Ms'nifolds, Tensor Analysis, and Applications
RAt.PH ABRAHAM
Department of Mathematics Univei'sity of California, Santa Cruz
Santa Cruz, California JERROLD E. �ARSDEN
Department of Mathematics University of California, Berkeley Berkeley, California
TUDOR RATIU Df!partment of Mathematics University of Arizona Tucson, Arizona
1983 ADDISON-WESLEY PUBLISHING COMPANY, INC.
Advanced Book ProgramjWorld SCience Division Reading, Massachusetts London· Amsterdam. Don Mills, Ontario· Syd n ey. Tokyo
Some figures and text reproduced from "Foundations of Mechanics," 2nd Edition, by Ralph Abraham and Jerrold E. Marsden, (') 1978 by Addison-Wesley Publishing Company.
Library of Congress Cataloging In Publication Data Abraham, Ralph. Manifolds, tensor analysis, and applications. (Global analysis, pure and applied) Bibliography: p. Includes index. 1. Global analysis (Mathematics) 2. Manifolds (Mathematics) 3. Calculus of tensors. I. Marsden, Jerrold E. II. Ratiu, Tudor S. III. Title. IV. Series. QA614.A28 1983 514'.3 82-13737 ISBN 0-201-10168-8
American Mathematical Society (MOS) Subject Classification (1980): 34. 58, 70,76.93. Copyright· f 1983 by Addison-Wesley Publishing Company. Inc. Published simultaneously in Canada.
Ait rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means. electronic, mechanical, photocopying, recording, or otherwise, with'HJt thr' prtor written permission of the publisher, Addison-Wesley Publishing Company, Inc., Advanced Book ProgramjWorid Science Division, Reading, Massachusetts 01867, U.S.A. ABCDEFGHIJ-HA-S9S7654l
Manufactured in the United States of America
Contents
SarI. Editors' Foreword . ........................... xl Preface . ........................................ xv Background Notation . . . . . . . . . . .................... xix
Topology............................. .. 1 1.1. Topological Spaces . . . . . . . . . . . . . . . . . '. . ... . . . . I 1.2 Metric Spaces. . . . . . . . .. : . . . . . . . . . . . . . . . . . .9 1.3 Continuity ..................... " ........ 14 1.4 Subspaces, Products, and Quotients ............ 18 ) .5 Compactness ............................ 2A 1.6 Connectedness ......... ~ ................. 31
Ch8f»ter 1
Chapter 2
Banach Spaces and Differential Calculus ..... . 37
2.1 Banach Spaces ........................... 37 2.2 Linear and Multilinear Mappings .............. 52 2.3 The Derivative .......................... ; 67 2.4 Properties of the Derivative .................. 75 2.5 The Inverse and Implicit Function Theorems ..... 102 vii
v;;;
CONTENTS
Chapter 3
Manifolds and Vector Bundles ............ . 122
3.1 3.2 3.3 3.4 3.5
Manifolds..................... ......... 122 Submanifolds. Products, and Mappings. '........ 129 Vector Bundles .......................... 133 The Tangent Bundle ...................... 150 Submersions. Immersions and Transversality ..... 161
Chapter 4
Vector Fields and Dynamical Systems ....... 184
4.1 Vector Fields and Flows. . . . . . . . . . . . . . . . . ... 184 4.2 Vector Fields as Differential Operators ......... 207 4.3 An Introduction to Dynamical Systems ......... 240 4.4 Frobenius' Theorem and Foliations ............ 260 Chapter 5 Tensors .................. . . . . . . . . . . . .271
5.1 5.2 5.3 5.4 5.5
Tensors in Linear Spaces ................... 271 Tensor Bundles and Tensor Fields ............. 283 The Lie Derivative: Algebraic Approach ........ 293 The Lie Derivative: Dynamic Approach ......... 304 Partitions of Unity ....................... 309
Chapter 6 Differential Forms ...................... . 323
6.1 6.2
Exterior Algebra ......................... 323 Determinants. Volumes. and the Hodge Star Operator .............................. 335 6.3 Differential Forms ....................... 351 6.4 The Exterior Derivative. Interior Product, and Lie Derivative ........................ 356 6.5 Orientation. Volume Elements. and the Codifferential ........................... 381
Chapter 7 Integration on Manifolds . ................ . 395
7.1 The Definition of the Integral ................ 395 7.2 Stokes'Theorem ......................... 403 7.3 The Classical Theorems of Green. Gauss. and Stokes ................................ 424 7.4 Induced Flows on Function Spaces and Ergodicity .433 7.5 Introduction to Hodge-deRham Theory and Topological Applications of Differential Forms.... 440
CONTENTS
Chapter 8
ix
Applications ........................... 461
8.1 Hamiltonian Mechanics . . . . . . . . ............ 462 8.2 Fluid Mechanics . . . . . . . . . . . . . . . .......... 481 8.3 Maxwell's Equations ...................... 498 8.4 Thermodynamics, Constraints. and Control ...... 511 Appendix A.
Zorn's Lemma and Balre Spaces ......... 522
Appendix B.
The Three Pillars of Linear Analysis ....... 526
Appendix C.
Unbounded and Self-adjoint Operators ..... 530
Appendix D.
Stone's Th~rem ..................... 541
Appendix E.
The Sard and Smale Theorems ........... 551
References ..................................... 562 Index ............... ~ ......................... 570
Foreword
AIM AND SCOPE OF THE SERIES
What II Global Analyall? From ancient times till Newton. mathematics meant geometry and, algebra. Then analysis (now called classical) was born. along with the foundations of physics, engineering, and modem science. Arnong the outstanding events of modem mathematics are the syntheses of these fields along common frontiers. The synthesis of classical analysis and geometry is DOW called global analysis. The Hietory of Global AM.,... .... Ita AppIIcatIOni Importint pioneers in the synthesis of global analysis were Henri Poincar6 (l11Os), Oeorp Birlmofl (19208). Marston Morse '(1930&). and Hassler Whitney (19405). The technical tools of differential topology (19505) made the rmal synthesis possible (19605). Through the efforts of Solomon Lefshetz (1950s), the work of the Russian school (Liapounov. Andronov. Pontriagin) on dynamics became widely known in the west and included in this synthesis. A veritable explosion of new results and applications followed in the 1910s. From the earliest work of Poincare and Liapounov onward. the applications of g~metry and analysis to astronomy. physics. and engin~ring provided the explicit motivation for much of this work. The current form of the theory reflects this pervasive influence in its direct applicability to these
xi
xii
FOREWORD
fields. It has already created new and powerful methods of applied mathematics. which complement existin& tools such as perturbation methods. asymptotics. and numerical techniques. Far from being the exclusive preserve of pure mathematicians. globill analysis has its roots in physical problems and can he redirected to these problems once again. often with startling results. ' Target of the SerI..: The Accessibility of Global Analysis
There is a great contrast between the potential importance of global analysis and the great dirriculty of learning about it. A growing number of scientists of all disciplines have discovered that the techniques of global analysis have important applications in their own fields: they are looking seriously for keys to these techniques. This series will attempt to provide the keys. Needed are books that introduce the basic concepts and their applications. texts that develop the prerequisites for more serious study accessibly and compactly. and advanced monographs which make the research frontier available to a wide audience of scientists and engineers who have acquired these prerequisites. To these ends. this series will deal with such suhjects as Theory Linear algebra and representation theory Calculus on manifolds and bundles Differential geometry and Lie theory Manifolds of mappings and sections Transversal approximations Calculus of variations in the large Dynamical systems theory and nonlinear oscillations Nonlinear actions of Lie groups Applications Classical mechanics and field theory Geometric quantization Hydrodynamics Elastomechanics Econometrics Social theory Morphogenesis Network theory and other topics of pure and applied glohal analysis.
FOREWORD
xiii
AUDIENCES
Sub-Series A. The advanced texts will provide reports on theory or applications from the research frontier in expository style for /ipecialists. or for nonspecialists who have the prerequisite mathematical background. For example. graduate students of science or engineering as well as mathematics will find them manageable. , Sub-Series B. The basic texts will provide a complete curriculum of essential prerequisites. starting with advanced linear algebra and calculu,s. for the advanced texts of Sub-Series A. These texts will be suitable for advanced undergraduate courses in pure and applied mathematics. or as reference works for research in engineering. the ~cicnl:l~~. or mathematics. UNIQUE FEATURES
Through the basic texts (Sub-Series B) covering all the prerequisites in a uniform style and the advanced texts (Sub-Series A) building on this foundation, it will be possible for readers to study the deta,i1ed applications of global analysis to their own fields (as they appear in the series). to form independent evaluations of the new methods, and to master the techniques for their own use if it is justi(ied. Sub-Series B starts from the post-calculus level. in textbook format with worked examples, exercises and adequate illustrations. The series will give a complete library of prerequisites. together with new contributions to global analysis and some outstanding examples of its applications. illustrating the n~w methods in applied mathematics. All the texts will be in English and conform as far as possible to a common notational scheme. ' RALPH ABRAHAM PHILIP JERROLD
E.
J.
HOI.MES
MARSDEN
PREFACE
The purpose of this book is to provide cOre background material in global analysis for mathematicians sensitive to applications and to physicists, engineers, and mathematical biologists. The main goal is to provide a working knowledge of manifolds, dynamical systems, tensors. and differential forms. Some applications to Hamiltonian mechanics, fluid mechanics. electromagnetism. and control theory are given in Chapter 8. using both invariant and index notation. Detailed treatments of these and additional applications are planned for other volumes in the series. The book does not deal with Riemannian geometry in detail or with Lie groups. or Morse theory. These too are planned for a subsequent volume. Throughout the text special or supplementary topics occur in ,boxes. This device enables the reader to skip various topics without disturbing the main flow of the text. Additional background material in the appen4ices is given for completeness. to minimize the necessity of consulting too many outside refetences. We treat finite and infinite-dimensional manifolds simultaneously. This is partly for efficiency of exposition. Without advanced applications. using, say. manifolds of mappings. the study of infinite-dimensional manifolds is hard to motivate. except for its intrinsic interest. Chapter 8 gives a hint of these applications. In fact, some readers may wish to skip the infinitedimensional case altogether. To aid in this we have separated into boxes many of the technical points peculiar to the infinite-dimensional case.
xv
xv;
PREFACE
Our own research interests lean heavily toward physical applications, and the choice of topics is partly molded by what has been and remains useful for this kind of research. Some interesting technical side issues not consistent with these goals have been omitted or relegated to boxes. We have tried to be as sympathetic to our readers as possible, by providing ample interesting examples, exercises, and applications. When a computation in coordinates is easiest, we give it and don't try to hide things behind complicated invariant notation. On the other hand, index-free notation can often provide valuable geometric, and sometimes computational, insight so we have tried to simultaneously convey this navor. The only prerequisites required are a solid undergraduate course in linear algebra and a classical course in advanced calculus. At isolated points in the text some contacts are made with other subjects. For students. this provides a good way to link this material with other courses. These links do not require extra background material. but it is more meaningful when the bond is made. For example. Chapter I (and Appendix A) link with point-set topology, the boxes in Chapter 2 and Appendices B, C, D, are connected with functional analysis, Section 4.3 relates to ordinary differential equations. Chapter 3, Section 7.5, and Appendix E are linked to differential topology and algebraic topology. and finally Chapter 8 on applications is connected with applied mathematics, physics, and engineering. ' This book is intended to be used in courses as well as for reference. The sections are, as far as possible, lesson sized, if the boxed material is omitted. For some sections, like 2.5, 4.2, or 7.5 two lecture hours are required. A standard course for mathematics graduate students, for example, could omit Chapter I and the boxes entirely and do Chapters 2 through 7 in one semester with the possible exception of Section 7.4. The instructor could then assign certain boxes or appendices for reading and choose among the applications of Chapter 8 according to taste. A shorter course, or a course for advanced undergraduates probably should omit all boxes, spend about two lectures on Chapter I for reviewing background point set topology, and cover Chapters 2 through 7 with the exception of Sections 4.4, 7.4, 7.5 and all the material relevant to volume elements induced by metrics, the Hodge star, and codifferential operators in Sections 6.2, 6.4, 6.5, and 7.2. A more applications oriented course could omit Chapter I. review without proofs the material of Chapter 2 and cover Chapters 3 to 8 omitting the boxed materials and Sections 7.4 and 7.5. For such a course the instructor should keep in mind that while Sections 8.1 and 8.2 use only elementary material, Section 8.3 relies heavily on the Hodge star and codifferential operators, while Section 8.4 consists primarily of applications of Frobenius' theorem dealt with in Section 4.4. The appendices, included for completeness, contain technical proofs of facts used in isolated places in the text.
PREFACE
xvii
The notation in the book is as standard as conflicting usages in the literature allow. We have had to compromise among utility. clarity. clumsiness, and absolute precision. Some possible notations would have required too much interpretation on the part of the novice while others. while precise. would have been so dressed up in symbolic decorations that even an expert in the field would not recognize them. We have used hold face symhols to help the reader distinguish objects; for the most part. linear spaces. linear operators and abstract tensor fields are in boldface italics. while manifolds, points. point mappings. and tensor components are in lightface itaIlcs. Strict compliance is not always possible. This usage is only to help the reader distinguish symbols to which no further significance should be attributed. In a subject as developed and extensive as this one. an accurate history and crediting of theorems is a monumental task. espcl:ially when so many results are folklore and reside in private notes. We have indicated some of the important credits where we know of them. but we did not undertake this task systematically. We hope our readers will inform us of these and other shortcomings of the book so that. if necessary. corrected printings wm be possible. The reference list at the back of the book is confined to works actually cited in the text. These works are cited by author and year like this: deRham (1955). During the preparation of the book. much valuable advice was pro- ' vided hy Alan Weinstein. Our other teachers and collaborators from whom we learned the material and who inspired. directly and indirectly. various portions of the text are too numerous to mention individually. We hereby thank them all collectively. Finally. we thank Connie C'alica for her careful typing of the manuscript. RALPH ABRAHAM JERROLDE.
~ARSDEN
TUDOR RATIU
BACKGROUND NOTATION
The reader is assumed to be familiar with usual notations of set theory such as e. u. n and with the concept of a mapping. If A and B are sets and I:. A -- B is a mapping. we often write a ...... I( a) for the e,ffect of the mapping on the element a e A; "iff' stands for "if and only if' (="if' in definitions). Other notations we shall ,use without explanation include the following: R.C Z. Q
real. complex numbers integers. rational numbers AX B Cartesian product R". C " Euclidean n-space. complex n-space (xl •...• x")eR" pointinR" , A\B set theoretic difference I or Id identity map rl(B) inverse image of B under I rf = «x. I(x)~x e domain of f} graph of I infinimum (greatest lower hound) of A c R infA supremum (least upper hound) of A c R sup A basis of an n-dimensional vector space:: el.···.e" kernel and range of a linear transformation T ker T. range T end of an example end of a proof proof of a lemma is done. but the proof of the theorem goes on.
••
These modifications of the Halmos symbol • are notations of Alan Weinstein.
xix
Manifolds, Tensor Analysis, and Applications
CHAPTER
1
Topology
The purpose of this chapter is to introduce just the right amount of topology for later requirements. It is assumed that the reader has had a course in advanced calculus and so is acquainted with open, closed, compact, and connected sets in Euclidean space (see for ex~ple Marsden (197_] and Rudin (1976». If this background is weak. the reader may find the pace of this chapter too fast If the background is under control. the cbapter sbould serve to collect, review, and solidify concepts in a more· general context. • A key concept in manifold tbeory is tbat of a dirlerentiable map between manifolds. However, manifolds are also topological spaces and . differentiable maps are continuous. Topology is the study of continuity in a general context; it is therefore appropriate to begin with it. Topology often involves interesting excursions into pathological spaces and exotic.theorems. Sucb excursions are deliberately minimized here. The examples will be ones most relevant to later developments, and tbe main thrust will be to obtain a working knowledge of continuity, connectedness. and compactness.
1.1 TOPOLOGICAL SPACES Abstracting our ideas about open sets in R", we shall first consider the notion of a topological space. 1
2
1.1.1
TOPOLOGY
Definition. A top%giclll space is a set S together with a colle(·tion 0
0/ subsets called 0fIDt .Is such that
.
(Tl) IZJE fl and S E fl: (TI) 1/ VI' V z E fl. then VI n V2 E (T3) The union 01 any collection 01 open sets is open.
e:
A basic example is the real line. We choose S ... R. with fl consisting of all sets that are unions of open intervals. Thus, as exceptional cases. the empty set IZJ E and R itself belong to fl. Thus (TI) holds. For (TI). let VI and V2 E fl: to show VI n Vz E fl. we can suppose VI n V2 * "". If x E VI n Vz then x lies in an open interval )a l • hll C VI and x E )(12' hzl C Vl . Let lal.bl[nlaz.bz[=la.b[ (so a=max(a l .a 2 ) and h=min(h l .b2 »). Thus x E la. b[ C VI n Vz. Hence VI n Vz is the union of such intervals. so is open. Finally. (TJ) is clear by definition. Similarly. R" may be topologized by declaring a set to be open if it is a union of open rectangles. An argument similar to the one just given shows that this is a topology. called the standard topology on R". Open intervals in . R and open rectangles in Rn are each examples of a basis for a topology; i.e.• every open set is a union of sets in '!1\. Any set S can be topologized in an obvious manner in two ways. The trivial topology on S consists of fl = { IZJ • S}. The discrete topology on S is defined by fl = {A IA C S}; i.e.• r consists of all subsets of S. Topological spaces are specified by a pair (S. ~): we shall. however. just write S if there is no danger of confusion.
e
'i'
1.1.2 DeflnlUon. Let S be a topological space. A set A c S is called cto.d if its complement S \
A is open.
The collection 01 closed sets will be denoted e.
For example. the closed interval [0. I] c R is closed as it is the complement of the open set ]-oo.O[U]I.oo[.
1.1.3 PropollUon. (el)
(e2) (e3)
Proof.
The closed sets in a topological space satisfv:
IZJE e and SE e; . If AI' A2 E t' then AI U A2 E t'; the intersection 0/ any collection of closed sets ;s dosed.
(C1) follows from (TI) since IZJ = S\S. S = S\ 0. The relations S\(A I U A z ) = (S\A I )U(S\A z )
and
S\C~\ B,) = ,~/S\B,)
TOPOLOGICAL SPACES
3
e'
for {Bj }, a family of closed sets show that (C2).«(,3) are equivalent to (T2).(TJ), respectively. • Closed rectangles in R n are closed sets. as are clo~cd halls. one-point sets. and spheres. Not every set is either open or dosed. For example. the interval [0, II is neither an open nor a closed set. In the discrete topology on S any set A c S is both open and closed. whereas in the trivial topology any A . 0 or S is neither. Closed sets can be used to introduce a topology just as well as open ones. Thus, if is a collection satisfying
e
e.
1.1.4 Definition. A n open Mig/Iborhood of a point u in a topological space
S is an open set V such that u E V. Similarly,for Q subset A of S. V is an open neighborhood of A if V is open and A c V. A Mig/Iborhood of a point (or a subset) is a set containing some open neighborhood of the point (or subset).
If x E R, neighborhoods of x are. for example, Jx -I. x + 3J. Jx - e. x + e( for any E> 0, and R itself; only the last two are open neighborhoods. The set (x, x + 2( contains the point x but is not one of its neighborhoods. In the trivial topology on a set S, there is only one neighborhood of any point, namely S itself. In the discrete topology any subset contaiping p is a neighborhood of the point pES, since {p} is an open set. 1.1.5 Definition. A topological space is called lint COfIIIttlbk if for each u E S there is a sequence {V•• V2 • ••• } - {V,.} of neighborhoods of U SUM that for any neiglJborhood V of u, there is an n such that Vn c V. The topology is called ,econd COfIIIttlbk if it has a countable basi,.
Topological spaces of interest to us will largely be second countable. For example R n is second countable since it has the countable basis formed by rectangles with rational side length. centered at points all of whose coordinates are rational. Clearly every second-countahle space is also first countable, but the converse is false. For example if S is an infinite noncountable set, the discrete topology is not second countable. but S'is first countable, since { p} is a neighborhood of pES. The trivial topology on S is second countable (see Exercises 1.11. I.IJ for more interesting counterexamples). . A basic fact about second-countable spaces is the following statement due to LindelOf..
.-
TOPOLOGY
1.1.8 PropoelUon. E~ry covering of a set A in a .,econd-countable space S by a family of open sets Vft (that is U "V" ::> A) ("Ontains a countable subcollection {Va,. Va, • ... } also covering A. Proof. Let 'i'l == {Bn} be a countable basis for the topology of S. For each pEA there are indices n and a such that p E Bn C Va' let ~~' = {B"I there
exists an II such that Bn C Va}' Now let Va be one of the Vft that includes the element B" of'i'l'. Since ~' is a covering of A. the countable collection {Vft } covers A. • • The terminology of closure and interior is also very useful.
1.1.7 DefinlUon. ut S be a topological space and A C S. Then the elos",. of A, denoted cl( A) is the intersection of all closed sets containing A. The interior of A. denoted int(A) is the union of all open sets contained in A. The boMtJmy of A. denoted bd(A) is defined by bd( A ) = cl( A) ncl( S \ A) By (e3). cl(A) is closed and by (n). int( A) is open. Note that as bd(A) is the intersection of closed sets. bd(A) is closed. and bd(A) - bd(S\A). Note that A is open iff A = int( A) and closed iff A == c1( A). For example. on R. c1([0.1[) = [O.J).
int([O. I[) = (0.1)
and
bd([O.I[)-{O.I}.
The reader is assumed to be familiar with examples of this type from advanced calculus. Some notions building on these are as follows.
1.1.8 DellnlUoM. A subset A of S is called dense in S if d( A) - S and is called IIOwllere detue if S\cl(A);s dense in S. S ;s called septII'fIbk if it has a countable dense .,uhsel. A point in S is called an tlCcum_,ion poin' of the set A if eoc-h of its neighborhoods contains points of A other than itself. The sel of uuunw/Illion /lfJillts of A is called the derived se' of A and is denoted hy der( A). . A point of A is said to be UoIll'ed if it has a neighborhood in S COIttuining no other point of A than itself. The set A == [0, I[U {2} in R has 2 as its only isolated point. iot( A) - JO, 1(. cl(A) - (0, I]U{2) and def( A) == [0.1). In the discrete topolou OIl • let S. int{p}-c1{p} == {p}. for any peS.
TOPOLOGICAL SPACES
5
Since the set Q of rational numbers is dense in R and is countable. R is separable. Similarly R" is separable. A set S with the trivial toPology is separable since cl{p} = S for any pes. But S = R with the discrete topology is not separable, since c1( A) = A for any A e S. It is not hard to see that any second-countable space is separable. The converse is false: see Exercises 1.11, 1.11. A basic characterization of the operations of closure, interior. and boundary is as follows.
1.1.' Propolltlon. Let S he a topological spare and A e S. Then (i) u E c1( A) iff for every neighborhood U of u, unA. 121; (ii) u E int( A) iff there is a neighborhood U of u .vuch that U e A; (iii)' u E bd(A) iff for every neighborhood U of u, UnA.IZI and un
(S\A).IZI.
(i) u ~ cl( A) iff there exists a closed set C :l A such that u ~ C. But this is equivalent to the existence of a neighborhood of u not intersecting A. namely S\C. (ii) and (iii) are proved in a similar way.. .
Proof.
The basic properties of the operations c1. into bd. and der are given in the following.
1.1.10 PrVpoaltlon. Let A, B and A,. i E I be subsets of S. (i) If AeB, then int(A)eint(B). c1(A)ecl(B). der(A)eder(B) and bd(A) e bd(B); (ii) S\cl(A) "" int(S\A), S\int(A) = cI(S\A). and'cl(A) = A U der(A). (iii) cl(IZI) ... int(IZI)'" 121. cl(S) - int(S) - S. cl(c1(A»'" cl(A) and int(int( A» - int( A); (iv) c1(A U B)=-cI(A)Ucl(B). der(A U B)=der(A)Udcr(B). int(A U B) ::> int(A)Uint( B); (v) cl(A n B)ecI(A)ncl(B). der(A n B)e der(A)ndcr( B). int(A n B) = int(A)nint(B); (vi) CI(U;E,~;)::>U;E,cl(A;), cl(n;E,A,)en'E,c1(A;). int(U;E,A;) ::> U IE ,int(A i ), int( n;E ,Ai) e :';E ,int(A i ). Proof. (i), (ii). and (iii) are immediate consequences of the definition and of Proposition 1.1.9. Since for each i E I. Ai e U IE ,A;. by (i) cl(A,)c c1(U iE ,A i ) and hence u;E,cI(A,)ecl(UiE,A;). Similarly. since niElA, e Ai e cI( A;) for each i E I. it follows that n iE,A; is a subset of the ~Iosed set n;E,cl(A,); thus by (i) cl(n;ElA;)ecl(n'E,cI(A,»=-n;e,cl(A;).
6
TOPOLOGY
The other formulas of (vi) follow from these and (ii). This also proves all the other formulas in (iv) and (v) except the ones with equalitiC5. Since c1(A)Ucl( B) is closed by (e2) and AU Be c1( A)ucI( B). it follows hy (i) that c1(AUB)ccl(A)Ucl(B) and hence equality by (vi). The formula int( A n B) = int( A)nint( B) is a corollary of the previous formula via (ii).
•
The inclusions in the above proposition can be strict. For example. if 14=]0.1[. B"'[I.2[. then c1(A)= der(A) = [0.1). cl(B)-der(B)-[1.2). int( A) -= ]0.1[. int( B) = )1.2[. A U B = )0.2[. and A n B - (lJ so that int( A)U int(B) -)0. I(U)I.2[" JO.2( - int( A U B). cl( A n B) - 0 .. (I) - cl(A)n c1(B). Let A
"
=]-.!. .!.[
n" n '
n= 1.2 •... :
then oc
n A" = {O}, ,,-I and int(
int( A,,) = A" for all n
nA,,) -0-
,,-I
{O} -
n
,,-I
int(A,,).
Dualizing this via (ii) gives
Uc1(R\A,,)=R\{O}*R=c1( ,,-I U(R\A,,»).
,,-I
Limits of sequences can be defined in topological spaces as follows. 1.1.11 Definition. Let S be a topological space and {un} a seqUl'lIc(, of points in S. The sequence is said to etHIW~ if there i.v a point u E S .fuch (hat for every neighborhood V of II. there is an N such that n ~ N implies urI E V. We say that u" CtIIIfJIrIft t" II, or II is Q limit po;"1 of {II,,}.
For example. the sequence {I / n} in R converges to O. I t is obvious that limit points of sequences II" of distinct points are accumulation points of the set {II,,}. In a first-countable topological space any accumulation point of a set A is a limit of a sequence of elements of A. Indeed. if {V,,} denotes the countable collection of neighborhoods of a E der( A) given by definition U.S, then choosing for each n an element a" E V" n A. a,," Q, we see that {all} converges to a. We have proved the following.
TOPOLOGICAL SPACES
7
1.1.12 PropoalUon. Let S be a first-countable space and A c S. Then u E c1( A) iff there is a sequence of points of A that converge to u (in the topology of S). It should be noted that a sequence can be divergent and still have accumulation points. For example-{2.0.3/2. -1/2.4/3. - 2/3 •... } does not converge but has both I and - I as accumulation points. Also. in arbitrary topological spaces. limit points of sequences are in general not unique. For example. in the trivial topology of S any sequence converges to aU points of S. In order to avoid such situations several separation axioms have been introduced. or which the three most impt)rtant ones will he mentioned.
1.1.13 DeftnlUon. A topological space S is called Hau.fdotff if each two distinct points have disjoint neighborhoods (that is. with empty intersection). S is called regu/Iu if it is Hausdorff and if each closed set and point not in this set have disjoint neighborhoods. Similarly. S is called IIOrmal if it is Hausdorff and if each two disjoint closed sets have disjoint neighborhoods.
AU standard spaces in analysis are normal. The discrete topology on any set is normal. but th:: trivial topology is not even Hausdorfr. It turns out that "Hausdorff' is the necessary and sufficient condition for uniqueness of limit points of sequences in first-countable spaces (see Exercise J.JE). Since in Hausdorff spaces single points are closed (Exercise 1.1 F). we haye the implications, normal =- regular =- Hausdorff. Counterexamples for each of the converse' of these implications are given in Exercises 1.11 and I.IJ.
1.1.14 Propoaltlon. A regular second-countable space is normal. Proof. Let A. B be two disjoint closed sets in S. By regularity. for every point pEA there are disjoint open neighborhoods Up of p and UB of B. Hence d( Up)n B ""' 0. Since {Up Ip E A} is an open covering of A. by the Lindel~r theorem 1.1.6. there exists a countable collection {Uklk = 1.2•... } covering A. Thus 00
l::.I
Uk:J A
and
d( Uk ) n B
= 0 .
4-1
Similarly. find a family {Vk } such that 00
U Then the sets
Vk:J B
and
cl(Vk)n If ==0.
k-O II
GII + 1-Un + 1\ U cl(V.). .-0
II
HII-v,,\ U cl(U.), Go -lIo, 4-0
8
TOPOLOGY
are open and
00
G- U G~:::lA,
.-0
H- U H.:::lB
,,-0
are also open and disjoint. •
ConventIon In the remainder of this book Euclidean n-space R· will be understood to have the standard topology (unless explicitly stated to the contrary).
