Serge Lang Real and Functional Analysis Third Edition 6 SpringerSpringer New York Berlin Heidelberg Barcelona Budapest Hong Kong London Milan Paris Santa Clara Singapore Tokyo HCINauUCS Ae Editorial Board S. Axler F.W. Gehring P.R. HalmosBoos OF RELATED INTEREST BY SERGE LANG Fundamentals of Diophantine Geometry ‘A systematic account of fundamentals, including the basic theory of heights, Roth and Siegel's theorems, the Néron-Tate quadratic form, the Mordell-Weill theorem, Weil and Néron functions, and the canonical form on a curve as it related to the Jacobian via the theta function. Introduction to Complex Hyperbolic Spaces Since its introduction by Kobayashi, the theory of complex hyperbolic spaces thas progressed considerably. This book gives an account of some of the most important results, such as Brody's theorem, hyperbolic imbeddings, curvature properties, and some Nevanlinna theory. It also includes Cartan’s proof for the Sccond Main Thcorem, which was elegant and short. Pili ie Curves: Diophantine Analysis is systematic account of the basic diophantine theory on elliptic curves starts with the classical Weierstrass parametrization, complemented by the basic theory ‘of Néron functions, and goes on to the formal group, heights and the Mordell- Weil theorem, and bounds for integral points. A second part gives an extensive account of Baker's method in diophantine approximation and diophantine in- ‘equalities which were applied to get the bounds for the integral points in the first part. Cyclotomic Fields 1 and I ‘This volume provides an up-to-date introduction to the theory of a concrete and classically very interesting example of number fields. It is of special interest to number theorists, algebraic geometers, topologists, and algebraists who work in Kaheory. This book is a combined edition of Cyclotomic Fields (GYM 59) and Cyclotomic Fields I! (GTM 69) which are out of print. In addition to some minor corrections, this edition contains an appendix by Karl Rubin proving the Mazur-Wiles theorem (the “main conjecture’’) in a self-contained way. Other Books By LANG PUBLISHED BY SPRINGER-VERLAG Introduction to Arakelov Theory ¢ Riemann-Roch Algebra (with William Fulton) * ‘Complex Multiplication * Introduction to Modular Forms * Modular Units (with Daniel Kubert) * Introduction to Algebraic and Abelian Functions * Cyclotomic Fields I and I Varieties * Differential Manifolds * Complex Analysis * Real Analysis * Undergraduate Analysis * Undergraduate Algebra * Linear Algebra * Introduction to Linear Algebra * Calculus of Several Variables © First Course in Calculus * Basic Mathematics * Geometry: (with Gene Murrow) * Math! Encounters with High School Students * The Beauty of Doing Mathematics e THE FILESerge Lang Real and Functional Analysis Third Edition With 37 Illustrations & : SpringerSerge Lang Department of Mathematics Yale University New Haven, CT 06520 USA Editorial Board S. Axler F.W. Gehring PR. Halmos Department of Department of Department of Mathematics Mathematics ‘Mathematics Michigan State University University of Michigan Santa Clara University East Lansing, MI 48824 ‘Ann Arbor, MI 48109 Santa Clara, CA 95053 USA USA USA MSC 1991: Subject Classification: 26-01, 28-01, 46-01 Library of Congress Cataloging-in-Publication Data Lang, Serge, 1927- Real and functional analysis / Serge Lang. — 3rd ed. P. em. — (Graduate texts in mathematics ; 142) Includes bibliographical references and index. ISBN 0-387-94001-4 |, Mathematical analysis. 1. Title. M Series. QA300.1274 1993 ‘51S—de20 92-21208 ‘The previous edition was published as Real Analysis. Copyright 1983 by Addizon Wesley. Printed on acid-free paper. © 1993 Springer-Verlag New York, Inc. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer-Verlag New York, Inc, 175 Fifth Avenue, New York, NY 10010, USA) except for brief excerpts in connection with reviews or schol arly analysis. Use in connection with any form of information storage and retrieval, elec- tronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. ‘The use of general descriptive names, trade names, trademarks, tc, in this publication, even if the former are not especially identified, is not to be taken as a sign that such names. as understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone. Production coordinated by rian Howe and managed by Terry Kornak; manufacturing supervised by Vincent Scelta. ‘Typeset by Asco Trade Typesetting Ltd., North Point, Hong Kong. Printed and bound by R. R. Donnelley & Sons, Harrisonburg, VA. Printed in the United States of America. 98765432 ISBN 0-38 /-94001-4 Springer-Veriag New York Beriin Heidetberg ISBN 3-540-94001-4 Springer-Verlag Berlin Heidelberg New York SPIN 10545036Foreword This book is meant as a text for a first year graduate course in analysis. Any standard course in undergraduate analysis will constitute sufficient preparation for its understanding, for instance, my Undergraduate Anal- ysis. I assume that the reader is acquainted with notions of uniform con- vergence and the like. In this third edition, I have reorganized the book by covering inte- gration before functional analysis. Such a rearrangement fits the way courses are taught in all the places I know of. I have added a number of examples and exercises, as well as some material about integration on the real line (e.g. on Dirac sequence approximation and on Fouricr analysis), and some material on functional analysis (e.g. the theory of the Gelfand transform in Chapter XVI). These upgrade previous exercises to sections in the text. In a sense, the subject matter covers the same topics as elementary calculus, viz. linear algebra, differentiation and integration. This time, however, these subjects are treated in a manner suitable for the training of professionals, i.e. people who will use the tools in further investiga- tions, be it in mathematics, or physics, or what have you. In the first part, we begin with point set topology, essential for all analysis, and we cover the most important results. I am selective here, since this part is regarded as a tool, especially Chapters I and I. Many results are easy, and are less essential than those in the text. They have been given in exercises, which are designed to acquire facility in routine techniques and to give flexibility for those who want to cover some of them at greater length. The point set topol- ogy 2 hs of continnity, open and closed sets, connectedness, compactness, and continuous functions. The chapter ply de:vi FOREWORD concerning continuous functions on compact sets properly emphasizes results which already mix analysis and uniform convergence with the language of point set topology. In the second part, Chapters IV and V, we describe briefly the two basic linear spaces of analysis, namely Banach spaces and Hilbert spaces. The next part deals extensively with integration. We begin with the development of the integral. The fashion has been to emphasize positivity and ordering properties (increasing and decreas- ing sequences). I find this excessive. The treatment given here attempts to give a proper balance between L'-convergence and positivity. For more detailed comments, see the introduction to Part Three and Chapter VI. The chapters on applications of integration and distributions provide concrete examples and choives for leading the course in other directions, at the taste of the lecturer. The general theory of integration in mea- sured spaces (with respect to a given positive measure) alternates with chapters giving specific results of integration on euclidean spaces or the real line. Neither is slighted at the expense of the other. In this third edition, I have added some material on functions of bounded variation, and I have emphasized convolutions and the approximation by Dirac sequences or families even more than in the previous editions, for in- stance, in Chapter VIII, §2. For want of a better place, the calculus (with values in a Banach space) now occurs as a separate part after dealing with integration, and before the functional analysis. The differential calculus is done because at best, most people will only be acquainted with it only in euclidean space, and incompletely at that. More importantly, the calculus in Banach spaces has acquired consider- able importance in the last two decades, because of many applications like Morse theory, the calculus of variations, and the Nash—Moser im- plicit mapping theorem, which lies even further in this direction since one has to deal with more general spaces than Banach spaces, These results pertain to the geometry of function spaces. CL. the exercises of Chapter XIV for simpler applications. The next part deals with functional analysis. The purpose here is twofold, We place the linear algebra in an infinite dimensional setting ¥ raps, and we show how one can “linearize” a problem by taking derivatives, again in a setting where the theory can be applied to function spaces. This part includes several major spectral theorems of analysis, showing how we can extend to the infinite dimensional case certain results of finite dimen- sional linear algebra. The compact and Fredholm operators have appli- cations to integral operators and partial differential elliptic operators (e.g. in papers of Atiyah—Singer and Atiyah—Bott), Chapters XIX and XXEX. on unbounded hermitian operators, combineFOREWORD vii both the linear algebra and integration theory in the study of such operators. One may view the treatment of spectral measures as providing an example of general integration theory on locally compact spaces, whereby a measure is obtained from a functional on the space of contin- uous functions with compact support. I find it appropriate to introduce students to differentiable manifolds during this first year graduate analysis course, not only because these objects are of interest to differential geometers or differential topologists, but because global analysis on manifolds has come into its own, both in its integral and differential aspects. It is therefore desirable to integrate manifolds in analysis courses, and I have done this in the last part, which may also be viewed as providing a good application of integration theory. A number of examples are given in the text but many interesting examples are also given in the exercises (for instance, explicit formulas for approximations whose existence one knows abstractly by the Weierstrass— Stone theorem; integral operators of various kinds; etc). The exercises should be viewed as an integral part of the book. Note that Chapters XIX and XX, giving the spectral measure, can be viewed as providing an example for many notions which have been discussed previously: operators in Hilbert space, measures, and convolutions. At the same time, these results lead directly into the real analysis of the working mathematician. ‘As usual, I have avoided as far as possible building long chains of logical interdependence, and have made chapters as logically independent as possible, so that courses which run rapidly through certain chapters, omitting some material, can cover later chapters without being logically inconvenienced. The present book can be used for a two-semester course, omitting some material. I hope I have given a suitable overview of the basic tools of analysis. There might be some reason to include other topics, such as the basic theorems concerning elliptic operators. I have omitted this topic and some others, partly because the appendices to my SL,(R) constitutes a sub-book which contains these topics, and partly because there is no time to cover them in the basic one year course addressed to graduate students. The present book can also be used as a reference for basic analysis, since it offers the reader the opportunity to select various topics without reading the entire book. The subject matter is organized so that it makes the topics available to as wide an audience as possible. There are many very good books in intermediate analysis, and inter- esting research papers, which can be read immediately after the present course. A partial list is given in the Bibliography. In fact, the determina- tion of the material included in this Real and Functional Analysis has ndviii FOREWORD Finally, I thank all those people who have made valuable comments and corrections, especially Keith Conrad, Martin Mohlenkamp, Takesi Yamanaka, and Stephen Chiappari, who reviewed the book for Springer- Verlag. New Haven 1993/1996 SERGE LANGContents PART ONE General Topology CHAPTER | Sets . §1. Some Basic Terminology . §2. Denumerable Sets. §3. Zorn’s Lemma .. CHAPTER It Topological Spaces §1. Open and Closed Sets. §2. Connected Sets . §3. Compact Spaces §4. Separation by Continuous Functions §5. Exercises .. CHAPTER IIL Continuous Functions on Compact Sets .. §1. The Stone—Weierstrass Theorem §2. Ideals of Continuous Functions §3. Ascoli’s Theorem M4. Exercises .... 17 17 27 3 43 st St 55 37 59x CONTENTS PART TWO Banach and Hilbert Spaces CHAPTER IV Banach Spaces §1. Definitions, the Dual Space, and the Hahn-Banach Theorem §2. Banach Algebras .. §3. The Linear Extension Theorem §4. Completion of a Normed Vector Space . §5. Spaces with Operators ..........5 Appendix: Convex Sets ..... 1. The Krein—Milman Theorem 2. Mazur’s Theorem £6. Exercises CHAPTER V Hilbert Space §1. Hermitian Forms .....- §2. Functionals and Operators §3. Exercises PART THREE Integration CHAPTER VI §1. Measured Spaces, Measurable Maps, and Positive Measures §2. The Integral of Step Maps §3. The L!-Completion ... §4. Properties of the Integral: First Part §5. Properties of the Integral: Second Part §6. Approximations .. §7. Extension of Positive Measures from Algebras to o-Algebras §8. Product Measures and Integration on a Product Space . §9. The Lebesgue Integral in R? . §10. Exercises CHAPTER vii Duality and Representation Theorems ..................002..-- : §1. The Hilbert Space L7(u) .. §2. Duality Between L'(ji) and L(y) §3. Complex and Vectorial Measures §4. Complex or Vectorial Measures and Duality §5. The L? Spaces, 1

