FUN C'IIONAL ANALYSIS

Walter Rudin

Professor of Mathematics

University of Wisconsin

McGraw-Hill Book Company New York St. Louis San Francisco DUsseldorf Johannesburg Kuala Lumpur London Mexico Montreal New Delhi Panama Rio de Janeiro Singapore Sydney Toronto

-

J

I ,jbrary of Congress Cataloging jn Publication Data

I. Functional analysis. I. Title.

FUNCTIONAL ANAI.vSIS

publication may be reproduced, stored in a retrieval system,

Printed in the United States of America. No par t of this

Copyright © 1973 by McGraw Hill, Inc. All rights reserved.

or transmitted, in any form or by any means, electronic,

the prior '."ritten permi~~ion of the publisher.

mechanical, photocopying, recording, or otherwise, without

This book was set in Times New Roman. The editors were Jack L. Farnsworth, Bradford Bayne, and M. E. Margolies; and the production supervisor was Ted Agrillo.

f I

CONTENTS

Preface xi

-' -'

Part Ol)e-GENERAL THEORY

Chapter 1

Topological Vector Spaces 3

Introduction 3 Separation pI OpCI tics 9

Lineal mappiligs 13

Firrite=dirrrerrsiotral spaces 14

Metlization 17

Boundedness and continuity 2]

Seminorms and local convexity 24

Quotient spaces 29

Examples 31 Exercises 36

\ Chapter 2 Completeness 41

Baire category 41

The Banach-Steinhaus theorem 43

The open nmpping theorem 46

The closed graph theorem 49

Exercises ~-

Biline~r ma~:in~s 5 I

Chapter 3 Convexity 55

The Hahn-Banach theorems 55 Weak topologies 60

Compact convex sets 66

vi CONTENTS

Vector-valued integration 73

Chapter 4 Duality in Ranach Spaces 87

The normed dual of a normed space 87 Adjoints 92

Compact operators 97 Exercises 105

Chapter 5 Some Applications 110

A continuity theorenl 110

Closed subspaces of LP-spaces I I I

The range of a vector-valued measure 113

A generalized Stone-Weierstrass theorem 115

Two interpolation theorems ] 17

A fixed point theorem 120

Haar measure on compact groups 122 Uncomplemented subspaces 125 Exercises 130

Part TWO-DISTRIBUTIONS AND FOURIER TRANSFORMS

Chapter 6

Test Functions and Distributions 135

Introduction 135

Test function spaces 136

Calculus with distributions 142

Localization 147

Supports of distributions 149 Distributions as derivatives 152 Convolutions 155

Exercises 162

CONTENTS vii

Chapter 7 Fourier Transforms 166

/

Basic properties 166

Tempered distributions 173

Paley-Wiener theorems 180 Sobolev's lemma 185 Exercises 187

Chapter 8 Applications to Differential Equations 192

Fundamental solutions 192

Elliptic equations 197

Exerci ses 204

Chapter 9

Taube} ian Them y 208

Wiener's theorem 208

The prime number theorem 212 The renewal equation 218 Exercises 221

Part Three BANACH ALGEBRAS AND

SPECTRAL THEORY

Chapter 10 Banach Algebras 227

Introduction ,,227

Complex homon:orphisIIIS 231 Basic properties of spectra 234 Symbolic calculus 240 Differentiation 248

The group of invertible elements 257 Exercises 259

viii CONTENTS

Chapter 11 Commutative Banach Algebras 263

Ideals and homomorphisms 263

(iemma transforms 268

Involutions 275

Applications to noncommutative algebras 279 Positive functionals 283

Exercises 288

Chapter 12

Bounded Operators on a Hilbert Space 292

Basic facts 292

Bounded operators 295

A commutativity theorem 300

Resolutions of the identity 301

The spectral theorem 305

Eigenvalues of normal operators 311 Positive operators and square roots 313 The group of invertible operators 316 A characterization of B*-algebras 319 Exercises 323

Chapter 13 Ilnhollnded Operators 329

Introduction 329

Graphs and symmetric operators 333

The Cayley transform 338

Resolutions of the identity 341 The spectral theorem 348 Semigroups of operators 355 Exercises 363

Appendix A

Compactness and Continuity 367

Appendix B

Notes and Comments 372

Bibliography 384

List of Special Symbols 386 Index 389'

gEFACE

Functional analysis is the study of certain topological-algebraic structures and of the methods by which knowledge of these structures can be applied to analytic problems.

A good introductory text on this subject should include a presentation of its axiomatics (i.e., of the general theory of topological vector spaces), it should treat at least a few topics in some depth, and it should contain some interesting applications to other branches of

mathematics.· I hope that the present book meets these cntena,

The subject is huge and is growing rapidly. (The bibliography

in volume I of [4] contains 96 pages and goes only to 1957.) In order

to write a book of moderate SIze, It was therefore necessary to select

certain areas and to ignore others. I fully realize that almost any

expert who looks at the table of contents will find that some of his

(and my) favorite topics are mISSIng, but thIS seems unaVOIdable. It

was not my intention to write an encyclopedic treatise. I wanted to write a book that would open the way to further exploration.

This is the reason for omitting many of the more esoteric topics that might have been included in the presentation of the general theory of topological vector spaces. For instance, there is no discussion of uniform spaces, of Moore-Smith convergence, of nets, or of filters. The notion of completeness occurs only in the context of metric spaces. Bornological spaces are not mentioned, nor are

barreled ones. DualIty IS of course presented, but not In ItS utmost

generality. Integration of vector=vahred functions is treated strictly as

a tool; attention is confined to continuous integrands, with values in

a Frechet space.

Neverthel~s, the material of Part 1 is fully adequate fOI almost

all applications to concrete problems. And this is what ought to be

stressed In such a- course: I he close mter la between the abstract

and the concrete is not only the most useful aspect of the whole subject but also the most fascinating one.

Here are some further features of the selected material. A fairly large part of the general theory is presented without the assumption of local convexity. The basic properties of compact operators are derived from the duality theory in Banach spaces. The Krein-Milman theorem on the existence of extreme points is used in several ways in

•

xii PREFACE

algebras); this is perhaps not the shortest way, but it is an easy one. The symbolic calculus in Banach algebras is discussed in considerable detail; so are involutions and positive functionals. Several fairly recent results on Banach algebras that have not found their way into other textbooks as yet are included.

1 assume familiarity with the theory of measure and Lebesgue integration (including such facts as the completeness of the LPspaces), with SOllie basic properties of holomorpfilc functIOns (such

with the elementary topological background that goes with these two

as the general form of Cauchy's theorem, and Runge's theorem), and

analytic topics. Some olIrel topological facts are brIefly presented In

Appendix A Almost no algebraic background is needed, beyond

the knowledge of what a homomorphism is

H istorieal rerel elices aI e gathered in AppendIX B. Some of

these refer to the original sources, and some to more recent books, papers, or expository articles in which further references can be found. There are, of course, many items that are not documented at all. In no case does the absence of a specific reference imply any claim to originality on my part.

Most of the applications are in Chapters 5, 8, and 9. Some are in Chapter 11 and in the more than 250 exercises; many of these are supplied with hints. The interdependence of the chapters is indicated in the following diagI am.

I 7

/\

8 9

4

I --------10

5 I

11 I 12 I 13

It is a pleasure to Iecord my thanks to them.

Patrick Ahern, Paul Rabinowitz, Daniel Shea, and Robelt TUllier.

various topics in it with some of my colleagues, especIally WIth

sity of \VisconsilI. I have had many fruitful conversatiom about

This book grew out of a conrse that I have taught at the Univer-

I

Walter Rudin

-' - .

PART ONE

General Theory

TOPOLOGICAL VECTOR SPACES

..' - '

Introduction

1.1 Many problems that analysts study are not primarily concerned with a single

object sl]ch as a function, a measure, or an operator, bllt they deal instead with large

classes of such objects. Most 01 the mteresting classes that occur in this way turn out

to be vector spaces, either with real scalars or with complex ones. Since limit processes

playa role in every analytic problem (explicitly or implicitly), it should be no surprise

that these vector spaces are supplIed WIth metncs, or at least with topologies, that

beal some natm al relation to the objects of which the spaces are made up. The

simplest and most important way of doing this is to introduce a norm. The resulting structure (defined below) is called a normed vector space, or a normed linear space, or simply a normed space.

Throughout this book, the term vector space will refer to a vector space over the complex field f; or over the real field R. For the sake of completeness, detailed definitions are given in Section 104.

4 GENERAL THEORY

1.2 Normed spaces A vector space X is said to be a normed space if to every

(b) Ilaxll Ictlllxli ifxE X and tJ; is a scalar,

(c) Ilxll > 0 if x ::fi O.

The word" norm" is also used to denote the function that maps x to [x]. Every normed space may be regarded as a metric space, in which the distance d(x, y) between x and y is Iix - YII. The relevant properties of dare:

(i) 0 < d(x, y) < 00 for all x and y,

(ii) d(x, y) = 0 if and only if x = y,

(iii) d(x, y) dey, x) for all x and y,

(iv) d(x, z) < d(x, y) + dey, z) far an x, y, z

In any metric space, the open ball ,,"ith center at x and radius r is the set

BrCx) {y: d(x, y) < r}.

In partIcular, If X IS a normed space, the sets j

BICo) {x: Ilxll < I}

and

BI(O) (x: Ilxll < I}

are the open unit ball and the closed unit ball of X, respectively.

By declaring a subset of a metric space to be open if and only if it is a (possibly empty) union of open balls, a topology is obtained. (See Section 1.5.) It is quite easy to verify that the vector space operations (addition and scalar multiplication) are continuous in this topology, if the metric is derived from a norm, as above.

A Banach space is a narmed space which is complete in the metric defined by its norm; this means that every Cauchy sequence is required to converge.

1.3 Many of the best known function spaces are Banach spaces. Let us mention

Just a few types: spaces of continuous functions on compact spaces; the familiar

H'-spaces that occur in integration theory; Hilbert spaces the closest relatives of

euclidean spaces; certain spaces of differentiable functions; spaces of continuous linear

mappmgs from one Banach space into another; Banach algebras All of these will

occur later on in the text.

But there are also many important spaces that do not fit into this framework.

Here are some examples:

(a) C(n), the space of all continuous complex functions on some open set n in a euclidean space R".

(b) H(n), the space of all holomorphic functions in some open set n in the complex plane.

TOPOLOGICAL VECTOR: SPACES 5

(c) CR, the space of all infinitely diffelentiable complex functions on RR that vamsh

outside some fixed compact set f( with nonempty interior.

These spaces carry natural topologIes that cannot be induced by norms, as we shall see later. They, as well as the normed spaces, are examples of topological vector spaces, a concept 'that pervades all of functional analysis.

After this brief attempt at motivation, here are the detailed definitions, followed (in Section 1.9) by a preview of some of the results of Chapter 1.

1.4 Vector spaces The letters Rand fl will always denote the field of rea! numbers and the field of complex numbers, respectively, For the moment, let <D stand for either

whose elements are called vectors, and III WhICh tw<? operatIOns, addition and scalar

R or e:. A scalar is a member of the 'icalarfield <D. A vector space over <I> is a set X,

multiplica tion, are defined, 'Nith the following familial algebI aic properties:

(a) To every pair of vectors x and y corresponds a vector x + y, in such a way that

x+y y+x

and

x I (y I z) (x+y)+z,

X contains a unique vector 0 (the zero vector or ortqtn of X) such that x + 0 x for every x EX; and to each x E X corresponds a unique vector - x such that x+(-:-x)=O.

(b) To every pair (ex, x) with ex E <D and x E X corresponds a vector «x, in such a way that

Ix:::;:: x,

ex(fJx) :::;:: (exfJ)x,

and such that the two distributive laws

a(x + y) ax I ay,

(0: + f3)x exx + {]x

o .

The symbol 0 ] • .."ill of course also be used fOl the zelO element of the scalar field.

A real vector space is one for which <D - R; a complex t'ector space is one for

which <I> fl. Any statement about vector spaces III WhICh the scalar field is not

explicitly mentioned is to be u9~erstood to apply to both of these cases.

If X IS a vector space, A c X, B c X, X E X, and I. E $, the follO'.ving notations

will be used:

x + A = {x + a: a E A},

x - A = {x - a: a E A}, A+B={a+b:aEA, bEB},

AA = {Aa: a E A}.

In particular (taking A:

1), A denotes the set of all additive inverses of member s

A word of warning: With these conventions, it may happen that 2A "# A + A

A set Y c X is called a subspace of X if Y is itself a vector space (with respect to

the same operations, of course). One checks easily that this happens if and only if o E Yand

aY+{3Yc Y

for all scalars a and {3.

A set C c X is said to be convex if

IC + (1 - t)C c C

(0 ~ I ~ 1).

In other words, it is required that C should contain tx + (1 1))1 if x E C, Y E C, and

A set D c X is said to be balanced if aD c D for every a E <D with I IX I < 1.

A vector space ,y has dimension n (dim ,r - n) if ,r has a bmiv {uu

,u}

II

I hIS means that every x E X has a umque representatIOn of the form

(ex, E <1».

If dim X = n for some n, X is said to have finite dimension. If X = {O}, then dim X = o.

Example If X = fl (a one-dimensional vector space over the scalar field fl), the balanced sets are: fl, the empty set 0, and every circular disc (open or closed) centered at O. If X = R2 (a two-dimensional vector space over the scalar field R), there are many more balanced sets; any line segment with midpoint at (0,0) will do. The point is that in spite ofthe well-known and obvious identifica-

tion of (j; v/ith R2, these two are entirely different as far as their vector space

stmcture is concerned

1.5 Topological spaces A topological space is a set S in which a collection r of

subsets (caned open sfiV) has been specified, with the [oHowing properties: S is open,

o IS open, the intersection of any two open sets is open, and the union of every coJIec-

tion of open sets is open. Such a collection 1: is called a topology on S. When clarity seems to demand it, the topological space corresponding to the topology 1 will be written (S, 1) rather than S.

Here is some of the standard vocabulary that will be used, if Sand r are as above. A set E c S is closed if and only if its complement is open. The closure E of E is the intersection of all closed sets that contain E. The interior EO of E is the union

TOPOLOGICAL vECIOR SPACES 7

of all Ope" sets that are sub",ts of E. A neighborhood of a pointy E S IS an~ 0rn set

that contains p. (S, r) IS a Hausdorffspace, and 1" is a HausdorfFtopology, ~istinct

points of S have disjoint neighborhoods. A set K c S IS compact if every open cover

of K has a finite sllbcover. A collection f' c r is a base for r if every member of r

(that is, every open set) IS a union of members of r '. A collection y 'of neighborhoods of a point PES is a local base at p if every neighborhood of p contains a member of y.

If E c S ancf If o is the collection of all intersections E n V, with V E r , then a is a topology on E, as is easily verified; we call this the topology that E inherits from S.

If a topology r is induced by a metric d (see Section 1.2) we say that d and r are compatible with each other.

A sequence {xn} in a Hausdorff space X converges to a point x E X (or: limn-+co

XI1 x) if every neighborhood of x contains all but finitely many of the points Xn•

1.6 Topological vector spaces Suppose 1" is a topology on a vectOI space X such

(a) every point oj X is a clO'ied set, and

(b) the vecto, space operatIOns are continuous with respect to 1".

Under these conditions, 'C is said to be a vector topology on X, and X is a

topological vector space.

Here is a more precise way of stating (a): For every x E X, the set {x} which has x as its only member is a closed set.

In many texts, (a) is omitted from the definition of a topological vector space.

Since (a) is satisfied in almost every application, and since most theorems of interest require (a) in their hypotheses, it seems best to include it in the axioms. [Theorem 1.12 will show that (a) and (b) together imply that T is a Hausdorff topology.]

To say that addition is continuous means, by definition, that the mapping

IS a neIghborhood of Xl + x2, there should exist netgnbortrcorts V. of x. such that

mapping

SImIlarly, the aSSllmption that scalar multiplication 1S continuous means that the

((X, x) -> (Xx

of cI> x X into X is continuous: If x E X, ex is a scalar, and V is a neighborhood of (Xx, then for some r > 0 and some neighborhood W of x we have f3W c V whenever I f3 - (X I < r.

8 GENERAL THEORY

A subset E o[a topological vector space is said to be bounded it to every nelghbor-

hood V of 0 in X corresponds a number s > 0 such that E c t V for every t > s:

1.7 In·.,rariance Let X be a topological veclor space. Associate to each a E X and

to each scalar J ::f 0 the translation operator Ta and the multiplication operator l'.f;.,

by the formulas

TaCx) = a + x,

Mix) = AX

(x EX).

The following simple proposition is very important:

Proposition Ta and M). are homeomorphisms of X onto X.

PROOF The vector space axioms alone imply that Ta and M;. are one-to-one, that they map X onto X, and that their inverses are T a and MIl;" respectively.

The assumed continuity of the vector space operations implies that these four

mappmgs are continuous, Hence each of them is a homeomorphism (a con-

linaoas mapping whose inverse is also continuous). III!

One consequence of thIS proposItIOn IS that every vector topology I is translation-

invariant (or simply invaJ iant; [01 brevity): A set E c X is open if and only if each of

its translates a + E is open Thus l' is completely determined by any local base.

In the vector space context, the term local base will always mean a local base at O. A local base of a topological vector space X is thus a collection !!I of neighborhoods of 0 such that every neighborhood of 0 contains a member of !!I. The open sets of X are then precisely those that are unions of translates of members of f!l.

A metric d on a vector space X will be called invariant if

d(x + z, y + z) = d(x, y)

for all x, y, z in X.

1.8 . Types of topological vector spaces In the following definitions, X always

denotes a topologIcal vector space, with topology T

(a) X is locally convex if there is a loeal base fJj whose membels ale convex.

(b) X is locally bounded if 0 has a hounded neighborhood.

(c) X is locally compact if 0 has a neighborhood whose closure IS compact.

(d) X is metrizable if 1: is compatible with some metl ic d.

(e) X is an F-space if its topology T is induced by a complete invariant metric d. (Compare Section 1.25.)

(f) X is a Frechet space if X is a locally convex F-space.

(g) X is normable if a norm exists on X such that the metric induced by the norm is compatible with T.

TOPOLOGICAl VECTOR SPACES 9

(h) Normed spaces and Banach spaces have already been defined (Section 1.2).

0) X has the Reme-Borel property If every closed and bounded subset of 1'" is

The termmology of (e) and (f) IS not universally agreed upon: In some texts,

local convexity is omitted from the definition of a Frechet space, whereas others use F-space to describe what we have called Frechet space.

1.9 Here is a list of some relations between these properties of a topological vector space X.

(a) If X is locally bounded, then X has a countable local base [part (c) of Theorem 1.15].

(b) X is metnzable If and only If X has a countable local base (Theorem 1.24).

(c) X is nefmable if and orily if X is locally conve~ and locally bounded 0 fieorem

(d) X has tillIte dimenSIOn If and only If X IS locally compact (Theorems 1 21, 1.22).

(e) If a locally bounded space X has the Heine- Borel property, then X has fimte

dimension (Theorem 1 23).

The spaces H(ft) and C': mentioned in Section 1.3 are mfimte-dImensIOnaI

Frechet spaces with the Heine-Borel property (Sections 1.45, 1.46). They are therefore not locally bounded, hence not normable; they also show that the converse of (a) is false.

On the other hand, there exist locally bounded F-spaces that are not locally convex (Section 1.47).

Separation Properties

1.10 Theorem Suppose K alld C Q/ e subsets of a topological vector space X, K ts

compact, C jll cla<;ed, and K n C 0. Then 0 has a neighborhood V such that

(K + V) n (C + V) 0.

open set that contains K. The theOIem thus implies the existence of disjoint open sets

Note that K + v is a uruon of translates x + Vof V ex E K) Thus K + V is an

that contain K and C, respectively.

PROOF We begin with the following proposition, which will be useful in other contexts as well:

If W is a neighborhood 0[0 in X, then there is a neighborhood U 0[0 which is symmetric (in the sense that U = - U) and which satisfies U + U c W.

10 GENERAL THEORY

To see this, note that 0 I 0 0, that addition is continuous, and that 0

therefore has neighborhoods VI' V2 SJlch that Vi + V2 C W If

then U has the required properties.

The proposition can now be applied to U in place of Wand yields a new

symmetric neighborhood V of 0 such that V+V+V+VcW.

It is clear how this can be continued.

If K = 0, then K + V = 0, and the conclusion of the theorem is obvious.

We therefore assume that K i= 0, and consider a point x E K. Since C is closed, since x is not in C, and since the topology of X is invariant under translations, the preceding proposition shows that 0 has a symmetric neighborhood Vx such

that x + Vx + Vx + Vx does not intersect C; the symmetry of Vx then shows that

(1)

(x + J'X + J'X) I \ (C + f~) 0.

Since K is compact, there are finitely many points Xl' ... , x, in /( such that

K C (Xl + VXI) U ..• U (X" + F:J.

Put V - V (\'" r, V . Then

n n

K + V c U (Xi + VXi + V) c U (Xi + VX/ + Vx),

i= 1 i= 1

and no term in this last union intersects C + V, by (1). This completes the proof.