Exerct... l.lA 1.1 B l.1C 1.1 D 1.1 E 1.1 F
1.10
l.lH
Let A - {(x, y, z) e R310 ""' x < I and y2 + Z2 ""' I}. Find int(A). Show that any finite set in R" is closed. Find the closure of {Ijnln = 1,2, ... } in R. Let A cR. Show that sup( A) e cI( A) where sup( A) is the supremum (l.u.b.) of A. Show that a first-countable space is Hausdorff iff all sequences have at most one limit point. (i) Prove that in a Hausdorff space, single points are closed. (ii) Prove that a topological space is Hausdorff iff the intersection of all closed neighborhoods of a point equals the point itself. Show that in a Hausdorff space S the following are equivalent: (i) S is regular; (ii) for every point peS and any of its neighborhoods U, there exists a closed neighborhood V of p such that V c U; (iii) for any closed set A the intersection of all closed neighborhoods of A equals A. (i) Show that if "Ir( p) denotes the set of all neighborhoods of peS, then the following are satisfied: (VI) If A:::l U and U E C'f{( p), then A E 'V(p); (V2) Every finite intersection of elements in 'V( p) is an element of
CV(p); (V3) P belongs to all elements of 'V( p); (V4) If V e 'V( p) then there is a set U
E 'V( p)
such that for all
qEU, Ue'V(q). (ii) If for each peS there is a family 'V( p) of suh:,clS of S satisfying (VI)-(V4), prove there is a unique topology eonS such that for each peS, 'V( p) is the set of neighborhoods of p
e. e
in the topology (Hint: Prove first uniqueness and then' define elements of as being subsets A c S satisfying: for each peA, we have A e'V(p).)
METRIC SPACES
1.11
9
Let S={p=(x.y)ER2IY~O} and let D,(p)={qlllq-plI~E} denote the usual e-disk about p in the plane R2. Define .
( {D,(P)()S. if p=(x,y) with y>O B. p)= {(x,y)ED,(p)IY>O}U{p). if p-(x,Q). Pr~ve the following: (i) 'Y( p) = {U C Sithere exists H,( p) C U} satisfies (VI)-(V4) of Exercise 1.1 H. Thus S becomes a topological space. (ii) S is first countable. (iii) S is Hausdorff. (iv) S is separable. (Hint: {(x. y)E Six. r"E Q. r > O} is dense in S.)
\.\ J
(v) S is not second countable. (Hint: Assume the contrary and get· a contradiction by looking at the points (\.0) of S.) (vi)· S is not regular. (Hint: Try to separate (x(l.!» from {( x.O) Ix E R}\{( x o, On.) With the same notations as in the preceding exercise. except changing B,(p) to if P"" (x, y), Y>O D.(x. e)U{p}. if p = (x.O).
D.(P)()S. B (P) = { . ,
.\ K
Show that (i)-(v) of 1.11 remain valid and that (vi) S is regular. (Hint: Use 1.10.) (vii) S is not normal. (Hint: Try to separate cI( B,( p» for p = (xo.O) from {(x.O)lx E R}\{( xo.O)}.) Prove the following properties of the boundary operation and show by example that each inclusion cannot be replaced by equality (i) bd(A)=bd(S\A); (ii) bd(bd( A» c bd( A); (iii) bd( A U B) c bd( A)ubd(B) c bd(A U B)U A U B; (iv) bd(bd(bd( A))) = bd(bd( A». (These properties may be used to characteri7.e the topology.) 1.2 METRIC SPACES
One of the common ways to form a topological space is through the use of a distance function. also called a metric. For example. on R" the standard distance d(x.y)= (x,1/2
(t,-I
y,)2)
10
TOPOLOGY
between x == (XI'" .,xn ) and y = (YI" .. ,Yn) can be used to construct the open disks and from them the topology. The abstraction of this proceeds as follows.
1.2.1
Definition. Let M he a set. A metric on M is a function d: M X M ..... R such that
d(ml' m2) == 0 iff m l = m 2; (definiteness); (symmetry); and (M3) d(ml' m) .. d( mi' m;)+ d(m 2 • m J ); (triangle inequality). (MI)
(M2) d(m l ,m2)==d(m 2,m l );
A metric spIICt! is the pair ( M, d); if there is no danger of confusion just write M for (M,d).
Taking m l == m) in (M3) shows that d is necessarily a non-negative function. It is proved in advanced calculus courses (and is geometrically clear) that the standard distance on R n satisfies (M I )-(M~). The topology determined by a metric is ddined as follows.
1.2.2 Definition. For r> O. and m EM. the t-disk about m is defined by D.{m) = {m'E Mld(m'. m) < t}. The collection of subsets of M that are unions of such disks is the metric topology of the metric space (M. d). Two metrics on a set are called equivalent if the.' induce the same metric topology.
1.2.3 Propolltlon.
(i) The de.~cripti(/n of open Jets Killen in the precedinK definition is a topology. (ii) A set U C M i.f open if and only if for each m E U there is an r > 0 such that D.( m) C U.
(i) (TI) and (TJ) are clearly satisfied. To prove (T2). it suffices to show that the intersection of two disks is a union of disks, which in tum is implied by the fact that any point in the intersection of two disks sits in a smaller disk included in this intersection. To verify this. let p E D.(m)n D,(n) and O
METRIC SPACES
11
Every set M can be made into a metric space by the discrete metric given by demo n) = I for all m * n. The metric topology of M is the dis.crete topology. A pseudometric on the set M is a function d: M x M ..... R that satisfies (M2). (M3). and (PMI) demo m) = 0 for all m. Thus the distance between distinct points can be zero for a pseudometric. The pseudometric topology is defined exactly as the metric space topology. Any set M can be made into a pseudometric space by the trivial pseudometric demo n) = 0 for all m. n EM; . the pseudometric topology on M is the trivial topology. Note that a pseudometric space is Hausdorff iff it is a metric space. If M is a metric space (or pseudometric space) and /I'M. A ( M. we define d(u. A) = inf{d( u. v)lv E A} if A * 12). and d( /I. (1) - 'Xl. The (liameter of a set A eM is defined by diam( A) = slIp{d( II. " )1/1," E A}. A set is called bounded if its diameter is finite. Clearly metric spaces are first-countahle and Hausdorff; in fact:
1.2.4 Propoaltlon. Every metric space is lIorma/. 'Proof.
Let A and B be closed. disjoint subsets of M. and let
U=
(UE
V= (v
E
Mld(u. A) < d(u, B)}; Mld( v. A) > d(v. B)}.
It' is verified that U and V are open. disjoint. and A c U. B c V.
•
In a, metric space we have the important notion of completeness.
1.2.5 Definition. Let M he a metric space with metrh' d lIlId (u,,) a sequence in M. Then {un} is a Cauchy .~equence if for all r > O. fER. there is an integer N such that n. m ~ N implies d( u". u m ) < f. The ,~pa(''e M is ca/It'd complete if ever)' Cauchy sequence converges. A sequence {un} converges to u iff for every E> 0 there is an integer N such that n ~ N implies d(u". u) < E. This follows readilv from the definitions 1.1.11 and 1.2.2. A convergent sequence {u,,} is a Cauchy sequence. To see this. let E> 0 be given. Choose N such that n ~ N implies d( II". II) < f /2. Thus n. m ~ N implies d(u". u m ) " d(u". u)+ d(u. /I,,) <' r/2+ r/2 = f hy the triangle inequality. Completeness requires that collverselv. every CaudlV, sequence converges. A basic fact about R n is that with the standard metric. it is complete. The proof is found in any textbook on advanced cakulus. One of the most useful applications of the notion of completeness is Banach's contraction mapping theorem. We shall use it several times later in the book .
.
12
TOPOLOGY
1.2.8. Theorem. Let M be a complete metric spa('e and f: M ..... M a mapping. Assume there Is a constant k, where 0" k < I .fuch that d(j(m).j(n» " kd(m. n). for all m. n E M; (such an f is called a contrtlction). Then f has a unique fixed point; i.e .• there exists Q unique m. E M .fuch that f(m.) = m •. Proof. Let mo be an arbitrary point of M and define recursively m, + , = f(m;). i = 0.1.2..... Inductively it follows that d( m,. m,. ,) ".k;d( mo. m,). so that for i < }.
Since 0" k < I. we have the convergent geometric series 00
L k' = 1/(1 -
k).
,-0.
so that in particular k' + ... + k I , approaches zero. as i . .i -400. This shows that the sequence {m,} is Cauchy and thus by completeness of M it converges to a point m •. Since
is arbitrarily small it follows that m. = f( m .), proving thus the existence of a fixed point of f. To show uniqueness, let m' be another fixed point of f. Then
d(m', m.) = d(j(m'), f(m.»" kd(m'. m.), which. by virtue of 0 or;; k < I, implies d( m'. m.) = 0, so m' = m..
•
We note that the condition k < 1 is necessary. For if M = Rand f(x) = x + I, then k = I, but f has no fixed point. At this point the true significance of 1.2.6 cannot be demonstrated. When applied to the right spaces, however it will yield the inverse function theorem (Chapter 2) and the basic existence theorem for differential equations (Chapter 4), A hint of this is given in Exercise 1.21.
METRIC SPACES
13
Exercl... 1.2A
I.2B
Let d«x" YI).(x2' Y2»'" sup( IXI - x21. IYI - Y21). Show that d is a metric on R 2 and is equivalent to the standard metric. Let I(x) = sin(\/x). x> O. Find the distance hetween the graph of I and (0.0).
12C Show that every separable metric space is sC'l'Ond countahle. 1.2D Show that every metric space has an equivalent metric in which the diameter of the space is I. (Hint: Consider d l (",. fl) = d( m. n )/(1 + demo n».) 1.2E In a metric space M. let '\.( m) = {V c MI there exists E > 0 such that D,(m)CU}. Show that 'V(m) satisfies (VI)-(V4) of Exercise I.IH. This shows how the metric topology can be defined in an alternative way starting from neighborhoods. 1.2F
In a metric space show that d( A) = {u E Mld( u. A) = O}.
Exercises 1.2G-I use the notion of continuity from calculus (see Section 1.3). 1.2G
Let M denote the set of continuous functions f: [0. I interval [0. I). Show that d(f. g) =
r
-+ ~
on the
11/(x) - g(.\')1 d:c 1
II
1.2H
is a metric. Let M denote the set of continuous functionsJ: [0.1] -+ R. Set
d (f. g ) = sup {If ( x ) - g ( x)1 I 0 .;:; x .;:; I).
1.21
(i) Show that d is a metric on M. (ii) Show that /" -+ I in M iff In converges un;rorm~v to f. . (Iii) By consulting theorems on uniform convergence from your advanced calculus text. show that M is a complete metric space. Let M be as in the previous exercise and define T: M -+ M by
'.
T(f)(x)=a+ fo'K(x.)')/(Y)d.Y. where a is a constant and K is a continuous function of two variables. Let k
=
and suppose k < I.
sup{
10'1 K ( x. Y)I dy I0 .;:; x ~ I}
14
TOPOLOGY
(i) (ii)
Show that T is a contraction. Deduce the eltistence of a unique solution of the integral, equation f(x)=a+l'K(x.)')f(y)dy. (I
(iii) Taking a special case of (ii). prove the "existence of e' ".
1.3 CONTINUITY Continuity is one of the basic ingredients used in manifold theory. We begin with the definition in the context of topological spaces.
1.3.1
Definition. Let Sand T be topological spaces and cp: S -+ T be a . mapping. Then cp is COIftilluoru tit u E S if for every neighborhood V of cp(u) there is a neighborhood U of u such that cp(U) C v. If. for every open set Vof T. cp-I(V)={uESlcp(U)E V} is open in S. cp is COIftilluoru. (Thus. cp is continuous iff cp is continuous at each u E S.) If cp: S -+ T is a bijfftiott (that is. one-to-one and onto). cp. and cp 1 are continuous. then cp is a Itomeomorpllism and Sand Tare Itomeomorpltic. For example. notice that any map from a discrete topological space to any topological space is continuous. Similarly. any map from an arbitrary topological space to the trivial topological space is continuous. Hence the identity map from the set S topologized with the discrete topology to S with the trivial topology is bijective and continuous. but its inverse is not continuous. hence it is not a homeomorphism. From 1.3.1 it follows by taking complements and using the identity S\!p-I(A) =11'- I(T\A). that 11': S -+ T is continuous iff the inverse image of every closed set is closed.
1.3.2 Propottltlon. Let S, T be topological spaces and 11': S -+ T. The following are equivalent: (i) II' is continuous; (ii) cp(cl(A»C cl(cp(A» for every A C S; (iii) cp - I(int( 8» C int( cp - 1( 8» for every 8 C T.
Proof. If II' is continuous, then II' Icl(!p( A» is closed. But A c!p Icl(!p( A» and hence cl( A) C II' - Icl(!p( A». or 11'(cl( A) C cl(!p( A». Conversely. let B C T be closed and A = !p-I(B). Then cl(A)C !p-I(B)= A. so A is closed. A similar argument shows that (ii) and (iii) are equivalent. • This proposition combined with l.I.l2 (or a direct argument) gives the following.
CONTINUITY
15
1.3.3 Corollary. Let Sand T he topological spaces with S lirst cOimtable and 'P: S -+ T. Then 'P i.f continuous iff lor every sequence {u m } converging to u. ('P( u,,)} converges to 'P( u), lor all u E S.
1.3.4 Propoeftlon. ous map.
The composition
01 two continuous maps is a
continu-
Proof. If 'PI: SI -+ S2 and 'P2: S2 -+ S) are continuous maps and if V is open in S3' then ('P2 0 'PI)- I(V) = 'PI I( 'Pi I( is open in SI since 'P2 I(U) is open in s,. by continuity of 'P2 and hence its inverse image by 'PI is open in SI by continuity of 'PI' •
V»
1.3.5 CorOlla". The set 01 all homeomorphisms of a topological space to itself lorms a group under composition.
Proof. Composition of maps is associative and has for an identity element the identity mapping. Since the,inverse of a homeomorphism is a homeomorphism by definition, and since for any two homeomorphisms 'PI' 'P2 of S. 'PI 0 'P2 and ('PI 0 'P2)- I = 'Pi 10 'PI I are continuous by 1.3.4. the corollary follows. •
1.3.8 Pr0P,08lUon. The space 01 continuous mUfI.f algebra under pointwise addition and multiplication.
I: S -+ R
lorm.I' un
Proof. We' have to show that if I and g are continuous. then so are I + g and Ig. Let So E S be fixed and e> O. By continuity of I and g at ,fo. there exists an open set V in S such that lI(s)- l(so)1 < e/2, Ig( s)- g(.fu)1 < E/2 for all s E V. Then· 1(f+g)(s)-(f+g)(so)IEOI/(s)-/(so)I+I.,,(.f)
g(so)l<e
Similarly. for e> O. choose a neighborhood V o(.so such that I/(s)- l(so)1 < 6.lg(s)- g(so)1 < 6 for all s E V ,where 6 is any positive number satisfying (6 + I/(so)l)6 + Ig(so)16 < e. Then l(fg )(s)- Ug )(so)1 EO lI(s )Ug( s) - g(so)1
+ if( s)- l(so)llg(so)1
< (6 + I/(so)1)6 + 6Ig(so)1 < e.
Therefore, I + g and Ig are continuous at so' •
16
TOPOLOGY
Continuity is defined by requiring that inverse images of open (closed) sets are open (closed). In many situations it is important to know whether the image of an open (closed) set is open (closed).
1.3.7 Deftnltlon. A map 'P: S -+ T. where Sand T are topological spaces. is called opm (cloud) if the image of every open (closed) set in S is open ( closed) in T. Thus a homeomorphism is a bijective continuous open (closed) map. It should be noted however that not every open (closed) map is closed (open). Examples and applications of this notion will be given in the next section. For the moment note that the identity map of S topologized with the trivial and discrete topologies on the domain and range. respectively. is not continuous but is both open and closed. Let us now turn our attention to continuous maps between metric spaces. For these spaces. continuity may be expressed in terms of f'S and 6's familiar from calculus.
1.3.8 Propo8lt1on. Let (MI' d l ) and (M2 • d 2 ) be metric spaces. and 'P: MI -+ M2 a given mapping. Then 'P: MI -+ M2 is continuous at u l E MI iff for every f> 0 there is a 6> 0 such that dl( u l ' u;) < 6 implies d 2('P( u l ). 'P( U;» <E.
Proof.
Let 'P be continuous at
UI
and consider D,2( 'P( UI»' the f-disk at
«p( u l ) in M 2 • Then there exists a 6-disk D~( u l ) at u l in MI such that
»
'P(Dl(ul»CD/('P(u l (by 1.3.1); i.e.. d(u l,u;)<6 implies d 2 ('P(u l ). 'P( u;» < f. Conversely. assume this latter condition is satisfied and let V be a neighborhood of lJ'(u l ) in M 2 • Choosing an r-disk D/<'P(ul»C V there exists 6 > 0 such that 'P( D~( u l c D,2( (fi( III » by the fl. . regoing argument. Thus'P is continuous at u l . •
»
In a metric space we also have the notions of uniform continuity and uniform convergence.
1.3.9 Deftnltlon. (i) Let (MI.d l ) and (M 2 .d2 ) be metric spaces and 'P: MI -+ M 2· We say 'P is lIIIi/ormly COlttinuous if for every f> 0 there is a 6 > 0 mch that dl( u. v) < 6 implies d 2 ( 'P( u). 'P( v» < E. (ii) Let S be a set. M a metric space and 'PII: S -+ M. II = 1.2.... and 'P~ S -+ M be given mappings. We say 'Pn -+ 'P lIIIi/orml), if for every 10 > 0 there is an N such that d('P,,(u).'P(u» < f for all n ~ N and all uE S. For ellample. a map satisfying d( 'P( u). 'P( v» ~ Kd( u. v) for a constant K is uniformly continuous. Uniform continuity and uniform convergence
CONTlNUI.TY
17
ideas come up in the construction of a metric of the space of continuous maps. This is considered next. .
1.3.10 ~ltIon. Let M be a topological space and (N.d) a complete metric space. Then the collection C( M. N) 01 all bounded continuous maps cp: M - N lorms a complete metric space with the metric
Proof. It is readily verified that dO is a metric. Convergence of a sequence In E C( M. N) to IE C( M. N) in the metric dO is the same as unilorm convergence. as is readily checked. (See Exercise 1.211" Now if In is a Cauchy sequence in C( M. N). then I .. (x) is Cauchy for each x E M since.
Thus In converges pointwise. defining a function II x). We must show that I" - I uniformly and that I is continuous. First of all. given e> O. choose N such that dOC In. 1m) < e/2 if n. m ~ N. Then for any x E M. pick N" ~ N so that d(fm(x)./(x» < f'/2 if m., N", Thus with n ~ Nand m ~ N".
dU,,(x). I(x») ~ dUne x). Im(x »)+ dUm(x). I{x») < e/2+ e/2 = e.
so I" - I uniformly, The reader can similarly verify that f is continuous (see Exercise l.3F; look in any advanced calculus text such as Marsden'[1974a] for the case in R" if you get stuck). _ . Exercl... 1.3A Show that a map cp: S - T between the topological spaces Sand Tis continuous iff for every set BeT. cI(cp-I(B»ccp-l(cl(B». Show that continuity of cp does not imply any inclusion relations between cp(int A) and int( cp( A». 1.38 Show that a map cp: S - T is continuous and dosed if for every subset U C S. cp( cl( U» = c1( cp( U » 1.3C Show that compositions of open (closed) mappings are also open (closed) mappings. I.3D Show that cp: ]0. oc[ - ]0. oc[ defined by cp( x) = 1/x is continuous but not uniformly continuous. 1.3E Show that if d is a pseudometric on M, then the map d( '. A): M - R for A C M a fixed subset is continuous.
18
TOPOLOGY
l.3F If S is a topological space. T a metric space and CPn: S -+ T a sequence of continuous functions uniformly convergent to a mapping cP: S -. T. then cP is continuous. 1.4 SUBSPACES, PRODUCTS, AND QUOTIENTS
This section concerns the construction of new topological spaces from old ones. The first basic operation of this type we consider is the formation of subset topologies. 1.4.1 Definition. If A is a suhset of a topological space S. the relative topology on A is defined by
From the identities l. 0nA=0.SnA=A. 2. (V. n A)n(U2 n A) = (V. n U2 )n A. and 3. ua(VanA)=(UPa)nA. we see that ~A is indeed a topology. Thus. for example. the topology on sn I = ( X E R" Iti( x. 0) = I} is the relative topology induced from Rn; i.e .. a neighhorhood of /I point x F S" I is a subset of sn -. containing D,< x)n sn I for some E> O. Note that an open (closed) set in the relative topology of A is in general not open
Products or more than two factors can be considered in a similar way. u. v). w) ..... (u. (v. w» is a
It is straightforward to veriry that the map
«
SUBSPACES. PRODUCTS. AND QUOTIENTS
19
homeomorphism of (S x T)X Z onto S X(T x Z). Similarly one sees that S X T is homeomorphic to T X S. Thus one can take products of any number of topological spaces and the factors can he grouped in any order; we simply write SI X ••• x S" for such a finite produl:t. For example. A" has the product topology of AX ... XA (n times). Indl'cd using the metric d(x, y) == max (Ix i
-
y'l)
I .. ; " "
we see that the f-disk at x coincides with (Xl - f, Xl + fIx . ~. + f). For a generalization to metric spaces see Excrdsc I.4N.
X
(.1(11 -
f,
x"
I'·
1.4.3 ProposlDon. Let S and T be topolog;cal ,~pa((.',~ and dellote hy PI: . S X T -+ Sand P2: S X T -+ T the clllfOilicol projection,.. , PI(S, t) = s. and P2(s! t) = I. Then (i)
PI and P2 are open mappings, and
(ii) a mapping cp: X -+ S X T, where X is CI lop%gim/ space. i.l· iff both PI 0 cp: X -+ S. P2 0 cp: X -+ Tare COIIII!IllOlIS.
('(~I11II11WUS
(i) follows directly from the definitions. X (ii) cp is continuous iff cp-I(U X V) is open in S X t, for U VeT open sets. Since
Proof.
the assertion follows.
' C
Sand
_
.It should be noted that the maps Pi' ; = 1,2 are not closed. Fo! example. ' if S=T=A the set A={(.I(,y)lxy=I,x>O} is closed in SXT==A2. but PI( A) = )0. oo[ which is not closed in S. An important characterization of Hausdorff spaces is given in the following.
1.4.4 ProposlDon. A lopologica/ space S is Huu.fdorff iff 'he diagonal 4 s '" {(so s)ls E S} c S X S is a closed subspace of S X S. with the product topology.
Proof. It is enough to remark that S is Hausdorff iff for every two.distinct points p. q,E S there exists neighborhoods Up. Uq of p. q, respectively. such, that
10
TOPOLOGY
In • number or places later in the book we are going to form new lopologicaJ Ip8CIeS by collapsing old ones. We define this process now and pve some examples.
1.4.5 Deflntuon. Let S be a set. An equivalence relation - on S is a binary relation such that for all u, v, wE S, (i) (ii)
(iii)
u - u, v iff v - u, and u - v and v - w implies u - w. U -
(identity) (reflexivity) (transitivity)
The equivalence claM containing u, denoted (u], is defined hy [u]={vESlu-v). The set of equivalence classes is denoted S/ -, and the mapping 'IT: S ---- S/ -; u ..... I·u] is called the CIIltOlliclll projection.
It follows easily that S is the disjoint union of its equivalence classes. The collection of subsets V of S/ - such that 'IT-I(V) is open in S is a topology because
1. 'IT- I (f2J)=f2J,'IT- I (S/-)=S 2. 'IT-I(V. n V z ) - 'IT-I(VI )(')'IT - I(V2), and
3.
'IT- I ( U "V,,) = U ,,'IT- I(V,,)
1.4.8 Definition.
Let S be a topological space and - an equivalence relation on S. Then (V c S/ - 1 'IT- I(V) is open in S} is called the quotient topology on S/ -.
1.4.7 Examples A. Consider R 2 and the relation - defined by
(Z denotes the integers). Then T Z =R 2/-is called the 2-10rus. In addition to the quotieqt topology, it inherits a group structure in the usual way:
The n-dimensional torus Tn is defined in a similar manner. The torus T Z m,il' he obtained in two other ways. First. let 0 be the unit square in R Z \',Illl the subspace topology. Define - t· f - Y iff any of
SUBSPACES, PRODUCTS, AND QUOTIENTS
21
Figure 1.4.1
the following hold: (i) x = y; (ii) XI = YI' x 2 = 0.Y2 = I; (iii) x, = YI' X2 == I. Y2 = 0; (iv) x2=Y2.XI=0'YI-I;or (v) x 2 == Y2' XI = I. YI = 0;
as indicated in Fig. 1.4.1. Then T 2 = OJ - . Second. define T 2 = S I X S I. also shown in Fig. 1.4.1. B. The Klein bottle is obtained by reversing one of the orientations on O. as indicated in Fig. 1.4.2. Then K 2 = 0/ - (the equivalence relation
\.
-'
:r
-
--- ·J!IIl-( : .. 4/
"';' : ,
..
-
.
,'.,,' .:. -:'
-.',:
P
',;
'j: -, ;
.
.'
~
. ;. . ,'. Iii ....... ·
.~
Figure 1.4.2
22
TOPOLOGY
indicated) is the Klein bottle. Although it is realizable as a subset ()f R 4 , it is convenient to picture it in R] as shown. In a sense we will make precise in Chapter 6, one can show that K 2 is not "orientable." Also note that K I does not inherit a group structure from R 2, as did T 2. C. On R"\{O) define x - y if there is a nonzero real constant A such that x = Ay. Then R"\{O}/- is called real projective (n -I) space and is denoted by RP"~I. Alternatively, Rp .. -I can be defined as S .. -I (the "nit sphere in R") with antipodal points x and - x identified. (It is easy to see that this gives a homeomorphic space.) One defines complex projective space Cp .. -I in an analogous way where now A is complex. ..
1.4.8 Propoaltlon. Let - be an equivalence relation on the topological space Sand w: S ...... S/ - the canonical projection. A map q>: S/ - ...... T, where T is another topological space, is continuous iff q> 0 w: s . . . T is continuous. Proof. q> is continuous iff for every open set VeT, S/ -, i.e., iff (q> 0 w) -I( V) is open in S. •
q> - I( V)
is open in
1.4.9 Definition. The set r = {(s, s')ls - s') c S X S is called the graph of the equivalence relation -. The equivalence relation is called open (closed) if the canonical projection w: S ...... S/- is open (closed). We note that - is open (closed) iff for any open (closed) subset A of S the set w- I( w( is open (closed). As in 1.4.8 we see that for an open (closed) equivalence relation - on S, a map q>: S/- ...... T is open (closed) iff q> w: S ...... T is open (closed). In particular, if - is an open (closed) equivalence relation on Sand q>: S/- ...... T is a bijective continu()us map, then q> is a homeomorphism iff q> w is open (closed). It is sometimes important to know whether or not a quotient space is Hausdorff. The standard criterion is the following.
A»
0
0
1.4.10 Pr0p08IUon. If S/- is Hausdorff, then the graph r of - is closed in S X S. If the equivalence relation - is open and r is closed (as a subset of S X S). then S/- is Hausdorff.
!'roo/. If S/- is Hausdorff, then
r .. (ft X 7T)-I(l:i. sl _ )
is closed in S
l:i. SI _ is closed by 1.4.4 and hence S, where w X 7T: S X S ...... (S/- )X
X
(S/- ) is given by (7T X 7T)(X. y)= ([x),[y)).
Assume now that r is closed and - is open. If S/ - is not Hausdorff then there are distinct points (x),[y) E S/ - such that for any pair of neighborhoods U, and Ul' of (x I and [y J. respectively, we have U, n Ul' '*' 0 . Let V, and v,. be any open neighhorhoods of x and .1', respectively. Since ,
SUBSPACES. PRODUCTS. AND QUOTIENTS
23
is an open equivalence relation. '1T( Vx ) = Vx and '1T( V,. ) = V •. are open neighborhoods of [x] and [y) in SI-. Since Vx n V, * i2J. there exist xE V, and jiE V. such that[X] = [ji); i.e. (x. ji)E r. Thus, (~. y)E c1(f) by I.I.9(i). As r is dosed, (x, y) E r. i.e., [x) = (y), a contradiction. • Exercl...
1.4A Show that the sequence x .. "" lin in the topological space (0.1) (with the relative topology from R) does not converge. I AD If 4p: S ... T is continuous and T is Hausdorff. show that the graph 'of f. r, "" {(so f(s»ls E S} is closed in S X T. 1.4C Let X and Y be topological spaces with Y Hausdorff. Then show that for any continuous maps f, g: X -+ Y. (x E Xlf(x) = g(x)} is closed. (Hint: Consider the mapping x .... (f(x). g(.\') and use 1.4.4.) Thus. if f( x) = g( x) at all points of a dense subset of X, th~n f = g. 1.4D Define a topological manifold to be a space locally homeomorphic to R". Find a topological manifold that is not Hausdorff. (Hint: Consider R with "extra origins.") . 1.4E Show that a mapping 4p: S ... T is continuous irf tht mapping s, ...... (s./(s» of S to the graph r,={(s./(s»lsES}CSXS is·a homeomorphism of S with r, (r, with the subspace topology induced from the product topology of S x T). 1.4F Show that every subspace of a Hausdorff (regular) space is Hausdorff (regular). Conversely, if each point of a topological space has a closed neighborhood that is Hausdorff (regular) in the subspace topology. then the topological space is Hausdorff (regular). (Hint: use Exercises 1.1 F and l.1G.) I.4G Show that a product of topological spaces i~ Hausdorff iff each factor is Hausdorff. 1.4H Let S. T be topological spaces and -. '" he elluivalence relations on Sand T. respectively. Let 4p: S -+ T be continuous and such that SI-'S2 implies 4p(SI)""4p(S2)' Show then that the induced mapping q>: SI- ... TI'" is continuous. 1.41 Let S be a Hausdorff space and assume there is a continuous map a: SI - ... S such that '1T 0 a =- i s/_ • the identity. Show SI- is Hausdorff and a(SI- ) is closed in S. 1.4J Let M. N be metric spaces, N complete, and 4p: A ... N .uniformly continuous (A with the induced metric topology). Show that 4p has a unique extension q>: X ... N that is uniformly continuous. l.4K Let S be a set. To a family of topological spaces, and 4p.. : S ... To a family of mappings. Let ~'6 be the collection of finite intersections of
24
TOPOLOGY 5CtS 01 the form.; 'W.)' for l~ ~ ~ . . ~.. S - T.
cpen in T._ TbI! i,.;tkJl tt'Pt"t"!.'· \.'lCl S _ as. Nsis the u-.llc\."1h"tTI .-r.. ~ thltt Ihlli lopolo,), IIi charactcri7.ed by the 'act that any mapping cP: R ...:. S (rom a topological space R is continuous iff all CP.. 0 cP: R - T,. are continuous. Show that the subspace and product \oP9logies ar~ initial tQPoloajes,
Let T be-a set and CP.. :~ S.. - T a family of mappings. S.. topological spaces with topologies Let = (U c Tlcp; I( U) E eO. for each a}. Show that is a topology on T. called the final topology on T given by the family CPa: So. - T. Show that this topology is characterized by the fact that any mapping cP: T - R is continuous iff cP 0 CPa: Sa - R are all continuous. Show that the quotient topology is a final topology. I.4M Show that in a complete metric space a subspace is closed iff it is complete. 1.4N Show that a product of two metric spaces is also a metric space by finding at least three different equivalent metrics. Show that the product is complete iff each factor is complete. I.4L
e
ea'
e
1.5 COMPACTNESS Some basic theorems of calculus. such as "every real valued continuous function on [a. b) attains its maximum and minimum" implicitly use the fact that [a. b) is compact. The general theory of compactness is described in this section.
