure on the Borel sets of X preserved by τ and
ν(X) = 1, then ν must equal µ.
To see that this so in the present case, let Pν be the orthogonal projec-
tion guaranteed by Theorem 5.1, on the space L2 (X, ν). Then (29) shows
again that the range of Pν on the continuous functions, andRthen on all
1
of L2 (X, ν), is the subspace of constants, and thus Pν (f ) = 0 f dν.
R1 R1
This means also that 0 f (x) dx = 0 f dν whenever f is continuous
and periodic. By a simple limiting argument we then get that the mea-
sure dx = dµ and ν agree on all open intervals, and thus on all open
sets. As we have seen, this then proves that the two measures are then
identical.
In general, uniquely ergodic measure-preserving transformations are
ergodic, but the converse need not be true, as we shall see below.
b) The doubling mapping
2 See also Section 2, Chapter 4 in Book I.

We now consider the mapping x 7→ 2x mod 1 for x ∈ (0, 1], with µ

Lebesgue measure, that arose in example (iv) at the beginning of Sec-
tion 5*. We shall prove that τ is ergodic and in fact satisfies a different
and stronger property called mixing.3 It is defined as follows.
If τ is a measure-preserving transformation on the space (X, µ), it is
said to be mixing if whenever E and F are a pair of measurable subsets
then
(30) µ(τ −n (E) ∩ F ) → µ(E)µ(F ) as n → ∞.
The meaning of (30) can be understood as follows. In probability theory

one often encounters a “universe” of possible events to which probabilities
are assigned. These events are represented as measurable subsets E of
some space (X, µ) with µ(X) = 1. The probability of each event is then
µ(E). Two events E and F are “independent” if the probability that
they both occur is the product of their separate probabilities, that is,
µ(E ∩ F ) = µ(E)µ(F ). The assertion (30) of mixing is then that in the
limit as time n tends to infinity, the sets τ −n (E) and F are asymptotically
independent, whatever the choices of E and F .
We shall next observe that the mixing condition is implied by the
seemingly stronger condition
(31) (T n f, g) → (f, 1)(1, g) as n → ∞,
where T n (f )(x) = f (τ n (x)) whenever f and g belong to L2 (X, µ). This

implication follows immediately upon taking f = χE and g = χF . The
converse is also true, but we leave its proof as an exercise to the reader.
We now remark that the mixing condition implies the ergodicity of τ .
Indeed, by (31)
n−1
1X k
(An (f ), g) = (T f, g) converges to (f, 1)(1, g).
n
k=0
This means (P (f ), g) = (f, 1)(1, g), and hence P (f ) is orthogonal to all

g that are orthogonal to constants. This of course means that P is the
orthogonal projection on constants, and hence τ is ergodic.
We next observe that the doubling map is mixing. Indeed, if f (x) =
2πimx
e , g(x) = e2πikx , then (f, 1)(1, g) = 0, unless both m and k are 0,
in which case this product equals 1. However, in this case (T n f, g) =
R 1 2πim2n x −2πikx
0
e e dx, and this vanishes for sufficiently large n, unless
3 This property is often referred to as a “strongly mixing” to distinguish it from still
another kind of ergodicity called “weakly mixing.”

both m and k are 0, in which case the integral equals 1. Thus (31)
holds for all exponentials f (x) = e2πimx , g(x) = e2πikx , and therefore by
linearity for all trigonometric polynomials f and g. It is from there an
easy step to use the completeness in Chapter 4 to pass to all f and g in
L2 ((0, 1]) by approximating these functions in the L2 -norm by trigono-
metric polynomials.
Let us observe that the action of rotations τ : x 7→ x + α of the unit
circle for irrational α, although ergodic, is not mixing. Indeed, if we take
f (x) = g(x) = e2πimx , m 6= 0, then (T n f, g) = e2πinmα (f, g) = e2πinmα ,
while (f, 1) = (1, g) = 0; thus (T n f, g) does not converge to (f, 1)(1, g)
as n → ∞.
Finally, we note that the doubling map τ : x 7→ 2x mod 1 on (0, 1]
is not uniquely ergodic. Besides the Lesbesgue measure, the measure ν
with ν{1} = 1 but ν(E) = 0 if 1 ∈ / E is also preserved by τ .
Further examples of ergodic transformations are given below.
6* Appendix: the spectral theorem

The purpose of this appendix is to present an outline of the proof of the spectral
theorem for bounded symmetric operators on a Hilbert space. Details that are
not central to the proof of the theorem will be left to the interested reader to fill
in. The theorem provides an interesting application of the ideas related to the
Lebesgue-Stieltjes integrals that are treated in this chapter.
6.1 Statement of the theorem

A basic notion is that of a spectral resolution (or spectral family) on a Hilbert
space H. This is a function λ 7→ E(λ) from R to orthogonal projections on H that
satisfies the following:
(i) E(λ) is increasing in the sense that kE(λ)f k is an increasing function of λ

for every f ∈ H.
(ii) There is an interval [a, b] such that E(λ) = 0 if λ < a, and E(λ) = I if λ ≥ b.
Here I denotes the identity operator on H.
(iii) E(λ) is right-continuous, that is, for every λ one has
lim E(µ)f = E(λ)f for every f ∈ H.

µ → λ
µ > λ
Observe that property (i) is equivalent with each of the following three assertions
(holding for all pairs λ, µ with µ > λ): (a) the range of E(µ) contains the range of
E(λ); (b) E(µ)E(λ) = E(λ); (c) E(µ) − E(λ) is an orthogonal projection.
Now given a spectral resolution {E(λ)} and an element f ∈ H, note that the
function λ 7→ (E(λ)f, f ) = kE(λ)f k2 is also increasing. As a result, the polar-
ization identity (see Section 5 in Chapter 4) shows that for every pair f, g ∈ H,
6*. Appendix: the spectral theorem 307
the function F (λ) = (E(λ)f, g) is of bounded variation, and is moreover right-

continuous. With these two observations we can now state the main result.
Theorem 6.1 Suppose T is a bounded symmetric operator on a Hilbert space H.

Then there exists a spectral resolution {E(λ)} such that
Z b
T = λ dE(λ)
a−
in the sense that for every f, g ∈ H

Z b Z b
(32) (T f, g) = λ d(E(λ)f, g) = λ dF (λ).
a− a−
The integral on the right-hand side is taken in the Lebesgue-Stieltjes sense, as

in (iii) and (iv) of Section 3.3.
The result encompasses the spectral theorem for compact symmetric operators T
in the following sense. Let {ϕk } be an orthonormal basis of eigenvectors of T with
corresponding eigenvalues λk , as guaranteed by Theorem 6.2 in Chapter 4. In this
case, we take the spectral resolution to be defined via this orthogonal expansion
by
X
E(λ)f ∼ (f, ϕk )ϕk ,
λk ≤λ
and one easily verifies P

that it satisfies conditions (i), (ii) and (iii) above. We also
note that kE(λ)f k2 = λk ≤λ |(f, ϕk )|2 , and thus F (λ) = (E(λ)f, g) is a pure jump
function as in Section 3.3 in Chapter 3.
6.2 Positive operators

The proof of the theorem depends on the concept of positivity of operators. We
say that T is positive, written as T ≥ 0, if T is symmetric and (T f, f ) ≥ 0 for
all f ∈ H. (Note that (T f, f ) is automatically real if T is symmetric.) One then
writes T1 ≥ T2 to mean that T1 − T2 ≥ 0. Note that for two orthogonal projections
we have E2 ≥ E1 if and only if kE2 f k ≥ kE1 f k for all f ∈ H, and that is then
equivalent with the corresponding properties (a)−(c) described above. Notice also
that if S is symmetric, then S 2 = T is positive. Now for T symmetric, let us write
(33) a = min(T f, f ) and b = max(T f, f ) for kf k ≤ 1.
Proposition 6.2 Suppose T is symmetric. Then kT k ≤ M if and only if −M I ≤

T ≤ M I. As a result, kT k = max(|a|, |b|).
This is a consequence of (7) in Chapter 4.
Proposition 6.3 Suppose T is positive. Then there exists a symmetric operator

S (which can be written as T 1/2 ) such that S 2 = T and S commutes with every
operator that commutes with T .
The last assertion means that if for some operator A we have AT = T A, then
AS = SA.
The existence of S is seen as follows. After multiplying by a suitable positive
scalar, we may assume that PkT k ≤ k1. Consider the binomial expansion of (1 −
t)1/2 , given by (1 − t)1/2 = ∞ bk t , for |t| < 1. The relevant fact that is needed
k=0P
here is that the bk are real and ∞ k=0 |bk | < ∞. Indeed, by direct calculation of
the power series expansion of (1 − t)1/2 we find that b0 = 1, b1 = −1/2, b2 = −1/8,
and more generally, bk = −1/2 · 1/2 · · · (k − 3/2)/k!, if k ≥ 2, from which it follows
that bk = O(k−3/2 ). Or morePsimply, since bk < 0Pwhen k ≥ 1, if we let t → 1 in
the definition, we see that − ∞ k=1 bk = 1 and so
∞
k=0 |bk | = 2.
Pn k
Now let sn (t) denote the polynomial k=0 bk t . Then the polynomial
2n
X
(34) s2n (t) − (1 − t) = cn
kt
k
k=0
P2n
has the property
P∞that k=0 |cn
k | → 0 as n → ∞. In fact, sn (t) = (1 − t)
1/2
− rn (t),
k 2 2
with rn (t) = k=n+1 bk t , so sn (t) − (1 − t) = −rn (t) − 2sn (t)rn (t). Now the left-
hand side is clearly a polynomial of degree ≤ 2n, and so comparing P coefficients with
n
those on the right-hand side shows
P that the c
Pk are majorized by 3 j>n |bj | |bk−j |.
From this it is immediate that k |cn k | = O( j>n |b j |) → 0 as n → ∞, as asserted.
To apply this, set T1 = I −PT ; then 0 ≤ T1 ≤ I, and thus kT1 k ≤ 1, by Proposi-
tion 6.2. Let Sn = snP (T1 ) = n k 0
k=0 bk T1 , with T1 = I. Then in terms of operator
norms, kSn − Sm k ≤ k≥min(n,m) |bk | → 0 as n, m → ∞, because kT1k k ≤ kT1 kk ≤
1. Hence Sn converges to some operator S. Clearly Sn is symmetric
P for each n,
and thus S isPalso symmetric. Moreover, by (34), Sn2 − T = 2n n k
k=0 ck T1 , therefore
kSn2 − T k ≤ |cn 2
k | → 0 as n → ∞, which implies that S = T . Finally, if A com-
mutes with T it clearly commutes with every polynomial in T , hence with Sn , and
thus with S. The proof of the proposition is therefore complete.
Proposition 6.4 If T1 and T2 are positive operators that commute, then T1 T2 is

also positive.
Indeed, if S is a square root of T1 given in the previous proposition, then T1 T2 =

SST2 = ST2 S, and hence (T1 T2 f, f ) = (ST2 Sf, f ) = (T2 Sf, Sf ), since S is sym-
metric, and thus the last term is positive.
Proposition 6.5 Suppose T is symmetric and a and b are given by (33). If p(t) =
P n
tk is a real polynomial which is positive for t ∈ [a, b], then the operator
k=0 ckP
p(T ) = n k=0 ck T
k
is positive.
Q Q Q
To see this, write p(t) = c j (t − ρj ) k (ρ0k − t) ` ((t − µ` )2 + ν` ), where c is pos-
itive and the third factor corresponds to the non-real roots of p(t) (arising in con-
jugate pairs), and the real roots of p(t) lying in (a, b) which are necessarily of
even order. The first factor contains the real roots ρj with ρj ≤ a, and the second
factor the real roots ρ0k with ρ0k ≥ b. Since each of the factors T − ρj I, ρ0j I − T
and (T − µ` I)2 + ν`2 I is positive and these commute, the desired conclusion follows
from the previous proposition.
Corollary 6.6 If p(t) is a real polynomial, then
kp(T )k ≤ sup |p(t)|.

t∈[a,b]
This is an immediate consequence using Proposition 6.2, since −M ≤ p(t) ≤ M ,

where M = supt∈[a,b] |p(t)|, and thus −M I ≤ p(T ) ≤ M I.
Proposition 6.7 Suppose {Tn } is a sequence of positive operators that satisfy

Tn ≥ Tn+1 for all n. Then there is a positive operator T , such that Tn f → T f as
n → ∞ for every f ∈ H.
Proof. We note that for each fixed f ∈ H the sequence of positive numbers
(Tn f, f ) is decreasing and hence convergent. Now observe that for any positive
operator S with kSk ≤ M we have
(35) kS(f )k2 ≤ (Sf, f )1/2 M 3/2 kf k.
In fact, the quadratic function (S(tI + S)f, (tI + S)f ) = t2 (Sf, f ) + 2t(Sf, Sf ) +
(S 2 f, Sf ) is positive for all real t. Hence its discriminant is negative, that is,
kS(f )k4 ≤ (Sf, f )(S 2 f, Sf ), and (35) follows. We apply this to S = Tn − Tm with
n ≤ m; then kTn − Tm k ≤ kTn k ≤ kT1 k = M , and since ((Tn − Tm )f, f ) → 0 as
n, m → ∞ we see that kTn f − Tm f k → 0 as n, m → ∞. Thus limn→∞ Tn (f ) =
T (f ) exists, and T is also clearly positive.
6.3 Proof of the theorem

Starting with a given symmetric operator T , and with a, b given by (33), we shall
now exploit further the idea of associating to each suitable function Φ on [a, b] a
symmetric operator Φ(T P). We do this in increasing order of generality.
Pn First, kif
Φ is a real polynomial n k
k=0 ck t , then, as before, Φ(T ) is defined as k=0 ck T .
Notice that this association is a homomorphism: if Φ = Φ1 + Φ2 , then Φ(T ) =
Φ1 (T ) + Φ2 (T ); also if Φ = Φ1 · Φ2 , then Φ(T ) = Φ1 (T ) · Φ2 (T ). Moreover, since
Φ is real (and the ck are real), Φ(T ) is symmetric.
Next, because every real-valued continuous function Φ on [a, b] can be approx-
imated uniformly by polynomials pn (see, for instance, Section 1.8, Chapter 5 of
Book I), we see by Corollary 6.6 that the sequence pn (T ) converges, in the norm of
operators, to a limit which we call Φ(T ), and moreover this limit does not depend
on the sequence of polynomials approximating Φ. Also, Φ(T ) is automatically a
symmetric operator. If Φ(t) ≥ 0 on [a, b] we can always take the approximating
sequence to be positive on [a, b], and as a result Φ(T ) ≥ 0.
Finally, we define Φ(T ) whenever Φ arises as a limit, Φ(t) = limn→∞ Φn (t),
where {Φn (t)} is a decreasing sequence of positive continuous functions on [a, b]. In
fact, by Proposition 6.7 the limit limn→∞ Φn (T ) exists by what we have established
above for Φn . To show that this limit is independent of the sequence {Φn } and
thus that Φ(t) is well-defined as the limit above, let {Φ0n } be another sequence of
decreasing continuous functions converging to Φ. Then whenever ² > 0 is given and
k is fixed, Φ0n (t) ≤ Φk (t) + ² for all n sufficiently large. Thus Φ0n (T ) ≤ Φk (T ) + ²I
for these n, and passing to the limit first in n, then in k, and then with ² → 0, we get
limn→∞ Φ0n (T ) ≤ limk→∞ Φk (T ). By symmetry, the reverse inequality holds, and

the two limits are the same. Note also that for a pair of these limiting functions,
if Φ1 (t) ≤ Φ2 (t) for t ∈ [a, b], then Φ1 (T ) ≤ Φ2 (T ).
The basic functions Φ, Φ = ϕλ , that give us the spectral resolution are defined
for each real λ by
ϕλ (t) = 1 if t ≤ λ and ϕλ (t) = 0 if λ < t.
We note that ϕλ (t) = lim ϕλn (t), where ϕλn (t) = 1 if t ≤ λ, ϕλn (t) = 0 if t ≥ λ + 1/n,
and ϕλn (t) is linear for t ∈ [λ, λ + 1/n]. Thus each ϕλ (t) is a limit of a decreasing
sequence of continuous functions. In accordance with the above we set
E(λ) = ϕλ (T ).
Since limn→∞ ϕλn1 (t)ϕλn2 (t) = ϕλn1 (t) whenever λ1 ≤ λ2 , we see that E(λ1 )E(λ2 ) =
E(λ1 ). Thus E(λ)2 = E(λ) for every λ, and because E(λ) is symmetric it is
therefore an orthogonal projection. Moreover, for every f ∈ H
kE(λ1 )f k = kE(λ1 )E(λ2 )f k ≤ kE(λ2 )f k,
thus E(λ) is increasing. Clearly E(λ) = 0 if λ < a, since for those λ, ϕλ (t) = 0 on
[a, b]. Similarly, E(λ) = I for λ ≥ b.
Next we note that E(λ) is right-continuous. In fact, fix f ∈ H and ² > 0. Then
for some n, which we now keep fixed, kE(λ)f − ϕλn (T )f k < ². However, ϕµ n (t)
converges to ϕλn (t) uniformly in t as µ → λ. Hence supt |ϕµ λ
n (t) − ϕn (t)| < ², if
|µ − λ| < δ, for an appropriate δ. Thus by the corollary kϕµ λ
n (T ) − ϕn (T )k < ²
µ
and therefore kE(λ)f − ϕn (T )k < 2². Now with µ ≥ λ we have that E(µ)E(λ) =
E(λ) and E(µ)ϕµ n (T ) = E(µ). As a result kE(λ)f − E(µ)f k < 2², if λ ≤ µ ≤ λ +
δ. Since ² was arbitrary, the right continuity is established.
Finally we verify the spectral representation (32). Let a = λ0 < λ1 < · · · < λk =
b be any partition of [a, b] for which supj (λj − λj−1 ) < δ. Then since
k
X
t= t(ϕλj (t) − ϕλj−1 (t)) + tϕλ0 (t)
j=1
we note that
k
X
t≤ λj (ϕλj (t) − ϕλj−1 (t)) + λ0 ϕλ0 (t) ≤ t + δ.
j=1
Applying these functions to the operator T we obtain
k
X
T ≤ λj (E(λj ) − E(λj−1 )) + λ0 E(λ0 ) ≤ T + δI,
j=1
and thus T differs in norm from the sum above by at most δ. As a result
˛ Z ˛
˛ k
X ˛
˛ ˛
˛(T f, f ) − λj d(E(λ)f, f ) − λ0 (E(λ0 )f, f )˛ ≤ δkf k2 .
˛ (λj−1 ,λj ] ˛
j=1
But as we vary the partitions of [a, b], letting their meshes δ tend to zero, the
Rb Rb
above sum tends to a− λ d(E(λ)f, f ). Therefore (T f, f ) = a− λ d(E(λ)f, f ), and
the polarization identity gives (32).
A similar argument shows that if Φ is continuous on [a, b], then the operator
Φ(T ) has an analogous spectral representation
Z b
(36) (Φ(T )f, g) = Φ(λ) d(E(λ)f, g).
a−
P
This is because |Φ(t) − kj=1 Φ(λj )(ϕλj (t) − ϕλj−1 (t)) − Φ(λ0 )ϕλ0 (t)| < δ 0 , where
δ 0 = sup|t−t0 |≤δ |Φ(t) − Φ(t0 )|, which tends to zero as δ → 0.
This representation also extends to continuous Φ that are complex-valued (by
considering the real and imaginary parts separately) or for Φ that are limits of
decreasing pointwise continuous functions.
6.4 Spectrum
We say that a bounded operator S on H is invertible if S is a bijection of H
and its inverse, S −1 , is also bounded. Note that S −1 satisfies S −1 S = SS −1 = I.
The spectrum of S, denoted by σ(S), is the set of complex numbers z for which
S − zI is not invertible.
Proposition 6.8 If T is symmetric, then σ(T ) is a closed subset of the interval
[a, b] given by (33).
Note that if z ∈/ [a, b], the function Φ(t) = (t − z)−1 is continuous on [a, b] and
Φ(T )(T − zI) = (T − zI)Φ(T ) = I, so Φ(T ) is the inverse of T − zI. Now suppose
T0 = T − λ0 I is invertible. Then we claim that T0 − ²I is invertible for all (com-
plex) ² that are sufficiently small. This will prove that the complement of σ(T ) is
open. Indeed, T0 − ²I = T0 (I − ²T0−1 ), and we can invert the operator (I − ²T0−1 )
(formally) by writing its inverse as a sum
∞
X
²n (T0−1 )n+1 .
n=0
P P n −1 n+1
Since ∞ n −1 n+1
n=0 k² (T0 ) k≤ |²| kT0 k , the series converges when |²| < kT0−1 k−1 ,
and the sum is majorized by
1
(37) kT0−1 k .
1 − |²|kT0−1 k
P
Thus we can define the operator (T0 − ²I)−1 as limN →∞ T0−1 N n −1 n+1
n=0 ² (T0 ) ,
and it gives the desired inverse, as is easily verified.
Our last assertion connects the spectrum σ(T ) with the spectral resolution
{E(λ)}.
Proposition 6.9 For each f ∈ H, the Lebesgue-Stieltjes measure corresponding

to F (λ) = (E(λ)f, f ) is supported on σ(T ).
To put it another way, F (λ) is constant on each open interval of the complement
of σ(T ).
To prove this, let J be one of the open intervals in the complement of σ(T ),
x0 ∈ J, and J0 the sub-interval centered at x0 of length 2², with ² < k(T − x0 I)−1 k.
First note that if z has non-vanishing imaginary part then (T − zI)−1 is given by
Φz (T ), with Φz (t) = (t − z)−1 . Hence (T − zI)−1 (T − zI)−1 is given by Ψz (T ),
with Ψz (t) = 1/|t − z|2 . Therefore by the estimate given in (37) and the represen-
tation (36) applied to Φ = Ψz , we obtain
Z
dF (λ)
≤ A0 ,
|λ − z|2
as long as z is complex and |x0 − z| < ². We can therefore obtain the same in-
equality
R for x real, |x0 − x| < ². NowR integration in x ∈ J0 using the fact that
dx
J² |λ−x|2 = ∞ for every λ ∈ J² , gives J² dF (λ) = 0. Thus F (λ) is constant in J² ,
but since x0 was an arbitrary point of J the function F (λ) is constant throughout
J and the proposition is proved.
7 Exercises
1. Let X be a set and M a non-empty collection of subsets of X. Prove that if

M is closed under complements and countable unions of disjoint sets, then M is
a σ-algebra.
[Hint: Any countable union of sets can be written as a countable union of disjoint
sets.]
2. Let (X, M, µ) be a measure space. One can define the completion of this
space as follows. Let M be the collection of sets of the form E ∪ Z, where E ∈ M,
and Z ⊂ F with F ∈ M and µ(F ) = 0. Also, define µ(E ∪ Z) = µ(E). Then:
(a) M is the smallest σ-algebra containing M and all subsets of elements of M

of measure zero.
(b) The function µ is a measure on M, and this measure is complete.
[Hint: To prove M is a σ-algebra it suffices to see that if E1 ⊂ M, then E1c ⊂ M.

Write E1 = E ∪ Z with Z ⊂ F , E and F in M. Then E1c = (E ∪ F )c ∪ (F − Z).]
3. Consider the exterior Lebesgue measure m∗ introduced in Chapter 1. Prove that

a set E in Rd is Carathéodory measurable if and only if E is Lebesgue measurable
in the sense of Chapter 1.
[Hint: If E is Lebesgue measurable and A is any set, choose a Gδ set G such
that A ⊂ G and m∗ (A) = m(G). Conversely, if E is Carathéodory measurable and
m∗ (E) < ∞, choose a Gδ set G with E ⊂ G and m∗ (E) = m∗ (G). Then G − E
has exterior measure 0.]
7. Exercises 313
4. Let r be a rotation of Rd . Using the fact that the mapping x 7→ r(x) preserves
Lebesgue measure (see Problem 4 in Chapter 2 and Exercise 26 in Chapter 3), show
that it induces a measure-preserving map of the sphere S d−1 with its measure dσ.
A converse is stated in Problem 4.
5. Use the polar coordinate formula to prove the following:

R 2
(a) e−π|x| dx = 1, when d = 2. Deduce from this that the same identity
Rd
holds for all d.
“R ”
∞ −πr 2 d−1
(b) 0
e r dr σ(S d−1 ) = 1, and as a result, σ(S d−1 ) = 2π d/2 /Γ(d/2).
(c) If B is“ the unit ball, vd = m(B) = π d/2 /Γ(d/2 + 1), since this quantity
R 1 d−1 ”
equals 0 r dr σ(S d−1 ). (See Exercise 14 in Chapter 2.)
6. A version of Green’s formula for the unit ball B in Rd can be stated as follows.
Suppose u and v are a pair of functions that are in C 2 (B). Then one has
Z Z „ «
∂u ∂v
(v4u − u4v) dx = v −u dσ.
B S d−1 ∂n ∂n
Here S d−1 is the unit sphere with dσ the measure defined in Section 3.2, and
∂u/∂n, ∂v/∂n denote the directional derivatives of u and v (respectively) along
the inner normals to S d−1 .
Show that the above can be derived from Lemma 4.5 of the previous chapter by
taking η = η²+ and letting ² → 0.
7. There is an alternate version of the mean-value property given in (21) of Chap-

ter 5. It can be stated as follows. Suppose u is harmonic in Ω, and B is any ball
of center x0 and radius r whose closure is contained in Ω. Then
Z
u(x0 ) = c u(x0 + ry) dσ(y), with c−1 = σ(S d−1 ).
S d−1
Conversely, a continuous function satisfying this mean-value property is harmonic.