IIII

Since C + V is open, it is even true that the closure of K + V does not intersect C + V; in particular, the closure of K + V does not intersect C. The following special case of this, obtained by taking K = {O}, is of considerable interest.

1.11 Theorem If~ is a local base for a topological vector space X then every member

off!)) contains the closure ofsome member of (JjJ

So far we have oat llsed the assumption that every point of :r is a closed set

We now use it and apply Theorem 1.10 to a paIr of dIstInct POInts In place of K and C.

The eonclusion is that these points have disjoint neighborhoods. In other WOlds, the

Hausdorff separati60 axiom holds:

1.12 Theorem Every topological vector space is a Hausdorffspace.

We now derive some simple properties of closures and interiors in a topological vector space. See Section 1.5 for the notations E and EO, Observe that a point p belongs to E if and only if every neighborhood of p intersects E.

TOPor OGle" L VEcmR SPACES 11

1.13 Tbeorem Let X be a topological vector space.

(b) lfA c XandB eX, then A + 13 c it + B.

(a) /fA c X then A - A (A + V), where V runs through all neighborhoods o.lO.

( c) 1] Y is a subspace of Y, so is y.

(d) If C is a convex subset oj X, so are C and CO.

(e) If B is a balanced subset of X, so is 13; if also ° E BO then BO is balanced. (f) If E is a bounded subset of X, so is E.

PROOF (a) x E A if and only if (x + V) n A ¥= 0 for every neighborhood V of 0, and this happens if and only if x E A - V for every such V. Since _ V is a neighborhood of 0 if and only if V is one, the proof is complete.

(b) Take a E A, b e 13; let W be a neighborhood of a + b. There are

X E A n WI and Y E B A W2, since a E A and b E B. Then x + v lies in

neighborhoods WI and tV2 of a and b such that WI + W2 C W. There exist

(A + B) n If; so that thi, intersection is not empty. consequentlr; + iF E A ~ ~.

from (b) that

aY IX Y if a i= 0; if a 0, these two sets are obviously equal. Hence it follows

aY+{3Y aY+{3YcaY+/3YcY;

the assumption that Y is a subspace was used in the last inclusion.

The proofs that convex sets have convex closures and that balanced sets have balanced closures are so similar to this proof of (c) that we shall omit them from (d) and (e).

(d) Since Co c C and C is convex, we have

tCO + (1 t)CO c C

every open subset of C is a subset of Co, it follows that Co is convex.

if b < t < I. The two sets on the left are open; hence so is their sum. Smce

contains the origin, then aB~ c BO even for Ct: a

Hence aBO c aB c B, since B is balanced. But aBO is open. So aBO c Be. If BO

(e) If 0 < lal < 1, then aBO (aBt, since x ~ ax IS a homeomorphism.

(f) Let V be a neighbol hood of O. By Theorem 1.11, We V for some

neighborhood W of O. Since E is bounded, E c t W for all sufficiently large t. For these t, we have E c tW c tv. IIII

1.14 Theorem In a topological vector space X,

(a) every neighborhood % contains a balanced neighborhood ofO, and

(b) every convex neighborhood of ° contains a balanced convex neighborhood of O.

12 GENERAL THEORY

PROOF (a) Suppose U is a neighborhood of 0 ill X. Since scalar multiplication

is continuous, there is a b > 0 and there is a neighborhood V of 0 in 1'sHch that

a V c: V whenever IIX I < b. Let W be the uruon of' all these sets a V. I hen W is

a neighborhood of 0, W is balanced, and W c u.

(b) Suppose U is a convex neighborhood of 0 in X. Let A na [~ where

a ranges over the scalars of absolute value 1. (:hoose Was III part (a). SIllce W is balanced, a-1 W = W when lal = 1; hence W c: «U, Thus W c: A, which implies that the interior AO of A is a neighborhood of 0. Clearly AO c: U. Being an intersection of convex sets, A is convex; hence so is AO. To prove that AO is a neighborhood with the desired properties, we have to show that AO is balanced; for this it suffices to prove that A is balanced. Choose rand 13 so that Osrsl, 1131 =1. Then

which completes the proof. llll

Them eIll 1.14 can be restated in terms oflocal bases. Let us say that a local base

@J is balanced if its members are balanced sets, and let liS call @ convex if its members

are convex sets.

Corollary

(a) Every topological vector space has a balanced local base.

(b) Every locally convex space has a balanced convex local base.

Recall also that Theorem 1.11 holds for each of these local bases.

1.15 Theorem Suppose V is a neighborhood ufO in a topological vector space X.

(a) If 0 < '1 < '2 < ... and r" ~ 00 as n ~ 00, then

0) Every compact subset K of X is bounded

(c) ,Ifbl > (52 > ... and ()II ~ 0 as n ~ 00, and 1/ V IS bounded, then the collection

{bn V: n = 1,2,3, ... }

is a local base for X.

PROOF (a) Fix x E X. Since a ~ ax is a continuous mapping of the scalar field into X, the set of all a with ax E V is open, contains 0, hence contains l/rn for all large n. Thus (l/rn)x E V, or x E rn V, for large n.

TOPOLOGICAL VECTOR SPACES 13

(b) Let JV be a balanced neighborhood of 0 slIch that We V. By (a),

Since K is compact, there are Integers nl < ... < ns such that.

nW.

-' - ~

The equality holds because W is balanced. If t > ns, it follows that K e tWe tV ..

(c) Let V be a neighborhood of 0 in X. If V is bounded, there exists s> 0 such that Ve iu for all t > s. If n is so large that sOn < 1, it follows that Ve (I/on)V, Hence V actually contains all but finitely many of the sets on V.

IIII

I.inear Mappings

and the inverse image or pre image f lEE) of B are defined by

f(A) = {f(x): x E A},

f-l(B) = {x:f(x) E B}.

Suppose now that X and Yare vector spaces over the same scalar field. A mapping A: X -+ Yis said to be linear if

A(ax + py) = aAx + /3Ay

for all x and y in X and all scalars a and /3. Note that one often writes Ax, rather than A(x), when A is linear.

translation operators Ta are not, except when a 0

For example, the multiplication operators Afrz of Section 1.7 are linear, but the

linear mappings of X into its scalar field are called linear Iunciianals

Here are some properties of linear mappings A: X -+ Y whose proofs are so

(0 If A IS a subspace (or a conveX: set, or a balanced set) the same is true of A(A).

(c) If B is a subspace (or a convex set, or a balanced set) the same is true of A -l(B). (d) In particular, the set

A -l({O}) = {x E X: Ax = O} = SeA)

is a subspace of X, called the null space of A.

We now turn to continuity properties of linear mappings.

14 GENERAL THEORY

1.17 Theorem Let X and Y be topological vector spaces. if A' ;Y --4 Y is linear

and continuous at 0, then A is continuous. in fact, A IS umformly continuous, In the

y X E V implies A.y A.x E W.

PROOF Once W is chosen, the continuity of A at 0 shows that AV c W for some neighborhood V of O. If now y - x E V, the linearity of A shows that Ay - Ax = A(y - x) E W. Thus A maps the neighborhood x + V of x into the preassigned neighborhood Ax + W of Ax, which says that A is continuous at x.

IIII

1.18 Theorem Let A be a linear junctional on a topological vector space X.

(b) The null space ;A/'(A) is closed.

(c) .¥(A) is not dense in ;r

(d) A IS bounded In some neighborhood P 0[0.

PROOF Since ..;V(A) = A -1({O}) and {O} is a closed subset of the scalar field <1>, (a) implies (b). By hypothesis, ..;V(A) =1= X. Hence (b) implies (c).

Assume (c) holds; i.e., assume that the complement of ..;V(A) has nonempty interior. By Theorem 1.14,

(1)

(x + V) n ..;V(A) = 0

for some x E X and some balanced neighborhood V of O. Then AV is a balanced subset of the field <1>. Thus either AV is bounded, in which case (d) holds, or

AfT <I>. In the latter case, there exists or E V such that l\Y Ax, and so

x + Y E .¥(A), in contradiction to (1) Thus (c) implies (if)

Fmally, If (d) holds then I Ax I < M for all x in P and for some Iv! < 00.

If 1 > 0 and if W (1/1.1) V, then I Ax I < r for every x in W. Hence A is con

tinuollS at the origin By Theorem 1 17, this imp1ies (a) 1///

Finite-dimensional Spaces

1.19 Among the simplest Banach spaces are Rn and rz;n, the standard n-dimensional vector spaces over Rand €, respectively, normed by means of the usual euclidean metric: If, for example,

TOPOLOGICAL VECTOR SPACES 15

is a vector in (r, then

Other norms can be defined on f]n. FOI example,

I/zl1 Izd + ... + IZnl

or

/lzil max Clzd : 1 < t« n).

These norms correspond, of course, to different metrics on fln (when n > 1) but one can see very easily that they all induce the same topology on fln. Actually, more is true:

If X is a topological vector space over fl, and dim X = n, then every basis of X induces an isomorphism of X onto fln. Theorem 1.21 will prove that this isomorphism must be a homeomorphism. In other words, this says that the topology of (f;n is the only

vector topology that an n-dimensionai complex topological vector space can bUl)e

complex ones.

Everything in the precedmg dIscussion remains true with real scalars in place of

We shall also see that finite-dimensional subspaces are always closed.

We start with a lemma, which will be superseded by Theorems 1.21 and 1.22.

locally compact, in the topology inherited from X. Then Y is a closed subspace of X.

1.20· Lemma Suppose Y is a subspace of a topologlcal vector space X, and Y is

PROOF There is a compact set KeY whose interior (relative to Y) contains O. Hence there is a neighborhood U of 0 in X such that U () Y c K. Choose a symmetric neighborhood Vof 0 in X such that V + V c U. We claim that the set

(1)

Y () (x + V)

is compact, for every x E X.

To see this, fix Yo in (1). For any y ill (1),

y Yo (y x) + ex )'0) € v + v c U.

Also, Y Yo E Y, sinee Y is a subspace. Thus

which implies that (1) lies in the compact set Yo + K. But (1) is also a closed

subset of Y, since x + V is closed in X and since Y inherits its topology from X. Thus (1) is a closed subset of a compact set and is therefore compact.

Now fix x E Y. Let!!l be the collection of all open sets W in X such that o E Wand We V, and associate with each WE!!l the set

Ew = Y () (x + W).

16 GENERAL THEORY

Since We V, each Ew is compact. Since x E Y, no Ew is empty. Since inter

sections of finItely many members of.@ belong to .oJ, it follows that {Ew' WE @}

is a collection of compact sets with the finite intersection property. Therefore

there exists Z E nEw. This z lies in Y. On the other hand, Z E x I W for every

WE f!JJ. 'lhus z - x (Theorem I 12) Hence x E Y This proves that Y - Y,

and so Yis closed. jill

1.21 Theorem Suppose X is a complex topological vector space, Y is a subspace of X, n is a positive integer, and dim Y = n. Then

(a) every isomorphism of en onto Y is a homeomorphism, and (b) Y is closed.

'l he term "homeomorphIsm" refers, of course, to the euclidean topology of

(j;n on the one hand, and to the topology that Y inherits from X on the other. Since

fln is locally compact, Lemma 1.20 shows that (b) follows from (a). The proof that

follows also YIelds the analogous theorem with rea] scalars in place of complex ones

PROOF Let Pn be the theorem as stated. We first plOve Pl' Let A. fl;"4 Y be

an isomorphism (i e , a one-to-one linear mapping of ({; onto Y). Put U l\1.

Then Aa - au. The continurty of the vector space operations III Y Imphes that A is continuous. Note that A -1 is a linear functional on Y with null space {O}, a closed set. By Theorem 1.18, A -1 is continuous. This proves Pl'

Assume next that n > 1 and Pn-1 is true. Let A: en -+ Y be an isomorphism. Let {eb •.• , en} be a basis of en; the kth coordinate of ek is 1; the others are O. Put Uk = Aek, for k = 1, ... , n. Then

and the continuity of the vector space operations in Y implies again that l\ is

continuous. Since A is an isomorphism, {Vi' , VII} is a basis of Y Hence

there are linear functionals 1'1' ... , Yn on Y such that every x E Y has a unique

representation of the form

X ydX)u1 + ... + Yn(x)u",

Each )" has a null space in Y, of dimension Jl 1, which is cl1Jsed in Y, by the

assumed truth of Pn-1• Hence I'i is continuous, by Theorem 1.18. Since

(x E Y),

it follows that A -1 is continuous. Hence P; is true, and the proof is complete.

jill

TOPOLOGICAl VECTOR SPA,CES 17

1.22 TheOI em Every locally compact topological vector space X has finite di-

Theorem 1.15, V IS bounded, and the sets 2 nV(n _ I, 2, 3, ... ) fOlm a local

PROOF The origin of X has a neighborhood V whose c10sure IS compact. By

base for X.

The compactness of V shows that there exist Xl' ... , Xm in X such that

V C (Xl + tV) u ... U (Xm + tV).

Let Y be the vector- space spanned by Xl, ..• , xm• Then dim Y ~ m. By Theorem 1.21, Y is a closed subspace of X.

Since V C Y + t V and since ),Y = Y for every scalar), =f= 0, it follows that

VcY+~VcY+Y+1V y+lv.

If we contmue m this way, we see that

v C n (Y + 2 n V).

n=l

Since {2~nv} is a local base, it now follows from~Y"O'f Theorem 1.13 that V c Y. But Y = Y. Thus V c Y, which implies that kV c Yfor k = 1,2, 3, .... Hence Y = X, by (a) of Theorem 1.15, and consequently dim X::;; m. IIII

1.23 Theorem If X is a locally bounded topological vector space with the HeineBorel property, then X has finite dimension.

finite-dimensional, by Theorem 1.22.

pact, by the Heine-Borel property. 1hIS says that X is locally compact, hence

Metrization

We recall that a topology r on a set X is said to be metrizable if there is a metric don X which is compatible with r. In that case, the balls with radius lin centered at x form a local base at x. This gives a necessary condition for metrizability which, for topological vector spaces, turns out to be also sufficient.

18 GENERAL THEORY

] .24 Theorem if Y is a topological t'ector space with a countable local base, then

there IS a metrlc d on X such that

(a) d is compatible with the topology of X,

(b) the open balls centered at 0 are balanced, and

(c) d is invariant: d(x + z, y + z) = d(x, y)for x, y, z EX.

If, in addition, X is locally convex, then d can be chosen so as to satisfy (a), (b), (c), and also

(d) all open balls are convex.

PROOF By Theorem 1.14, X has a balanced local base {v,.} such that

(1)

(n = 1,2,3, ... );

Let D be the set of all ratIOnal numbers r of the form

r

where each of the" digits " ciCr) is 0 or 1 and only finitely many are 1. 1hus

each rED satisfies the inequalities 0 :::; r < 1.

Put A(r) = X if r ::::: 1; for any rED, define

(3) A(r) = c1(r)V1 + c2(r)V2 + c3(r)V3 + ....

Note that each of these sums is actually finite. Define

(4)

f(x) = inf {r: x E A(r)}

(XE X)

and

(5)

d(x, y) f(x Y)

(xE X,YE X).

The proof that this d h~s the desired properties depends on the inclusions

(6)

A(r) + A(s) c A(r I s)

(r e D, s E IJ).

Before proving (6), let us see how the theolem follows flom it. Since

every d(s) contains 0, (6) implies

(7)

A(r) c A(r) I ACt r) c ACt)

if

r < t.

Thus {A(r)} is totally ordered by set inclusion. We claim that

(8)

f(x + y) :::;f(x) + fey)

(xe X,yeX).

In the proof of (8) we may, of course, assume that the right side is < 1. Fix e > O. There exist rand s in D such that

f(x) < r,

fey) < s,

r + s <f(x) + fey) + e.

TOPOLOGICAL -VECTOR SPACES 19

Thus x E A(r), Y E A(~), and (6) implies x lYE A(i + s). Now (8) follows,

/(x + y) < r + s <f(x) +f(y) + e,

and e was arbItrary.

Since each A(r) is balanced, f(x) f( x). Jt is obvious that /(0) O.

If x =1= 0, then ~ 1. ~ = A(2 n) for some n, and so f(x) 2: 2 n > O.

These properties off show that (5) defines a translation-invariant metric d on X. The open balls centered at 0 are the open sets

(9)

BoCO) = {x:f(x) < c:5} = U A(r).

r</i

If c:5 < 2-n, then BoCO) c v". Hence {B/i(O)} is a local base for the topology of X. This proves (a). Since each A(r) is balanced, so is each BoCO). If each ~ is convex, so is each A(r); and (7) implies that the sallie is tl ue of each B/i(O), hence

also of each translate of BoCO).

The proof of (6) will be by induction. Let PN be the statement:

if r + s < 1 and cn(r) cn(s) 0 for all n > iV, then

A(r) + A(s) c A(r + s).

PI is true, by InspectIOn. Assume P..v 1 is tme, for some N > 1. Choose

s' by

reD, sED, so that I + S < I and enCr) cn(s) 0 If n > N, and define 1" and

(11 ) Then

(12)

A(s) = A(s') + CN(S)VN.

By PN-1, A(r') + A(s') c A(r' + s'). Hence

(13) A(r) + A(s) c A(r' + s') + cN(r)VN + CN(S)VN.

and c,v(~) 1, the right side of (13) is

If eNEr) cN(s) 0, then r r'; s s', and (13) gIves (10). If cN(r) 0

A(r' Is') + VN A(I' + s' + 2 IV) A(r + s),

If cN(r) cN(s) - I, the right side of (13) is

so that (10) holds again, The.case cN(r) 1, c/V(s) 0 is handled the same way

A(t' + s') + v:V + vN c A(r' + s') + VN 1

= A(r' + s') + A(2-N+1) c A(r' + s' + 2-N+1) = ACr + s).

The last inclusion depended on PN-1•

Thus PN-1 implies PN• Hence (6) is correct, and the proof is complete.

IIII

1.25 Cauchy sequences (a) Suppose d is a metric on a set X. A sequence {xn} in

X is a Cauchy sequence if to every e > 0 there corresponds an integer N such that

d(xm' Xn) < 8 whenever m > Nand n > N If every Cauchy sequence in X converges

to a pomt of X, then d IS said to be a complete metric on X.

(b) Let! be the topology of a topological vector space X. The notion of Cauchy

sequence can be defined in this setting without reference to any metric: Fix a local base f!IJ for r. A sequence {xn} in X is then said to be a Cauchy sequence if to every V E f!IJ corresponds an N such that x, - Xm E V if n > Nand m > N.

It is clear that different local bases for the same '[' give rise to the same class of Cauchy sequences.

(c) Suppose now that X is a topological vector space whose topology '[' is compatible with an invariant metric d. Let us temporarily use the terms d-Cauchy

sequence and ,[,-Cauchy sequence for the concepts defined in (0) and (b), respectively.

and since the d: balls centered at the origin form a local base for 't, we conclude:

A sequence {xn} in X is a d-Cauchy sequence ({and only ifit is a t Cauchy sequence.

Consequently, any two invariant metrics on X that are compatible with 't" have the same Cauchy sequences. They clearly also have the same convergent sequences (namely, the r-convergent ones). These remarks prove the following theorem:

1.26 Theorem If d1 and d2 are invariant metrics on a vector space X which induce the same topology on X, then

(a) d1 and d2 have the same Cauchy sequences, and (b) d1 is complete if and only if d2 is complete.

Invariance is needed in the hypothesis (Exercise 12).

The next theorem is an analogue of Lemma 1.20, with completeness in place of

local compactness Note that the two proofs are quite similar.

1.27 Theorem Suppose Y is a subsnace of a tonolngical vector snace X, and Y is

an }'..space (In the topology Inherzted from X). Then .Y IS a closed subspace of X.

, PROOF Choose an invariant metric don Y, compatible with, its topology. Let '-, B1/n = {Y E Y: dey, 0) < ~} ,

let U; be a neighborhood of 0 in X such that Y n U; = B1/n, and choose symmetric neighborhoods Vn of 0 in X such that Vn + v,. C Un'

TOPOLOGICAL 'vECTOR SPACES 21

SUppose x E Y, and define

(n 1,2,3, ... ).

in B11n· lhe dIameters of the sets En therefole tend to O. Smce each En is

If Yl E En and )'2 E En, then Yl Y2 lies in Y and also in 1';, + J!;, C Un, hence

nonempty and since Y is complete, it follows that the Y-cIosures of the sets En have exactly _<?ne point Yo in common.

Let -W be a neighborhood of 0 in X, and define

F; = Y n (x + W n v,,).

The preceding argument shows that the Y-cIosures of the sets F; have one common point Yw· But Fn C En· Hence Yw = Yo . Since F; C x + W, it follows that Yo lies in the X-closure ofx + W, for e'velY W. This implies Yo = x.

Thus x E Y. This proves that Y = Y.

The following simple facts are sometimes useful

1.28 Theorem

(a) Tfd is a tran8iation inva/ iant mel ric on a vector space X then

d{nx, 0) < nd(x, 0) for every x EX and for n = 1,2,3, ....

(b) If {xn} is a sequence in a metrizable topological vector space X and if Xn -4- 0 as n -4- 00, then there are positice scalars ')In such that ')In -4- 00 and ')In Xn -4- O.