5" to denote the fact that S’ is a subset of S. Let S, T be sets. A mapping or map f: TS is an association which to each element xe 1° associates an element of 3, denoted by f(x), and called the value of f at x, or the image of x under f. If T’ is a subset of T, we denote by f(T’) the subset of S consisting of all elements f(x) for xeT". The association of f(x) to x is denoted by the special arrow xe flx). We usually reserve the word function for a mapping whose values are in the real or complex numbers. The characteristic function of a subset S' of 8 is the function y such that y(x)=1 if xe S’ and y(x)=0 if x¢ 5S’. We often write x5 for this function. Let X, Y be sets. A map f: X + ¥ is said to be injective if for all x, x'eX with x #x’ we have f(x) # f(x). We say that f is surjective if F(X) . if the image of f is all of ¥. We say that f is bijective if it is both injective and surjective. As usual, one should index a map f by its set of arrival and set of departure to have absolutely correct notation, but this is too clumsy, and the context is supposed to make it clear what these sets are. For instance, let R denote the real numbers, and R’ the4 SETS OL §1] real numbers 2 0. The map RoR given by xt-+x? is not surjective, but the map Sg: ROR given by the same formula is surjective. If f:X + Y is a map and S a subset of X, we denote by SIs the restriction of f to S, namely the map f viewed as a map defined only on S. For instance, if f: RR’ is the map x++x?, then f is not injec- tive, but /IR’ is iniective. We often let fe = fre be the function equal to f on § and 0 outside 5. ‘A composite of injective maps is injective, and a composite of surjec- tive maps is surjective. Hence a composite of bijective maps is bijective. We denote by Q, Z the sets of rational numbers and integers respec- tively. We denote by Z* the set of positive integers (integers > 0), and similarly by R' the set of positive reals. We denote by N the set of natural numbers (integers > 0), and by C the complex numbers. A map- ping into R or C will be called a function. Let S and I be sets. By a family of elements of S, indexed by I, one means simply a map f:1 +S, However, when we speak of a family, we write f(i) as fz, and also use the notation {f,},¢1 to denote the family. Example 1. Let S be the set consisting of the single element 3. Let T= {1,...,n} be the set of integers from 1 to n. A family of elements of S, indexed by J, can then be written {aj}i-1,.... with each a,= 3. Note that a family is different from a subset. The same element of S may receive distinct indices. A family of clements of a set $ indexed by positive integers, or non- negative integers, is also called a sequence. Example 2, A sequence of real numbers is written frequently in the form (ae Xa} or Ohaze and stands for the map f:Z* +R such that f(i) =x, As before, note that a sequence can have alll its elements equal to each other, that is (41 is a sequence of integers, with x, = 1 for each ie Z*0,91 SOME BASIC TERMINOLOGY 5 ‘We define a family of sets indexed by a set 7 in the same manner, that is, a family of sets indexed by J is an assignment ins, which to each i€ I associates a set S,. The sets S, may or may not have elements in common, and it is conceivable that they may all be equal. As before, we write the family {S,}rer- We can define the intersection and union of families of sets, just as for the intersection and union of a finite number of sets. Thus, if {S,}se, is a family of sets, we define the intersection of this family to be the set Os fet consisting of all elements x which lie in all S, We define the union 5, Us to be the set consisting of all x such that x lies in some 5,. If S, S’ are sets, we define S x S’ to be the set of all pairs (x, y) with xeS and yeS'’. We can define finite products in a similar way. If S,, S,,... is a sequence of sets, we define the product HS to be the set of all sequences (x1, xz,-..) with x,¢ 5; Similarly, if J is an indexing set, and {5;},<1 a family of sets, we define the product ns fel to be the set of all families {x;}1er with x;€ 5. Let X, Y, Z be sets. We have the formula (XUY) x Z=(X x DY x Z). To prove this, let (w, z)¢(X UY) x Z with weXUY and zeZ. Then weX orweY. Say weX. Then (w,z)eX x Z. Thus (XUY)x Zo (X x Zu(¥ x Z). ly, Xv Z 1Y) x Z and en is Y x Z Hence their union is contained in (XU Y) x Z, thereby proving our assertion.6 SETS Ls) We say that two sets X, Y are tif their intersection is empty. We say that a union Xu ¥ is disjoint if X and Y are disjoint. Note that if X, Y are disjoint, then (X x Z) and (Y x Z) are disjoint. We can take products with arbitrary families. For instance, if {Xi}ier is a family of sets, then (yx)xz- Yours If the family {X;},,1 is disjoint (that is X,9 X, is empty if i#/ for i, je), then the sets X, x Z are also disjoint. We have similar formulas for intersections. For instance, (XN Y) x Z=(X x Z)A(Y x Z). We leave the proof to the reader. Let X be a set and Y a subset. The complement of Y in X, denoted by GY, or X — Y, is the set of all elements xX such that x¢ ¥. If ¥, Z are subsets of X, then we have the following formulas: CAV UZ) = GY OG eZ, GV AZ) = GY OG. These are essentially reformulations of definitions. For instance, suppose xeX and x¢(¥UZ), Then x¢¥ and x¢Z. Hence xe ¥AGqZ. Conversely, if xe@Y AZ, then x lies neither in Y nor in Z, and hence x €6(Y UZ). This proves the first formula. We leave the second to the reader. Exercise: Formulate these formulas for the complement of the union of a family of sets, and the complement of the intersection of a family of sets. Let A, B be sets and f: A+B a mapping. If Y is a subset of B, we define f-*(¥) to be the set of all x € A such that f(x)e ¥. It may be that 1 -*(Y) is empty, of course. We call f-!(¥) the inverse image of Y (under ‘f). If f is injective, and Y consists of one element y, then f~*({y}) is either empty or has precisely one element. The following statements are easily proved: If f: A+B is a map, and Y, Z are subsets of B, then IV UZ)=fNvf"(Z), INV OD =I "(YN 0S(), More generally, if {¥%};¢1 is a family of subsets of B, thenOL §2] DENUMERABLE SETS 7 and simiiariy for the intersection. Furthermore, if we denote by ¥ — the set of all elements ye Y and y ¢ Z, then LV =2Z)= IN - I". IGx2) = Ef (2)- In particular, Thus the operation f~! commutes with all set theoretic operations. |, §2. DENUMERABLE SETS Let n be a positive integer. Let J, be the set consisting of all integers k, 1 Z*, and hence go f is a bijection of S with Z*. Let T be a sct. A sequence of elements of T is simply a mapping of Z* into T. If the map is given by the association nt+x,, we also write the sequence as {x,}no1» OF also {x,,x2,...}. For simplicity, we also write {x,} for the sequence. Thus we think of the sequence as prescrib- ing a first, sccond, ..., x-th element of T. We use the same braces for sequences as for sets, but the context will always make our meaning clear. Examples. The even positive integers may be viewed as a sequence {x,} if we put x, =2n for n=1, 2,.... The odd positive integers may also be viewed as a sequence {y,} if we put y, = 2n—1 forn=1, 2,.... In each case, the sequence gives an enumeration of the given set. We also use the word sequence for mappings of the natural numbers into a set, thus allowing our sequences to start from 0 instead of 1. If we8 SETS O, §2] need to specify whether a sequence starts with the 0-th term or the first term, we write {%Juzo Of (nha according to the desired case. Unless otherwise specified, however, we always assume that a sequence will start with the first term. Note that from a sequence {x,}, 39 we can define a new sequence by letting Yn =%q-1 for nZ 1. Then y; =o, y2=%4,.... Thus there is no essen- tial difference between the two kinds of sequences. Given a sequence {x,}, we call x, the n-th term of the sequence. A sequence may very well be such that all its terms are equal. For in- stance, if we let x, = 1 for all n= 1, we obtain the sequence {1, 1, 1, ...}. Thus there is a difference between a sequence of elements in a set 7, and a subset of T. In the example just given, the set of all terms of the sequence consists of one element, namely the single number 1. Let {x1,2,-.-} be a sequence in a set S. By a subsequence we shall mean a sequence {x,,,X,,)---} such that my k,. We define k,4, to be the smallest element of D which is > k,. Then the map n++k, is the desired enumeration of D. 22, Let § be a derumeruble sei und D an infinite subset of S. Then D is denumerable.OF, §2] DENUMERABLE SETS 9 subset of Z* in this enumeration. Using Proposition 2.1 we conclude that we can enumerate D. Proposition 2.3. Every infinite set contains a denumerable subset. Proof. Let $ be a infinite set. For every non-empty subset T of S, we select a definite element az in 7. We then proceed by induction. We let x, be the chosen element as. Suppose that we have chosen x1, having the property that for each k=2, ...,n the element x, is the selected element in the subset which is the complement of {x,,...4X,-1}- We let x,4; be the selected element in the complement of the set {x,,--.,%,}. By induction, we thus obtain an association nt+x, for all positive integers n, and since x, # x, for all k1. Then the left-hand side is even, but the right-hand side is odd, so the assumption r s, then which is impossible. Similarly, we cannot have s>m, whence m=. Hence our map is injective, as was to be proved. Proposition 2.6. Let {D,,D2,...} be a sequence of denumerable sets. Let S be the union of all sets D; (i= 1,2,...). Then S is denumerable. Proof. For each i=1, 2, ... we enumerate the elements of D,, as indicated in the following notation: Dy: {X14 X125 X13 -—+} Dz: (X15 Xa2-¥aae--} The map f:Z* x Z* + D given by MG) = is then a surjective map of Z* x Z* onto S. By Proposition 24, it follows that S is denumerable. Corollary 2.7. Let F be a non-empty finite set and D a denumerable set. Then F x D is denumerable. If S,, S;, ... are a sequence of sets, each of which is finite or denumerable, then the union S,US,U"~ is denumerable or finite. Proof. There is an injection of F into Z* and a bijection of D with Z*. Hence there is an injection of F x D into Z* x Z* and we can apply Corollary 2.2 and Proposition 2.6 to prove the first statement. One could also define a surjective map of Z* x Z* onto F x D. As for the second statement, each finite set is contained in some denumerable set, so that the second statement follows from Propositions 2.1 and 2.6. For convenience, we shall say that a set is countable if it is either finite or denumerable. 