[Hint: This can be proved as a direct consequence of the corresponding result
for averages over balls (Theorem 4.27 in Chapter 5), or can be deduced from
Exercise 6.]
8. The fact that the Lebesgue measure is uniquely characterized by its translation
invariance can be made precise by the following assertion: If µ is a Borel measure
on Rd that is translation-invariant, and is finite on compact sets, then µ is a
multiple of Lebesgue measure m. Prove this theorem by proceeding as follows.
(a) Suppose Qa denotes a translate of the cube {x : 0 < xj ≤ a, j = 1, 2, . . . , d}

of side length a. If we let µ(Q1 ) = c, then µ(Q1/n ) = cn−d for each integer n.
(b) As a result µ is absolutely continuous with respect to m, and there is a

locally integrable function f such that
Z
µ(E) = f dx.
E
(c) By the differentiation theorem (Corollary 1.7 in Chapter 3) it follows that

f (x) = c a.e., and hence µ = cm.
[Hint: Q1 can be written as a disjoint union of nd translates of Q1/n .]
9. Let C([a, b]) denote the vector space of continuous functions on the closed and
bounded interval [a, b]. Suppose we are given a Borel measure µ on this interval,
with µ([a, b]) < ∞. Then
Z b
f 7→ `(f ) = f (x) dµ(x)
a
is a linear functional on C([a, b]), with ` positive in the sense that `(f ) ≥ 0 if f ≥ 0.
Prove that, conversely, for any linear functional ` on C([a, b]) that is positive in
Rb
the above sense, there is a unique finite Borel measure µ so that `(f ) = a f dµ for
f ∈ C([a, b]).
[Hint: Suppose a = 0 and u ≥ 0. Define F (u) by F (u) = lim²→0 `(f² ), where

1 for 0 ≤ x ≤ u,
f² (x) =
0 for u + ² ≤ x,
and f² is linear between u and u + ². (See Figure 3.) Then F is increasing and
Rb
right-continuous, and `(f ) can be written as a f (x) dF (x) via Theorem 3.5.]
The result also holds if [a, b] is replaced by a closed infinite interval; we then
assume that ` is defined on the continuous functions of bounded support, and
obtain that the resulting µ is finite on all bounded intervals.
A generalization is given in Problem 5.
10. Suppose ν, ν1 , ν2 are signed measures on (X, M) and µ a (positive) measure

on M. Using the symbols ⊥ and ¿ defined in Section 4.2, prove:
(a) If ν1 ⊥ µ and ν2 ⊥ µ, then ν1 + ν2 ⊥ µ.
(b) If ν1 ¿ µ and ν2 ¿ µ, then ν1 + ν2 ¿ µ.
(c) ν1 ⊥ ν2 implies |ν1 | ⊥ |ν2 |.
(d) ν ¿ |ν|.
(e) If ν ⊥ µ and ν ¿ µ, then ν = 0.
11. Suppose that F is an increasing normalized function on R, and let F =

FA + FC + FJ be the decomposition of F in Exercise 24 in Chapter 3; here FA is
7. Exercises 315
1
f²
0 u u+² b
Figure 3. The function f² in Exercise 9
absolutely continuous, FC is continuous with FC0 = 0 a.e, and FJ is a pure jump

function. Let µ = µA + µC + µJ with µ, µA , µC , and µJ the Borel measures
associated to F , FA , FC , and FJ , respectively. Verify that:
(i) µR A is0 absolutely continuous with respect to Lebesgue measure and µA (E) =
E
F (x) dx for every Lebesgue measurable set E.
R R
(ii) RAs a result, if F is absolutely continuous, then f dµ = f dF =
f (x)F 0 (x) dx whenever f and f F 0 are integrable.
(iii) µC + µJ and Lebesgue measure are mutually singular.
12. Suppose Rd − {0} is represented as R+ × S d−1 , with R+ = {0 < r < ∞}.

Then every open set in Rd − {0} can be written as a countable union of open
rectangles of this product.
[Hint: Consider the countable collection of rectangles of the form
{rj < r < rk0 } × {γ ∈ S d−1 : |γ − γ` | < 1/n}.
Here rj and rk0 range over all positive rationals, and {γ` } is a countable dense set
of S d−1 .]
13. Let mj be the Lebesgue measure for the space Rdj , j = 1, 2. Consider the
product Rd = Rd1 × Rd2 (d = d1 + d2 ), with m the Lebesgue measure on Rd . Show
that m is the completion (in the sense of Exercise 2) of the product measure
m1 × m2 .
14. Suppose (Xj , Mj , µj ), 1 ≤ j ≤ k, is a finite collection of measure spaces.

Show that parallel with the case k = 2 considered in Section 3 one can construct
a product measure µ1 × µ2 × · · · × µk on X = X1 × X2 × · · · × Xk . In fact, for
any set EQ⊂ X such that E = E1 × E2 × · · · × Ek , with Ej ⊂ Mj for all j, define
µ0 (E) = kj=1 µj (Ej ). Verify that µ0 extends to a premeasure on the algebra A
of finite disjoint unions of such sets, and then apply Theorem 1.5.
15. The product theory extends to infinitely many factors, under the requisite
assumptions. We consider measure spaces (Xj , Mj , µj ) with µj (Xj ) = 1 for all
but finitely many j. Define a cylinder set E as
{x = (xj ), xj ∈ Ej , Ej ∈ Mj , but Ej = Xj for all but finitely many j}.

Q
For such a set define µ0 (E) = ∞ j=1 µj (Ej ). If A is the algebra generated by the
cylinder sets, µ0 extends to a premeasure on A, and we can apply Theorem 1.5
again.
16. Consider the d-dimensional torus Td = Rd /Zd . Identify Td as T1 × · · · × T1

(d factors) and let µ be the product measure on Td given by µ = µ1 × µ2 × · · · × µd ,
where µj is Lebesgue measure on Xj identified with the circle T. That is, if we
represent each point in Xj uniquely as xj with 0 < xj ≤ 1, then the measure µj is
the induced Lebesgue measure on R1 restricted to (0, 1].
(a) Check that the completion µ is Lebesgue measure induced on the cube
Q = {x : 0 < xj ≤ 1, j = 1, . . . , d}.
(b) For each function f on Q let f˜ be its extension to Rd which is periodic, that
is, f˜(x + z) = f˜(x) for every z ∈ Zd . Then f is measurable on Td if and
only if f˜ is measurable on Rd , and f is continuous on Td if and only if f˜ is
continuous on Rd .
d
(c) Suppose f andR g are integrable on T . Show that the integral defining
(f ∗ g)(x) = Td f (x − y)g(y) dy is finite for a.e. x, that f ∗ g is integrable
over Td , and that f ∗ g = g ∗ f .
(d) For any integrable function f on Td , write

X
f∼ an e2πin·x
n∈Zd
R
to mean that an = Td f (x)e−2πin·x dx. Prove that if g is also integrable,
P
and g ∼ n∈Zd bn e2πin·x , then
X
f ∗g ∼ an bn e2πin·x .
n∈Zd
2πin·x
(e) Verify that {eP }n∈Zd is an orthonormal basis for L2 (Td ). As a result
kf kL2 (Td ) = n∈Zd |an |2 .
(f) Let f be any continuous periodic function on Td . Then f can be uniformly

approximated by finite linear combinations of the exponentials {e2πin·x }n∈Zd .
[Hint: For (e), reduce to the case d = 1 by Fubini’s theorem. To prove (f) let
g(x) = g² (x) = ²−d , if 0 < xj ≤ ², j = 1, . . . , d, and g² (x) = 0 elsewhere
P in Q. Then
(f ∗ gR² )(x) → f (x) uniformly asP² → 0. However (f ∗ g² )(x) = an bn e2πinx with
bn = Td g² (x)e−2πin·x dx, and |an bn | < ∞.]
7. Exercises 317
17. By reducing to the case d = 1, show that each “rotation” x 7→ x + α of the

torus Td = Rd /Zd is measure preserving, for any α ∈ Rd .
18. Suppose τ is a measure-preserving transformation on a measure space (X, µ)

with µ(X) = 1. Recall that a measurable set E is invariant if τ −1 (E) and E differ
by a set of measure zero. A sharper notion is to require that τ −1 (E) equal E.
Prove that if E is any invariant set, there is a set E 0 so that E 0 = τ −1 (E 0 ), and E
and E 0 differ by a set of measure zero.
T “S ”
[Hint: Let E 0 = lim supn→∞ {τ −n (E)} = ∞ n=0 k≥n τ
−k
(E) .]
19. Let τ be a measure-preserving transformation on (X, µ) with µ(X) = 1. Then

τ is ergodic if and only if whenever ν is absolutely continuous with respect to µ and
ν is invariant (that is, ν(τ −1 (E)) = ν(E) for all measurable sets E), then ν = cµ,
with c a constant.
20. Suppose τ is a measure-preserving transformation on (X, µ). If
µ(τ −n (E) ∩ F ) → µ(E)µ(F )
as n → ∞ for all measurable sets E and F , then (T n f, g) → (f, 1)(1, g) whenever

f, g ∈ L2 (X) with (T f )(x) = f (τ (x)). Thus τ is mixing.
[Hint: By linearity the hypothesis implies the conclusion whenever f and g are
simple functions.]
21. Let Td be the torus, and τ : x 7→ x + α the mapping arising in Exercise 17.
Then τ is ergodic if and only if α = (α1 , . . . , αd ) with α1 , α2 , . . . , αd , and 1 are
linearly independent over the rationals. To do this show that:
m−1 Z
1 X
(a) f (τ k (x)) → f (x) dx as m → ∞, for each x ∈ Td , whenever f is
m Td
k=0
continuous and periodic and α satisfies the hypothesis.
(b) Prove as a result that in this case τ is uniquely ergodic.
[Hint: Use (f) in Exercise 16.]
Q
22. Let X = ∞ i=1 Xi , where each (Xi , µi ) is identical to (X1 , µ1 ), with µ1 (X1 ) =
1, and let µ be the corresponding product measure defined in Exercise Q 15. Define
the shift τ : X → X by τ ((x1 , x2 , . . .)) = (x2 , x3 , . . .) for x = (xi ) ∈ ∞
i=1 Xi .
(a) Verify that τ is a measure-preserving transformation.
(b) Prove that τ is ergodic by showing that it is mixing.
(c) Note that in general τ is not uniquely ergodic.
If we define the corresponding shift on the two-sided infinite product, then τ is

also a measure-preserving isomorphism.
[Hint: For (b) note that µ(τ −n (E ∩ F )) = µ(E)µ(F ) whenever E and F are cylin-
der sets and n is sufficiently large. For (c) note that, for example, if we fix a point
x ∈ X1 , the set E = {(xi ) : xj = x all j} is invariant.]
Q
23. Let X = ∞ i=1 Z(2), where each factor is the two-point space Z(2) = {0, 1}
with µ1 (0) = µ1 (1) = 1/2, and suppose µ denotes the product
P measure on X. Con-
aj
sider the mapping D : X → [0, 1] given by D({aj }) → ∞ j=1 2j . Then there are
denumerable sets Z1 ⊂ X and Z2 ⊂ [0, 1], such that:
(a) D is a bijection from X − Z1 to [0, 1] − Z2 .
(b) A set E in X is measurable if and only if D(E) is measurable in [0, 1], and
µ(E) = m(D(E)), where m is Lebesgue measure on [0, 1].
Q
(c) The shift map on ∞ i=1 Z(2) then becomes the doubling map of example (b)
in Section 5.4.
24. Consider the following generalization of the doubling map. For each integer
m, m ≥ 2, we define the map τm of (0, 1] by τ (x) = mx mod 1.
(a) Verify that τ is measure-preserving for Lebesgue measure.

(b) Show that τ is mixing, hence ergodic.
(c) Prove as a consequence that almost every number x is normal in the scale m,
in the following sense. Consider the m-adic expansion of x,
X∞
aj
x= j
, where each aj is an integer 0 ≤ aj ≤ m − 1.
j=1
m
Then x is normal if for each integer k, 0 ≤ k ≤ m − 1,
#{j : aj = k, 1 ≤ j ≤ n} 1
→ as N → ∞.
N m
Note the analogy with the equidistribution statements in Section 2, Chap-
ter 4, of Book I.
25. Show that the mean ergodic theorem still holds if we replace the assumption
that T is an isometry by the assumption that T is a contraction, that is, kT f k ≤
kf k for all f ∈ H.
[Hint: Prove that T is a contraction if and only if T ∗ is a contraction, and use the
identity (f, T ∗ f ) = (T f, f ).]
26. There is an L2 version of the maximal ergodic theorem. Suppose τ is a

measure-preserving transformation on (X, µ). Here we do not assume that µ(X) <
∞. Then
m−1
1 X
f ∗ (x) = sup |f (τ k (x))|
m
k=0
8. Problems 319
satisfies
kf ∗ kL2 (X) ≤ ckf kL2 (X) , whenever f ∈ L2 (X).
The proof is the same as outlined in Problem 6, Chapter 5 for the maximal function
on Rd . With this, extend the pointwise ergodic theorem to the case where µ(X) =
∞, as follows:
1
Pm−1 k
(a) Show that limm→∞ m k=0 f (τ (x)) converges for a.e. x to P (f )(x) for
every f ∈ L (X), because this holds for a dense subspace of L2 (X).
2
(b) Prove that the conclusion holds for every f ∈ L1 (X), because it holds for
the dense subspace L1 (X) ∩ L2 (X).
27. We saw that if kfn kL2 ≤ 1, then fnn(x) → 0 as n → ∞ for a.e. x. However, show
that the analogue where one replaces the L2 -norm by the L1 -norm fails, by con-
structing a sequence {fn }, fn ∈ L1 (X), kfn kL1 ≤ 1, but with lim supn→∞ fnn(x) =
∞ for a.e. x.
[Hint: Find intervals In ⊂ [0, 1], so that m(In ) = 1/(n log n) but lim supn→∞ {In } =
[0, 1]. Then take fn (x) = n log nχIn .]
28. We know by the Borel-Cantelli lemma

P that if {En } is a collection of measurable
sets in a measure a space (X, µ) and ∞ n=1 µ(En ) < ∞ then E = lim supn→∞ {En }
has measure zero.
In the opposite direction, if τ is a mixing
P measure-preserving transformation
on X (with µ(X) = 1), then whenever ∞ n=1 µ(E n ) = ∞, there are integers m =
mn so that if En0 = τ −mn (En ), then lim supn→∞ (En0 ) = X, except for a set of
measure 0.
8 Problems
1. Suppose Φ is a C 1 bijection of an open set O in Rd onto another open set O0

in Rd .
(a) If E is a measurable subset of O, then Φ(E) is also measurable.

R
(b) m(Φ(E)) = E | det Φ0 (x)| dx, where Φ0 is the Jacobian of Φ.
R R
(c) O0 f (y) dy = O f (Φ(x)) | det Φ0 (x)| dx whenever f is integrable on O0 .
[Hint: To prove (a) follow the argument in Exercise

S 8, Chapter 1. For (b) assume
E is a bounded open set, and write E as ∞ j=1 Qj , where Qj are cubes whose
interiors are disjoint, and whose diameters are less than ². Let zk be the center of
Qk . Then if x ∈ Qk ,
Φ(x) = Φ(zk ) + Φ0 (zk )(x − zk ) + o(²),

hence Φ(Qk ) = Φ(zk ) + Φ0 (zk )(Qk − zk ) + o(²), and as a result (1 − η(²))Φ0 (zk )(Qk −
zk ) ⊂ Φ(Qk ) − Φ(zk ) ⊂ (1 + η(²))Φ0 (zk )(Qk − zk ), where η(²) → 0 as ² → 0. This
means that
X X
m(Φ(O)) = m(Φ(Qk )) = | det(Φ0 (zk ))| m(Qk ) + o(1) as ² → 0
k k
on account of the linear transformation property of the Lebesgue measure given in

Problem 4 of Chapter 2. Note that (b) is (c) for f (Φ(x)) = χE (x).]
2. Show as a consequence of the previous problem: the measure dµ = dxdyy2

in the
2
upper half-plane R+ = {z = x + „ iy, y > 0}
« is preserved by any fractional linear
az+b a b
transformation z 7→ cz+d , where belongs to SL2 (R).
c d
3. Let S be a hypersurface in Rd = Rd−1 × R, given by
S = {(x, y) ∈ Rd−1 × R : y = F (x)},
with F a C 1 function defined on an open set Ω in Rd−1 . For each subset E ⊂ Ω

we write Eb for the corresponding subset of S given by E b = {(x, F (x)) x ∈ E}. We
note that the Borel sets of S can be defined in terms of the metric on S (which is
the restriction of the Euclidean metric on Rd ). Thus if E is a Borel set in Ω, then
Eb is a Borel subset of S.
(a) Let µ be the Borel measure on S given by

Z p
b =
µ(E) 1 + |∇F |2 dx.
E
b δ = {(x, y) ∈ Rd , d((x, y), B)

If B is a ball in Ω, let B b < δ}. Show that
b = lim 1 b δ ),
µ(B) m((B)
δ→0 2δ
where m denotes the d-dimensional Lebesgue measure. This result is anal-

ogous to Theorem 4.4 in Chapter 3.
(b) One may apply (a) to the case when S is the (upper) half of the unit sphere
in Rd , given by y = F (x), F (x) = (1 − |x|2 )1/2 , |x| < 1, x ∈ Rd−1 . Show
that in this case dµ = dσ, the measure on the sphere arising in the polar
coordinate formula in Section 3.2.
(c) The above conclusion allows one to write an explicit formula for dσ in
terms of spherical coordinates. Take, for example, the case d = 3, and
write y = cos θ, x = (x1 , x2 ) = (sin θ cos ϕ, sin θ sin ϕ) with 0 ≤ θ < π/2, 0 ≤
ϕ < 2π. Then according to (a) and (b) the element of area dσ equals
(1 − |x|2 )−1/2 dx. Use the change of variable theorem in Problem 1 to deduce
that in this case dσ = sin θ dθ dϕ. This may be generalized to d dimensions,
d ≥ 2, to obtain the formulas in Section 2.4 of the appendix in Book I.
8. Problems 321
4.∗ Let µ be a Borel measure on the sphere S d−1 which is rotation-invariant in the
following sense: µ(r(E)) = µ(E), for every rotation r of Rd and each Borel subset
E of S d−1 . If µ(S d−1 ) < ∞, then µ is a constant multiple of the measure σ arising
in the polar coordinate integration formula.
[Hint: Show that
Z
Yk (x) dµ(x) = 0
S d−1
for every surface spherical harmonic of degree k ≥ 1. As a result, there is a constant

c so that
Z Z
f dµ = c f dσ
S d−1 S d−1
for every continuous function f on S d−1 .]
5.∗ Suppose X is a metric space, and µ is a Borel measure on X with the property
that µ(B) < ∞ for every ball B. Define C0 (X) to be the vector space of continuous
R
functions on X that are each supported in some closed ball. Then `(f ) = X f dµ
defines a linear functional on C0 (X) that is positive, that is, `(f ) ≥ 0 if f ≥ 0.
Conversely, for any positive linear functional ` on C0 (X),Rthere exists a unique
Borel measure µ that is finite on all balls, such that `(f ) = f dµ.
6. Consider an automorphism A of Td = Rd /Zd , that is, A is a linear isomorphism

of Rd that preserves the lattice Zd . Note that A can be written as a d × d matrix
whose entries are integers, with det A = ±1. Define the mapping τ : Td → Td by
τ (x) = A(x).
(a) Observe that τ is a measure-preserving isomorphism of Td .
(b) Show that τ is ergodic (in fact, mixing) if and only if A has no eigenvalues
of the form e2πip/q , where p and q are integers.
(c) Note that τ is never uniquely ergodic.
[Hint: The condition (b) is the same as (At )q has no invariant vectors, where At is
t k
the transpose of A. Note also that f (τ k (x)) = e2πi(A ) (n)·x where f (x) = e2πin·x .]
7.∗ There is a version of the maximal ergodic theorem that is akin to the “rising
sun lemma” and Exercise 6 in Chapter 3. Pm−1
Suppose f is real-valued, and f # (x) = supm m 1 k
k=0 f (τ (x)). Let E0 = {x :
#
f (x) > 0}. Then
Z
f (x) dx ≥ 0.
E0
As a result (when we apply this to f (x) − α), we get when f ≥ 0 that

Z
1
µ{x : f ∗ (x) > α} ≤ f (x) dx.
α {f ∗ (x)>α}
In particular, the constant A in Theorem 5.3 can be taken to be 1.
8. Let X = [0, 1), τ (x) = h1/xi, x 6= 0, τ (0) = 0. Here hxi denotes the fractional
part of x. With the measure dµ = log1 2 1+x
dx
, we have of course µ(X) = 1.
Show that τ is a measure-preserving transformation.
P∞ 1 1
[Hint: k=1 (x+k)(x+k+1) = 1+x .]
9.∗ The transformation τ in the previous problem is ergodic.
10.∗ The connection between continued fractions and the transformation τ (x) =
h1/xi will now be described. A continued fraction, a0 + 1/(a1 + 1/a2 ) · · · , also
written as [a0 a1 a2 · · · ], where the aj are positive integers, can be assigned to any
positive real number x in the following way. Starting with x, we successively
transform it by two alternating operations: reducing it modulo 1 to lie in [0, 1),
and then taking the reciprocal of that number. The integers aj that arise then
define the continued fraction of x.
Thus we set x = a0 + r0 , where a0 = [x] = the greatest integer in x, and r0 ∈
[0, 1). Next we write 1/r0 = a1 + r1 , with a1 = [1/r0 ], r1 ∈ [0, 1), to obtain suc-
cessively 1/rn−1 = an + rn , where an = [1/rn−1 ], rn ∈ [0, 1). If rn = 0 for some n,
we write ak = 0 for all k > n, and say that such a continued fraction terminates.
Note that if 0 ≤ x < 1, then r0 = x and a1 = [1/x], while r1 = h1/xi = τ (x).
More generally then, ak (x) = [1/τ k−1 (x)] = a1 τ k−1 (x). The following properties
of continued fractions of positive real numbers x are known:
(a) The continued fraction of x terminates if and only if x is rational.
(b) If x = [a0 a1 · · · an · · · ], and xN = [a0 a1 · · · aN 00 · · · ], then xN → x as N →

∞. The sequence {xN } gives essentially an optimal approximation of x by
rationals.
(c) The continued fraction is periodic, that is, ak+N = ak for some N ≥ 1, and
all sufficiently large k, if and only if x is an algebraic number of degree ≤ 2
over the rationals.
(d) One can conclude that a1 +a2 +···+a

n
n
→ ∞ as n → ∞ for almost every x. In
particular, the set of numbers x whose continued fractions [a0 a1 · · · an · · · ]
are bounded has measure zero.
[Hint: For (d) apply a consequence