PROOF Statement (a) follows from

n

d(nx, 0) ~ L d(kx, (k - l)x) = nd(x, 0).

k=l

For such n,

Boundedness and Continuity

1.29 Bounded sets The notion of a bounded subset of a topological vector space X was defined in Section 1.6 and has been encountered several times since then. When X is metrizable, there is a possibility of misunderstanding, since another very familiar notion of boundedness exists in metric spaces:

22 GENERAL THEORY

If d is a metric on a set ¥, a set E c K is said to he d-hollnded if there is a

number M < 00 such that d{x, y) < M for all x and y In E.

If X is a topological vector space with a compatible metric d, the bounded sets

and the d-bolJnded ones need not be the same, even ifd is invariant For instance, ifd

IS a metnc such as the one constructed In Theorem 1.24, then X itself is d-bounded

(with M = 1) but, as we shall see presently, X cannot be bounded, unless X = {O}. If X is a normed space and d is the metric induced by the norm, then the two notions of boundedness coincide; but if d is replaced by d1 = dl(l + d) (an invariant metric which induces the same topology) they do not.

Whenever bounded subsets of a topological vector space are discussed, it will be understood that the definition is as in Section 1.6: A set E is bounded if, for every neighborhood V of 0, we have E c t V for all sufficiently large t.

We already saw (Theorem 1 15) that compact set,> are bounded. To see another

type of example, let us prove that Cauchy sequences are bounded (hence convergent

sequences are bounded): If {xn} is a Cauchy sequence in X, and Vand Ware balanced

neighborhoods of 0 wjth V + V c W, then [part (b) of Section 1.25] there exists N

such that x, E XN + J7 for all n > N. Take s » 1 so that XN E sV. Then

XnESV I VcsV I sVcsW

En> IV).

Hence x, E t W for all n ~ I, If t IS suffiCIently large.

Also, closures of bounded sets are bounded (Theorem 1.13).

On the other hand, if x =1= 0 and E = {nx: n = 1,2,3, ... }, then E is not bounded, because there is a neighborhood V of 0 that does not contain x; hence nx is not in n V; it follows that no n V contains E.

Consequently, no subspace of X (other than {O}) can be bounded. The next theorem characterizes boundedness in terms of sequences.

1.30 Theorem The following two properties of a set E in a topological vector

space are equivalent:

(a) E is bounded.

(b) if {x,,} is a sequence in E and {all} is a sequence ot scalars such that an ~ 0 as

n ~ 00, then (XnXn ~O as n ~ co.

PROOF Suppose E is bounded. Let V be a balanced neighborb ood of 0 in X.

Then E c t V for some t. If x, E E and an ~ 0, there eXIsts N such that I an It < 1

if n > N. Since t -1 E c V and V is balanced, an x, E V for all n > N. Thus '-.

anxn ~O.

Conversely, if E is not bounded, there is a neighborhood V of 0 and a sequence rn ~ 00 such that no rn V contains E. Choose x, E E such that Xn 1: r.; V. Then no r;; 1xn isin V, so that, {r;; 1xn} does not converge to o. IIII

TOPOLOGICAL VECTOR SPACES 23

1.31 Bounded tineal transformations Suppose Xand Yare topological vector

one whose range i~ a bounded set. In that sense, no linear functIOn (other than 0)

This definition conflicts with the usual notion of a bounded function as being

could ever be bounded. Thus when bounded linear mappings (or transformations) are discussed, it is to .be understood that the definition is in terms of bounded sets, as above.

,

1.32 Theorem Suppose X and Yare topological vector spaces and A: X ~ Y is linear. Among the following four properties of A, the implications

hold. 1] X is metrizable, then also

so that all tour properties are equimlent

(a) . A is continuous.

(b) A is bounded

(c) If Xn ~O then {Axn: n = 1,2,3, ... } is bounded. (d) If xn ~ 0 then AXn ~ O.

Exercise 13 contains an example in which (b) holds but (a) does not.

PROOF Assume (a), let E be a bounded set in X, and let W be a neighborhood of 0 in Y. Since A is continuous (and AO = 0) there is a neighborhood VofO in X such that A(V) c: W. Since E is bounded, E c t V for all large t, so that

l\(£) c: ACtf') zA(V) c tWo

This shows that A(E) IS a bounded set in Y.

Thus (a) "'"4 (b). Since convergent sequences are bounded, (b) =t (c).

A ssume now that X is metrizable, that 1\ satisfies (c), and that Xn ~ O. By

{A(yn xn)} is a bounded set in Y, and now I heorem 1.30 implies that

Theorem 1.28, there are positive scalars l'n ~ 00 such that Tn Xn "'"4 O. HelIce

as n ~ 00.

Finally, assume that (a) fails. Then there is a neighborhood W of 0 in Y such that A -leW) contains no neighborhood of 0 in X. If X has a countable local base, there is therefore a sequence {xn} in X so that Xn ~ 0 but AXn ¢ W. Thus (d) fails. IIII

\

24 GENERAL THEORY

Seminorms and Local Convexity

Ea) p(x + y) ~ pEx) + p(y)

(b) p(ax) = I alp(x)

for all x and y in X and all scalars IX.

Property (a) is called subadditioity, Theorem 1.34 will show that a seminorm p is a norm if it satisfies

(c) p(x)#Oifx#O.

A family (lJJ of semi norms on X is said to be separating if to each x # 0 corresponds at least one p E fjl with p(x) # O.

Next, consider a convex set d c ¥ which is absorbing, in the sense that every

X E X lies in tA for some t - t(x) > O. [For example, (a) of Iheorem 1.15 Imphes that

e"v'Cry neighborhood of 0 in a topological vectol space is absorbing. EvelY absorbing

set obviollsly contains 0] The Minkowskifunctional fLA of 4 is defined by

PA(X) inf{t > O. t lXEA}

(x EX).

Note that f.1A(X) < 00 for all x E X, since A is absorbing. The seminorms on X will turn out to be precisely the Minkowski functionals of balanced convex absorbing sets.

Semi norms are closely related to local convexity, in two ways: In every locally convex space there exists a separating family of continuous seminorms. Conversely, if f!i' is a separating family of seminorms on a vector space X, then (lJJ can be used to define a locally convex topology on X with the property that every p E (lJJ is continuous. This is a frequently used method of introducing a topology. The details are contained in Theorems 1.36 and 1.37.

(e) The set B - {X" p(x) < I} is COnt'~X, balanced, absorbing, andp - fiB'

PROOF Statement (a) follows from p(ax) = IIX Ip(x), with IX = O. The subadditivity of p shows that

p(x) = p(x - y + y) ::s;p(x - y) + p(y)

so that p(x) - p(y) ::s; p(x - y). This also holds with x and y interchanged.

0, (b) Implies (c) If p(X)

IOPouxirCAL VECTOR SPACES 25

o <p(ax + py) < Icxlp(x) + IPlp(y) O.

As to (e), it is clear that B is balanced. If x E B, Y E B, and 0 < t < 1, then

-' _ ...

p(tx + (1 - t)y) ~ tp(x) + (1 - t)p(y) < 1.

Thus B is convex. If x E X and s > p(x) thenp(s-lx) = S-lp(X) < 1. This shows that B is absorbing and also that /lB(X) < s. Hence JlB ~ p. But if 0 < t < pex) then p(t-IX) ~ I, and so rlx is not in B. This implies p(x) ~ /lBeX) and completes the proof. 11I1

1.35 Theorem Suppose A is a convex absorbing set in a vectOJ space X. Then

(c) filA is a seminorm if A is balanced.

(d) It B {x- JiA(X) < I} and C Ex: PA(X) < I}, then B c: A c: C and PH 14. Pc

PROOF Associate with each x E X the set

HA(x) = {t > 0: t 1XEA}.

Suppose t E HA(x) and s » t. Since 0 E A and A is convex, it' follows that s E HAeX). Each HA(x) is a half line whose left end point is /lA(X).

Suppose /lA(X) < s, /lACy) < t, u = S + t. Then S-l x E A, t-1y EA. Since A IS convex,

obvious.

lies in A Hence flA(X ~ y) < u .. This gives (a). PropertIes (b) and (c) are now

, so tha

If PA(X) < 1, then 1 E !fAlx), and so x E A Likewise, if x E A, then

To prove that equality holds, Suppose /lcex) < s < t. Then SIX E C, hence ,lA(S-IX) ~ 1, so that

s fiA(t-IX) < - < 1. t

Thus t-Ix E B, /lBU-IX) ~ 1, /lBeX) st. This completes the proof. 11I1

26 GENERAL THroR I

\

\

\

1.36 Theorem Suppose!7J IS a convex balanced local base tn a topological vector

space X. Associate to eve" V E g(J its Alinko wski functional J1v. Theil {Pv: V E 86'} is a

separating family of continuolls seminorms on l'

PROOF Since V is convex, balanced, and absorbing, fly is a seminorm. If x E X

and x i= 0, then x 1; V for some V E f!4. For this V we have flv(X) :;:::: 1. Thus {fly} is a separating family. If x E V, then tx E V for some t > 1, since V is open. Hence fly < 1 in V. If r > 0, it follows from Theorem 1.34 that

I flv(X) - flv(Y) I :-:; flv(X - y) < r

if x - Y E r V. This proves that each fly is continuous.

IIII

1.37 Theorem Suppose [!P is a separating family ofseminorms on a vector space r

ASSOCiate to each p E !!J and to each positive inteqer n the set

V(p, n) {x: p(X) < ~}.

Let FA be the collectIOn of allfimte zntersectlOns o[the sets v(p, n). Then tJJ IS a convex

balanced local base for a topol-ogy 'f on X, which turns X into a locally convex space

such that

(a) every p E f!J is continuous, and

(b) a set E c X is bounded if and only if every P E f!J is bounded on E.

PROOF Declare a set A c X to be open if and only if A is a (possibly empty) union of translates of members of f!4. This clearly defines a translation-invariant topology 7: on X; each member of f!4 is convex and balanced, and f!g is a local base for r.

Suppose x E X, x i= O. Then p(x) > 0 fm some p E 9. Since x is not in

V(p, 11) if I1p(X) > 1, we see that 0 is not in the neighborhood x V(p, 11) of x,

so that x IS not III the closure of {Of. Thus {OJ IS a closed set, and since r IS

translation-invariant, evely point of X is a closed set.

Next we show that addition and scalar multiplication are continuous.

Let U be a neIghborhood of 0 III X. Then

(1)

for some Pl' ... ,Pm E P and some positive integers nl> ... , nm. Put

(2)

..

Since every P E f!J is subadditive, V + V c U. This proves that addition IS

continuous.

TOPOLOGICAL vl':CIOR SPACES 27

then

X E sV for some s > O. Put t s!(1 + IIX Is) If Y E X I tV and I P ex I < lis,

Suppose now that x EX, 0: is a scalar, and U and V are as above. Then

fly !Xx P(y x) + (P IX)X

which lies in

'~- 1.8ltV+ 1.8-lXlsvc V+ Vc U

SInce 1.8 It -< 1 and V is balanced. This proves that scalar multiplication is continuous.

Thus X is a locally convex space. The definition of V(p, n) shows that every P E f!J' is continuous at O. Hence p is continuous on X, by (b) of Theorem 1.34.

It follows that every p E pj> IS bounded on E

borhood of 0, E c: kV(p, 1) for some k < 00. Hence p(x) < k for every x E E

Finally, suppose E c X is bounded. FIx P E f!J'. Since V(p, 1) is a neigh

1.38 Remarks (a) It was necessary to take finite intersections of the sets V(p, n) in Theorem 1.37; the sets V(P, n) themselves need not form a local base. (They do form what is usually called a subbase for the constructed topology.) To see an example of this, take X = R2, and let f!J' consist of the seminorms PI and P2 defined by Pi(X) = I x;/ ; here x = (Xl' x2). Exercise 8 develops this comment further.

(b) Theorems 1.36 and 1.37 raise a natural problem: If f!4 is a convex balanced local base for the topology r of a locally convex space X, then f!J generates a separating family f!J' of continuous semi norms on X, as in Theorem] 36 This f?J' in turn induces a

topology 't'l on X, by the process described in Theorem 1.37. Is r 't'l ? ~

TIle answer is affirmatIve. To see this, note that every p E # is r-continuous, so

and P llw, then

that the sets V(p, n) of Theorem 1.37 are m r. Hence If c T Conversely, if JVE fA

W {X'llwW < 1) - V(P, 1).

'I hus W E If for every WE f!#; this implies that 't' c 't'J.

(1)

(c) If&' {p i: I 1, 2, 3, ... } is a countable separating family of seminorms on X, Theorem 1.37 shows that f!J' induces a topology 't' with a countable local basco By Theorem 1.24, 't is metrizable. In the present situation, a compatible translationinvariant metric can be defined directly in terms of {pJ Define

d( ) _ ~ 2-iplx - y)

X, Y - L., •

i=ll+Pi(X-y)

It is easy to verify that d is a metric on X. To prove that d is compatible with 1", we

show that the balls

(2)

B, [x: d(x, 0) < r}

(r> 0)

form a local base fOI L.

Since each Pi is continuous (Theorem 1.37) and since the senes (1) converges uniformly on X x X, d is continuous; hence each B, is open. If W is a neighborhood of 0, then W contains the intersection of appropriately chosen sets

(3)

(l sis k).

If x E Br, then

(4)

(I 1, 2, 3, ... ).

If r is small enongh, (4) forces Pl(x)'

, Pk(x) to be so small that Br lIes III each of the

sets (3); hence Br C w.

This proves that d is compatible with 't.

Formula (1) has considerable advantages over the more comphcated construc-

tion of Theorem 1.24. Of course, (1) is applicable only ill locally COIl vex spaces, and

it has a flaw even there: The balls which it defines need not be convex. An example of this is given in Exercise 18.

1.39 Theorem A topological vector space X is normable if and only if its origin has a convex bounded neighborhood.

PROOF If X is normable, and if 11'11 is a norm that is compatible with the topology of X, then the open unit ball {x: IIxll < 1} is convex and bounded.

For the converse, assume V is a convex bounded neighborhood of O. By

Theorem 1.14, V contains a convex balanced neighborhood flofO; of course, U

is also bounded. Define

(1)

Ilxll fl(X)

(XE X)

where J1 is the Minkowski functional of U.

By (c) of Theorem 1.15, the sets rU (r > 0) form a local base for the topoI-

ogy of X. If x i= 0, then x iJ rU for some r » 0; hence Ilxll > r. It now follows

from Theorem 1.35 that (1) defines a norm. The definition of the Minkowski functional, together with the fact that U is open, implies that

(2)

{x: Ilxll < r} = rU

for every r > O. The norm topology coincides therefore with the given one.

, IIII

TOPOLOGICAL VECTOR SPACES 29

. Quotient Spaces

n(x) be the coset of LV that contams x: thus

1.40 Definitions Let N be a subspace of a vector space X. For every x E X, let

These cosets are the elements of a vector space X/ N, called the quotient space of X modulo N, in which addition and scalar multiplication are defined by

(1)

n(x).+ n(y) = n(x + y),

an(x) = n(ax).

[Note that now an(x) = N when ex = O. This differs from the usual notation, as introduced in Section 1.4.] Since N is a vector space, the operations (1) are well

defined. This means that if n(x) n(x') (that is, x' x E N) and n(y) n(y') then

(2)

n(x) + n( y} nex') + n(y'),

o:n(x') Ctn(x).

lf as its null space; tt IS often called the quotient map of X onto xtN.

The origin of XfN is nEo) 1'1. By (1), n is a linear mappmg of X onto XjN with

Suppose now that l' is a vector topology on X and that N is a closed subspace of

x · Let I,'I' be the collection of all sets E c XliV for which tt 1(E) E T. Then IX turns o~t ;0 "be a topology on X,IN, called the quotient topology. Some of its properties are

listed in the next theorem. Recall that an open mapping is one that maps open sets to open sets.

1.41 Theorem Let N be a closed subspace of a topological vector space X. Let T be the topology of X and define TN as above.

(a) TN is a vector topology on X/N; the quotient map n: X ~ X/N is linear, continuous, and open.

(b) If f!8 is a local base for I, then the collection &/ all sets irCV) with V E ~ is a local

N'

(d) If X is an I=space; or a Frechet space, or a Banach "'pace, so is X/N.

PROOF Si nce n 1 (A A B)~':'" 1C 1 (A) n n 1 (B) and

tt 1(UE;.) Un 1 (E;.),

TN is a topology. A set Fc X/N is TN-dosed if and only if n-\F) is r-closed, In particular, every point of X/N is closed, since

n-1(n(x» = N + x

and N was assumed to be closed.

The continuity of n follows directly from the definition of IN' Next,

suppose V E I. Since

and N + V E I, it follows that n(V) E IN' Thus tt is an open mapping.

If now W is a neighborhood of 0 in XI N, there is a neighborhood V of o in X such that

Hence n(V) + n(V) c: W. Since tt is open, n(V) is a neighborhood of 0 in XIN. Addition is therefore continuous in XI N.

The continuity of scalar multiplication in XI1Vis proved in the same manner.

This establishes (a)

It IS clear that (a) ImplIes (b). WIth the aId of Theorems 1.32, 1.24, and

1.39, it is just as easy to see that (b) implies (c).

Suppose next that d is an invariant metric on ¥, compatible with 1:

p(n(:x), n(y)) iar {u'(x y, z). Z E lv'}.

This may be interpreted as the distance from x - y to N. We omit the verifications that are now needed to show that p is well defined and that it is an invariant metric on XI N. Since

n({x: d(x, 0) < r}) = {u: p(u, 0) < r},

it follows from (b) that p is compatible with IN'

If ¥ is normed, this definition of p specializes to yield what is usually

called the quotIent norm of XjN:

Iln(x)11 inf{llx zll: z EN}.

To prove (d) we have to show that p is a complete metric whenever d is

Suppose {unf IS a Cauchy sequence In X/N, relatIve to p. There is a

subsequence {un) with p(uni, uni + J < 2 - '. One can then inductively choose Xi E X such that n(xi) = uni and d(xi, Xi+1) < z:'. If d is complete, the Cauchy sequence {xJ converges to some XE X. The continuity ofnimplies that uni -+ n(x) as i -+ 00. But if a Cauchy sequence has a convergent subsequence then the full sequence must converge. Hence p is complete, and so is the proof of Theorem

1.41. II11

TOPOLOGICAL VECTOR SPACES 31

Here is' an easy application of these concepts:

1.42 TlleOIem Suppose Nand Fare subspaces ot a topological vector space X,

N is closed and F has finite dimension. Theil 1'/ + F is closed.

PROOF Let t: be the quotIent map of X onto X/N, and give X/N its quotient topology. Then n(F) is a finite-dimensional subspace of X/N; since X/N is a topologicalvector space, Theorem 1.21 implies that n(F) is closed in X / N. Since N + F = 1l;,-l(n(F» and tt is continuous, we conclude that N + Fis closed. (Compare Exercise 20). . / / / /

1.43 Seminorms and quotient spaces Suppose p is a semi norm on a vector space Xand

If 1t(x) n(v), then p(x y) 0, and since

Ip(x) - p(y) I ::;; p(x - y)

it follows that jJ(n(x» = jJ(n(y». Thus jJ is well defined on X/N, and it is now easy to verify that jJ is a norm on X/No

Here is a familiar example of this. Fix r, 1 :::;; r < OCJ; let E be the space of all Lebesgue measurable functions on [0, 1 J for which

{ 1 } l/r

pel) = 1I!llr = {If(t)lr dt < OCJ.

This defines a seminorm on IJ, not a norm, since Ilfllr 0 whene-ver f 0 almost

everywhere. Let N be the set of these" null functions" Then J:jN is the Banach

space that is usually called E. The norm of E is obtained by the above passage from

Examples

1.44 The spaces C(Q) If Q is a nonempty open set in some euclidean space, then Q is the union of countably many compact sets K; i= 0 which can be chosen so that K; lies in the interior of Kn+l (n = 1,2,3, ... ). C(Q) is the vector space of all complexvalued continuous functions on Q, topologized by the separating family of serninorms

(1)

Pn(f) = sup {1/(x)I : x E Kn},

ill aeeo! danee with TheOl em 1. 37. Si nee P 1 5:, P 2 5:, •• " the se ts

(2)

(n - 1,2,3, ... )

form a convex local base for C(Q). .A~ccording to remark (c) of Section 1.38, the

topology of C(Q) is compatible with the metric

(3)

d(f, g) = f 2-npnCf - g) . n = 1 1 + Pn(f - g)

If {Ji} is a Cauchy sequence relative to this metric, then PiJi - fj) -+ 0 for every n, as i, j -+ 00, so that {~} converges uniformly on Kn, to a function J E Cen). An easy computation then shows deJ,~) -+ O. Thus d is a complete metric. We have now ploved that C(fl) is a Fdchet space.

By (0) of Theorem 1.37, a set E c C(Q) is bounded if and only if there are

numbers M,. < 00 such that Pn( f) < Mn for aIlfE E; explicitly,

(4)

if IE E and x E Kn'

Since every V;; contains an/for WhICh Pn±l (t) IS as large as we please, It follows that

no Vi, is bounded. Thus CeQ) is not locally bounded, hence is not normable:

1.45 The spaces H(n) Let Q now be a nonempty open subset of the complex plane, define C(Q) as in Section 1.44, and let H(n) be the subspace of C(Q) that consists of the holomorphic functions in Q. Since sequences of holomorphic functions that converge uniformly on compact sets have holomorphic limits, H(Q) is a closed subspace of C(Q). Hence H(Q) is a Frechet space.