1, §3. ZORN’S LEMMA ¥ seis simuliancousiy, one needs a special property. To state it, we need some more terminology. Let S be a set. An ordering (also called partial ordering) of (or on) $U, §3] ZORN’S LEMMA rT is a relation, written x < y, among some pairs of elements of 5, having the following properties. ORD 1. We have x Sx. ORD 2. If x Sy and ySz thenx S$ ORD 3. If x Sy and yx then x=y. ‘We sometimes write y = x for x Sy. Note that we don't require that the relation x S y or y S x hold for every pair of elements (x, y) of S. Some pairs may not be comparable. If the ordering satisfies this additional property, then we say that it is a total ordering. Example 1, Let G be a group. Let $ be the set of subgroups. If H, H’ are subgroups of G, we define HSH if H is a subgroup of H’. One verifies immediately that this relation defines an ordering on S. Given two subgroups, H, H’ of G, we do not necessarily have H SH’ or H’ SH. Example 2. Let R be a ring, and let S be the set of left ideals of R. We define an ordering in $ in a way similar to the above, namely if L, L’ are left ideals of R, we define LSU #LcL. Example 3. Let X be a set, and S the set of subsets of X. If ¥, Z are subsets of X, we define Y SZ if Y is a subset of Z. This defines an ordering on S. In all these examples, the relation of ordering is said to be that of inclusion. In an ordered set, if x $ y and x # y we then write x < y. Let A be an ordered set, and B a subset. Then we can define an ordering on B by defining x Sy for x, ye B to hold if and only if xs y in A. We shall say that it is the ordering on B induced by the ordering on A, or is the restriction to B of the partial ordering of A. Let S be an ordered set. By a least element of S (or a smallest clement) one means an clement a¢S such that a 0, we mean the set of all x E such that |x —v| 0) there exists N such that for all m,n = N we have Ix, — 1 <6. This sequence is said to converge to an element x if given «, there exists N such that for all n > N we have Ix~xl 0 such1, §1) OPEN AND CLOSED SETS 19 that [J (x) S : for all x€S. If f is bounded, define Ills = 71 = sup Lf), sup meaning least upper bound. It can be easily shown that the set of bounded maps B(S, F) of S into F is a vector space, and that || || is a norm on this space, called the sup norm. The L'-Norm. Let E be the space of continuous functions on [0, 1]. For fe E define rts = [*i7enrae Then || Mh is a norm on E, called the L'-norm. This norm, will be a Much of this book is devoted to studying the convergence of se- quences for one or the other of the above two norms. For instance, consider the sup norm. A sequence of maps {f,} is said to be uniformly Cauchy on S if given ¢ there exists N such that for all m,n > N we have Wn — Snlls < & It is said to be uniformly convergent to a map f if given e there exists N such that for all n 2 N we have I — Sls <& In the second example, we would use the expressions L'-Cauchy and L'-convergent instead of uniformly Cauchy and uniformly convergent, if we replace the sup riorm by the L!-norm in these definitions. Up to a point, one can generalize the notion of subset of a normed vector space as follows. Let X be a set. A distance function (also called a metric) on X is a tnap (x, y)++ d(x, y) from X x X into R satistying the following conditioris: DIS 2. For all x, y, we have d(x, y) = d(y, x). DIS 3. For all x, y, z, we have A(x, 2) S d(x, y) + aly, 2). A set with a metric is called a metric space. We can then define open20 TOPOLOGICAL SPACES OW §1] balls just as we did in the case of normed vector spaces, and also define a topology in a metric space just as we did for a normed vector space. Every open set is then a union of open balls. This topology is said to be determined by the metric. In a normed vector space, we can define the distance between elements x, y to be d(x, y) = |x — yl. It is immediately verificd that this is a metric on the space. Conversely, the reader will see in Exercise 5 how a metric space can be embedded naturally in a normed vector space, in a manner preserving the metric, so that the “generality” of metric spaces is illusory. For convenience, we also make here the following definition: If A, B are subsets of a normed vector space, we define their distance to be aA, B)=inf|x-yl, xeA,yeB. Basic theorems concerning subsets of normed vector spaces hold just as well for metric spaces. However, almost all metric spaces which arise naturally (and certainly all of those in this course) occur in a normed vector space with a natural linear structure. There is enough of a change of notation from |x — y| to d(x, y) to warrant carrying out proofs with the norm notation rather than the other. Let F and 7’ be topologies on a set X. One verifies at once that they are equal if and only if the following condition is satisfied: For each xeéX and each set U open in J containing x, there exists a set U’ ‘open in J” such that xe U'c U, and conversely, given U’ open in 7’ containing x, there cxists U open in F such that xe U < U’. Example, The reader will verify easily that two norms | |, and | |, on a vector space E give rise to the same topology if and only if they satisly the following condition: There exist C,, C, > 0 such that for all xe E we have Clxh Six S$ Clxh. If this is the case, the norms are called equivalent. Just to fix terminology, we define the closed ball centered at v and of radius r 2 0 to be the set of all x € E such that Iv — alse. We define the sphere centered at v, of radius r, to be the set of points x such that Ix — ol =r. Warning. In some books, what we call a ball is called a sphere. This jot good i ogy used here is now essen- tially universally adopted.(0, §1) OPEN AND CLOSED SETS 21 Examples of normed vector spaces are given in the exercises. The standard properties of subsets of normed vector spaces having to do with limits are also valid in metric spaces (cf. Exercise 5). We can define balls and spheres in metric spaces just as in normed vector spaces. We can also define the notion of Cauchy sequence in a metric space X as usual (again cf. Exercise 5), and X is said to be complete if every Cauchy sequence converges, i. has a limit in X. Example 5. Let G be a group. We define a subset U of G to be open if for each element xe U there exists a subgroup H of G, of finite index, such that xH is contained in U. It is a simple exercise in algebra to show that this defines a topology, which is called the profinite topology. Example 6. Let R bc a commutative ring (which according to stan- dard conventions has a unit element). We define a subset U of R to be open if for each x€U there exists an ideal J in R such that x+J is contained in U. It is a simple exercise in algebra to show that this defines a topology, which is called the ideal topology. Note. The topologies of Examples 5 and 6 will not occur in any significant way in this course, and may thus be disregarded by anyone uninterested in this type of algebra. A set together with a topology is called a topological space. In this chapter we develop a large number of basic trivialities about topological spaces, and except for the numbered theorems, it is recommended that readers work out the proofs for all other assertions by themselves, even though we have given most of them. The duality between intersections and unions with respect to taking the complement of a subsct allows us to define a topology by means of the complements of open sets, called closed sets. In any topological space, the closed sets satisfy the following conditions: CL i. The empty set and the whole space are closed. CL 2. The finite union of closed sets is closed. CL 3. The arbitrary intersection of closed sets is closed. The first condition is clear, and the other two come from the fact that the complement of the union of subsets is equal to the intersection of their complements, and that the complement of the intersection of subsets is equal to the union of their complements C ly, Bi cae topological space), we say that it defines a topology on X by means of of a set X (not yet a22 TOPOLOGICAL SPACES (1, §1] closed sets if its elements satisfy the three conditions UL 1, 2 3. We can then define an open set to be the complement of a set in F. Example 7, Let X =R". Let f(x,,..-,X,) be a polynomial in n vari- ables. A point a =(a,,...,a,) in R" is called a zero of f if f(a)=0. We define a subset S of R” to be closed if there exists a family {fi}ier of polynomials in n variables (with real coefficients) such that S consists precisely of the common zeros of all f; in the family (in other words, all points ae R" such that f(a) = 0 for all i). The reader may assume here the result that, for any such closed set S, there exists a finite number of polynomials f,, ...,f such that S is already the set of zeros of the set {fis}. It is easy to prove that we have defined a topology by means of closed sets, and this topology is called the Zariski topology on R”. It is a topology which is adjusted to the study of algebraic sets, that is sets which are zeros of polynomials. It will not reappear in this course, and again a disinterested reader may omit it. It does become important in subsequent courses, however. In 2-space, a closed set consists of a finite number of points and algebraic curves. In 3-space, a closed set consists of a finite number of points, algebraic curves, and algebraic surfaces. Let X be a topological space, and § a subset. A point x €X is said to be adherent to S if given an open set U containing x, there is some point of S lying in U. In particular, every element of S is adherent to S. A point of X is called a boundary point of 5 if every open set containing this point also contains a point of S and a point not in S. Thus an adherent point of S which does not lie in S is a boundary point of S. An interior point of S is a point of S which does not lie in the boundary of S. The set Int(S) of interior points of S is open. A subset $ of X is closed if and only if it contains all its boundary points. This follows at once from the definitions. By the closure of a subset S of X we mean the union of S and all its boundary points. The closure of S, denoted by 5, is therefore the set of adherent points of S. It is also immediately verified that 5 is closed, and is equal to the intersection of all closed sets containing S. In particular, we have Ss. As an exercise, the reader should prove that for subsets 5, T of X we have: SUT=SoT and SAT CSnT Eg: does Equal (Example?)OF, §1] OPEN AND CLOSED SETS 2B A subset S of a space X is said to be dense (in X) is S=X. For instance, the rationals are dense in the reals. Let X be a topological space and S a subset. We define a topology on S by prescribing a subset V of S to be open in S if there exists an open set U in X such that Y= UNS. The conditions for a topology on S are immediately verified, and this topology is called the induced topology. With this topology, S is called a subspace. Note. A subset of 5 which is open in $ may not be open in X. For instance, the real line is open in itself, but definitely not open in R?. Similarly for closed sets. On the other hand, if U is an open subset of X, then a subset of U is open in U in the induced topology if and only if it is open in X. Similarly, if S is a closed subset of X, a subset of S is closed in S$ if and only if it is closed in X. If F is a certain property of certain topological spaces (e.g. connected, or compact as we shall define later), then we say that a subset has property P if it has this property as a subspace. A topology on a set is often defined by means of a base for the open sets. By a base for the open sets we mean a collection @ of open sets such that any open set U is a union (possibly infinite) of elements of 2. There is an easy criterion for a collection of subsets to be a base for a topology. Let X be a set and @ a collection of subsets satisfying: B 1. Every element of X lies in some set in B. B2. If B, B’ are in @ and xe BB then there exists some B" in @ such that x € B” and B’ c BOB’, If @ satisfies these two conditions, then there exists a unique topology whose open sets are the unions of sets in @. Indeed, such a topology is uniquely determined, and it exists because we can define a set to be open if it is a union of sets in 4. The axioms for open sets are trivially verified. Example. The open balls in a normed vector space form a base for the ordinary topology of that space. Example. Let X be a set and let %, ¥ be topologies on X, that is collections of open sets satisfying the axioms for a topology. We say that ¥ is a refinement of %, or that @ is coarser than ¥, if every set open in % is also open in ¥. Thus & has fewer open sets than ¥ (“fewer” in the weak sense since % may be equal to ¥). pace Fbea ei ty of mappings Let @ be the family of all subsets of X consisting24 TOPOLOGICAL SPACES CH, §1) of the sets f-"(W), where W is open in Y and f ranges over ¥. Then we leave to the reader the verification of the following facts: 1. @ is a base for a topology on X, ic. satisfies conditions B 1, B 2. 