R of the pointwise ergodic theorem,
Pm−1 which is as
1 k
follows: Suppose f ≥ 0, and f dµ = ∞. If τ is ergodic, then m k=0 f (τ (x)) →
∞ for a.e. x as m → ∞. In the present case take f (x) = [1/x].]
7 Hausdorff
Fractals
Measure and
Carathéodory developed a remarkably simple general-

ization of Lebesgue’s measure theory which in particu-
lar allowed him to define the p-dimensional measure of
a set in q-dimensional space. In what follows, I present
a small addition.... a clarification of p-dimensional
measure that leads immediately to an extension to
non-integral p, and thus gives rise to sets of fractional
dimension.
F. Hausdorff, 1919
I coined fractal from the Latin adjective fractus. The

corresponding Latin verb frangere means to “break”:
to create irregular fragments.
B. Mandelbrot, 1977
The deeper study of the geometric properties of sets often requires

an analysis of their extent or “mass” that goes beyond what can be
expressed in terms of Lebesgue measure. It is here that the notions
of the dimension of a set (which can be fractional) and an associated
measure play a crucial role.
Two initial ideas may help to provide an intuitive grasp of the concept
of the dimension of a set. The first can be understood in terms of how
the set replicates under scalings. Given the set E, let us suppose that
for some positive number n we have that nE = E1 ∪ · · · ∪ Em , where the
sets Ej are m essentially disjoint congruent copies of E. Note that if
E were a line segment this would hold with m = n; if E were a square,
we would have m = n2 ; if E were a cube, then m = n3 ; etc. Thus, more
generally, we might be tempted to say that E has dimension α if m = nα .
Observe that if E is the Cantor set C in [0, 1], then 3C consists of 2 copies
of C, one in [0, 1] and the other in [2, 3]. Here n = 3, m = 2, and we would
be led to conclude that log 2/ log 3 is the dimension of the Cantor set.
324 Chapter 7. HAUSDORFF MEASURE AND FRACTALS
Another approach is relevant for curves that are not necessarily rec-
tifiable. Start with a curve Γ = {γ(t) : a ≤ t ≤ b}, and for each ² > 0
consider polygonal lines joining γ(a) to γ(b), whose vertices lie on suc-
cessive points of Γ, with each segment not exceeding ² in length. Denote
by #(²) the least number of segments that arise for such polygonal lines.
If #(²) ≈ ²−1 as ² → 0, then Γ is rectifiable. However, #(²) may well
grow more rapidly than ²−1 as ² → 0. If we had #(²) ≈ ²−α , 1 < α,
then, in the spirit of the previous example, it would be natural to say
that Γ has dimension α. These considerations have even an interest in
other parts of science. For instance, in studying the question of determin-
ing the length of the border of a country or its coastline, L.F. Richardson
found that the length of the west coast of Britain obeyed the empirical
law #(²) ≈ ²−α , with α approximately 1.5. Thus one might conclude
that the coast has fractional dimension!
While there are a number of different ways to make some of these
heuristic notions precise, the theory that has the widest scope and great-
est flexibility is the one involving Hausdorff measure and Hausdorff di-
mension. Probably the most elegant and simplest illustration of this
theory can be seen in terms of its application to a general class of self-
similar sets, and this is what we consider first. Among these are the
curves of von Koch type, and these can have any dimension between 1
and 2.
Next, we turn to an example of a space-filling curve, which, broadly
speaking, falls under the scope of self-replicating constructions. Not
only does this curve have an intrinsic interest, but its nature reveals the
important fact that from the point of view of measure theory the unit
interval and the unit square are the same.
Our final topic is of a somewhat different nature. It begins with the
realization of an unexpected regularity that all subsets of Rd (of finite
Lebesgue measure) enjoy, when d ≥ 3. This property fails in two di-
mensions, and the key counter-example is the Besicovitch set. This set
appears also in a number of other problems. While it has measure zero,
this is barely so, since its Hausdorff dimension is necessarily 2.
1 Hausdorff measure
The theory begins with the introduction of a new notion of volume or
mass. This “measure” is closely tied with the idea of dimension which
prevails throughout the subject. More precisely, following Hausdorff,
one considers for each appropriate set E and each α > 0 the quantity
mα (E), which can be interpreted as the α-dimensional mass of E among
sets of dimension α, where the word “dimension” carries (for now) only
1. Hausdorff measure 325
an intuitive meaning. Then, if α is larger than the dimension of the set

E, the set has a negligible mass, and we have mα (E) = 0. If α is smaller
than the dimension of E, then E is very large (comparatively), hence
mα (E) = ∞. For the critical case when α is the dimension of E, the
quantity mα (E) describes the actual α-dimensional size of the set.
Two examples, to which we shall return in more detail later, illustrate
this circle of ideas.
First, recall that the standard Cantor set C in [0, 1] has zero Lebesgue
measure. This statement expresses the fact that C has one-dimensional
mass or length equal to zero. However, we shall prove that C has a
well-defined fractional Hausdorff dimension of log 2/ log 3, and that the
corresponding Hausdorff measure of the Cantor set is positive and finite.
Another illustration of the theory developed below consists of starting
with Γ, a rectifiable curve in the plane. Then Γ has zero two-dimensional
Lebesgue measure. This is intuitively clear, since Γ is a one-dimensional
object in a two-dimensional space. This is where the Hausdorff measure
comes into play: the quantity m1 (Γ) is not only finite, but precisely equal
to the length of Γ as we defined it in Section 3.1 of Chapter 3.
We first consider the relevant exterior measure, defined in terms of
coverings, whose restriction to the Borel sets is the desired Hausdorff
measure.
For any subset E of Rd , we define the exterior α-dimensional Haus-
dorff measure of E by
( ∞
)
X [
m∗α (E) = lim inf (diam Fk )α : E ⊂ Fk , diam Fk ≤ δ all k ,
δ→0
k k=1
where diam S denotes the diameter of the set S, that is, diam S =
sup{|x − y| : x, y ∈ S}. In other words, for each δ > 0 we consider covers
of E by countable families of (arbitrary)
P sets with diameter less than δ,
and take the infimum of the sum k (diam Fk )α . We then define m∗α (E)
as the limit of these infimums as δ tends to 0. We note that the quantity
( ∞
)
X [
δ α
Hα (E) = inf (diam Fk ) : E ⊂ Fk , diam Fk ≤ δ all k
k k=1
is increasing as δ decreases, so that the limit

δ
m∗α (E) = lim Hα (E)
δ→0
exists, although m∗α (E) could be infinite. We note that in particu-

δ ∗
lar, one has Hα (E) ≤ mα (E) for all δ > 0. When defining the exte-
rior measure m∗α (E) it is important to require that the coverings be of
sets of arbitrarily small diameters; this is the thrust of the definition

m∗α (E) = limδ→0 Hα δ
(E). This requirement, which is not relevant for
Lebesgue measure, is needed to ensure the basic additive feature stated
in Property 3 below. (See also Exercise 12.)
Scaling is the key notion that appears at the heart of the definition of
the exterior Hausdorff measure. Loosely speaking, the measure of a set
scales according to its dimension. For instance, if Γ is a one-dimensional
subset of Rd , say a smooth curve of length L, then rΓ has total length
rL. If Q is a cube in Rd , the volume of rQ is rd |Q|. This feature is
captured in the definition of exterior Hausdorff measure by the fact that
if the set F is scaled by r, then (diam F )α scales by rα . This key idea
reappears in the study of self-similar sets in Section 2.2.
We begin with a list of properties satisfied by the Hausdorff exterior
measure.
Property 1 (Monotonicity) If E1 ⊂ E2 , then m∗α (E1 ) ≤ m∗α (E2 ).
This is straightforward, since any cover of E2 is also a cover of E1 .

S∞ P∞
Property 2 (Sub-additivity) m∗α ( j=1 Ej ) ≤ j=1 m∗α (Ej ) for any
countable family {Ej } of sets in Rd .
∞
For the proof, fix δ, and choose for each
P j a cover α{Fj,k }δk=1 of Ej by
diameter less than δ such that k (diam Fj,k ) ≤ Hα (Ej ) + ²/2j .
sets ofS
Since j,k Fj,k is a cover of E by sets of diameter less than δ, we must
have
∞
X
δ δ
Hα (E) ≤ Hα (Ej ) + ²
j=1
X∞
≤ m∗α (Ej ) + ².
j=1
δ
P ∗
Since ² is arbitrary, the inequality Hα (E) ≤ mα (Ej ) holds, and we let
δ tend to 0 to prove the countable sub-additivity of m∗α .
Property 3 If d(E1 , E2 ) > 0, then m∗α (E1 ∪ E2 ) = m∗α (E1 ) + m∗α (E2 ).
It suffices to prove that m∗α (E1 ∪ E2 ) ≥ m∗α (E1 ) + m∗α (E2 ) since the re-
verse inequality is guaranteed by sub-additivity. Fix ² > 0 with ² <
d(E1 , E2 ). Given any cover of E1 ∪ E2 with sets F1 , F2 . . . , of diame-
ter less than δ, where δ < ², we let
Fj0 = E1 ∩ Fj and Fj00 = E2 ∩ Fj .

1. Hausdorff measure 327
Then {Fj0 } and {Fj00 } are covers for E1 and E2 , respectively, and are
disjoint. Hence,
X X X
(diam Fj0 )α + (diam Fi00 )α ≤ (diam Fk )α .
j i k
Taking the infimum over the coverings, and then letting δ tend to zero
yields the desired inequality.
At this point, we note that m∗α satisfies all the properties of a metric
Carathéodory exterior measure as discussed in Chapter 6. Thus m∗α
is a countably additive measure when restricted to the Borel sets. We
shall therefore restrict ourselves to Borel sets and write mα (E) instead
of m∗α (E). The measure mα is called the α-dimensional Hausdorff
measure.
Property
S∞ 4 If {Ej } is a countable family of disjoint Borel sets, and
E = j=1 Ej , then
∞
X
mα (E) = mα (Ej ).
j=1
For what follows in this chapter, the full additivity in the above prop-
erty is not needed, and we can manage with a weaker form whose proof
is elementary and not dependent on the developments of Chapter 6. (See
Exercise 2.)
Property 5 Hausdorff measure is invariant under translations
mα (E + h) = mα (E) for all h ∈ Rd ,
and rotations
mα (rE) = mα (E),
where r is a rotation in Rd .
Moreover, it scales as follows:
mα (λE) = λα mα (E) for all λ > 0.
These conclusions follow once we observe that the diameter of a set S

is invariant under translations and rotations, and satisfies diam(λS) =
λ diam(S) for λ > 0.
We describe next a series of properties of Hausdorff measure, the first
of which is immediate from the definitions.
Property 6 The quantity m0 (E) counts the number of points in E,

while m1 (E) = m(E) for all Borel sets E ⊂ R. (Here m denotes the
Lebesgue measure on R.)
In fact, note that in one dimension every set of diameter δ is contained in

an interval of length δ (and for an interval its length equals its Lebesgue
measure).
In general, d-dimensional Hausdorff measure in Rd is, up to a constant
factor, equal to Lebesgue measure.
Property 7 If E is a Borel subset of Rd , then cd md (E) = m(E) for

some constant cd that depends only on the dimension d.
The constant cd equals m(B)/(diam B)d , for the unit ball B; note that
this ratio is the same for all balls B in Rd , and so cd = vd /2d (where vd
denotes the volume of the unit ball). The proof of this property relies on
the so-called iso-diametric inequality, which states that among all sets of
a given diameter, the ball has largest volume. (See Problem 2.) Without
using this geometric fact one can prove the following substitute.
Property 7 0 If E is a Borel subset of Rd and m(E) is its Lebesgue

measure, then md (E) ≈ m(E), in the sense that
cd md (E) ≤ m(E) ≤ 2d cd md (E).
Using Exercise 26 in Chapter 3 we can find for every

P ², δ > 0, a covering
of E by balls {Bj }, such that diam Bj < δ, while j m(Bj ) ≤ m(E) + ².
Now,
X X
Hdδ (E) ≤ (diam Bj )d = c−1
d m(Bj ) ≤ c−1
d (m(E) + ²).
j j
−1
Letting δ and ² tend
S to 0, we get md (E) ≤ Pcd m(E). For the reverse
d
direction, let E ⊂ j Fj be a covering with j (diam Fj ) ≤ md (E) + ².
We can always find closed balls Bj centered at a point
P of Fj so that
Bj ⊃SFj and diam Bj = 2 diam Fj . However, m(E) ≤ j m(Bj ), since
E ⊂ j Bj , and the last sum equals
X X
cd (diam Bj )d = 2d cd (diam Fj )d ≤ 2d cd (md (E) + ²) .
Letting ² → 0 gives m(E) ≤ 2d cd md (E).
Property 8 If m∗α (E) < ∞ and β > α, then m∗β (E) = 0. Also, if m∗α (E) >
0 and β < α, then m∗β (E) = ∞.
2. Hausdorff dimension 329
Indeed, if diam F ≤ δ, and β > α, then
(diam F )β = (diam F )β−α (diam F )α ≤ δ β−α (diam F )α .
Consequently
Hβδ (E) ≤ δ β−α Hα

δ
(E) ≤ δ β−α m∗α (E).
Since m∗α (E) < ∞ and β − α > 0, we find in the limit as δ tends to 0,
that m∗β (E) = 0.
The contrapositive gives m∗β (E) = ∞ whenever m∗α (E) > 0 and β < α.
We now make some easy observations that are consequences of the
above properties.
1. If I is a finite line segment in Rd , then 0 < m1 (I) < ∞.
2. More generally, if Q is a k-cube in Rd (that is, Q is the product of
k non-trivial intervals and d − k points), then 0 < mk (Q) < ∞.
3. If O is a non-empty open set in Rd , then mα (O) = ∞ whenever
α < d. Indeed, this follows because md (O) > 0.
4. Note that we can always take α ≤ d. This is because when α > d,
mα vanishes on every ball, and hence on all of Rd .
2 Hausdorff dimension
Given a Borel subset E of Rd , we deduce from Property 8 that there
exists a unique α such that
½
∞ if β < α,
mβ (E) =
0 if α < β.
In other words, α is given by
α = sup{β : mβ (E) = ∞} = inf{β : mβ (E) = 0}.
We say that E has Hausdorff dimension α, or more succinctly, that

E has dimension α. We shall write α = dim E. At the critical value α
we can say no more than that in general the quantity mα (E) satisfies
0 ≤ mα (E) ≤ ∞. If E is bounded and the inequalities are strict, that is,
0 < mα (E) < ∞, we say that E has strict Hausdorff dimension α.
The term fractal is commonly applied to sets of fractional dimension.
In general, calculating the Hausdorff measure of a set is a difficult
problem. However, it is possible in some cases to bound this measure
from above and below, and hence determine the dimension of the set in
question. A few examples will illustrate these new concepts.
2.1 Examples
The Cantor set
The first striking example consists of the Cantor set C, which was con-
structed in Chapter 1 by successively removing the middle-third intervals
in [0, 1].
Theorem 2.1 The Cantor set C has strict Hausdorff dimension α =

log 2/ log 3.
The inequality
mα (C) ≤ 1
follows from the construction

T of C and the definitions. Indeed, recall from
Chapter 1 that C = Ck , where each Ck is a finite union of 2k intervals of
length 3−k . Given δ > 0, we first choose K so large that 3−K < δ. Since
the set CK covers C and consists of 2K intervals of diameter 3−K < δ,
we must have
δ
Hα (C) ≤ 2K (3−K )α .
However, α satisfies precisely 3α = 2, hence 2K (3−K )α = 1, and therefore

mα (C) ≤ 1.
The reverse inequality, which consists of proving that 0 < mα (C), re-
quires a further idea. Here we rely on the Cantor-Lebesgue function,
which maps C surjectively onto [0, 1]. The key fact we shall use about
this function is that it satisfies a precise continuity condition that reflects
the dimension of the Cantor set.
A function f defined on a subset E of Rd satisfies a Lipschitz con-
dition on E if there exists M > 0 such that
|f (x) − f (y)| ≤ M |x − y| for all x, y ∈ E.
More generally, a function f satisfies a Lipschitz condition with ex-

ponent γ (or is Hölder γ) if
|f (x) − f (y)| ≤ M |x − y|γ for all x, y ∈ E.
The only interesting case is when 0 < γ ≤ 1. (See Exercise 3.)
Lemma 2.2 Suppose a function f defined on a compact set E satisfies

a Lipschitz condition with exponent γ. Then
(i) mβ (f (E)) ≤ M β mα (E) if β = α/γ.
1
(ii) dim f (E) ≤ γ dim E.
Proof. Suppose {Fk } is a countable family of sets that covers E.

Then {f (E ∩ Fk )} covers f (E) and, moreover, f (E ∩ Fk ) has diameter
less than M (diam Fk )γ . Hence
X X
(diam f (E ∩ Fk ))α/γ ≤ M α/γ (diam Fk )α ,
k k
and part (i) follows. This result now immediately implies conclusion (ii).
Lemma 2.3 The Cantor-Lebesgue function F on C satisfies a Lipschitz

condition with exponent γ = log 2/ log 3.
Proof. The function F was constructed in Section 3.1 of Chapter 3 as

the limit of a sequence {Fn } of piecewise linear functions. The function
Fn increases by at most 2−n on each interval of length 3−n . So the slope
of Fn is always bounded by (3/2)n , and hence
µ ¶n
3
|Fn (x) − Fn (y)| ≤ |x − y|.
2
Moreover, the approximating sequence also satisfies |F (x) − Fn (x)| ≤

1/2n . These two estimates together with an application of the triangle
inequality give
|F (x) − F (y)| ≤ |Fn (x) − Fn (y)| + |F (x) − Fn (x)| + |F (y) − Fn (y)|

µ ¶n
3 2
≤ |x − y| + n .
2 2
Having fixed x and y, we then minimize the right hand side by choosing
n so that both terms have the same order of magnitude. This is achieved
by taking n so that 3n |x − y| is between 1 and 3. Then, we see that
|F (x) − F (y)| ≤ c2−n = c(3−n )γ ≤ M |x − y|γ ,
since 3γ = 2 and 3−n is not greater than |x − y|. This argument is re-
peated in Lemma 2.8 below.
With E = C, f the Cantor-Lebesgue function, and α = γ = log 2/ log 3,
the two lemmas give
m1 ([0, 1]) ≤ M β mα (C).

Thus mα (C) > 0, and we find that dim C = log 2/ log 3.

The proof of this example is typical in the sense that the inequal-
ity mα (C) < ∞ is usually easier to obtain than 0 < mα (C). Also, with
some extra effort, it is possible to show that the log 2/ log 3-dimensional
Hausdorff measure of C is precisely 1. (See Exercise 7.)
Rectifiable curves
A further example of the role of dimension comes from looking at con-
tinuous curves in Rd . Recall that a continuous curve γ : [a, b] → Rd is
said to be simple if γ(t1 ) 6= γ(t2 ) whenever t1 6= t2 , and quasi-simple
if the mapping t 7→ z(t) is injective for t in the complement of finitely
many points.
Theorem 2.4 Suppose the curve γ is continuous and quasi-simple. Then

γ is rectifiable if and only if Γ = {γ(t) : a ≤ t ≤ b} has strict Hausdorff
dimension one. Moreover, in this case the length of the curve is precisely
its one-dimensional measure m1 (Γ).
Proof. Suppose to begin with that Γ is a rectifiable curve of length L,

and consider an arc-length parametrization γ̃ such that Γ = {γ̃(t) : 0 ≤
t ≤ L}. This parametrization satisfies the Lipschitz condition
|γ̃(t1 ) − γ̃(t2 )| ≤ |t1 − t2 |.
This follows since |t1 − t2 | is the length of the curve between t1 and t2 ,
which is greater than the distance from γ̃(t1 ) to γ̃(t2 ). Since γ̃ satisfies
the conditions of Lemma 2.2 with exponent 1 and M = 1, we find that
m1 (Γ) ≤ L.
To prove the reverse inequality, we let a = t0 < t1 < · · · < tN = b denote

a partition of [a, b] and let
Γj = {γ(t) : tj ≤ t ≤ tj+1 },
SN −1
so that Γ = j=0 Γj , and hence
N
X −1
m1 (Γ) = m1 (Γj )
j=0
by an application of Property 4 of the Hausdorff measure and the fact

that Γ is quasi-simple. Indeed, by removing finitely many points the
SN −1
union j=0 Γj becomes disjoint, while the points removed clearly have
zero m1 -measure. We next claim that m1 (Γj ) ≥ `j , where `j is the dis-
tance from γ(tj ) to γ(tj+1 ), that is, `j = |γ(tj+1 ) − γ(tj )|. To see this,
recall that Hausdorff measure is rotation-invariant, and introduce new or-
thogonal coordinates x and y such that [γ(tj ), γ(tj + 1)] is the segment
[0, `j ] on the x-axis. The projection π(x, y) = x satisfies the Lipschitz
condition
|π(P ) − π(Q)| ≤ |P − Q|,
and clearly the segment [0, `j ] on the x-axis is contained in the image
π(Γj ). Therefore, Lemma 2.2 guarantees
`j ≤ m1 (Γj ),
P
and thus m1 (Γ) ≥ `j .PSince by definition the length L of Γ is the
supremum of the sums `j over all partitions of [a, b], we find that
m1 (Γ) ≥ L, as desired.
Conversely, if Γ has strict Hausdorff dimension 1, then m1 (Γ) < ∞,
and the above argument shows that Γ is rectifiable.
The reader may note the resemblance of this characterization of rec-
tifiability and an earlier one in terms of Minkowski content, given in
Chapter 3. In this connection we point out that there is a different
notion of dimension that is sometimes used instead of Hausdorff dimen-
sion. For a compact set E, this dimension is given in terms of the size
of E δ = {x ∈ Rd : d(x, E) < δ} as δ → 0. One observes that if E is a
k-dimensional cube in Rd , then m(E δ ) ≤ cδ d−k as δ → 0, with m the
Lebesgue measure of Rd . With this in mind, the Minkowski dimen-
sion of E is defined by
inf {β : m(E δ ) = O(δ d−β ) as δ → 0}.
One can show that the Hausdorff dimension of a set does not exceed its
Minkowski dimension, but that equality does not hold in general. More
details may be found in Exercises 17 and 18.
The Sierpinski triangle

A Cantor-like set can be constructed in the plane as follows. We begin
with a (solid) closed equilateral triangle S0 , whose sides have unit length.
Then, as a first step we remove the shaded open equilateral triangle
pictured in Figure 1.
Figure 1. Construction of the Sierpinski triangle
This leaves three closed triangles whose union we denote by S1 . Each

triangle is half the size of the original (or parent) triangle S0 , and these
smaller closed triangles are said to be of the first generation: the trian-
gles in S1 are the children of the parent S0 . In the second step, we repeat
the process in each triangle of the first generation. Each such triangle
has three children of the second generation. We denote by S2 the union
of the three triangles in the second generation. We then repeat this pro-
cess to find a sequence Sk of compact sets which satisfy the following
properties:
(a) Each Sk is a union of 3k closed equilateral triangles of side length

2−k . (These are the triangles of the k th generation.)
(b) {Sk } is a decreasing sequence of compact sets; that is, Sk+1 ⊂ Sk
for all k ≥ 0.
The Sierpinski triangle is the compact set defined by

∞
\
S= Sk .
k=0
Theorem 2.5 The Sierpinski triangle S has strict Hausdorff dimension

α = log 3/ log 2.
The inequality mα (S) ≤ 1 follows immediately from the construction.

Given δ > 0, choose K so that 2−K < δ. Since the set SK covers S and
consists of 3K triangles each of diameter 2−K < δ, we must have
δ
Hα (S) ≤ 3K (2−K )α .
But since 2α = 3, we find Hαδ

(S) ≤ 1, hence mα (S) ≤ 1.
The inequality mα (S) > 0 is more subtle. For its proof we need to fix
a special point in each triangle that appears in the construction of S.
We choose to call the lower left vertex of a triangle the vertex of that
triangle. With this choice there are 3k vertices of the k th generation.
The argument that follows is based on the important fact that all these
vertices belong to
S∞ S.
Suppose S ⊂ j=1 Fj , with diam Fj < δ. We wish to prove that
X
(diam Fj )α ≥ c > 0
j
for some constant c. Clearly, each Fj is contained in a ball of twice the

diameter of Fj , so upon replacing 2δ by δ and noting that S is compact,
SN
it suffices to show that if S ⊂ j=1 Bj , where B = {Bj }N j=1 is a finite
collection of balls whose diameters are less than δ, then
N
X
(diam Bj )α ≥ c > 0.
j=1
Suppose we have such a covering by balls. Consider the minimum diam-

eter of the Bj , and choose k so that
2−k ≤ min diam Bj < 2−k+1 .