1.5.1 Definition. Let S be a topological space. Then S is called compact -if for every covering of S by open sets Ua (that is U "Ua = S) there is a finite subcollection of the Ua also covering S. A subset A c S is called compact if A is compact in the relative topology. A subset A is called relat~1y compact if cl( A) is compact. A space is called loctIIIy compact if it is Hausdorff and each point has a relatively compact neighborhood. Some basic properties are given in the following.
1.5.2 Propoeltlon. (i) If S is compact and A c S is closed. then A is compact. (ii) If cp: S - T is continuous and S is compact. then cp(S) is compact. Proof. (i) Let {Ua} be an open covering of A. Then {Ua. S\A} is an open covering of S and hence contains a finite subcollection of this covering also covering S. The elements of this collection. except S\A. cover A.
COMPACTNESS
25
(ii) Let (U,,) be an open covering of cp(S). Then (cp-I(Va )} is an open l"Ovmng of S and thu!i h;.- compaC'tnes.s of S :\ rinite suhroll«tioo {cp - I( Va,) •...• cp I( Va.)} covers S. But then {VR ,., . . . l .• ) covers !p( S) and thus ,,( S) is compact. • In a Hausdorff space, compact subsets are dll~~d (nercise), rhu~ If S is compact, T is Hausdorff and cp is continuous. then cp is closed; if!p is also ' bijective. then it is a homeomorphism.
1.5,3 Propos.Uon. A product space S x Tis ('ompact iff both Sand Tare compact. Proof. In view of 1.5.2 all W6 have to show is thai if ,S' and T are compact,' so is S x T. Let {A,,> be a covering of S x T by open sets, Each Aa is the union of sets of the form V x V with V and V open in Sand T, respectively. Let {Vp x Vp} be a covering of S x T by open rectangles. If we show that there exists a finite subcollection of Vp x Vp covering S x T. then clearly also a finite subcollection of {A,,} will cover S x T. A finite subcollection of {Vp x Vp } is found in the following way. Since {s}X T is .compact there exists a finite collection
covering it. If I,
then V, is open. contains sand U. x Vp" ... ; V, x Vp, covers {s}X T. Let W, = V, x T; {w,} is an open covering of S x T and if 'we show that only a finite number of these W, cover S x T, then since I.
W,
=
U (V, x Vp ).
j-I
I
it follows that a finite number of Vp x VII will cover S x T. Now look at S x {t}, for t E T fixed, Since this is compact. a finite suhcolle~tion W" ..... W" covers it. But then k
U W, =SxT. • I-I
"
As we shall see shortly in 1.5.9, [-1,1) is compact. Thus TI is compact.' It follows from 1.5.3 that the torus T2. and inductively Tn, are compact. Thus if 71': R2 -+ T2 is the canonical projection we see that T2 is com-
26
TOPOLOGY
pact without R2 being compact; i.e.• the converse of 1.5.2 (ii) is false. Nevertheless it sometimes occurs that one does have a converse: this leads to the following.
1.5.4 Definition. A mapping
!p: S ..... T is proper if it is continuous and if the inverse image of every compact set is compact.
Note that the composition of proper mappings is again a proper mapping. Without the explicit assumption of continuity. such mappin'gs may be discontinuous. For example. the identity mapping of a finite set with the trivial and discrete topologies is not continuous. but the inverse images of compact sets are compact since any finite set with the trivial topology is compact. Note that any continuous map defined on a compact space is proper in view of 1.5.2(i). Compactness is also characterized in terms of convergence of sequences.
1.5.5 Boluno-Welerstras. Theorem. If S is a compact Hausdorff space. then every sequence has a convergent subsequence. The converse is also true in metric and second-countable Hausdorff spaces. Proof. Suppose S is compact and {un} contains no convergent subsequences. Then we may assume that the points of the sequence are distinct. Then each un has a neighborhood v" that contains no other Urn. for otherwise un would be a limit of a subsequence. Since the single-point sets {un} are closed (Exercise I.IF) the sets S\{u,,} form an open covering of S with no finite subcollection covering S. a contradiction. Let S now be second countable. Hausdorff. and such that every sequence has a convergent subsequence. If {Ua} is an open covering of S. by the Lindelof theorem there exists a countable collection {V"I n = 1.2.... } also covering S. Thus we have to show that {V"I n = I. 2.... } contains a finite collection covering S. If this is not the case. the family (S\ U ~-IV,) consists of closed nonempty sets and S\ U 7-1V, ::> S\ u 7: IV, for m ~ n. Let p" E S \ u 7- IV,. If {p"ln ... 1,2•... } is infinite. by hypothesis it contains a convergent subsequence; let its limit point be denoted by p. But then pES \ U 7- IV, for all 'I, contradicting the fact that {U"ln = I. 2•... } covers S. Thus {p"ln = 1.2•... } must be a finite set; i.e .• for all n ~ no, Pn = PIOn' But then again p" .. E S \ U :_ IV, for all n, contradicting the fact that {V"I n = 1,2.... } covers S. Hence S is compact. Let S be a metric space such that every sequence has a convergent subsequence. If we show that S is separable. then since S is a metric space it is second countable (Exercise 1.2(,). and by the preceding paragraph. it will be compact. Separability of S is proved in two steps.
COMPACTNESS
27
First we show that for any l > 0 there exists a fi"ife set { PI" .. • p,,} such that S - U ;'. ,D,( p,). If this were false. there would exist an l > 0 such that no finite number of e-disks cover S. Let p, E S he arhilrary. Since D,( PI)" S.there exists P2 E S\D,< p,). Since D,( p,)U D,( P~ j""" S. there exists P.l E S\(D.(p,)UD,(P2»' etc. The sequence {p"ln=I.2.... } is infinite and d( Pi' Pj) ~ e. But this sequence has a convergent subsequence by hypothesis. so this subsequence must be Cauchy. co~tradicting d( Pi' Pi) ~ e for all i. j. Second. we show that the existence for every l > 0 of a finite set {p, .....p"",} such that S = U 7~',)D,( p;) implies S is separable. Let -1" denote this finite set for l = 1/ n and A = U :'_, A". Thus A is countable and it is easily verified that d( A) ... S. • A property that came up in the preceding proof turnll out to be important.
1.5.8 DeflnlUon. Let S be a metric space. A suhset A c S IIoruIMd if for any
E>
i.~ called totally 0 there exists a finite set {Pl .... •p,,} in S .~uch that
..
Ac
U
D,(p;).
;-1
1.5.7 Corollary. A metric space is compact iff it is complete and total!.l· bounded: A subset of a complete metric space is relatil'ely compact iff it is tota/~v bounded. Proof. The previous proof shows that compactness implies total boundedness. As for compactness implying completeness. it is enough to remark that in this context. a Cauchy sequence having a convergent subsequence is itself convergent. Conversely. if S is complete and totally bounded. let {p"ln = I. 2.... } bj: a sequence in S. By total bounded ness. this scquenl'e contains a Cauchy subsequence. which by completeness. converges. I'hu .. S is compact hy the 8olzano-:-Weierstress theorem. The second statement 'hlW readily follows .•
1.1.1 PropoalUon. In a metric space £'Ompact S('t.V arC' c/o.fed and hounded.
Proof. The first assertion is a particular case of the previous corollary but can be easily proved directly in Hausdorff spaces. If A is compact. it can be finitely covered bye-disks: A"" U 7_ID.< Pi)' Thus
.
diam(A).,;
E diam(D.(p;»=2ne. •
From 1.5.2 and 1.5.8. if S is compact and 'P: S -0 R is continuous. then 'P(S}, is closed and bounded. Thus 'P attains its sup and inC ,<see Exercise I.ID).
28
TOPOLOGY
1.5.8 Helne-8oreI Th.."......
In R" a closed and bounded set is compact.
Proof. By 1.5.2(i) it is enough to show that closed bounded rectangles are compact in R". which in tum is implied via 1.5.3 by the fact that closed bounded intervals are compact in R. To show that [- a, a], a > 0 is compact, it suffices to prove (by 1.5.7) that for any given e> 0, [- a, a] can be finitely covered by intervals of the form (p - e. p + e). since we are accepting completeness of R. Let n be a positive integer such that a < ne. Let t E [- a, a] and k be the largest positive or negative integer satisfying kt: E; t. Then -n4iOk4iOn and ke4iOt«k+l)e. Thus any tEI-a.a] ~ils in an interval of the form (ke-f!.ke+e). for k--n ..... O•.••• n and hence {eke - f. ke + e)lk ... o. ± I ..... ± n} is a finite covering of (- a. a]. • This theorem is also proved in virtually every textbook on advanced calculus. As is known from calculus. continuity of a function on an interval [a. b] implies uniform continuity. The generalization to metric spaces is the following.
1.5.10 ProposlUon. A continuous mapping cp: MI ..... M 1 • where MI and M2 are metric spaces and MI is compact. is uniformly continuous. Proof. The metrics in MI and M z are denoted d l and d z• Fix e> O. Then for each p E MI' by continuity of cp there exists 8, > 0 such that if d I ( p, q) <8,. then d z(cp(p).cp(q»<e/2. Let D,,/z(PI)•...• D,./z(p.. ) cover the compact space MI and let 8 = min{ 81/2, ... ,8.. /2}. Then if p. q E MI are such thatd l (p,q)<8, there exists an index i,l < i E; n such thatdl(p, PI)< 81 /2 and thus
Thus
One of the most useful analytical applications of Corollary 1.5.7 concerns relatively compact sets in C( M. N). for metric spaces (M, d M ) and (N. d N ) with M compact and N complete.
1.5.11" DeflnlUon. A subset tJ c C( M. N) is called equicOfftillllous lit mo EM, i/givell e> O.there exists 8 > Osuch that wheneverd,.,(m, mol < 8, we have d N ( cp( m ), cp( m0» < e for every cp E ';'1" (8 is independent "of cp). ~1 is called .,.,kMt"""', i/ it is equicontinuous at every point 0/ M.
COMPACTNESS
29
1.5.12 Arzela-Aecoll Theorem. Let (M. d M) alld (N. d N ) be metril' spaces. with M compact and N complete. A set ':1 C C( M. N) is relative(v compact iff it is equicontinuous and all the sets ':1( m ) = (!J!( m) I!J! E ~'t} are relatively. compact in N. Proof. If ~·t is relatively compact. it is totally bounded and hence so are all the sets ·'!f(m). Since N is complete. by Corollary I.S.7 the sets ·~'t(m) are relatively compact. Let (!J! ...... !J!,,} be the centers of the f-disks covering ~1. Then there exists 6>0 such that if d M (m.m')<6. we have d ... lcr,(m). !J!,(m'» ~ F/3. for i-I ..... n aod hence if!J! E \'t is arhitrary. q> lies in one of the e-disks whose center. say is !J!,. so that
d/V( !J!(m).!J!( m'l) :E;; d/V (!J!( m). !J!,(m» + d .... (cr, (m).!J!, (m')).
+ d/V(!J!;(m').!J!(m ' » < E. This shows that ~ is equicontinuous. Conversely, since C( M. N) is complete. by Corollary 1.5.7 we nee(l only show that ~'t is totally bounded. For E> O. find a neighborhood U", of mE M such that for all m'E U",. d/V(q>(m).!J!(m' < E/4 for all cp E ~'t(this is possible by equicontinuity). Let Un/I •...• U"'. be a finite collection of these neighborhoods covering the compact space M. By assumption each ~1(m) is relatively compact. hence ~·t(m.)u ... U ~'t(m,,) is also relatively compact. and thus totally bounded. Let D, /4 (X. )..... D./4(x 4 ) cover th,is union. If a denotes the set of all mappings a: {l ..... n} -+ {I ..... k}. then a is finite and
»
~1=
u
~\.
where \'t.. = {cpE\'tld/V(cp(m,),x,,(i)<e/4 •
.. en
for all But if CPo '" E ,:t.. and
mE
d N (cp( m). '" (m»
:E;;
M. then
d N (cp( m). cp( m,»
+ dN ( X i.e .• the diameter of \'t.. is
mE D,/4(x;)
:E;; E.
a(,)' '" (
i.., I. .... n}
for some i. and thus
+ d N (cp( III, ). Xn(il)
m, ) ) + d>oj ( '" ( 11/, ). '" ( m » < f:
so ':f is totally bounded.
•
The.Heine-Borel theorem then implies the following. 1.5.13 Corollary. If M is a compact metric space. a set ~'t C C( M.R") is relatively compact iff it is equicontinuous and uniform(r hounded (i.e .• llcp(m~1 :E;; constant for all cr E ~. m EM).
30
TOPOLOGY
The following example shows how to use the Arzela-Ascoli theorem. 1.5.14 Example.
Let /,,: [O.IJ -- R be continuous and be such that
1I.(x)1 ~ 100 and the derivatives /,: exist and are uniformly bounded on
10.1[. Prove I. has a uniformly convergent subsequence. We verify that the set Un> is equicontinuous and bounded. The hypothesis is that 1f~(x)1 ~ M for a constant M. Thus by the mean-value theorem. I/',(x)- I.(Y)I~ Mlx-
Yi.
so given t we can choose 8 = tiM. independent of x, y. and n. Thus {/,,} is equicontinuous.1t is bounded because II In II = sUPo .. , .. dln(x>l" 100. • Exercl...
l.SA Show that a topological space S is compact iff every family of dosed subsets of S whose intersection is empty contains a finite subfamily whose intersection is empty. !.SB Show that every compact metric space is separable. (Hint: Use total boundedness.) I.SC Show that the space of Exercise 1.11 is not locally compact. (Hint: look at the sequence I/n).O).) !.SD Show that every closed subset of a locally compact space is locally compact. I.SE Show that S x T is locally compact if h(lth Sand T arc locally compact. 1.5F Let S be a compact topological space and - an equivalence relation on S. so that SI - is compact. Prove that the following conditions are equivalent (d. 1.4.10): (i) The graph P of - is closed in S x S; (ii) - is a closed equivalence relation; (iii) SI - is Hausdorff. I.5G Let S be a Hausdorff space that is locally homeomorphic to a locally compact Hausdorff space (that is, for each u E S. there is a neighborhood of S homeomorphic. in the subspace topology. to an open subset of a locally compact Hausdorff space). Then show that S is locally compact. In particular, Hausdorff spaces locally homeomorphic to R· are locally compact. Is the conclusion true without the Hausdorff assumption?
«
CONNECTEDNESS
31
1.5H Let MJ be the set of all 3 X 3 matrices with the topology obtained by , regarding M3 as R 9 , Let S0(3) = {A e M~IA is orthogonal and det A = -I}. (i) Show that S0(3) is compact. (ii) Let P={QeS0(3)IQ is symmetric} and let cp: Rp2->S0(3) be given by cp(l) = rotation by 17 about the line Ie R3. Show that cp maps Rp2 homeomorphically onto P. 1.51
Let fn: [a, b) -> R be uniformly bounded continuous functions. Set
F" (x ) = [fn ( t ) dt.
(/ ~ x '" h.
u
Prove that
F" has a uniformly convergent subs~qllence,
1.8 CONNECTEDNESS
There are three basic varieties of connectedness that topological spaces may have. These are referred to as arcwise connectedness. connectedness. and simple connectedness. 1.8.1 Definition. Let S be a topological space and I = [0. I] c R, A n are cp in S is a continuous mapping cp: I ~ S. If cp(O) = u. cp( I) = v. we ,WlI' cp joins u and v; S .is called are woe connected if evet:v two points /1/ S can he joined by an arc in S. A space is called locaHy arewise connected if each poiqt has an arcwise connected neighborhood (in the relative topology),
For example. R n is arcwise and locally arcwise connected: 'any two points of R n can be joined by the straight line segment connecting them. A set A c R n is called convex if this property holds for any two of its points. Thus convex sets in Rn are arcwise and locally arcwise connected, A set with the trivial topology is arcwise and locally arcwise connected. but in the discrete topology it is neither (unless it has only one point). 1.8.2 Definition. A topological space S is corrnected if (lj and S are the only subsets of S that are both open and closed . .if subset of Sis ctJM«IMI if it is connected in the relative topology. A COIrIptIMnt A of S is a nonempty connected subset of S such that the only connected .rubset of S containing A ;s A; S is called locaHy corrnected if each point has a connected neighborhood. ' The comptIIfeIIts of a subset T c S are the components of the points of T in the relative topology of T in S.
For example, R n and any convex subset of R n are connected and locally connected. The union of two disjoint open convex sets is discon-
32
TOPOLOGY
nected but is locally connected: its components are the two convex sets. The trivial topology is connected and locally connected. whereas the discrete topology is neither: its components are all the one-point sets. Connected spaces are characterized by the following. 1.8.3 Proposition. A space S is not connected iff either of the following holds. There is a nonempty proper subset of S that is both open and closed. (ii) S is the disjoint union of two nonempty open sets. (iii) S is the disjoint union of two nonempty closed sets. (i)
The sets in (ii) or (iii) are said to dUCOIIIIect S. Proof. If there is a nonempty proper set A that is both open and closed. then S=AU(S\A) with A.S\A open and nonempty. Conversely. if S == AU B with A. B open and nonempty. then A is also closed. and thus A
is a proper nonempty set of S that is both open and closed. The equivalence of (ii) and (iii) is a direct consequence of the definition of closed sets as complements of open sets. • The behavior of these two notions of connectedness under mappings is as follows. 1.8.4 ProposlUon. Iff: S -+ T is a continuous map of topological spaces and S.is connected (resp. arcwise connected) then so is f(S). Proof. Let S be arcwise connected and consider f(s,), f(s2) E f(S) cT. If c: 1-+ S. c(O) == Sl' c(I)'" s2 is an arc connecting Sl to S2' then clearly f 0 c: 1-+ T is an arc connectingf(s,) to f(s2); i.e.• f(S) is arcwise connected. Let S be connected and assume f(S) c U U V. where U and V are open and un V=0. Thenrl(U) andr'(V) are open by continuity off.r I(U)U r I(V)= rl(U U V);:)r 1(f(S»= S. and rl(U)nr I(V)= rl(U n V) ... r I( 0)'" 0. thus contradicting connectedness of S by 1.6.3. Hence f(S) is connected. •
We now prove that arcwise connected spaces are connected. To do so, however, requires the following. 1.6.5 Lemma. The on~v connected sets of R are the intervals (finite, infinite, open. closed, or half-open).
CONNECTEDNESS
33
Let us prove that [a. b[ is connected; all other possibilities have identical proofs. If not. [a. b[ = V U V with U. V nonempty closed sets in [a, b[. Assume that a E V. If x = suP(U). tben x E V since V is closed in [a. b[ and x < b since V. f2J. But then Ix. b[ c V and since V is closed. X E V. Hence x E V n V. a contradiction. , Conversely. let A be a connected set in R. We claim tbat [x. yJc A whenever x. yEA, which implies tbat A is an interval. If not. there exists zE[x,y] witb zEA. But then ]-oo.z[nA and )z,oo[nA are open nonempty sets disconnecting A. • Proof.
1.8.8 Proposition. If Sis arcwi.fe connected then it is connected. Proof. If not. there are two nonvoid. disjoint open sets Vo and VI whose union is S. Let Xo E Vo and XI e VI and let fP be an arc joining' Xo to XI' Then Vo = fP-I(VO) and VI = fP -I(VI ) disconnect [0,1). •
A sf,andard example of a space that is connected but is not arcwise connected nor locally connected, is {(x,y)eR2Ix>0 and y=sin(l/x))U{(O.y)I-I
1.8.7 Proposition. If a space is connected and /oca/fl' arcwise connected, it ;.1 arcwise connected. In particular, a space lo('(J/~l' homeomorphic to A" is connected ifI it ;.1 arcw;se connected. Proof. Fix xeS. The set A - (y e Sly can be connected to x hy an arc) is nonempty and open since S is locally arcwise connected. For the same reason, S\A is open. Since S is connected we must have S\A=f2J; i.e., A = S; i.e.• S is arcwise connected. •
Also, we have the following.
1.8.8 Intermediate Value Theorem. Let S be a connected space and I: S -+ R be continuous. Then I assumes every value between any two values I( u) and/(v). Proof. Suppose I( u) < a < I( v) and I does not assume the value' a. Then V = (uol/(uo) < a) is both open and closed. ill
An alternative proof uses the fact that f( S) is connected in A and therefore is an interval.
34
TOPOLOGY
1.6.9 Proposl8on. Let S be a topological space and Be S be connected. Then (i) (ii)
if Be A c cI( B). then A is connected; if B.. is a fami(y of connected subsets of Sand B.. n B * 0. then B U( U "B.. ) is connected.
Proof. If A is not connected. A is the disjoint union of VI n A and V 2 n A where VI and V2 are open in S. Then from 1.1.9(i), VI n B *0,V2 n B ok0. so B is not connected. We leave (ii) as an exercise. •
1.6.10 Corollary. The components of a topological space are closed. Also. S is the disjoint union of its components. If S is local(y connected. the components are open as well as closed.
1.6.11 Proposl8on.
Let S he a first countable compact Hausdorff space and {An} a sequence of closed. connected subsets of S with An C An I' Then A = n:_IAn is connected.
PI'OOf. As S is normal. if A is not connected. A lies in two disjoint open subsets VI and V 2 of S. If A n n(S\VI )n(S\V2 )*0 for all n. then there is a sequence Un E An n( S\V I )n( S\V2 ) with a suhsequem:e converging to u. As An. S\VI • S\V2 are closed. u E A n(S\V I )n(S\V1 ). a contradiction. Hence some An is not connected. •
Finally we consider the notion of simple connectivity.
1.6.12 Definition. Let S be a topological space and c: 10. I J -+ S a continuous map such that c(O) = c( I) = pES. We call c' a loop in S based at p. The loop c is called COIItracti/Jk if there is a continuous map H: [0. l]x[O.I) -+ S such that H( t, 0) = c(t) and H( t. I) = P for all t E [0. I). (See Fig. 1.6.1).
We think of c,(t) = H(t . .~) as a family of arcs connecting Co = c to c i • a constant arc: see Fig. 1.6.1. Roughly speaking. a loop is contractihle when it can be shrunk continuously to p. The study of loops leads naturally to homotopy theory. In fact. the loops at p can. hy sucl'essively traversing them. he made into a group called the flllle/amelltal ~rollp.
1.6.13 Definition. A .fflll'(' S et'lf'I)' loop in S is contractihle.
H .~;mp~r
connec·ted if S is ('/llI/leefed (lnd
In the plane R 2 there is an alternate approach to simple connectedness. by way of the Jordan curve theorem: namely, that every simple (noninter-
CONNECTEDNESS
35
Figure 1.6.1
seeting) loop in R 2 divides R 2 (that is, its complement has two components). The bounded component of the complement is called the interior. and. a subset A of R 2 is simply connected iff the interior of every loop in A lies in A.
BOX 1.fA CONDITIONS FOR SEPARABILITY
We close this section with a theorem often Invoked in Riemannian geometry (to show that a Riemannian manif"IJ is second countable)' that illustrates very well the interplay hctwecn various notions introduced in this chapter. 1.6.14 Proposition (P. S. AlexillldroJf). An arcwise-connected locally compact metric space is separable and hence is second countable. Proof. (A. pnuger (1957)). Since the metric space M is locally compact, each m E M has compact neighborhoods that are disks. Denote by rem) the least upper bound of the radii of such disks. If r(m)'" 00, since every metric space is first countable, M can be written as a <;ountable union of compact disks. But since each compact metric space is separable (Exercise 1.5B), these disks and also their union will be separable, and so the proposition is proved in this case. If r( mil) < 00, since r(m)~r(mo)+d(m.mo). we see that r(m)
36
TOPOLOGY
hoods of m. Fix mo E M and put Ao = Gmo' and inductively. define A".I - U: e A Gm • Since Mis arcwise connected. every point me M can be connected by an arc to mo. which in tum is covered by finitely many Gm • This shows that M - U :"_oA". But Ao is compact. So all that remains to be shown is that the A" are all compact. Assume inductively that A" is compact and let {m;} be an infinite sequence of points in A". I' But then there exist m; E A" such that m; E Gm ;. Since A" is assumed to be compact there is a subsequence that converges to a point m' E A". But
m;,
d(ml' m') < d(ml' m;)+d(m;. m') < r(m;)/2+ d(m;. m') < r(m')/2+ 3d(m;. m')/2
Hence for· i" big enough, all m;, are in the compact set (n E Mld(n. m')< 3r(m')/2}. so m;, has a subsequence converging to a point m. The preceding inequality shows that in fact mEA". I' By the Bolzano-Weierstrass theorem, A". 1 is compact. •
Exercl... 1.6A Let M be a topological space and H: M -+ R continuous. Suppose e E H(M). Then show H-1(e) divides M; that is, M\H-1(e) has at least two components. J.68 Let 0(3) be the set of orthogonal 3x3 matrices. Show that 0(3) is not connected and that in fact is has two components. 1.6C Show that S X T is connected (locally connected. arcwise connected. locally arcwise connected) iff both Sand Tare. (Hint: For connectedness write S X T- U tET[(S X {/})U({So}X T)] for So E S fixed and use 1.6.9{ii).) 1.60 Show that S is locally connected if every component of an open set is open. 1.6E Show that the quotient space of a connected (locally connected. arcwise connected). space is also connected (etc.). (Hin/: For local connectedness use Exercise 1.60 and show that the inverse image by ." of a component of an open set is a union of components.)
CHAPTER
2
Banach Spaces and Differential Calculus
Manifolds have enough structure to allow differentiation of maps between them. To set the stage for these concepts requires a development of differential calculus in linear spaces from a geometric point of view. The goal of this chapter is to provide. such a perspective. Perhaps the most important theorem for later use is. the implicit function theorem. A fairly detailed exposition of this topic will' be given with examples appropriate for use in manifold theory. However. the hasic language of tangents. the derivative as a linear map. and the chain rule. while elementary. are important for developing intuitive and linguistic skills necessary for mastering manifold theory. The main goal is to develop the theory of finite-dimensional manifolds. However, it is most efficient to do the infinite-dimensional theory simultaneously. It is also very instructive to see the compari!.on of the finite-and infinite-dimensional theory. To avoid being sidetracked hy infinite-dimensional technicalities at this stage, however. some (If the necessan [wll'tional analysis background is reserved for Appendix A .md other topics special to the infinite-dimensional case are presented in boxes. With this arrangement readers who wish to concentrate on the finite-dimensional theory can do so with a minimum of distraction. 2.1
BANACH SPACES
It is assumed the reader of this book is familiar with the concept of an abstract vector space over the real or complex numbers. Banach spaces are,
37
38
BANACH SPACES AND DIFFERENTIAL CALCULUS
vector spaces with additional structure. Most of this book is concerned only with finite-dimensional spaces. However, much of the theory is really no harder in the general case, and the infinite-dimensional case is essential in more advanced work. Thus it makes sense to work in the setting of Banach. spaces. In addition, although the primary concern is with real Banach spaces, the basic concepts for complex spaces are introduced with little extra effort. 2.1.1 Definition. A IfOI7JI on a real (complex) vector space E is a mapping from E into the real numbers. 11'11: E .... R; e ...... lIeli. such that (N I) lIeli ~ 0 for all e E E and lIeli = 0 iff e = 0 ( positive definiteness); (N2) IIAeli -IAllleli for all e E E and AE R (homogeneity); (N3) lIe l + e211 E;; lIedl+ lIe211 for all e l •. e 2 E E (triangle inequality). The pair (E.II·II) is sometimes called a normed space. If there is no danger of confwion. we sometimes jwt say "E is a normed space". To distinguish different norms. different notations are sometimes used. e.g .. 1I·IIE.II·II,.III·III. etc.• for the norm.
The triangle inequality (N3) has the following important consequence:
IlIedl-lIezlll E;; lIe l - e211 for all el •e 2 E E. which is proved in the following way:
lIe 211 = lIe l +(e2 -e, )11 E;; lIedl+lle l - e21i lIedl = lIe 2+ (e l - (2)11 E;; lIe211 + lIel - e211 so that both lIe 211-lIedl and lIedl-lIe211 are smaller than or equal to lIe l - e211. If (N I) in 2.1.1 is replaced by (N I)'
lIeli ~ 0 for all e E E and e = 0 implies lIeli = O.
the mapping 11'11: E .... R is called a seminorm. For example. R" with the standard norm 2 2)1/2 _( ) 11,1"11= ( xl+···+x" . , 1 " - xl.· ... x"
is a normed space. Actually. the standard norm on R" comes from a more special structure defined as follows. 2.1.2 Definition. An iIrMr prodltct on a real vector space E is a mapping (·.·):EXE ..... R:
(e l .e2) ...... (e l .e2)
BANACH SPACES
39
such that
(e,e, + e2)'" (e,e,)+(e,e 2); (e,cre,) - cr(e,e,); (e"e2) - (e2,e,);
(11) (12) (13) (14)
(e,e)~Oand(e,e)=Oiffe=O.