2. This topology is the coarsest topology (the one with the fewest open sets) such that every map f € ¥ is continuous. We call this topology the weak topology on X determined by #. For an application of the weak topology, see Chapter IV, §1 and also the appendix of Chapter IV. There is a generalization of the weak topology as follows. Instead of considering one spacc Y, we consider a family of spaces {Yj}, for i ranging in some index set. We let ¥ be a family of mappings f;: X - ¥. We let @ be the family of all subsets of X consisting of finite interse tions of sets j; *(U;) where U, is open in ¥. Then again it is easi verified that @ is a base for a topology, called the weak topology deter- mined by the family ¥, The product topology defined below will provide an example of this more general case, when the family ¥ is the family of projections on the factors of a product. A topological space is said to be separable if it has a countable base. (By countable we mean finite or denumerable) Exercises on separable spaces designed to acquaint the reader with them, and essentially all trivial, are given at the end of the chapter. It is easy to see that the real numbers have a countable base. Indeed, we can take for basis elements the open intervals of rational radius, centered at rational points. Simi- larly, R" has a countable base. Note. In most cases, the property defining separability is equivalent with the property that there exists a countable dense subset (cf. Exer 15), and this second property is sometimes used to define separability. We find our definition to be more useful but the reader is warned on the discrepancy with some other texts. An open set containing a point x is called an open neighborhood of this point Ry a neighhorhood of x we mean any set conte: = set containing x. In a normed vector space, one speaks of an e-neighbor- hood of a point x as being a ball of radius ¢ centered at x. Let X, ¥ be topological spaces. A map f: X — Y is said to be contin- uous if the inverse image of an open set (in ¥) is open in X. In other words, if V is open in Y then f~'(V) is open in X. Equivalently, we see that a map f is continuous if and only if the inverse image of a closedHt, $13 OPEN AND CLOSED SETS 25 Proposition 1.1. Let E, F be normed vector spaces and let f:E+K be a map. This map is continuous if and only if the usual (s, 5) definition is satisfied at every point of E. We prove one of the two implications. Assume that f is continuous and let x€E. Given e, let V be the open ball of radius ¢ centered at f(x). The open set U=f~1(V) contains an open ball B of radius 6 centered at x for some 6. In particular, if ye E and |x — y| <6, then SY) EV and f(y) — fO0| <6. This proves the (,6) property. The con- verse is equally clear and is left to the reader. Actually, this (e, 5) property can be formulated analogously in arbi- tary topological spaces, as follows: The map f:X—+Y is said to be continuous at a point xe X if given a neighborhood V of f(x) there exists a neighborhood U of x such that f(U)< V. It is then verified at once that f is continuous if and only if it is continuous at every point. Proposition 1.2. Let X be a metric space (or a subset of a normed vector space) and let f:X +E be a map into a normed vector space. ‘Then f is continuous if and only if the following condition is satisfied. Let {xj} be a sequence in X converging to a point x. Then {f(%,)} converges to f(x). The proof will be left as an exercise to the reader. A composite of continuous maps is continuous. Indeed, if f: X + ¥ and g: Yr Z are continuous maps and V is open in Z, then oN =s-o WV) is seen to be open. As usual, we observe that a continuous image of an open set is not necessarily open. A continuous map f: XY which admits a continuous inverse map g: YX is called a homeomorphism, or topological isomorphism. It is clear that a composite of homeomorphisms is also a homeomorphism. As usual, we ubserve ilai a cuiiinuous bijective map need not be a homeomorphism. In fact, later in this course, we meet many examples of. vector spaces with two different norms on them such that the identity map is continuous but not bicontinuous. Let {X,}ier be a family of topological spaces and Ict x=[T[X,26 TOPOLOGICAL SPACES OH, §1] be their product. We define a topology on X, caiied tite prow ogy, by characterizing a subset U of X to be open if for each x ¢ U there exists a finite number of indices i,, ...,4, and open sets U,,, ...,U,, in the spaces X;,, .-.+X,, respectively such that xeU, xox Ux T] XU. ity The product for i # i, is taken for all indices i unequal to i,, other words, we can say that the product topology is the one having as a base all sets of the form U, x x U;,, x TT X- ithe Such sets have arbitrary open sets at a finite number of components, and the full spa The product topology is the unique topology with the fewest open sets in X which makes each projection map me X +X, continuous. Indeed, for each open set U; in X;, the set HU) = Ux [Xs ; must be open if 1; is continuous, and our previous assertion follows. In other words, it is the weak topology determined by the family of all projections on the factors. More generally, given a set and a family of mappings of this set into topological spaces, one can define a unique topology on the set making all these mappings continuous, and having the fewest open sets doing namely the weak topology. If S is a set, and {SiS Yhter is a family of maps into topological spaces ¥,, then the map HIyoTLy tet such that f(x) = {/(x)} is continuous for this topology. Example 8 We can give R" the product topology, which is called the ordinary topology. We define the sup norm on R” by xl] = max|x/](U1, 823 CONNECTED SETS 27 if x =(%1,...4%,) is given in terms of its coordinates. Ihen the topology determined by this norm is clearly the same as the product topology. Remark. A map f:X—Y which maps open sets onto open sets is said to be open. A map which maps closed sets onto closed sets is said to be closed. A continuous map need not be either. For instance, the graph of the tangent is closed in the plane, but the projection map on the x-axis maps it on an open interval: Figure 2.1 The map which folds the plane over the real axis maps the open plane on the closed half plane. If f: X + Y is continuous and bijective, then a necessary and sufficient condition that f be a homeomorphism is that f be open. This is simply a rephrasing of the continuity of the inverse mapping f~'. §2. CONNECTED SETS A topological space X is said to be connected if it is not possible to express X as a union of two disjoint non-empty open sets. Of course, we can formulate the definition in terms of closed sets instead of open sets. The reader's intuition of connectedness probably comes from the pos- sibility of connecting two points of a set by a path. We shall discuss the relation between this notion and the general notion later, after developing first some basic properties of connected sets. Proposition 2.1. Let f-X + ¥ he a continuous map. If X is connected then the image of X is connected.28 TOPOLOGICAL SPACES LIL §2] Proof. Without ioss of generality we may assume that ¥ is the of f. Suppose that ¥ is not connected, so that we can write Y= UUV where U, V are open, non-empty, and disjoint. Then x= f(s), which is impossible, This proves our assertion. Proposition 2.2. A topological space X is connected if and only if every continuous map of X into a discrete space having at least two elements is constant. Proof. Assume that X is connected, and that f is a continuous map of X into a discrete space with at least two elements. If f is not constant, we can write the image of f as a union of two disjoint non-empty sets, open by definition, and this contradicts our previous result. Conversely, suppose that we can write X = UUV as a disjoint union of non-empty open sets. Let p, q be two distinct objects and let the set {p, q} have the discrete topology. If we define f:X {p,q} to be the map such that SU) ={P} and f(V) = {4}, then f is continuous and not constant, as was to be shown. Observe that our proof shows that instead of taking a discrete space having at least two points, we can take a space with exactly two points in characterizing a connected set, as we have just done. Proposition 2.3. Let X be a topological space and let {i family of subspaces which are connected. If they have a point in com- mon then their union is connected. Proof. Let a lie in the intersection of all S;. If we can write Us where U, V are open in this union, then 5; U and S,n\V are open in S, for each i and hence S,< U or S,cV. If for some i we have Sc U, then a U and consequently we must have 5, U for all i, thus proving our assertion, uuy,CH, §2] CONNECTED SETS 29 As a consequence of the preceding statement, we define the connected component of a point a in X to be the union of all connected subspaces of X containing a. component is actually not empty, because the set consisting of a alone is connected. Proposition 2.4. Let X be a topological space and S a connected subset. Then the closure of S is connected. In fact, if S X is a continuous map defined on a closed interval [a;, 5) such that 4 (b;) = O41 (141). We say that this path goes from x to y if a(a)=x and a,(b,)=y. Of course, if such a path exists, then it is easy to define just one continu- ‘ous map a: [a,b] +X from some interval [a, b] into X such that (a) =x and a(b) = y. One can even take the interval [a, b] to be [0, 1]. Proposition 2.6. Any interval of real numbers is connected. Proof. We give the proof for a closed interval J = [a, b] and leave the other cases (open, half-open, infinite intervals) as exercises. Suppose that we can write J = AUB where A, B are closed, disjoint, and non-empty. A Leto he the est lower hound of B. Then c lies in30 TOPOLOGICAL SPACES LL §2] the ciosure of & and since B 1s closed, ce B, so C#a. For any xe J with a 0 there exists a finite number of open balls of radius r which cover S. Proof. We have already proved that compactness implies sequential compactness. Conversely, assume that 5 is sequentially compact. Then certainly S is complete, and we shall prove that the other condition stated in (ii) is satisfied. Suppose it is not. Let r>0. Let x, €S and let B, be the open ball of radius r centered at x,. Then B, does not contain 5, and there is some x2€S, x. ¢B,. Proceeding inductively, suppose that we have found open balls B,, ...,B, of radius r, and points x,, ss-5%q with x,€B, such that x,4, does not lie in By U-"*U B. We can then find x,., which does not lie in B, U---UB,, and we let B,,, be the open ball of radius r centered at x,41- Let v be a point of accumulation of the sequence {x,}. By definition, there exist positive integers m, k with k>m such that [xq — vl < 7/2 and [xq — of < 1/2. Then |x; — ql <1 and this contradicts the property of our sequence (x,} because x, lies in the ball B,. This proves that S satisfies the condition of (ii. Now assume this condition. Let {U;};¢ be an open covering of S, and suppose that there is no finite subcovering. We construct a sequence fs} ve We hat S is covered by a finite number of closed balls of radius }. Hence there exists at least one closedLH, §3] COMPACT SPACES 35 ball C, of radius $ such tai Cy 15 is nui covered by a finite number of U,. We let x, be a point of C, oS. Suppose that we have obtained a sequence of closed balls Qa3G such that C, has radius 1/2", with a point x,¢C,7S, and such that CG, is not covered by a finite number of U;. Since S itself can be covered by a finite number of closed balls of radius 1/(2"*), it follows that C,7S can also be so covered, and hence there exists a closed ball Cysr of radius 1/(2"*) and such that C,,,S cannot be covered by a finite number of U;. We let x,., be a point of C,4,°S. This constructs our sequence as desired. We see that {x,} is a Cauchy sequence in S, which coverges to a point x in S. But x lies in some U,, which contains C, for all sufficiently large n, a contradiction which proves our theorem. A subset $ of a metric space, or a normed vector space, which can be covered by a finite number of open balls of given radius r > 0 is said to be totally bounded, We can phrase (ji) by saying that S is compact if and only if it is complete and totally bounded. A subset of a topological space is said to be relatively compact if its closure is compact. From (ii) we get a convenient criterion for relative compactness. Corollary 3.9. Let S be a subset of a complete normed vector space. Assume that given r > 0 there exists a finite covering of S by balls of radius r. Then S is relatively compact. Proof. The closure 5 of S has the same property, because if S is covered by a finite number of balls of radius 7/2, then the closure of S is covered by a finite number of balls of radius r (centered at the same points). Also 5 is complete. Hence we conclude that the closure of S is compact. As an application of Theorem 3.8, we recall that a closed (bounded) interval in R has the Weierstrass—Bolzano property. Hence it is compact, and therefore so is any closed bounded subset of R (being a closed subset of a compact set). The converse is also true, since a compact set is closed, and must be bounded, otherwise one can find an infinite sequence tending to infinity, and not having a point of accumulation. One can also prove the compactness of a closed interval directly from the least upper bound axiom, as follows. Let a , a contradiction, proving that c= and that [a, b] is compact. One can generalize to arbitrary compact sets some standard theorems on closed intervals, eg.: Proposition 3.10. Let A be a compact set, and f: A-+R a continuous function on A. Then f has a maximum (a point ce A such that SO Z Sle) for all xe A). Proof. The image f(A) is compact, so closed and bounded. The least upper bound of f(A) lies in f(A), thus proving our assertion. If A is a subset of a normed vector space, and if f:A>F is a continuous map into some normed vector space F, then we say that f is uniformly continuous on A if given ¢ there exists 6 such that whenever x, yeA and |x—y| <6, then [f(x) — f(y) 0 he such that if |y—x| 0 is compact, and hence the unit ball is compact by multiplication with a positive number.) Corollary 3.15 (F. Riesz). A normed vector space is locally compact if and only if it is finite dimensional. Proof. Let E be a locally compact normed vector space, and let B be the closed ball of radius 1 centered at 0. We can find a finite number of points x,, ...,x,€B such that B is covered by the open balls of radius $ centered at these points. We contend that x,, ...,%, generate E. Let F be the subspace generated by x1, ...,%,- Then F is finite dimensional, hence closed in E as a trivial consequence of Corollary 3.14. Suppose that xe E and x¢F. Let d(x, F) = inf |x — yl. Drawing a closed ball around x intersecting F, and using the fact that the intersection of F and this ball is compact, we conclude that there is some zé F such that d(x, F) =|x —z|, and we have x —2z #0 since F is closed in E. Then there is some x; such that |x-z lix=ai =xl b if and only if x ¢ U, for some r > b, so that S61 = YT, This proves our theorem. Since a compact Hausdorff space is normal. Urysohn’s lemma applies in this case. One neods it frequently in the locally compact case in the following form. Corollary 4.3. Let X be a locally compact Hausdorff space, and K a compact subset. There exists a continuous function g on X which is 142 TOPOLOGICAL SPACES CI §4] Proof. Each xéK has an open neighborhood ¥, with compact cio- sure. A finite number of such neighborhoods V,,, ..