1≤j≤N
Lemma 2.6 Suppose B is a ball in the covering B that satisfies
2−` ≤ diam B < 2−`+1 for some ` ≤ k.
Then B contains at most c3k−` vertices of the k th generation.
In this chapter, we shall continue use the common practice of denoting

by c, c0 , . . . generic constants whose values are unimportant and may
change from one usage to another. We also use A ≈ B to denote that
the quantities A and B are comparable, that is, cB ≤ A ≤ c0 B, for
appropriate constants c and c0 .
Proof of Lemma 2.6. Let B ∗ denote the ball with same center as B but
three times its diameter, and let 4k be a triangle of the k th generation
whose vertex v lies in B. If 40` denotes the triangle of the `th generation
that contains 4k , then since diam B ≥ 2−` ,
v ∈ 4k ⊂ 40` ⊂ B ∗ ,
as shown in Figure 2.
Next, there is a positive constant c such that B ∗ can contain at most
c distinct triangles of the `th generation. This is because triangles of the
4k
40`
v
B B∗
Figure 2. The setting in Lemma 2.6
`th generation have disjoint interiors and area equal to c0 4−` , while B ∗
has area at most equal to c00 4−` . Finally, each 40` contains 3k−` triangles
of the k th generation, hence B can contain at most c3k−` vertices of
triangles of the k th generation.
PN
To complete the proof that j=1 (diam Bj )α ≥ c > 0, note that
N
X X
(diam Bj )α ≥ N` 2−`α ,
j=1 `
where N` denotes the number of balls in B that satisfy 2−` ≤ diam Bj ≤

2−`+1 . By the lemma, we see that the total number of vertices of triangles
in the k th generation
P that can be covered by the collection B can be no
more than c ` N` 3 . Since all 3k vertices of triangles in the k th
k−`
generation belong to S, P and all vertices of the k th generation must be

covered, we must have c ` N` 3k−` ≥ 3k . Hence
X
N` 3−` ≥ c.
`
It now suffices to recall the definition of α which guarantees 2−`α = 3−` ,

and therefore
N
X
(diam Bj )α ≥ c,
j=1
as desired.
We give a final example that exhibits properties similar to the Cantor
set and Sierpinski triangle. It is the curve discovered by von Koch in 1904.
The von Koch curve

Consider the unit interval K0 = [0, 1], which we may think of as lying
on the x-axis in the xy-plane. Then consider the polygonal path K1
illustrated in Figure 3, which consists of four equal line segments of
length 1/3.
K0
K1
K2
K3
Figure 3. The first few stages in the construction of the von Koch curve
Let K1 (t), for 0 ≤ t ≤ 1, denote the parametrization of K1 that has

constant speed. In other words, as t travels from 0 to 1/4, the point
K1 (t) travels on the first line segment. As t travels from 1/4 to 1/2, the
point K1 (t) travels on the second line segment, and so on. In particular,
we see that K1 (`/4) for 0 ≤ ` ≤ 4 correspond to the five vertices of K1 .
At the second stage of the construction we repeat the process of re-
placing each line segment in stage one by the corresponding polygonal
line. We then obtain the polygonal curve K2 illustrated in Figure 3. It
has 16 = 42 segments of length 1/9 = 3−2 . We choose a parametrization
K2 (t) (0 ≤ t ≤ 1) of K2 that has constant speed. Observe that K2 (`/42 )

for 0 ≤ ` ≤ 42 gives all vertices of K2 , and that the vertices of K1 belong
to K2 , with
K2 (`/4) = K1 (`/4) for 0 ≤ ` ≤ 4.
Repeating this process indefinitely, we obtain a sequence of continuous

polygonal curves {Kj }, where Kj consists of 4j segments of length 3−j
each. If Kj (t) (0 ≤ t ≤ 1) is the parametrization of Kj that has constant
speed, then the vertices are precisely at the points Kj (`/4j ), and
Kj 0 (`/4j ) = Kj (`/4j ) for 0 ≤ ` ≤ 4j
whenever j 0 ≥ j.
In the limit as j tends to infinity, the polygonal lines Kj tend to the
von Koch curve K. Indeed, we have
|Kj+1 (t) − Kj (t)| ≤ 3−j for all 0 ≤ t ≤ 1 and j ≥ 0.
This is clear when j = 0, and follows by induction in j when we consider

the nature of the construction of the j th stage. Since we may write
J−1
X
KJ (t) = K1 (t) + (Kj+1 (t) − Kj (t)),
j=1
the above estimate proves that the series

∞
X
K1 (t) + (Kj+1 (t) − Kj (t))
j=1
converges absolutely and uniformly to a continuous function K(t) that is

a parametrization of K. Besides continuity, the function K(t) satisfies a
regularity assumption that takes the form of a Lipschitz condition, as in
the case of the Cantor-Lebesgue function.
Theorem 2.7 The function K(t) satisfies a Lipschitz condition of expo-
nent γ = log 3/ log 4, that is:
|K(t) − K(s)| ≤ M |t − s|γ for all t, s ∈ [0, 1].
We have already observed that |Kj+1 (t) − Kj (t)| ≤ 3−j . Since Kj travels
a distance of 3−j in 4−j units of time, we see that
µ ¶j
4
|Kj0 (t)| ≤ except when t = `/4j .
3
Consequently we must have

µ ¶j
4
|Kj (t) − Kj (s)| ≤ |t − s|.
3
P∞
Moreover, K(t) = K1 (t) + j=1 (Kj+1 (t) − Kj (t)). We now find our-
selves in precisely the same situation as in the proof that the Cantor-
Lebesgue function satisfies a Lipschitz condition with exponent log 2/ log 3.
We generalize that argument in the following lemma.
Lemma 2.8 Suppose {fj } is a sequence of continuous functions on the

interval [0, 1] that satisfy
|fj (t) − fj (s)| ≤ Aj |t − s| for some A > 1,
and
|fj (t) − fj+1 (t)| ≤ B −j for some B > 1.
Then the limit f (t) = limj→∞ fj (t) exists and satisfies
|f (t) − f (s)| ≤ M |t − s|γ ,
where γ = log B/ log(AB).
Proof. The continuous limit f is given by the uniformly convergent

series
∞
X
f (t) = f1 (t) + (fk+1 (t) − fk (t)),
k=1
and therefore
∞
X ∞
X
|f (t) − fj (t)| ≤ |fk+1 (t) − fk (t)| ≤ B −k ≤ cB −j .
k=j k=j
The triangle inequality, an application of the inequality just obtained,

and the inequality in the statement of the lemma give
|f (t) − f (s)| ≤ |fj (t) − fj (s)| + |(f − fj )(t)| + |(f − fj )(s)|

≤ c(Aj |t − s| + B −j ).
For a fixed pair of numbers t and s with t 6= s, we choose j to minimize

the sum Aj |t − s| + B −j . This is essentially achieved by picking j so that
two terms Aj |t − s| and B −j are comparable. More precisely, we choose

a j that satisfies
(AB)j |t − s| ≤ 1 and 1 ≤ (AB)j+1 |t − s|.
Since |t − s| ≤ 2 and AB > 1, such a j must exist. The first inequality

then gives
Aj |t − s| ≤ B −j ,
while raising the second inequality to the power γ, and using the fact
that (AB)γ = B gives
1 ≤ B j |t − s|γ .
Thus B −j ≤ |t − s|γ , and consequently
|f (t) − f (s)| ≤ c(Aj |t − s| + B −j ) ≤ M |t − s|γ ,
as was to be shown.
In particular, this result with Lemma 2.2 implies that
1 log 4
dim K ≤ = .
γ log 3
To prove that mγ (K) > 0 and hence dim K = log 4/ log 3 requires an ar-
gument similar to the one given for the Sierpinski triangle. In fact,
this argument generalizes to cover a general family of sets that have a
self-similarity property. We therefore turn our attention to this general
theory next.
Remarks. We mention some further facts about the von Koch curve.
More details can be found in Exercises 13, 14, and 15 below.
1. The curve K is one in a family of similarly constructed curves. For

each `, 1/4 < ` < 1/2, consider at the first stage the curve K1` (t)
given by four line segments each of length `, the first and last on the
x-axis, and the second and third forming two sides of an isoceles
triangle whose base lies on the x-axis. (See Figure 4.) The case
` = 1/3 corresponds to the previously defined von Koch curve.
Proceeding as in the case ` = 1/3, one obtains a curve K` , and it
can be seen that
log 4
dim(K` ) = .
log 1/`
` `
` `
Figure 4. The curve K1` (t)
Thus for every α, 1 < α < 2, we have a curve of this kind of dimen-
sion α. Note that when ` → 1/4 the limiting curve is a straight line
segment, which has dimension 1. When ` → 1/2, the limit can be
seen to correspond to a “space-filling” curve.
2. The curves t 7→ K` (t), 1/4 < ` ≤ 1/2, are each nowhere differen-
tiable. One can also show that each curve is simple when 1/4 ≤
` < 1/2.
2.2 Self-similarity
The Cantor set C, the Sierpinski triangle S, and von Koch curve K all
share an important property: each of these sets contains scaled copies
of itself. Moreover, each of these examples was constructed by iterating
a process closely tied to its scaling. For instance, the interval [0, 1/3]
contains a copy of the Cantor set scaled by a factor of 1/3. The same is
true for the interval [2/3, 1], and therefore
C = C1 ∪ C 2 ,
where C1 and C2 are scaled versions of C. Also, each interval [0, 1/9],
[2/9, 3/9], [6/9, 7/9] and [8/9, 1] contains a copy of C scaled by a factor
of 1/9, and so on.
In the case of the Sierpinski triangle, each of the three triangles in the
first generation contains a copy of S scaled by the factor of 1/2. Hence
S = S1 ∪ S 2 ∪ S 3 ,
where each Sj , j = 1, 2, 3, is obtained by scaling and translating the

original Sierpinski triangle. More generally, every triangle in the k th
generation is a copy of S scaled by the factor of 1/2k .
Finally, each line segment in the initial stage of the construction of the
von Koch curve gives rise to a scaled and possibly rotated copy of the
von Koch curve. In fact
K = K1 ∪ K2 ∪ K3 ∪ K4 ,
where Kj , j = 1, 2, 3, 4, is obtained by scaling K by the factor of 1/3 and

translating and rotating it.
Thus these examples each contain replicas of themselves, but on a
smaller scale. In this section, we give a precise definition of the resulting
notion of self-similarity and prove a theorem determining the Hausdorff
dimension of these sets.
A mapping S : Rd → Rd is said to be a similarity with ratio r > 0 if
|S(x) − S(y)| = r|x − y|.
It can be shown that every similarity of Rd is the composition of a trans-

lation, a rotation, and a dilation by r. (See Problem 3.)
Given finitely many similarities S1 , . . . , Sm with the same ratio r, we
say that the set F ⊂ Rd is self-similar if
F = S1 (F ) ∪ · · · ∪ Sm (F ).
We point out the relevance of the various examples we have already seen.
When F = C is the Cantor set, there are two similarities given by
S1 (x) = x/3 and S2 (x) = x/3 + 2/3
of ratio 1/3. So m = 2 and r = 1/3.

In the case of F = S, the Sierpinski triangle, the ratio is r = 1/2 and
there are m = 3 similarities given by
x x x
S1 (x) = , S2 (x) = +α and S3 (x) = + β.
2 2 2
Here, α and β are the points drawn in the first diagram in Figure 5.
If F = K, the von Koch curve, we have
x x x
S1 (x) = , S2 (x) = ρ + α, S3 (x) = ρ−1 + β,
3 3 3
and
x
S4 (x) = + γ,
3
β
α
0 β 0 α γ 1
Figure 5. Similarities of the Sierpinski triangle and von Koch curve
where ρ is the rotation centered at the origin and of angle π/3. There
are m = 4 similarities which have ratio r = 1/3. The points α, β, and γ
are shown in the second diagram in Figure 5.
Another example, sometimes called the Cantor dust D, is another

two-dimensional version of the standard Cantor set. For each fixed 0 <
µ < 1/2, the set D may be constructed by starting with the unit square
Q = [0, 1] × [0, 1]. At the first stage we remove everything but the four
open squares in the corners of Q that have side length µ. This yields a
union D1 of four squares, as illustrated in Figure 6.
D1 D2
Figure 6. Construction of the Cantor dust
We repeat this process in each sub-square of D1 ; that is, we remove

everything but the four squares in the corner, each of side length µ2 .
This gives a union D2 of 16 squares. Repeating this process, we obtain
a family D1 ⊃ D2 ⊃ · · · ⊃ Dk ⊃ · · · of compact sets whose intersection
defines the Cantor dust corresponding to the parameter µ.
There are here m = 4 similarities of ratio µ given by
S1 (x) = µx,
S2 (x) = µx + (0, 1 − µ),
S3 (x) = µx + (1 − µ, 1 − µ),
S4 (x) = µx + (1 − µ, 0).
It is to be noted that D is the product Cξ × Cξ , with Cξ the Cantor set

of constant dissection ξ, as defined in Exercise 3, of Chapter 1. Here
ξ = 1 − 2µ.
The first result we prove guarantees the existence of self-similar sets
under the assumption that the similarities are contracting, that is, that
their ratio satisfies r < 1.
Theorem 2.9 Suppose S1 , S2 , . . . , Sm are m similartities, each with the

same ratio r that satisfies 0 < r < 1. Then there exists a unique non-
empty compact set F such that
F = S1 (F ) ∪ · · · ∪ Sm (F ).
The proof of this theorem is in the nature of a fixed point argument.

We shall begin with some large ball B and iteratively apply the mappings
S1 , . . . , Sm . The fact that the similarities have ratio r < 1 will suffice to
imply that this process contracts to a unique set F with the desired
property.
Lemma 2.10 There exists a closed ball B so that Sj (B) ⊂ B for all
j = 1, . . . , m.
Proof. Indeed, we note that if S is a similarity with ratio r, then
|S(x)| ≤ |S(x) − S(0)| + |S(0)|

≤ r|x| + |S(0)|.
If we require that |x| ≤ R implies |S(x)| ≤ R, it suffices to choose R

so that rR + |S(0)| ≤ R, that is, R ≥ |S(0)|/(1 − r). In this fashion,
we obtain for each Sj a ball Bj centered at the origin that satisfies
Sj (Bj ) ⊂ Bj . If B denotes the ball among the Bj with the largest radius,
then the above shows that Sj (B) ⊂ B for all j.
Now for any set A, let S̃(A) denote the set given by
S̃(A) = S1 (A) ∪ · · · ∪ Sm (A).

Note that if A ⊂ A0 , then S̃(A) ⊂ S̃(A0 ).

Also observe that while each Sj is a mapping from Rd to Rd , the
mapping S̃ is not a point mapping, but takes subsets of Rd to subsets of
Rd .
To exploit the notion of contraction with a ratio less than 1, we intro-
duce the distance between two compact sets as follows. For each δ > 0
and set A, we let
Aδ = {x : d(x, A) < δ}.
Hence Aδ is a set that contains A but which is slightly larger in terms of δ.

If A and B are two compact sets, we define the Hausdorff distance as
dist(A, B) = inf{δ : B ⊂ Aδ and A ⊂ B δ }.
Lemma 2.11 The distance function dist defined on compact subsets of

Rd satisfies
(i) dist(A, B) = 0 if and only if A = B.
(ii) dist(A, B) = dist(B, A).
(iii) dist(A, B) ≤ dist(A, C) + dist(C, B).
If S1 , . . . , Sm are similarities with ratio r, then
(iv) dist(S̃(A), S̃(B)) ≤ r dist(A, B).
The proof of the lemma is simple and may be left to the reader.
Using both lemmas we may now prove Theorem 2.9. We first choose
B as in Lemma 2.10, and let Fk = S̃ k (B), where S̃ k denotes the k th com-
position of S̃, that is, S̃ k = S̃ k−1 ◦ S̃ with S̃ 1 = S̃. Each Fk is compact,
non-empty, and Fk ⊂ Fk−1 , since S̃(B) ⊂ B. If we let
∞
\
F = Fk ,
k=1
then
T∞ F is compact,Tnon-empty, and clearly S̃(F ) = F , since applying S̃
∞
to k=1 Fk yields k=2 Fk , which also equals F .
Uniqueness of the set F is proved as follows. Suppose G is another
compact set so that S̃(G) = G. Then, an application of part (iv) in
Lemma 2.11 yields dist(F, G) ≤ r dist(F, G). Since r < 1, this forces
dist(F, G) = 0, so that F = G, and the proof of Theorem 2.9 is com-
plete.
Under an additional technical condition, one can calculate the precise

Hausdorff dimension of the self-similar set F . Loosely speaking, the
restriction holds if the sets S1 (F ), . . . , Sm (F ) do not overlap too much.
Indeed, if these sets were disjoint, then we could argue that
m
X
mα (F ) = mα (Sj (F )).
j=1
Since each Sj scales by r, we would then have mα (Sj (F )) = rα mα (F ).

Hence
mα (F ) = mrα mα (F ).
If mα (F ) were finite, then we would have that mrα = 1; thus
log m
α= .
log 1/r
The restriction we impose is as follows. We say that the similarities

S1 , . . . , Sm are separated if there is an bounded open set O so that
O ⊃ S1 (O) ∪ · · · ∪ Sm (O),
and the Sj (O) are disjoint. It is not assumed that O contains F .
Theorem 2.12 Suppose S1 , S2 , . . . , Sm are m separated similarities with

the common ratio r that satisfies 0 < r < 1. Then the set F has Haus-
dorff dimension equal to log m/ log(1/r).
Observe first that when F is the Cantor set we may take O to be

the open unit interval, and note that we have already proved that its
dimension is log 2/ log 3. For the Sierpinski triangle the open unit triangle
will do, and dim S = log 3/ log 2. In the example of the Cantor dust the
open unit square works, and dim D = log m/ log µ−1 . Finally, for the von
Koch curve we may take the interior of the triangle pictured in Figure 7,
and we will have dim K = log 4/ log 3.
We now turn to the proof of Theorem 2.12, which will follow the same
approach used in the case of the Sierpinski triangle. If α = log m/ log(1/r),
we claim that mα (F ) < ∞, hence dim F ≤ α. Moreover, this inequality
holds even without the separation assumption. Indeed, recall that
Fk = S̃ k (B),
Figure 7. Open set in the separation of the von Koch similarities
and S̃ k (B) is the union of mk sets of diameter less than crk (with c =
diam B), each of the form
Sn1 ◦ Sn2 ◦ · · · ◦ Snk (B), where 1 ≤ ni ≤ m and 1 ≤ i ≤ k.
Consequently, if crk ≤ δ, then

X α
δ
Hα (F ) ≤ (diam Sn1 ◦ · · · ◦ Snk (B))
n1 ,...,nk
≤ c0 mk rαk
≤ c0 ,
since mrα = 1, because α = log m/ log(1/r). Since c0 is independent of

δ, we get mα (F ) ≤ c0 .
To prove mα (F ) > 0, we now use the separation condition. We argue
in parallel with the earlier calculation of the Hausdorff dimension of the
Sierpinski triangle.
Fix a point x in F . We define the “vertices” of the k th generation as
the mk points that lie in F and are given by
Sn1 ◦ · · · ◦ Snk (x), where 1 ≤ n1 ≤ m, . . . , 1 ≤ nk ≤ m.
Each vertex is labeled by (n1 , . . . , nk ). Vertices need not be distinct, so

they are counted with their multiplicities.
Similarly, we define the “open sets” of the k th generation to be the mk
sets given by
Sn1 ◦ · · · ◦ Snk (O), where 1 ≤ n1 ≤ m, . . . , 1 ≤ nk ≤ m,
and where O is fixed and chosen to satisfy the separation condition.

Such open sets are again labeled by multi-indices (n1 , n2 , . . . , nk ) with
1 ≤ nj ≤ m, 1 ≤ j ≤ k.
Then the open sets of the k th generation are disjoint, since those of
the first generation are disjoint. Moreover if k ≥ `, each open set of the
`th generation contains mk−` open sets of the k th generation.
Suppose v is a vertex of the k th generation, and let O(v) denote the
open set in the k th generation which is associated to v, that is, v and
O(v) carry the same label (n1 , n2 , . . . , nk ). Since x is at a fixed distance
from the original open set O, and O has a finite diameter, we find that
(a) d(v, O(v)) ≤ crk .
(b) c0 rk ≤ diam O(v) ≤ crk .
As in the case of the Sierpinski triangle, it suffices to prove that if
B = {Bj }N
j=1 is a finite collection of balls whose diameters are less than
δ and whose union covers F , then
N
X
(diam Bj )α ≥ c > 0.
j=1
Suppose we have such a covering by balls, and choose k so that
rk ≤ min diam Bj < rk−1 .

1≤j≤N
Lemma 2.13 Suppose B is a ball in the covering B that satisfies
r` ≤ diam B < r`−1 for some ` ≤ k.
Then B contains at most cmk−` vertices of the k th generation.
Proof. If v is a vertex of the k th generation with v ∈ B, and O(v)

denotes the corresponding open set of the k th generation, then, for some
fixed dilate B ∗ of B, properties (a) and (b) above guarantee that O(v) ⊂
B ∗ , and B ∗ also contains the open set of generation ` that contains O(v).
Since B ∗ has volume crd` , and each open set in the `th generation has
volume ≈ rd` (by property (b) above), B ∗ can contain at most c open
sets of generation `. Hence B ∗ contains at most cmk−` open sets of the
k th generation. Consequently, B can contain at most cmk−` vertices of
the k th generation, and the lemma is proved.
For the final argument, let N` denote the number of balls in B so that
r` ≤ diam Bj ≤ r`−1 .
By the lemma, we see that the total number of vertices of the k th gen-
eration that can be covered by the collection B can be no more than
3. Space-filling curves 349
P
c ` N` mk−` . Since mk vertices of the k th generation belong to F ,
P all k−`
we must have c ` N` m ≥ mk , and hence
X
N` m−` ≥ c.
`
The definition of α gives r`α = m−` , and therefore

N
X X
(diam Bj )α ≥ N` r`α ≥ c,
j=1 `
and the proof of Theorem 2.12 is complete.
3 Space-filling curves
The year 1890 heralded an important discovery: Peano constructed a
continuous curve that filled an entire square in the plane. Since then,
many variants of his construction have been given. We shall describe here
a construction that has the feature of elucidating an additional significant
fact. It is that from the point of measure theory, speaking broadly, the
unit interval and unit square are “isomorphic.”
Theorem 3.1 There exists a curve t 7→ P(t) from the unit interval to
the unit square with the following properties:
(i) P maps [0, 1] to [0, 1] × [0, 1] continuously and surjectively.
(ii) P satisfies a Lipschitz condition of exponent 1/2, that is,
|P(t) − P(s)| ≤ M |t − s|1/2 .
(iii) The image under P of any sub-interval [a, b] is a compact subset of

the square of (two-dimensional) Lebesgue measure exactly b − a.
The third conclusion can be elaborated further.
Corollary 3.2 There are subsets Z1 ⊂ [0, 1] and Z2 ⊂ [0, 1] × [0, 1], each
of measure zero, such that P is bijective from
[0, 1] − Z1 to [0, 1] × [0, 1] − Z2
and measure preserving. In other words, E is measurable if and only if

P(E) is measurable, and
m1 (E) = m2 (P(E)).
Here m1 and m2 denote the Lebesgue measures in R1 and R2 , respec-

tively.
We shall call the function t 7→ P(t) the Peano mapping. Its image
is called the Peano curve.
Several observations help clarify the nature of the conclusions of the
theorem. Suppose that F : [0, 1] → [0, 1] × [0, 1] is continuous and sur-
jective. Then:
(a) F cannot be Lipschitz of exponent γ > 1/2. This follows at once
from Lemma 2.2, which states that
1
dim F ([0, 1]) ≤ dim[0, 1],
γ
so that 2 ≤ 1/γ as desired.

(b) F cannot be injective. Indeed, if this were the case, then the in-
verse G of F would exist and would be continuous. Given any two
points a 6= b in [0, 1], we would get a contradiction by looking at
two distinct curves in the square that join F (a) and F (b), since the
image of these two curves under G would have to intersect at points
between a and b. In fact, given any open disc D in the square, there
always exists x ∈ D so that F (t) = F (s) = x yet t 6= s.
The proof of Theorem 3.1 will follow from a careful study of a natu-
ral class of mappings that associate sub-squares in [0, 1] × [0, 1] to sub-
intervals in [0, 1]. This implements the approach implicit in Hilbert’s
iterative procedure, which he set forth in the first three stages in Fig-
ure 8.
Figure 8. Construction of the Peano curve
We turn now to the study of the general class of mappings.