We shall now prove that H(n) has the Heine-Borel property. It will then follow from Theorem 1.23 that H(n) is not locally bounded, hence is not normable.

Let E be a closed and bounded subset of HEQ). Then E satisfies inequalities

such as (4) of Section l.44. Mantel's classical theorem about normal families (Th 146

of [23]1) implies therefore that every sequence {fa c E has a subsequence that con-

yerges uniformly on compact subsets of n [hence in the topology of H(Q)] to some

IE H(O). Since E is closed, IE E This proves that E is compact

l.46 The spaces coo(Q) and @K We begin this section by introducing some

termmology that WIn be used III our later work with distributions.

-In any discussion of functions of n variables, the term multi-index denotes an ordered n-tuple

(1)

1 Numbers in brackets refer to sources listed in the Bibliography.

mpOIOQIC<\L \lBCTOR SPACES 33

operator

of nonnegative integers ai• WIth each multi-index (X is associated the differentIal

(2)

whose order is

~/

(3) If I a I = 0, Dj = f A complex functionj defined in some nonempty open set Q c R" is said to belong to COO(Q) if DJ E C(Q) for every multi-index a.

I a I = al + ... + an .

{x:f(x) ~ O}. -

The Support of a complex fllnction/Eon any topological space) is the closure of

If K is a compact set in Rn, then ~K denotes the s12ace o[al1 IE 9°O(RII)~vhose support lies in R. (The letter q; has been used for thes; spaces ever sIllceSchwartz

We now define a topology on C<X>(Q) which makes COO(ft) into a Frechet space with

the Heine-Borel property, such that f»K is a closed subspace of COO(Q) whenever K c Q.

To do this, choose compact sets K, (i = 1,2,3, ... ) such that K, lies in the interior of Kj+l and Q = UKj• Define seminorms PN on COO(Q), N = 1,2, 3, ... , by setting

(4)

PN(f) = max {I Dj(x) I : x E KN, lal ~ N}.

They define a metrizable locally convex topology on C<X>(O); see Theorem 1.37 and remark (c) of Section 1.38. For each xEQ, the functional/~f(x) is continuous in

ranges over the complement of K, it follows that f!JK is closed iII COO(Q). •

thIS topology Since ~ is the intersection of the null spaces of these functionals, as x

A local base is given by the sets

(5)

(N 1,2,3, ... ).

If Ui} is a Cauchy sequence III e"'(!!) (see Section 1.25) and if N is fixed, thenj', _ Jj E VN if i and j are sufficiently large. Thus f trj, - DJj I < liN on KN, if I a I s N. It follows that each D'}; converges (uniformly on compact subsets of Q) to a function gGt. In particular, /;(x) ~ go(x). It is now evident that go E COO(O), that gGt = DGtgo, and that/; ~ 9 in the topology of COO(O).

Thus COO(Q) is a Frechet space. The same- is true of each of its closed subspaces f» K •

34 GEl'ERAL THEORY

Suppose next that E c CCXl(Q) is closed and bounded. By Iheorem 1.37, the

boundedness of E is equivalent to the existence of numbers MN < 00 such that

PNG,) < MN for N 1, 2, 3, ... and for all IE E. The inequalities I D'il < M,"1, valid

on K,'" when I Ct I < N, imply the equicontinuity of {DPj': j E E} on KN l' If I B I < N - 1.

It now follows from Ascoli's theorem (proved ill Appendix A) and CantoI's diagonal

process that every sequence in E contains a subsequence {Ii} for which {DPli} converges, uniformly on compact subsets of Q, for each multi-index {3. Hence {[;} converges in the topology of CCXl(Q). This proves that E is compact.

Hence CCXl(Q) has the Heine-Borel property. It follows from Theorem 1.23 that CCXl(Q) is not locally bounded, hence not normable. The same conclusion holds for !!)K whenever Khas nonempty interior (otherwise zsy = {O}), because dim!!)K = 00 in that case. This last statement is a consequence of the following proposition:

!fBl andB2 me COllcentlic closed balls in Rn, with Bl in the interior ofBz, then

there exists ¢ € CCXl(Rn) such that 4>(X) 1 for every x E 81 and ¢(x:) 0 for every x

To find sllch a ¢, we constmct g E CCXl(Rl) such thatg(x) 0 for x < a, g{x) 1

for x > b (where 0 < a < b < 00 are preassigned) and put

(6)

The following construction of 9 has the advantage that suitable choices of {<5J can lead to functions with other desired properties.

Suppose 0 < a < b < 00. Choose positive numbers <50, b1, b2, ••• , with Lbi = b - a; put

.. ~

(7)

(n= 1,2,3, ... );

let fo be a continuous monotonic function sl]ch that loCX:) 0 when x < a, hex) 1

when x > a + (j . and define

(8)

fix)

in let) dt

Cn 1,2,3, )

Differentiation of this integral shows, by induction, thatJi, has 11 continuous derivatives,

and that I nnfn I < mn, If n > r, then

(9)

1 filM

D1n(x) = y (Drfn_l)(X - t) dt,

n 0

so that

(10)

(n ;?: r),

TOPOLOGICAL VECTOR SPACES 35

again by induction on n. The mean value theorem, applied to (9), shows that

(11)

(n> r + 2).

Since !<5n < 00, each {D7,J converges, uniformly on ( 00, (0), as 11 ~ 00. Hence fin)

converges to a function g, with J D'gl < m, for r 1, 2, 3, ... , such that g(x) 0 for x < a and g(x) = 1 for x > b .

. ~ .

1.47 The space~ LP with 0 < p < 1 Consider a fixed p in this range. The elements of II are those Lebesgue measurable functions / on [0, 1] for which

(1)

1

11([) = f I/(t) I Pdt < 00,

o

o <p < 1, the inequality

WIth the usual identification of flJnctions that coincide almost everywhere. Since

holds when a > 0 and b > O. ThIS gIves

(3) ;

AU' + g) < Ji(f) + n.Cg),

so that

(4)

dtf; g) = 11([ - g)

defines an invariant metric on Il'. That this d is complete is proved in the same way as in the familiar case p ;:::: 1. The balls

(5)

B, = {/E I!': 11(f) < r}

Thu£ I! is a locally bounded F space.

We claIm that D' contains no convex open sets, other than 0 and 1!'.

To prove this, suppose V 1-= 0 is open and convex III I!'. Assume 0 E V. WIthout

of If I P, there are points

positive Integer n such that nP t;!lU) < r. By the continnity of the indefinite integral

Joss of generality Then V:;) Br, for some r> O. PieklE I!. Since p < 1, thele is a

o = Xo < Xl < ... < x, = 1

such that

(6)

IXi If(t)IP dt = n-1 A(f)

Xi-l

(1 :::;; i S; n).

Define g/t) - nf(e) if Xi-l < t < Xi> g/t) - 0 otherwise. Then OJ E P, since (6) shows

(7)

(1 < i < n)

and V:::oJ Br• Since V is COil vex and

(8)

it follows that j s V. Hence V = If.

This lack of convex open sets has a curious consequence.

Suppose A : If ~ Y is a continuous linear mapping of If into some locally convex space Y. Let PJJ be a convex local base for Y. If WE PJJ, then A -leW) is convex, open, not empty. Hence A -leW) = If. Consequently, A(If) c W for every WE PJJ. We conclude that ,Iif 0 for every IE I1'.

Thus 0 is the only continuous linear mapping of IT into any locally convex space t

i/O < p < 1. In particular, 0 is the only continuous linear functional on these I!-spaces.

This is, of course, in violent contrast to the familiar case]J > 1.

1 Suppose X IS a vector space. All sets mentioned below are understood to be subsets of X.

Prove the following statements from the axioms as given in Section 1.4. (Some of these are tacitly used in the text.)

(a) If x E X and y E X there is a unique Z E X such that x + Z = y. (b) Ox = 0 = exO if x E X and ex is a scalar.

(c) 2A c A + A; it may happen that 2A -::j: A + A.

(d) \(f is convex if and only if (s + I)A = sA + IA for all positive scalars s and I. (e) Every union (and intersection) of balanced sets is balanced.

(f) Every intersection of convex sets is convex.

(0) If I' IS a collectIon of convex sets that IS totally ordered by set mcluslOn, then the

union of all members of r is convex.

(h) If A and B are convex, so IS A + B.

(i) If A and B are balanced, so is A + B.

(j) Show that parts Cn, (g), and (h) hold with subspaces in place of convex sets.

2 The convex hua of a set A in a vector space X is the set of all convex combinations of

members of A, that is, the set of al1 sums

in which x, E A, t, > 0,2: I, = 1; n is arbitrary. Prove that the convex hull of A is convex and that it is the intersection of all convex sets that contain A.

3 Let X be a topological vector space. All sets mentioned below are understood to be subsets of X. Prove the following statements.

(a) The convex hull of every open set is open .

. .

TOPOLOGICAL VECTOR SPACES 37

(b) If X IS locally convex then the convex hull of every bounded set is bounded. (I Iiis IS

(c) an B are bounded, so IS .4 + B.

(d) If A and B are compact, so is A + B.

(e) If A is compact and B is closed, then A I B is closed.

(f) The sum of two closed sets may fail to be closed. [The inclusion -in (b) of Theorem 1.13 may therefore be strict.]

4 Let B = {(zt, Z2)-E (2'2: I ztl < IZ2 D. Show that B is balanced but that its interior is not. [Compare with (e) of Theorem 1.13.]

5 Consider the definition of "bounded set" given in Section 1.6. Would the content of this definition be altered if it were required merely that to every neighborhood V of 0 corresponds some t > 0 such that E c tV?

6 Prove that a set E in a topological vector space is bounded if and only if every cOlluta ble

subset of E IS bounded.

7 Let X be the vector spaCe of all complex functions on the unit interval [0, 1], topologized

by the frumly of seminorms

Px(f) If(x) I

(O<x<l).

This topology IS called the topology a/pointwise convergence. Justify this terminology.

~ ;~7 ~hat there is a sequence {t.} in X such that (a) U:'} converges to 0 as n =* 00,

ut (b 1 Yn IS any sequence of scalars such that Yn -+ 00 then {Ynf,,} does not converge to O. (Use the fact that the collection of all complex sequences converging to 0 has the same cardinality as [0, 1].)

This shows that metrizability cannot be omitted in (b) of Theorem 1.28.

8 (a) Suppose f!J' is a separating family of seminorms on a vector space X. Let fl be the smallest family of seminorms on X that contains f!J' and is closed under max. [This means: IfplEfl,P2Efl, andp=max(Pl,P2), thenpEfl.] If the construction of Theorem 1.37 is applied to f!J' and to fl, show that the two resulting topologies coincide. The main difference is that f2 leads directly to a base, rather than to a subbase

[See Remark (a) of Section 1.38.J

(6) Suppose f2 is as in part (a) and A is a linear functional on X. Show that A is contin,

uous If and only if there exists apE q such that I Axl < ,Hp(x) fOI all x E Xand some

(b) A: X -+ Yis linea

(a) X and Yare topological vector spaces,

?

(c) N is a closed subspace of X, _

(d) 1T: X -+ X/N is the quotient map, and (e) Ax = 0 for every x E N.

Prove that there is a unique f: X/ N -+ Y which satisfies A = / 0 1T, that is, Ax = /(1T(X» for all x E X. Prove that this/is linear and that A is continuous if and only if/is continuous. Also, A is open if and only if / is open.

38 GENERAL THEORY

10

(b) Assume, in addition, that the null space of " is closed, and prove that A is then

continuous.

11 If]'v' is a subspace of a vector space X, the codimension of N in ¥ is, by definition, the

dimension of the quotient space X/No

Suppose 0 <p < 1 and prove that every subspace of finite codimension is dense in

U. (See Section 1.47.)

12 Suppose d1(x, y) = Ix - YI, d2(x, y) = 14>(x) - 4>(Y)I, where 4>(x) = x/(l + Ix\). Prove that d, and d, are metrics on R which induce the same topology, although di is complete

and d2 is not.

13 Let C be the vector space of all complex continuous functions on [0, 1]. Define

Let (C, a) be C with the topology induced by this metric Let (c, 7) be the topological

vector space defined by the semmorms

pi]) If(x) I

(O<x<l),

in accordance WIth Theorem 1.37.

(a) Prove that every T bounded set in C is also a bounded and that the identity map

id: (C, 7)-l>-(C, a) therefore carries bounded sets into bounded sets.

(b) Prove that id: (C, 7)-l>-(C, a) is nevertheless not continuous, although it is sequentially continuous (by Lebesgue's dominated convergence theorem). Hence (C, 7) is not metrizable. (See Appendix A6, or Theorem 1.32.) Show also directly that (C, 7) has no countable local base.

(c) Prove that every continuous linear functional on (C, 7) is of the form

for some choice of Xl, ••• , X" in [0, 1] and some CI E Q'.

(d) Prove that (C, a) contains no convex open sets other than 0 and C.

(e) Prove that id: (C, a) --* (C, '1") is not continuous

14 Put K [0, 1] and define !YJ K as III SectIOn 1.46. Show that the following thIee families

of seminorms (v,l1ere n 0, 1, 2, ... ) define the same topology on !l)K, if D dldx:

(a) IID"flloo -suP {I D"f(x) I : co <x< co}.

(c) IID"fI12 = {( ID"f(X)12dXr2.

15 Prove that the spaces C(O) (Section 1.44) do not have the Heine-Borel property.

16 Prove that the topology of C(O) does not depend on the particular choice of {Kn}, as long as this sequencesatisfies the conditions specified in Section 1.44. Do the same for Coo(D) (Section 1.46).

IUPOLOGICAL VECTOR SPACES 39

17 In the setting of Section 1.46, pIOve that f -- D"j IS a contmuous mapping of Coo(D) into

C''''(U) and also of !0" into 9f[;., for every multi index 01:.

18 The semi norms

pi!) Slip { lAx) I : n < x < n)

induce the metric

d(f,g) = ~ 2-npn(f - g) n=1 1 +Pn(f-g)

in the space C(R); compare Section 1.46 and remark (c) of Section 1.38. Define

f(x) = max (0, 1 -Ixl),

g(x) = 100f(x - 2),

2h =Lv a,

and compute that

d( f, 0)

2'

d(g,O)

d(h, 0)

The halls with radius t are therefore not convex, although a'is compatible wIth the

usual locally convex topology of C(R).

Is there any r < 1 for vmich the balls ofIadius r are convex?

19 Suppose M IS a dcnse subspace of a topological vector space X, Y is an F space, and

A: M -- Y is continuous (relative to the topology that M inherits from X) and linear.

Prove that A has a continuous linear extension A: X __ Y.

Suggestion: Let Vn be balanced neighborhoods of 0 in X such that Vn + Vn C Vn-l and such that d(O, Ax) < 2-n if x E M n Vn. If x E X and Xn E (x + Vn) n M, show that {Axn} is a Cauchy sequence in Y, and define Ax to be its limit. Show that A is well defined, that Ax = Ax if x E M, and that A is linear and continuous.

20 For each real number t and each integer n, define en(t) = e1nt, and define

(n = 1, 2, 3, ... ).

L2 that contains /1,/2 ,f3, .... Show that Xl + X2 IS dense in L2 but not closed For

space of L2 that contains Po, Pl, e2, , and let X;, be the smallest closed subspace of

Regard these functions as members of L2( 1T,1T). Let Xl be the smallest closed sub-

instance, the vector

n

is in L1. but not in Xl + X2• (Compare with Theorem 1.42.)

21 Let Vbe a neighborhood of 0 in a topological vector space X. Prove that there is a real continuous function f on X such that f(O) = 0 and f(x) = 1 outside V. (Thus X is a completely regular topological space.) Suggestion: Let Vn be balanced neighborhoods of 0 such that VI + VI C V and Vn+I + Vn+1 C Vn• Construct f as in the proof of Theorem 1.24. Show that f is continuous at 0 and that

If(x) - fey) I < f(x - y).

wo(f) sap {I/(x) f(y)I:lx y\<o,xEI,yEI}

22 If/is a complex function defined on the compact interval I [0,11 c; R, define

If 0 < IX < 1, the correspondmg LipschItz space Lip IX consists of all f for which

II!II 1/(0)1 + SliP {o «w.(f): (5 > O}

is finite. Define

lip IX = {fE Lip IX: lim 8-«woCf) = O}.

0 .... 0

Prove that Lip cx is a Banach space and that lip cx is a closed subspace of Lip cx.

23 Let X be the vector space of all continuous functions on the open segment (0, 1). For fE X and r > 0, let V(t, r) consist ofallg E Xsuch that Ig(x) - f(x) 1< r for all x E (0, 1). Let 7" be the topology on X that these sets V(t, r) generate. Show that addition is 7"-

24 Show that the set W that occurs in the proof of Theorem 1 14 need not be convex, and

continuolls but scalar multiplicatIOn IS not.

that A need Dot be balanced unless U is convex.

COMPLETENESS

. - -

The validity of many important theorems of analysis depends on the completeness of the systems with which they deal. This accounts for the inadequacy of the rational number system and of the Riemann integral (to mention just the two best-known examples) and for the success encountered by their replacements, the real numbers and the Lebesgue integral. Baire's theorem about complete metric spaces (often called the category theorem) is the basic tool in this area. In order to emphasize the role

Theorems 2.7 and 2.11) are stated ill a little more generalIty than IS usually needed

played by the concept of category, some theorems of this chapter (for instance,

applications) are also gIven.

When this is done, simpler versions (more easily remembered but sufficient for most

Baire Category

2.1 Definition Let S be a topological space. A set E c S is said to be nowhere dense if its closure E has empty interior. The sets of the first category in S are those that are countable unions of nowhere dense sets. Any subset of S that is not of the first category is said to be of the second category in S.

This terminology (due to Baire) is admittedly rather bland and unsuggestive.

42 GENERAL THEORY

A/eager and nonmeager bave been IIsed instead in some texts. But" category argu-

ments " are so entrenched in the mathematical literature and are so well known that it

seems pointless to insist on a change.

Here are some obviolls propertie'l of category that will be freely used III the

se ue :

(a) If A c Band B is of the first category in S, so is A.

(b) Any countable union of sets of the first category is of the first category. (c) Any closed set E c S whose interior is empty is of the first category in S ..

(d) If h is a homeomorphism of S onto S and if E c S, then E and h(E) have the same category in S.

2.2 Baire's theorem If S is either

(a) a complete metric space , or

(b) a locally compact Hausdorff space,

then the intersection of every countable collection qf dense open subsets oj S is dense

in .

This is often called the category theorem, for the following reason.

If {EJ is a countable collection of nowhere dense subsets of S, and if Vi is the complement of Ei, then each Vi is dense, and the conclusion of Baire's theorem is that n Vi i= 0· Hence S i= UEi•

Therefore, complete metric spaces, as well as locally compact Hausdorff spaces,

are of the second category in themselves.

PROOF Suppose Vi' V2, V3, ••• are dense open subsets of S. Let Bo be an arbitrary non empty open set in S. If n ~ 1 and an open Bn-i i= 0 has been

chosen, then (because P n IS dense) there exists an open B; ¥ 0 with

n r-

In case (a), Bn may be taken to be a ball ofradius < lin; in case (b) the choi ce can

be made so that ~ is compact Pllt

n=1

Tn case (a), the centers of the nested balls B; form a Cauchy sequence which converges to some point of K, and so K i= 0. In case (b), K i= 0 by compactness. OUf construction shows that K c Bo and K c Vn for each n. Hence Bo intersects

n Vn• ' I111

The Banach-Steinhaus Theorem

collection of linear mappings from X into Y. We say that F is equicontilluollS if to

2.3 Equieontinuity Suppose X and Yare topological vector spaces and I IS a

every neighborhood W of 0 In Y there corresponds a neighborhood V of 0 in X £uch

that A(V} c W for all A E F.

If r contains .only one A, equicontinuity is, of course, the same as continuity (Theorem 1.17). We already saw (Theorem 1.32) that continuous linear mappings are bounded. Equicontinuous collections have this bounded ness property in a uniform manner (Theorem 2.4). It is 'for this reason that the Banach-Steinhaus theorem (2.5) is often referred to as the uniform boundedness principle.

2.4 Theorem Suppose ¥ and Yare topological vector spaces, F is all eqtttcott=

tintrotts collection of ttn ear mappmgs from X into Y, and E is a bounded wb"et of l' Then

Y has a bounded subset F Stich that A(E) c Flor every A E F.

PROOF Let} be the uruon of the sets A(E), for t\ E r I et W be a neighborhood

of 0 in Y. Since F is equicontinuous, there is a neighborhood V of 0 III X such

that t\(V) c W for all A. E r. Since E is bounded, E c tV for all sufficiently

large t. For these t,

A(E) c A(tV) = tA(V) c tW,

so that F c t W. Hence Fis bounded.

IIII

2.5 Theorem (Banach-Steinhaus) Suppose X and Yare topological vector spaces, r is a collection oj continuous linear mappings from X into Y, and B is the set oj all x E X whose orbits

II B lS ot the second category in X, then B X and r is equicontinuous.