-5, covers K. Let Vehwuruk,. Then the closure of V is compact. There exists a continuous function g20 on V (compact Hausdorff, hence normal) which is 1 on K and 0 outside V, ie. 0 on Vn GV. We define g to be 0 on the complement of 7 in X. Then g is continuous at every point in the complement of V, and as function on X is also continuous on V. This proves our corollary. Theorem 4.4 (Tietze Extension Theorem). Let A be a closed subset of a normal space X and let f be a continuous (real valued) function on A, Then there exists a continuous function f* on X whose restriction to A is equal to f. If f has values in [0, 11, then we can choose f* to have values in [0, 1] also. Proof. Assume first that f has valucs in [0,1]. If A, B are disjoint closed subsets of X, we denote by g,.2 a function with values in [0,1] such that (4) = 0 and g(B) = 1. Such a function exists by Theorem 4.2. We shall now define functions f, on A and g, on X. We let fo =f and define sets Ao, Bo by the conditions: Ao = {xe A such that f(x) <4}, By = {xe A such that f(x) 2 3}. We let Go = 404,09 and define f, = fo ~ go. Inductively, suppose that we have defined f,; we have A, = {x A such that f,(x) S(4)(3)"}, B, = {x € A such that J,(x) 2 (3)(3)"}. We then define In = (YVI.2. and let fuss = Sg —9n- (Here of course, we understand by g, its restric tion to A.) Then in particular: Invs =F — (Go + °°" + Gy). We have (*) OSg,S33" and O 0. One distinguishes the three cases in which for a given xe A we have xe A,, or x¢A, but x¢B,, or xeB,. The desired inequality of f, is then obvious in each case, using the inductive hypothesis. From our inequalities («), we then conclude that the series Go+ Oy bie + yb converges pointwise, and furthermore converges to f on A. The uniform bounds imply at once that the limit function is continuous, thus proving Theorem 4.4, when f has values in [0, 1]. Remark 1. The restriction to the interval [0,1] is of course unneces- sary, and the theorem extends at once to any other closed bounded on 0, 1}. terval, for instance by mapping cuch an Now suppose that f is unbounded. Using the arctangent map we reduce the theorem to the case when f takes values in the open interval (—1, 1) and we must then know that the extension can be so chosen that its values also lie in the open interval (1,1). Let B be the closed set where the extension f* (which we have constructed with i [-1,1]) takes on the values 1 or —1. Then A and B are di that by Urysohn’s lemma there exists a continuous function h on X with values in [0,1] such that h is 1 on A and 0 on B. Then hf* has values in the open interval (—1, 1), as desired. This concludes the proof of Theorem 4.4. Remark 2. The theorem also holds in the complex case dealing sepa- rately with the real and imaginary parts. The extra condition on the restriction of the values can then be formulated analogously by requiring that If" s yi. Indeed, suppose that we have extended f to a bounded continuous com- plex valued function g. Let b= |[/l. Let h be the function such that h@) =z if |z| Sb, and h(z) = bz/\z| if [2] >. Then h is continuous, Whi) Y be a mapping that is uniformly cor jous.44 TOPOLOGICAL SPACES CH, §5] Let A be the closure of A in X. Show that there exists a unique extension of f to a continuous map f: A-> ¥, and that f is uniformly continuous. You may assume that X, Y are subsets of a Banach space if you wish, in order to write the distance function in terms of the absolute value sign. 2. Seminorms. Let E be a vector space. A function o:E-R is called a seminorm if it satisfies the same conditions as a norm except that we allow a(x) =0 without necessarily having x=0. In other words, o satisfies the following conditions: SN 1. We have o(x) 2 0 for all xe E. SN 2. If xe E and a is a number, then o(ax) = |a|o(x). SN 3. We have o(x + y) S o(x) + o(y) for all x, ye E. We also denote a seminorm by the symbols | |. (a) If | | is a seminorm on E, show that the set Eo of elements xe E with iio (b) Define open balls with respect to a seminorm as with a norm. Show that the topology whose base is the family of open balls is Hausdorff if and only if the seminorm is a norm. (c) Let {0,} be a sequence of seminorms on E such that the values ¢,(x) are bounded. Let {a,} be a sequence of positive numbers such that 'a, converges. Show that J\a,0, is a seminorm, (@) Let {o;}:e1 be a family of seminorms on a vector space E. Let xo ¢E and let iy, ---si, be a finite number of indices. Let r>0. We call the set of all x € E such that ox — x9) 0, there exists N such that for m,n N, we have Wn Salle <6 where ll IIx is the sup norm on K. In other words, the sequence restricted to K is uniformly Cauchy. The sequence is said to be uniformly convergent on ‘compact sets if there is some map f having the following property. Given a compact set K and ¢, there exists N such that for n 2 N, we have Wa — Se <& In other words, the sequence restricted to K is uniformly convergent. We shall now make M(X, E) into a metric space for which the above convergence is the same as convergence with respect to this metric, in certain cases. A sequence {K;} of compact subsets of X said to be exhaustive if their union is equal to X, and if every compact subset of X is contained in some K,. We assume that there exists such a sequence {K,}. (a) Define atg)= $2 mint, 1) If f is unbounded on K, then we set [Ifllx=0o and min(1, [Iflx) Show that d(f) satisfies two of the properties of a norm, namely: d(f)=0 ifand only if f= Af +g) Sd(f) + d(g). (b) Define d(f, a) by d(f~ 4). Show that d(f.a) is a metric on M(X, F) (©) Show that 2*inf(L, Lf) SAP) and dN) SS a, + 2% (d) Show that a sequence {f,} converges uniformly on compact sets if and only if it converges in the above metric. (e) Let K be a compact set and e>0. Given f, let V(Jf, K,e) be the set of naps y such that if —gix 1. If for all 1< k k girls, marry off one boy and refer the others to the induction hypothesis. If for some k with 1 Sk —6, or in other words, Sle) = 6 0 such that —csfa)se for all x€S. The absolute value function can be uniformly approximated by ordinary polynomials on the interval [—c,c] by Exercises 6, 7, or 8, which are very simple ad hoc proof. Given ¢, let P be a polynomial such mt [PQ - [tll Z(s'). We ask to what extent the set of zeros of an ideal deter- mines this ideal, and answer this question in an important case. Theorem 21. Let X be a compact space, and let R be the ring of continuous functions on X, with the sup norm. Let J be a closed ideal (ie. an ideai, ciosed under tie sup norm). if fe R is such thar for all zeros x of J (ie. if f vanishes on the set of zeros of J), then f lies in J. Proof. Given &, let U be the subset of X consisting of all x eX such that |f(x)| <# Then U is open, and the complement $ of U is closed, and hence compact. Note that U contains Z,. For each yeS, we can find a function g, in J such that g,(x) #0 in some open neighborhood ¥, of y (by continuity). There is some finite covering {V,,,.-.4,,.} of S corresponding to functions g,,,....9,,,- Let I= Gt + Dh Then g is in J. is continuous, is nowhere 0 on S, and 2 0. Since g has a minimum on S, there is a number a >0 such that g(x) 2 a for all xeS. The function ng T+ ng lies in J, because 1+ ng is nowhere 0 on X, its inverse is continuous on X, so in R, and hence (1+ ng)nge J. For n large, the function nail + na) tends uniformly to 1 on S. and hence the function ng Tr1g lies in J, and approximates f within ¢ on S. Since 0S ng/(1 + ng) <1 it follows that on U we have the estimate 0 Sl fnail + no)| <5,(HH, §3] ASCOLI’S THEOREM 57 and so fng/(1 + ng) lies within 2¢ of f. Thus we have shown that f lies in the closure of J. Since J is assumed closed, we conclude that f lies in J, thereby proving our theorem. Remark 1. Situations analogous to that of Theorem 2.1 arise fre- quently in mathematics. For instance, let R be the ring of polynomials in n variables over the complex numbers, R=C[t,,...,f,]. Let J be an ideal of R, and define zeros of J to be n-tuples of complex numbers x such that f(x)=0 for all feJ. It is shown in algebraic geometry courses that if f is a polynomial in R which vanishes on Z(J), then Jf eJ for some positive integer m. This is called Hilbert’s Nullstellensatz. Remark 2. Theorem 2.1 is but an example of a type of theorem which describes the topology of a space and describes properties of a space in terms of the ‘ting of continuous functions on that space. (Cf. also Exercise 53 TI topological spaces. I, §3. ASCOLI'S THEOREM In the examples of Chapter XVIII, §4, we shall deal with compact subsets of function spaces, and we need a criterion for compactness, which is provided by Ascoli’s theorem. It is also used in other places in analysis, for instance in a proof of the Riemann mapping theorem in complex analysis. Therefore, we give a proof here in the general discussion of compact spaces. Let X be a subset of a metric space, and Jet F be a Banach space. Let ® be a subset of the space of continuous maps C(X, F). We shall say that ® is (or its elements are) equicontinuous at a point xo € X if given ¢, there exists 6 such that whenever x € X and d(x, x9) < 6, then I) - fo) 0. By equicontinuity, for each x € X we select an ‘open neighborhood V(x) such that if ye V(x), then [f(y) — SQ) 20). 2 Let S be a closed subset of X. Let J be the set of all feR such that f vanishes on S. Show that J is a closed ideal. Assume that X is Hausdorff. continuous functions on S. (We assume that you have had the notion of a factor ring in an algebra course.) 3. Let X be a compact space and let J be an ideal of CLA). If the set of zeros of tis the complex case.)10. CONTINUOUS FUNCTIONS ON COMPACT SETS CHIT, §4] 1. Let X be a compact Hausdorff space. Show that a maximal ideal of C(X) has only onc zero, and is closed. (Recall that an ideal M is said to be maximal if M # C(X), and if there is no ideal J such that Mc Jc C(X) other than M and C(X) itself) Thus if M is maximal, then there exists p¢ X such that M consists of all continuous functions f vanishing at p. 3. Let X be a normal space, and let R be the ring of continuous functions on X. Show that the topology on X is the one having the least amount of open sets making every function in R continuous. Give a Taylor formula type proof that the absolute value can be approxi- mated uniformly by polynomials. First, reduce it to the interval {~1, 1] by multiplying the variable by c or c™! as the case may be. Then write |t|= Jt. Select 6 small, 0<6 <1. If we can approximate (t? + 5)", then we can approximate ,/t?. Now to get (t? + 6)" either use the Taylor series approximation for the square root function, or if you don’t like the binomial expansion, first approximate log(t? + 4)" = dtog(e? + 6) by a polynomial P, Then take a sufficiently large number of terms from the Taylor formula for the exponential function, say a polynomial Q, and use Q©P to solve your problems. . Give another proof for the preceding fact, by using the sequence of poly- nomials {P,}, starting with Fo(t) = 0 and letting Prralt) = Plt) + 4(t— Py). Show that {F,} tends to ,/t uniformly on (0, 1}, showing by induction that os Vi-nos, V4 2+nfe whence 0S \/t ~ P(t) $ 2/n 3. Look at Example 1 of Chapter VIII, §3 to see another explicit way of proving Weierstrass’ approximation theorem for a continuous function on a finite closed interval. Do Exercise 1 of that chapter. ). Let X be a compact set in a normed vector space, and let {J} be a sequence of co and ! Sy and such that { f,} is a monotone increasing sequence, Show that the convergence is uniform (Dini’s theorem; cf. Chapter 1X, §1). Let X be a compact metric space (whence separable). Show that the Banach space C(X, R) or C(X, C) of continuous functions on X is separable. (Hint: Let {x,} be a countable dense set in X and let g, he the function on X given by Galx) = d(x. X,),(1, §4] EXERCISES 61 where d is the distance function. Use the Stone-Weierstrass theorem applied to the algebra generated by all functions g, to conclude that C(X,R) is separable.] Note: Since a compact Hausdorff space is normal, and since a normal separable space is metrizable, one can adjust the statement of the theorem proved in the exercise as follows: Let X be a compact Hausdorff separable space. Then C(X, R) is separable. IL, Let X, Y be compact Hausdorff spaces. If f, g are continuous functions on X and ¥ respectively, we denote by f@g the function such that (F@ Mx y) = fx)aly). Show that every continuous function on X x Y can be uniformly approxi- mated by sums J f-, f,®@g, where f, is continuous on X and g, is continuous on BR Lei X be compact Hausdorit, By an aigebra auiomorpiism of C(x) we mean a map 0: C(X)— C(X) such that o leaves the constants fixed, and satisfies Aft Q=of+o(9, (fg) = o(fo(g). =I. 13, Let X be a compact Hausdorff space and let A be a subalgebra of C(X,R). Show that there exists a continuous map g:X—Y of X onto a compact space Y such that every element of A can be written in the form go g, where @ is a continuous function on ¥. 