3.1 Quartic intervals and dyadic squares

The quartic intervals arise when [0, 1] is successively sub-divided by
powers of 4. For instance, the first generation quartic intervals are the
closed intervals
I1 = [0, 1/4], I2 = [1/4, 1/2], I3 = [1/2, 3/4], I4 = [3/4, 1].
The second generation quartic intervals are obtained by sub-dividing each

interval of the first generation by 4. Hence there are 16 = 42 quartic in-
tervals of the second generation. In general, there are 4k quartic intervals
of the k th generation, each of the form [ 4`k , `+1
4k
], where ` is integral with
0 ≤ ` < 4k .
A chain of quartic intervals is a decreasing sequence of intervals
I1 ⊃ I2 ⊃ · · · ⊃ Ik ⊃ · · · ,
where I k is a quartic interval of the k th generation (hence |I k | = 4−k ).
Proposition 3.3 Chains of quartic intervals satisfy the following prop-

erties:
(i) If {I k } is a chain of quartic

T intervals, then there exists a unique
t ∈ [0, 1] such that t ∈ k I k .
t ∈ [0, 1], there is a chain {I k } of quartic inter-

(ii) Conversely, given T
vals such that t ∈ k I k .
(iii) The set of t for which the chain in part (ii) is not unique is a set
of measure zero (in fact, this set is countable).
Proof. Part (i) follows from the fact that {I k } is a decreasing sequence
of compact sets whose diameters go to 0.
For part (ii), we fix t and note that for each k there exists at least one
quartic interval I k with t ∈ I k . If t is of the form `/4k , where 0 < ` < 4k ,
then there are exactly two quartic intervals of the k th generation that
contain t. Hence, the set of points for which the chain is not unique is
precisely the set of dyadic rationals
`
, where 1 ≤ k, and 0 < ` < 4k .
4k
Note that of course, these fractions are the same as those of the form
0 0
`0 /2k with 0 < `0 < 2k . This set is countable, hence has measure 0.
It is clear that each chain {I k } of quartic intervals can be represented

naturally by a string .a1 a2 · · · ak · · · , where each ak is either 0, 1, 2, or 3.
Then the point t corresponding to this chain is given by
∞
X ak
t= .
4k
k=1
The points where ambiguity occurs are precisely those where ak = 3 for
all sufficiently large k, or equivalently where ak = 0 for all sufficiently
large k.
Part of our description of the Peano mapping will follow from associ-
ating to each quartic interval a dyadic square. These dyadic squares
are obtained by sub-dividing the unit square [0, 1] × [0, 1] in the plane by
successively bisecting the sides.
For instance, dyadic squares of the first generation arise from bisecting
the sides of the unit square. This yields four closed squares S1 , S2 , S3
and S4 , each of side length 1/2 and area |Si | = 1/4, for i = 1, . . . , 4.
The dyadic squares of the second generation are obtained by bisecting
each dyadic square of the first generation, and so on. In general, there
are 4k squares of the k th generation, each of side length 1/2k and area
1/4k .
A chain of dyadic squares is a decreasing sequence of squares
S1 ⊃ S2 ⊃ · · · ⊃ Sk ⊃ · · · ,
where S k is a dyadic square of the k th generation.
Proposition 3.4 Chains of dyadic squares have the following proper-

ties:
(i) If {S k } is a chain of dyadic squares,

T then there exists a unique
x ∈ [0, 1] × [0, 1] such that x ∈ k S k .
[0, 1] × [0, 1], there is a chain {S k } of dyadic
(ii) Conversely, given x ∈ T
squares such that x ∈ k S k .
(iii) The set of x for which the chain in part (ii) is not unique is a set
of measure zero.
In this case, the set of ambiguities consists of all points (x1 , x2 ) where
one of the coordinates is a dyadic rational. Geometrically, this set is
the (countable) union of vertical and horizontal segments in [0, 1] × [0, 1]
determined by the grid of dyadic rationals. This set has measure zero.
Moreover, each chain of dyadic squares can be represented by a string

.b1 b2 · · · , where each bk is either 0, 1, 2 or 3. Then
∞
X bk
(1) x= ,
2k
k=1
where
bk = (0, 0) if bk = 0,
bk = (0, 1) if bk = 1,
bk = (1, 0) if bk = 2,
bk = (1, 1) if bk = 3.
3.2 Dyadic correspondence

A dyadic correspondence is a mapping Φ from quartic intervals to
dyadic squares that satisfies:
(1) Φ is bijective.
(2) Φ respects generations.
(3) Φ respects inclusion.
By (2), we mean that if I is a quartic interval of the k th generation, then
Φ(I) is a dyadic square of the k th generation. By (3), we mean that if
I ⊂ J, then Φ(I) ⊂ Φ(J).
For example, the trivial, or standard correspondence assigns to the
string .a1 a2 · · · the string .b1 b2 · · · with bk = ak .
∗
Given a dyadic correspondence Φ, the induced mapping T k Φ maps
[0, 1] to [0, 1] × [0, 1] and is given as follows. If {t} = I where {I k }
is a chain of quartic intervals, then, since {Φ(I k )} is a chain of dyadic
squares, we may let
\
Φ∗ (t) = x = Φ(I k ).
We note that Φ∗ is well-defined except on a (countable) set of measure

zero, (those points t that are represented by more than one quartic chain.)
A moment’s reflection will show that if I 0 is a quartic interval of the
k generation, then the images Φ∗ (I 0 ) = {Φ∗ (t), t ∈ I 0 }, comprise the
th
dyadic square of the k th generation Φ(I 0 ). Thus Φ∗ (I 0 ) = Φ(I 0 ), and

hence m1 (I 0 ) = m2 (Φ∗ (I 0 )).
Theorem 3.5 Given a dyadic correspondence Φ, there exist sets Z1 ⊂

[0, 1] and Z2 ⊂ [0, 1] × [0, 1], each of measure zero, so that:
(i) Φ∗ is a bijection on [0, 1] − Z1 to [0, 1] × [0, 1] − Z2 .

(ii) E is measurable if and only if Φ∗ (E) is measurable.
(iii) m1 (E) = m2 (Φ∗ (E)).
Proof. First, let N1 denote the collection of chains of those quartic

intervals arising in (iii) of Proposition 3.3, those for which the points in
I = [0, 1] are not uniquely representable. Similarly, let N2 denote the
collection of chains of those dyadic squares for which the corresponding
points in the square I × I are not uniquely representable.
Since Φ is a bijection from chains of quartic intervals to chains of dyadic
squares, it is also a bijection from N1 ∪ Φ−1 (N2 ) to Φ(N1 ) ∪ N2 , and
hence also of their complements. Let Z1 be the subset of I consisting of
all points in I that can be represented (according to (i) of Proposition 3.3)
by the chains in N1 ∪ Φ−1 (N2 ), and let Z2 be the set of points in the
square that can be represented by dyadic squares in Φ(N1 ) ∪ N2 . Then
Φ∗ , the induced mapping, is well-defined on I − Z1 , and gives a bijection
of I − Z1 to (I × I) − Z2 . To prove that both Z1 and Z2 have measure
zero, we invoke the following lemma. We suppose {fk }∞ k=1 is a fixed given
sequence, with each fk either 0, 1, 2, or 3.
Lemma 3.6 Let

∞
X
E0 = {x = ak /4k , where ak 6= fk for all sufficiently large k}.
k=1
Then m(E0 ) = 0.
Indeed, if we fix r, then m({x : ar 6= fr }) = 3/4, and
m({x : ar 6= fr and ar+1 6= fr+1 }) = (3/4)2 , etc.
Thus m({x : ak 6= fk , all k ≥ r}) = 0, and E0 is a countable union of

such sets, from which the lemma follows.
There is a similar statement for points in the square S = I × I in terms
of the representation (1).
Note that as a result the set of points in I corresponding to chains in
N1 form a set of measure zero. In fact, we may use the lemma for the
sequence for which fk = 1, for all k, since the elements of N1 correspond
to sequences {ak } with ak = 0 for all sufficiently large k, or ak = 3 for
all sufficiently large k.
Similarly, the points in the square S corresponding to N2 form a set of
measure zero. To see this, take for example fk = 1 for k odd, and fk = 2
for k even, and note that N2 corresponds to all sequences {ak } where
one of the following four exclusive alternatives holds for all sufficiently
large k: either ak is 0 or 1; or ak is 2 or 3; or ak is 0 or 2; or ak is 1
or 3. By similar reasoning the points Φ−1 (N2 ) and Φ(N1 ) form sets of
measure zero in I and I × I respectively.
We now turn to the proof that Φ∗ (which is a bijection from I − Z1
to (I × I) − Z2 ) is measure preserving. For this it is useful to recall
Theorem 1.4 in Chapter 1, whereby any S∞ open set O in the unit interval
I can be realized as a countable union j=1 Ij , where each Ij is a closed
interval and the Ij have disjoint interiors. Moreover, an examination of
the proof shows that the intervals can be taken to be dyadic, that is, of the
form [`/2j , (` + 1)/2j ], for appropriate integers ` and j. Further, such an
interval is itself a quartic interval if j is even, j = 2k, or the union of two
quartic intervals [(2`)/22k , (2` + 1)/22k ] and [(2` + 1)/22k , (2` + 2)/22k ],
if j is odd, j = 2k − 1. Thus any open set in I can be given as a union of
quartic intervals whose interiors are disjoint. Similarly, any open set in
the square I × I is a union of dyadic squares whose interiors are disjoint.
Now let E beSany set of measure zero in I − Z1 and ²P > 0. Then we
can cover E ⊂ j S Ij , where Ij are quartic intervals and j m1 (Ij ) < ².
Because Φ∗ (E) ⊂ j Φ∗ (Ij ), then
X X
m2 (Φ∗ (E)) ≤ m2 (Φ∗ (Ij )) = m1 (Ij ) < ².
Thus Φ∗ (E) is measurable and m2 (Φ∗ (E)) = 0. Similarly, (Φ∗ )−1 maps
sets of measure zero in (I × I) − Z2 to sets of measure zero in I.
Now the argument above also shows that if O is any open set in I,
then Φ∗ (O − Z1 ) is measurable, and m2 (Φ∗ (O − Z1 )) = m1 (O). Thus
this identity goes over to Gδ sets in I. Since any measurable set differs
from a Gδ set by a set of measure zero, we see that we have established
that m2 (Φ∗ (E)) = m1 (E) for any measurable subset of E of I − Z1 . The
same argument can be applied to (Φ∗ )−1 , and this completes the proof
of the theorem.
The Peano mapping will be obtained as Φ∗ for a special correspon-
dence Φ.
3.3 Construction of the Peano mapping

The particular dyadic correspondence we now present provides us with
the steps to follow when tracing the approximations of the Peano curve.
The main idea behind its construction is that as we go from one quartic
interval in the k th generation to the next quartic interval in the same
generation, we move from a dyadic square of the k th generation to another

square of the k th generation that shares a common side.
More precisely, we say that two quartic intervals in the same generation
are adjacent if they share a point in common. Also, two squares in the
same generation are adjacent if they share a side in common.
Lemma 3.7 There is a unique dyadic correspondence Φ so that:
(i) If I and J are two adjacent intervals of the same generation, then
Φ(I) and Φ(J) are two adjacent squares (of the same generation).
(ii) In generation k, if I− is the left-most interval and I+ the right-
most interval, then Φ(I− ) is the left-lower square and Φ(I+ ) is the
right-lower square.
Part (ii) of the lemma is illustrated in Figure 9.
I− I+
0 1
Figure 9. Special dyadic correspondence
Given a square S and its four immediate sub-squares, an acceptable

traverse is an ordering of the sub-squares S1 , S2 , S3 , and S4 , so that
Sj and Sj+1 are adjacent for j = 1, 2, 3. With such an ordering, we note
that if we color S1 white, and then alternate black and white, the square
S3 is also white, while S2 and S4 are black. The important point to
remember is that if the first square in a traverse is white, then the last
square is black.
The key observation is the following. Suppose we are given a square
S, and a side σ of S. If S1 is any of the immediate four sub-squares in
S, then there exists a unique traverse S1 , S2 , S3 , and S4 so that the last
square S4 has a side in common with σ. With the initial square S1 in
the lower-left corner of S, the four possibilities which correspond to the
four choices of σ, are illustrated in Figure 10.
We may now begin the inductive description of the dyadic correspon-
dence satisfying the conditions in the lemma. On quartic intervals of the
first generation we assign the square Sj = Φ(Ij ), as pictured in Figure 11.
S2 S3 S2 S3
S1 S4 S1 S4
σ
σ
S4 S3 S4 S3
S1 S2 S1 S2
Figure 10. Traverses
S2 S3
I1 I2 I3 I4
S1 S4
Figure 11. Initial step of the correspondence
Now suppose Φ has been defined for all quartic intervals of generation
less than or equal to k. We now write the intervals in generation k in
increasing order as I1 , . . . , I4k , and let Sj = Φ(Ij ). We then divide I1
into four quartic intervals of generation k + 1 and denote them by I1,1 ,
I1,2 , I1,3 , and I1,4 , where the intervals are chosen in increasing order.
Then, we assign to each interval I1,j a dyadic square Φ(I1,j ) = Sj of
generation k + 1 contained in S1 so that:
(a) S1,1 is the lower-left sub-square of S1 ,
(b) S1,4 touches the side that S1 shares with S2 ,
(c) S1,1 , S1,2 , S1,3 , and S1,4 is a traverse.
This is possible, since the induction hypothesis guarantees that S2 is
adjacent to S1 .
This settles the assignments for the sub-squares of S1 , so we now turn

our attention to S2 . Let I2,1 , I2,2 , I2,3 , and I2,4 denote the quartic
intervals of generation k + 1 in I2 , written in increasing order. First, we
take S2,1 = Φ(I2,1 ) to be the sub-square of S2 which is adjacent to S1,4 .
This can be done because S1,4 touches S2 by construction. Note that
we leave S1 from a black square (S1,4 ), and enter S2 in a white square
(S2,1 ). Since S3 is adjacent to S2 , we may now find a traverse S2,1 , S2,2 ,
S2,3 and S2,4 so that S2,4 touches S3 .
We may then repeat this process in each interval Ij and square Sj ,
j = 3, . . . , 4k . Note that at each stage the square Sj,1 (the “entering”
square) is white, while Sj,4 (the “exiting” square) is black.
In the final step, the induction hypothesis guarantees that S4k is the
lower-right corner square. Moreover, since S4k −1 must be adjacent to
S4k it must be either above it, or to the left of it, so we enter a square of
the (k + 1)st generation along an upper or left side. The entering square
is a white square, and we traverse to the lower right corner sub-square
of S4k , which is a black square.
This concludes the inductive step, hence the proof of Lemma 3.7.
We may now begin the actual description of the Peano curve. For each
generation k we construct a polygonal line which consists of vertical and
horizontal line segments connecting the centers of consecutive squares.
More precisely, let Φ denote the dyadic correspondence in Lemma 3.7,
and let S1 , . . . , S4k be the squares of the k th generation ordered according
to Φ, that is, Φ(Ij ) = Sj . Let tj denote the middle point of Ij ,
j − 12
tj = for j = 1, . . . , 4k .
4k
Let xj be the center of the square Sj , and define
Pk (tj ) = xj .
Also set
Pk (0) = (0, 1/2k+1 ) = x0 where t0 = 0,
and
Pk (1) = (1, 1/2k+1 ) = x4k+1 where t4k +1 = 1.
Then, we extend Pk (t) to the unit interval 0 ≤ t ≤ 1 by linearity along

the sub-intervals determined by the division points t0 , . . . , t4k +1 .
Note that the distance |xj − xj+1 | = 1/2k , while |tj − tj+1 | = 1/4k for
0 ≤ j ≤ 4k . Also
1
|x1 − x0 | = |x4k − x4k+1 | = ,
2 · 2k
while
1
|t1 − t0 | = |t4k − t4k+1 | = .
2 · 4k
Therefore Pk0 (t) = 4k 2−k = 2k except when t = tj .

As a result,
|Pk (t) − Pk (s)| ≤ 2k |t − s|.
However,
√
|Pk+1 (t) − Pk (t)| ≤ 2 2−k ,
because when `/4k ≤ t ≤ (` + 1)/4k , then Pk+1 (t) and Pk (t) belong to
the same dyadic square of generation k.
Therefore the limit
∞
X
P(t) = lim Pk (t) = P1 (t) + Pj+1 (t) − Pj (t)
k→∞
j=1
exists, and defines a continuous function in view of the uniform conver-

gence. By Lemma 2.8 we conclude that
|P(t) − P(s)| ≤ M |t − s|1/2 ,
and P satisfies a Lipschitz condition of exponent of 1/2.

Moreover, each Pk (t) visits each dyadic square of generation k as t
ranges in [0, 1]. Hence P is dense in the unit square, and by continuity
we find that t 7→ P(t) is a surjection.
Finally, to prove the measure preserving property of P, it suffices to
establish P = Φ∗ .
Lemma 3.8 If Φ is the dyadic correspondence in Lemma 3.7, then Φ∗ (t) =

P(t) for every 0 ≤ t ≤ 1.
∗
Proof. First, we observe that
T kΦ (t) is unambiguously
T k defined for
every t. Indeed, suppose t ∈ k I and t ∈ k J are two chains of
quartic intervals; then I k and J k must be adjacent for sufficiently large
k. Thus Φ(I k ) and Φ(J k ) must be adjacent squares for all sufficiently
large k. Hence
\ \
Φ(I k ) = Φ(J k ).
k k
Next, directly from our construction we have

\
Φ(I k ) = lim Pk (t) = P(t).
k
This gives the desired conclusion.

The argument also shows that P(I) = Φ(I) for anySquartic interval I.
Now recall that any interval (a, b) can be written as j Ij , where the Ij
are quartic intervals with disjoint interiors. Because P(Ij ) = Φ(I
Sj ), these
are then dyadic squares with disjoint interiors. Since P(a, b) = k P(Ij ),
we have
X∞ X ∞ ∞
X
m2 (P(a, b)) = m2 (P(Ij )) = m2 (Φ(Ij )) = m1 (Ij ) = m1 (a, b).
j=1 j=1 j=1
This proves conclusion (iii) of Theorem 3.1. The other conclusions hav-
ing already been established, we need only note that the corollary is
contained in Theorem 3.5.
As a result, we conclude that t 7→ P(t) also induces a measure pre-
serving mapping from [0, 1] to [0, 1] × [0, 1]. This concludes the proof of
Theorem 3.1.
4* Besicovitch sets and regularity

We begin by presenting a surprising regularity property enjoyed by all
measurable subsets (of finite measure) of Rd when d ≥ 3. As we shall
see, the fact that the corresponding phenomenon does not hold for d =
2 is due to the existence of a remarkable set that was discovered by
Besicovitch. A construction of a set of this kind will be detailed in
Section 4.4.
We first fix some notation. For each unit vector γ on the sphere,
γ ∈ S d−1 , and each t ∈ R we consider the plane Pt,γ , which is defined
as the (d − 1)-dimensional affine hyperplane perpendicular to γ and of
“signed distance” t from the origin.1 The plane Pt,γ is given by
Pt,γ = {x ∈ Rd : x · γ = t}.
1 Note that there are two planes perpendicular to γ and of distance |t| from the origin;
this accounts for the fact that t may be either positive or negative.
4*. Besicovitch sets and regularity 361
We observe that each Pt,γ carries a natural (d − 1) Lebesgue measure,

denoted by md−1 . In fact, if we complete γ to an orthonormal basis
e1 , e2 , . . . , ed−1 , γ of Rd , then we can write any x ∈ Rd in terms of the
corresponding coordinates as x = x1 e1 + x2 e2 + · · · + xd γ. When we set
x ∈ Rd = Rd−1 × R with (x1 , . . . , xd−1 ) ∈ Rd−1 , xd ∈ R, then the mea-
sure md−1 on Pt,γ is the Lebesgue measure on Rd−1 . This definition of
md−1 is independent of the choice of orthonormal vectors e1 , e2 , . . . , ed−1 ,
since Lebesgue measure is invariant under rotations. (See Problem 4,
Chapter 2, or Exercise 26, Chapter 3.)
With these preliminaries out of the way, we define for each subset
E ⊂ Rd the slice of E cut out by the plane Pt,γ as
Et,γ = E ∩ Pt,γ .
We now consider the slices Et,γ as t varies, where E is measurable and

γ is fixed. (See Figure 12.)
Et1 ,γ
Pt2 ,γ
Pt1 ,γ
Figure 12. The slices E ∩ Pt,γ as t varies
We observe that for almost every t the set Et,γ is md−1 measurable
and, moreover, md−1 (Et,γ ) is a measurable function of t. This is a
direct consequence of Fubini’s theorem and the above decomposition,
Rd = Rd−1 × R. In fact, so long as the direction γ is pre-assigned, not
much more can be said in general about the function t 7→ md−1 (Et,γ ).
However, when d ≥ 3 the nature of the function is dramatically different

for “most” γ. This is contained in the following theorem.
Theorem 4.1 Suppose E is of finite measure in Rd , with d ≥ 3. Then
for almost every γ ∈ S d−1 :
(i) Et,γ is measurable for all t ∈ R.
(ii) md−1 (Et,γ ) is continuous in t ∈ R.
Moreover, the function of t defined by µ(t, γ) = md−1 (Et,γ ) satisfies a
Lipschitz condition with exponent α for any α with 0 < α < 1/2.
The almost everywhere assertion is with respect to the natural measure
dσ on S d−1 that arises in the polar coordinate formula in Section 3.2 of
the previous chapter.
We recall that a function f is Lipschitz with exponent α if
|f (t1 ) − f (t2 )| ≤ A|t1 − t2 |α for some A.
A significant part of (i) is that for a.e. γ, the slice Et,γ is measurable
for all values of the parameter t. In particular, one has the following.
Corollary 4.2 Suppose E is a set of measure zero in Rd with d ≥ 3.
Then, for almost every γ ∈ S d−1 , the slice Et,γ has zero measure for all
t ∈ R.
The fact that there is no analogue of this when d = 2 is a consequence of
the existence of a Besicovitch set, (also called a “Kakeya set”), which is
defined as a set that satisfies the three conditions in the theorem below.
Theorem 4.3 There exists a set B in R2 that:
(i) is compact,
(ii) has Lebesgue measure zero,
(iii) contains a translate of every unit line segment.
Note that with F = B and γ ∈ S 1 one has m1 (F ∩ Pt0 ,γ ) ≥ 1 for some t0 .
If m1 (F ∩ Pt,γ ) were continuous in t, then this measure would be strictly
positive for an interval in t containing t0 , and thus we would have
m2 (F ) > 0, by Fubini’s theorem. This contradiction shows that the ana-
logue of Theorem 4.1 cannot hold for d = 2.
While the set B has zero two-dimensional measure, this assertion can-
not be improved by replacing this measure by α-dimensional Hausdorff
measure, with α < 2.
Theorem 4.4 Suppose F is any set that satisfies the conclusions (i)
and (iii) of Theorem 4.3. Then F has Hausdorff dimension 2.
4.1 The Radon transform

Theorems 4.1 and 4.4 will be derived by an analysis of the regularity
properties of the Radon transform R. The operator R arises in a number
of problems in analysis, and was already considered in Chapter 6 of
Book I.
For an appropriate function f on Rd , the Radon transform of f is
defined by
Z
R(f )(t, γ) = f.
Pt,γ
The integration is performed over the plane Pt,γ with respect to the
measure md−1 discussed above. We first make the following simple ob-
servation:
1. If f is continuous and has compact support, then f is of course
integrable on every plane Pt,γ , and so R(f )(t, γ) is defined for all
(t, γ) ∈ R × S d−1 . Moreover it is a continuous function of the pair
(t, γ) and has compact support in the t-variable.
2. If f is merely Lebesgue integrable, then f may fail to be measurable
or integrable on Pt,γ for some (t, γ), and thus R(f )(t, γ) is not
defined for those (t, γ).
3. Suppose f is the characteristic function of the set E, that is, f =
χE . Then R(f )(t, γ) = md−1 (Et,γ ) if Et,γ is measurable.
It is this last property that links the Radon transform to our problem.
Key estimates in this conclusion involve a maximal “Radon transform”
defined by
R∗ (f )(γ) = sup |R(f )(t, γ)|,
t∈R
as well as corresponding expressions controlling the Lipschitz character

of R(f )(t, γ) as a function of t. A basic fact inherent in our analysis
is that the regularity of the Radon transform actually improves as the
dimension of the underlying space increases.
Theorem 4.5 Suppose f is continuous and has compact support in Rd

with d ≥ 3. Then
Z
£ ¤
(2) R∗ (f )(γ) dσ(γ) ≤ c kf kL1 (Rd ) + kf kL2 (Rd )
S d−1
for some constant c > 0 that does not depend on f .

An inequality of this type is a typical “a priori” estimate. It is obtained

first under some regularity assumption on the function f , and then a
limiting argument allows one to pass to the more general case when f
belongs to L1 ∩ L2 .
We make some comments about the appearance of both the L1 -norm
and L2 -norm in (2). The L2 -norm imposes a crucial local control of
the kind that is necessary for the desired regularity. (See Exercise 27.)
However, without some restriction on f of a global nature, the function
f might fail to be integrable on any plane Pt,γ , as the example f (x) =
1/(1 + |x|d−1 ) shows. Note that this function belongs to L2 (Rd ) if d ≥ 3,
but not to L1 (Rd ).
The proof of Theorem 4.5 actually gives an essentially stronger result,
which we state as a corollary.
Corollary 4.6 Suppose f is continuous and has compact support in

Rd , d ≥ 3. Then for any α, 0 < α < 1/2, the inequality (2) holds with
R∗ (f )(γ) replaced by
|R(f )(t1 , γ) − R(f )(t2 , γ)|

(3) sup .
t1 6=t2 |t1 − t2 |α
The proof of the theorem relies on the interplay between the Radon
transform and the Fourier transform.
For fixed γ ∈ S d−1 , we let R̂(f )(λ, γ) denote the Fourier transform of
R(f )(t, γ) in the t-variable
Z ∞
R̂(f )(λ, γ) = R(f )(t, γ)e−2πiλt dt.
−∞
In particular, we use λ ∈ R to denote the dual variable of t.