PROOF Pick balanced neighborhoods Wand U of 0 in Y such that U + U c w.

ut

If x E B, then rex) c nU for some n, so that x EnE. Consequently,

00

Be U nE.

n= 1

At least one nE is of the second category in X, since this is true of B. Since x -+ nx is a homeomorphism of X onto X, E is itself of the second category in X.

44 GENERAL THEORY

But E is closed because each ,A .. is continuous. Therefore E has an interior point

x Then x - E contains a neighborhood VofO III X, and

A(V) c Ax A(E) c U U c J¥

for every A E F.

This proves that F is equicontinuous. By Theorem 2.4, T is uniformly

bounded; in particular, each rex) is bounded in Y. Hence B = X. III/

In many applications, the hypothesis that B is of the second category is a consequence of Baire's theorem. For example, F-spaces are of the second category. This gives the following corollary of the Banach-Steinhaus theorem:

2.6 Theorem If r is a collection ~f continuous linear mappings from an F ~pace y

are bounded in Y,for every x E X, then r is equicontinllolls

Briefly, pointwise boulldedness implies uniform boundedness (Theorem 2.4.)

As a special case of Theorem 2.6, let X and Y be Banach spaces, and suppose

that

(1)

sup IIAxl1 < CIJ

AEr

for every x E X.

The conclusion is that there exists M < CIJ such that

(2)

II Axil ~ M if IIxll ~ 1 and A E r.

IIAxll < M IIxll If x E X and A E r.

The following theorem establishes the continuity of limits of sequences of

continuous linear mappings

2.7 Theorem Suppose Xand Yare topolagica/vector spaces, and {All} is a sequence

of continuous linear mappings of X into Y.

.

(a) If C is the set of all x EX for which {A,«} is a Cauchy sequence in Y, and if C

is of the second category in X, then C = X.

(b) If L is the set of all x E X at which

Ax = lim A,,»

,,-+00

ana 1\ • ~ IS continuo

x

PROOF (a) Since Cauchy sequences are bounded (Section 1.29) the Banach-

Steinhaus theorem asserts that {An} is eqmcontmuous.

One checks easily that C is a subspace of X. Hence C IS dense. (Other-

WIse, C is a J~~~oper subspace of X; proper subspaces have empty interior; thus C would be of the first category.)

Fix X'E X; let Wbe a neighborhood of 0 in Y. Since {An} is equicontinuous, there is a neighborhood V of 0 in X such that An(V) c: W for n = 1, 2, 3, .... Since C is dense, there exists x' E C n (x + V). If nand m are so large that

sequence III ,and x E .

shows th~t :." x Am~E W! W+ w. Consequently, {Anx} is a Cauchy

that A(V) c: W Thus A is continuous. ///1

If V and Ware as above, the inclusion An(V) c: W, valId for all n, implies now

(b) The completeness of Y implIes that L C Hence L X, by (a).

The hypotheses of (b) of Theorem 2.7 can be modified in various ways. Here is an easily remembered version:

2.8 Theorem If {An} is a sequence of continuous linear mappings from an F-space X into a topological vector space Y, and if

exists /01 every x E X, then A is cantinuau S'

PROOF Theorem 2.6 ImplIes that {All} is equicontinuous. Therefore if W is a

continuous. jill

In the following variant of the Banach-Steinhaus theorem the category argument is applied to a compact set, rather than to a complete metric one. Convexity also enters here in an essential way (Exercise 8).

46 GENERAL THEORY

2.9 Theorem Suppose X and Yare topological vector spaces, K IS a compact

r ts a collectlOn of continuous linear mappings at X into Y, and the

rex) {Ax: A En

are bounded subsets of' Y, for every x c K.

Then there is a bounded set BeY such that A(K) c B Jor every A E r.

PROOF Let B be the union of all sets r(x), for x c K. Pick balanced neighborhoods Wand U of 0 in Y such that V + [J c W, Put

(1)

If xc K, then r(x) c nU for some n, so that x EnE. Consequently,

00

(2)

K U (Kn nE).

n 1

Since E is closed, Balre's theorem shows that K nnE has nonempty interiOI

(relative to K) for at least one n.

\Ve fix such an fl, we fix an interior point Xo of K nnE, we fix a balanced

neighborhood V of 0 III X such that

(3)

K n (xo + V) enE,

and we fix a p > 1 such that

(4)

K c Xo +pV.

Such a p exists since K is compact.

If now x is any point of K and

(5)

then Z E K SInce K IS convex. Also,

(6)

Z Xo p ex xo) c V,

by (4). Hence Z EnE, by (3). Since A(nE) c nU for every A c r and SInce

x -pz (p

l)x me have

0' y~ ,

Ax EpnU (p l)nV cpn(U + U) cpnW.

. Thus B c pn W, which proves that B is bounded.

iIiI

The Open Mapping Theorem

2.10 Open mappings. SupposeJmaps S into T, where Sand T are topological spaces. We say that f is open at a point PES if f(V) contains a neighborhood of

COMPLETENESS 47

f(P) wheneve.r'v ~ a. ~ghborhood ofp. We say that t IS open if/(U) is open, in T

whenever U 15 op n In '.

Let us also note that a one-to-one continuous mappingj of S onto T is a homeomorphism precisely- when j is open.

logical vector space into another is open if and only if it is open at the origin,

of the invanance of vector topologies, it follows that a lInear mapping of one topo-

It is c~ear that f is open if and only It f is open at every point of S. Because

2.11 The open mapping theorem Suppose (a) X is an F-space,

(b) Y is a topological vector space,

Cd) t\(X) is of the second category In Y.

(c) A. X""""+ Y is continUous and linear, and

T.

y

(iii) Yis an Fspace.

PROOF Note that (U) implies (i), since Y is the only open subspace of Y. To prove (ii), let V be a neighborhood of 0 in X. We have to show that A(V) contains a neighborhood of 0 in Y.

Let d be an invariant metric on X that is compatible with the topology of X. Define

(1)

VII = {x: d(x, 0) < 2-nr}

(n = 0, 1 , 2, ... ),

of 0 in Y satisfies

where 1 > 0 is so small that Vo c V We 'NiH prove that some neighborhood W

Since Vl ::5 V2 - V 2 ' statement (b) of Theorem 1.13 implies

(3)

The first part of (2) will therefore be proved if we can show that A( V2) has nonempty interior. But

(4)

co

A(X) = U kA(V2),

k=l

48 GENERAL THEORY

because V2 is a neighborhood of O. At least one kA( V2) is therefore of the second

category in Y. Since Y -+ ky is a homeomorphism of Yonto Y, A( V2) is of the

second category in Y. 1ts closure therefore has nonempty interior.

To prove the second inclusion in (2), fix )'~ E l\( ~). Assume n > 1 and

Yn has been chosen in A( Vn). What was just proved for VI holds equally well for Vn+l, so that A(Vn+l) contains a neighborhood ofO. Hence

(5)

This says that there exists x; E Vn such that

(6)

Put Yn+l Yn Axn· Theil Yn+l EA(Vn+1), and the construction proceeds:

Since d(xn' 0) < 2 -nr, for n 1, 2, 3, ... , the sums Xl + ... + Xn form a

Cauchy sequence whIch converges (by the completeness of X) to some x E X,

with dix; 0) < r. Hence x E V. Since

(7)

Yn+l) - Yl Ym+l,

and since Ym+ 1 -+ a as m -+ 00 (by the continurty of A), we conclude that Yl - Ax E A( V). This gives the second part of (2), and (ii) is proved.

Theorem 1.41 shows that XIN is an F-space, if N is the null space of A.

Hence (iii) will follow as soon as we exhibit an isomorphism f of XjN onto Y which is also a homeomorphism. This can be done by defining

(8)

f(x + N) = Ax

(x EX).

It is trivial that this f is an isomorphism and that Ax = f(n(x», where tt is the quotient map described in Section 1 40 If V is open in Y, then

(9)

f - J ( V) n(A - 1 ( V»)

I( E) - A(zr (E»)

is open, since n is continuous and A is open. Consequelitly, f is a homcomor-

.phism. IIII

2.12 Corollaries

(a) If A is a continuous linear mapping of an F-space X onto an F-space Y, then A is open.

for every x E X.

(d) If rIc r 2 are cector topologies on a vector space X and if both (X, r 1) and (X, r 2) are F-spaces, then r 1 = r 2 •

PROOF Statement (a)'follows from Theorem 2.1] and Baire's theorem, since Y is now of the second category in itself. Statement (b) is an immediate con-

sequence of (a), and (c) follows from (b). The two inequalities in (c) simply express the continuity of /\. 1 and of A. Statement (d) is obtained by applying

(b) to the identity mapprng of (X, r2) onto (X, Ta. 1177

The Closed Graph Theorem

if X x f'" is gi ven the usual product topology (the smallest topology that contains all sets U x V with U and V open in X and Y, respectively), and if J: X ~ Y is continuous, one would expect the graph of J to be closed in X x Y (Proposition 2.14): For linear mappings between F-spaces this trivial necessary condition is also sufficient to assure continuity. This important fact is proved in Theorem 2.15.

2.14 Proposition If X is a topological space, Y is a Hausdorff space, and f: X ~ Y is continuous, then the graph G oj J is closed.

~o =f itxo)· Thus Yo and !(xo) have disjoint neighborhoods V and HZ ill Y.

PROOF Let!l be the complement of G in X x Y, fix (xo, Yo) E !l. 1 hene

neighborhood U x V of (xo , Yo) lies therefore in n. This pi aves that 12 IS open.

mce f is continuous, Xo has a neighborhood U such that I(ll) c W The

consider an arbitrary topological space X, and let J: X ~ X be the identity. Its graph is the diagonal

N01e. One cannot omit the hypothesis that Yis a Hausdotif space. To see thiS,

D = {(x, x): x E X} C X x X.

The statement" D is closed in X x X" is just a rewording of the Hausdorff separation axiom.

2.15 lhe closed graph theorem Suppose

(a) X and Yare F spaces,

(b) A: X ~ Y is linear,

(c) G {(x, Ax). x E X} is closed in X x Y.

Then A is continuous.

PROOF X x Y is a vector space if addition and scalar multiplication are defined componentwise:

There are complete invariant metrics dx and dy on X and Y, respectively, which induce their topologies. If

then d is an invariant metric on X x Y which is compatible with Its product

topology and which makes X x Y into an F-space. (The easy bat tedious

verifications that are needed here are left as an exercise.)

Since A is linear, G is a subspace of X x Y. Closed subsets of complete

metric spaces are complete. Therefore G is an F-space.

Define 1! 1 : G ~ X and 1! 2: X x Y ~ Y by

Now 1!1 is a continuous linear one-to-one mapping of the F-space G onto the F-space X. It follows from the open mapping theorem that

1!ll:X~G

is continuous. But A = 1!2 0 1!il and 1T2 is continuous. Hence A is continuous.

Remark The crucial hypothesis (c), that G is closed, is often verified in

applications by showing that A satisfies property (e/) belew:

( c' ) It 1 XII} IS a sequence in X such that the mum

x lim x"

and

exist, then y = Ax.

Let us prove that (c') implies (c). Pick a limit point (x, y) of G. Since X x Y is metrizable,

n-+ 00

for some- sequence {Xn}' It follows [!Om the defimtIOn of the prodllct topology

that Xn -+ x and 1\xn -+ y. Hence y Ax, by (c'), and so (x, Y) E G, and G IS

COMPLETENESS 51

closed.

It is jllst as easy to prove that (c) implies (c').

Bilinear- Mappings

, -'

2.16 Definitions Suppose X, Y, Z are vector spaces and B maps X x Y into Z. Associate to each x E X and to each Y E Y the mappings

n.. Y -+Z and BY: X-+Z

by defining

BAy) B(x, y) BY(x).

B is said to be bilinear if every B; and every BY are linear

If X, Y, Z are topological vector spaces and if 'every Bx and every BY is con-

situatIOns, the converse can be proved '.vith the aid of the Banach-Stemhaus theorem

product topology of X x Y) then B is ObVIOusly separately continuous. In certain

tinuous, then B is said to be sevarately continuous. If B is continuous (relative to the

2.17 Theorem Suppose B: X x Y -+ Z is bilinear and separately continuous, X is an F-space, and Y and Z are topological vector spaces. Then

(1)

whenever Xn -+ Xo in X and Yn -+ Yo in Y. If Y is metrizable, it follows that B is continuous.

PROOF Let U and W be neighborhoods of 0 m Z such that U + U c W. Define

ex E X, 11 1, 2, 3, ... ).

Since B is continuous as a fUBction of y,

n->

(x E X).

implies that {bn} is equicontinuous, Hence there is a neighborhood V of 0 in X such that

tinuous linear mapping of the F-space Y, the Banach Steinhaus theolem 2.6

Thus {bn(xU IS a bOllnded subset of Z, for each x E X. Since each bn IS a con-

(n = 1,2,3, ... ).

Note that

52 GENERAL THEORY

If n is sufficiently la [ge, then (i) xn E Xo + V, so that bn(xn xo) E U, and (ii)

B(xo' VII Yo) E V, since B is continuous in y and B(xo, 0) O. Hence

B(xn, y,J 8(xo , Yo) E U + U c: W

for all large n. This gives (1).

If Y is metrizable, so is X x Y, and the continuity of B then follows

from (1). (See Appendix A6.) 1111

Exercises

1 If X is an infinite-dimensional topological vector space which is the union of countably many finite-dimensional subspaces, prove that X is of the first category in itself. Prove that theIefore nO infinite dimensional F space has a countable Hamel basis.

subset of X. Al temativ ely , f3 is a Hamel basis if every x E K has a unique representatIOn

(A set f3 is a Hamel basis for a vector space X iff$ is a maximal linearly independent

2 Sets of first and second categOIY ate" small" and" large" in a topological sense. These

as a finite linear combination of elements of t3.)

notions are different when" small" and" large " are understood in the sense of measure,

even when the measure is intimately Ielated to the topology. To see this, construct a

subset of the unit interval which is of the first category but whose Lebesgue measme is 1.

3 Put K = [-1, 1]; define !!1J K as in Section 1.46 (with R in place of R·). Suppose {I.} is a sequence of Lebesgue integrable functions such that

1

A1> = !~ L/·(t)1>(t)dt

exists for every 1> E!!1J K. Show that A is a continuous linear functional on!!1J K· Show that there is a positive integer p and a number M < CIJ such that

foI alln. For example, if /.(!) n3 t on [ 1 !n, l,ln] and 0 elsewhere, show that this can be

done with p 1. Construct an example where it can be done with p 2 but not with

4 Let 11 and U be the usual Lebesgue spaces on the unit inlet val. Prove that L2 is of the

first categor y in V, in three ways:

(a) Show that {f. S III2 < n} is closed III V but has empty interior.

(b) Put gn - non [0, n: 3], and show that

for every f E L2 but not for every f ELI.

(c) Note that the inclusion map of U into V is continuous but not onto.

Do the same for V' and Lq if p <q.

COMPLETENESS 53

5 Prove results analogous to those of Exercise 4 for the spaces (P, where tP is the Banach

. space of all complex functions x on {D, 1, 2, .. } 'Nhose norm

6 Define the Fourier coefficients len) of a function IE £l(T) (T is the unit circle) by

for all n E Z (the integers): Put

•

A.I = 2: l(k).

k=-n

category.

Prove that {IE peT): Iim,,-.oo -l\cn I exists} is a dense subspace of P(T) of the first

7 Let C(T) be the set of all continuous complex functIOn"s on the unit circle T Suppose

whose Fourier coefficients are

{Y.} (n E Z) is a complex sequence that associates to each IE GET) a fUllction AjE C(T)

(The notation is as in ExerCise 6.) Prove that {y,,} has this multiplier propetty if and

(Aj)"(n) y. An)

(nEZ)

only if there is a complex Borel measure p.. on T such that

(n EZ).

Suggestion: With the supremum norm, C(T) is a Banach space. Apply the closed graph theorem. Then consider the functional

00

1--+ (A/)(1) = 2: Yn len)

-00

8 Define functionals Am on p. (see Exercise 5) by

not converge; use it only for trigonometric polynomials.)

and apply the Riesz representatIOn theorem ([23J, Th 619). (The above series may ,

m

n 1

(m 1,2, 3, ).

PIOve that K IS compact. Compute Ana XII Show that (:A:mx} is bounded for each x E K

Define x. E (2 by x.(n) l/n, xnor 0 if l ¥ n. Let K c (2 consist of 0, Xl, X.z, X3, ••••

but {Am Xm} is not. Convexity can therefore not be omitted from the hypotheses of Theorem 2.9.

Choose c; > 0 so that 2: c; = 1, 2: nco = if). Take x = 2: c; XII' Show that x lies in the closed convex hull of K (by definition, this is the closure of the convex hull) and that {Am X} is not bounded.

Show that the convex hull of K is not closed.

54 GENERAL THEOR"t

9 Suppose A, }', Z are Banach spaces and

is bilinear and continuous. Prove that there exists },1 < ex) such that

IIB(x,y)1I <Mllxllilyll

(XE X,YE Y).

Is completeness needed here?

10 Prove that a bilinear mapping is continuous if it is continuous at the origin (0, 0).

11 Define Btx«, X2 ; y) = (XlY, X2Y). Show that B is a bilinear continuous mapping of R2 x R onto R2 which is not open at (1, 1; 0). Find all points where this B is open.

12 Let X be the nonned space of all real polynomials in one variable, with

IIfil

t

J~ I fEt) I at.

Put B(f, g) fA f(tm(t) at, and sho .... that B is a bilinear functional on X x X which is

separately continuous but is not continuous.

13 Suppose X is a topological vector space which is of tire second category in itself. Let K

be a closed, convex, absorbing subset of X Prove that K contains a neighborhood of O.

Suggestion. Show first that H K (\ ( K) is absorbing. By a categOIy argument, H

has interior Then use

2H=H+H=H-H.

Show that the result is false without convexity of K, even if X = R2. Show that the result is false if X is U topologized by the V-norm (as in Exercise 4).

14 (a) Suppose X and Yare topological vector spaces, {An} is an equicontinuous sequence of linear mappings of X into Y, and C is the set of all x at which {An(x)} is a Cauchy sequence in Y. Prove that C is a closed subspace of X.

(b) Assume, in addition to the hypotheses of (a), that Y is an F-space and that {A. (x)} converges in some dense subset of X. Prove that then

exists for every x E X and that l\ is continuous.

CONVEXITY

, -'

This chapter deals primarily (though not exclusively) with the most important class of topological vector spaces, namely, the 10ca11y convex ones. The highlights, from the theoretical as well as the applied standpoints, are (a) the Hahn-Banach theorems (assuring a supply of continuous linear functionals that is adequate for a highly developed duality theory), (b) the Banach-Alaoglu compactness theorem in dual

various problems in analysis are postponed to Chapter 5.

spaces, and (c) the Krein-Milman theorem about extreme points. .Applications to,

The Hahn Banach TheOI ems

applied to several closely related results. Among these are the dominated extension

The plural IS used here because the term "Hahn-Banach theorem " IS customarily

theorems 3.2 and 3.3 (in which no topology is involved), the separation theorem 3.4, and the continuous extension theorem 3.6. Another separation theorem (which implies 3.4) is stated as Exercise 3.

3.1 Definitions The dual space of a topological vector space X is the vector space X* Whose elements are the continuous linear functionals on X.

56 GENERAL THEORY

Note that addition and scalar multiplication are defined in X* by

(ctf .. )x 0:: • Ax.

It IS clear that these operations do indeed make X* into a vector space.

It will be necessary to use the obvious fact that every complex vector space is

also a real vector space, and it will be convenient to use the folIowmg (temporary) terminology: An additive functional A on a complex vector space X is called reallinear (complex-linear) if A(IXx) = IXAx for every x E X and for every real (complex) scalar IX. Our standing rule that any statement about vector spaces in which no scalar field is mentioned applies to both cases is unaffected by this temporary terminology and is stilI in force.

If u is the real part of a complex-linear functional! on X, then u is real-linear and

(1)

lEx) u(x) iu(ix)

Ex E X)

because Z Re Z 1 Re (lZ) for every Z E C.

Conv'Cfsely, if tt: X ~ R is real linear on a complex vector space X and if f is

defined by (1), a straightforward computation shows that JIS complex-Imear.

Suppose now that .x is a complex topological vector space. The abo vC facts

imply that a complex-linear functional on X is in X* if and only if its real part is

continuOHS, and that every contiUllOlIs real-linear u: X ~ R is the real part of a umque

!EX*.

3.2 Theorem Suppose

(a) M is a subspace of a real vector space X, (b) p: X ~ R satisfies

p(x + y) :s; p(x) + p(y) if x E X, Y E X, t > 0,

and

p(tx) = tp(x)

(c) j: M ~ R is linear and !(x) < p(x) on AI.

Then there exists a linear A. X=f R such that

Ax - f(x}

(x EM)

p( x) < Ax <p(x)

(x E X).

PROOF If },{ =I- X, choose Xl E X, Xl ¢ AI, and define

Ml = {x + tx.: x EM, t E R}.

It is clear that MI is a vector space. Since

fCx) + fCy) = f(x + y) s p(x + y) :s; p(x - Xl) + P(XI + y),

we have

(1)

(x,y EM).