14. Let X, Y be compact Hausdorff spaces. Show that X is homeomorphic to Y if and only if C(X, C) is algebra-isomorphic to C(¥, C). 15. Let X be a compact Hausdorff space. Let be the set of all maximal ideals in C(X, ©) Define a closed set in 4 to consist of all maximal ideals con- taining a given ideal. Show that this defines a topology on “. For each xeX, let M, be the ideal of functions in C(X,C) which vanish at x. Show that the map Show that an algebra automorphism is norm preserving, ic. llof| xteM, is a homeomorphism between X and 16. For aR let f,(x) = e“*e-*". Prove that any function @ which is C* and has 1n he uniformly approximated hy elements of the space generated by the functions J, over C. (Hint: If y is a function van- ‘outside a compact set, and IV is a large integer, let yy, be the extension of ¥ on [—N,N] to R by periodicity. Use the partial sums of a Fourier series to approximate such an extension of g(x)e*, and then multiply by e-"] Remark, Instead of e~** you could use any function h(x) > 0 which is C®, and tends to 0 at infinity. This would not be the case in Exercises 19 and 20 below. The next four exercises form a connected set.62 CONTINUOUS FUNCTIONS ON COMPACT SETS (HHT, §4] 17, Let X be compact Hausdorff and let p be a point of X. Let A be a subalgebra of C(X,R) consisting of functions g such that g(p) = 0. Assume that there is no point q#p such that g(q)=0 for all ge A, and that A separates the points of X — {p}. The the uniform closure of A is equal to the ideal of all functions vanishing at p. 18, Let X be locally compact Hausdorff, but not compact. Let C.,(X,R) be the algebra of continuous functions f on X such that f vanishes at infinity (meaning, given ¢ there exists a compact K such that |f(x)| 0 instead of e~* and e~” in Exercises 19 and 20. 21, Let X be a metric space and E a normed vector space. Let BC(X, E) be the space of bounded continuous maps of X into E. Let © be a bounded subset of BC(X, B). For xe X, let ev,:—+E be the map such that ev,(g) = g(x). Show that ev, is a continuous bounded map. Show that ® is equicontinuous at a point aeX if and only if the map xen, of X into BC(®, F) is continuous at a. 22. Let X be a compact subset of a normed vector space, and E a normed vector space. Show that any equicontinuous subset ® of C(X, £) is uniformly equi- continuous. [This means: Given ¢, there exists 5 such that |x — y| <5 im- plies [f(x) — f(y) <6 for all fe ©.) 23. Let X be a subset of a normed vector space and @ an equicontinuous subset of BC(X,R). Let Y be the set of points xe X such that ®(x) is bounded. Prove that Y is open and closed in X. If X is compact and connected, and if for some point aeX the set O(a) is bounded, show that © is relatively compact in C(X, R).PART TWO Banach and Hilbert Spaces ‘The two chapters of this part are absolutely basic for everything else that follows, and introduce the most useful of all the spaces encountered in analysis, namely Banach and Hilbert spaces. The reader who wishes to study integration theory as soon as possible may continue these chapters with Chapter VI, which will make essential use of the basic properties of these spaces, especially the completion of a normed vector space and the linear extension theorem. Indeed, the integral of the absolute value of a function defines a seminorm on a suitable space of functions, whose com- pletion will be the main object of study of the chapters on integration. On the other hand, readers may look directly at the functional anal- ysis, as a continuation of the linear theory of Banach and Hilbert spaces. ‘At some point, of course, these come together when we study the spectral theorems and the existence of spectral measures. As in the algebraic theory of vector spaces, we shall consider continu- ous linear maps L: E+ F of a normed vector space into another. The kernel and image of L are defined as in the algebraic theory, namely the kernel is the set of elements x€E such that L(x)=0. The image is simply L(E). Both Ker L and Im L are subspaces, of E and F respec- dively. Huwever, now iat we have ¢ a closed subspace (being the inverse image of the closed set {0}). Warn- ing: the image if not necessarily closed. For conditions under which the image is closed, see Chapter XV. For the integration theory, we do not need such considerations of subspace and factor space. However, we shall consider the dual space in the context of integration, showing that various spaces of functions are dual to each other. Thus we deal at somewhat greater length with the dual space in this chapter. An application of the duality theory in the context of Banach algebras will be given in Chapter XVI.CHAPTER IV Banach Spaces IV, §1. DEFINITIONS, THE DUAL SPACE, AND THE HAHN—BANACH THEOREM Let E be a Banach space, ie. a complete normed vector space. One can deal with series > x, in Banach spaces just as with series of numbers, or of functions, and the most frequent test for convergence (in fact absolute convergence) is the standard one: Let {a,} be a sequence of numbers 20 such that Ya, converges. If Ixql & @, for all n, then Y x, converges. The proof is standard and trivial. Let E, F be normed vector spaces. We denote by L(E, F) the space of continuous linear maps of E into F. It is easily verified that a linear map A: EF is continuous if and only if there exists C>0 such that JAG) S Clx| for all xe E. Indeed, if the C exists, continuity is obvious (even uniform continuity). Conversely, if 2 is continuous at 0, then there exists 5 such that if Ix! <6, then [A(xil<1. Hence for any non-zero xe E, we get ®) Al be whence we can take C = 2/6. Such a number C is called a bound for 4, and 2 is also said to be bounded. Let S, be the unit sphere in E (centered at the origin), that is 0 such that AGS Clix, all xe EB. Let E, F, G be normed vector spaces, let w€ L(E, F), and let ve L(F, G). Then vow is in L(E, G) and we have Jvoul Solu. Proof. A composite of continuous maps is continuous, and a compos- ite of linear maps is linear, so our first assertion is clear. As to the second, we have Jv 0 ue(x)] = Jo(u(x))] S Fell] S Tel el xh, so the desired inequality follows by definition. If F is complete, then L(E, F) is complete. This is but an exercise. If {4,} is a Cauchy sequence of elements in L(E,F), then for cach x€E one verifies that {4,(x)} is a Cauchy se- quence in F, and hence converges to an element which we define to be A(x). One then verifies that A is linear, and that if C= lim|A,|, then C is a bound for 4, so that 4 is continuous. Finally one verifies that {4,} converges to 2 in L(E, F). (Fill in the details as Exercise 1, or look them up in Undergraduate Analysis.) We give some terminology concerning the space L(E, F) which is used constantly in this book, and in analysis. A continuous (bounded) linear map of a Banach space into itself is called an endomorphism, or an operator. In the case of two spaces E, F, an element ue L(E, F) is said to be invertible if there exists ve L(F, E) such that uov=l, and vow=[g stance a map having a refers to invertibility in various contexts, for[tV, §1] DEFINITIONS, THE DUAL SPACE 67 continuous inverse, a finear inveise, a Thus in each case, one should add an adjective to the word isomorphism to make precise the kind of invertibility which is meant. In our present case, we shall call invertible elements of L(E, F) toplinear isomorphisms, the adjective toplinear referring to the topology and the linearity. The set of toplinear isomorphisms of E onto F is denoted by Lis(E, F). If E=F, then we call toplinear isomorphisms of E with itself toplinear automorphisms of E; the set of such automorphisms is denoted by Laut(E). (For euphony, the reader may prefer the adjective topolinear instead of toplinear.) A toplinear isomorphism u between Banach spaces E, F which also preserves the norm (that is Ju(x)] = |x| for all xe E) will be called a Banach isomorphism, or an isometry. We shall also be dealing with bilinear maps. Let E, F, G be normed vector spaces. A map fcreniiabie inverse, cic. ad fib. QgiEX FG is said to be bilinear if for each x € E the map y++ 9(x, y) is linear, and if for each ye F the map x(x, y) is linear. Such bilinear maps form a vector space. It is easily verified (in a manner similar to the case of linear maps) that ¢ is continuous if and only if there exists C such that lox, yl S Clxllyl for all xe E, ye F. The greatest lower bound of such C then defines a norm on the space of continuous bilinear maps, denoted by L(E, F; G), and this space is a Banach space if G is complete. (Cf. Exercise 3.) In the differential calculus, and other applications, we need an isomorphism between L(E,L(F,G)) and L(E,F;G) as follows. Let Ae L(E, L(F, @)) and define p, by Pals ¥) = AYY)- Then g, is obviously bilinear, and we have leas YE AGH Iyt SII so that loal Sl. On the other hand, given ¢ € L(E, F; G), we can define A, by ALY) = 90% ¥)- Then le) S lollxIy168 BANACH SPACES Uv, §1] so that by definition, AG S ella} Hence lag] S ll. Thus we get a Banach isomorphism L(E, L(F, G)) o L(E, F; G). As one example of a bilinear map, we have L(E, F) x E> F QL Ey XX By is a map which is linear in each variable. Such a map is continuous if and only if there exists C such that for all x,¢ E, we have 19(%15 ++-% nl S Cla ba) 1a We have a norm-preserving isomorphism L(E,, L(Ea, ---5L(Em F).-.)) + L(Ey,---E ns F) from the space of repeated continuous linear maps to the space of con- tinuous multilinear maps exactly as in the bilinear case. If F is complete, then all these spaces are also complete. We now consider a specially important space of linear maps. The normed vector space L(E,R) [or L(E,C) in the complex case] is called the dual space of E, and is denoted by E’. Elements of E’ are called functionals on E. Functionals can be used as substitutes for coor- dinates. Indeed, suppose that E = R*, and let 4, be the i-th coordinate function, that ts AX -+2Xn) Then it is easily verified that {A,,...,4,} is a basis for the dual space of R‘. Furthermore, the values of 4,, ...,4, on an element x € R* character- ize this element, Although we do not have such convenient bases in the infinite dimensional case, we still ha of E in terms of the values of functionals. This i is based on the following theorem,OY, §1] DEFINITIONS, THE DUAL SPACE 0 Theorem i.i. Let £ be a real normed vector space, and iet F be a subspace. Let 3: F +R be a functional, bounded by a number C > 0. Then there exists an extension of A to a functional of E, having the same bound. Proof. Changing the norm on F (multiplying it by a number) we see that it suffices to prove our theorem when C = 1. We first prove that if ve E and v¢ F, then we can extend A to F + Ro, and preserve the bound 1. Every element of F + Ru has a unique expression as x + tv with xe F and teR. Let ae R. The map 4* on F + Rv such that I(x + tv) = A(x) + ta is certainly linear. We must show that we can select a such that A* is bounded by 1. Dividing both sides by ¢ (if t #0), we see that it suffices to find a number a such that AQ) +a) Sly +0) for all ye F, or equivalently that for all ye F, Ay+aS|ytvo| and —Ay)—aSly +o} This determines inequalities for a, namely Ay) ly + ol Sas -Ay) + ly + oh and it suffices to show that the sct of real a satisfying such inequalities is not empty. But for all y, z€ F we have AQ) — 4@) = Ay — 2 Sly — 21 so that —A() — z+ ol S Ay) + 1y + ob From this we conclude that there is a non-empty interval of values of which satisfy our requirements, ‘We now use Zorn’s lemma. We consider the set of pairs (G, 2*) where G is a subspace of E containing F, and 2* is a functional on G having the same bound as 2, and extending 2. We order such pairs (Gi, 41) S (Ga Aa) if G, is a subspace of G, and Az is an extension