We also write fˆ for the Fourier transform of f as a function on Rd ,
namely
Z
ˆ
f (ξ) = f (x)e−2πix·ξ dx.
Rd
Lemma 4.7 If f is continuous with compact support, then for every

γ ∈ S d−1 we have
R̂(f )(λ, γ) = fˆ(λγ).
The right-hand side is just the Fourier transform of f evaluated at the

point λγ.
Proof. For each unit vector γ we use the adapted coordinate system
described above: x = (x1 , . . . , xd ) where γ coincides with the xd direc-
tion. We can then write each x ∈ Rd as x = (u, t) with u ∈ Rd−1 , t ∈ R,
where x · γ = t = xd and u = (x1 , . . . , xd−1 ). Moreover
Z Z
f= f (u, t) du,
Pt,γ Rd−1
R R ∞ ³R ´
and Fubini’s theorem shows that Rd
f (x) dx = −∞ Pt,γ
f dt. Ap-
−2πix·(λγ)
plying this to f (x)e in place of f (x) gives
Z Z ∞ µZ ¶
ˆ
f (λγ) = f (x)e−2πix·(λγ)
dx = f (u, t) du e−2πiλt dt
R d −∞ Rd−1
Z ÃZ !
∞
= f e−2πiλt dt.
−∞ Pt,γ
Therefore fˆ(λγ) = R̂(f )(λ, γ), and the lemma is proved.
Lemma 4.8 If f is continuous with compact support, then

Z µZ ∞ ¶ Z
2 d−1
|R̂(f )(λ, γ)| |λ| dλ dσ(γ) = 2 |f (x)|2 dx.
S d−1 −∞ Rd
Let us observe the crucial point that the greater the dimension d, the
larger the factor |λ|d−1 as |λ| tends to infinity. Hence the greater the
dimension, the better the decay of the Fourier transform R̂(f )(λ, γ),
and so the better the regularity of the Radon transform R(f )(t, γ) as a
function of t.
Proof. The Plancherel formula in Chapter 5 guarantees that
Z Z
2 2
|f (x)| dx = 2 |fˆ(ξ)|2 dξ.
Rd Rd
Changing to polar coordinates ξ = λγ where λ > 0 and γ ∈ S d−1 , we

obtain
Z Z Z ∞
2 ˆ 2
|f (ξ)| dξ = 2 |fˆ(λγ)|2 λd−1 dλ dσ(γ).
Rd S d−1 0
We now observe that a simple change of variables provides

Z Z ∞ Z Z 0
|fˆ(λγ)|2 λd−1 dλ dσ(γ) = |fˆ(λγ)|2 |λ|d−1 dλ dσ(γ),
S d−1 0 S d−1 −∞
and the proof is complete once we invoke the result of Lemma 4.7.
The final ingredient in the proof of Theorem 4.5 consists of the follow-
ing:
Lemma 4.9 Suppose

Z ∞
F (t) = F̂ (λ)e2πiλt dλ,
−∞
where
Z ∞
sup |F̂ (λ)| ≤ A and |F̂ (λ)|2 |λ|d−1 dλ ≤ B 2 .
λ∈R −∞
Then
(4) sup |F (t)| ≤ c(A + B).

t∈R
Moreover, if 0 < α < 1/2, then
(5) |F (t1 ) − F (t2 )| ≤ cα |t1 − t2 |α (A + B) for all t1 , t2 .
Proof. The first inequality is obtained by considering separately the

two cases |λ| ≤ 1 and |λ| > 1. We write
Z Z
F (t) = F̂ (λ)e2πiλt dλ + F̂ (λ)e2πiλt dλ.
|λ|≤1 |λ|>1
Clearly, the first integral Ris bounded by cA. To estimate the second inte-
gral it suffices to bound |λ|>1 |F̂ (λ)| dλ. An application of the Cauchy-
Schwarz inequality gives
Z µZ ¶1/2 µZ ¶1/2
2 d−1 −d+1
|F̂ (λ)|dλ ≤ |F̂ (λ)| |λ| dλ |λ| dλ .
|λ|>1 |λ|>1 |λ|>1
This last integral is convergent precisely when −d + 1 < −1, which is

equivalent to d > 2, namely d ≥ 3, which we assume. Hence |F (t)| ≤
c(A + B) as desired.
To establish Lipschitz continuity, we first note that
Z ∞ £ ¤
F (t1 ) − F (t2 ) = F̂ (λ) e2πiλt1 − e2πiλt2 dλ.
−∞
Since one has the inequality2 |eix − 1| ≤ |x|, we immediately see that
|e2πiλt1 − e2πiλt2 | ≤ c|t1 − t2 |α λα if 0 ≤ α < 1.
We may then write the difference F (t1 ) − F (t2 ) as a sum of two inte-
grals. The integral over |λ| ≤ 1 is clearly bounded by cA|t1 − t2 |α . The
second integral, the one over |λ| > 1, can be estimated from above by
Z
α
|t1 − t2 | |F̂ (λ)||λ|α dλ.
|λ|>1
An application of the Cauchy-Schwarz inequality show that this last in-

tegral is majorized by
µZ ¶1/2 µZ ¶1/2
2 d−1 −d+1+2α
|F̂ (λ)| |λ| dλ |λ| dλ ≤ cα B,
|λ|>1 |λ|>1
since the second integral is finite if −d + 1 + 2α < −1, and in particular

this holds if α < 1/2 when d ≥ 3. This concludes the proof of the lemma.
We now gather these results to prove the theorem. For each γ ∈ S d−1
let
F (t) = R(f )(t, γ).
Note that with this definition we have
sup |F (t)| = R∗ (f )(γ).

t∈R
Let
Z ∞
A(γ) = sup |F̂ (λ)| and B 2 (γ) = |F̂ (λ)|2 |λ|d−1 dλ.
λ −∞
Then by (4)
sup |F (t)| ≤ c(A(γ) + B(γ)).
t∈R
However, we observed that F̂ (λ) = fˆ(λγ), and hence
A(γ) ≤ kf kL1 (Rd ) .
2 The distance in the plane from the point eix to the point 1 is shorter than the length
of the arc on the unit circle joining them.

Therefore,
|R∗ (f )(γ)|2 ≤ c(A(γ)2 + B(γ)2 ),
and thus
Z
|R∗ (f )(γ)|2 dσ(γ) ≤ c(kf k2L1 (Rd ) + kf k2L2 (Rd ) ),
S d−1
R
since B 2 (γ) dσ(γ) = 2kf k2L2 by Lemma 4.8. Consequently,
Z
R∗ (f )(γ) dσ(γ) ≤ c(kf kL1 (Rd ) + kf kL2 (Rd ) ).
S d−1
Note that the identity we have used,

Z ∞
R(f )(t, γ) = F̂ (λ)e2πiλt dλ,
−∞
with F (t) = R(f )(t, γ), is justified for almost every γ ∈ S d−1 by the
Fourier inversion result in Theorem 4.2 of Chapter 2. Indeed, we have
seen that A(γ) and B(γ) are finite for almost every γ, and thus F̂ is
integrable for those γ. This completes the proof of the theorem. The
corollary follows the same way if we use (5) instead of (4).
We now return to the situation in the plane to see what information
we may deduce from the above analysis. The inequality (2) as it stands
does not hold when d = 2. However, a modification of it does hold, and
this will be used in the proof of Theorem 4.4.
If f ∈ L1 (Rd ) we define
Z t+δ
1
Rδ (f )(t, γ) = R(f )(s, γ) ds
2δ t−δ
Z
1
= f (x) dx.
2δ t−δ≤x·γ≤t+δ
In this definition of Rδ (f )(t, γ) we integrate the function f in a small

“strip” of width 2δ around the plane Pt,γ . Thus Rδ is an average of
Radon transforms.
We let
R∗δ (f )(γ) = sup |Rδ (f )(t, γ)|.

t∈R
Theorem 4.10 If f is continuous with compact support, then

Z
¡ ¢
R∗δ (f )(γ) dσ(γ) ≤ c(log 1/δ)1/2 kf kL1 (R2 ) + kf kL2 (R2 )
S1
when 0 < δ ≤ 1/2.
The same argument as in the proof of Theorem 4.5 applies here, except
that we need a modified version of Lemma 4.9. More precisely, let us set
Z ∞ µ 2πi(t+δ)λ ¶
e − e2πi(t−δ)λ
Fδ (t) = F̂ (λ) dλ,
−∞ 2πiλ(2δ)
and suppose that

Z ∞
sup |F̂ (λ)| ≤ A and |F̂ (λ)|2 |λ| dλ ≤ B.
λ −∞
Then we claim that
(6) sup |Fδ (t)| ≤ c(log 1/δ)1/2 (A + B).

t
Indeed, we use the fact that |(sin x)/x| ≤ 1 to see that, in the definition
of Fδ (t), the integral over |λ| ≤ 1 gives the cA. Also, the integral over
|λ| > 1 can be split and is bounded by the sum
Z Z
c
|F̂ (λ)| dλ + |F̂ (λ)||λ|−1 dλ.
1<|λ|≤1/δ δ 1/δ≤|λ|
The first integral above can be estimated by the Cauchy-Schwarz in-

equality, as follows
Z µZ ¶1/2µZ ¶1/2
2 −1
|F̂ (λ)| dλ ≤ c |F̂ (λ)| |λ| dλ |λ| dλ
1<|λ|≤1/δ 1<|λ|≤1/δ 1<|λ|≤1/δ
≤ cB(log 1/δ)1/2 .
Finally, we also note that

Z µZ ¶1/2 µZ ¶1/2
c −1 2 1 −3
|F̂ (λ)||λ| dλ ≤ c |F̂ (λ)| |λ| dλ |λ| dλ
δ 1/δ≤|λ| 1/δ≤|λ| δ 1/δ≤|λ|
≤ cB
and this establishes (6), and hence the theorem.

4.2 Regularity of sets when d ≥ 3

We now extend to the general context the basic estimates for the Radon
transform, proved for continuous functions of compact support. This will
yield the regularity result formulated in Theorem 4.1.
Proposition 4.11 Suppose d ≥ 3, and let f belong to L1 (Rd ) ∩ L2 (Rd ).

Then for a.e. γ ∈ S d−1 we can assert the following:
(a) f is measurable and integrable on the plane Pt,γ , for every t ∈ R.
(b) The function R(f )(t, γ) is continuous in t and satisfies a Lips-

chitz condition with exponent α for each α < 1/2. Moreover, the
inequality (2) of Theorem 4.5 and its variant with (3) hold for f .
We prove this in a series of steps.

Step 1. We consider f = χO , the characteristic function of a bounded
open set O. Here the assertion (a) is evident since O ∩ Pt,γ is an open
and bounded set in Pt,γ and is bounded. Thus R(f )(t, γ) is defined for
all (t, γ).
Next we can find a sequence {fn } of non-negative continuous func-
tions of compact support so that for every x, fn (x) increases to f (x) as
n → ∞. Thus R(fn )(t, γ) → R(f )(t, γ) for every (t, γ) by the monotone
convergence theorem, and also R∗ (fn )(γ) → R∗ (f )(γ) for each γ ∈ S d−1 .
As a result we see that the inequality (2) is valid for f = χO , with O
open and bounded.
Step 2. We now consider f = χE , where E is a set of measure zero,
and take first the case when the set E is bounded. Then we can find a
decreasing sequence {On } of open and bounded sets, such that E ⊂ On ,
while m(OnT ) → 0 as n → ∞.
Let Ẽ = On . Since Ẽ ∩ Pt,γ is measurable for every (t, γ), the func-
tions R(χẼ )(t, γ) and R∗ (χẼ )(γ) are well-defined. However, R∗ (χẼ )(γ) ≤
R∗ (χOn )(γ), while the R∗ (χOn ) decrease. Thus the inequality (2) we
have just proved for f = χOn shows that R∗ (χẼ )(γ) = 0 for a.e. γ. The
fact that E ⊂ Ẽ then implies that for a.e. γ, the set E ∩ Pt,γ has (d − 1)-
dimensional measure zero for every t ∈ R. This conclusion immediately
extends to the case when E is not necessarily bounded, by writing E as a
countable union of bounded sets of measure zero. Therefore Corollary 4.2
is proved.
Step 3. Here we assume that f is a bounded measurable function
supported on a bounded set. Then by familiar arguments we can find
a sequence {fn } of continuous functions that are uniformly bounded,
supported in a fixed compact set, and so that fn (x) → f (x) a.e. By the
bounded convergence theorem, kfn − f kL1 and kfn − f kL2 both tend to
zero as n → ∞, and upon selecting a subsequence if necessary, we can
suppose that kfn − f kL1 + kfn − f kL2 ≤ 2−n . By what we have just
proved in Step 2 we have, for a.e. γ ∈ S d−1 , that fn (x) → f (x) on Pt,γ
a.e. with respect to the measure md−1 , for each t ∈ R. Thus again by the
bounded convergence theorem for those (t, γ), we see that R(fn )(t, γ) →
R(f )(t, γ), and this limit defines R(f ). Now applying Theorem 4.5 to
fn − fn−1 gives
∞ Z
X ∞
X
R∗ (fn − fn−1 )(γ) dσ(γ) ≤ c 2−n < ∞.
n=1 S d−1 n=1
This means that

X
sup |R(fn )(t, γ) − R(fn−1 )(t, γ)| < ∞,
t
n
for a.e.γ ∈ S d−1 , and hence for those γ the sequence of functions R(fn )(t, γ)
converges uniformly. As a consequence, for those γ the function R(f )(t, γ)
is continuous in t, and the inequality (2) is valid for this f . The inequality
with (3) is deduced in the same way.
Finally, we deal with the general f in L1 ∩ L2 by approximating it by
a sequence of bounded functions each with bounded support. The details
of the argument are similar to the case treated above and are left to the
reader.
Observe that the special case f = χE of the proposition gives us The-
orem 4.1.
4.3 Besicovitch sets have dimension 2

Here we prove Theorem 4.4, that any Besicovitch set necessarily has
Hausdorff dimension 2. We use Theorem 4.10, namely, the inequality
Z
¡ ¢
R∗δ (f )(γ) dσ(γ) ≤ c(log 1/δ)1/2 kf kL1 (R2 ) + kf kL2 (R2 ) .
S1
This inequality was proved under the assumption that f was continuous
and had compact support. In the present situation it goes over without
difficulty to the general case where f ∈ L1 ∩ L2 , by an easy limiting
argument, since it is clear that R∗δ (fn )(γ) converges to R∗δ (f )(γ) for all
γ if fn → f in the L1 -norm.
Now suppose FSis a Besicovitch set and α is fixed with 0 < α < 2.
∞
Assume that F ⊂ i=1 Bi is a covering, where Bi are balls with diameter
less than a given number. We must show that
X
(diam Bi )α ≥ cα > 0.
i
We proceed in two steps, considering first a simple situation that will

make clear the idea of the proof.
Case 1. We suppose first that all the balls Bi have the same diameter
δ (with δ ≤ 1/2) and also that there are only a finite number, say N , of
balls in the covering. We must prove that N δ αS≥ cα .
Let Bi∗ denote the double of Bi and F ∗ = i Bi∗ . Then, we clearly
have
m(F ∗ ) ≤ cN δ 2 .
Since F is a Besicovitch set, for each γ ∈ S 1 there is a segment sγ of

unit length, perpendicular to γ, and which is contained in F . Also, by
construction, any translate by less than δ of a point in sγ must belong
to F ∗ . Hence
R∗δ (χF ∗ )(γ) ≥ 1 for every γ.
If we take f = χF ∗ in the inequality (6), and note that the Cauchy-

Schwarz inequality implies
kχF ∗ kL1 (R2 ) ≤ ckχF ∗ kL2 (R2 ) ≤ c(m(F ∗ ))1/2 ,
then we obtain
c ≤ N 1/2 δ(log 1/δ)1/2 .
This implies N δ α ≥ c for α < 2.

S∞
Case 2. We now treat the general case. Suppose F ⊂ i=1 Bi , where
the balls Bi each have diameter less than 1. For each integer k let Nk be
the number of balls in the collection {Bi } for which
2−k−1 ≤ diam Bi ≤ 2−k .

P∞
We need to show that k=0 Nk 2−kα ≥ cα . In fact, we shall prove the
0
stronger result that there exists a positive integer k 0 such that Nk0 2−k α ≥
cα .
Let
 
\ [
Fk = F  Bi  ,
2−k−1 ≤diam Bi ≤2−k
and let
[
Fk∗ = Bi∗ ,
2−k−1 ≤diam Bi ≤2−k
where Bi∗ denotes the double of Bi . Then we note that
m1 (Fk∗ ) ≤ cNk 2−2k for all k.
Since F is a Besicovitch set, for each γ ∈ S 1 there is a segment sγ of

unit length entirely contained in F . We now make precise the fact that
for some k, a large proportion of sγ belongs to Fk .
∞
PWe pick a sequence of real numbers {ak }k=0 such that 0 ≤ ak ≤ 1,
ak = 1, but ak does not tend to zero too quickly. For instance, we
may choose ak = c² 2−²k with c² = 1 − 2−² , and ² > 0 but ² sufficiently
small.
Then, for some k we must have
m1 (sγ ∩ Fk ) ≥ ak .
S
Otherwise, since F = Fk , we would have
X
m1 (sγ ∩ F ) < ak = 1,
and this contradicts the fact that m1 (sγ ∩ F ) = 1, since sγ is entirely

contained in F .
Therefore, with this k, we must have
R∗2−k (χFk∗ )(γ) ≥ ak ,
because any point of distance less than 2−k from Fk must belong to Fk∗ .
Since the choice of k may depend on γ, we let
Ek = {γ : R∗2−k (χFk∗ )(γ) ≥ ak }.
By our previous observations, we have

∞
[
S1 = Ek ,
k=1
and so for at least one k, which we denote by k 0 , we have
m(Ek0 ) ≥ 2πak0 ,
P
for otherwise m(S1 ) < 2π ak = 2π. As a result
2πa2k0 = 2πak0 ak0

Z
≤ ak0 dσ(γ)
Ek 0
Z
≤ R∗2−k0 (χFk∗0 )(γ) dσ(γ).
S1
By the fundamental inequality (6) we get

0
a2k0 ≤ c(log 2k )1/2 kχFk∗0 kL2 (R2 ) .
Recalling that by our choice ak ≈ 2−²k , and noting that kχFk∗0 kL2 ≤
1/2 0
cNk0 2−k , we obtain
0 0 1/2
2(1−2²)k ≤ c(log 2k )1/2 Nk0 .
0
Finally, this last inequality guarantees that Nk0 2−αk ≥ cα as long as
4² < 2 − α.
This concludes the proof of the theorem.
4.4 Construction of a Besicovitch set

There are a number of different constructions of Besicovitch sets. The one
we have chosen to describe here involves the concept of self-replicating
sets, an idea that permeates much of the discussion of this chapter.
We consider the Cantor set of constant dissection C1/2 , which for sim-
plicity we shall write
T∞ as C, and which is defined in Exercise 3, Chapter 1.
Note that C = k=0 Ck , where C0 = [0, 1], and Ck is the union of 2k
closed intervals of length 4−k obtained by removing from Ck−1 the 2k−1
centrally situated open intervals of length 21 · 4−k+1 . The set
PC∞can also
be represented as the set of points x ∈ [0, 1] of the form x = k=1 ²k /4k ,
with ²k either 0 or 3.
We now place a copy of C on the x-axis of the plane R2 = {(x, y)}, and a
copy of 21 C on the line y = 1. That is, we put E0 = {(x, y) : x ∈ C, y = 0}
and E1 = {(x, y) : 2x ∈ C, y = 1}. The set F that will play the central
role is defined as the union of all line segments that join a point of E0
with a point of E1 . (See Figure 13.)
Figure 13. Several line segments joining E0 with E1
Theorem 4.12 The set F is compact and of two-dimensional measure

zero. It contains a translate of any unit line segment whose slope is a
number s that lies outside the intervals (−1, 2).
Once the theorem is proved, our job is done. Indeed, a finite union of
rotations of the set F contains unit segments of any slope, and that set
is therefore a Besicovitch set.
The proof of the required properties of the set F amounts to showing
the following paradoxical facts about the set C + λC, for λ > 0. Here
C + λC = {x1 + λx2 : x1 ∈ C, x2 ∈ C}:
• C + λC has one-dimensional measure zero, for a.e. λ.
• C + 12 C is the interval [0, 3/2].
Let us see how these two assertions imply the theorem. First, we note
that the set F is closed (and hence compact), because both E0 and E1
are closed. Next observe that with 0 < y < 1, the slice F y of the set
F is exactly (1 − y)C + y2 C. This set is obtained from the set C + λC,
where λ = y/(2(1 − y)), by scaling with the factor 1 − y. Hence F y is of
measure zero whenever C + λC is also of measure zero. Moreover, under
the mapping y 7→ λ, sets of measure zero in (0, ∞) correspond to sets of
measure zero in (0, 1). (For this see, for example, Exercise 8 in Chapter 1,
or Problem 1 in Chapter 6.) Therefore, the first assertion and Fubini’s
theorem prove that the (two-dimensional) measure of F is zero.
Finally the slope s of the segment joining the point (x0 , 0), with the
point (x1 , 1) is s = 1/(x1 − x0 ). Thus the quantity s can be realized if
x1 ∈ C/2 and x0 ∈ C, that is, if 1/s ∈ C/2 − C. However, by an obvious

symmetry C = 1 − C, and so the condition becomes 1/s ∈ C/2 + C − 1,
which by the second assertion is 1/s ∈ [−1, 1/2]. This last is equivalent
with s ∈
/ (−1, 2).
Our task therefore remains the proof of the two assertions above. The
proof of the second is nearly trivial. In fact,
µ ¶
2 1 2 1
C + C = C + C,
3 2 3 3
P∞ ³ ²0
´
and this set consists of all x of the form x = k=1 2²3k + 3k 4−k , where
²0
²k and ²0k are independently 0 or 3. Since ¡then 2²3¢k + 3k can take any
of the values 0, 1, 2, or 3, we have that 32 C + 21 C = [0, 1], and hence
C + 12 C = [0, 3/2].
The proof that m(C + λC) = 0 for a.e. λ

We come to the main point: that C + λC has measure zero for almost all
λ. We show this by examining the self-replicating properties of the sets
C and C + λC.
We know that C = C1 ∪ C2 , where C1 and C2 are two similar copies
of C, obtained with similarity ratio 1/4, and given by C1 = 41 C and
C2 = 14 C + 34 . Thus C1 ⊂ [0, 1/4] and C2 ⊂ [3/4, 1]. Iterating ` times this
decomposition of C, that is, reaching the `th “generation,” we can write
[
(7) C= Cj` ,
1≤j≤2`
with C1` = (1/4)` C and each Cj` a translate of C1` .

We consider in the same way the set
K(λ) = C + λC,
and we shall sometimes omit the λ and write K(λ) = K, when this causes
no confusion. By its definition we have
K = K1 ∪ K2 ∪ K3 ∪ K4 ,
where K1 = C1 + λC1 , K2 = C1 + λC2 , K3 = C2 + λC1 , and K4 = C2 + λC2 .

An iteration of this decomposition using (7) gives
[
(8) K= Ki` ,
1≤i≤4`
where each Ki` equals Cj`1 + λCj`2 for a pair of indices j1 , j2 . In fact,
this relation among the indices sets up a bijection between the i with
1 ≤ i ≤ 4` , and the pair j1 , j2 with 1 ≤ j1 ≤ 2` and 1 ≤ j2 ≤ 2` . Note
that each Ki` is a translate of K1` , and each Ki` is also
S obtained from K by
a similarity of ratio 4−` . Now note that C = C/4 (C/4 + 3/4) implies
that
λ λ 3λ
K(λ) = C + λC = (C + C) ∪ (C + C + )
4 4 4
3λ
= K(λ/4) ∪ (K(λ/4) + ).
4
Thus K(λ) has measure zero if and only if K(λ/4) has measure zero.
Hence it suffices to prove that K(λ) has measure zero for a.e. λ ∈ [1, 4].
After these preliminaries let us observe that we immediately obtain
that m(K(λ)) = 0 for some special λ’s, those for which the following
coincidence takes place: for some ` and a pair i and i0 with i 6= i0 ,
Ki` (λ) = Ki`0 (λ).
Indeed, if we have this coincidence, then (8) gives
`
4
X
m(K(λ)) ≤ m(Ki` (λ)) = (4` − 1)4−` m(K(λ)),
i=1, i6=i0
and this implies m(K(λ)) = 0.