The standard inner produc.t on R" is (x, Y)
=
L" X,Y/ ;-,
and (I I )-(14) are readily checked. For complex vector spaces the definition is modified slightly as follows.
2.1.2' DennlUon. A complex inner product (or a Hermitilut inMr produet) on a complex vector _'pace E is a mapping (-,:):EXE-+C such that (CII)
(e,e,+e2)=(e,e,)+(e,e 2 );
(CI2) (C13) (CI4)
(cre,e,) "" cr(e,e,); (e"e 2) = (e 2.e,) (so (e.e) is real); . (e.e)~O and (e,e)=O iff e=O.
These properties are to hold for ail e, e" e2 E E and a E C; denotes complex conjugation. Note that (CI2) and (CI3) imply that (e"cre 2) =ci(e,.e 2 ). Properties (CI 1)-(CI3) are also known in the literature under the name sesquilinearily. As customary. for a complex number a we shall denote hy Re a = ! (a + ii), 1m a"" (1/2i)(a - ii),lcrl = (crii)1/2, its real and imaginary parts and its absolute value. The standard inner product on Cit = C x ... x C is n
(z, w) =
L
;-, z;w;
and (CIl)-(CI4) are readily checked. Also C" is a normed space with
" Iz,12. liz II 2= L
;-,
40
BANACH SPACES AND DIFFERENTIAL CALCULUS
In R" or C". property (N3) is a little harder to check directly. However, as we shall show in Proposition 2.1.4, (N3) Collows from (11)-(14) or (CII )-(CI4). In a (real or complex) inner product space E. two vectors '1.'2 e E are called orthogonal denoted'I.l '2' iC ('1"2) == O. For any set AcE. A.l (,E EI(', x) -0 for all xE A} is called the orthogonal complement of A. Two sets A. BeE are called orthogonal, denoted A.l B. if (A, B) = 0; i.e.• '1 .1 '2 for all '1 E A and '2 E B. Next we tum to a fundamental inequality valid in inner product spaces.
2.1.3 Propoaltlon. (CauchJ'-Schwarz Inequality). In a (real or complex) inner product space we have
Equality holds iff '1,'2 are linearly dependent. Proof.
It suffices to prove the complex case. If a, P E C we have
,a., + P'2) -laI2(." •• )+2Re(aP(""2»)+ IPI 2('2"2)'
0< (a' i + p.2 If a'" ('2"2)'
p ... - ('1"2) this becomes 0< ('2' '2)2('1' ") - 2('2' '2) 1(""2) 12
+ 1(""2)1 2('2"2)'
or
('2"2)1(""2)1 2 EO; ('2,e2)2(.",.).
'2
If == 0 equality results in the statement of the proposition and there is nothing to prove. If'2 .. 0, the term ('2' '2) in the preceding inequality can be cancelled since ('2"2)" 0 by (CI4). Taking square roots yields the statement of the proposition. Finally. equality results iCC ('2' '2)" (""2)'2 ==0. •
2.1.4 PropoaItIon. Let (E. ( .•. » be a (real or complex) innt'r product space and set
Then (E.II·II) is a normed space. Proof. (NI) and (N2) are trivial veriCications. As for (N3). the Cauchy-Schwartz inequality and the obvious inequality Re«.,.e2»EO;
BANACH SPACES
II., + .211
2
41
-1I.dI 2 +2 Re«.".2»+11.211 2 ~
lI.dl 2+21(.".2)1 +11. 211 2
~
lI.dl 2+ 211.,1111.211+11.211 2
=
(lIedl+lle211~)~ •
Some other useful facts about inner products are given next.
2.1.5 Propolltlon. Let (E.(·,·» be an inner product space and 11·11 the corresponding norm. Then (i)
(Polarization)
(ii) (Parallelogram law)
Proof.
(i)
II., + .211
2
-lie, - e211 2+ ille, + ie2112 - ille, - ie 2 11 2 =
11."1 2 +2 Re«.".2»+11.211 2-lIedl2 +2 Re«e".2» -11.211 2 + ;11.,11 2 +2; Re«e,. i.2»+;lIe2112 - ill.dl 2 +2; Re«.,. ;.2»- ille l ll 2
=
4 Re«.".2»+4; Re( -
;(e,.e 2 »
"" 4 Re«.".2»+4; Im«4!".2» =
4(."e 2 )·
42
BANACH SPACES AND DIFFERENTIAL CALCULUS
The real case is proved in a similar way.
lie, + e211 2+ lie, - ezll z = lIedl z+ 2 Re( (e,. ez») +lIe 2 11 2
(ii)
+ lie II 2-2 Re«e,.e 2 »+lIe zII 2 ... 21ledl 2+21le zll z. • Not all norms come from an inner product. For example in R".
" Ix;1 IUxlll = E i-I
is a norm. but there is no inner product with this as norm. since this norm fails to satisfy the parallelogram law (see Exercise 2.18' for a discussion).
A 2.1.8 Propo8lt1on. Let (E.II·II) be a normed (seminormed) space and let d(e,.e2) =
lie, - e211. Then (E. d) is a metric (pseudometric space).
Proof. The only nonobvious verification is the triangle inequality. By (N3) we have d(e,.~)
= lie, - e~1I = lI(e, - e2)+ (e z- eJ)IIEtlle, - 1'211+111'2 - elll '"' d(e,. e 2 )+ d(e 2• ell.
•
Thus we have the following hierarchy of generality more general inner product c normed c metric c topological spaces spaces spaces spaces +- more special Since inner product and normed spaces are metric spaces. we can use the concepts from Chapter I. In a normed space. (N I) and (N2) imply that the maps (e,.e 2 ) ..... e,+e2. (a.e) ...... ae of ExE-E. and CxE-E. respectively. are continuous. Hence for eo E E. ao E C. a o ... 0 fixed. the mappings e ...... eo + e. e ...... aoe are homeomorphisms. Thus U is a neighborhood of the origin iff e + U = {I' + xix E U} is a neighborhood of e E E. In other words all the neighborhoods of e E E are sets that contain translates of disks centered at the origin. This constitutes a complete description of the topology of a normed vector space ( E .11'11 ).
BANACH SPACES
. 43
Finally. note that the inequality 1111'.11-111'2111 lit II', - '211 implies Jhat the norm is uniformly continuous on E. In inner product spaces, the Cauchy-Schwartz inequality implies the continuity of the inner product as a function of two variables. 2.1.7 DeflnlUon. Let (E.II·II) be a normed space. If the correspondinf{ metric d is complete. we say (E.II·II) is a Batulch sptlCe. If (E. (' •. » is an inner product space whose corresponding metric is complete. we say (E. ( " . » is a HilM" sptlCe. For example. it is proved in books on advanced calculus that R" is complete. Thus. RIt with the standard norm is a Banach space and with the standard ·inner product is a Hilbert space. Not only is the standard norm on R n complete, but so is the nonstandard one n
IlIxlll=
L
Ix,l·
i-I
However. it is equivalent to the standard one in the following sense. 2.1.1. Definition. Two norms on a vector space Earl' equivlllelft if they induce the same topology on E. 2.1.9 ProposWon. Two norms 11'11 and 111'111 on Earl' equitlolent iff there is a constant M such that M- '1111'111 lit lieU lit Milleill for all I' E E.
Proof. Let D,'(x)={yEEIIIY-xll.s;;r}. D/(x)={yEElllly-xlll.s;;r} denote the two closed disks of radius r centered at x E E in the two metrics defined by the norms 11'11 and 111'111. respectively. Since neighborhoods of an arbitrary point are translates of neighborhoods of the origin. the two topologies are the same iff for every R > O. there exist /II,. M. > 0 such that Df" (0) c Dk (0) C The first inclusion says that if IlIxlll lit IIxll.s;; RIM,. Thus. if 1'''' O. then
D~t: (0)
M,. then IIxll lit R. i.e .• ir IlIxlll lit I. then
II.
I' 111'11 R 1111'111 = lIIelll.s;; MI' "
that is. 111'11 lit (R I M\) 1111'111 for all I' E E. Similarly. the second inclusion is equivalent to (MdR) 1111'111 lit 111'11 for all I' E E. Thus the two topologies are
44
BANACH SPACES AND DIFFERENTIAL CALCULUS
the same if there exist constants NI > O. N2 > 0 such that
for all , E E. Taking M = max( N2 .1/ N I ) gives the stalemenl or Ihe proposition. •
If E and F are normed vector spaces. the map 11'11: E x F -+ A defined by 11(.,.')11-11.11+11.11 is a norm on Ex F inducing the product topology. Equivalent norms on Ex Fare (.,.') .... max(I'",1I.11> and ( ••••) .... (11.11 2 +11.11 2 )1/2. The normed vector space E X F is usually denoted by eeF and called the dir.ct sum of E and F. Note that EeF is a Banach space iff both ' E and F are. These statements are readily checked. In finite dimensions some special things occur as follows. 2.1.10 Propo8lt1on. ut E be a finite-dimensional real or complex vector
space. Then (i)
there is a norm on E.
(ii) all norms on E are equivalent; (iii) all norms on B are complete.
Proof. Let ., .... ,." denote a basis of E. where n is the dimension of E. (i) A norm
on E is given for example by 11-11-
"
E loll.
n
E 0/./.
where - -
i-I
Note that this norm coincides with the A" or e"-norm
" loll. 111(0' ..... 0")111= E i-I
(ii)
Let 11'11' be any other norm on E. If
.= E "
the inequality
a l• i
and
I-I
1=
n
L bi.,. i-I
111.11'-11/11'1< II. -/1r..;; L" la i - billle;II' i-I
..;; max 01 •. :~:( a l , .... 0")- (b l , ... ,b")\\\. 1..,/< "
BANACH SPACES
45
shows that the map (xl •...• x")ec" ....
II;~IX;'.r ero.co[
is continuous with respect ~o the III·III-norm on C". Since the set S = .{ x e C "\lIIxlll"" I} is closed and bounded. it is compact. The restriction of this map to S is a continuous. strictly positive function. so it attains its minimum MI and maximum M2 on S; that is.
for all (xl •...• x")eC" such that IIKx' •...• x"~1I = I. Thus
or
for
Taking M - max(M2 .I/M1). Proposition 2.1.9 shows that 11·11 and equivalent. (iii) It is enough to observe that
n·U' are
" (XI ••.•• X") e C" .... Ex'" e I:: i-I
is a nonn-preserving map (i.e.• an isometry) between (e".III·III> and (E.II·II). •
•
The unit spheres for the three common norms on R 2 are shown in Fig.
2.1.1 .. The foregoing proof shows that compactness of tbe unit sphere in a finite-dimensional space is· crucial. This fact is exploited in the following box.
If.
IANACH '/lAC'S AND DIFFERENTIAl. CAI.CUI.US ..
Y
Y
x
muCx.
Y
x
x
y)
figure 2.1.1
BOX 2.fA A CHARACTERIZATION OF FINITE-DIAfENSlONAL SPACES
2.1.11 PropoaIIIon. .A normed vector space is finite dimensional iff ;t ;s locally compact. (Local comptlCtneu is equivalent to the fact that the closed un;t disk is compact.)
Proof. If E is finite dimensional. the previous proof of
(iii) shows that E is locally compact. Conversely, assume the closed unit disk D,(O) c ~ is compact and let {D I / 2 (x,)I; -I, .. .• n} be a finite cover with open disks of radii 1/2. Let F - span (Xl.' .. ,x,,). Then F is finite dimensional. hence homeomorphic to CA (or AA) for some k < n. and thus complete. Being a complete subspace of the metric space (E.II'II). it is closed. We shall prove F - E, Suppose the contrary. that is. there exists "E E, V. ~ F. Since F- cl(F). the number d - inf{II"- ell Ie E F} > O. Let r > 0 be such that D,(,,)nF.flJ. The set D,(,,)nF is closed and bounded in the fmite-dimensional space F. and thus is compact. Since inf{II" - ell I e E F}'" inf{lI" - '111' E D,( ,,) ~ F} and the continuous function. E D,( ,,)n F .... 11" - '11 E )0. co[ attains its minimum. there exists a point '0 E D,( ,,)n F such that d'"" II" - '011. But then there exists Xi such that
,,- eo II 1 II 11"eoll- Xi < 2' so that
'
BANACH SPACES
47 .
Since eo + 110 - eollx; E F. it follows that 110 - "" - ( II v - eo II )x;1I ~ d. which is a contradiction. •
2.1.12 Exampl... A. Let X be a set andF a nonned vector space. Let B(X. F) = {f: X-+FlsuPxExll/(x)lI
2.1.13 ProposlUon. Let E be a normed vector space. F a closed subspace. ElF the quotient vector space and '/1': E -+ ElF the ctlllOlliCill projection ' delined by '/I'(e) = (e) = e + F E ElF. Then (i) The mapping 11'11: EIF-+R.
II [ e ]11 = inf{ lie + 01110 E F} delines a norm on ElF. (ii) '/I' is continuous and in lactthe topology on ElF defined by (he norm coincides with the quotient topology. In particular '/I' is open. (iii) II E is a Banach space. so is ElF. Proof. (i) Clearly lI[e]1I ~ 0 for all [eJE ElF and IIrOl1l = inf{lIolilDE F) ... O. If lI[e]lI- 0, then there exists a sequence {o,,} C F such that lim ...... oolle + 0,,11- O. Thus lim" .... 00 v" = - e and since F is closed e e F; i.e., Ie] - O.
48
BANACH SPACES AND DIFFERENTIAL CALCULUS
Thus (NI) is verified and the necessity of having F closed becomes apparent. (N2) and (N3) are straightforward verifications. (ii) Since 11l~]II" II~II, it is obvious that lim"-.""e,, = e implies lim" _ """.(4111 ) -lim" _ ",,[41,,]'" [411 and hence". is continuous. Translation by a fixed vector is a homeomorphism. Thus to show that the topology of ElF is the quotient topology, it suffices to show that if [0] e U and". - I( U) is a neighborhood of zero in E, then U is a neighborhood ot [0] in ElF. Since ".-I(U) is a neighborhood of zero in E, there exists a disk D,(O) c ".-I(U). But then "'(D,(O»cU and ".(D,(O»={[ell~eD,(O)}-{[ellll[e)lI
2.1.14 Deftnltlon. The closed subspace F of the Banach space E is said to be IpNt. or ~ed, if th~re is a closed subspace GeE such that E-FeG.
BOX 2.1. SPLIT SUBSPACES
Definition 2.1.14 implicitly asks that the topology of E coincide with the product topology of FeG. We shall show in Box 2.2C that this topological condition can be dropped; i.e., F is split iff E is the algebraic direct sum of F and the closed subspace G. We note that if E - FeG then G is isomorphic to ElF. However, F need not split for ElF to be a Banach space, as we proved in 2.1.13. In finite-dimensional spaces. any subspace is closed and splits; however, in infinite dimensions this is false. t Box 2.2B and Appendix B give some general criteria useful in nonlinear analysis for a subspace to be split. But the simplest situation occurs in Hilbert spaces.
2.1.15 Propoeltlon. If E is a Hilbert spac'e an.! F a closed subspace then E"'"' FeF J.. Thus every closed subspace oj a Hilbert space splits. tFor example ('0 c I"" is Dot complemented. If every subspace of a Banach space is complemented, the space must be isomorphic to a Hilbert space. See J. LiDdmstrauss aDd L. Tzafriri (1971).
BANACH SPACES
49
The proof of this theorem is done in three steps, the first two . being important results in their own right.
m.. (
2.1.18 ...... F. ";".IIU tIIIII 1III/qIIIMB of ",btbrttll IIOrIIf .k".,,18 ill doutI C"()'.... elC s"s). 1/ C is a closed. convex set (i.e., x, "E C and 0" t "I implIes IX +(1- t)" E C) in E. then there exists a unique eo E C such that lIeoll = inf{lIellle E C}.
Proof. Let {d ... inf{lIellle E C}. ·Then there exists a sequClnce {ell} satisfying d" lie" II 2 < d + lin; h~nce lIe,,1I 2 -+ d. Since (e" + e",)/1 E C. C being convex. it follows that lI(e" + e", )/211 2 ~ d. By the parallelogr~m law .
did I . <-+-+-+--d 2 2n 2 2m
that is, {ell} is a Cauchy sequence in E. Let lim" • yell ,. eo. Continuity of the norm implies that Id "" lim" _ ",,'II ell II = lIeoli. and so the existence of an element of minimum norm in C is proved. Finally. if is such that lIeoll""lIeoll ... {d. the pa,rallelogram law implies
eo
d
d
""'-+--d=O' ... 2 2 ' i.e.• eo =
eo· •
2.1.17 Lemma. Let FeE. F. E be a closed subspace there exists a nonzero element eo E E such that eo .1. F.
0/ E.
Then
Proof. Let e E E, e ~ F. The set e - F - {e - 01" E F} is convex and closed. so by the previous lemma 'it contains a unique element eo ... e - "0 E e - F of minimum norm. Since F is closed and e ~ F. it follows that eo • O. We shall prove that eo.1. F.
50
BANACH SPACES AND DIFFERENTIAL CALCULUS
Since eo is of minimal norm in e - F. for any "E F and we have
~E
C(R).
If ~,.. a(e o• ,,). a E R. a"" 0, this becomes
for all "E F. a E R. a "" O. This forces (v. eo) = 0 for all vEl'. since if - 2/11 vII 2 < fJ < O. the preceding expression is negative. • Proof of Proposition 2.1.15. First of all. it is easy to see that F J. is closed (Exercise 2.1C). We now show that FeF J. is a closed suhspace. If D,;j
{en + e~} c FeF J. .{e n} C
F.(e~}
c F J..
the relation
shows that {en + e;"} is Cauchy iff both {en} C F. and {e~} C F.J. are Cauchy. Thus if {en + e~} converges. then there exist e E F. e' E F.J. such that Iimn_ooen=e.limn_ooe~=e'. Thus limn_oo(en+e~)==e+ e'EFeFJ.. If FeF.J. .,. E. then by the previous lemma there exists eo E E. eo ~FeF.J..eo"O.eo.J..(FeF.J.). Hence eoEF.J. and eoEF so that (eo. eo) -lleoll2 - 0; i.e., eo - O. a contradiction. • Exercl... 2.IA
Show that a normed space is an inner product space iff the norm satisfies the parallelogram law. (-Hint: Use the polarization identities over R and C to guess to corresponding inner products.) Conclude that IlIxlil = Elxil on R n does not arise from an inner product. 2.1 B Let Co be the space of real sequences {an} such that an --+ 0 as n --+ 00. Show that Co is a closed subspace of the space c" of bounded sequences (see Example 2.1.I2A) and conclude that Co is a Banach space.
BANACH SPACES
2.1C
51
Let EI be the set of all C l functionsf: [0.1) -- H with the norm
IIfll= sup If(x)l+ sup 1f'(x)l . .• e (0.1)
.\
F
In.11
Prove that EI is a Banach space. Let Eo be the space of CO maps f: [0. I) -> H. as in 2.1.12. Show that the inclusion map EI -+ Eo is wmpal·t; i.e.. the unit ball in EI has compact closure in Eo. (Hint: lise the Arzela-Ascoli theorem.) 2.1 D Let (E. (-. be an inner product space amI A. B subseis of E. Let A + B = {a + bla E A. b E B}. Show that '(i) Ac B implies B.I. c A.I.; (ii) A.I. is a closed subspace of E; (iii) A.I. = (cl(span(A))).I.. (A.I. ).1. = cl(span( A»; (iv) (A+B).I.=A.I.()B.I.; (v) (cl(span(A»()cl(span(B»))l =Al +B J (not neces~arily a direct sum). 2.1 E A sequence {en} c E. where E is an inner product space. is said to be weakly convergent to I' E E iff all the numerical sequences (v. en) converge to (v.e) for all vEE. Let f2(C)={{an>ianEC and E:'_lla n I2 < co} and put ({an}.{bn }) = E:'_la"f;n. Show that (i) in any inner product space convergence implies weak convergence; (ii) f 2(C) is an inner product space; (iii) the sequence (1.0.0.... ).(0.1.0•... ).(0.0. I •... ). . .. is not convergent but is weakly convergent to 0 in f 2 (C). Note.f 2 (e) is in fact complete. so it is a Hilbert space.'The ambitious reader can attempt a direct proof or consult 1\ book on real analysis such as Royden [t 968). 2. t F Show that a normed vector space is a Banach space iff every ~bsolutely convergent series is convergent. (A series E:'_IXn is call~ absofutely convergent if E:'_ I IIxnll converges.) 2.1 G Let E be a Banach space and FI c F2 C E be dosed subspaces such . that F2 splits in E. Show that FI splits in E iff FI splits in F2. 2.1 H Let F be closed in E of finite codimension. Show that if G is a subspace of E containing F. then G is closed. 2.11 Let E be a Hilbert space. A set {e j } " I is called orthonormal if (e j • ei ) = 8, ,. the Kronecker delta. An orthonormal set {e,}, e I ,is (i) (ii)
-»
52
BANACH SPACES AND DIFFERENTIAL CALCULUS
called a Hilbert basis. if c1(span{ej }j e I) = E. (i) Let {_/},., be an orthonormal set and (,'o ..... ,,) be any finite subset. Show that
for any I' E E. (Hint: 1" - I' - Ej_,( e. e j )e j is orthogonal to all (ej[j=I ..... n}.) , I , (ii) Deduce from (i) that for any positive integer n. the set {i E 111(e.ej)l> lin} has at most n 2 11ell 2 elements. Hence at most countable many i E 1 satisfy <e, 1',) '* 0, for any e E E. (iii) Show that any Hilbert space has a Hilbert hasis. (Hint: Use Zorn's lemma (Appendix A) and 2.1.17.) (iv) If {ej}je 1 is a Hilbert basis in E,e E E. and {e,} is the
L I(e.e j-I
(v)
(vi)
)1 2 = 111'112.
j
'
(Hint: If e'=Ej_l(e.ej)ej • show that (ej.e-e')=O for all i E J and then use maXirriality of {e/}/.,.) Show that E is separable iff any Hilbert basis is at most countable. (Hint: For the "if part," show that {E._,a"l"la,,ale + ib", ale' bIt rational} is dense in E. For the "only if' part show that since lIe j - ejll2 == 2, the disks DI/{2(e j ) are all dis-
joint.) If E is a separable Hilbert space it is algebraically isomorphic either with C" or 12(C) (R n or 12(R» and the algebraic isomorphism can be chosen to be norm preserving.
2.2 LINEAR AND MULTILlNEAA MAPPINGS This section deals with various aspects of linear and multilinear maps between Banach spaces. We begin with a study of continuity and go on to study spaces of continuous linear and multilinear maps and some related fundamental theorems of linear analysis.
2.2.1 PropoIItIon. Let A: E -+ F he a lineal map of normed spaces. Then A is continuous if and on(v if there ;\' ,J (.onstco,,, M> 0 such that
LINEAR AND MUL TIL/NEAR MAPPINGS
53
Proof. 'Continuity of A at ~o E E means thaI for O. there e"isls '
p > 0 such that
A( ~o + D,,(OE»)
C
A~o + D,(OF)
(OE denotes the zero element in E and D,(O)E denotes the closed disk of radius s centered at the origin in E). Since A is linear. this is equivalent to: if II~II .. P. then IIA~II" r. If M == pro continuity of A is thus equivalent to the following: lIellE .. 1 implies IIA~IIF" M. which in turn is the same as: there exists M> 0 such that IIA~IIF" MII~IIE' which is seen by taking ~/II~IIE in the preceding implication. • Because of this proposition one says tbat a continuous linear map is bounded,
2.2.2 Proposition.
If E is finite dimensional and A: E
-+
F is linear. then
A is continuous.
Proof. Let ~I ..... e" be a basis for E. Letting M, = max(IIAe"I..... IIA~nll) and ~ = a'~1 + ... a"e". we see that
IIAell = lIa'A~, + ... anA~nll "la'IIiAetll+ ... + lo"IIIAe,,1I .. M,(la'I+," +Ianl)· Since E is finite dimensional. all norms on it are equivalent. Since IlIelll = 1:la i l is a norm. it follows that lIIelll" Cllell for a constant C. Let M = M,C and use 2.2.1. •
2.2.3 DefinlUon. If E and Fare normed spaces and A: E
->
F is a
continuous linear map. let the operator no"" of A he J(:f;ned by
(which is finite by 2.2.1.) Let L( E. F) denote the .~pace of all continuous linear maps of E to F. If F- C (resp. R). then L(E.C)(resp.I.(E.R» is denoted by E· and is called the complex (resp. retII) dIIIII space (~r E. (It will always be clear from the context whether L(E. F) or E· meolll' the real or complex linear maps or dual space; in most of the work later in this book it will mean the real case). '
'
54
BANACH SPACES AND DIFFERENTIAL CALCULUS
A straightforward verification gives the following equivalent definitions ofliAII: IIAII=inf{M>OIIiAell~Mllell
forall
eEE}
... sup{ IIAelllliell ~ I} - sup{ IIAelllllell '"' I} In particular,lIAell ~ IIAliliell. If A E L(E, F) and BE L(F,G), where E, F, and G are normed spaces, then '
which shows that
liB All ~ IIBIlIiAIl. 0
Equality does not hold in general. A simple example is with E = F = G = R 2, A(x, y) '"' (x,O), and B(x, y) = (0, y), so that BoA = 0 and IIAII = IIBII = I.
2.2.4 Proposition. L(E, F) with the norm just defined is a normed space. 11 is a Banach space if F is. Proof. Clearly IIAII ~ 0 and 11011 = O. If IIAII = 0, then for any e E E.IIAell ~ IIAlillell- 0, so that A = 0 and thus (N I) (see Definition 2.1.1) is verified. (N2) and (N3) are also straightforward to check. Now let F be a Banach space and {An)C L(E. F) be a Cauchy sequence. Since IIA"e - Amell ~ IIAn - Amllilell for each e E E, the sequence (Aile} is Cauchy in F and hence is convergent. Let Ae = Iimn_ooAne. This defines a map A: E -. F, which is evidently linear. It remains to be shown . that A is continuous and IIAn - AII-' O. If e> 0 is given, there exists a natural number N( e) such that for all m, n ~ N(e) we have IIA" - Amll < e. Ifllell ~ I, this impliesllAne - Amell < e and now letting m -. 00, it follows that IIA"e - Aell ~ e for all e with lIell ~ I. Thus A" - A E L(E, F), hence A E L(E, F) and IIAn - All ~ e for all n ~ N(e); i.e.,IIA" - AII-' O. •
If a sequence {A,,} converges to A in L(E, F) in the sense that IIA" - AII-' 0, i.e. if A" -. A in the norm topology, we say An -. A in norm. This phrase is necessary since other topologies on L(E, F) are possible. For example, we say that A" -. A Slrongly, if A"e -. Ae for each e E E. Since IIA,.. ""A,, - Allil-ib norm convergence implies strong convergence.
A_"
LINEAR AND MUL TILINEAR MAPPIflJGS
55
The converse is false as the following example shows. Let E = 12(R) = = r:~_.a~bn. Let
{{a .. }Ir:~_.a~ < oo} with inner product ({an},{bn})
The sequence {An} is not Cauchy in norm since IIA" - A ",II = Ii. hut if e'" {am}' An(e) == (en' e) = an -+ O,i.e., An -+ 0 strongly. If both E and Fare finite dimensional, strong convergence implies norm convergence. (To see this, choose a basis e., ... ,en of E and note that strong convergence is equ~valent to Ake; -+ Ae; as k -+ 00 for i-I, ... ,n. Hence max;IIAe;1I = IIIAIiI is a norm yielding strong convergence. But all norms are equivalent in finite dimensions.)
BOX 2.2A
DUAL SPACES
Recall from elementary linear algebra that the dual space of C n(R n) may be identified with itself. For general Banach spaces this is no longer true (see Appendix B). An important result asserts that this fact still holds for Hilbert space.
2.2.5. RI..z RepresentaUon Theorem. Let E he a real (resp. complex) Hilbert space. The map e"'" (·,e) is a linear (resp. anlilinear t ) norm-preseroing isomorphism of E with E·; for short. E == E·. Let/, = (-,e). Then 11/,11 = lIell and thus/, E: E*. The map A: E·, Ae = /, is clearly linear (resp. antilinear), norm preserving,
Proof.
E
-+
and thus injective. It remains to prove surjectivity. Let 1 e E· and ker 1- (e e EIf( e) - oj, ker I is a closed subspace in E. If ker/=E, then/'""O·and/-A(O) and there is nothing to prove. If ker I"" E, then by Lemma 2.1.17 there exists e" 0, such that e 1. ker I. Then we claim that 1 = A (f( e)e IlIeIl 2 ). Indeed, any vEE can be written as
., = ., -
1(.,) .. + 1(.,) e and v I( v) e E ke 1 • I( e ) '" I( e ) - I( e) r.
tA map A: E .... F between complex vector spaces is called anlilinear if A (.e + e')-
Ae+Ae', andA(cre)-aAe.