of 4. This is an ordering, and our set of pairs is inductively ordered. The proof of this is the usual proof: Given a totally ordered set of pairs as above, say {(G;, A)}, we let G be the union of ali G. We can define a functional 2*10 BANACH SPACES UY, $1) on G extending ali 4,: Any x€G 1s in some Gj, and we define A*(x) = A(x). This is independent of the choice of i such that x ¢ G,, and the pair (G, 2*) is an upper bound for our family. By Zorn’s lemma, let (G, 4*) be a maximal element. Then G = E, for otherwise, there is some ve E, v¢G, and we can use the first part of the proof to get a bigger pair. This proves our theorem. Corollary 1.2. Let E be a normed vector space, and ve E,v #0. Then there exists a functional 4 on E such that H(v) #0. Proof. Let F be the one-dimensional space generated by v. We define 4 on F taking any non-zero value on v, and extend A to E using Theorem 1.1. Theorem 1.1, or its Corollary, is referred to as the Hahn-Banach theorem. We have formulated it over the reals, but it is also valid for complex Banach spaces, and the complex case is casily reduced to the real case. Indeed, given a complex functional 4 on a complex subspace F, let @ be its real part, Let op" be a real extension of y to E, and define A) 9") —ip'v) — forve E. You can verify as Exercise 2 that 1’ is a desired complex extension of 2. The dual space £’ is a special case of the space of linear maps L(E, F) when F is the space of scalars. As such, we have seen that it is a Banach space with its natural norm. Furthermore, we can form the double dual E” in a similar fashion, and E” is also a Banach space. Note that each element x € E gives rise to a functional f, € E”, given by Se E'—scalarsR or — such that (A) = A(x), continuous for the topology defined by the norm on E’. Proposition 1.3. The map x++f, is an injective linear map of E into E", which is norm preserving, i.e. |x| = |f-l Proof. Suppose x, ye E and x #y. Then x—y £0. Ry the Hahn Banach theorem, there exists A€ E’ such that A(x — y) #0, so A(x) # A(y). ‘This proves that f, # f,, whence the map x1 f, is injective. ‘The inequality JAG} S lllx] shows that |/,|S|x| We leave to the reader the opposite inequality ' i at we have au i i Ud, hat we have au isumetric em- bedding of E in E”.CV, §1] DEFINITIONS, THE DUAL SPACE n In Chapter Ii, §1 we defined the weak topoiogy on a space, deter- mined by a set of mappings into a topological space. We now apply this notion to the dual space. We let # be the family of functions on E’ given by F =(hdsck as above. The weak topology on E’ determined by this family # is called simply the weak topology on E’. The next theorem gives one of its most important properties. Theorem 1.4 (Alnoglu’s Theorem). Let E be a Banach space, and let E be the unit ball in the dual space E'. Then E, is compact for the weak topology. Proof For each ve F, let K, he the closed dise of ras K=[[K fist be the Cartesian product of all closed discs of radius 1, taken over all xeE satisfying |x| <1. We give K the product topology, so that by Tychonoft’s theorem, K is compact. We map Ej into K by the map SEK such that Ars TT] 40) = TJ £0) Immediately from the definition, one sees that the map f is injective, and thus gives an embedding of E, into the product space. Furthermore, also from the definition of the weak topology defined in Chapter II, §1, we observe that the weak topology determined by the family F is the same as the weak topology determined by the family F, of functionals f, with x€E, (the closed unit ball in E), because any xe E, x #0 is a scalar multiple of a unit vector. More precisely, we also have an imbedding fiES Ic given by = Ars T] 20, and the following diagram is commutative: Bec. TC ek ii.2 BANACH SPACES OV, §2] The product topology induced on || K, is the same as the topoiogy induced by viewing this product as a subspace of []C,. Therefore, it follows that the weak topology on E', is the topology induced by viewing E, as a subspace of K via the embedding f, or also as a subspace of TC. (eB), via the embedding of E’ in []C,. To show that E; is ‘compact, it suffices therefore to show that f(E}) is closed in K. To do this, we first prove that E’ is closed in []C.(x€E). Let T] x) (xe E) be an element of the product which lies in the closure of SUE). Given elements x, ye E, we have to show that xt+7(x) is a bounded functional. By definition of the weak topology, given there exists A€ E’ such that Hx) = yO <8, AQ) — WW) WE) oF is continuous and linear (it even preserves the norm) and consequently, by the linear extension theorem, it has a unique continuous linear exten- sion of E into F, which we denote by 2. Similarly, the continuous linea map pow: WE) o(E) cE has a continuous linear extension of F into E, which we denote by u. ‘Then pro A: E+E gives the identity when restricted to (E), and hence is equal to the identity on E itself by continuity (or by the uniqueness part of the lincar extension theorem). Similarly, 10: FF is the identity. This proves the uniqueness of the completion. We observe that our toplinear isomorphism A preserves norms, that is lax] = Ix] for all xe E. This again follows by continuity. We shall now give two proofs of the existence of a completion. So let E be a normed vector space and let E’ be its dual. As we saw in Proposition 1.3, we have a natural norm-preserving injection E->E". But E” is complete because E” = L(E’, F) with complete F (F = scalars). ‘of E ie ei Ein F" (Do Exercise 15) Next we give another proof, based on the same construction as the8 BANACH SPACES Ov, §4) Teal numbers from the rational numbers. {his construction be used in the integration theory. See the examples after the construction. The Cauchy sequences of elements of E form a vector space, which we denote by S, As usual, we have the notion of null sequences, that is sequences {x,} in E such that given ¢, there exists N such that for all n> N we have |x,| <¢. The null sequences form a subspace. We define two Cauchy sequences £ = {x,} and = {y,} to be equivalent if there exists a null sequence «= {a,} such that €=1 +a (in other words = yp, +a, for all n). This is an equivalence relation, and we denote the equivalence class of & by Z. Then the equivalence classes of Cauchy sequences form a vector space in a natural way, and we have (for c R): Gen ley and Raed. We denote the vector space of equivalence classes of Cauchy sequences by E. (It is nothing but the factor space of Cauchy sequences modulo the subspace of null sequences.) If E = {x,} is a Cauchy sequence and 9 = {y,} is equivalent to E, then lim |x| = lim |y,1. Then we define {| = lim |x, It is verified at once that this is a norm of E, which is thus a normed vector space. We let EE be the map such that g(x) is the class of the Cauchy sequence {x, x, ...}. Then is clear that_p is linear, and preserves norms. Furthermore, one sees at once that if 2 is the class of a Cauchy sequence €, and x = {x,}, then F = lim ox,). Hence g(E) is dense in E. All that remains to prove is that E is complete. To do this, let {Z,} be a Cauchy sequence in E. For each n there exists an element x, €E such that [En 9x41 < In, The scyueuce {x,} is then Cauchy (in £).OV, $4] COMPLETION OF A NORMED VECTOR SPACE 9 Indeed, we have 19%, — OXml 19%, — EN + 1B, — End + En — Xm Ixq = which gives a 3e-proof of the fact that {x,} is a Cauchy sequence. Let § = {xy}. Then {€,} converges to & because given &, 1, — 81S 16, — 9x4 + 19X_— 8 < 2e for n sufficiently large. This proves that proof for the existence of a completion of E. complete, and concludes the Example 1. In integration theory, covered later in this book, one Fo, 11 fo, 3, 1 Ill, = If) at. 0 One can also take the vector space of continuous functions on R, van- ishing outside some bounded interval, and define the L!-norm similarly. Then this space is not complete, and its completion is called L'. It then becomes a problem to identify elements of L* with certain functions, and this is what we shall do. Example 1 points to the need of a slight generalization of our normed vector spaces. Indeed, even in elementary integration theory, one deals with step functions, or piecewise continuous functions, which are such that if If], =0, then f may not be the zero function. For instance, if f is 0 except at a finite number of points, then we do have |[/], =0. In view of this, one defines a seminorm on a vector space E to be a function satisfying all properties of a norm, except that we require Ixl20 for all xeF, but we allow Ix!=0 without having necessarily x =0. Then it is clear that the set of all x€E such that |x| =0 is a subspace Eo. The terminology of open and closed sets applies in the present context, and the topology defined by a seminorm is simply not Haus- dorff. In fact, the closure of 0 is obviously the space Eo itself In defining the completion, we can just as well define the comple- of a space with a80 BANACH SPACES OV, §4] sequences, and we still get a map jE +E, the only difference being that j has a kernel, which the reader will verify to be precisely Ey. In fact, we have a norm on the factor space E/Ey if we define the norm of a coset |x + Eol to be |x| (independent of the coset representative x since we have tet yl = Il for all y€E). Thus we can say that if E has a seminorm, the comple- tion E is simply the completion of E/E as discussed in this section. A vector space E with a seminorm | | can be called a seminormed space, We can define Cauchy sequences using the same definition as in the normed case. We shall say that E is complete if every Cauchy sequence in E converges—in other words, if given a Cauchy sequence {x,} in E, there exists xe E such that given e, there exists N such that for all n = N we have bey al siiuaiiun. Let E be a normed vector space. Elements of L(E, E) are also called operators on E. Let S be a set of operators on E. By an S-invariant subspace F we mean a subspace such that for every Ae€S we have AF c F, ic. if xe F and AcS, then Axe F. It is clear that if F is an S-invariant subspace, then its closure is also S-invariant because if x, ¢ F and x, x, then Ax, — Ax, so Ax lies in the closure of F. An operator B is said to commute with S if AB = BA for all AcS. If B commutes with S, then both the kernel of B and its image are S-invariant subspaces. Proof. If xe E and Bx — 0, then ABx — BAx = 0 for all A eS, so the kernel of B is S-invariant. Similarly, also from the relation ABx = BAx, we see that the image of B is S-invariant. If A is an operator on E, and co operator fq, are numbers, we may form the P(A) = ¢,A" +--+ Col, where POO) = et + +9 is the polynomial having the numbers as coefficients. If p, q are polyno- mials and pq denotes the ordinary product of polynomials, then we have (p + (A) = (A) +.q(A) and (pq)(A) = p(A)a( A). Indeed, if g(t) = bt” + °** + bo, then pitvatt) = Yat’,82 BANACH SPACES LIV, §5) where a= Yb But P(A)q(A) = Yd," since associativity, commutativity, and distributivity hold in multiplying powers of A. The statement concerning the sum p+q is even more trivial to see. Also, if ¢ is a number, then (cp)(A) = cplA). All these rules are useful when considering the evaluation of polynomials on operators. In algebraic terminology, they express the fact that the map r+ P(A) is a ring-homomorphism from the ring of polynomials into the ring of operators. If F is an A-invariant subspace, then it is clear that F is also p(A)}- invariant for all polynomials p. Thus if F is in fact a subspace of E which is invariant for an operator A, then it is also invariant for the set of all polynomials in A, called also the ring of operators generated by A. The same holds for any set of operators S, letting the ring of operators generated by S be the set of all operators expressed as finite sums Leet Ay -+4q ate elements of S, and the coefficients are numbers. F is A- and B-invariant, then it is also (A + B)-invariant and AB-invariant. If an operator B commutes with all elements of S, then it is clear that B also commutes with all elements in the ring of operators generated by S, because if B commutes with A, and A,, then B commutes with A, + Aq and also with A,A,. Furthermore, if F is a closed subspace and is S-invariant, then it is also S-invariant, where 5 is the closure of S. Indeed, if {R,} is a sequence of operators in § conver operator B, and if xe F, then the sequence {B,x} is Cauchy, and hence converges to Bx which lies in F. In Chapters XVII and XVIII we study a pair (E, A) consisting of a space E and an operator A, and analyze this pair, describing its structure completely in important cases. The idea is to apply in the present con- text an all: pervasive point of view in mathematics, which is to decompose jouis. in ihe present context, iet g to some us make some general definitions.IV, App] THE KREIN-MILMAN THEOREM 83 Let E be a Banach space, and F, G closed subspaces. We know that the product F x G consisting of all pairs (y,z) with ye F and zeG is also a Banach space, say under the sup norm. If the map FxG+E given by