Let IX be the least upper bound of the left side of (1), as x ranges over AI. Then

(2)

(xEM)

(3)

fey) + IX :5: p(y + Xl) Define II on MI by

ft(x + txl) = f(x) + fIX

(y EM).

(4)

(X E M, t E R).

Then I. = f on M, and ft is linear on MI.

Take t > 0, replace X by t -1 X in (2), replace y by t -1 Y in (3), and multiply the resulting inequalities by t. In combination with (4), this proves that

method of trarrsfmtre inductIOn IS; one can use well-ordering, or Zorn's lemma,

or Hausdorff' s maximality theoI em .

of X that contains M and f' IS a lmear functional on M' that extends f and

Let PJl be the collection or all ordered pairs U.!',f'), wheIe lvf' is a subspace

mean that ' c M" and f" !' on M'. By Hausdorff's maximality theorem

there exists a maximal totally ordered subcollection n of 9.

Let <D be the collection of all M' such that (M',!') E O. Then <1> is totally ordered by set inclusion, and the union £if of all members of <1> is therefore a subspace of X. If X EM then X E M' for some M' E <1>; define Ax = f'(x), where!' is the function which occurs in the pair (M',f') E Q.

It is now easy to check that A is well defined on £if, that A is linear, and that A <p. If £if were a proper subspace of X, the first part of the proof would give a further extension of l'i., and this would contradIct the maximality of {)

ThusM ~

Finally, the inequalIty A < p implies that

fill 171/

3.3 Theorem Suppose M is a subspace of a vector space X, p is a seminorm on X, and f is a linear functional on M such that

I[(x) I :5: p(x)

(X EM).

Then f extends to a linear functional A on X that satisfies

[Ax] :5:p(x)

(x EX).

58 GENERAL THEORY

PROOF If the scalar field is R, this is contained in Theorem 3.2, since p now

satisfies pC x) p(x).

Assume that the scalar field is v: Put u - Re.£ By Theorem 3.2 there IS

a real-1m ear U on X such that U u on M and V < P on X. Let A be the

complex-linear funetional on X whose real part is U. The discussion in Section

3.1 implies that A = f on M.

Finally, to every x E X corresponds an a E (2:', I a I = 1, such that aAx =

I Ax I. Hence

[Ax] = A(ax) = U(ax) :::;; p(ax) = p(x).

IIII

Corollary If X is a normed space and Xo E X, there exists A E X* such that

and

[Ax] :::;; Ilxll

for all x E X.

PROOF If Xo - 0, take A - 0 If Xo =1= 0, apply Theorem 3.3, WIth pIx)

IIxll,

M the one-dimensional space generated by Xo , andf{axo) ex Ilxo Ii 011 ]yJ.

fin

J 1/,

Suppose A and Bare disjoint. nonempty, convex sets m a topologIcal

(a) if A is open there exist A. E X* and '}' E R such that

Re Ax < ')I:::;; Re Ay for every x E A andfor every y E B.

(b) If A is compact, B is closed, and X is locally convex, then there exist A E X*, ')11 E R, ')12 E R, such that

ReJ\x < 'Yl < 'Y2 < Re Ay for every x E A and for every y E B.

Note that this is stated without specifying the scalar field; if it is R, then Re A -

,0 course.

PROOF It is enough to prove this for real scalars For if the scalar field is <Z

and the real case has been proved, then there IS a continuous real-linear Ai on

X that gives the requited separation; if A is the unique eomplex linear functional

on X whose real part is At, then A E y* (See Section 3 1 ) Assume real scalars.

(a) FIx ao E A, bo E B. Put Xo - bo ao; put C - A - B + Xo' Then

C is a convex neighborhood of 0 in X. Let p be the Minkowski functional --. of C. By Theorem 1.35, p satisfies hypothesis (b) of Theorem 3.2. Since

A n B = 0, Xo ¢ C, and so p(xo) ~ 1.

Define f(txo) = t on the subspace M of X generated by Xo. If t ~ 0 then

f(txo) = t s; tp(xo) = p(txo);

CONVEX

if t < 0 thenj(txo) < 0 <p(txo). Thus/< p on M. By Theorem 3.2, /extends

:;n.;:~ t\ > 1 en ~ C, so that 11\1 < I on the neighborhood C ( \ ( ) 0 .

to a linear funotionalA on X that also satisfies A < p. In particular: A "c! o~ i.

eorem 1.18, A E ¥*.

If now a E A and b E B, we have

.: Aa - Ab + 1 = A(a - b + xo) ~ pea - b + xo) < 1 since Axo = 1, a - b + Xo E C, and C is open. Thus Aa < Ab.

It follows that i\(A) and A(B) are disjoint convex subsets of R, with A(A) to the left of A(B). Also, A(A) is an open set since A is open and since every nonconstant linear functional on X is an open mapping. Let y be the right end point of A(A) to get the conclusion of part (a).

(b) By Theorem l.10 there is a convex neighboIilOod P of 0 m X such

exists A E X* such that A(A + V) and A(lJ) are dIs/omt convex subsets of R,

that (A + V) A B 0. Par t (a), with A + V in' place of A, shows that there

CoroUary If X is a locally convex space then ;f* S€pm·ates point,., on X.

PROOF If Xl E X, x2 E X, and Xl =1= x2, apply (b) of Theorem 3.4 with A

{Xl}' B = {x2}· IIII

3.5 Theorem Suppose M is a subspace of a locally convex space X, and Xo EX. If Xo is not in the closure of M, then there exists A E X* such that Axo = 1 but Ax = 0 for every x E M.

PROOF By (b) of Theorem 3.4, with A = {xo} and B = M, there exists A E X* such that Ax, and A(M) arc disjoint. Thus A(~ is a:oper ~~spa,: of the,

obtained by dividing A by Axo.

for every continuous linear functIOnal A on X that vanishes on lvf.

some subspace M of X It suffices (if X is locally convex) to show that Axo 0

approxImation problems: In 9Tder to prove that an Xo E X lies in the closure of

Remark This theot em is the baSIS of a standard method of tI eating certain

3.6 Theorem If f is a continuous linear functional on a subspace M of a locally convex space X, then there exists A E X* such that A = f on M.

Remark For normed spaces this is an immediate corollary of Theorem 3.3. The general case could also be obtained from 3.3, by relating the continuity

60 GENERAL THEORY

of linear functionals to seminorms (see Exercise 8, Chapter 1). The proof

given below shows that Theorem 36 depends only on the separatIOn property

of 1 heorem 3.5.

PROOF Assume, without loss of generality, that f is not identically 0 on M.

}"10 = {x E M:f(x) = O}

and pick Xo E M such that f(xo) = I. Since f is continuous, Xo is not in the M-closure of Mo, and since M inherits its topology from X, it follows that Xo is not in the X-closure of Mo·

Theorem 3.5 therefore assures the existence of a A E X* such that

Axo = I and A = 0 on Mo·

If x E M, then x f(x)xo E Mo, since f(xo) I. Hence

Ax I(X) Ax f(x)Axo A(x f(x)xo) O.

Thus A

f on M.

J

'''Ie conclude this discussion with another useful corollary of the separation

m.

3.7 Theorem Suppose B is a convex, balanced, closed set in a locally convex space X, Xo E X, but Xo ¢ B. Then there exists A E X* such that [Ax] ~ 1 for all

x E B, but Axo > 1.

PROOF Apply (b) of Theorem 3.4, with A = {xo}, note that A(B) is convex and balanced, and multiply the corresponding A by an appropriate scalar. / / / /

'''leak Topologies

3.8 Topological preliminaries The purpose of this section is to explam and

illustrate some of the phenomena that occur when a set is topologized in several ways.

than Xl [Note that (in accordance WIth the meanmg of the inclusion symbol c) the_,

trOpen set is also 't'z open. Then we say that 'tt is weaker than 1'2' or that 't"2 IS stronger

Let "t'1 and "t' 2 be two topologies on a set X, and assume 't'l C T2; that is, every

terms "weaker" and "stronger" do not exclude equality.] In this situation, the identity mapping on X is continuous from eX, 1'2) to (X, 't"1) and is an open mapping

from eX, 1'1) to (X, 't"2)'

As a first illustration, let us prove that the topology of a compact Hausdorff

space has a certain rigidity, in the sense that it cannot be weakened without losing the Hausdorff separation axiom and cannot be strengthened without losing compactness:

.(a) If It C' Xl are topok}gies all a set X, if'l IS a Rausdorff topology, and if 1'2 is

compact, then 't'

CONVEXITY 61

'1 C '2, It follows that F is xrcompact. (Every .1-open cover of f IS also a '2-open

To see this, let Fc.X be 'trcIosed. Smce X IS '2-compact, so is F. Since

cover.) Since '1 is a Hausdorff topology, it follows that F is 'rcIosed.

As another)l1.ustration, consider the quotient topology 't'N of XIN, as defined in Section 1.40, and the quotient map n: X ~ XI N. By its very definition, 't' N is the strongest topology on XI N that makes tt continuous, and it is the weakest one that makes tt an open mapping. Explicitly, if " and r" are topologies on XI N, and if n is continuous relative to " and open relative to r", then r c 'N C r".

Suppose next that X is a set and !F is a nonempty family of mappings f: X ~ Yf, where each Yf is a topological space. (In man im ortan

that property, then , c I' This T is cal1ed the weak topology on .x mduced by !F, or,

more succinctly, the :F -topology of X.

one topologIzes the cartesian product X of a collection of topologIcal spaces Xa'

,: The best-known example of this situation is undoubtedly the lIsual way in which

If nix) denotes the «th coordinate of a point x E X, then na. maps X onto Xa, and the product topology, of X is, by definition, its {na.}-topology, the weakest one that makes every na. continuous. Assume now that every Xa. is a compact Hausdorff space. Then 't' is a compact topology on X (by Tychonoff's theorem), and proposition (a) implies that, cannot be strengthened without spoiling Tychonoff's theorem.

In the last sentence a special case of the following proposition was tacitly used:

(b) 1/!F is a/amity of mappinqs f; X ~ Yf, where X is a set and each Yf is a Hausdorff space, and if/F separates pOints on X, then the !lP -topology of X is a Haus-

dorff topology.

Here is an application of these ideas to a metrization theorem.

(c) If X ts.a compact topological space and if some sequence {f,,} 0/ continuous realvalued functions separates points on X, then X is metrizable.

Let r be the given topology of X. Suppose, without loss of generality, that If" I :::; 1 for all n, and let 't'd be the topology induced on X by the metric

C1:)

d(p, q) = L 2-nlfn(p) - f,,(q)l.

n=1

62 GENERAL THEORY

This is indeed a metric, since {};.} separates points Since each};. is r-continuous and the

series converges uniformly on X x X, d is a r-continuous function on X x X. The

g(P) - {q E X : d( p, q) < r}

are therefore r-open. Thus Id C I. Since rd is induced by a metric, Id is a Hausdorff topology, and now (a) implies that I = rd'

The following lemma has applications in the study of vector topologies. In fact, the case n = 1 was needed (and proved) at the end of Theorem 3.6.

3.9 Lemma Suppose Al, ... , An and A are linear functionals on a vector space X. Let

The following three properties are then equivalent:

(a) There are scalars (Xl' •.• , an such that

(b) There exists y < <Xl such that

IAxl ~ y max IAixl

1:5: i:$n

(x EX).

(c) Ax = O/or every x EN.

PROOF It is clear that (a) implies (b) and that (b) implies (c). Assume (c) holds. Let (J> be the scalar field. Define n: X --» (J>n by

n(x) = (A1X, ... , An x).

If n(x) n(x') then (c) implies Ax Ax' . Hence A Fen, for some function

Fan <I>n This F is a linear fllnctional on <Dn Hence there exist rt.i E <D such that

us

n

i= 1

which is (a).

IIII

3.10 Theorem Suppose X is a vector space and X' is a separating vector space of linear functionals on X. '[hen the X' -topology r' makes X into a locally convex space whose dual space is X'.

· The assumptIons on X' are, more explicitly, that X' is closed under addition and

distinct points of ;[

scalar mUltiplication and that AXl ¥ AX2 for some A E ;[' whenever Xl and X2 are

PROOF Since Rand C' are Hausdorff spaces, (b) of Section 3.8 shows that r is a Hausdorff topology. The linearity of the members of X' shows that r is translation-invariant. If A1, " . ,An E X', if ri > 0, and if

(1)

V={x: IAiXI <r, for 1 S;isn},

then V is convex, balanced, and V E t '. In fact, the collection of all V of the form (1) is a local base for ,'. Thus " is a 10ca11y convex topology on X. If (1) holds, then I V + } V V. This proves that addition is contInUOus.

and y x E r V then

Suppose x E X and a is a scalar. Then x E .'tV for some 8 > O. If I fJ a I < I

fly ax (f3 a)}' I ct(y x)

lies in V, pro' v ided that r is so small that

Hence scalar multiplication is continuous.

We have now proved that r is a 10caIIy convex vector topology. Every A E X' is ,'-continuous. Conversely, suppose A is a ,'-continuous linear functional on X. Then I Ax I < 1 for all x in some set V of the form (1). Condition (b) of Lemma 3.9 therefore holds; hence so does (a): A = ~)iAi' Since Ai E X' and X' is a vector space, A E X'. This completes the proof. fill

and the separating family of seminorms PA(A E X') given by p,,(x) IAx/.

Note: The first pal t of this proof could have been based on Theorem 1..37

(We know that this happens In every 10caIly convex X. It also happens in some others;

topologIcal vector space (with topology -r) Whose dual X* separates POInts on X.

3.11 The weak topology of a topological vector space Suppose X is a

see Exercise 5.) The X*-topology of X is called the weak topology of X. We shall let Xw denote X topologized by this weak topology 'w' Theorem 3.10 implies that Xw is a locally convex space whose dual is also X*.

Since every A E X* is t-continuous and since 'w is the weakest topology on X with that property, we have 'w c: r. In this context, the given topology, will often be called the original topology of X.

to make it clear ',!lith respect to which topology these terms are to be understood.1

original closure, weak closure; oIiginally bO\1nded, weakly bounded, etc., will be used

Self-explanatory expreSSIOns such as ongmal neighborhood, weak neighbOl hood,

For mstance, let {xnf be a sequence m X. To say that Xn --)c 0 originally means

that every original neighbOlhood of 0 contains all Xn with sufficiently large n. To say

that x; --)c 0 weakly means that every weak neighborhood of 0 contains all x; with sufficiently large n. Since every weak neighborhood of 0 contains a neighborhood of the form

(1)

where Ai E X* and r,'> 0, it is easy to see that x, --)c 0 weakly if and only if' tvx; --)c 0 for every A E X*.

Hence every originally convergent sequence converges weakly. (The converse

IS usual; false; see Exercises 5 and 6.) .

Si ilarly, a set E c X is weakly bounded (that is, E is a bounded subset of X,:) if

and only if every Vas in (I) contains tE for some t te V) > O. This happens if and

only 11 there corresponds to each A E X* a number yCA) < CfJ such that IAxl < y(.A.)

fOI every x E E. In other words, a set E G X is weakly bounded if and only if every

A E 1'* is a bounded function on E.

Let V again be as in (1), and put

N = {x: Alx = ... = Anx = O}.

Since x --)c (Alx, ... , An X) maps Xinto en, with nullspace N, we see that dim X::;; n + dim N. Since N c V, this leads to the following conclusion.

If X is infinite-dimensional then every weak neighborhood of 0 contains an infinitedimensional subspace; hence Xw is not locally bounded.

This implies in many cases that the weak topology is strietly v,reaker than the

original one. Of course, the two may coincide: Theorem 3 10 impJi es that (X .. ).. Xw .

We now come to a more interestmg result.

3.12 Theorem Suppose E is a convex subset of {1 locally convex space X. Then

the weak closure Ew olE is equal to its original closure E

1 When X is a Frechet space (hence, in particular, when X is a Banach space) the original topology of X is usually called its strong topology. In that context, the terms" strong" and" strongly" will be used in place of .. original" and" originally." For locally convex spaces in general, the term .. strong topology" has been given a specific technical meaning. See [15], pp. 256-268; also [14], p. 104. It seems therefore advisable to use" original" in the present general discussion.

the opposite inclusion, choose Xo E X, XO 1: E. Par t (b) of the separation theorem

FROOF £.;... is weakly closed, hence origmally closed, so that E c: E". To obtain

3.4 shows that there exist A E X* and'}' E R such that, for every x E E,

ReAxo < y < Refix.

The" set {x: Re Ax < y} is therefore a weak neighborhood of Xo that does not intersect E. TP.lI.s Xo is not in e; This proves t; c: E. / / / /

Corollaries . Let X be a locally convex space.

(a) A subspace of X is originally closed If and only if it is weakly closed.

(b) A convex subset of X is originally dense if and only if it is weakly dense.

The proofs are obvious. Here is another notevt'orthy consequence of Theorem

in X that converges weakly to some x E X, then there is a sequence {yJ in X such that

3.13 Theorem SUPP9se X is a metrizable locally convex space. If {xn} is a sequence

(a) each Yi is a connex combination ofjinitely many Xn, and

(If) Yi -7 X orzgmaUy.

Conclusion (a) says, more explicitly, that there exist numbers !Xin :::::: 0, such that

co

L!Xin = 1,

n=1

co

Yi = L !Xinxn' n=1

and, for each i, only finitely many !Xin are =f= O.

PROOF Let H be the convex hull of the set of all Xn; let K be the weak closure of

H. Then x E K. By Theorem 3.12, x IS also In the ongmal closure of H. Since

the original topology of X is assumed to be mettizable, it follows that there is a,

sequence {Yi} in H that converges originally to x. 70'/1

To get a feeling for 'Nhat is involved hel e, consider the following example.

sufficiently interesting one), and as;ume that t and t~ (n 1,2, 3, ... ) are continuolls

Let K be a compact Hausdorff space (the unit interval on the real line is a

complex functions on K such that fn(x) -+ fex) for every x E K, as n -+ 00, and such that Ifnex) I ~ 1 for all n and all x E K. Theorem 3.13 asserts that there are convex combinations of the I, that converge uniformly to f'

To see this, let C(K) be the Banach space of all complex continuous functions on K, normed by the supremum. Then strong convergence is the same as uniform convergence on K. If Jl is any complex Borel measure on .K, Lebesgue's dominated

,

66 GENERAL THEORY

convergence theorem implies that J1~ dg -+ J t 'dii. Hence Tn -+ f weakly, by the Riesz

representation theorem which identiftes the dual of C(K) 'vVith the space of all regular

complex Borel measures on K. Now Theorem 3 13 can be applied

After this short detour we now return to our main line of development.

3.14 The weak*-topology of a dual space Let X again be a topological vector space whose dual is X*. For the definitions that follow, it is irrelevant whether X* separates points on X or not. The important observation to make is that every x E X induces a linear functional fx on X*, defined by

and that {Ix: x EX} sepUlates pointS 011 X*.

The linearity of each Ix is obvious; if fx A fx i\' for all x E X, then Ax A' x

for all x. and so A A' by the very defillltIOn of what it means for two functions to be

equa.

We are now in the situation described by Theorem 3 10, with X* in place of X

and with X in place of X'.

The X-topology of X* is called the weak*-topology of X* (pronunciation: weak

star topology).

Theorem 3.10 implies that this is a locally convex vector topology on X* and that every linear functional on X* that is weak*-continuous has the form A -+ Ax for some x E X.

The weak*-topologies have a very important compactness property to which we now turn our attention. Various pathological features of the weak- a~p weak*topologies are described in Exercises 9 and 10.

Compact Convex Sets

3.15 The Banach-Alaoglu theorem It V IS a neighborhood of 0 in a topological

vector space X and if

K {A E X*: IAxl < 1 for every x E V}

then K is weak*-compact.

Note: K is sometimes called the polar of V. It is clear that K is convex and balanced, because this is true of the unit disc in e (and of the interval [-1, 1] in R). There is some redundancy in the definition of K, since every linear functional on X that is bounded on V .is continuous, hence is in X*.

X E X a nnmbel rex) < IX) such that x E y(x) V Hence

PROOF SIllce neighborhoods of 0 are absorbing, thele corresponds to each

(1)

I Axl < y(x)

(x E X, li E K).

Let Dx be the set of alI scalars a such that I rx I < y(x) I et r be the product

topology on P, the cartesian product of all Dx, one for each x E X. Since each Ii; is compact, so is P, by Tychonoff's theorem. The elements of P are the functions I on X (linear or not) that satisfy

(2)

I/(x) I ::; rex)

(x EX).

Thus K c X* n P. It follows that K inherits two topologies: one from X* (its weak*-topology, to which the conclusion of the theorem refers) and the other, r, from P. We will see that

(b) K is a closed subset of P.

(a) these two topologies coincide on K, and

K I wea *-compact.

Fix some Ao E K. Choose Xi E X, for I < j < n; choose ~ > O. Put

(3)

and

(4)

W2 = {/EP: I/(x;) - Aoxd < b for 1 s i < n}.

Let n, Xi' and b range over all admissible values. The resulting sets Wi then form a local base for the weak*-topology of X* at Ao and the sets W2 form a local base for the product topology. of P at Ao. Since K c P n X*, we have

This proves (a).