The key insight below is that, in a quantitative sense, the λ’s for which
this coincidence takes place are “dense” relative to the size of `. More
precisely, we have the following.
Proposition 4.13 Suppose λ0 and ` are given, with 1 ≤ λ0 ≤ 4 and `

a positive integer. Then, there exist a λ and a pair i, i0 with i 6= i0 such
that
(9) Ki` (λ) = Ki`0 (λ) and |λ − λ0 | ≤ c4−` .
Here c is a constant independent of λ0 and `.
This is proved on the basis of the following observation.
Lemma 4.14 For every λ0 there is a pair 1 ≤ i1 , i2 ≤ 4, with i1 6= i2

such that Ki1 (λ0 ) and Ki2 (λ0 ) intersect.
Proof. Indeed, if the Ki are disjoint for 1 ≤ i ≤ 4 then for sufficiently

small δ the Kiδ are also disjoint. Here we have used the notation that F δ
denotes the set of points of distance
S4 less than δ from F . (See Lemma 3.1
δ
in Chapter 1.) However, K = i=1 Ki , and by similarity m(K4δ ) =
δ
4m(Kiδ ). Thus by the disjointness of the Kiδ we have m(Kδ ) = m(K4δ ),

which is a contradiction, since K4δ − Kδ contains an open ball (of radius
3δ/2). The lemma is therefore proved.
Now apply the lemma for our given λ0 and write Ki1 = Cµ1 + λ0 Cν1 ,
Ki2 = Cµ2 + λ0 Cν2 , where the µ’s and ν’s are either 1 or 2. However, since
i1 6= i2 we have µ1 6= µ2 or ν1 6= ν2 (or both). Assume for the moment
that ν1 6= ν2 .
The fact that Ki1 (λ0 ) and Ki2 (λ0 ) intersect means that there are pairs
of numbers (a, b) and (a0 , b0 ), with a ∈ Cµ1 , b ∈ Cν1 , a0 ∈ Cµ2 , and b0 ∈ Cν2
such that
(10) a + λ0 b = a0 + λ0 b0 .
Note that the fact that ν1 6= ν2 means that |b − b0 | ≥ 1/2. Next, look-
ing at the `th generation we find via (7) that there are indices 1 ≤
j1 , j2 , j10 , j20 ≤ 2` , so that a ∈ Cj`1 ⊂ Cµ1 , b ∈ Cj`2 ⊂ Cν1 , a0 ∈ Cj`0 ⊂ Cµ2 , b0 ∈
1
Cj`0 ⊂ Cν2 . We also observe that the above sets are translates of each
2
other, that is, Cj`1 = Cj`0 + τ1 and Cj`2 = Cj`0 + τ2 , with |τk | ≤ 1. Hence if
1 2
i and i0 correspond to the pairs (j1 , j2 ) and (j10 , j20 ), respectively, we have
(11) Ki` (λ) = Ki`0 (λ) + τ (λ) with τ (λ) = τ1 + λτ2 .
Now let (A, B) be the pair that corresponds to (a0 , b0 ) under the above
translations, namely
(12) A = a0 + τ1 , B = b0 + τ2 .
We claim there is a λ such that
(13) A + λB = a0 + λb0 .
In fact, by (12) we have put B in Cj`2 ⊂ Cν1 , while b0 is in Cj`0 ⊂ Cν2 . Thus
2
|B − b0 | ≥ 1/2, since ν1 6= ν2 . We can therefore solve (13) by taking
λ = (A − a0 )/(b0 − B). Now we compare this with (10), and get λ0 =
(a − a0 )/(b0 − b). Moreover, |A − a| ≤ 4−` and |B − b| ≤ 4−` , since A
and a both lie in Cj`1 , and B and b lie in Cj`2 . This yields the inequality
(14) |λ − λ0 | ≤ c4−` .
Also, (12) and (13) clearly imply τ (λ) = τ1 + λτ2 = 0, and this together
with (11) proves the coincidence.
Therefore our proposition is proved under the restriction we made
earlier that ν1 6= ν2 . The situation where instead µ1 6= µ2 is obtained
from the case ν1 6= ν2 if we replace λ0 by λ−1 `
0 . Note that Ki (λ0 ) =
` ` ` ` `
Ki0 (λ0 ) if and only if Cj1 + λ0 Cj2 = Cj 0 + λ0 Cj 0 and this is the same as
1 2
Cj`2 + λ−1 ` ` −1 `
0 Cj1 = Cj20 + λ0 Cj10 . This allows us to reduce to the case µ1 6=
µ2 , since Cj`1 ⊂ Cµ1 and Cj`0 ⊂ Cµ2 . Here the fact that 1 ≤ λ0 ≤ 4 gives
1
λ−1
0 ≤ 1 and guarantees that the constant c in (9) can be taken to be
independent of λ0 . The proposition is therefore established.
Note that as a consequence, the following holds near the points λ where
the coincidence (9) takes place: If |λ − λ| ≤ ²4−` , then
(15) Ki` (λ) = Ki`0 (λ) + τ (λ) with |τ (λ)| ≤ ²4−` .
In fact, this is (11) together with the observation that
|τ (λ)| = |τ (λ) − τ (λ)| ≤ |λ − λ|,
since |τ (λ)| = τ1 + λτ2 and |τ2 | ≤ 1.

The assertion (15) leads to the following more elaborate version of
itself:
There is a set Λ of full measure such that whenever λ ∈ Λ

and ² > 0 are given, there are ` and a pair i, i0 so that (15)
holds.3
Indeed, for fixed ² > 0, let Λ² denote the set of λ that satisfies (15) for
some `, i and i0 . For any interval I of length not exceeding 1, we have
m(Λ² ∩ I) ≥ ²4−` ≥ c−1 ²m(I),
because of (9) and (15). Thus Λc² has no points of Lebesgue density,
hence Λc² has measure zero, and thus ΛT ² is a set of full measure. (See
Corollary 1.5 in Chapter 3.) Since Λ = ² Λ² , and Λ² decreases with ²,
we see that Λ also has full measure and our assertion is proved.
Finally, our theorem will be established once we show that m(K(λ)) =
0 whenever λ ∈ Λ. To prove this, we assume contrariwise that m(K(λ)) >
0. Using again the point of density argument, there must be for any
3 The terminology that Λ has “full measure” means that its complement has measure
zero.
0 < δ < 1, a non-empty open interval I with m(K(λ) ∩ I) ≥ δm(I). We

then fix δ with 1/2 < δ < 1 and proceed. With this fixed δ, we select
² used below as ² = m(I)(1 − δ). Next, find `, i, and i0 for which (15)
holds. The existence of such indices is guaranteed by the hypothesis that
λ ∈ Λ.
We then consider the two similarities (of ratio 4−` ) that map K(λ) to
Ki (λ) and Ki`0 (λ), respectively. These take the interval I to correspond-
`
ing intervals Ii and Ii0 , respectively, with m(Ii ) = m(Ii0 ) = 4−` m(I).
Moreover,
m(Ki` ∩ Ii ) ≥ δm(Ii ) and m(Ki`0 ∩ Ii0 ) ≥ δm(Ii0 ).
Also, as in (15), Ii0 = Ii + τ (λ), with |τ (λ)| ≤ ²4−` . This shows that
m(Ii ∩ Ii0 ) ≥ m(Ii ) − τ (λ) ≥ 4−` m(I) − ²4−` ≥ δm(Ii ),
since ²4−` = (1 − δ)m(Ii ). Thus m(Ii − Ii ∩ Ii0 ) ≤ (1 − δ)m(Ii ), and
m(Ki` ∩ Ii ∩ Ii0 ) ≥ m(Ki` ∩ Ii ) − m(Ii − Ii ∩ Ii0 )

≥ (2δ − 1)m(Ii )
1 1
> m(Ii ) ≥ m(Ii ∩ Ii0 ).
2 2
So m(Ki` ∩ Ii ∩ Ii0 ) > 12 m(Ii ∩ Ii0 ) and the same holds for i0 in place of i.
Hence m(Ki` ∩ Ki`0 ) > 0, and this contradicts the decomposition (8) and
the fact that m(Ki` ) = 4−` m(K) for every i. Therefore we obtain that
m(K(λ)) = 0 for every λ ∈ Λ, and the proof of Theorem 4.12 is now
complete.
5 Exercises
1. Show that the measure mα is not σ-finite on Rd if α < d.
2. Suppose E1 and E2 are two compact subsets of Rd such that E1 ∩ E2 contains

at most one point. Show directly from the definition of the exterior measure that
if 0 < α ≤ d, and E = E1 ∪ E2 , then
m∗α (E) = m∗α (E1 ) + m∗α (E2 ).
[Hint: Suppose E1 ∩ E2 = {x}, let B² denote the open ball centered at x and of
diameter ², and let E ² = E ∩ B²c . Show that
m∗α (E ² ) ≥ Hα
²
(E) ≥ m∗α (E) − µ(²) − ²α ,
5. Exercises 381
where µ(²) → 0. Hence m∗α (E ² ) → m∗α (E).]
3. Prove that if f : [0, 1] → R satisfies a Lipschitz condition of exponent γ > 1,

then f is a constant.
4. Suppose f : [0, 1] → [0, 1] × [0, 1] is surjective and satisfies a Lipschitz condition
|f (x) − f (y)| ≤ C|x − y|γ .
Prove that γ ≤ 1/2 directly, without using Theorem 2.2.

[Hint: Divide [0, 1] into N intervals of equal length. The image of each sub-interval
is contained in a ball of volume O(N −2γ ), and the union of all these balls must
cover the square.]
5. Let f (x) = xk be defined on R, where k is a positive integer and let E be a

Borel subset of R.
(a) Show that if mα (E) = 0 for some α, then mα (f (E)) = 0.
(b) Prove that dim(E) = dim f (E).
6. Let {Ek } be a sequence of Borel sets in Rd . Show that if dim Ek ≤ α for some
α and all k, then
[
dim Ek ≤ α.
k
7. Prove that the (log 2/ log 3)-Hausdorff measure of the Cantor set is precisely
equal to 1.
[Hint: Suppose we have a covering of C by finitely many closed intervals {Ij }.
Then there exists P
another covering
P of C by intervals {I`0 } each of length 3−k for
some k, such that j |Ij |α ≥ ` |I`0 |α ≥ 1, where α = log 2/ log 3.]
8. Show that the Cantor set of constant dissection, Cξ , in Exercise 3 of Chapter 1

has strict Hausdorff dimension log 2/ log(2/(1 − ξ)).
9. Consider the set Cξ1 × Cξ2 in R2 , with Cξ as in the previous exercise. Show that
Cξ1 × Cξ2 has strict Hausdorff dimension dim(Cξ1 ) + dim(Cξ2 ).
10. Construct a Cantor-like set (as in Exercise 4, Chapter 1) that has Lebesgue
measure zero, yet Hausdorff dimension 1.
P
[Hint: Choose `1 , `2 , . . . , `k , . . . so that 1 − kj=1 2j−1 `j tends to zero sufficiently
slowly as k → ∞.]
11. Let D = Dµ be the Cantor dust in R2 given as the product Cξ × Cξ , with

µ = (1 − ξ)/2.
(a) Show that for any real number λ, the set Cξ + λCξ is similar to the projection
of D on the line in R2 with slope λ = tan θ.
(b) Note that among the Cantor sets Cξ , the value ξ = 1/2 is critical in the
construction of the Besicovitch set in Section 4.4. In fact, prove that with
ξ > 1/2, then Cξ + λCξ has Lebesgue measure zero for every λ. See also
Problem 10 below.
[Hint: mα (Cξ + λCξ ) < ∞ for α = dim Dµ .]
12. Define a primitive one-dimensional “measure” m̃1 as

∞
X ∞
[
m̃1 = inf diam Fk , E⊂ Fk .
k=1 k=1
This is akin to the one-dimensional exterior measure m∗α , α = 1, except that no

restriction is placed on the size of the diameters Fk .
Suppose I1 and I2 are two disjoint unit segments in Rd , d ≥ 2, with I1 = I2 + h,
and |h| < ². Then observe that m̃1 (I1 ) = m̃1 (I2 ) = 1, while m̃1 (I1 ∪ I2 ) ≤ 1 + ².
Thus
m̃1 (I1 ∪ I2 ) < m̃1 (I1 ) + m̃1 (I2 ) when ² < 1;
hence m̃1 fails to be additive.
13. Consider the von Koch curve K` , 1/4 < ` < 1/2, as defined in Section 2.1.
Prove for it the analogue of Theorem 2.7: the function t 7→ K` (t) satisfies a Lip-
schitz condition of exponent γ = log(1/`)/ log 4. Moreover, show that the set K`
has strict Hausdorff dimension α = 1/γ.
[Hint: Show that if O is the shaded open triangle indicated in Figure 14, then O ⊃
S0 (O) ∪ S1 (O) ∪ S2 (O) ∪ S3 (O), where S0 (x) = `x, S1 (x) = ρθ (`x) + a, S2 (x) =
ρ−1
θ (`x) + c, and S3 (x) = `x + b, with ρθ the rotation of angle θ. Note that the
sets Sj (O) are disjoint.]
` `
` a b `
Figure 14. The open set O in Exercise 13
14. Show that if ` < 1/2, the von Koch curve t 7→ K` (t) in Exercise 13 is a simple
curve.
5. Exercises 383
P∞
[Hint: Observe that if t = j=1 aj /4j , with aj = 0, 1, 2, or 3, then
∞
\ ` ` ´´
{K(t)} = Saj · · · Sa2 Sa1 (O) .]
j=1
15. Note that if we take ` = 1/2 in the definition of the von Koch curve in
Exercise 13 we get a “space-filling” curve, one that fills the right triangle whose
vertices are (0, 0), (1, 0), and (1/2, 1/2). The first three steps of the construction
are as in Figure 15, with the intervals traced out in the indicated order.
8 9
2 3 7 10
6 11
2 3 5 12 14 15
1 4 1 4 13 16
Figure 15. The first three steps of the von Koch curve when ` = 1/2
16. Prove that the von Koch curve t 7→ K` (t), 1/4 < ` ≤ 1/2 is continuous but
nowhere differentiable.
[Hint: If K0 (t) exists for some t, then
K(un ) − K(vn )
lim
n→∞ un − vn
must exist, where un ≤ t ≤ vn , and un − vn → 0. Choose un = k/4n and vn =

(k + 1)/4n .]
17. For a compact set E in Rd , define #(²) to be the least number of balls of
radius ² that cover E. Note that we always have #(²) = O(²−d ) as ² → 0, and
#(²) = O(1) if E is finite.
One defines the covering dimension of E, denoted by dimC (E), as inf β such
that #(²) = O(²−β ), as ² → 0. Show that dimC (E) = dimM (E), where dimM is the
Minkowski dimension discussed in Section 2.1, by proving the following inequalities
for all δ > 0:
(i) m(E δ ) ≤ c#(δ)δ d .
(ii) #(δ) ≤ c0 m(E δ )δ −d .
[Hint: To prove (ii), use Lemma 1.2 in Chapter 3 to find a collection of disjoint
balls B1 , B2 , . . . , BN of radius δ/3, each centered at E, such that their “triples”
B̃1 , B̃2 , . . . , B̃N (of radius δ) cover E. Then #(δ) ≤ N , while N m(Bj ) = cN δ d ≤
m(E δ ), since the balls Bj are disjoint and are contained in E δ .]
18. Let E be a compact set in Rd .
(a) Prove that dim(E) ≤ dimM (E), where dim and dimM are the Hausdorff and
Minkowski dimensions, respectively.
(b) However, prove that if E = {0, 1/ log 2, 1/ log 3, . . . , 1/ log n, . . .}, then
dimM E = 1, yet dim E = 0.
19. Show that there is a constant cd , dependent only on the dimension d, such
that whenever E is a compact set,
m(E 2δ ) ≤ cd m(E δ ).
[Hint: Consider the maximal function f ∗ , with f = χE δ , and take cd = 6d .]
20. Show that if F is the self-similar set considered in Theorem 2.12, then it has
the same Minkowski dimension as Hausdorff dimension.
[Hint: Each Fk is the union of mk balls of radius cr k . In the converse direction one
sees by Lemma 2.13 that if ² = rk , then each ball of radius ² can contain at most
c0 vertices of the kth generation. So it takes at least mk /c0 such balls to cover F .]
21. From the unit interval, remove the second and fourth quarters (open intervals).
Repeat this process in the remaining two closed intervals, and so on. Let F be the
limiting set, so that
∞
X
F = {x : x = ak /4k ak = 0 or 2}.
k=1
Prove that 0 < m1/2 (F ) < ∞.
22. Suppose F is the self-similar set arising in Theorem 2.9.
(a) Show that if m ≤ 1/rd , then md (Fi ∩ Fj ) = 0 if i 6= j.
(b) However, if m ≥ 1/rd , prove that Fi ∩ Fj is not empty for some i 6= j.
(c) Prove that under the hypothesis of Theorem 2.12
mα (Fi ∩ Fj ) = 0, with α = log m/ log(1/r), whenever i 6= j.

6. Problems 385
23. Suppose S1 , . . . , Sm are similarities with ratio r, 0 < r < 1. For each set E,
let
S̃(E) = S1 (E) ∪ · · · ∪ Sm (E),
and suppose F denotes the unique non-empty compact set with S̃(F ) = F .
(a) If x ∈ F , show that the set of points {S̃ n (x)}∞
n=1 is dense in F .
(b) Show that F is homogeneous in the following sense: if x0 ∈ F and B is

any open ball centered at x0 , then F ∩ B contains a set similar to F .
24. Suppose E is a Borel subset of Rd with dim E < 1. Prove that E is totally
disconnected, that is, any two distinct points in E belong to different connected
components.
[Hint: Fix x, y ∈ E, and show that f (t) = |t − x| is Lipschitz of order 1, and hence
dim f (E) < 1. Conclude that f (E) has a dense complement in R. Pick r in the
complement of f (E) so that 0 < r < f (y), and use the fact that E = {t ∈ E :
|t − x| < r} ∪ {t ∈ E : |t − x| > r}.]
25. Let F (t) be an arbitrary non-negative measurable function on R, and γ ∈ S d−1 .

Then there exists a measurable set E in Rd , such that F (t) = md−1 (E ∩ Pt,γ ).
26. Theorem 4.1 can be refined for d ≥ 4 as follows.

Define C k,α to be the class of functions F (t) on R that are C k and for which
F (k) (t) satisfies a Lipschitz condition of exponent α.
If E has finite measure, then for a.e. γ ∈ S d−1 the function m(E ∩ Pt,γ ) is in
k,α
C for k = (d − 3)/2, α < 1/2, if d is odd, d ≥ 3; and for, k = (d − 4)/2, α < 1,
if d is even, d ≥ 4.
27. Show that the modification of the inequality (2) of Theorem 4.5 fails if we
drop kf kL2 (Rd ) from the right-hand side.
[Hint: Consider R∗ (f² ), with f² defined by f² (x) = (|x| + ²)−d+δ , for |x| ≤ 1, with
δ fixed, 0 < δ < 1, and ² → 0.]
28. Construct a compact set E ⊂ Rd , d ≥ 3, such that md (E) = 0, yet E contains

translates of any segment of unit length in Rd . (While particular examples of such
sets can be easily obtained from the case d = 2, the determination of the least
Hausdorff dimension among all such sets is an open problem.)
6 Problems
1. Carry out the construction below of two sets U and V so that
dim U = dim V = 0 but dim(U × V ) ≥ 1.
Let I1 , . . . , In , . . . be given as follows:

• Each Ij is a finite sequence of consecutive positive integers; that is, for all j
Ij = {n ∈ N : Aj ≤ n ≤ Bj } for some given Aj and Bj .
• For each j, Ij+1 is to the right of Ij ; that is, Aj+1 > Bj .

Let U ⊂ [0, 1] consist of all x which when written
S dyadically x = .a1 a2 · · · an · · ·
have the property that an = 0 whenever n ∈ j Ij . Assume also that Aj and Bj
tend to infinity (as j → ∞) rapidly enough, say Bj /Aj → ∞ and Aj+1 /Bj → ∞.
Also, let Jj be the complementary blocks of integers, that is,
Jj = {n ∈ N : Bj < n < Aj+1 }.

S
Let V ⊂ [0, 1] consist of those x = .a1 a2 · · · an · · · with an = 0 if n ∈ j Jj .
Prove that U and V have the desired property.
2.∗ The iso-diametric inequality states the following: If E is a bounded subset of

Rd and diam E = sup{|x − y| : x, y ∈ E}, then
„ «d
diam E
m(E) ≤ vd ,
2
where vd denotes the volume of the unit ball in Rd . In other words, among sets of
a given diameter, the ball has maximum volume. Clearly, it suffices to prove the
inequality for E instead of E, so we can assume that E is compact.
(a) Prove the inequality in the special case when E is symmetric, that is, −x ∈ E
whenever x ∈ E.
In general, one reduces to the symmetric case by using a technique called Steiner
symmetrization. If e is a unit vector in Rd , and P is a plane perpendicular to e,
the Steiner symmetrization of E with respect to E is defined by
1
S(E, e) = {x + te : x ∈ P, |t| ≤ L(E; e; x)},
2
where L(E; e; x) = m ({t ∈ R : x + t · e ∈ E}), and m denotes the Lebesgue mea-

sure. Note that x + te ∈ S(E, e) if and only if x − te ∈ S(E, e).
(b) Prove that S(E, e) is a bounded measurable subset of Rd that satisfies
m(S(E, e)) = m(E).
[Hint: Use Fubini’s theorem.]
(c) Show that diam S(E, e) ≤ diam E.
(d) If ρ is a rotation that leaves E and P invariant, show that ρS(E, e) =

S(E, e).
(e) Finally, consider the standard basis {e1 , . . . , ed } of Rd . Let E0 = E, E1 =

S(E0 , e1 ), E2 = S(E1 , e2 ), and so on. Use the fact that Ed is symmetric to
prove the iso-diametric inequality.
6. Problems 387
vd
(f) Use the iso-diametric inequality to show that m(E) = 2d
md (E) for any
Borel set E in Rd .
3. Suppose S is a similarity.
(a) Show that S maps a line segment to a line segment.
(b) Show that if L1 and L2 are two segments that make an angle α, then S(L1 )
and S(L2 ) make an angle α or −α.
(c) Show that every similarity is a composition of a translation, a rotation

(possibly improper), and a dilation.
4.∗ The following gives a generalization of the construction of the Cantor-Lebesgue

function.
Let F be the compact set in Theorem 2.9 defined in terms of m similarities
S1 , S2 , . . . , Sm with ratio 0 < r < 1. There exists a unique Borel measure µ sup-
ported on F such that µ(F ) = 1 and
m
1 X
µ(E) = µ(Sj−1 (E)) for any Borel set E.
m j=1
In the case when F is the Cantor set, the Cantor-Lebesgue function is µ([0, x]).
5. Prove a theorem of Hausdorff: Any compact subset K of Rd is a continuous

image of the Cantor set C.
[Hint: Cover K by 2n1 (some n1 ) open balls of radius 1, say B1 , . . . , B` (with
possible repetitions). Let Kj1 = K ∩ Bj1 and cover each Kj1 with 2n2 balls of
radius 1/2 to obtain compact sets Kj1 ,j2 , and so on. Express t ∈ C as a ternary
expansion, and assign to t a unique point in K defined by the intersection Kj1 ∩
Kj1 ,j2 ∩ · · · for appropriate j1 , j2 , . . .. To prove continuity, observe that if two
points in the Cantor set are close, then their ternary expansions agree to high
order.]
6. A compact subset K of Rd is uniformly locally connected if given ² > 0

there exists δ > 0 so that whenever x, y ∈ K and |x − y| < δ, there is a continuous
curve γ in K joining x to y, such that γ ⊂ B² (x) and γ ⊂ B² (y).
Using the previous problem, one can show that a compact subset K of Rd is
the continuous image of the unit interval [0, 1] if and only if K is uniformly locally
connected.
7. Formulate and prove a generalization of Theorem 3.5 to the effect that once
appropriate sets of measure zero are removed, there is a measure-preserving iso-
morphism of the unit interval in R and the unit cube in Rd .
8.∗ There exists a simple continuous curve in the plane of positive two-dimensional
measure.
9. Let E be a compact set in Rd−1 . Show that dim(E × I) = dim(E) + 1, where

I is the unit interval in R.
10.∗ Let Cξ be the Cantor set considered in Exercises 8 and 11. If ξ < 1/2, then
Cξ + λCξ has positive Lebesgue measure for almost every λ.
Notes and References
There are several excellent books that cover many of the subjects treated here.
Among these texts are Riesz and Nagy [27], Wheeden and Zygmund [33], Fol-
land [13], and Bruckner et al. [4].
Introduction
The citation is a translation of a passage in a letter from Hermite to Stieltjes [18].
Chapter 1
The citation is a translation from the French of a passage in [3].
We refer to Devlin [7] for more details about the axiom of choice, Hausdorff
maximal principle, and well-ordering principle.
See the expository paper of Gardner [14] for a survey of results regarding the
Brunn-Minkowski inequality.
Chapter 2
The citation is a passage from the preface to the first edition of Lebesgue’s book
on integration [20].
Devlin [7] contains a discussion of the continuum hypothesis.
Chapter 3
The citation is from Hardy and Littlewood’s paper [15].
Hardy and Littlewood proved Theorem 1.1 in the one-dimensional case by
using the idea of rearrangements. The present form is due to Wiener.
Our treatment of the isoperimetric inequality is based on Federer [11]. This
work also contains significant generalizations and much additional material on
geometric measure theory.
A proof of the Besicovitch covering in the lemma in Problem 3∗ is in Mat-
tila [22].
For an account of functions of bounded variations in Rd , see Evans and
Gariepy [8].
An outline of the proof of Problem 7 (b)∗ can be found at the end of Chapter 5
in Book I.
The result in part (b) of Problem 8∗ is a theorem of S. Saks, and its proof as
a consequence of part (a) can be found in Stein [31].
Chapter 4
The citation is translated from the introduction of Plancherel’s article [25].
An account of the theory of almost periodic functions which is touched upon
in Problem 2∗ can be found in Bohr [2].
The results in Problems 4∗ and 5∗ are in Zygmund [35], in Chapters V and VII,
respectively.
Consult Birkhoff and Rota [1] for more on Sturm-Liouville systems, Legendre
polynomials, and Hermite functions.
Chapter 5
389
390 NOTES AND REFERENCES
See Courant [6] for an account of the Dirichlet principle and some of its applica-
tions. The solution of the Dirichlet problem for general domains in R2 and the
related notion of logarithmic capacity of sets are treated in Ransford [26]. Fol-
land [12] contains another solution to the Dirichlet problem (valid in Rd , d ≥ 2)
by methods which do not use the Dirichlet principle.
The result regarding the existence of the conformal mapping stated in Prob-
lem 3∗ is in Chapter VII of Zygmund [35].
Chapter 6
The citation is a translation from the German of a passage in C. Carathéodory [5].
Petersen [24] gives a systematic presentation of ergodic theory, including a
proof of the theorem in Problem 7∗ .
The facts about spherical harmonics needed in Problem 4∗ can be found in
Chapter 4 in Stein and Weiss [32].
We refer to Hardy and Wright [16] for an introduction to continued fractions.
Their connection to ergodic theory is discussed in Ryll-Nardzewski [28].
Chapter 7
The citation is a translation from the German of a passage in Hausdorff’s arti-
cle [17], while Mandelbrot’s citation is from his book [21].
Mandelbrot’s book also contains many interesting examples of fractals arising
in a variety of different settings, including a discussion of Richardson’s work on
the length of coastlines. (See in particular Chapter 5.)
Falconer [10] gives a systematic treatment of fractals and Hausdorff dimension.
We refer to Sagan [29] for further details on space-filling curves, including the
construction of a curve arising in Problem 8∗ .
The monograph of Falconer [10] also contains an alternate construction of the
Besicovitch set, as well as the fact that such sets must necessarily have dimension
two. The particular Besicovitch set described in the text appears in Kahane [19],
but the fact that it has measure zero required further ideas which are contained,
for instance, in Peres et al. [30].
Regularity of sets in Rd , d ≥ 3, and the estimates for the maximal function
associated to the Radon transform are in Falconer [9], and Oberlin and Stein [23].
The theory of Besicovitch sets in higher dimensions, as well as a number of
interesting related topics can be found in the survey of Wolff [34].
Bibliography
[1] G. Birkhoff and G. C. Rota. Ordinary differential equations. Wiley,