56
BANACH SPACES AND DIFFERENTIAL CALCULUS
Thus. in a real Hilbert space E every continuous linear function /: E ..... R can be written
for some eo e E and II/II = lIeoli. In a general Banach space E we do not have such a concrete realization of E*. However one should not always attempt to identify E and E*. even in finite dimensions. In fact, distinguishing these ' spaces is fundamental in tensor analysis. We have a canonical map ;: E ..... E** defined by ;(e)(/) -/(e). (Pause and look again at this strange but natural formula: lee) e E** - (E*)*, so 1(,) is applied to the element I e E*). It is easy to check that; is norm preserving. One calls E reflexive when; is onto. Hilbert spaces are reflexive, as is seen by using 2.2.5. Next we shall discuss integration of vector valued functions. We shall require the following. 2.2.8 Lemma. (Lbv", Extmsiolr Tlteorem). Let E, F. and G be normed vector spaces where (i) FeE (ii) G is a Banach space and (iii) T e L(F, G). Then the closure cl( F) of F is a normed vector subspace of E and T can be uniquely extended 10 a map Te L(c1(F). G). Moreover II Til = 111i1. Proof. The fact that c1(F) is a linear subspace of E is easily checked. Note that if T exists it is unique by continuity. Let us prove the existence of T. If e e c1(F). we can write e = limn_""e n, where en e F, SO that
which shows that the sequence (Te.> is Cauchy in the Banach space G. Let Te==lim"_""Te,,. This limit is independent of the sequence {en}' for if e ... lim ,~. then
liT,,, -
Te~lI.,:; II Til (lien - ell+lle - e~II).
which proves that limn_",,(Te n ) = Iimn_",,(Te~). It i~ simple to check the linearity of f. Since Te = Te for e E F (because, = Iim,,_ooe), T is an extension of T. Finally. IITell = II lim (Te n )1I = lim IITe"II":; IITII lim lIenll = II Til lIeil n-oo
" .... IX)
n-oo
LINEAR AND MUL TlLINEAR MAPPINGS
57
shows that TE L(c1(F). G) and ufil "it UTU. The inequality UTU "it ufil is obvious since T extends T. • ' As an application of this lemma we define a Banach space valued integral that will be of use later on. Fix the closed interval [a. bJc R and the Banach space E. A mapj: [a. b) -+ E. is called a .~tep jUllction. if there exists a partition a -'0 <'1 < '2 < ... <'" = b such that j is constant on each interval ['I' '/+1[' Using the standard notion of a refinement of a partition it is clear that the sum of two step functions and the scalar multiples of step functions are also step functions. Thus the set ~(Ia. b). E) of step functions is a vector subspace of B(la. b). E). the Banach space of all bounded functions (see Example 2.1.12). The integral of a step function j is defined by
It is easily verified that this definition is independent of the partition. Also note that.
I ,tjll~tUjU"(b-a)lIjU"" a a
whereuiu",,= sup V(/)I a.;,.;b
t: ~([Q.b].E)-+E
that is.
.
a
is continuous and linear. By the linear extension the(lrem. it extends to a continuous linear map
t
E
L(c1(~([a. b). E». E).
"
2.2.7 DetlnlUon.
f: is called (he Ctllldt)'-1IocItMr bt,. . . .
NoJe that The usual properties of the integral such as
f"bj = f"u j + Icbj.
and
I j 1"j b = -
a
h
are easily verified since they clearly hold for step functions. The space cl(~([a. b). E) contain~ eflou," interesting functions for our purposes. namely
CO([a. b). E) C cl(~([a. b). E»
C
B([a. b). E).
68
BANACH SPACES AND DIFFERENTIAL CALCULUS
The first inclusion is proved in the following way. Since [a, hI is compact, each / E CO([a, b), E) is uniformly continuous. For E> O. let 8 > 0 he given by uniform continuity of/fou/2. Then take a partition a - to < ... < til - b such that Iti til < 8 and define a step function g by gil""" ,I"" /(ti)' Then the ,-disk D,U) in 8«(a. bl, E) contains g. Finally, note that if E and F are Banach spaces, A E L( E. F) and / E c1(~([a, hI, E», we have A 0/ E c1(f,([a. bl. since IIA 0/.. - A 0III ~ IIAIIII/" - III"" for /" step functions in E. Moreover ItA 0/= A( It!> since this relation is obtained as the limit of the same (easily verified) relation for step functions. The reader versed in Riemann integration should notice that this integral for E = R is less general than the Riemann integral: i.e .• the Riemann integral exists also for functions outside of c1(~([a. hl.R». For purposes of this book. however. this integral will suffice. Next we tum to multilinear mappings. If E ...... E" and F are linear spaces. a map
+. -
F»
A : E. x ...
X
Ek
--+
F
is called k-multilinear if A(e ...... e,,) is linear in each argument separately. Linearity in the first argument means that
We shall study multilinear mappings in detail in our study of tensors. They also come up in the study of differentiation. and we shall require a few facts about them for that purpose. ~efinition. The space 0/ continuous k-multilinear maps 0/ E ••... ,Ek to F ;s denoted L(E., ... ,E,,; F). 1/ Ei = E. I ~ i ~ k, this space is denoted Lk(E, F).
2.2.8
As in 2.2.1, a k-multilinear map A is continuous if and only if there is an M > 0 such that
for all e,
E Ei ,l ~; ~ k.
We set
which makes L( E., ... ,E,,: F) into a normed space that is complete if F is.
LINEAR AND MUL TlLlNEAR MAPPINGS
59
Again IIAII can also be defineO as
IIAII- inf{ M > OIIiA ('1.··· .en)1I c; Mlledl' . 'lIenll} =
sup{ IIA(el •...• en)lIllIe"l
Eo;;
1•.... lIe,,1I
Eo;;
I}
2.2.9 Proposition. There is a (natural) norm-preserl'ing isomorphism
Proof.
For A
E
L(EI • L(E2..... E It ; F» we define
AE
L(EI, ...• E k; 'F) by
A-( e" .. .,elt) = A (e l )( e~ •...• e lt )
The association A ..... A is clearly linear and IIAII = IIAII. • In a similar way. we can identify L(R. F)( or L(C. F) if F is complex) with F: to A E L(R. F) we associate A(I) E F; again IIAII = IIA(I)II. As a special case of 2.2.9 note that L(E. E*)=: L 2 (E,R) (or L2(E.C), if E is complex). This isomorphism will be useful when we consider second derivatives. We shall need a few facts about the permutation group on k elements. The information we cite is obtainable from virtually any elementary algebra book. The permutation' group on k elements, denoted Sk' consists of, all bijections a: {I, ... ,k}-+{I, ... ,k} together with the structure of, a group under composition. Clearly, Sit has order kL Letting (R, X) denote R\{O} with tlte'multiplicative group structure, we have a homomorphism sign: Sit -+ (R, X). That is. for o. T E Sit' sign( a 0 T) = (sign a )(sign T). The image of sign is the subgroup {-I. I}; while its kernel consists of the subgroup of even permutations. Thus. a permutation a is even when sign 0 = + I and is odd when sign 0 - - 1. A transposition is a permutation that swaps two elements of {I •... ,k}. leaving the remainder fixed. An even (odd) permutation can be written as the product of an even (odd) number of transpositions. The group Sk acts on Lk(E, F); i.e.• each 11 E .\'A defines a map a: Lk(E.F)-+Lk(E,F) by (oA)(#.'I' ... '#.'k)=A(""") ..... e,,(k). Note that (To)A = :"(oA) for all T.O E Sit. Accordingly, A E L A(I:. F) is called ~ym metric (antisymmetric) if for any permutation a E Sit. oA = A (respectively aA = (sign a')A.)
60
BANACH SPACES AND DIFFERENTIAL CALCULUS
2.2.10 Definition. Let E and F be normed vector spaces. Let L:(E. F) and L!(E. F) denote the subspaces of symmetric and anlisymmetric elements of LIr(E.F). Wrile SO(E.F)-Lo(E.F)-F and SIr.(E.F)={p: E-+ Flp(e)- A(e•...• e). A e LIr.(E. F» We call SIr(E. F) the '/'Me 0/ "omogeMOIlS poIyrwm_ of MIne k from E to F. Note that L:(E. F) and L!( E. F) are closed in L Ir( E. F); thus if F is a Banach space. so are L~( E. F). and L:( E. F). The antisymmetric maps L!(E. F) will be studied in detail in Chapter 7. For technical purposes later in this chapter we will need a few facts about SAC E. F) which are given in the following box. BOX 2.2B HOMOGENEOUS POLYNOMIALS
2.2.11
ProposIUon. (i) SIr(E. F) is a normed vector space with
respect to the following norm:
IIfll = inf{ M > 0 IlIf( e )11 :E; Mllell lr } = sup{ II/( e )11 III ell :E; I} = sup{ II/(e)lIllIell = I}. It is--eo7nplete il E is. (ii) II leS"(E.F) and geS'(F.G). then go/eSIrI(E.G) and IIg 0 III :E; II gil 11/11· (iii) (Polarization). The mapping A: L"(E.F)-+S"(E.F) defined by iCe) = A(e•...• e) restricted to L:(E. F) has an inverse': SIr(E. F) -+ L~(E. F) given by I
air
I
j(el.···.e,,)- k! atl ... at. I./(tle l + .,. +t"elr.) (note that l(tle, + ... + tlrelr.) is polynomial in t l •. .•• 1. so there is no problem in understanding what the derivatives on the right-hand side mean). (iv) For A e L"(E. F). lIill ~ IIAII:E;; (klr/k!~lill. which implies that the maps •• ' are continuous. Proof. (i) and (ii) are proved exactly as for L(E. F) = SI(E. F). (iii) For A e L~(E. F) we have
i(tle l + =
... + tire,,)
E
a,+ ... +aJr.- k 0.. .. );..
a! .k.!. 0 ' I
~
tf"" t~.t4( el ... ·.e~ •...• e¥ ....~ '--y-01
0.
~
LINEAR AND MUL TILINEAR MAPPINGS
and
It follows that
This means that (Af ... A for any A E L~(E. F). Conversely. if f ES"(E. F). then
(iv) IIA(e)1I = IIA(e..... e)1I ~ IIAllllell". so IIAII ~ the other inequality. note that if A E L:(E. F). then
where the sum is taken over all the 2" possibilitieSE, = Put lIelll- ... -lIe"II-1 and get
IIAII. To prove
± 1... .. e" = ± I.
~ II All (je.llle.ll + ... + le.lllekll)"
=IIAllk 4 • whence
.
61
"
BANACH SPACES AND DIFFERENTIAL CALCULUS
Let E - R". F.., R. and
Ie SA(R".R). set
If, = tllrl
'1 ..... '" be the standard basis in Rn. For
+ ... + t"Ir". the proof of (iii) shows that
I(')-/(,.·.·.Ir)=
L
tI, + ... +u,,-k
c.. , ......tf'···t:·;
i.e.• I is a homogeneous polynomial of degree k in t I' ...• t" in the usual algebraic sense. The constant kit/ k! in (iv) is the best possible. as the following example shows. Write elements of RIt as x'" (Xl, .. .• x A ) and introduce the norm IIKxl •...• XA~II-lxll + ... + IxAI. Define A e L~(RA.R) hy 1 A(xl.···.xlt)== k!
LX:,"
.xt·
where xl-(x: •... ,xt)eR" and the sum is taken over all permutations of {I •.. . ,k}. It is easily verified that IIAII == I/k!. and IIAII = I/k"; i.e.• IIAII- (k"/k!)IIAII. Thus, except for k -1. the isomorphism • is not norm preserving. (This is a source of annoyance in the theory of formal power series and infinite-dimensional holomorphic mappings.)
BOX 2.2C.
THE THREE PILLARS OF LINEAR ANALYSIS
The three fundamental theorems of linear analysis are the HDhn- Banach theo"m. the open mapping theorem and the unilorm boundedness principle. Some corollaries that will be used later on will be derived. In finite dimensions these corollaries are all ··obvious." The reader especially interested in infinite dimensions can find the proofs in Appendix B. We begin with the first pillar. 2.2.12 Hehn-8enech Theorem. II FeE is a subspace 01 the nomred oector spau E and I e there exists Ie E· such that II F == I artdll/ll-II/II.
r.
LINEAR AND MUL TlLINEAR MAPPINGS
63
2.2.13 Corollary. Let E be a normed vector space and 1''' O. Then there exists / E E* such that /(1')" O. In other words if /(1') = 0 for all / E E*, then I' = 0; i.e. E* separates points of E. Proof.
•
Apply 2.2.12 to the linear Cunction {aela E C} -+ C. a~ ...... a .
11Je second pillar is as follows.
2.2.14 Open Mapping Theorem. A
E
If E, F are Banach spaces and L(E, F) is surjective, then A is an open mapping. An immediate consequence is the Collowing,
2.2.15 Banach'alaomorphlam Theorem. A continuous linear isomorphism of Banach spaces is a homeomorphism.
Thus, if F, G are closed subspaces of the Banach space E and E is the algebraic direct sum of F and G, then the mapping (e. e') E F x G ..... I' + 1" e E is a continuous isomorphism, and hence a homeomorphism; i.e., E = FeG; this proves the comment at the beginning of Box 2.18.
2.2.18 CloHd Graph Theorem. Let E, F be Banach spaces and A: E -+ F a linear map. Then A is continuous iff its graph E E
r. . = {(I', AI')
x Ell' E E} is a closed subspace of EeF.
PI'fHJf. It is readily veriCied that r. . is a linear subspace of EeF. If A E L(E. F). then r. . is closed (see Exercise I.4B). Conversely, if r. . is closed. then it is a Banach subspace of EeF and since the mapping (e. Ae) E I' E E is a continuous isomorphism, its ~nverse I' E E ..... (I', AI') E is also continuous by 2.2.15. Since (e. AI') E AI' E F is clearly continuous. so is the composition e ...... l 1', Ae) ...... AI'. •
r.........
r. .
r........
The closed graph theorem is often used in the following way. To show that a linear map A: E -+ F is continuous Cor E. F Banach spaces, it suffices to show that if en -+ 0 and Ae" -+ e', then 1" = O.
2.2.17 Corollary. Let E be a Banach space and F a closed .fubspace of E. Then F is split iff there exists P E L(E. E) sU('h that pop = P and F= {I' E EIPe - e}.
64
BANACH SPACES AND DIFFERENTIAL CALCULUS
Proof. If such a P exists. then clearly ker P is a closed subspace of E that is an algebraic complement of F: any ~ E E is of the form ~ - ~ - P~ + P~ with ~ - P~ E ker P and p~ E F. Conversely. if E.,. F$G. define p:Eq....... E by p(~) == ~., where ~ .. ~. + e2'~1 E F'~2 E G. P is clearly linear, p2 = P and F= {~E J:IP~ == ~}. so all there is to show is that P is continuous. Let e.... e' lf + e 21f ....... 0 and P(~") = ~'If ~'; i.e., - ~2" ....... ~' and since F and G are closed this implies that e' E F n G = {O}. By the closed graph theorem, P E L(E, E). •
....
2.2.18 Fundamental laomorphlsm Theorem. Let A
E L(E, F) be surjective where E and Fare Banach spaces. Then E/ker A and Fare isomorphic Banach spaces.
Proof.
The map [e) ..... A ~ is bijective and continuous (its norm is •
< IIAII), so it is a homeomorphism. A sequence of maps A,
AI,.
I
.......... E;_ • ....... E; ....... Ei+ I
..........
or Banach spaces is said to be split exact if for all i, ker A;+ I = range AI and both ker A; and range A; split. With this terminology. 2.2.18 can be reformulated in the following way: If 0 ....... G ....... E ....... F ....... 0 is a split" exact sequence of Banach spaces. then E/G is Banach space isomorphic to F(thus E a GeF). Now we are ready for the third pillar.
2.2.11 Uniform Bounded.,... Principle. Let E be a Banach space. Fa normed o«Ior space (over R or C) and (A;};EI c L(E, F). If for each e E E the set {Ale);.1 c F is bounded. then the set {IIAIII}IE/;s a bounded set of ( positive) numbers.
2.2.20 Corollary. If (A"}C L(E. F) ;s a strongly convergent sequence; i.e.• lim" _ OOA"~ = Ae exists for every ~ E E, then A E L(E, F). A is clearly a linear map. Since {A"e} is convergent, it is a bounded set for each e E E. so that by 2.2.20 (IIA"I!) i~ hounded hy, i ~ O. But then ./q
Proof. );"
<
IIAell- lim IIA".',:
,,--0
i.e. A
E L(E. F).
•
<;, (
lim SUPIIA"II)lIell< Mllell;
n-oo
LINEAR AND MUL TILINEAR MAPPINGS
65
Exercl... 2.2A
2.2B
If E "" R" and F = R m with the standard norms. and A: E ..... F is a linear map, show that (i) IIAII is the largest eigenvalue of AA T, where' AT is the transpose of A and (ii) this norm does not come from an inner product (use Exercise 2.1 A). Let E = R" and F = R" with the standard norms and A. BEL (E, F). Let (A. B)
2.2C
=
trace(AB T )
Show that this is an inner product on L(E, F). Show that the map L(E, F) x L( F. E) ..... R,(A, B) ...... trace(AB)
gives a (natural) isomorphism L(E. F)·
=L(F, E) ..
2.20 (L. HOrmander). Let E, F. G he Banach spaces, A E L( E, F) and BE L(E, G). Show there exist constants MI' M] > 0 such that IIA41I1 ~ MI
IIIkIl ~ M 2 (I\A41Il+l\el\)
for all eE E. (Hint: Norm EeG by lI(e,g)II-l\ell+lIglI and define T: r B ..... G by T(e, Be) = Ae. Use the closed graph theorem to show that T E L(rB , G).) 2.2E (Linear transversality). Let E, F be Banach spaces, Fo c F a closed subspace, and T E L(E, F). T is said to be transversal to Fo' if T-1(Fo) splits in E and T(E)+ Fo "" {Te + lie E E.I E F} -.F. Prove the following . (i) T is transversal to Fo iff 'IT 0 T E L(E. F/Fo) is surjective with split kernel; here 'IT: F ..... F/Fo is the projection. . (ii) If 'If 0 TEL (E, F/ Fo) is surjective and #;) has finite codimension, then ker( 'If 0 T) has the same codimension and T is transversal to I,. \ 11,,1/: Use the algebraic isomorphism T( E )/( Fo n T( E» "'" (T( E)+ Fo)/Fo to show E/ker('If 0 T) ... F/Fo; now use 2.2.18.) . (iii) If 'If 0 T E L(E, F/Fo) is surjective and if kerTand Fo are finite dimensional, then ker('If 0 T) is finite dimensional and T is
66
BANACH SPACES AND DIFFERENTIAL CALCULUS
transversal to 1'0' (Hinl: Use the exact sequence 0 -> ker T-> ker(tr 0 T) .... 1'0 (") T(E) .... 0.) 2.2F
Let E and I' be Banach spaces. Prove the following. (i) If IE c1(Si([a. hI. L(E. 1'))) and ~ E E. then
(I(t)edt= ((I(t}dt)(e).
(ii)
(Hint: T >-+ Te is in 1.( L( E. F).
n.)
If IE c1(:;([a. hl.H) and vE
then
(/(I)vdl
F. =
((/(1) dt )v.
(Hinl: I ..... multiplication by I in Fis in L(R. L(F. 1'»; apply (i).) (iii) Let X be a topological space and I: [a. b)X X .... E be continuous. Then the mapping g: X -> E. g(x) = 1:1(1. x) dl is continuous. (Hinl: For IE R. Xo E X and E> 0 given. II/(s. x)-, 1(1. xo)1I < E if (s. x) E U, X Ux •• ,; use compactness of [a. hI to
find Uxo as a finite intersection and such that 1I/(I.x)1(1. xo)1I < E for all t E [a. b). x E Ux ".) 2.2G Show that the Banach isomorphism theorem is false for normed incomplete vector spaces in the following way. Let E be the space of all polynomials over H normed by lIa o + alx + ... + anx"lImax(laol ... ·.Ianl}· (i) Show that E is not complete. (ii) Define A: E-+EbyA
a ( Lna;x') =a,,+ Ln ...J.. x' ;-0
(iii) 2.2H
2.21
and show that A E L( E. E). Prove that A Show that A - I is not continuous.
,-I
I:
I
E -> E exists.
Let E and I' be Banach spaces and A E L(E. F). If A(E) has finite codimension. show that it is closed. (Hinl: If 1'0 is an algehraic complement to A (E) in F. show there exists a continuous linear isomorphism ElkerA "" FIFo; compose its inverse with Elker A .... A(E).) (Symmetrization operator). Define Sym": L k(E. F) .... L k(E. F). by . k
Sym A
=
I kI
~
~ aA •
. .. es,
THE DERIVA T1VE
67
where (aA)(e., ... ,eli:)'" A(e,,(I)' .,e,,(k»' Show that (i) Syrrr'«Lk(E,F»=L~(E,F). (ii) (Syrrr'<)2 = Sym". (iii) IISyrrr'<1I ~ \. (iv) If F is Banach. then L~(E. F) splits in L"(E. F). (Hint: Use 2.2.17.) . (v) (Sym"A)' = A 00
2.21
Show that a k-multilinear map continuous in eac~ argument separately is continuous. (Hint for k = 2: If lit'. II ~ I. then IIA(e •• e2)11 ~ IIA(·. e 2)1I. which by the uniform bounded ness principle implies th~t IIA(e •• ·)1I ~ M for lie. II ~ \.)
2.3 THE DERIVATIVE For a differentiable function I: VcR - R. the usual interpretation of the derivative at "0 E V is the slope of the line tangent to the graph of I at "0' To generalize this. we interpret DI( uo) = f'( III) as a linear map ac'tins on the vector (" - "0)' Then we can say that D/( un) is' the unique linear m~p from R into R such that the mapping
"0
(see Fig. 2.3.1). This version can he extended to vector is tangent to I at valued mappings of a vector variable as follows.
R
,V
R 110
figure 2.3.1
68
BANACH SPACES AND DIFFERENTIAL CALCULUS
2.3.1
DefinlUon.
Let E, F be nornted vector spaces, with maps I,g: UcE-+F
where U ;s open in E. We say I and gore tQIIgent at Uo E U il
lim 11/(11)- g(II)1I = 0, ,,- "0
1111 - 11011
where 11'11 represents the norm on the appropriate space. Pro~ltIon. For I: U c E -+ Fond "0 E U there is at most one L E L(E, F) such that the map gL: U c E -+ F given by gL(II) = 1(11 0 )+ L(II- " 0 ) is tangent to I at " 0 ,
2.3.2
Proof. Let LI and L2 E L(E. F) satisfy the conditions of the proposition. If e E E, lIell == I and II == "0 + 'Ae for 'A E R (or C), then for 'A small and II E U, we have
IIL,e - L 2ell
""
=
IIL 1(1I- u o )- L 2 (1I- uo)lI/lIu - uoll
11/(")- I(u o )- L,(u - uo)1I
II" - uoll
+
II/(u)':'" I(uo)- L 2(u - "0)11
lIu - uoll
.
As 'A -+ 0, the right-hand side approaches zero so that IKL, - L2 )ell = 0 for all e E E with lIell-I; i.e.IIL , - L211 == 0 and thus LI =- L 2 • •
2.3.3 Definition. II, in 2.3.2. there is such an L E L(E. F) we say I is dillemrtiabk at "0' and define the derivative 0/ I at Uo to be D/("o) = L. II I is differentiable at each "0 E U, the map D/: U-+L(E.F);" ...... DI(II) is the derivative 01 f. Moreover, il DI is a continuous map (where L (E, F) has the norm topology), we say I;s 0/ c/ass C· (or is contin"ously dif/erentiabk). Proceeding inductively we deline D'I-D{D'-'/}: UcE-+L'(E.F) il it exists. where we have identilied L(E, L,-I(E. F» with L'(E, F) (see 2.2.9). II D'I exists and is norm continuous we say I is 01 c/ass C'.
We shall reformulate the definition of the dt, '!v(' with the aid of the conwnient LIIIIIIIuI symbol: o(e A ) will dellote lm\' COII/IIIlIOUS map ill a
THE DERIVATIVE
69
neighborhood 01 the orlgm 01 a normed vector space E. satisfying lim ......( o( e k }/llell k) - O. The collection of these functions forms a vector space. Clearly I: U c E ... F is differentiable at Un E U iff there exists a linear map DI(uo} E L( E. F} such that
This formulation of 2.3.3 can save a lot of writing in technical proofs. We shall next apply it to derive some of the most elementary properties of the derivative. First of all. note that if DI(u o } exists. then 1 is continuous at"o since
Secondly we prove the sum and constant multiple rules for the derivative. 2.3.4 'Proposition (Linearity 0/ t"e Derivative). Let I. g: u. c E ..... F be r times differentiable mappings and a a real (or complex) consttmt.' Then al and 1 + g: U c E ..... Fare r times differentiable' with D'(I + g) = D'I + D'g and D'(a/) = aD'/. Proof.
If" E U and e E E. then
1( U + e) ... 1(U ) + Df ( II ) . e + () ( e ) g(u + e) = g(,,)+ Dg(II)·e + lI(e).
and
so that adding these two relations yields
(f + g)(" + e).,. (I + g)(,,)+ (DI(")+ Dg(u »·e + o(e). The case r > I follows by induction. Similarly a/(u + e) = a/(II}+ aDI(u}'e' + ao(e)'" al(u)+ aDI(u)'e + o(e}. • 2.3.5 Proposition. Let /;: U c E ..... F;.1
u»
D'I = D'/I
X ••• X
Drl,,·
70
BANACH SPACES AND DIFFERENTIAL CALCULUS
Proof.
For _ e U• •
e £ we have .
+.»
1(- +.) - (f,(- +. )... ·.1,,(" - (f,(.)+ 0/,(.)'. + 0(. ) .....1,,(.)+ 01"(,, )". + 0('» - (/,(. ) •... •1..
(-» + (0/,(") •...• 01.. (" » .• + (0(.) •...• 0(.»
-/(")+ 0/(,,),. + 0(.). the last equality follows using the sum norm in F.
x ... x F,,:
11(0(.) ..... 0(.»11-110(.)11+ ... +110(')11. so (0(.) ..... 0(.» = 0(').
•
Notice from the definition that for LEL(E.F).DL(,,)=I. for any
• e E. It is also clear that the derivative of a constant map is zero. Usually all our spaces will be real and linearity will mean real-linearity. In the complex case. differentiable ma~pings are the subject of analytic function theory. a subject we shall not pursue in this book (See Exercise 2.3F for a hint of why there is a relationship with analytic function theory.) In addition to the foregoing approach. there is a more traditional way to differentiate a function I: U C R" - R .... We can write it in component form I(x' ..... x .. )- (f'(x' ..... x").....I"'(x' •...• x"» and compute partial derivatives. ap/ax l for j=I ..... m and i-I ..... n. where the symbol ap/ax; means that we compute the usual derivative of p with respect to Xl while keeping the other variables x' ..... XI-'. xl+' ..... x" fixed. For I: R -R. O/(x) is just the linear map "multiplication by dl/dx." i.e.• dl/dx ... O/(x),1, This fact. which is obvious from the definitions. can be generalized to the following theorem.
2.3.1 Propoeltion. Suppose U C R" is an open set and I: U - R'" is di/l"entiable. Then the partial derivalives ap/ax; exist. and the matrix 01 the linear map Ol(x) wilh respect 10 the slandard bases in R" and R m is given by
!£ ax'
~ ax 2
~) ax"
!£ !£ ax'
ax 2
a/ 2 ax"
al'" ax'
aIm ax 2
al no ax"
THE DERIVA TlVE
'71
where each partial derivative is evaluated at x == (x I •...• X"). This matrix is called the Jacobillll matrix 01 f. Proof. By the usual definit~on of the matrix of a linear mapping from linear algebra. the jith matrix element 0 1; of DI(x) is given by the jth component of the vector DI(x)e;. where e l... .• e" is the standard basis of A". Lettingy -= x + he;. we see that
II/( y)- I(x)- DI(x)( y - x)1I lIy-xll
h =
Xll/(xl ..... x; + h ..... x")- l(xl ......l"")_vDI( x)e,1I
approaches zero as h -+ O. so the jth component of the numerator does as well; i.e., '
1-
1 I/J{ x I ..... x ; + h , .... x ") - li{ x I ..... x " ) - ha,j - 0 • Ii m -h
h .... O
which means that ail = ap/ax;. • In doing practical computations one can usually compute the Jacobian matrix easily. and this proposition then gives Df. In some books, D/is called' the differential or the total derivative of f. 2.3.7 Example. Let I: A2 -+ Al, I(x, y) = (x 2, xly, X 4y 2). Then D/(x, y) is the linear map whose matrix in the standard basis is
ax
all
all ay
al 2 ax
ap ay
all ax
all ay
=
( 2.2
3x y 4xly2
o 1
Xl
2x4y
where/l(x, y) ... x 2,f2(x. y) = xly,fl(X, y) = X·y2. • One should take special note when m -1, in which case we have a real-valued function of n variables. Then DI has the matrix
... YL) (YL axl ax"
72
BANACH SPACES AND DIFFERENTIAL CALCULUS
and the derivative applied to a vector t' = (a l ••••• a n ) is Df(x)·e=
af i • L" -.a i-I
ax'
It should be emphasized that DI is a linear mapping at each x E U and the definition of D/(x) is independent of the basis used. If we change the basis from the standard basis to another one. the matrix elements will of course change. If one examines the definition of the matrix of a linear transformation. it can be seen that the columns of the matrix relative to the new basis win be the derivative DI(x) applied to the new basis in R n with this image vector expressed in the new basis in R"'. or course. the linear map DI(x) itself does not change from basis to basis. In the case m = I. Df( x) is. in the standard basis. a I x n matrix. The vector whose components are the same as those of Df(x) is called the gradient of f. and is denoted grad for vj. Thus for f: U eRn -+ R.
gradf =
( : : . •...•
:!n ).
(Sometimes it is said that grad f is just Df with commas inserted!). The formation of gradients makes sense in a general inner product space as follows. 2.3.8 Definition. (i) ut E be a normed space and I: U c E --+ R be differentiable so thaI D/(II) E L( E.R) = E*. In this case we sometimes write d/(lI) for Df(lI) and call df the differential off. Thus d/: U -+ E*.
(ii) II E is a Hilbert space. the Ilnulia, 01 I is
(Jej"r~'J
by
grad/-v/: U-+E (v/(II).e) - dl(II)·e. where d/(II)'e means the linear map d/(lI) applied to the vector e.