Next, suppose fo is in the .-closure of K Choose x E X, Y E X, scalaI S

lis linear, we have

fO(rxx + .By) - aloex) - /3I000 - (/0 f)(rxx + fly) + !XCf - fo)(x) + flU - io)(Y),

so that

Ifo(ax + Py) - alo(x) - Plo(Y)I < (l + lal + IPI)e.

Since B was arbitrary, we see that 10 is linear. Finally, if x E V and e> 0, the same argument shows that there is an IE K such that I/(x) - lo(x) I < B.

68 GENERAL THEORY

Since IlEx) I < 1, by the definition of K, it follows that Ifo(X) I < 1 We conclude

that Io E K. This proves (b) and hence the theorem. /iii

When X is separable (i.e., when there is a countable dense set in X) then the

conclusion of the Banach-Alaoglu theorem can be strengthened by combining it with the following fact:

3.16 Theorem If X is a separable topological vector space, if K c X* and if K is weak*-compact, then K is metrizable, in the weak*-topology.

Warning: It does not follow that X* itself is metrizable in its weak*-topology. In fact, this is false whenever X is an infinite-dimensional Banach space. See Exercise 15.

PROOF Let {xn} be a countable dense set in X Put .,t;,(t\) Ax,,, for /\ E X*.

Each j~ IS weak*-contmuous, by the defimtlOn of the weak*-topology. If

inCA) J~(A') fOI all II, then AXn A'xn fOI all 11, which implies that A IV,

since both are continuous on X and coincide on a dense set

Thus {t~} IS a countable family of continuous functions that separates

points on X*. The metlizability of Know follows from (c) of Section 3.8. HN

3.17 Theorem If V is a neighborhood of 0 in a separable topological vector space X, and if {An} is a sequence in X* such that

(x E V, n = 1,2,3, ... ),

then there is a subsequence {AnJ and there is a A E X* such that

(x EX).

In other words, the polar of V is sequentially compact in the weak*-topology.

PROOF Combme Theorems 3.15 and 3.16.

71//

The next application of the Banach Alaoglu theorem involves the Hahn-

Banach theorem and a category argument

3.18 Theorem In a locally convex space X, every weakly bounded set is originally :-. bounded, and vice versa.

Part (d) of Exercise 5 shows that the local convexity of X cannot be omitted from the hypotheses. '

CONVEXITY

subset of Xis weakly bounded. The converse IS the nontrivial part of the theorem.

0[0, it is obvious from the definition of" bounded" that evelY originally bounded

PROOF Smce every weak neighborhood of 0 in X is an ongmal neighborhood

o in X.

Suppose E e .¥ is weakly bounded and U is an ongmaI neIghborhood of

- Since X is locally convex, there is a convex, balanced, original neighborhood VofO inX such that Ye U. Let K e X* be the polar of V:

(1)

K = {A E X*: [Ax] :s;: 1 for all x E V}.

We claim that

(2)

Y = {x EX: [Ax] :s;: 1 for all A E K}.

It is clear that V is a subset of the right side of (2) and hence so is ~~ since the

Y , ere corresponds to each A E X* a number

yCA) < if) such that

IAxl < yCA)

(x E E).

Since K is convex and weak*-compact (Theorem 3.15) and since the functions

A ~ Ax are weak*-continuous, we can apply Theorem 2.9 (with X* in place of X and the scalar field in place of Y) to conclude from (3) that there is a constant y < if) such that

(4)

[Ax] ~y

(x EE,A EK).

(5)

Now (2) and (4) show that y - 1 X EYe U for all x E E. Since V is balanced,

Ee tYe tU

(t » y).

Thus E is originally bounded.

Corollary IIX is a nvrmed space, 1/ E e X, and if

sup I Axl < OJ

(Ii: E X*)

then there exists y < en such that

(7)

II:\';/I < y

(x E E).

PROOF Normed spaces are locally convex; (6) says that E is weakly bounded, and (7) says that E is originally bounded. JIJI

The following analogue of part (b) of the separation theorem 3.4 will be used in the proof of the Krein-Milman theorem.

70 GENERAL THEORY

3.19 Theorem Suppose X ts a topological vector space on which X* separates

points. Suppose A and B are disjoint, nonempty, compact, convex sets in X Then

there exi<it<i A E ;y* such that

(1)

sup Re Ax < inf Re Ay.

xeA yeB

Note that part of the hypothesis is weaker than in (b) of Theorem 3.4 (since local convexity of X implies that X* separates points on X); to make up for this, it is now assumed that both A and B are compact.

PROOF Let Xw be X with its weak topology. The sets A and B are evidently compact in Xw. They are also closed in Xw (because Xw is a Hausdorff space). Since Xw is locally convex, (b) of Theorem 3.4 can be applied to Xw in place of

X; it gives us a /\ E (Xw)* that satisfies (1). But we saw in Section 3 11 (as a

consequence of Theorem 3.10) that (XlI'r x*. JIll

3.20 Extreme points Let K be a subset of a vector space X. A nonempty set

S c K is called an extreme set of K if no point of S is an inter nal point of a line interval

whose end points are in K but not in S. Analytically, the condition can be expressed

as follows· If x E K, Y E K, 0 < t < ], and

tx + (l - t)y E S,

then XES and YES.

The extreme points of K are the extreme sets that consist of just one point. Recall that the convex hull of a set E c X is the smallest convex set in X that contains E, and that the closed convex hull of E is the closure of its convex hull.

At present it is not known whether it is true in every topological vector space X that every compact convex set has an extreme point. In a large class of spaces, the supply of extreme points is actually very abundant. This is shown by Theorems

3.21 The Krein-MUrnan theorem Suppose X is a topological vectOl space on

which X* separates points. If K is a compact COfWex set in X, then K is the clMed

C0nllPX hull of the set of its extreme points.

PROOF Let fll' be the collection of all compact extreme sets of K. Since K E 9,

[l}J =f:. 0. We shall use the following two properties of [l}J:

(a) The intersection S of any nonempty subcollection of [l}J is a member of [l}J," unless S = 0.

(b) If S E [l}J, A E X*, J1 is the maximum ofRe A on S, and

SA = {x E S: ReAx = J1},

)' E S. Hence Re Ax < /1, Re Ay < /1. Smce Re J\z tl and A is linear, we

2 ESA, XEK, YEK, 0<1< J Since ZES and SEfY', we have XES and

The' proof of (a) is immediate. To prove (b), suppose tx + (I t)y

conclude' Re Ax J1 Re Ay. Hence x E SA and Y E SA'

Choose some S E 9. Let fllJ' be the collection of all membel s of;?P that are

subsets of S. Since S E 9',9' is not empty. Partially order P/' by set inclusion, let Q be a maximal totally ordered subcollection of 9', and let M be the intersection of al~ members of Q. Since Q is a collection of compact sets with the finite intersection property, M t= 0. By (a), ME 9'. The maximality of Q implies that no proper subset of M belongs to 9. It now follows from (b) that every A E X* is constant on M. Since X* separates points on X, M has only one point. Therefore M is an extreme point of K.

. We ~av~;~ prC~1:::xr:;:f: e:eme set 0/ K contains an

every S E .9, that ff A S is not empty.

If H is the convex hull of the set of extreme points of K, it [oIlO'.>,'s, for

3.19 there is a t\ E X* such that Re Ax < Re Axo for every x E H. If KA is

Assume (to get a eOlltladiction) that some Xo E K is not in H By Theorem

Since K IS compact and convex, we have H c K. I lenee lJ is compact.

defined as in (b), then KA E P/. Since H does not intersect KA we have our contradiction. IIII

Remark The convexity of K was used only to show that H is compact. If X were assumed to be locally convex, the compactness of H would not be needed, since one could use (b) of Theorem 3.4 in place of Theorem 3.19. The above argument then proves that K c H. The following version of the Krein- Milman theorem is thus obtained:

ala compact set K in X, then K lIes m the closed convex hull olE.

3.22 Theorem I} X is a locally convex 8pace and ifE is the set of extreme points,

are situations in which H is not compact (Exelcises 20, 22). In Frechet spaces, the

hull H of a compact set K? Even m a HIlbert space, H need Dot be closed, and there

The precedmg remark raises a question: What can one say about the convex

Equivalently, E and K have the same closed convex hull.

latter pathology does not occur (Theorem 3.25). The proof of this will depend on the fact that a subset of a complete metric space is compact if and only if it is closed and totally bounded. (Appendix A4.)

Let us recall that a subset E of a metric space X is said to be totally bounded if E is contained in the union of finitely many open balls of radius e, for every e > O. The same concept can be defined in any topological vector space X, metrizable or not.

3.23 Definition A set E in a topological vector space X is said to be totally

If X happens to bc a metrizable topological vector space, then these two notions

of total boundedness coincide, provided that we restnct ourselves to Invariant metrics that are compatible with the topology of X. (The proof of this is as in Theorem 1.26.)

3.24 Theorem If X is a locally convex space and H is the convex hull of a totally bounded set E c X, then H is totally bounded.

PROOF Let U be a neighborhood of 0 in X. There is a convex neighborhood V of 0 in X such that V + V c U, and there is a finite set E1 c X such that

E C £1 I V. Let HI be the convex hull of £1

If e~, ... , em are the pomts of E1 and If S is the simplex in Rm consisting

of all t (ll"'" 1m) that satisfy t; > 0 and Iti 1, then

t, e,

maps the compact set S continuously onto lIt. Hence Bt is compact

There are points Yi E E1 such that Xi - Yi E V; this follows from the choice of EI. Decompose x into the sum

x=x' + x",

where x' = LtXiYi and x" = 2)i(Xi - yJ The convexity of V implies that x" E V. It is clear that x' E HI. Hence

Smce V was arbItrary, It follows that H is totally bounded.

fill

WI

3.25 TheOIem Suppose H is the convex hull af a compact set K in a top()logical

oector space X.

(a) If X is a Frechet space, then H is compact. (b) If X = R", then H is compact.

PROOF (a) By Theorem 3.24, H is totally bounded. Since Frechet spaces are complete metri~ spaces, the closure H of H is compact.

with tj > 0 and 2), 1. By the lemma that follows, x E H If and o~ly if

(b) Let S be the simplex In Rn+1 conSlstmg of all t (t1> ... , tn+1)

n

x t·x.

j 1

for some t E S and Xi E K (1 sis n + 1). In other words, H is the image of SxKx···xK

(K occurs n + 1 times) under the continuous mapping

n+l

(t, Xl' •.• , Xn+1) --,} L t i x], j= 1

Hence R IS compact.

1m

. of some r I 1 vectols Xi E E, then X is actually a convex combination of some

PROOF It is enough to show that if r > n and x 2), x, is a convex combination

r of these vectors.

Assume, without loss of generality, that t, > ° for 1 sis r + 1. The r vectors x, - xr+1 (1 sis r) arc linearly dependent, since r > n. It follows that there arc real numbers a., not all 0, such that

r+ 1

L a.x, = 0

j= 1

and

r+ 1

Laj=O.

i= 1

Choose m so that I a jt, I < I «J«; I, for I < is r + 1, and define

Vector-valued Integration

~.

Sometimes it is desirable to be able to integrate functions f that are defined on some measure space Q (with a real or complex measure Ii) and whose values lie in some topological vector space X. The first problem is to associate with these data a vector in X that deserves to be called

J f du,

Q

74 GENERAL THEORY

n

J (Ai) dJ1

Q

ought to hold for every A E X*, because it does hold for sums, and because integrals are (or ought to be) limits of sums in some sense or other. In fact, our definition will be based on this single requirement.

Many other approaches to vector-valued integration have been studied in great detail; in some of these, the integrals are defined more directly as limits of sums (see Exercise 23).

3.26 Definition Suppose fl is a measure on a measure space Q, X is a topological

vector space on which X* separates points, andfis a function from Q into X such that

(A f)(q) A(f(q))

(q E Q).

If there eXIsts a vector y E X such that

(2)

Ay = f (AI) dii

Q

for every A E X*, then we define

(3)

f f du. = y.

Q

Remarks It is clear that there is at most one such y, because X* separates points on X. Thus there is no uniqueness problem.

Existence will be proved only in the rattIeI special case (sufficient for many

applieations) in which Q is compact andfis continuous. In that case,f(Q) is compact,

and the only other requirement that will be imposed IS that the closed convex hull of

{( Q) should be compact. By Theorem 3.25, this additional requirement is automatically

satisfied when X is a Fr6chet space.

Recal1 that a Borel measure on a compact (or locally compact) Hausdorff

space Q is a measure defined on the O"-algebra of all BOle} sets in Q, this is the smallest

e-algebra that contains all open subsets of Q. A probability measure is a positive. measure of total mass 1.

3.27 Theorem Suppose

(a) X is a topoloqical, vector space on which X* separates points, and (b) J1 is a Borel probability measure on a compact Hausdorff space Q.

CONVEXITY 75

exists, jn the sense ~f Definition 3.26.

Mo'reover, Y E H. : -;

Remark If. v is any positive Borel measure on Q, then some scalar multiple of v is a probability measure. The theorem therefore holds (except for its last sentence) with v in place of /1. It can then be extended to real-valued Borel measures (by the Jordan decomposition theorem) and (if the scalar field of X is (2:') to complex ones.

Exercise 24 gives another generalization.

Y E H such that

PROOF Regard X as a real vector space. Vie have to prove that there exists

(2)

Ay - ~ (AI) dp

Let L - {AJ, ... , An} be a finite subset of X*. Let EL be the set of all Y E H that satisfy (2) for every A E L. Each EL is closed (by the continuity of A) and is therefore compact, since H is compact. If no EL is empty, the collection of all EL has the finite intersection property. The intersection of all EL is therefore not empty, and any y in it satisfies (2) for every A E X*. It is therefore enough to prove EL =1= 0.

Regard L = (A1, "', An) as a mapping from X into R", and put K = L(f(Q»). Define

(3)

m· - ] (A.f) d~l

(l < i < 12).

We claim that the point m (m!> ... , mn) lies in the convex huH of K.

If t (11, .. " tn) E R" IS not in this hull, then [by TheoIem 3.25 and

are real numbers c1, , en such that

(Ii) of Theorem 3.4 and the known form of the lInear fUDctionais on RnJ there

(4)

n n

I c, u, < I c, t,

i==1 i= 1

if U = (ut, "', un) E K. Hence

(5)

n n

I CiAif(q) < I c.t,

j==1 i=1

(q E Q).

Smce J.l is a probability measure, integration of the left side of (5) gIves

This shows that m lies in the convex hull of K. Smce K L ( f(Q)) and L

It follows that m Ly for some y in the convex lrulll1 of f(Q). For

(6)

Aiy = m, = f (AJ) dii Q

Hence y E EL. This completes the proof.

(1 ~ i ~ n).

IIII

3.28 Theorem Suppose

(a) X is a topological vector space on which X* separates points,

(b) Q is a cvmpact subset of X, and

(c) the closed convex hull H of Q is compact

Then y E H 11 and only if there is a regular EOI el plobability measure Jl on Q

(1)

Remarks The integral is to be understood as in Definition 3.26, withf(x) = x.

Recall that a positive Borel measure on Q is said to be regular if

(2)

}leE) = sup {}l(K): K c E} = inf {}leG): E c G}

for every Borel set E c Q, where K ranges over the compact subsets of E and G ranges over the open supersets of E.

The integral (1) represents every Y E H as a "weighted average" of Q,

or as the" center of mass" of a certain unit-mass distributed over Q. ,

We stress once more that ee) follows from (b) if OX is a Frechet space.

PROOF Regard X again as a real vector space. Let C(Q) be the Banach space

of all real continuolls functions on Q, with the supremum norm. The Riesz

representation theorem identifies the dual space C(Q)* with the spaee of all

real Borel measures on Q that are differences of regular positive ones With this

identification in mind, we define a mappmg

(3) by

if>: C(Q)* -+ X

(4)

if>(Jl) = f x dJl(x).

Q

CONVEXITY 77

theorem asser ts that p(P) H.

Let P be the set of all regular BOlel probabIhty measures on Q. The

that He ¢(P), where R IS the convex hull of Q. By Theorem 3.27, p(P) c R.

¢(i5x) x, we see that Q c ¢(P). Since ¢ is linear and P IS convex, it follows

For each x E Q, the unit mass <5r concentrated at x belongs to P. Since

Therefore all that remains to be done is to show that ¢(P) is closed in X.

This isa consequence of the following two facts:

(i) Pis weak*-compact in C(Q)*.

(ii) The mapping ¢ 'defined by (4) is continuous if C(Q)* is given its weak*topology and if X is given its weak topology.

Once we have (i) and (if), it follows that ¢(P) is weakly compact, hence weakly closed, and since weakly closed sets ate strongly closed, we have the

desired conclusion.

To prove (i), note that

(5)

-and that this larger set IS weak*-compact, by the Banach-Alaoglu theorem.

It is therefore enough to show that P is weak*-c1osed.

If hE CEQ) and h ~ 0, put

(6) e, = (P: fQh dp ~ 0).

Since p -> J h dp is continuous, by the definition of the weak*-topology, each Eh is weak*-closed. So is the set

(7)

Since P is the intersection of E and the sets Eh, P is 'Neak*-c1osed.

(8)

since ismear. Every weak neighborhood of 0 in X con tams a set of the form

where Ai E A: * and ri > O. The restrictions of the Ai to Q lie III C( Q). Hence

W {y E X: /Aiy/ < ri for 1 < i < n},

v = (Jl E C(Q).: I Jo Ai dJl/ < r, for 1,,; i ,,; n) is a weak*-neighborhood of 0 in C(Q)*. But

(9)

(10)

78 GENERAL THEORY

by Definition 3.26. It follows from (8), (9), and (10) that 4>(V) c WHence

¢ IS contmuous. i iii

The following simple inequalIty sharpens the last assertion III the statement of

3.29 Theorem Suppose Q is a compact Hausdorff space, X is a Banach space, I: Q -? X is continuous, and p, is a positive Borel measure on Q. Then

PROOF. Put Y = J I dp,. By the corollary to Theorem 3.3, there is a A E X* such that l\.y Ilyil and IAxl < Ilxll for all x e X. In particular,

IAt(s)1 < II t(s)1I

for all SEQ. By Theorem 3.27, it follows that

Ilyll /\y /Q(AD dl1 < JQllill djJ.

1//1

11/1

Holomorphic Functions

In the study of Banach algebras, as well as in some other contexts, it is useful to enlarge the concept of holomorphic function from complex-valued ones to vectorvalued ones. (Of course, one can also generalize the domains, by going from e to en and even beyond. But this is another story.) There are at least two very natural definitions of" holomorphic " available in this general setting, a "weak" one and a " strong" one. They turn out to define the same class of functions if the values are assumed to lie in a Frechet space.

3.30 Definition let 0 be an open set in e: and let X be a complex topologIcal

vector s ace.

(a) A functioni' n -? :ris said to be weakly hoZomorphic in n if AI IS holomorphlc

III the ordinary sense for every A E X *.

(b) A function f: Q --4 X is said to be strongly holomorphic in Q if

1. few) - fez)

1m _~--C...-_

w .... z

w-z

exists (in the topology of X) for every ZEn.

Note that the above quotient is the product of the scalar (w - Z)-1 and the vector I(w) - I(z) in X:

CONVEXITY

important role in this proof, as wi1I Theorem 3.18.

quences may very well fail to converge original1y.) The Cauchy theorem will play an

is a Fnkhet space, but it is faJ from ObVIOUS. (Recall that weakly convergent se-

strongly huloIIlOrphic functIOn IS weakly holomorphic. The conveJse is true when X

. Ihe confinuity of the functionaIs A that occur in (a) makes It obvious that every

The index of a point Z E (l with respect to a closed path r that does not pass through z will be denoted by Ind.. (z). We recall that

1 f. de

Ind,(z) =-. __ .

- 2m, e - z

3.31 Theorem Let n be open in (2', let X be a complex Frechet space, and assume that

i8 v;eakiy holomorphir; The/ollowlng conclusion» hold:

(a) t IS strongly continuous in n.

(b) The Cauchy tlteorem and the Cauchy formula hold: If F is a dosed path in n

;such that Ind, (tv) 8/or every W 1= n, then

(1)

I, f(C) d( = 0,

and

(2)

f(z) = ~ f. (( - z)-11(0 s;

2m ,

if ZEn and Ind.. (z) = 1. If r1 and r 2 are dosed paths in 0 such that Indr1 (w) = Indr2 (w)

for every w 1: n, then

(3)

(c) t is strongly h%morphic in n.

The integrals in (b) are to he understood in the sense of Theorem 3.27. Either

one can regard d~ as a complex measure OIl the range of r (a compact subset of (2'), or one can parame1nze r and integrate with respect to Lebesgue measure on a compact

interval in R.

PROOF (a) Assume 0 EO. We shall prove thatfis strongly continuous at O. Define

(4) t1r = {z E eo: [z] :::; r}.

Then A2r C fl fOI some r > O. Let r be the positively oriented boundary of

FIX A E X*. Smce Aj IS liolomorphic,

(5)

z 2rri Jr (( - z)(

if 0 < JzJ < 2r. Let M(A) be the maximum of JA/J on Lllr. If 0 < JzJ $; r, it follows that

(Af)(z) (Af)(O) 1 r (Af)(O d(

(6)

The set of all quotients

(7)

IS therefore weakly bounded in X. By Theorem 3 18, this set is also strongly

(8)

fez) flO) E ztP

(0 < J zJ < r).