Oarytz is a toplinear isomorphism, then we say that E is the direct sum of the subspaces F and G. Observe that our requirements involve both an algebraic and a topological condition. It follows from our conditions that E=F+G and FoG={0}. © proved iater that, in fact, these two conditions are sufticient; in other words, if they are satisfied, then the map (naeyte not only has an algebraic inverse, but this inverse is continuous (corol- lary of the open mapping theorem). When E is a direct sum of F and G, we write E=F@G. If A is an operator on E, then we are interested in expressing E as a direct sum of A-invariant subspaces. Subsequent chapters give examples of this situation. APPENDIX: CONVEX SETS APP., §1. THE KREIN-MILMAN THEOREM Although we shall not use the theorem of this section latcr in the book (except for some exercises). it is worthwhile giving it since it is used at the beginning of more advanced and specialized courses, in a wide variety of contexts. The exposition follows that of Artin (cf. Collected Works). Throughout this section, we let E be a vector space over the reals (not normed). We let E* be a vector space of linear r maps of E into R (nor 1at E* coparates E, that is given xe E, x#0 there exists: Je E* “auch that 1G) #0. We84 BANACH SPACES 1, App.] give E the topology having the smailest amount of open seis muking uit A«€E* continuous. A base for this topology is therefore given by the following sets: We take xe E, and A,,...,4,€E*, and ¢>0. We let B be the set of all y¢E such that IAQ) - AG) < & The set of all such B is a base for the E*-topology. A subset S of E is said to be convex if given x, y €S, the line segment (l-ox+ty, OsStS1, joining x to y is contained in S. We observe that an arbitrary intersection of convex sets is convex. Lemma 1.1. Let x,, ...,%,€5S. Any convex set containing xy, ...,X, also contains all linear combinations tye bob x, with 0S $1 for all i, and t, +++ +t, = 1. Conversely, the set of all such linear combinations is convex. Proof. If t, #1, then the above linear combination is equal to “The first assertion follows at once by induction. The converse is also an immediate consequence of the definitions. The following properties of convex sets also follow at once from the definitions. Let A: E- F be a linear map. If S is convex in E. then MS) is convex in F. If T is convex in F, then 4*(T) is convex in E. In other words, the image and inverse image of a convex set under a linear map are convex. Let A€ E*, 2#0, and let Ho be the kernel of A (ie. the set of all xe E such that A(x) = 0). Then Ho is a closed subspace, and if ve E is such that I(r) #0, E=H,+Rv.CIV, App.] ‘THE KREIN-MILMAN THEOREM 85 If A,, A, are non-zero functionals with the same kernel H, then there exists ceR, c #0 such that 4,=cA,. Indced, one sces at once that c=A,Wi/Aa(v). Let 40 be an element of E*, and let ce R. By the hyperplane H, we mean the set of all x €£ such that A(x) =c. In other words, H, = 4-*(c). If Ho is the kernel of 4, then H, consists of all elements y+ yo with y €Hp and yo any fixed element of E such that A(yo) = c. The set of x €E such that A(x)2c will be called a closed half space determined by the hyperplane, and so will the set of all x such that A(x) $c. Similarly, we have the open half spaces, determined by the inequalities A(x) > c and A(x) 0. But C is convex. Therefore C contains the sct A of elements (—a + x)/2 with x € B. It is easy to see that A contains the ball centered at the origin and of radius r/2. This and the fact that C is a cone imply that C = E, a contradiction. It follows that H = Cn(—C) is closed, is a cone, is convex, and H = —H. Therefore H is immediately seen to be a closed subspace. We have H#E because —a¢C, so —a¢H. We have E=Cu(-Q. To see this, let xeE and suppose x¢C, x¢—C. Since C is maximal, the cone consisting of all elements ¢ + tx with ceC, t20 contains —a, and so does the cone of all elements c+t(—x), c€ C,t 20. Hence we can write ma sey +X = Cy — tak with c, c,€C and t,, t, 20. Consequently ey + lt, + xe.90 BANACH SPACES UV, App. owever, Cp + tyx — 1 (t, + t2)x, and thus lies in C, a contradiction which proves that E = Cu(-O). Now suppose that xe¢C. Then the line segment between x and —a contains a point of H. For instance, on the sogment x + t(—a — x) with died 5 let t be the sup of all ¢ such that x + (+a — x) lies in C. xe (nx) fies ta Hy tid ty ty otherwise ce M1, whitch Iv ree ‘Wa therefore have (l—)x—ta=heH, whence Working also with —x instead of x. we conclude that E is generated by H and a, so that the factor space E/H has dimension 1 and hence H is a closed hyperplane. Finally, H does not intersect U, for otherwise let he HU. Since U is open, for small s > 0 we have h—saeU so h—saeC. But —heC, whence —saeC and —aeC, which is impossible. This proves our lemma. We now prove: Let b be a point of a Banach space E, which does not belong to the norm-closed non-empty convex set A. Then there exists a functional A and a number « such that A(x)>o for all xe A and 2) 0 for all z¢ A— B, and therefore Ax 2 Ay for all xeA, yeB. Let B= inf dx for xe. The map A is an open map- ping, for instance because 4 gives an isomorphism of a one-dimensional subspace of E univ R. Therefore 4y 0 such that 1G, WS Clxilyl for all xe E. ye F. Let L(E, F;G) be the set of continuous bilinear maps of if the norm of numbers C as above. Show that if G is complete, then L(E, F; G) is complete. 4. Let E be a Banach space and F a closed subspace. For each coset x + F of F, define |x + F] = inflx + yl for ye F. Show that this defines a norm on the factor space E/F, and that the natural map E-» E/F is continuous linear. (Cf. Chapter XV, §1.) 5. Let A be a Banach algebra. Suppose that there is a unit element ¢ #0, but that we do not necessarily have |e] = 1. {a) Show that |e] 2 1. (b) Define a new norm || [| on A by putting bt =p Show that [| || is in fact a norm and that [ell = 1. (©) Show that A is a Banach algebra under this new norm. 6. (a) Show that a finite dimensional subspace of a normed vector space is closed. (b) Let E be a Banach space and F a finite dimensional subspace. Show that there exists a closed subspace G such that F 4 G= E and FoG= {0}. ‘You will have to use the Hahn-Banach theorem. 7. Let F be a closed subspace of a normed vector space E, and let ve E, vg F. Show that F + Ro is closed, If E=F + Ro, show that E is the direct sum of F and Rv. (You can give a simple ad hoc proof for this. A more general result will be proved later as a consequence of the open mapping theorem)92 BANACH SPACES QV, §6] E, F, G be nor ¢ that G is complete. Let x F-+G be a continuous bilinear map. Show that A can be extended to a continuous bilinear map of the completions E x F + G, which has the same norm as 2. (Identify E, F as subspaces of their completions.) 9. Let A be a Banach algebra, commutative, and with unit clement. Let J be an ideal. Show that the closure of J is also an ideal. (The definition of an ideal is the same as in the case of rings of continuous functions. If the algebra is not commutative, then the same result is valid if we replace ideals by left ideals.) 10, Let A be a commutative Banach algebra and let M be a maximal ideal, Show that M is closed. 11. Give the proof of the inequality left to the reader in Proposition 1.3. 12, Let E be an infinite dimensional Banach space, and let {x,} be a sequence of linearly independent elements of norm 1. Show that there exists an element ihe civsure of ihe space yeneraied by ali x, which does nui jie in any subspace generated by a finite number of x,. [Hint: Construct this element as an absolutely convergent sum Y° ¢,x,-] 13. Let {E,} be a sequence of Banach spaces, Let E be the set of all sequences €={x,} with x,eE, such that > |x,| converges. Show that E is a vector space, and that if we define Wl= Lbs) then this is a norm, and E is complete. 14, Let E be a Banach space, and P, Q two operators on E such that P+ Q=1, and PQ = OP = 0. Show that E=KerP+KerQ, and that Ker P=ImQ. Show that Ker Pry Ker Q = {0}, and that Ker P and Im P are closed subspaces. 15. Let E be a Banach space and let F be a vector subspace. Let F be the closure of F. Prove that F is a subspace, and is complete. 16. Let A be a subset of a Banach space. By c(A) we denote the convex closure of A, ie. the intersection of all convex sets containing A. We let 2(A) denote is also compact. [Hint: Show c(K) is totally bounded as follows. First find a finite number of points x;, ...,x, such that K is contained in the union of the balls of radius 6 around these points. Let C be the convex closure of the set {x;,....%,}- Show that C is compact, expressing C as a continuous image of a compact set. Let ys, ...,Yq be points of C such that C is con- tained in the union of balls of radius ¢ around these points. Then get the desired result.]EY, §6] EXERCISES 93 17. Let F be the complete normed vector space of continuous periodic functions on [—2, x] of period 2, with the sup norm, Let E be the vector space of all real sequences a = {a,} such that Y|a,| converges. Define tel = ¥ tah Show that this is a norm on E. Let La(x) = ¥, a, cos nx, so that L: E— F is a linear map. Show that L has norm 1. Let B the closed unit ball of radius 1 centered at the origin in E. Show that L(B) is closed in F. [Hint: Let {f,} (k=1,2,...) be a sequence of elements in L(B) which converges uniformly to a function f in F. Let b, be the Fourier coefficient of f respect to cosnx. Let 6 = {b,}. Show that B is in E and that L(B)=f 18, Let K be a continuous function of two variables defined for (x,y) in the square [a, 6] x [a,b]. Assume that ||K|| $ C for some constant C > 0, where Il [Lis the sup norm. Let E be the Banach space of continuous functions on [a,b], and let T: E+ E be the linear map such that . a0 = [ (6, x)g(0) dt. Show that T is bounded and [|TI| C denoted by Os Yr Os y> which is linear in its first variable, and semilinear or antilinear in its second variable, meaning that for x, y, yi, 2 €E, 2€C, we have Os vty = Oo + Osa) and — Cx, ay) = Ex y). If in addition we have for all x, ye E Y=, we say that the form is hermitian. if furthermore we have 0ifx #0. We shall a assume through out that our form ¢ , > is positive, but not necessarily definite, We ob- serve that a sesquilinear form is always R-bilinear. We define v to be perpendicular or orthogonal to w if = 0 for all ve E. Proof. Let t be real, and consider OS = (v, 0) + 2t Rec, wy. If Re. Then Cv, v) + 2t Rev, w) is negative, a contradiction. Hence Reto, w) = 0. This is true for all ve E. Hence Rediv,w)=0 for all ve E, whence Im(v, w) = 0. Hence + Bw, Bw) = av, v) + BaKw, v) + aBKv, w) + BBCw, w). Note that « = |w|?. Substituting the values for a, B, we obtain 0 S |wl* |v? — 2hwl?)?. Hence IW? Ido, WP S bwI*]01. If |w|=0, then weE by Theorem 1.1 and the Schwarz inequality is obvious. If |w| #0, then we can divide this last relation by |w/?, and taking the square roots yields the proof of the theorem. Theorem 1.3, The function v+-+|v| is a seminorm on E, that is: We have |v| 2 0, and |v| =0 if and only if ve Eo. For every complex a, we have |av| = |al|0} For v, we E we have |v + w| $ Jv] + |wh Proof. The first assertion follows from Theorem 1.1. The second is left to the reader. The third is proved with the Schwarz inequality. It suffices to prove that Jo + wh S ({o| + fw). To do this, we have |v + wl? = Cv + w, 0 + W) = C0, 0) + Cy D> + Cv, W) + CW, WD. But ]. Hence by Schwarz, Jv + wl? S |v? + 210, w>] + Iwi? Slo? + 2|o|}w] + bwl? = (lol + |wl)?. Taking the square root of each side yields what we want. We call | | the L?-norm (or we should really say the L-seminorm). An clement of E is said to be a unit vector if |v] = 1. If |v] #0, then v/|v| is a unit vector. Let we E be an element such that |w| #0, and let ve E. There exists @ unique number c such that v— cw is perpendicular to w. Indeed, for v — ew te be perpendicular to + . & Yy N Figure 5.1 We let a= t, with ¢ real #0. We can then cancel ¢ and get a contradiction for small ¢. if ] & bre ft must approach x also. If x= x A with x,¢, then we apply the projection P, (which is continuous!) to conclude that P,x = x,, thus proving the uniqueness. It is convenient to call the family (F} an orthogonal decomposition of F in the preceding theorem. If F =E, then we call it an orthogonal decomposition of E, of course. Suppose that the Hilbert space E has a denumerable total family {v,}, which we assume to be orthonormal. Then every element can be written as a convergent series where a, is the Fourier coefficient of x with respect to v,, and the convergence is of course with respect to the L?-norm. Namely, we take the spaces F, in the previous discussion to be the 1-dimensional spaces generated by v,. In particular, we see that Y Ja,|? converges, and that Siar. IxP If {v,} is merely an orthonormal system, not necessarily a Hilbert basis, then of course we don't get the equality, merely the inequality Is la, $ IxP. 4 This is called the Bessel inequal , and it is essentially obvious from