Note that the existence of v/(lI) requires the Riel%. re.Y1II5entation theorem (se. ~,2·5) The notation 81/"" instead of (grad J )(U) II) is also in widt' II,"~ ("l)CCially in the case in which E is a space of functions.
=,-/(
(.»
2.3.8 Example. Let (E. be a real inner product space and / ( II) = 11_11 2, Since 1111112=11110112+2(110.11-110)+1111-110112. we obtain df(lIo)·e
THE DERIVATIVE
73
== 2(uo'~> and thus vf(u) = 2u. Hence f is of class C l . But since Of(u) = 2(u, . >E E· is a continuous linear map in u E E, it follows that 02f(u) ... Of E L(E, E·) and thus OAf = 0 for k ;;d. Thus f is of class CfX>. The mapping f considered here is a special case of a polynomial mapping (see 2.2.10) .. We close this section with the Fundamental Theorem or Calculus in real Banach spaces. First a bit or notation. If cp: U c R -- F is differentia· ble, then Dcp(t)E L(R, F). Now L(R, F) is isomorphic to F by A ..... A(I). IE R; note thatIlAII"'IIA(I~I. We denote cp'(/)'"
~~ =
Ocp(/)·I,
IE R.
Clearly
'( ) . I. cp(t + h)-cp(t) cp t = 1m h 11-0
and cp is differentiable iff cp' exists.
2.3.10 Fundamental Theorem of Calculua. (i) If g: [a, b] ~ F;s continuous, where F is a real Banach space, the map f: ]a. b[ -- F;f(t),. f~g(s)ds is differe!'tiable and we have f' = g. (0) Iff: [a, b] ..... F is continuous, is differentiable on ]a, b[ and/, extends to a continuous map on (a, b], then f(b)....,f(a)
= tf'(S)df. a
Proof.
(i) Let to E ]0, b[. Since the integral is linear and
continuous~
IIf(t o + h)- f(to)-hg(to)II .... IIJ,~o+II(g(s)- g(/o» dr~
4iOlhi sup{ IIg(s)-g(t o )lllto ..;s..;to +h} .. The sup on the right-hand side -- 0 as Ihl-- 0 by continuity of g at ' 0 • (ii) Let h(t)- (ff'(S)ds)- f(t).
By (i) h'(t) == 0 on la, b[ and h is continuous on [a, b]. If for some.t E [a, b] h( t) =I> h( a), then by the Hahn-Banach theorem there exists tJ E such that (tJ 0 h Xt) =I> ( tJ 0 h Xa). Moreover 11 0 h is differentiable on ]a, .b( and its
r
74
BANACH SPACES AND DIFFERENTIAL CALCULUS
derivative is zero (Exercise 2.3D). Thus by elementary calculus. 11 0 h is constant on [a.b]. a contradiction. Hence h(t)=h(a) for all tE[a.b]. In particular heal - h(b). • Exercl... 2.3A
Let B: E X F -. G be a continuous bilinear map of normed spaces. Show that B is COO and that
DB(" • .,)(e, f)=B(", f)+B(e,.,). 2.3B
Show that the derivative of the map is unaltered if the spaces are renormed with equivalent norms.
2.3C
If IE Sk(E. F), show that
and 2.3D
Let I: VeE -. F be a differentiable (respectively C) map and A E L( F. G). Show that A 0 I: VeE --+ G is differentiahle (respectively C) and D'( A 0/)(") = A 0 D'/( u). (Hint: Use induction.)
2.3E
Let I: VeE -. F be r times differentiable and A E L(G. E). Show that
2.3F
exists for all i ... r, where.,E A -I(V). and IJI •••.• IJ, E G. (i) Show that a complex linear map A E L(C.C) is necessarily of the form A(z) = >.z, for some>. E C. (ii) Show that the matrix of A E L(C.C) when A is regarded as a real linear map in L (R 2. R 2) is of the form
(~ -a h). (Hint: >. = a
(iii)
+ ih.)
Show that a map I: V c C --C.I = g + ih. g. h: V CR2--+R is complex differentiable iff the Cauchy- Riemann equations
ag = ah ! . La ax ay'ay
_ ah
are satisfied. (Hint: Use (ii) and 2.3.6.)
ax
PROPERTIES OF THE DERIVATIVE
75
Let (E.(.» be a complex inner product space. Show that the map . 1(11)= 1111112 is not differentiable. Contrast this with 2.3.9. (Hint: DI(II). if it exists. should equal 2 Re«II,' ».) 2.3H Show that the matrix of D 2/(x)eL 2 (R".R) for I: VeR"-R. is given hy 2.3G
~
~
ax'ax 2
a2n ax'ax"
~
a21 ax"ax2
a21 ax"a.J("
ax'ax'
ax"ax'
(Hint: Apply 2.3.6. Recall from linear algehra that the matrix of a bilinear mapping BeL(R".Rm;R) is given by the entries B(e,. ") (first index = row index. second index = column index). where {e, •...• e,,} and {/, •...•lm} are ordered bases of R" and R m respectively.) 2.4
PROPERTIES OF THE DERIVATIVE
In this section some of the fundamental properties of the derivative are developed. These properties are analogues of rules familiar from elementary calculus. Let us begin by strengthening the fact that differentiability implies continuity. 2.4.1 Propoaltlon. Suppose VeE is open and I: V ~ F;s differentiable on V. Then I is continuous. In lact. lor each "0 e V there i.~ a constant M> 0 and a 80 > 0 such that 1111 - 11011 < 80 implies 11/(11)- I(IIo~1 ~ MIlII - 110 11. (This is called the Lipsmit:. property.)
Proof.
1111(11)- 1(lIo)II-IIDI(uoHu -11 0 )111 =lIo(u-u()II~lIu-uoll
for
~
11/(11)- I(u o )- DI(uo)(u -11 0 )11
IIII-u olI<80 •
where 80 is some positive constant depending on " 0 ; this holds since lim.~.lIo(II-lIo)/liu - 11011 = O. thus
II/( u)- I( " 0 )11 ~ II D/( uo)( II -
"0 )11+1111 -
~ (IIDI(lIo)lI+ 1)1111- uoll
for 1111 - 11011 <
80. •
11011
76
BANACH SPACES AND DIFFERENTIAL CALCULUS
Perhaps the most important rule of differential calculus is the chain rule. To facilitate its statement, the notion of the tangent of a map is introduced. The text will begin conceptually distinguishing POin/.f in U from vectors in E. We aenerall), use hold face for vectors like, E E and lightface for paints II • (I, At Ihlll l'III"1 Ihlll ill jUlit nututiun. hut it will help with the transition to manirolds in Chapter 3.
2.4.2 OeflnlUon. Suppose f: U c E .... F is of class C I . Define the t.gmt of f to be the map Tf: UxE .... FxF given by Tf(u.e) = (f(u), Df(u)·e) where Df(u)·e denotes Df(u) applied to e E E as a linear map. Iff is of class C', define T'f ... T(T'- If) inductively. From a geometric point of view, T is more natural than D. If we think of (u,e) as a vector with base point u, then (f(u), Df(u)·e) is the image vector with its base point I(u). See Fig. 2.4.1. Another reason for this is its simple behavior under composition, as given in the next theorem.
l
E
u
I
~
F
\(U) DI(u)·.
Figure 2.4.1
2.4.3 Theorem. (C' Composite MIIPJIDtg lMoIWm). Suppose f: U c E .... V c F and g: V c F .... G are differentiable (respectively C') maps. Then the composite g f: U c E .... 'is also differentiable (respectively C') and 0
rr
PROPERTIES OF THE DERIVATIVE
77
(nspectively, T'(g 0 I) - T'g 0 T'I).
The lormula T(g 0 f) - Tg 0 TI is equivalent to the chain rwle in tmns
01 D:
D(g 0 I}(u)·e - Dg(f(u»· (DI(u»·e) (For a formula for D'(g 0 f) if r
~
2, see Box 2.4.A.)
Since I is differentiable at u E U and g is differentiable at I( u) E V,
Proof.
we have
l(u+e)=/(u)+D/(u)e+o(e)
for eEE
and for v = I(u) we have g( v + w) = g( v)+ Dg( v )·w + o( w). Thus
(g 0 I)(u + e) == g(f(u)+ D/(u)·e + o(e» - (g 0 f)(u)+ Dg(f(u»· (DI(u),. + o(e»+o(DI(u)·e + o(e» ==
'.
(g 0 f)(u)+ Dg(f(u»·(DI(u)·e)+ Dg(f(u»(o(e» +o(D/(u)·e+o(e».
For e in a neighborhood of the origin,
IIDI(u)·e + o(e)1I
lIell
EO
(1IDI( )11 'J
o(e») M u + lIell EO
for some constant M> O. and
IIDg(f(u»·o(e}1I EO IIDg(f(u»lIlIo(e)lI. Thus'
lIo(D/( u )·e + o(e »11 =
lIe o( DI( u )·e + o(e »11 . IIDI(u)·e + o(e )11 IID/(u)·e + o(e)1I
II ell
. Mllo(DI(u).e + o(e»11 IID/(u)·e + o(e)1I •
EO
Hence we conclude that
Dg(f(u)(o(e»+ o(D/(u)·e + o(e»'" o(e)
lIell
78
BANACH SPACES AND DIFFERENTIAL CALCULUS
and hence D(g 0 f)(u)·e = Dg(f(u»· (DI( u ).t!).
Denote by..,: L(F.G)XL(E.F)-+L(E.G) the bilinear mapping ..,(B. A) .. BoA and note that.., E L(L(F. G). L(E. F); L(E. G» since liB 0 All" IIBIlIlAII; i.e.• II..,n" I. Let (Dg 0 j)x DI: U -+ L(F. G)X L(E. F) be defined by «Dgoj)XDI)(u)=(Dg(f(u»,DI(u»; note that this map is continuous if I and g are C I . Thus the composite.., 0 Dg 0 j) x Dj) = D(gof): U-+L(E.G) is continuous if I and g are CI. i.e.• gol is C I. Inductively suppose I and g are C'. Then g 0 I is C' I and Dg 0 I is C' -- I. so that (Dg 0 f)x DI is C- I (see 2.3.5). Since.., is C'" (Exercise 2.3A). again the inductive hypothesis forces f{J 0 Dg 0 / ) x D/) = D( go/) to he C' ,; i.e.• g 0 I is C. . The formula T'( go I) = T'g 0 T'I is a direct verification for r = I using the chain rule. and the rest follows by induction. •
«
«
Ji:
If E ... R'". F- R". G ... R' and/- (/ 1••••• /"). g = (gl •...• g,). where U -+ R. gl: V -+ R. by 2.3.6 the chain rule becomes
a(gOf)I(X) ... a(gof)l(x) axl
ax'"
a(gof)'(x)
a(g 0 f)'(x) ax'"
ax' agl(f(x»
iJi
iJgl(f(x» ay"
iJ/I(x) ... iJI'(x) iJx" iJx l
iJgP(f(x» iJyl
iJg'(f(x» iJy"
af"(x) axl
which. when read componentwise. becomes the usual chain rule from calculus:
i1(gof)I(X) ... axi
r.
~ _,
iJg 1(f(x» iJI"(x) iJ yk ax;'
The chain rule applied to B E L( Fl' F2 ; G) and II F2 yields the following.
i= I •...• p. X 12 :
Uc E
-+ FI X
2.4.4 Corollary. (1M ui/mi'{. R_). Let /;: U c E -+ F;. i ... 1.2. be dillerentiable (respectively C) maps and BE L(FI • F2; G). Then the mapping'
PROPERTIES OF THE DERIVA TlVE
BU" f2)" B °U, x f2): U c E
-+
G, is differentiable
(respective~v
79
C) and
In the case F, = F2 = Rand B is multiplication, 2.4.4 reduces to the usual . product rule for derivatives. Leibniz' rule can easily be extended to multilinear mappings (Exercise 2.4C). The first of further consequences of the chain rule involves the directional'derivative.
2.4.5 Definition. Let f: U c E -+ F and lei u E U. We derivative in the direction e E E at U If
sa)'
that f hat' a
exists. We call this element of F the directional derivatit)e.
Sometimes a function all of whose directional derivatives exist is called Gdteaux differentiable. whereas a function differcntiahle in the sense we have defined is called Frechet differentiable. The laller is stronger. according to the following. (See also Exercise 2.4{)
2.4.8 Proposition. If f is differentiable at tives off exist at u and are given by d
II,
then the directional deriva-
.
d/(u + te)I,_o = Df(u)·e. Proof. A path in E is a map from I into E. where I i~ an open interval of R. Thus, if cis difl.:rentiable, for tEl we have Ik(t)E/.(R.E). by definition: Recall that we identify L(R. E) with E hy associating Dc(t) with Dc(t)·1 (I E R). Let dc dt(t)=Dc(t).1.
For f: U c E -+ F of class C' we consider f from the chain rule that
0
c, where c: 1-+ U. It follows
d de dt(f(c(t»)=D(foc)(t).I ... Df(e(/»' dt'
The proposition follows by choosing e(/)-'II+le(lI.eEE,I-(-A,A). and A sufficiently small). •
80
BANACH SPACES AND DIFFERENTIAL CALCULUS
For I: U c R" .... R. the directional derivative is given in terms of the standard basis " ..... ' .. by
X'"
where, + ... + x"'... This follows from 2.3.6 and 2.4.6. The formula in 2.4.6 is sometimes a convenient method for computing DI(u)·,. For example. let us compute the differential of a 'homogeneous polynomial of degree 2 from E to F. Let I(~) = A(" ,). where A E L2(E; F). By the chain and Leibniz rules.
d
DI(u)·~ = dtA(u + t'. u + t~ )1,-0; =
A(u.,)+ A(" u).
DI(u)'~ = 2A(u.e). One of the basic tools for finding estimates is the following.
If A is symmetric. then
I
2.4.7 PropoeHlon. Let E and F he real Banach spaces. I: U c E
Fa C'-map, x. y E U. and c a C I arc in U conneclinJ{ x 10 y; i.e .• c is a continuous map c: (0. I) -+ U, which is ('Ion )0,1(. c(O)'" x, and c( I) = y. Then -+
I(y)- I(x) = fo'DI(c(t»'c'(t) dt. II U is convex and c(t) -= (1- t)x + ty, then I(y)- I(x) == fo'D/«I- t)x + ty)·(y - x) dt;
=
Proof. If g( t) ,. (I
0
(10' DI« 1- t)x + ty) dt). (y -
x).
c)( t), the chain rule implies
g'(t) = D/(c(t»·c'(t).
and the fundamental theorem of calculus gives g(l)- J{(O) = fo'8'(t) dt,
which is the first equality. The second equality for U convex and c(t) = + ty is Exercise 2.2F(i). •
(J - t)x
PROPERTIES OF THE DERIVATIVE
81
Coral...,.
2.4.1 (MtNIII Ya/w IMpaIi'Y). SuppoSe U c E ;s convex and f: U c E ...... F ;s C I . Then for all x, y E U IIf(y)- f(x)1I ~ suplIDf(I - t)x + tv)IIII·\, - yll. Thus. if II Df( u ~I is uniformly bounded on U by a x.yEU'
('mWtllll
M > O. the.:n for all '
IIf(y)- f(x)1I ~ MIIY- xII. If F=R, then f(y)- f(x) = Df(c)(y - x) for some c on the line joining x toy.
Proof.
The inequality follows directly from 2.4.7. The last assertion follows from the intermediate value theorem as in elementary calculus. •
2.4.1
Corol"".
Let U c E be an open set; then the following are equiva-
lent (i) (ii)
U is connected. every differentiable map f: U
c E ...... F sati,ffy/nll, Df - 0 on U i" ('{In-
stant.
f(u)=
{O,~fUEUI
e,.f uE U2 ,
where e E F, e .. 0 is a fixed vector. has Df = O. yet is not constant. Con\'l·rsely. assume that U is connected and Df = O. Then f is in fact Coo. Let "oEU be fixed and consider the set S={uEUI/(u)=f(uo)}' Then S .. (i'J (uo E S), S c U and S is closed since f is continuous. We shall show that S is also open. If u E S. consider v E D,( u) c U ~d apply 2.4.8 to get IIf(u)- f(v)1I ~ sup{ IIDf«l- t)u+ tv)IIlt E [O.I]}IIu-vll=O; i.e., f(v) = feu) = f(uo) and 'hence D,(u) c S. Connectedness of U implieS S=U. • If fis GAteaux differentiable and the GAteaux derivative is in L(E, F); i.e., for each u E V there exists Gil E L(E, F) such thaI
d d,f( U + te) = G"e. and if u ..... Gil is continuous, we say
f
is C1-Gateaux. The mean value
82
BANACH SPACES AND DIFFERENTIAL CALCULUS
inequality holds. replacing CI everyWhere by "CI-GAteaux" and the identical proofs work. When studyinl differentiability the folJowinl is often useful.
2.4.10 CoroIl.". II I: U c E
-+
F is CI-Gdleaux then it is
C I and the
two tkrivatives coincide.
Proof.
Let II E U and work in a disk centered at II. Proposition 2.4.7 gives
11/(11 + e)- I(II)-Gy ell
=11(fo'( G
HI
+I•
-
GU> dt
"sup{ IIGII + 1•
-
Gllilit E [O.I]}lIell
y
and the sup converges to zero as. e -- O. by uniform continuity of the map t E (0.1] .... GY+l. E L(E. F). This says that DI( II)·e exists and equals Gu'. • It wilJ be convenient to consider partial derivatives in this context. We shall discuss only functions of two variables. the generalization to n variables being obvious.
2.4.11
DefInition. Let I: U -- F be a mapping defined on the open set U c EleEz and let "0 - (11 01 , IIoz) E U. The derivatives 01 the mappings 0, .... 1(0,. "0Z)' "2 .... /(1101' "z). 0, E E,. 0z E Ez• if they exist. are called,.,.. tkrit:Jtltiva 01 I at "0 E U and are denoted by D.t(uo) E. L(E,. F). Dz/(u o ) E L(Ez • F).
,ill'
2.4.12 Propoaltlon. Let U C EI eEz be open and I: U -- F. (i)
II I is differe"tiable. then the partial derivatives exist a"d are equal
to
D.t(u)·e, = DI(u)'(el'O) Dz/(II)'ez = DI(u)·(O.e z )·
(ii) II I is differe"tiable. the" DI(u)·(el.ez) = D.t(u)·el
(iii) I is 01 class C' iff DJ: class C,-I.
+ Dz/(II)'e2'
U-- L(E;. F). i =
1.2 both exist and are 01
Let j~: E, -+ EleEz be defined by j~( v,) = (° 1, " 2 ), where Thenj~ is C GO and Dj~ (II I )=/,EL(E,.E,eEz ) is given by I, ( e I) ... (e ,. 0). Thus by the chain rule. Proof.
(i)
II-(U I.U Z )'
D.t( u) = D(f 0 j~ )( u l ) = D/( II) 0 II_
PROPERTIES OF THE DERIVA TlVE
83
which proves the first relation in (i). One similarly defines j;, 12 and proves the second relation. (ii) Let P;(_I'_2)-""i-l.2 be the canonical projections. Comp~se the relation 1 1 0 PI + 12 0 P2 == identity on E l eE2 with D/( u) on the left and use (i). .(iii) Let.; E L(L(E l eE2 • F), L(E;, F» and -t, E L(L(E" F), L(E I eE2 • F» be defined by .;( A) = A 0 I; and -t;( B;) = B, 0 p;, ; = 1.2. then (i) and (ii) become
DJ= Dlo." This shows that if I is differentiable, then I is C' iff DI I and D21 are C· I.' Thus to conclude the proof we need to show that if DI I and D21 exist and are continuous. then DI exists. By 2.4.7 applied consecutively to the two arguments we get .
I(u l +e"u2 +1'2)- l(u l ,u 2 )- Dd(U I,u 2 )'e l - Dd(U I,U 2 )'''2 =
I(u l + 1'1' u 2 + "2)- l(u l , u 2 + 1'2)- D,/( UI , U2 )'1'1 + I( UI' U2 + 1'2):- I(UI' U2)- Dd( UI' U2 )'1'2
=
(fol(DI/(U I + lei' U2 + 1'2)- Dd(u l , 11 2 ) dl )'1'1
+
(10 (Dd( u" U2+ 11'2)- Dd(u " U2)) dl) '1'2' 1
Taking norms and using in each term the obvious inequality 111';1140; lIedl + . 111'211' == IKe l , 1'2 ~I, we see that
II/(u l + 1'1' u 2 + 1'2)- l(u l , u 2 )- Dd(u l , u2)'e l - Dd(u l .u 2 )·e211 40;( sup IIDd(ul+lel,u2+e2)-Dd(ul,u2+e2)1I 0 •.,., I
+ sup IIDd(ul. u 2 + 11'2)- Dd(ul' U 2 )1I)II(e l ,t'2)1I. 0<1<1
.
Both sups in the parenthesis converge to zero as (e I' "2) -+ (0.0) by.continuity of the .partial derivatives. •
If EI
= E2 == Rand {e l .e2} is the standard basis in R2 we see that
~( l(x+h.y)-/(x,y)=DI( !l x. Y )-1' 1m h I x, Y ). 1'1 EF• (IX h-O
84
BANACH SPACES AND DIFFERENTIAL CALCULUS
Similarly, (aflayXx, y)= D2(x. y)·e 2 E F. Define inductively higher derivatives
2.4.13 Exampl•• As an application of the formalism just introduced we shall prove that for f: U c R 2 -- F
D 2f(u)·(", w) = (D(Df)(u)''')'w Applying the chain rule to Df(' )'w - T", 0 Df. where T",: A E L(R 2, F) ..... A 'w E F, the above equals D(Df(' )'wXu)'"
- D( Dd(' }w1e l + DzI(' ),w2e2)(U )," =
D(W1 ax af + w2 af)(u)." ay
- wl( DI(
:~)(u)'vlel + D2( :~)(U}'02"2)
+ w 2( DI( =
(by 2.4.12(ii»
:~)(u)'olel + D2( :~)(u)'02e2)
af a_ f (U) OIWI_(U)+ 02w l _ 2
2
ax 2
axay
PROPERTIES OF THE DERIVATIVE
B5
'For computation of higher derivatives. let us note that by repeated . . application of 2.4.6. D'/(u)(', •...••,) .. : , ... :, {/(u
+
t ti.;)}I·'1-'.'.,,.-0 .
I-I
In particular for I: U c R m - R" the components of D'IC,,) in terms of the standard basis are
Thus I is of class C' iff all its rth order partial derivatives exist and are continuous.
2.4.14 propoeltlon. (L &ler). L=(E, F); i.e.• DrIC,,) is symmetric.
II I: U c E -
F is C. then D'/( u) E
Proof. Consider fllst the
C{lSe r=2. We want to show DZ/(u)·(o.w)= DZ/(u)'(w, tI) where u E U, 0, wEE. o. W'" O. Define A E L(Rz. E) by 0, Aez " w. (., •• z) ~ng the standard basis of R2. Since for (x; y) E A-'(U), A(x. y) ... u, we have D 2 (fo A)(x. y)(e •• "2) = D1/(uXtI, w) CExercise 2.3E), it suffices to prove that D2(f 0 A) (x. YXe~.ez) - D2(f 0 A)(x. Y)(.2,e,); i.e.•
A.,-
iJ2(foA) iJ2(foA) iJxay ... iJyiJx
(see 2.4.13). Let I-loA: V- A -leU) c R2 - F. Since for any). E r, a2().o/)/iJxay->..(a zlliJxiJy). using the Hahn--Banach theorem it suffices to prove that ~_ iJ 21 iJxiJy iJyiJx where ]-l\ ./: VcR 2 - R, which is a sta~dard result from calculus. For the sake of completeness we recall the proof. Applying the mean value theorem twice, we get .
Sit." - [J(x + h. y+ k)- }(x,y + k)]- [/(x + h, y)-l(x, y)] =
(.u(clt.",Y+k)-
;~(CII,I<.Y»)k
.. ~ iJxiJy (Cit.", dlt .• )hk
86
BANACH SPACES AND DIFFERENTIAL CALCULUS
for some Cir • A' d lr • 1r lying between x and x + hand y and y + k, respectively. By interchanging the two middle terms in SIr.A we can derive in the same way that
Equating these two formulas for Sir k' cancelling h. k. and letting h -+ 0, k ..... 0, the continuity of D2j gives the' result. For general r, proceed by induction:
= D2(D'-2f)(u),(~,VI)'(":3"",vn) =
D'/( u)· (~, VI' ":3, ... ,vn )·
Let a be any permutation of {2, ... , n}, so by the inductive hypothesis
Take the derivative of this relation with respect to u E U keeping fixed and get (Exercise 2.4F)
U2'''''U"
Since any permutation can be written as a product of the transposition {t,2,3 ..... n} ..... (2.1.3, .... n} (if necessary) and a permutation of (2 .... ,n},
the result follows.
•
Suppose U c E is an open set. Then as +: E X E -+ E is continuous, there exists an open set ifc E X E with (i) U X {O} c if. (ii) u + ~h E U for all (u. h) E if and ~ ~ ~ 1, and (iii) (u. ") E if implies u E U. For example, let if= {( +) -1(U)}n{U X E}
°
Let us call a set if, temporarily, a thickening of U. See Figure 2.4.2. 2.4.15 Taylor'_
Theorem.
A map f: U c E
-+
F is of class C iff there
an C'OIIIinuous mappings 4Jlp : U
c E ..... Lf(E. F).
p=t, .... r
PROPERTIES OF THE DERIVA TlVE
87
v [; = I hickening or V
E
Flgunt 2.4.2
and where
if is some thickening of U.
such that for 0/1 ( u. h) E
if.
where hR=(h ..... h) (p times) and R(u.O)=O. If f is C then necessarily II'P = DPf and
~
____.
R(u.h)=
1
1 (I-t)'
0
r!
' f _,
~_
~......,'
(D'f(u+th)-D'f(u»dt.
Proof. We shall now prove the "only if' part. The converse is proved in Box 2.4B. Leibniz' rule immediately gives the. following integration hy partJ formula: If (a. bJ cUe Rand 1/1,: U c R ..... E,. i = I. 2. are C l mappings and B E.L(EI' E2 ; F) is a bilinear map of E. x £2 tn F. then
tB( I/I,(t). 1/12(t») dt = B( I/I.(b). I/Ilb»-B(I/!.(O).1/!2(O» a
88
BANACH SPACES AND DIFFERENTIAL CALCULU$
Assume/is a C' mapping. If r = I, then by 2.4.7
)'h = I(u)+ DI(u)"" + (fo'( DI( u+ th)- DI( u» dt)'h
feu + A) - l(u)+(fo'DI(U + tAl dt
and the formula is proved. For general k ~ r proceed by induction choosing in the integration by parts formula EI=R, E 2 =E, B(s,e)=se, .p,(t)= DIr/(u+tA)·A Ir , and "'2(t)=-(I-t)lr/k!. Since Dk/(u)EL~(E.F) by 2.4.13, Taylor's formula follows. • Note that R(u, h)·A'", o(h') since R(u, A) .... 0 as A .... O. If I is C+ I then the mean value inequality and a bound on D'+ II gives R( u, ")'"'''' o(A'+ I). See Exercise 2.4M for the differentiability of R. If / is COD (that is, is C for all r) then we may be able to extend' Taylor's formula into a convergent power series. If we can, we say f is of class cw, or analytic. A standard example of a COD function that is not analytic is the following function from R to R (Fig. 2.4.3) 8(x)
=
{exp{ -1/(I-l x I2 )},
0,
Ixl < I. Ixl;;d
This function is COD, and all derivatives are 0 at x =
y
-\
Figure 2.4.3
± l.
Hence all
PROPERTIES OF THE DERIVATIVE
89
coefficients of the Taylor series around these points vanish. Since the function is not identically 0 in any neighborhood or ± I. it cannot be analytic there. 2.4.18 Example. (Dif/erelllillting 1IIUle, the integral). Let U c E be open and I: (a •. bIXU - F. For each 1 E (a. hI define g(l): U -+ F by g(/)(u) = 1(/, u). If for each 1 E (a. hI, g(l) is of class C and if the maps (I, u) E (a. h] xU .... Dj( g(l»( u) E Lf( E. F) are continuous, then h: U -+ F. defined by h(u)" tlf/.u)dl= tg(/)(Il)df U
tI
is C' and D'h(u)- tDU(/.U)dt,
)=1 ..... "
U
where Du means the partial derivative in u. For, = I. write
lih(u + e)- h(u)- tD(g(/»(U)'edlll Q
=
Ilf(fol
(D(g(t»(u + se )·e - D(g(l)( u )'e)ds) dIll
4I;(h-a)lIell sup IID(g(/»(u+se)-D(g(t»)(u)lI-=o(e).
Q" 0 .. , , ..... " 1
For , > lone can also use an argument like this. but the converse to Taylor's tlleorem also yields the result rather easily. Indeed, if R(t, u,e) , denotes the remainder for the C' Taylor expansion of g( I), then with 'P, ... DPh = J:DP(g(l)] dl, the, remainder for h is clearly R(u. e) "" J:R(/.u.e)dl. But R(/.u,e)-O as e-O uniformly in I. so R(u.e) is continuous and R(u,O) = O. Thus his C. •
BOX 2.4A
THE LEIBNIZ AND CHAIN RULES
Here the explicit formulas are given for the" lh other derivatives of products and compositions. The proofs are straightforward but quite messy induction arguments, which will be left to the interested reader. 1. The Lelbnlz Nle. Let E. F" F2 , and G be Banach spaces. U c E an open set. I: U - Fl' and g: U -+ F2 of class C A and
90
BANACH SPACES AND DIFFERENTIAL CALCULUS
BE L(F•• F2 ; G). Let f x g: U -+ F. X F2 denote the mapping (f X g)(e) - (f(e). g(e» and let B(f. g) = B.o(f x g). Thus B(f. g) is of class C" and by Leibniz' rule.