New York, 1989.
[2] H. A. Bohr. Almost periodic functions. Chelsea Publishing Com-
pany, New York, 1947.
[3] E. Borel. Leçons sur la théorie des fonctions. Gauthiers-Villars,
Paris, 1898.
[4] J. B. Bruckner, A. M. Bruckner, and B. S. Thomson. Real Analysis.
Prentice Hall, Upper Saddle River, NJ, 1997.
[5] C. Carathéodory. Vorlesungen über reelle Funktionen. Leipzig,
Berlin, B. G. Teubner, Leipzig and Berlin, 1918.
[6] R. Courant. Dirichlet’s principle, conformal mappings, and minimal
surfaces. Interscience Publishers, New York, 1950.
[7] K. J. Devlin. The joy of sets: fundamentals of contemporary set
theory. Springer-Verlag, New York, 1997.
[8] L. C. Evans and R. F. Gariepy. Measure theory and fine properties
of functions. CRC Press, Boca Raton, 1992.
[9] K. J. Falconer. Continuity properties of k -plane integrals and Besi-
covitch sets. Math. Proc. Cambridge Philos. Soc, 87:221–226, 1980.
[10] K. J. Falconer. The geometry of fractal sets. Cambridge University
Press, 1985.
[11] H. Federer. Geometric measure theory. Springer, Berlin and New
York, 1996.
[12] G. B. Folland. Introduction to partial differential equations. Prince-
ton University Press, Princeton, NJ, second edition, 1995.
[13] G. B. Folland. Real Analysis: modern techniques and their applica-
tions. Wiley, New York, second edition, 1999.
[14] R. J. Gardner. The Brunn-Minkowski inequality. Bull. Amer. Math.
Soc, 39:355–405, 2002.
[15] G. H. Hardy and J. E. Littlewood. A maximal theorem with function
theoretic applications. Acta. Math, 54:81–116, 1930.
391
392 BIBLIOGRAPHY
[16] G. H. Hardy and E. M. Wright. An introduction to the Theory of

Numbers. Oxford University Press, London, fifth edition, 1979.
[17] F. Hausdorff. Dimension und äusseres Mass. Math. Annalen, 79:157–

179, 1919.
[18] C. Hermite. Correspondance d’Hermite et de Stieltjes. Gauthier-

Villars, Paris, 1905. Edited by B. Baillaud and H. Bourget.
[19] J. P. Kahane. Trois notes sur les ensembles parfaits linéaires. En-
seignement Math., 15:185–192, 1969.
[20] H. Lebesgue. Leçons sur l’integration et la recherche des fonctions

primitives. Gauthier-Villars, Paris, 1904. Preface to the first edition.
[21] B. B. Mandelbrot. The fractal geometry of nature. W. H. Freeman,

San Francisco, 1982.
[22] P. Mattila. Geometry of sets and measures in Euclidean spaces.

Cambridge University Press, Cambridge, 1995.
[23] D. M. Oberlin and E. M. Stein. Mapping properties of the Radon

transform. Indiana Univ. Math. J, 31:641–650, 1982.
[24] K. E. Petersen. Ergodic theory. Cambridge University Press, Cam-

bridge, 1983.
[25] M. Plancherel. La théorie des équations intégrales. L’Enseignement

math., 14e Année:89–107, 1912.
[26] T. J. Ransford. Potential theory in the complex plane. London Math-

ematical Society student texts, 28. Cambridge, New York: Press
Syndicate of the University of Cambridge, 1995.
[27] F. Riesz and B. Sz.-Nagy. Functional Analysis. New York, Ungar,

1955.
[28] C. Ryll-Nardzewski. On the ergodic theorem. ii. Ergodic theory of

continued fractions. Studia Math., 12:74–79, 1951.
[29] H. Sagan. Space-filling curves. Universitext. Springer-Verlag, New

York, 1994.
[30] Y. Peres, K. Simon, and B. Solomyak. Fractals with positive length

and zero Buffon needle probability. Amer. Math. Monthly, 110:314–
325, 2003.
BIBLIOGRAPHY 393
[31] E. M. Stein. Harmonic analysis: real-variable methods, orthogonal-

ity, and oscillatory integrals. Princeton University Press, Princeton,
NJ, 1993.
[32] E. M. Stein and G. Weiss. Introduction to Fourier Analysis on
Euclidean Spaces. Princeton University Press, Princeton, NJ, 1971.
[33] R. L. Wheeden and A. Zygmund. Measure and integral: an intro-
duction to real analysis. Marcel Dekker, New York, 1977.
[34] T. Wolff. Recent work connected with the Kakeya problem.

Prospects in Mathematics,Princeton, NJ, 31:129–162, 1996. Amer.
Math. Soc., Providence, RI, 1999.
[35] A. Zygmund. Trigonometric Series, volume I and II. Cambridge
University Press, Cambridge, second edition, 1959. Reprinted 1993.
Symbol Glossary
The page numbers on the right indicate the first time the symbol or
notation is defined or used. As usual, Z, Q, R, and C denote the integers,
the rationals, the reals, and the complex numbers respectively.
|x| (Euclidean) Norm of x 2

Ec, E − F Complements and relative complements of 2
sets
d(E, F ) Distance between two sets 2
Br (x), Br (x) Open and closed balls 2
E, ∂E Closure and boundary of E, respectively 3
|R| Volume of the rectangle R 3
O(· · · ) O notation 12
C, Cξ , Cˆ Cantor sets 9, 38
m∗ (E) Exterior (Lebesgue) measure of the set E 10
Ek % E, Ek & E Increasing and decreasing sequences of sets 20
E4F Symmetric difference of E and F 21
Eh = E + h Translation by h of the set E 22
BRd Borel σ-algebra on Rd 23
Gδ , Fσ Sets of type Gδ or Fσ 23
N Non-measurable set 24
a.e. Almost everywhere 30
f + (x), f − (x) Positive and negative parts of f 31, 64
A+B Sum of two sets 35
vd Volume of the unit ball in Rd 39
supp(f ) Support of the function f 53
fk % f , fk & f Increasing and decreasing sequences of func- 62
tions
fh Translation by h of the function f 73
L1 (Rd ), L1loc (Rd ) Integrable and locally integrable functions 69, 105
f ∗g Convolution of f and g 74
f y , fx , E y , Ex Slices of the function f and set E 75
fˆ, F(f ) Fourier transform of f 87, 208
f∗ Maximal functions of f 100, 296
L(γ) Length of the (rectifiable) curve γ 115
TF , PF , NF Total, positive, and negative variations of F 117, 118
L(A, B) Length of a curve between t = A and t = B 120
D+ (F ), . . . , D− (F ) Dini numbers of F 123
396 SYMBOL GLOSSARY
M(K) Minkowski content of K 138

Ω+ (δ), Ω− (δ) Outer and inner set of Ω 143
L2 (Rd ) Square integrable functions 156
`2 (Z), `2 (N) Square summable sequences 163
H Hilbert space 161
f ⊥g Orthogonal elements 164
D Unit disc 173
H 2 (D), H 2 (R2+ ) Hardy spaces 174, 213
S⊥ Orthogonal complement of S 177
A⊕B Direct sum of A and B 177
PS Orthogonal projection onto S 178
T ∗ , L∗ Adjoint of operators 183, 222
S(Rd ) Schwartz space 208
C0∞ (Ω) Smooth functions with compact support 222
in Ω
C n (Ω), C n (Ω) Functions with n continuous derivatives on 223
Ω and Ω
4u Laplacian of u 230
(X, M, µ), (X, µ) Measure space 263
µ, µ∗ , µ0 Measure, exterior measure, premeasure 263, 264, 270
µ1 × µ2 Product measure 276
S d−1 Unit sphere in Rd 279
σ, dσ(γ) Surface measure on the sphere 280
dF Lebesgue-Stieltjes measure 282
|ν|, ν + , ν − Total, positive, and negative variations of ν 286, 287
ν⊥µ Mutually singular measures 288
ν¿µ Absolutely continuous measures 289
σ(S) Spectrum of S 311
m∗α (E) Exterior α-dimensional Hausdorff measure 325
diam S Diameter of S 325
dim E Hausdorff dimension of E 329
S Sierpinski triangle 334
A≈B A comparable to B 335
K, K` Von Koch curves 338, 340
dist(A, B) Hausdorff distance 345
P(t) Peano mapping 349
Pt,γ Hyperplane 360
R(f ), Rδ (f ) Radon transform 363, 368
R∗ (f ), R∗δ (f ) Maximal Radon transform 363, 368
Index
Relevant items that also arose in Book I or Book II are listed in this
index, preceeded by the numerals I or II, respectively.
Fσ , 23 Borel
Gδ , 23 σ-algebra, 23, 267
σ-algebra measure, 269
Borel, 23 on R, 281
of sets, 23 sets, 23, 267
Borel, 267 Borel-Cantelli lemma, 42, 63
σ-finite, 263 boundary, 3
σ-finite signed measure, 288 boundary-value function, 217
O notation, 12 bounded convergence theorem, 56
bounded set, 3
absolute continuity bounded variation, 116
of the Lebesgue integral, 66 Brunn-Minkowski inequality, 34, 48
absolutely continuous
functions, 127 canonical form, 50
measures, 288 Cantor dust, 47, 343
adjoint, 183, 222 Cantor set, 8, 38, 126, 330, 387
algebra of sets, 270 constant dissection, 38
almost disjoint (union), 4 Cantor-Lebesgue
almost everywhere, a.e., 30 function, 38, 126, 331, 387
almost periodic function, 202 theorem, 95
approximation to the identity, 109; Carathéodory measurable, 264
(I)49 Cauchy
arc-length parametrization, 136; in measure, 95
(I)103 integral, 179, 220; (II)48
area of unit sphere, 313 sequence, 159; (I)24; (II)24
area under graph, 85 Cauchy-Schwarz inequality, 157,
averaging problem, 100 162; (I)72
axiom of choice, 26, 48 chain
of dyadic squares, 352
basis of quartic intervals, 351
algebraic, 202 change of variable formula, 149;
orthonormal, 164 (I)292
Bergman kernel, 254 characteristic
Besicovitch function, 27
covering lemma, 153 polynomial, 221, 258
set, 360, 362, 374 closed set, 2, 267; (II)6
Bessel’s inequality, 166; (I)80 closure, 3
Blaschke factors, 227; (I)26, 153, coincidence, 377
219 compact linear operator, 188
398 INDEX
compact set, 3, 188; (II)6 Dirichlet

comparable, 335 integral, 230
complement of a set, 2 kernel, 179; (I)37
complete principle, 229, 243
L2 , 159 problem, 230; (I)10, 28, 64, 170;
measure space, 266 (II)212, 216
mectric space, 69 distance
completion between two points, 2
Borel σ-algebra, 23 between two sets, 2, 267
Hilbert space, 170; (I)74 Hausdorff, 345
measure space, 312 dominated convergence theorem, 67
complex-valued function, 67 doubling mapping, 304
conjugate Poisson kernel, 255 dyadic
continued fraction, 293, 322 correspondence, 353
continuum hypothesis, 96 induced mapping, 353
contraction, 318 rationals, 351
convergence in measure, 96 square, 352
convex
function, 153 Egorov’s theorem, 33
set, 35 eigenvalue, 186; (I)233
convolution, 74, 94, 253; (I)44, 139, eigenvector, 186
239 equivalent functions, 69
countable unions, 19 ergodic, (I) 111
counting measure, 263 maximal theorem, 297
covering dimension, 383 mean theorem, 295
covering lemma measure-preserving
Vitali, 102, 128, 152 transformation, 302
cube, 4 pointwise theorem, 300
curve extension principle, 183, 210
closed and simple, 137; (I)102; exterior measure, 264
(II)20 Hausdorff, 325
length, 115 Lebesgue, 10
quasi-simple, 137, 332 metric, 267
rectifiable, 115, 134, 332
simple, 137, 332 Fatou’s lemma, 61
space-filling, 349, 383 Fatou’s theorem, 173
von Koch, 338, 340, 382 Fejér kernel, 112; (I)53, 163
cylinder set, 316 finite rank operator, 188
finite-valued function, 27
d’Alembert’s formula, 224 Fourier
dense family of functions, 71 coefficient, 170; (I)16, 34
difference set, 44 inversion formula, 86; (I)141, 182;
differentiation of the integral, 99 (II)115
dimension multiplier operator, 200, 220
Hausdorff, 329 series, 171, 316; (I)34; (II)101
Minkowski, 333 transform in L1 , 87
Dini numbers, 123 transform in L2 , 207, 211
Dirac delta function, 110, 285 fractal, 329
direct sum, 177 Fredholm alternative, 204
INDEX 399
Fubini’s theorem, 75, 276 measure, 327

function strict dimension, 329
absolutely continuous, 127, 285 heat kernel, 111; (I)120, 146, 209
almost periodic, 202 Heaviside function, 285
boundary-value, 217 Heine-Borel covering property, 3
bounded variation, 116, 154 Hermite functions, 205; (I)168, 173
Cantor-Lebesgue, 126, 331 Hermitian operator, 190
characteristic, 27 Hilbert space, 161; (I)75
complex-valued, 67 L2 , 156
convex, 153 finite dimensional, 168
Dirac delta, 110 infinite dimensional, 168
finite-valued, 27 orthonormal basis, 164
increasing, 117 Hilbert transform, 220, 255
integrable, 59, 275 Hilbert-Schmidt operator, 187
jump, 132 homogeneous set, 385
Lebesgue integrable, 59, 64, 68
Lipschitz (Hölder), 330; (I)43 identity operator, 180
measurable, 28 inequality
negative variation, 118 Bessel, 166; (I)80
normalized, 282 Brunn-Minkowski, 34, 48
nowhere differentiable, 154, 383 Cauchy-Schwarz, 157, 162; (I)72
positive variation, 118 iso-diametric, 328, 386
sawtooth, 200; (I)60, 83 isoperimetric, 143; (I)103
simple, 27, 50, 274 triangle, 157, 162
slice, 75 inner product, 157; (I)71
smooth, 222 integrable function, 59, 275
square integrable, 156 integral operator, 187
step, 27 kernel, 187
strictly increasing, 117 interior
support, 53 of a set, 3
total variation, 117 point, 3
fundamental theorem of the invariance of Lebesgue measure
calculus, 98 dilation, 22, 73
linear transformation, 96
Gaussian, 88; (I)135, 181 rotation, 96, 151
good kernel, 88, 108; (I)48 translation, 22, 73, 313
gradient, 236 invariant
Gram-Schmidt process, 167 function, 302
Green’s set, 302
formula, 313 vectors, 295
kernel, 204; (II)217 iso-diametric inequality, 328, 386
isolated point, 3
Hardy space, 174, 203, 213 isometry, 198
harmonic function, 234; (I)20; (II)27 isoperimetric inequality, 143; (I)103,
Hausdorff 122
dimension, 329
distance, 345 jump
exterior measure, 325 discontinuity, 131; (I)63
maximal principle, 48 function, 132
400 INDEX
Kakeya set, 362 linear ordering, 26, 48

kernel linearly independent
Dirichlet, 179; (I)37 elements, 167
Fejér, 112; (I)53 family, 167
heat, 111; (I)209 Lipschitz condition, 90, 147, 151,
Poisson, 111, 171, 217; (I)37, 55, 330, 362
149, 210; (II)67, 78, 109, 113, Littlewood’s principles, 33
216 locally integrable function, 105
Lusin’s theorem, 34
Laplacian, 230
Lebesgue maximal
decomposition, 150 function, 100, 261
density, 106 theorem, 101, 297
exterior measure, 10 maximum principle, 235; (II)92
integrable function, 59, 64, 68 mean-value property, 214, 234, 313;
integral, 50, 54, 58, 64 (I)152; (II)102
measurable set, 16 measurable
set, 106 Carathéodory, 264
Lebesgue differentiation theorem, function, 28, 273
104, 121 rectangle, 276
Lebesgue measure, 16 set, 16, 264
dilation-invariance, 22, 73 measure, 263
rotation-invariance, 96, 151 absolutely continuous, 288
translation-invariance, 22, 73, 313 counting, 263
Lebesgue-Radon-Nikodym theorem, exterior, 264
290 Hausdorff, 327
Lebesgue-Stieltjes integral, 281 Lebesgue, 16
Legendre polynomials, 205; (I)95 mutually singular, 288
limit outer, 264
non-tangential, 196 signed, 285
point, 3 support, 288
radial, 173 measure space, 263
linear functional, 181 complete, 266
null-space, 182 measure-preserving
linear operator (transformation), isomorphism, 292
180 transformation, 292
adjoint, 183 Mellin transform, 253; (II)177
bounded, 180 metric, 267
compact, 188 exterior measure, 267
continuous, 181 space, 266
diagonalized, 185 Minkowski
finite rank, 188 content, 138, 151
Hilbert-Schmidt, 187 dimension, 333
identity, 180 mixing, 305
invertible, 311 monotone convergence theorem, 62
norm, 180 multiplication formula, 88
positive, 307 multiplier, 220
spectrum, 311 multiplier sequence, 186, 200
symmetric, 190 mutually singular measures, 288
INDEX 401
negative variation perpendicular elements, 164

function, 118 Plancherel’s theorem, 208; (I)182
measure, 287 plane, 360
non-measurable set, 24, 44, 82 point in Rd , 2
non-tangential limit, 196 point of density, 106
norm Poisson
L1 (Rd ), 69 integral representation, 217;
L2 (Rd ), 157 (I)57; (II)45, 67, 109
Euclidean, 2 kernel, 111, 171, 217; (I)37, 55,
Hardy space, 174, 213 149, 210; (II)67, 78, 109, 113,
linear operator, 180 216
normal polar coordinates, 279; (I)179
number, 318 polarization, 168, 184
operator, 202 positive variation
normalized function, 118
increasing function, 282 measure, 287
nowhere differentiable function, 154, pre-Hilbert space, 169, 225; (I)75
383; (I)113, 126 premeasure, 270
product
open measure, 276
ball, 2, 267 sets, 83
set, 2, 267 Pythagorean theorem, 164; (I)72
ordered set
linear, 26, 48 quartic intervals, 351
partial, 48 chain, 351
orthogonal quasi-simple curve, 332
complement, 177
elements, 164 radial limit, 173
projection, 178 Radon transform, 363; (I)200, 203
orthonormal maximal, 363
basis, 164 rectangle, 3
set, 164 measurable, 276
outer volume, 3
Jordan content, 41 rectifiable curve, 115, 134, 332
measure, 10, 264 refinement (of a partition), 116;
outside-triangle condition, 248 (I)281, 290
regularity of sets, 360
Paley-Wiener theorem, 214, 259; regularization, 209
(II)122 Riemann integrable, 40, 47, 57;
parallelogram law, 176 (I)31, 281, 290
Parseval’s identity, 167, 172; (I)79 Riemann-Lebesgue lemma, 94
partial differential operator Riesz representation theorem, 182,
constant coefficient, 221 290
elliptic, 258 Riesz-Fischer theorem, 70
partitions of a set, 286 rising sun lemma, 121
Peano rotations of the circle, 303
curve, 350
mapping, 350 sawtooth function, 200; (I)60, 83
perfect set, 3 self-adjoint operator, 190
402 INDEX
self-similar, 342 closed, 175

separable Hilbert space, 160, 162 linear, 174
set support
bounded eccentricity, 108 function, 53
cylinder, 316 measure, 288
difference, 44 symmetric
self-similar, 342 difference, 21
shrink regularly, 108 linear operator, 184, 190
slice, 75
uniformly locally connected, 387 Tchebychev inequality, 91
shift, 317 Tietze extension principle, 246
Sierpinski triangle, 334 Tonelli’s theorem, 80
signed measure, 285 total variation
similarities function, 117
separated, 346 measure, 286
similarity, 342 translation, 73; (I)177
ratio, 342 continuity under, 74; (I)133
simple triangle inequality, 157, 162, 267
curve, 332
function, 27, 50, 274
uniquely ergodic, 304
slice, 361
unit disc, 173; (II)6
function, 75
unitary
set, 75
equivalence, 168
smooth function, 222
isomorphism, 168
Sobolev embedding, 257
mapping, 168; (I)143, 233
space L1 of integrable functions, 68
space-filling curve, 349, 383
span, 167 Vitali covering, 102, 128, 152
special triangle, 248 volume of unit ball, 92, 313; (I)208
spectral von Koch curve, 338, 340, 382
family, 306
resolution, 306 weak
theorem, 190, 307; (I)233 convergence, 197, 198
spectrum, 191, 311 solution, 223
square integrable functions, 156 weak-type inequality, 101, 146, 161
Steiner symmetrization, 386 weakly harmonic function, 234
step function, 27 well ordering
strong convergence, 198 principle, 26, 48
Sturm-Liouville, 185, 204 well-ordered set, 26
subspace Wronskian, 204

Elias M. Stein, Rami Shakarchi - Real Analysis - Measure Theory, Integration, and Hilbert Spaces. Vol.3.-Princeton University Press (2005)

Uploaded by

Copyright:

Available Formats

Elias M. Stein, Rami Shakarchi - Real Analysis - Measure Theory, Integration, and Hilbert Spaces. Vol.3.-Princeton University Press (2005)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Elias M. Stein, Rami Shakarchi - Real Analysis - Measure Theory, Integration, and Hilbert Spaces. Vol.3.-Princeton University Press (2005)

Uploaded by

Copyright:

Available Formats

REAL ANALYSIS

Ibookroot October 20, 2007

Princeton Lectures in Analysis

I Fourier Analysis: An Introduction

III Real Analysis:

IV Functional Analysis: Introduction

PRINCETON UNIVERSITY PRESS

All Rights Reserved

Library of Congress Control Number 2004114065

British Library Cataloging-in-Publication Data is available

Printed on acid-free paper. ∞

Printed in the United States of America

Beginning in the spring of 2000, a series of four one-semester courses

I. Fourier series and integrals.

II. Complex analysis.

However, this listing does not by itself give a complete picture of

problems in Book I, and reappear in Book III to play an important role in

It is with great pleasure that we express our appreciation to all who

Princeton, New Jersey

In this third volume we establish the basic facts concerning measure

Chapter 1. Measure Theory 1

Chapter 2. Integration Theory 49

Chapter 3. Differentiation and Integration 98

3 Differentiability of functions 114

Chapter 4. Hilbert Spaces: An Introduction 156

Chapter 5. Hilbert Spaces: Several Examples 207

Chapter 6. Abstract Measure and Integration Theory 262

Chapter 7. Hausdorff Measure and Fractals 323

Notes and References 389

Symbol Glossary 395

I turn away in fright and horror from this lamentable

Starting in about 1870 a revolutionary change in the conceptual frame-

1 Fourier series: completion

and saw then that one had Parseval’s identity,

However, the above relationship between functions and their Fourier

2 Limits of continuous functions

1 We use the notation of Chapter 3 in Book I.

(a) 0 ≤ fn (x) ≤ 1 for all x.

It is with Lebesgue integration that we can solve both this problem

The problems we are led to arise when we consider general curves.

4 Differentiation and integration

5 The problem of measure

(a) m(E) = b − a if E is the interval [a, b], a ≤ b, of length b − a.

This chapter is devoted to the construction of Lebesgue measure in Rd

Next, we prove two simple theorems that highlight the importance of

Addition of points is componentwise, and so is multiplication by a real

The distance between two points x and y is then simply |x − y|.

If E and F are two subsets of Rd , we denote the complement of F in E

The distance between two sets E and F is defined by

d(E, F ) = inf |x − y|,

where the infimum is taken over all x ∈ E and y ∈ F .

Open, closed, and compact sets

Br (x) = {y ∈ Rd : |y − x| < r}.

A subset E ⊂ Rd is open if for every x ∈ E there exists r > 0 with

In words, any covering of a compact set by a collection of open sets

Rectangles and cubes

R = [a1 , b1 ] × [a2 , b2 ] × · · · × [ad , bd ],

where aj ≤ bj are real numbers, j = 1, 2, . . . , d. In other words, we have

R = {(x1 , . . . , xd ) ∈ Rd : aj ≤ xj ≤ bj for all j = 1, 2, . . . , d}.