Consequently, I(z) ~ 1(0) strongly, as z ~ O.

This was the crux of the matter. The rest is now almost automatic.

(b). By (a) and Theorem 3.27, the integrals in (1) to (3) exist. These three formulas are correct (by the theory of ordinary holomorphic functions) if I is replaced in them by AI, where A is any member of X*. The formulas are therefore correct as stated, by Definition 3.26.

(c) Assume, as in the proof of (a), that Lllr C n, and choose r as in (a).

Define

(9)

The Cauchy formula (2) shows, aft er a small computation, that

(10)

f(i) f(O)

z

y + zg(z)

.If 0 < JzJ < 2r, where

g(z) = ~ r [2rei8(2rei8 - z)r1f(2rei8) so. 2rr -It

(11)

.

Let V be a, convex balanced neighborhood of 0 in X. Put K =

{/(O: J' J = 2r}. Then K is compact, so that K c t V for some t < 00. If s =

y(z) E S V if I z I < r. I he left side of (10) therefore converges strong to y, a

The following extension of LIouvIlle's theorem concerning bounded entire functions does not even depend on Theorem 3.31. It can be used in the study of spectra in Banach algebras. (See Exercise 4, Chapter 10.)

3.32 Theorem Suppose! is a complex topological vector space on which X* separates points. Suppose f: e ~ X is weakly holomorphir and f( C) is a weakly bounded subset of X. Then f is constant.

PROOF For every A E X*, AI is a bounded (complex-valued) entire functIOn.

If Z E e, it follows from LIOuville's theorem that

fill

m7

bounded, in an F-space X on which X * separates points. Compare with Theorem 3.18.

~art (d) of Exercise 5 describes a weakly bounded set which is not originally

Exercises

1 Call a set HeR· a hyperplane if there exist real numbers as, ... , an, c (with at :f; 0 for at least one i) such that H consists of all points x = (x., ... , x.) that satisfy .L at XI = c.

Suppose E is a convex set in R", with nonempty interior, and y is a boundary point of E. Prove that there is a hyperplane H such that y E Hand E lies entirely on one side of H. (State the conclusion more precisely.) Suggestion' Suppose 0 is an interior

(if ex ¥ fJ) which cannot be separated by any continuous linear functional A on L2.

each ~ is convex and that each is dense in L2. Thus Ea and Ep are dlslomt convex sets

Ea be the sel of all contInUOUS functions (on [ 1, 1 J such that f(!)) cx. Show that

ffint· What is A(E.)? 2 >

3 Suppose:X IS a real vector space (without topology). Call a point Xo E A c X an internal point of A if A - Xo is an absorbing set.

(a) Suppose A and B are disjoint convex sets in X, and A has an internal point. Prove that there is a nonconstant linear functional A on X such that A(A) n A(B) contains at most one point. (The proof is similar to that of Theorem 3.4.)

(b) Show (with X = R2, for example) that it may not be possible to have A(A) and A(B) disjoint, under the hypotheses of (a).

82. GENERAL THEOR Y

4 Let too be the space of all real bounded functions x on the positive integers. Let T be

the translation operator defined on too by the equation

(TX)(ll) x(n + 1)

en 1,2,3, ... )

Prove that theIe exists a lineal functional A. on loo (called a Banach limit) sueh that

(a) ATx = Ax, and

(b) lim inf x(n) < Ax < lim sup x(n)

n-t 00

n-too

for every x E too.

Suggestion: Define

x(1) + ... + x(n)

Anx=-----n

M = {x E too: lim An X = Ax exists}

n-+oo

and apply Theorem 3.2.

5 For 0 <p < 00, let tp be the space of all functions x (real or complex, as the case may

be) on the positive integers, such that

For 1 <p < 00, define Ilxllp = {L: Ix(n}/ p}l/p, and define Ilxlloo = sup, I x(n)l·

(a) Assume 1 <p < 00. Prove that Ilxllp and Ilxlloo make tp and too into Banach spaces.

If »:' + q-l = 1, prove that (tp)* = tq, in the following sense: There is a one-to-one correspondence A +-+ Y between (tp)* and tq, given by

Ax = 'L x(n)y(n)

(b) Assume 1 <p < 00 and prove that tp contains sequences that converge weakly but not strongly.

(c) On the other hand, prove that every weakly convergent sequence in t1 converges

strongly, in spite of the fact that the weak topology of /1 is different from i,ts strong

topology (which is induced by the norm).

(d) If 0 <p < 1, prove that a; metrized by

00

d(x, v) = 2: I xCn) - v(n) I P,

n=l

IS a locally bounded F-space which IS not locally convex but that (11))* nevertheless

separates points on tp• (Thus there are many convex open sets in tp but not enough

to form a base for its topology.) Show that (tp)* = too, in the. same sense as in (a)." Show also that the set of all x with L I x(n) I < 1 is weakly bounded but not originally bounded.

(e) For 0 <p < 1, let Tp be the weak*-topology induced on too by tp; see (a) and (d).

If 0 <p < r < 1, show that Tp and T, are different topologies (is one weaker than the other?) but that they induce the same topology on each norm-bounded subset of too. Hint: The norm-closed unit ball of too is weak*-compact.

6 Put In(t) e1ilt ( 1T < t < 7T); let LP L PC 77', 77'), with respect to Lebesgue measUi e.

If 1 <p < co, prove that/" >- 0 weakly in D', but not strongly.

7 L"'([O, 1]) has Its norm topology (11/11", is the essential supremum of IfD alid its weak*-

is dense In £00 In one of these topologies but not in the other. (Compare with the

topology as the dual ofD. Show that C, the space of all contmuous functions on [0, 1J,

corollaries to Theorem 3.12.) Show the same with" closed" in place of" dense."

8 Let t be the Banach space of all complex continuous functions on [0, 1], with the supremum nor~: 'Let B be the closed unit ball of C. Show that there exist continuous linear functionals A on C for which A(B) is an open subset of the complex plane; in particular, IAI attains no maximum on B.

9 Let E c L2( -1T, 1T) be the set of all functions

sequence in E converges weakly to g. (El IS called the weak sequential closure of E.)

where m, n are integers .and 0 < m < n. Let lh be the set of all 9 E L2 such that some

(a) Find all g E £.

(b) Find all 9 in the weak closure Ew of E.

(c) Show that 0 FE .. but 0 is not in £1, although 0 lies in the weak sequentIal closure

1·

closed. The passage from a set to its weak sequential closure is therefore not a closure operation, in the sense in which that term is usually used in topology. (See also Exercise 28.)

10 Represent {I as the space of all real functions x on S = {(m, n): m > 1, n > I}, such that

Tbis example shows that a 'Neak sequential closure need not be weakly sequentially

Ilxlll = L Ix(m, n)1 < 00.

Let Co be the space of all real functions y on S such that y(m, n) -+0 as m + n -+ 00, with norm IIYlloo = sup Iy(m, n)l.

Let M be the subspace of {I consisting of all x E (I that satisfy the equations

mx(m, 1) z: x(m, n)

(m 1, 2, 3, ... J.

(a) Prove that /1 (Co)*. (See also Exercise 24, Chapter 4.)

(Ii) Prove that M IS a norm-closed subspace of /1

(d) Let B be the norm-closed nnit baJJ of (I In spite of (e), prove that the weak*-

(c) Prove that Al is weak* dense in t! [xelalive to the weak*-topology gIven by (a)J.

closure of }off n B contains no ball. Suggestwn: If (5 > 0 and m > 2/8, then

Ilxll 8

Ix(m, 1)1 <- <-2 m

if x E M n B, although x(m, I) = 0 for some x E se. Thus oB is not in the weak*closure of M n B. Extend this to balls with other centers.

11 Let X be an infinite-dimensional Frechet space. Prove that X*, with its weak+-topology, is of the first category in itself.

o III the L2-norm. Find such a sequence. Show that gN N lUI I ... +h) will not do

By Theorem 3 13, some sequence of convex comblllatlOns of the iN converges to

14 (a) Suppose 0 is a locally compact Hausdorff space. For each compact K c n define a seminorm PK on C(O), the space of all complex continuous functions on 0, by

PK(f) = sup {J f(x)J : x E K}.

Give C(O) the topology induced by this collection of seminorms. Prove that to every A E C(O)* correspond a compact K c 0 and a complex Borel measure Ik on K

such that

Af= r I dlk

'K

(fE C(O),

(b) Suppose n is an open set in fl Find a countable collection l' of measures with

cO=i:ct su;!:rt III n such that B (0.) (the space of all holonrorphic functions in Q)

eons'sts of actly thosefE C(Q) which satisfy J f t4L 0 for every ME 1'.

15 I,et X be a topological vector space on which A: * separates points. Pr ave that the weak*

topology of X* is metrizable if and only if X hali a finite or countable Hamel baSIS.

16 Prove that the closed unit ball of Ll (relative to Lebesgue measure on the unit interval)

(See Exercise 1, Chapter 2 for the defimtlOn.)

has no extreme points but that every point on the "surface" of the unit ball in LP (1 <p < (0) is an extreme point of the ball.

17 Determine the extreme points of the closed unit ball of C, the space of all continuous functions on the unit interval, with the supremum norm. (The answer depends on the ,

choice of the scalar field.)

18 Let K be the smallest convex set in R3 that contains the points (1, 0,1), (1,0, -1), and

(cos e, sin e, 0), for 0 < (} < 27T. Show that K is compact but that the set of all extreme points of K is not compact. Does such an example exist in R2?

19 Suppose K is a compaet convex set in R'. Prove that every x E K is a cony ex combina-

I

\ \

lion of at most n + 1 extreme POllltS of K. SuggestIOn: Use induction on n. Draw a

line from some extreme point of K through x to where it leaves K lIse Exercise 1.

20 Suppose a topological vector space X contains a countable set E {el' ez , e3 , ... }

WIth the following proper ties.

t~ en ---':>- as n ---':>- 00.

IT Every x E X IS a finite lInear combination of members of E, x 2: Yn(X)en.

(c) No en is in the closed subspace of X generated by the other et:

For example, X could be the space of all complex polynomials

fez) = ao + alZ + .. , + anz',

with norm

. IIIII = {f" J f(ei8)J2 dof/2,

and with e.(z) = n=+z":' (n = 1, 2, 3, ., .).

CONVEXITY

points of I1 ale exactly the points of K.

Prove that the convex hlJJI l{ of K is closed but not compact and that the extreme

Prove that each Yn in (b) is in A: *. Put K E u {O}. Then K is compact

21 If 0 < [) < 1, every IE [ p (except f 0) is the arithmetic mean of two functions whose

distance froll1 0 is less than that of]. (See SectIOn 1.47.) Use th is to construct an explicit

example of a countable compact set K in LV (with 0 as its only limit point) which has no extreme point.

22 If 0 <p < 1,· show that tp contains a compact set K whose convex hull is unbounded.

This happens in spite ·of the fact that (tp)* separates points on /», see Exercise 5. Suggestion: Define Xn E t~ by

x.(n) = nP-1,

x.(m) =0

if m =1= n.

Let K consist of 0, Xl, X2, X3, •••• If

show that {YN} is unbounded in ('1'.

23 Suppose f.L is a Borel probability measure on a compact Hausdorff space Q, X is a

collection of disjoint Borel subsets of Q whose union is Q. PlOve that to every neighbor-

Bechet space, and f: Q -+ A: IS contmuous. A partition of Q is, by definition, a finite

hnod V of 0 in A: there corresponds a partition {E,} such that the difference

. lies in V for every choice of SI EEl. (This exhibits the integral as a strong limit of " Riemann sums.") Suggestion: Take V convex and balanced. If A E X* and if I Ax I < 1 for every X E V, then I Az] < 1, provided that the sets E, are chosen so that J(s) - J(t) E V whenever sand t lie in the same EI•

24 In addition to the hypotheses of Theorem 3.27, assume that T is a continuous linear mapping of X into a topological vector space Y on which y* separates points, and

reT!) dfL·

Hint: ATE X* for every A e Y*.

25 Let E be the set of all extreme points of a compact convex set K in a topological vectOi

Borel probability measure ,I( on Q E such that

spaee X on which X* separates "Points. Prove that to every y E K corresponds a regular

y = t X d{-t(x).

26 Suppose 0 is a region in e, X is a Frechet space, and J: 0 -+ X is holomorphic.

(a) State and prove a theorem concerning the power series representation of f, that is, concerning the formula J(z) = 2: (z - a)ncn, where c« E X.

(b) Generalize Morera's theorem to X-valued holornorphic functions.

(c) For a sequence of complex holomorphic functions in fl, uniform convergence on

compact subsets of Q implies that the limit is holomorpbic Does this generalize to

X-valued hoiomorphlc functIons '?

27 Suppose {!XI} is a bounded set of distinct complex numbers, fez) =,2;go cnzn is an entire function with every en i= 0, and

Prove that the vector space generated by the functions g I is dense in the Frechet space H(fl) defined in Section 1.45.

Suggestion: Assume f1- is a measure with compact support such that J gl dp: = 0

~(w) J f(wz) df1-(z)

(14' E (2')

Prove that crEw) 0 for all w. Deduce that S zn df1-(z) 0 for n 1, 2, 3, IIse

xercise 14.

Deser ibe the closed subspace of lIce) generated by the functions gt if some of

28 Suppose X is a Frechet space (or, more generally, a metrizable locally convex space).

Prove the following statements.

(a) X* is the union of countably many weak*-compact sets En.

(b) If X is separable, each En is metrizable. The weak*-topology of X* is therefore separable, and some countable subset of X* separates points on X. (Compare with Exercise 15.)

(c) If K is a weakly compact subset of X and if Xo E K is a weak limit point of some countable set E c K, then there is a sequence {x.} in E which converges weakly to Xo. Hint: Let Y be the smallest closed subspace of X that contains E. Apply

(b) to Y to conclude that the weak topology of K Po Y is metrizable.

Remark: The point of (c) is the existence of convergent subsequences rather

than saimets. Note that there exist compact Hausdorff spaces in which no sequence

of distinct points converges

29 Let C(K) be the Banach space of all continuous complex functions on the compact

Hausdorff space [(, with the supremum norm For p E K, define Ap E C(K)* by

Al' J J (p). Show that p --'J> Ap is a homeomorphism of K into C(K)*, equipped with

its weak*-topology. Part (c) of Exercise 28 can therefore not be extended to weak*compact sets.

DUALITY IN BANACH SPACES

-' ._"

The N ormed Dual of aN ormed Space

Introduction If X and Yare topological vector spaces, £?6'(X, Y) will denote the collection of all bounded linear mappings (or operators) of X into Y. For simplicity, £?6'(X, X) will be abbreviated to £?6'(X). Each £?6'(X, Y) is itself a vector space, with respect to the usual definitions of addition and scalar mUlti!fication of function;"

there are many ways in which 38(X, Y) can be made into a topological vector space.

(ThIS depends only on the vector space structure of Y, not on hat of X.) In genera "

the scalar field, so that 38(X, Y) IS the dual space X* of 1', the above mentioned

case, M(X, Y) can itselfhe Dormed in a very natural way. \Vh 1 Yis specialize to e

In the present chapter, we shal! deal only with normed s!aces X and Y. ~lLt~t

form the main topic of this chapter.

weak*-topology The relation£ bet\veen a Banach space X and its normed dual X *

norm on @(X, Y) defines a topol~gy OIl X* which turns out to be stronger than its

4.1 Theorem Suppose X and Yare norrned spaces. Associate to each A E .@(X, Y) the number

(1)

HAil = sup{IIAxil : x E X, "xII ~ I}.

88 GENERAL THEORY

This definition of IIAII makes &l(X, Y) into a normed space. If Y is a Banach space, so is

PROOF Since subsets of nOI med spaces are bounded if and only if they lie in

some multiple of the unit ball, IIAII < 00 for every A E ,18(X, Y). If o: IS a scalar,

then (rxA)(x) rx . Ax, so that

(2)

IIrxAIl = Irx\IIAIi.

The triangle inequality in Y shows that

II(Al + A2)xlI = IIAlx + A2xll < IIAlxll + IIA2xll s (IiAlll + IIA211)IIxil s; IIAlll + IIA211

for every x E X with IIxll < 1. Hence

(3)

If A =i= 0, then Ax ¥ 0 for some x EX, hence IIAII > O. Thus @(X, ¥) is a

W(X, l' ). Since

(4)

and since it is assumed that !IAn - Amll ~ 0 as nand m tend to 00, {An x} is a Cauchy sequence in Y for every x E X. Hence

(5)

Ax = Iim A,»

n ..... 00

exists. It is clear that A: X ~ Y is linear. If B > 0, the right side of (4) does not exceed ellxll, provided that m and n are sufficiently large. It follows that

(6) IIAx "m xII < ellxll

for all large m. Hence !lAx II < (IIAm!1 + S)lIx\l, so that A E @(X, ¥), and

pieteness of !?I( X, Y). 1111

IIA Am\l < B. Thus ,A'"m - .,l\. in the norm of.@(l:'. Y) This establishes the com-

4.2 Duality It will be convenient to designate elements of the dual space X* of 1:

by x* and to write

(1)

<x,x*)

in place of x*(x). This notation is well adapted to the symmetry (or duality) that exists between the action of X* on X on the one hand and the action of X on X* on the

,

other. The following theorem states some basic properties of this duality.

DUALIIY IN BANACH SPACES 89

4.3 Theorem Suppose B is the dosed unit ball 01 a normed space X Define

for ellery x* E X*.

IIx*II sup{l(x,x*)I'XEB}

(a) This norm makes X* into a Banach space.

(b) Let B* be the closed unit ball of X*. For every x EX,

[x] = sup{ I <x, x*): x* E B*}.

Consequently, x* ~ <x, x*) is a bounded linear functional on X*, of norm [x]. (c) B* is weak*-compact.

PROOF Since &6'(X, Y) - X*, when Y is the scalar field, (a) is a corollary of

TheOl em 4.1.

such that

Fix x E X. The corollary to Theorem 3.3 shows that there exists )1* E B*

On the other hand

(2)

I <x, x*) I < IIxil IIx*1I < IIxll

for every x* E B*. Part (b) follows from (1) and (2).

Since the open unit ball U of X is dense in B, the definition of IIx* II shows that x* E B* if and only if I <x, x*) I s 1 for every x E U. Part (c) now follows directly from Theorem 3.15. IIII

Remark The weak*-topology of X* is, by definition, the weakest one that makes all functionals

the proposition sLated at the end of SectIon 3.] 1 holds for the weak* topology as

than its weak*-topology; in fact, it is strictly stronger, unless dim X < 00, since

contimlOlJS. Part (Ji) shows the ref Ole that the norm topology of X* is stronger

Unless the contrary is explicitly stated, X* will from now on denote the

normed dual of X (whenever X is normed), and all topological concepts relating to X* will refer to its norm topology. This implies in no way that the weak"topology will not play an important role.

We now give an alternative description of the operator norm defined m Theorem 4.1.

90 GENERAL THEORY

4.4 Theorem If X and Yare normed spaces and if A € .@(x, Y), then

II All sup {I (Ax, y*) I : IIxll < 1, lIy* II < I}.

PROOF Apply (b) of Theorem 4.3 with Y in place of X This gIves

IIAxl1 sup{I(Ax,y*)I·lly*1I < I}

for every x EX. To complete the proof, recall that

IIAII = sup {\lAxll: IIxll ~ I}.

IIII

4.5 The second dual of a Banach space The normed dual X * of a Banach space X is itself a Banach space and hence has a normed dual of its own, denoted by X**. Statement (b) of Theorem 4.3 shows that every x E X defines a unique ¢x E X**,

(x, x*) (x*, rJ>x) (x* E X*),

(2)

II cfJxll IIxll (x eX).

It follows from (1) that cp: X ~ X** is linear; by (2), <t> IS an Isometry. Since X is now

assumed to be complete, 4>(X) is closed in X**.

Thus ¢ is an isometric isomorphism of X onto a closed subspace of X**. Frequently, X is identified with ¢(X); then X is regarded as a subspace of X**. The members of ¢(X) are exactly those linear functionals on X* that are con-

, tinuous relative to its weak*-topology. (See Section 3.14.) Since the norm topology of X* is stronger, it may happen that ¢(X) is a proper subspace of X**. But there are many important spaces X (for example, all LP-spaces with 1 < p < 00) for which 4>(X) = X**; these are called reflexive. Some oftheir properties are given in Exercise 1.

It should be stressed that, in order for X to be refiexive, the existence of some

isometric IsomorphIsm (jJ of X onto X** is not enough; it is Cl ucial that the identity (1)

be satisfied by cpo

subspace of X*; neither M nor N is assumed to be closed. Theil annihilators l.ll. and

4.6 Annihilators Sllppose X is a Banach space, M IS a subspace of X, and 1V is a

I N ate defined as follows:

M': = {x* E X*: <x, x*) = 0 for all xc M}, +N = {x c X: (x, x*) = 0 for all x* eN}.

Thus Ml. consists of all bounded linear functionals on X that vanish on M, and +N is the subset of X on which every member of Nvanishes. It is clear that Ml. and l.N'