DB{f. g)( p)·e = B(Df( P )·e. g( p»+ B(f(p). Dg( p)·e). Higher derivatives of f and g are maps
where
Denote by
the bilinear continuous mapping defined by
is defined by
for p
E
U. Leibniz' rule for k th derivatives is
where Sym": L"(E; G)-+L~(E; G) is the symmetrization operator. given by (see Exercise 2.21)
where S" is the group of permutations of {I ..... k}. Explicitly. taking
PROPERTIES OF THE DERIVATIVE
91
advantage of the symmetry of higher order derivatives. this formula is
DkB(f. g)( p). (e l , .. • !e k )
"es"
.. (I) < ... < "(i) ,,(;+I)<···<,,(k)
2. The cheln rule. Let E, E, G be Banach spaces and U c E, and VeE be open sets. Let f: U -+ V. and g: V -+ G be maps of class C k • By the usual chain rule, go f: U -+ G is of class C k and
D( go f)( p) = Dg( f( p» 0 Df( p) for pEU. For every tuple (i.}I .... ,},), where ;>1. and define the continuous multilinear map
r:.~_I},=k.
by
).:.1,··,j'(A, BI .... ,B,)·(el ..... ek )
"" A( BI(el, .... ell) ... ·.B,(e},+ ...• /,
," I .... ,ek ))
forAEL'(E; G), B,EL}'(E; £).I=I, ... ,i, and el ..... ekEE. Since
Dhf: U -+ LJ/(E; F). we can define
by
p .... ).:.1, ....1.( Dig(f( p», DJ'f( p) •.... D"l( p»). With these notations. the kth order chain rule is
k
__ k Sym·
0
~
't"' kl... L.. . I .... '1 A'·}'····,j· ; - I J, + ... I 1. - Ie 11· Ji' L..
0
( D'g .
0
i
X
. X ••• X Dii) D"f I
.
where Symk : Lie (E; G) -+ L! (E; G) is the symmetrization .operator.
•
I
92
BANACH SPACES AND DIFFERENTIAL CALCULUS
Taking into account the symmetry of higher order derivatives. the explicit formula is
D"(gof)(p)·(e, •...• e,,)=
E" i-I 1, + ... + j, - "
where the third sum is taken for
I". and p E U.el •...• e" E E. BOX 2.4B
II
< ... < I", ...• Ij , + ... +),
,+ I
< ... <
THE CONVERSE TO TA YLOR'S THEOREM
This theorem goes back to Marcinkiewicz and Zygmund (1936). Whitney [1943a). and Glaeser (1958). The proof of the converse that we shall follow is due to Nelson [1969]. Assume the formula in the theorem holds where ~p and Rare continuous. We shall show that f is cr. ~p = Dpf. I ~ p ~ r. and that R(u. II) has the desired expression. If r = I. the formula reduces to the definition of the derivative. Hence CJ11 = Df. f is C l and thus R(u. II> has the desired form. using 2.4.7. Inductively assume the theorem is true for r'" p -I. Thus ~j = Dif. for I ~ j ~ p -I. Let 11.1 E E be small in norm such that u + II + k E U. Write the formula in the theorem for f( u + II + k) in two different ways:
f(u + II + k) - f(u + 11)+ Df{u + II)·k
+ ... +
f( u + h + k)
=
I
(p -I)'
Dp-'f(u+II)'k p- 1
f( u) + Df( u)· (h + k)
+ ... +
I Dl'lf(u)'(h+k)1'1 ( p - I)!
I
+ p! ~1'(u)·(11 + k)
'J
+
1)·(11+1)1'.
PROPERTIES OF THE DERIVATIVE
93
Subtracting them and collecting terms homogeneous in Ai we get: 80(1.)+ 81(h)·k + ... + 8,_I(h)·k,-1 + 8,(h)·AP = R I ( u + It. k )·k P - R 2 (U. It + A)· (It + k)'.
where 8ilt) E Li(E. F). 8,(0) = 0 is given by
l[
.
8ilt) - -:"j' DJf(u + It)- D'f(u)-
J.
E
i-I
8p - I (It)';" (p
P-I-'I .. 7jDJ+'f(u)·It·'
(p~ j)!'Pp(U).ltP-l
~ I)! [ DP - If( U + It) -
and 8 p (It) = (l/p!)('Pp ( u + II) -
'l"
I.
OtfOjtfOp-2;
DP- 1/( u) - 'Pp ( u )·It ] ;
(u)).
*
Let IIkll satisfy lilt II ~ I'kll" tllh~ ~ince IIR.(u
+ It.A)k P - R 2 (u. It + k)·(h + k)P - gr( h)'k"l1
tfO (IIR I(u + It. k )11+ 118p ( It )1I)lIkIlP + IIRi( u. h + k )11' (lilt II + IIkll)P tfO {IIR I(u + It. k )11+ 118p (It)1I + IIR 2( u. It + k )II}( 1+ 3P)lIltllP/2 P and the quantity in braces { } ...... 0 as It ...... O. it follows that RI(u + It. A )·k P - R2(U. It + A )·(It + A)P - 8p (It)·k P - o(ltP).
Hence
go(It)+81(II)·k+ .. · +8,_I(h)·k p - I =0(ltP) .. We claim that subject to the condition *lIltll tfO IIkll tfO !lIhll. each term of this sum is o(h P ). 1(;\1'" .• Ap are distinct numbers. repIilce k by A}l in the foregoing. and get a p X P linear system in the unknowns 8o(l.)..... gp- l(h)·k P - 1 with Vandermonde determinant n i < j (A i Aj ) " 0 and right-hand side a column vector all of whose entries are o(h P ). Solving this system we get the result claimed. In particular
.
94
BANACH SPACES AND DIFFERENTIAL CALCULUS
Using polarization (see Box 2.2B) we get
..;; (p-l)P-I II(DP-I/(u+")-DP-I/(u)-cp (u»)"11 (p_I)! I' =-"(p-I)P-I sup II(DP- I/(u+")-DP I/(u)-cp (u»).~P III (p-I)! 11"11",1 I' "
..
=
(p-I),-I (-1)1
p
sup
II
. 11'11 '" IIAIV2
-I -I (2k)P-III (DP I(u+")-DP I(u)-'Pp(u»). -11"11
(2( p - I» P- I () ( "I' ). (p -1)!I"" P - I
Since 0(" P)/11"11 p .... 0 as " .... 0, this relation proves thai DP II is differentiable and D'/(u)=cp (u). Thus I is of class CP, 'PI' heing continuous, and the formula (or R follows by subtracting the given formula for I( u + ") from Taylor's expansion. • The converse of Taylor's theorem provides an alternative proof that D'/(u) e L~(£, F). Observe first that in the proof of Taylor's expansion for a C' map I the symmetry of Di/( u) was never used, so if one symmetrizes the Di/(u) and calls them 'Pi' the same expansion holds. But then the converse of 2.4.12 says that 'Pi ... Di{. We shall consider here simple versions of two theorems from global analysis, which shall be used in Box 4.1 C, namely the smoothness of the evaluation mapping and the "omega-lemma." Let 1- (0, I) and £ be a Banach space. The vector space C'( I; E) of C'-maps (r > 0) of 1 into £ is a Banach space with respect to the norm
(see Exercise 2.4H). It is easy to see that if U is open in £, then the set C'(/; U) - (f e C'(/: £)11(1) c U} is open in C'(I; E).
PROPERTIES OF THE DERIVATIVE
2.4.17 Proposition.
95
The evaluation map
cv: C(I; U)X]O.I[-U defined by
ev(f. I)
=
f( I)
is C' and
k
Dk/(t)·(SI.···.Sk)+
=
E Dk
Ig,(/)·(·~I.···.·~, 1•.~"I ..... Sk)
I-I
lor i= I •...• k.
Proof. For (g.s)EC(l;E)XR. define IKg.s~l=max(llgllk;lsl). Note that the right-hand side of the formula in the statement is symmetric in the arguments (g,. s,). i = I .... •k. We shall let this righthand side be denoted 'P,: C(l; U)X(O.I) ..... L~(C(/; E)xR;' E). Note that 'Po( I. I) = I( I) and that the proposition holds for r = 0 by uniform continuity of I on J since 11/(/)- g(s~l.s;; 11/(·1)- l(s~l+ III - gllo. Since
for alii
E
JO.I!
lim
_D"",,'g,-,,(....:.I),.....s_'
(11 •." .... (0.11)
lI(g.s)II'
=0
by Taylor's theorem for g we get
ev(f + g.1 + s) = I( I + s) + g( 1+ s)
,.;, t ~(DI/(t}.s; +
DIg(t)·s'}+ R(/ •.d·s'
; - 0 I.
= 1(/)+
, 1 ·
E -:-;-'P,(f./)·(g. s)' + R«(f./).(g. s»·(g •.v) i - I I.
•
where R ( ( I. I ) • ( g. s ) ) . ( ( g 1 • S I ) ••••• ( g,. s, ) )
, = R(/. S)·(SI •...• S,)+
E D'g/(/)·(SI .. ··.S,). i-I
r
•
96
BANACH SPACES AND DIFFERENTIAL CALCULUS
which is symmetric in its arguments and R«/. 1).(0.0»),- O. By the converse of Taylor's theorem, the proposition is proved if we show that every tp" 1< i < r, is continuous. Since
IIDAr-'g,(t)- DAr-'g,(s)1I < It - slsupIlD"gj(u)1I < It -slllg,ll, lIel
.
by the mean value theorem, the inequality
lIe tp" (f, 1)- tp,,(g, s»· «g,. s,), ... ,(g", s,,»11
+
···ls"l+
" IID"-'g/(t)-D"-'g/(s)lIls.l ···ls,_.1 ISH.I·· ·lsAI E
1-'
implies
IItp.. (f.I)- tp,,(g, s)1I < IID"/(I)- D"g(s )11+ kll- sl < IID"/(I)- D"/(s)II+IID"/(s)- D"g(s)lI+ kll- sl < IID"/(I)- D"/(s~I+2kll(f, 1)-(g,s)lI. Thus the uniform continuity of D"/ on J implies the continuity of tp" at (f. t). • Let us consider now the omega lemma. t Let M be a compact topological space and E, F be Banach spaces. With respect to the norm
11/11- sup 11/( m)11, ",eM
the vector space Co( M. E) of continuous E-valued maps on M, is a Banach space. If U is open in E, it is easy to see that CO(M. U) = {f E CO(M, E)I/(M) c U} is open.
2.4.1' PropoeItIon. (0rrrIgG Lerruna). Let g: U - F be a C' map. > O. The mIJfJ
r
0,: CO(M,U)-CO(M,F) de/inedby O,(/)=go/ tlbis is tmninolo&Y of Abraham (1963). Various results of this type can be traced badt to earlier works of SoboIev 11939) and Eells (1958).
PROPERTIES OF THE DERIVATIVE
97
is also 01 class C'. The derivative of D. is DO.(f)·h = [(Dg)o f]·h [DO.(f)·h ](x)'" Dg(f(x »·h( x).
Le ••
The formula for DO. is quite plausible. Indeed. we have d . DO.(n·h(x)"" deO.(f+eh)(x)I,-o
d - de/{ (f(x)+eh(\»I._u
By the chain rule this is Dg(f(x»·h(x). This shows that if 0,11 is differentiable. then DO. must be as stated in the propllsition.
Proof. Let! E
CO(M.U). By continuity of g and compactness of M.
1I0.(f)-O,(/')II- sup IIg(f(m»- g(f'(m)~·1 meM
is small as soon as point I .
III - f11 is small; i.e.• 0. is continuous at each
. Let
A,: CO{M. L~(E: F» .... L~(CO(M.E); CO(M. F»
be given by A;(H)(h ...... h;)(m) ... H(m)(h.(m) ..... h;(m») for H e CO(M. L'(E; F». h ...... h, eCo(M. B) and '" eM. The maps A; are dearly linear and are continuous with IIA,II<1. Since D'g:
U .... L~(E; F) is continuous. the maps
pr~g
argument shows that the
are continuous and hence Ai ODD'.: CO(M.U) .... L~(CO(M. E); CO(M.
F»)
is continuous. The Taylor theorem applied to g yields . , I . g(f(m)+ h(m» - g(f(m»+ -=tD'g(f(m»·h(m)'
E
•
i-I"
+ R(f(m). h(m»'h(m);
98
BANACH SPACES AND DIFFERENTIAL CALCULUS
so that defining
and
[R(f. h)· (hi'" .• h,)](m) - R(f(m). h(m»·( hl(m ) ..... h,(m» we see that R is continuous. R( /.0) ... O. and
O,(f+ h) - go(f+ h) =go/+
i:. ~(Digof)'hi+R(f.h)'h'
i - I I.
Thus by the converse of Taylor's theorem DiD, = Ai 0 Do., and 0, is of class C'. • This proposition can be generalized to the Banach space CU, E),I = [a, hI, with the norm 11'11, given by the maximum of the· norms of the first r derivatives; i.e., lilli, = max o .. i .. ,sup, E ,lIl'i)( I ~I. If g is Cr+I, then 0,: CU. E) ..... C,-hU, F) is C ' + h. Readers are' invited to convince themselves that the foregoing proof works with only trivial modifications in this case. This version of the omega lemma will be used in Box 4.1 B. For applications to partial differential equations, the most important generalizations of the two previous propositions is to the case of Sobolev maps of class H'; see for example Palais [1968], Marsden [I974b] and Marsden and Hughes (1982) for proofs and applications. Exercl...
2.4A Show that if g: VeE -+ L(F,G) is C'. then I: V X F ..... G, defined by 1(II.v)=(g(II»(v).ueU.veF is also C. (Hin/: Apply the Leibniz rule with L(F. G)X F ..... G the evaluation map.) 2.4B Show that if I: VeE ..... L(F,G).g: VeE ...... L(G,H) are C mappings then so is h: VeE ..... L(E.H). defined by h(u)= 8(u)0 I(u).
2.4C
Extend Leibniz' rule to multilinear mappings and find a formula for the derivative.
PROPERTIES OF THE DERIVATIVE
99
2.40 oefine a map I: U c E -+ F to be of class TI if I is differentiable. its tangent TI: U x E --+ F x F is continuous and IIDI(x~1 is locally bounded. (i) For E and F finite dimensional. show that this is equivalent to CI. (ii)
(L. Rosen) Let E be the space of real sequences x = (XI' x 2 • ••• )
such that In l/2x .. 1 is bounded and setll\lI-suPn;.aln1f2x"l. Check that E is a Banach space. Define f: E --+ E by I(x)" =, I,,(x,,). where I,,: R -+ R is a smooth wnvex function satisfying I,,(y) = 0 if Y ~ lin and My) = y - 2/n if v ~ 31n. Show that I is TI but is not CI. (Hint: You have to show DI(x) is strongly continuous but is not norm continuous.) (iii) (Project) Investigate the validity of the chain rule and Taylor's theorem for maps of class T'. (iv) (Project) Show that the function developed in Smale (1964] is T2 but is not C 2 as is claimed. 2.4E Suppose that I: E -+ F (where E. F are real) is homogeneous 01 degree k; i.e .• l(te) = tk/(e) for all t E R. and e E E. (i) Show that if I is differentiable. then D/(u)'u = k/(ri). (Hint: Let g(t) = I(tu) and compute dgldt.) (ii) If E = A", F = A, show that this relation is equivalent to n
aI' kf(x). '
L x'ax =
,-1
Show that maps multilinear in k variables are homogeneous of degree k. Give other examples. (iii) If I is C k show that I(e) - (Ilk!) Dk/(O)e k , i.e ..! E Sk( E. F) and thus it is C .... (Hint: 1(0) = 0; inductivrly applying Taylor's theorem and replacing at each step h by Ih. show that
2.4F
Let e l •... • e.. _ lEE be fixed and I: U c E -+ F be n times differentiable. Show that the map g: U c E --+ F defined by g( u) = D,,-I/(u)(e l •. ..• e,,_I) is differentiable and Dg(u)·e = D"/(u)" (e.el ... ·.e,,_ I)'
'
2.4G Prove the following refinement of 2.4.14. If I is C I and D,Dd(u), exists and is continuous in u. then D2 D I /( u) exists and these are equal.
100
BANACH SPACES AND DIFFERENTIAL CALCULUS
2.4H For I: U C E -. F, show that
T2/: (U x E)X(E x E) -. (F
X
F)X F x F:
(1I.~1.~2.~3) .... (1(11), DI(II)'~I' DI(II)'~2'
D2/(1I)' (~I' ~2)+ DI(II)'~3)'
I(x. y)=
{x
~ 2 • +y O.
X2*_y
x2 =
- y.
Show that (i) I is discontinuous at (0.0). hence is not differentiable at (0.0);' (ii) all directional derivatives exist at (0,0); i.e .• f is Gateaux differentiable. 2.41 (Differentiating sequences). Let I,,: U c E -. F be a sequence of C' maps, where E and F are Banach spaces. If {f,,} converges pointwise to I: U -. F and if (Djl,,},O < j < r, converges locally uniformly to a map gj: U -. LI(E, F). then show that I is c'. DJI- g and {f,,) converges locally uniformly to I. (Hint: For r -1 use the mean value inequality and continuity of gl to conclude that
11/(11 +/t)-/(II)- gl(II)'''1I .. 11/(11 +" )-1,,(11 + ,,)- [/(11)-1,,(11 >]11 +11/,,(11
+"
)-1,,(11)- DI,,(II)'''II
+IIDI,,(II)'" - gl(II)·/t1l
<ell"l1· For general r use the converse to Taylor's theorem.) 2.4K (ex lemma). In the context of 2.4.18, let
Show that ex, is continuous linear and hence is C". 2.4L Consider the map ~: CI([O. CO([O. given by
I» -.
1»
~(f)(x) - exp[/'(x)]
Show that
~
is Coo and compute
D~.
PROPERTIES OF THE DERIVATIVE
2.4M (H. Whitney [I 943aJ). Letf: U c E expansion
-+
101
Fbe of class Ck+p with Taylor
f(b) .. f(a)+Df(a}·(b-a}+ .. · +
~!Dkf(a)(b-a)·
(i) Show that the remainder Rk(a, b) is Ck+p for b ... a and C P for all a, bEE. If E = F = R, Rk(a, a) = 0, and lim (lb-al;D;+PRk(a,b»)=O,I~i~k.
h-u
.
(For generalizations to Banach spaces, see Toan and Ang [ 1979).) (ii) Show that the conclusion in (i) cannot be improved by consid- ' eringf(x)- Ixlk+p+1/2. 2.4N (H. Whitney (l943b)). Let f: R -+ R be an even (odd) function; i.e. f(x) - f( - x)(f(x) - - f( - x». (i) Show thatf(x) - g(x 2)(f(x) - xg(x 2 ) for some R. (ii) Show that if f is C lk (C 2 k+ I) then g is C k (Hint. Use the converse to Taylor's theorem) (iii) Show that (ii) is still true if k = 00 ,(iv) Let f(x) = Ix12k+ 1+112 to show that the conclusion in (ii) cannot be sharpened. 2.40 (M. Buchner, J. Marsden, and S. Schecter (1983)). Let E = L 4«(0, I)) and let cp: R -+ R be a Coo function such that ep'('-) = I, - I ~ '- ~ I and ep'(A) ... 0 if IAI ~ 2. We assume ep is monotone inc~easing with , ep'" - M for '- ~ - 2 and ep = M for A ~ 2. Let h: E -+ R be given by
h(u) = ~ fo'ep([u(x W)dx (a) ,Show that his C 1 using the converse to Taylor's theorem. (b) Show that if q is a positive integer and E = U«(O, I)), then h is c q - I but is not c q• (c) The formal L2 gradient of h is given by
Vh(u)
=
1
"31/1'(u)
where 1/1('-) = ep(>t). Show that vh: E -+ E is CO but is not CI.
102
BANACH SPACES AND DIFFERENTIAL CALCULUS
(d)
Let f(u)""
111
2
0
lu(x)1 2 dx+h(u). Show that on L 4 ,
f has a
formally non-degenerate critical point at O. yet this critical point is not isolated. 2.5 THE INVERSE AND IMPLICIT FUNCTION THEOREMS
The inverse and implicit function theorems are pillars of nonlinear analysis and geometry. so we give them special attention in this section. Throughout. E. F... . ,are assumed to be Banach spaces. In the finitedimensional case these theorems have a long and complex history; the infinite-dimensional version is due to Hildebrandt and Graves (1927). We first consider the inverse function theorem. This states that if the linearization of the equationf(x) = y is uniquely invertible then locally so is f; i.e. we can uniquely solvef(x)'" y for x as a function of y. To formulate the theorem, the following terminology is useful. 2.5.1 DefInition. A map f: U c E -+ V c F (U, V open) is a C'diffeo".""",_ iff is of class C, is a bijection (that is, one-to-one and onto V), and I is also of class c.
r
2.5.2 1m..... Mapping Theorem. Let f: U c E -+ F be of class C, r ~ I. Xo E U. and suppose Df(x o ) is a linear isomorphism. Then f is a C' diffeomorphism of some neighborhood of Xo onto some neighborhood of f(x o ) and Dr '( y) = [Df( y »)- , for y in this neighborhood of /( X o )'
r- '(
Although our immediate interest is the finite-dimensional case, for Banach spaces keep in mind the Banach isomorphism theorem: if T: E -+ F is linear, bijective, and continuous, then T- I is continuous. (See Box 2.21'.) c. Proof of17teomn 1.5.1. We begin by assembling a few standard lemmas. First recall the contraction mapping principle from Section 1.2. 2.5.3 Lemme. Let M be a complete metric space with distance function d: M X M -+ R. Let F: M -+ M and assume there is a constant >., 0 ~ A < I such that for all x, y E M, d( F(x), F(y».,. Ad(x, y). Then F has a unique fixed point Xo EM; that ;s F(xo) = xo'
This result is the basis of many important existence theorems in analysis. The other fundamental fixed point theorem in analysis is the Schauder fixed point theorem. which states that a continuous map of a
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
103
compact convex set (in a Banach space, say) to itself, has a fixed point-not necessarily unique however. 2.5.4 Lemma. Lei GU E. F) denole Ihe sel of linear isomorphisms from E onlo F. Then GUE. F) c L(E. F) is open. Proof. ' Let
lIall = sup lIa( e )11 ' I'E 1;.
111'11-1
be the norm on L(E. F) relative to given norms on E and F. We can assume E = F. Indeed. if CPo E GUE. F). the map"' ...... CPo 1 0 ", from L(E, F) to L(E, E) is continuous and GL(E. F) is the inverse image of GL({lE). For cp E GUE, E), we shall prove that any'" sufficiently near cp is also invertible, which will give the result. More precisely. II'" - cpll < IIcp III I implies", E GUE, E). The key is that 11'11 is an algehra norm. That is. liP 0 all ~ IIPliliall for aE L(E, E) and PE L(E, E) (see Section 2.2). Since '" = cp 0 (I -' cp - I 0 (cp _. '" cp is invertible. and our norm assumption shows that IIcp - I (cp - '" ~I < I. is sufficient to show that I - ~ is invertible whenever II~II <.1. (1 is the identity operator). Consider tbe following sequence (called the Neumann series):
»,
0
~=I
EI = 1+ E + ~ + ~a ~
~2 = I
Using the triangle inequality and the foregoing norm inequality. we can compare this sequence to the sequence of real numhers. I. I + II Ell. I + II Ell + 1I~1I2 •.•. , which we know is a Cauchy sequence since II Ell < I. Because L(E. E) is complete. E" must converge. The limit. say p. is the inverse of I-E. Indeed (1- ~)E" = I-(E 0 ~ 0 • • • 0 E). so the result follows. • 2.5.5 Lemma. LeI'l:GL(E.F) .... GL(F.E):cp .... cp I. COO and D~(cp)'''' = - cp-I",cp-I. (For Dr~, see Box 2.SC)
Then~isofdrJss
We may assume GL(E,F)-0. If we can show that D9(cp)'''' = ' " 0 cp - I, then it will follow from Leibniz' rule that 9 is of class COO. Indeed, D!J = B(9, 9) where BE L 2(L(F, E); L(L(E. F). L(F. E))) is de-
Proof.
- cp - I 0
104
BANACH SPACES AND DIFFERENTIAL CALCULUS
fined by B(+I' +ZXJf) - - +1 0 Jf 0 +Z. where +1' +z E L(F. E) and Jf E L(E, F), which shows inductively that if !f is C" then it is Chi. Since .-1. . -1 is linear E L(E. F», WI" -I ..how that
+... -
(+
lim
+-.
11+- 1-(.-1_.-1+.-1 + .-1•• -1)11_ 0 . 11+-.11
Note that
+-1_(.-1_.-1 ..-1 +.-".-1)-+-1-2.-1 +.-1+.-1 - +-1(+ _ .>.-1(+ _ .>.-1. Again,
usin& II' 0 alllIii lIallII'li for caE L(E, F), PE
L(F, G),
11+-1(+ _ .>.-I( + _ .>.-111" 11+-11111+ _ .11 211.- 111 2. With this inequality. the limit is clearly zero.
...
To prove the theorem it is useful to note that it is enough to prove it under the simplifying assumptions that Xo = 0, /(x o) = 0, E = F, and Df(O) is the identity. (Indeed, replace/by hex) = Df(xo)-I o(f(x + xo)- /(xo»)') Now let g(x) -"" x - lex) so Dg(O) = O. Choose, > 0 such that II xII'" implies IIDg(xllllii i, which i, ,.~ible by continuity of Dg. Thus by the mean value inequality, IIxll ~ r Implies IIg(x~11Iii ,/2. Let 0.(0) - {x E Flllxll IIii e}. For y E D,./2(0), let g,v(x) - y + x - f(x). By the mean value inequality, if then
and Thus by Lemma 2.5.3, g" has a unique fixed point x in 8,(0). This point x is the unique solution of /( x ) - y. Thus / has an inverse
From (ii) we have (iii) IIr I(YI)_
ous.
r
I(Y2~1" 211 VI
-
Y211. so
r
I is continu-
From Lemma 2.5.4 we can choose, small enough so that Df(x) -I will exist for x E D,(G). Moreover, by continuity, IID/(x)-Ill" M for some M
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
and all x E D,(O) can be assumed as well. If YI' Y2 and X2 = '(Y2)' then
r
IIr I(y,)_ r
l
E
D,/2(O),
XI =
105
r I(y,),
(Y2)- D/(X2)-'(Y' - Y2)1I =
IIx, - X2 - D/(x 2) -1(j(X,)_ /( x2)]11
-IlD/(x2)-'{D/(x 2Hx, - x 2)- /(x,)+ /(x 2)}11 .. MII/(x,)- /(xz)- DI(xz)(x, - x2)1I.
r'
This, together with (iii). shoM! that is differentiahle with derivative ' DI(x)-' at I(x); i.e., D(r ') = ~ 0 DI 0 r' on Vo = D,/l(O). This formula. is C· -, then the chain rule. and Lemma 2.5.5 show inductively that if I C· for I ~ k ~ r. •
r'
r is
This argument also proves the following: If /: U --. V is a C',homeomorphism where U c E, and V c F are open sets, and D/(U)E GL(E. F) for U E U, ,then / is a C' diffeomorphism.
BOX 2.M THE SIZE OF THE NEIGHBORHOODS IN THE INVERSE MAPPING THEOREM
An analysis of the preceding proof also gives explicit estimates on the size of the ball on which I(x) = Y is solvable. t Such estimates are sometimes useful in applications. The easiest one to use in examples involves estimates on the second derivative. '
2.5.1 Corollary. Suppose I: U c E a,.d D/(x o ) is an isomorphism. Let
-+
F is 01 class C', r ;;, 2, Xo E U
L-IiD/(xo)1I and M-II DI(xo)-'U. Assume IID2/(x)U~K for IIx-xoll~R and DR(XO)CU. Let
R, - rnin{ 2~M' R} , R 2 ...
rnin{ ~I
'
2M(L
~ KR I )}
and R) ...
tWe thank M. Buchner for providing this formulation.
:t·
106
BANACH SPACES AND DIFFERENTIAL CALCULUS
Then f maps the ballllx - xoil..;; R2 dil/eomorphically onto an open set containing the balllIy - f(xo~l..;; R]. For YI' Y2 E DR,(f(xo», we have II/I(YI)- II(Y2~1";; 2LIIYI - Y211. . Proof.
We can assume Xo = 0 and I(xo) = O. From
DI(x) = DI(O) + foID2/(tX)Xdt we get IIDI(x~1 ~ L + Kllxll for x
E
DR(O). From the identity
DI(x) - DI(O){ 1+ [DI(O)] " Ifo'D2/(tX)Xdt} and the fact that
11(1+ A)-'II ... I +IIAII+IIAII2 + ."
=
I-:IAII
for IIAII < I (see the proof of 2.5.5) we get
IIDI(x)-'1I..;;2L if IIxll ... R and
IIXII"'2~K'
i.e., if IIxll ..;; R I'
As in the proof of the inverse function theorem. let gy(x) = [DI(On- I(y + DI(O)x - I(x». Now '
IIg,,(x)lI..;; M(IIYII+ IlfoID2/(tX)Xdtll) ... M(IIYII+ Kllxll). Hence for lIylI";; RI/2M. S" maps DR,(O) to DR(O). We similarly get from the mean value inequality and the estimate
IIg,.(xl)-g,.(X2~1"'!lIxl-X211
if IIxll..;; R I • Thus. as in the previous proof. II: DR,/2M(O) -+ DR, is defined. Note that by the mean value inequality,
II/( x )11 ... (L + Kllxll)lIxll so if IIxll..;; R 2 , then II/(x~I'" R I /2M. The rest now follows as in the proof of the inverse mapping theorem. •
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
107
In the study of manifolds and submanifolds. the argument used in the following is of central importance.
2.5.7 Implicit Function Theorem. Let VeE, V c F be open and I: V X V - G be C', r ~ I. For some Xo E V, Yo E V assume Dzi(x o' Yo): F- G is an isomorphism. Then there are neighborhoods l.{) of XII and Wo of fCl(). Yo) and a unique C' map g: V() X!Vo - V such that for all (x. w) E Vo X Wo' I(x,g(x,w»=-w.
Proof. Consider the map 41: VXV-EXG,(x.y)'+(x'/(x.y». Then D4I(x o• Yo) is given by
~4I(xo. Yo)-(x
i•
YI) = (DI/(x/(). "0) ),1
D2 f( XII' 0 Yo
»)( YIXI')
which is easily seen to be an isomorphism of E X F with E X G. Thus 41 has a unique C' local inverse. say 41 I: l.{) X J.t() - V x V.( \. w) .... (x. g(x. w». The g so defined is the desired map. _ Applying the chain rule to the relation explicitly compute the derivatives of g: '
I( x, g( x, w» =
w.
one can
Dlg(x. w) = - [Dzf(x. g(x. w»r 10 Dd(x. g(x, w»
2.5.8 Corollary. Let VeE be open and I: V - F be C. r ~ l. Suppo.se DI(xo) is surjective and kerDI(x o ) is complemented. Then I(V) contains a neighborhood oll(x o )' .
Proof.
Let EI = ker Df(x o ) and E = E. X E2 • Then Dzf(xo}: E"l -+ F is' an isomorphism. Thus the hypotheses of Theorem 2.5.7 arc satisfied and SO I(V) con~ains Wo provided by that theorem. _ . Since in finite-dimensional spaces every subspace splits. the fotesoina corollary implies that if I: VcR II - R m, n ~ m and the Jacobian of I at every point of V has rank m, then I is an open mapping. 'Ibis statement generalizes directly to Banach spaces, but it is not· a consequence of the implicit function theorem any more. since not every subspace is split. This
108
BANACH SPACES AND DIFFERENTIAL CALCULUS
,0e5
result back to Graves (1950). The proof given in Box 20SB follows Luenberaer (1969).
2.5.9 Propoeltlon (L«Ill S"rjectivity).
Iff: VeE -+ F is C l and Df(uo) is onto for some U o E V, then f is loca/~v onto: i.e., there exist open neighborhoods VI of U o and VI of f( u) such that fl VI: VI -+ VI is onto. In particular, if Df(u) is onto for u E V, then f is an open mapping.
BOX 2.5B PROOF OF THE LOCAL SURJECTIVITY THEOREM Proof. The key is to recall from Section 2.1 that E/ker Df(u o ) = Eo is a Banach space with norm lI[xll = inf{lIx + uiliu E ker Df(u o )}, . where [x) is the equivalence class of x. To solvef(x) .. y we set up an iteration scheme in Eo and E simultaneously. Now Df(uo) induces an isomorphism T: Eo -+ F, so r- I E L(F, Eo) exists by the Banach isomorphism theorem. Write
f(x) = y
as f(uo+h)-y,
i.e.,
T-I(y-f(uo+h»-O,
where x .. Uo + h. To solve this equation, define a sequence L" E Ejker Df(u o ) (so L" is a coset of kerDf(uo» and h" E L" c E inductively by Loker Df(u o ), ho e Lo small, and
(I) and selectin, h" E L" such that
(2) The latter is possible since ilL" - L,,-.I1 = infOlh - h,,_ .Illh E Ln}' Since h"_1 E L"_I' L"_I = r-1(Df(uo)h,,_I)' so
Subtractin, this from the same expression for L" _ I gives L" - L,,_I - - r-1(f(UO + h,,_I)- f(uo - Df(uo)(h,,_I- h,,-2»'
+ h,,_2) (3)
109
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
For e> 0 given. there is a neighborhood U of U o such that II D/{ u)II e U, since/is ct. Assume inductively that 110 + e U arid "0 + h,,-2 e U. Then from the mean value inequality, '
h._,
D/{lIo~1 < e for
(4) By (2),
Thus if e is small.
Starting with ho small and inductively in U since
IIh, - holl < !lIholl, Uo + h"
remains
IIh"lI..; IIholl+llh,- holl+llh 2 - h.ll+ ... +lIh" - h,,_.l1 ..; ( 1+ ! + ... + 2..t-, )lIholi ..; 21lholl· It also follows that h" is a Cauchy sequence, so it converges to some point. say h. Correspondingly. L" converges to L aild h E L. Thus, from (I),
and so
This proves that for y near Yo = f(uo).!(x) = y has a solution. If there is a solution g(y)= x which is C', then Df(xo)o Dg(yo) = I and so rangeDg(yo) is an algebraic complement to ker Df(xo)' It follows that if range Dg(yo) is closed, then ker Df(x o } is split. In many applications to nonlinear partial differential equations. methods of functional analysis and elliptic operators can be used to show that ker D/(xo) does split. even in Banach spaces. Such a splitting theorem is called the Fredholm alternatwf', For illustrations of this idea in geometry and relativity, see hscher and Marsden [1915),[1919), and in elasticity, see Marsden and Hughes [1982. ch. 6). For such applications, 2.5.8 suffices.
The locally injective counterpart of this theorem is the following.
110
BANACH SPACES AND DIFFERENTIAL CALCULUS
2.5.10 PropoIltIon (LOCtIIllfjen;.,;ty). Let I: U c E - F be a C· map. DI(uoXE) be closed in F and DI(uo)E G1..(E. DI(uoXE». Then there exists a neighborhood V 01 u o• V c U on which I is injective. The inverse I( V) - U is Lipschitz continuous.
r-.:
Since (DI(uoW·
Proof.
E
L(D/(uo)(E). E). there exists M > 0 such that
II DI( Uo )ell ~ Mllell for all e E E. By continuity of DI. there exists r > 0 such that IID/(u)- D/(uo~1 < M/2 whenever lIu - uoll < 3r. By the mean value inequality, for e •• e2
E
D,(u o )
II/(e.)- l(e2)- D/(uo}(e. - e 2 )11 Eo;
sup IIDI{e.
+ f(e2
- e.»- DI(uo)II·lIe. -
e.~1 <
3r. Thus
e211 Eo; Mlle. - e211/2
IEIO •• )
since lIu o - e. -t(e2 -
i.e.•
which proves that I is injective on D,( uo) and that Lipschitz continuous. •
l .: I( D,( un» -+ U
is
We now give an example of the use of the implicit function theorem to prove an existence theorem for differential equations. For this and related examples we choose the spaces to be infinite dimensional. In fact, E. F. G, ... will be suitable spaces of functions. The map I will often be a nonlinear differential operator. The linear map D/(xo> is called the linearization of I about xo' (Phrases like "first variation," "first-order deformation." and so forth are also used.) 2.5.11
Example.
Let E = alI C· functions I: [0. I] -+ R with
11111. =
sup I/(x)1 + sup , E
(0 •• (
,
E
Id/(X) I ~
(0 •• 1
and F= all CO functions with 11/110 = sup, E' (Il.II I/(x)!. These are Banach spaces (see Exercise 2.1 G). Let F: E -+ F. F(/ ) = dfI dx + t-'. It is easy to check. that F is C«l and DF(O) = dldx: E -+ F. Clearly DF(O) is surjective (fundamental theorem of calculus). Also. ker DF(O) consists of E." all
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
111
constant functions. This is complemented because it is finite dimensIonal; explicitly, a complement consists of functio~s with zero integral. Thus Corollary 2.5.8 yields the following. There is an t> 0 such that if g: [0.1) -+ R. is a continuous lunction with Ig(x)1 < t. then there is a C' lunction I: [0.1) -+ R such that
The following series of consequences of the inverse function theorem are important technical tools in the study of manifolds, The first two results give. roughly speaking. sufficient conditions to "straighten out" the domain (respectively. the range) of I in a neighborhood of a point. thus \llaking I look like an inclusion (respectively. a projection), 2.5.12 Locallmm....lon Theorem. Let I: U c E -+ F be 01 ciass C. r ;. I. Uo E U and suppose that D/(uo) is one to one and has a closed split image' F, with closed complement F2, (II E = R m and F = R n. assume only that DI(u o ) has trivial kernel.) Then there are open sets U' c F. and V c EeF2 where I( uo) E U' and a C' diffeormorphism cp: U' - 0 V such that (cp 0 1)( e) (e.O) lor all e E V neE X{O}) c E. The intuition for E = £, = R2. £2 = R (i,e.• m = 2.,. = 3) is given in Fig. 2.5.1. The function cp nattens out the image of f. Notice that this is intuitively correct; we expect the range of I to he an m-dimensional' "surface" so it should be possible to natten it to a piel:e of R m , Note that the range' of a linear map of rank m is a linear suhspace of dimension eltactly m. so this result eltpresses. in a sense. 'a generalization of the linear case. Proof. Define g: UXF2 cExF2 -+F=F,eF2 by g(u.v)=/(u)+(O.o) and note that g(u.O)=/(u). Now Dg(uo,O)=(DI(uo),i,)eGL(Ee F2 • F) by the Banach isomorphism theorem (IFl denotes the identity mapping of F2 and (A. B) E L(EeE, FeF') for A e L(E. F). and Be L(E. F) is defined by (A. B)(e. e') = (Ae.&').) By the inverse function, theorem there eltist open sets U' and V such that (uo.O) E V c EeF2• and g(uo'O)I( U o ),E U' c £ and a C' diffeomorphism cp: U' -+ V such that cp - • - gl V. Hence for (e.O) E V. (cp 0 I)(e) = (cp 0 g)(e.O) ... (e.O),. .
2.5.13 Local Subm....lon Theorem. Let I: U c E -+ F be 01 closs C. r ~ I. U o e U and suppose D/(u o ) is surjective and has split kernel E2 with closed complement E •. (II E = R'" and F= R". assume only that rank (DI(u o n.) Then there are open sets U' and V such that eU'cU C B
»-
"0
112
BANACH SPACES AND DIFFERENTIAL CALCULUS
Figure 2.5.1
and V C FeE2and a C' diffeomorphism 1/1: V --+ U' such that (f 0 1/1)( u, v) = u for all (u, v) E V.
The intuition for EI ... E2 = F = R is given in Fig. 2.5.2, which should be compared to Fig. 2.5.1. Proof. By the fundamental isomorphism theorem (Section 2.2); Df(u o) E 0L(E1, F). Defineg: U C E 1eE2 --+ FeE2 by g(u l • u2)= (/(u l , U2)' u2)and note that
so that Dg(u o ) E OL(E, FeE2 ). By the inverse function theorem there exist open sets U' and V such that U o E U' cUe E, V c FeE2 and a C' diffeomorphism 1/1: V --+ U' such that 1/1 1 = gl U'. Hence if (u, v) E V. (u, v) - (g 0 1/1)( II. t') = (/( 1/1 ( u, t' )).1/I2( II. I'» where 1/1 = 1/11 X 1/12: i.e.. 1/12( II. ,,) = v . and (/0 I/I)(u. v)= u. • .
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
/0 '" -
/ =
constant
113
constant
.~ .J-------E,
F
F
F1gu... 2.5.2
2.5.14 Corollary (Loclll Representation Theorem). Let f: U c E -+ F be of class C. r ~ I. U o E U and suppose Df( u o) has closed split image F, with closed complement F2 and split kernel E2 with closed complement Pl' (II E - R"', F- R If , assume that rank (DI(u o k, k" n, k" m, so that F2 R If - k, F," Rk. E, - Rk, E2 - R",-k.) Then there are open sets U' and V such that Uo E U~ cUe E, V c F,eE2 and a C' difleomorphism "': V -+ (j' such that (f 0 '" Xu, v) = (u. 1J( u. v». where 1J: V -+ F2 is a C r map satisfying D1J(u o ) = o.
»"
Proof. Write f = f, X f2' where /;: U --+ F;. i -1,2. Then f, satisfies the conditions of 2.5.13, and thus there exists a C' diffeomorphism!/-: V c Fie E2 -+ U'c E such that U, !/-)(u. v) = u. Let 1J = 12 !/-. • 0
0
To use 2.5.12 (or 2.5.13) in finite dimensions, we must have the rank of Df equal to the dimension of its image space (or the domain space). However, we can also use the inverse function theorem to tell us that if Df(x) has constant rank k in a neighborhood of x(). then we can straighten out the domain of f with some invertible function", such that f 0 ' " depends only on k vanables. Then we can apply proposition 2.5.12. This is the essence of the following theorem. Roughly speaking, the theorem says that if DI has rank k on R "', then m - k variables are redundant and can be eliminated. As a trivial example, if I: R2 --+ R' is defined by I(x, Y)- x - Y. DI has rank I, and so we can express I using just one variable. namely, let !/-(x. Y) = (x + Y, y) so that (f o!/-)( x. y) ... x. which depends only on x. Let f: U c E -+ F be of class C', r ~ I, U o E U and suppose Df( u o) has closed split image F, with closed complement F2 and split kernel E2 with closed complement E,. In addition. Q.fsume thatlor'all u in a neighborhood 01 U o E U. Df(u)(E) is a closed subspace of F and DI(u)IE,: E,-+Df(u)(E) is a Banach spare isomorphism. (If E=R":' and F=RIf,
2.5.15 Rank Theorem.
114
BANACH SPACES AND DIFFERENTIAL CALCULUS
assume only that rank (D/(u» == k lor u in a neighborhood of uo.) Then there exist open sets VI C Fl eE2,v2 C E, VI C F. V2 C F and C' dilleomorphisms cP: VI - V2 and 1/1: VI - V2 such that (cp ° I ° 1/1 )(x,e) = (x.O).
The intuition is given by Fig. 2.5.3 for E = H2. F = H2 and k == 1.
£,
~~~~' f =
constant
---_. £2
range of /
.
~
v,
£2
F,
Figure 2.5.3
Rmlilric. It is clear that the theorem implies Eleker(O/(,,» = E and Ol( u)( E)eF2 - F for u in a neighborhood of U o in V. because cp 0 I ° 1/1 has these properties. These seemingly stronger conditions can in fact be shown directly to be equivalent to the hypotheses in the theorem by the use of the openness of 01..(E. E) in L(E. E).
By the local representation theorem there exists a C' diffeomorphism 1/1: V.cF.eE 2 -V2 cE such that j(x.y)=(/ol/l)(x.y)= (x.1J(x.y». Let P,:I F-F. be the projection. Since OJ(x.y)·(w.e)= _ ( w. 01J( x. y ).( w. e». it follows th~t (P" ° OJ( x. y»( w. e) == ( w.O). for we F•• ~E E2 • In particular P" ° DJ(x. y)IF. X{O}-l". which shows ~hat Proof.
· THE INVERSE AND IMPLICIT FUNCTION THEOREMS
115
Dj(x. y)lFj X {O}: FI X{O} -+ Dj(x, y)(F I EBF2 ) is injective. In finite dimensions this implies that it is an isomorphism, since dim FI = dim(Dj(x, y) (FI ED E 2 In infinite dimensions this is our hypothesis. Thus D/(x, y)o PF ID/(x, y)(FI eE2 ) = identity. Let (w, D.,,(x, y)(w, E Dj(x, y)(FI EB'E2 ). Since
».
e»
(Di(x. y) 0 PF ,)( w, D.,,(x, y). (w,
e»
= Di(x. y ).( w,O) = (w. D.,,(x. y ).( w.O» = (w, D I
,< x. y)·w),
we must have D21/(x. y)·e = 0 for all e E E 2 : i.e .. D21/(x. y) = O. But D 2 j(x.y)·e=(O.D21/(x,y)·e). says that D 2 i(x.y)=0; i.e.• i does not depend on y E E 2 • Define i(x) = j(x. y) = U 0 x. y),' so j: PF,(V) C FI -+ F where Pi,: FIEBE2 -+ FI is projection. Now f satisfies the conditions of 2.5.12 at Pi ('" I( Un» and hence there exists a C' diffeomorphism' ep: , VI -+ V2 , where VI' V2 c F. such that (ep 0 /)(z) = (z,O): i.e.• (ep 0 f 0 ' " )(x, y)
+)(
-
= (x.O).
•
2.5.16 Example (Functional Dependence). Let C' c R" he an open set and let the functions fl .... ./,,: U -+ R be smooth. The functions fl' ..../" are said to be functionally dependent at Xu E U if there i~ a neighhorhood. Vof the point UI(xo).... ./,,(xo))ER" and a smooth function F: V-+R such that DF'* 0 on a neighborhood of (fl( x o )•... •/,,( xu». an,d
F( fl ( X ) •... ./" ( x»
=
0
for all X in some neighborhood of xo' (i) Show that if fl .... ,f" are functionally dependent at determinant of Df, denoted
(ii)
XO'
then the
H.
on a neighborhood of Xo. then show that f ..... ./" are functionally dependent, and further, that
for some G.
116
BANACH SPACES AND DIFFERENTIAL CALCULUS
s.t.tio& Let 1- (f, •... •1,,)· (i) We have Fo/-O. so DF(f(x» ° DI(x) = O. Now ifJ/(xo). o. DI(x) would be invertible in a neighborhood of Xo. implying DF( I( x» - o. By the inverse function theorem. this implies DF(y)'" 0 on a whole neighborhood of I(x o)' (0) The conditions of (ii) imply that DI has rank n - I. Hence by the rank theorem. there are mappings fP and '" such that fP 0/ 0",(x, , .. . ,x,,) = (XI'" .• X,,_I'O).
Let F be the last component of fP. Then
F(/I •... .!,,) = 0 Since fP is invertible. DF. O. It .follows from the implicit function theorem that we can locally solve
F(f,.···.I,,)=O for I" -G(f, •...• I,,-I). provided we can show,;\ = saw before.
aF/ay" .0. Now, as we
DF(f(x» ° DI(x) ... 0, or, in components withy -/(x).
(;F ,... ,;F) YI
y"
= (0,0 •...• 0).
If aF/ay" == 0, we would have
- (0,0, ... ,0)
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
117
or
aF ( a"'" 'I
since the
~uare
aF ) -a- (0,0, ... ,0) ' .. -I
matrix is invertible by the assumption that
a(/I,···,/.. _I) a(xl" ... ,X.. _I )
.. 0
This implies D/ - 0, which is not truC!. Hence aF/ aY.. .. 0, and we have the desired result. The reader should note the analogy between linear dependence and functional dependence, where rank or determinant conditions are replaced , by the analogous conditions on the Jacobian matrix. • BOX 2.5C THE INVERSION !MAP
,
Let E and F be isomorphic Banach spaces and consider the inversion map 9: GL(E, F) -+ GL(F, E); 9(.)- .-1. We have shown that 9 is C"" and
+
for" e GL(E, F) and E L(E, F). We shall give below the formula for D"9. The proof is straightforward and done by a simple induction argument that will be left to the reader. Define alt+l:L(F,E) X ••• XL(F,E)-+'LIr(L(E.F);L(F,E»
(~ + 1) tim~s by
where
XI
e L(F, E), ; -I, ... ,k + I and +, E L( E. F). ; -I ..... k. Let
9x· .. x9:GL(E,F)-+GL(F,E)X ... xGL(F,E).
<,"1 + Iltim:'
\
(k + l)times
begivenby (9x .. · x9)(.)_(.-I, ... ,.-I).
j
118
BANACH SPACES AND DIFFERENTIAL CALCULUS
Then
where Synf denotes the symmetrization operator. Explicitly, for cp E GL(E, F), E L(E, F) this formula becomes
+1'" "+k
Dkg(qI)·(+I,···,+.)=(_I)k
1:
qI-lo+o(O°cp-l o ... ocp-lo+O(k)°ql-l,
oes. where Sir. is the group of permutations of {I, .. . ,k} (compare boxes 2.2B and 2.4A).
Exercl. .
2.SA
Let I: R4 -+ R2 be defined by I(x, y, u, v)- (u 3 + vx + y. uy + v 3 x). At what points can we solve/(x.y.u.v)=(O.O) for (u.v) in terms of (x. y)? Compute oulox.
2.SB
Let E be a Banach space. Using the inverse function theorem. show that each A in a neighborhood of the identity map ,in GL( E. E) has a unique square root. (ii) Show that for A E L(E. E) the series (i)
1 1 2 B= 1- -(J - A ) - - ( J - A) 2 22 .2!
- ... -
2.SC
(i)
1·3·S···(2n-3) (J-A)"-··· 2"n
is absolutely convergent for 11/- All < 1. Check directly that B2 =A. Let A E L(E. E) and let e A = E::'_oA"/n! Show this series is absolutely convergent and find an estimate for ileA II. A E L(E. E).
(ii) Show that if AB=BA. then eA+B=eAeB=eBeA. Conclude that (e A )- 1 = e- A ; i.e.• e A E GL(E. E). Show that e( .): L(E. E) -+ GL(E. E) is analytic. (iii) (iv) Use the inverse function theorem to conclude that A ..... e A has a unique inverse around the origin.
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
(v)
For B
E
119
L(E. E) define
. 10gB =
E (-1)"-1 (B-I)". n
Show that this series is absolutely convergent if 11/- BII < I. (vi) Show that if AB = BA. then log(AB) = log A + log B. If A E G L( E. E) conclude that log A 1 = - log A: log 1 = o. (vii) Verify that log A and e A are inverses to each other: i.e .. the . inverse in (iv) must be log. 2.5D Show that the implicit function theorem implies the inverse function theorem. (Hint: Apply the implicit function theorem to g: U x F -.. F. g(u. v) = I(u)- v. for f: U E E -.. F.) 2.5E Let f: R.2 - R 2 be Coo .and satisfy the Cauchy- Riemann equations . (see Exercise 2.3F): afl' a12 all ax - ay' ay -
i1 f2
-iix'
Show that Df(x. y) = 0 iff det(Df(x. y» = O. Show that the local inverse (where it exists) also satisfies the Cauchy-Riemann equatio.,s. Give a counterexample for the first statement. if I does not satisfy Cauchy- Riemann. 2.SF Let I: R -.. R be given by I(x) =.'C + x 2 cos( Ij.~) if x'- o. f(O) = O. Show that (i) f is continuous; (ii) f is differentiable at all points; (iii) the derivative is discontinuous at x -= 0; (iv) /,(0)'- 0; (v) fhas no inverse in any neighborhood of x = O. (This shows that in order to get the inverse function theorem. f must be at least C I .) 2.5G It is essential to have Banach spaces in the inverse function theorem . rather than more general spaces such as topoIopcal vector spaces or Fr!:chet spaces. (The following example of the failure of Theorem 2.5.2 in Fr!:chet spaces was kindly pointed out by M. McCracken.) Let X (.:1) denote the set of all analytic functions on the open unit disk in C, with the topology of uniform convergence on compaCt subsets. Let F: X (.:1) - X(.:1) be defined by . 00
GO
Ea"z" .... Ea!z". o
0
120
BANACH SPACES AND DIFFERENTIAL CALCULUS
Show that F is COlO and
If a o = I and a" = lin, n ~ I, then DF([.j'z"ln) is a bounded, linear isomorphism. However, since Z2
Zk - I
Zk
Zk + I
F ( z+-+ .. · + - - - - + - - + ... 2 k-I k k+1
)
~F
(
Z") EI n QQ
conclude that F is not locally I-I. (Consult Schwartz (1967], Sternberg (1969] and Hamilton (1982) for more sophisticated versions of the inverse function theorem valid in Fr6chet spaces.) .
2.SH (Generalized Lagrange Multiplier Theorem; Luenberger (1969». Let I: UcE ..... R and g: UcE ..... G be C I and suppose Dg(uo) is surjective. Suppose 1 has a local extremum (maximum or minimum) subject to the constraint g(u) - O. Then prove (i) DI(uo)'" - 0 for all" e ker Dg(u o ) and (ti) there is a ~ e G· such that
(See Box 3.58 for the geometry behin~ this result.) Show that if a C I map I: U c R m .... R" is injective (surjective). then n(m ~ n). (Hint: Use the rank theorem.) 2.5J (Uniform Contraction Principle; Hale (1969) and Chow and Hale (1982).) (i) Let T: c1(U)x V .... E be a C k map, where U c E and V c F are open sets. Suppose that for fixed ye V, T(x, y) is a contraction in x, uniformly in y. If g(y) denotes the unique fixed point of T(x, y), show that g is C k • (Hint. Proceed directly as in the proof of the inverse mapping theorem.) (ti) Use (i) to prove the inverse mapping theorem. 2.SK (Global Inverse Function Theorem; see Wu and Desoer (1972) for useful references to theory and applications). ! el I:R" .... R" be C", k ~ I and satisfy:
2.51
m"
THE INVERSE AND IMPLICIT FUNCTION THEOREMS
121
(i) Show that / is a global diffeomorphism. (Hint: Show that / is onto by showing its range is both oPen and
cloSed. The .. hard part" is to show / is one to one: do this by . assuming/(x,) - /(x 2 ) = 0, joining x, and X 2 by a straight line aCt) and letting p ... / 0 a. Shrink P continuously to 0 and define, using local invertibility of /, a corresponding deformation of a.) (il) What goes wrong with the examples (x, y) .... (eX, ye-X)
and
(x, y) .... (e'sin y, e'cosy)?
(iii) Show that properness of / can be replaced hv a uniform hound on [D/(nW '. (iv) Show that (i) and (iii) work in Banach spaces as well.
CHAPTER
3
Manifolds and Vector Bundles
We are now ready to study manifolds and the differential calculus of maps between manifolds. Manifolds are an abstraction of the idea of a surface in Euclidean space. This abstraction has proved useful since many sets that are smooth in some sense are not given to us as subsets of Euclidean space. The abstraction strips away any containing space and proceeds with all constructions intrinsic to the manifold itself. This point of view is well worth the geometric insight it provides. 3.1
MANIFOLDS
The basic idea of a manifold is to introduce a local object that will support differentiation processes and then to patch these local objects together smoothly. Before giving the formal definitions it is good to have in mind an example. In R""· consider the n-sphere S"; that is. all x E R"+· such that IIxll = I
MANIFOLDS
123
Figure 3.1.1
Each takes the overlap of the domains to an open subset of R". If we go from R" to the sphere by one of them. then back to R" by the otber, we get a smooth map from R" to R". In this way we can assign coordinate systems to S". Note, however, tbat no single homeomorphism can be used between S" and R", \lut we can cover S" using two of them. We demand that these be compatible; that is, in a region covered by both coordinate systems we must be able to cbange coordinates smoothly. For some studies of the sphere. two coordinate systems will not ~uffice. We thus allow all other coordinate systems compatible with these. For example on'S2 we want to allow spherical coordinates (8.,,) as they are convenient for many computations. Let us begin by defining the concept of an alias. 3.1.1 Definition. Let S be a set. A cIIart on S is a bijection "from a subset U of S to an open subset of a Banach space. We sometimes denote" by (U.,,), to indicate the domain U of ". An atltu on S is a family of charts (of = {( If;. )Ii E /) such that
"I
(MAl) S =:= U{lf;li E /}; (MA2) Any two charts in (£ are compatible in the sen.~e that the overlap maps between members of f1 are COD diffeOmorphisms: for two charts (V;. ,,;) and (~. "1) with V; n ~ .... '" we form the oonIap """,: "jl = 0 ";-II,,/(V; n ~). where fI',-'lfI';(lf; n~) me~ the restriction of,,; I to the set "/(V; n~). We require that "/(V; n~) is open and that "11 be a COD diffeomorphism. (See Fig. 3.1.2.)
"1
124
MANIFOLDS AND VECTOR BUNDLES
FIgure 3.1.2
3.1.2 Ex.mp.... Any Banach space F admits an atlas formed by the single chart (F. identity). A less trivial example is the atlas formed by the two charts of S" discussed previously. More explicitly. if N - (1.0•...• 0) and S - (-I •...• 0.0) are the north and south poles of S". the stereographic projections from N and S are
CPI: S"\{ N) .... RIO. CPI(X l •...• x"+ I) = (x 2/(I- Xl ) ••. •• x"+ 1/(1- Xl»). CP2: S"\{S} -+ RIO, CP2(XI ••.•• x"+ I) = (x 2/(1 + Xl )•..•• x"+ 1/(1 + xl») and the overlap map CP2 0 'PI I: R"\{O} -+ R"\{O} is given by ('P2 0 CPI I)(Z) == z/lIzII2.z E R"\{O}. which is clearly a COO diffeomorphism of R"\{O) to itself. 6. Now we are ready for the definition of a manifold. 3.1.3 DeftnlUon. Two atlases & I and (1' 2 are ~quivllintl if & I U & 2 is an Gllas. A difJnmtillbk slrIIctun ~ on S is an equivalence class of aliases on S. 17re union of th~ atlases in~. (:I'~ = U{&lff E~} is the maximal alias of 1. and a chart (U. cp) E ~ is an tldmissibl~ local chart. If ('I' i.~ at! atlas on S, then the union of all atlases equivalent to & is called the differentiable structure rawrated by &. A dill~lfttillbk lfIIIIIi/oId M is a pair (S. ~). where S is a set and ~ is a diff~rentiable structure on S. We shall often identify M with the und.er~ving set
MANIFOLDS
125
S for notational COmJenience. If a covering by charts takes their values in a Banach space E. then E ;s called the model s",," and we say that M is a IJtINM:j lIflIIIi/oItllllOlU/4'd Oft E. t If we make a choice of an atlas (:l on S then we obtain a maximal atlas just by including all charts whose overlap maps With those in (:l are Cf1O. In practice then it is only necessary to specify a particular atlas on S in order to detennine a manifold structure for S.
3.1.4 Example. An alternative atlas for SrI is given by the following 2(n
+ I) charts (U/.1{Il).i==I ..... n+l: U/={xES"I±x'>O} .
."i,.±( X I•••• ,x "+1) -_ (I x , ... ,x ;-1 ,.x ;+1 , ... ,x "+1)., 1{1; projects the hemisphere containing the pole (0 ..... ± 1..... 0) onto the open unit ball in the tangent space to the sphere at that pole. It can be easily verified that this atlas and the previous one in 3.1.2 with two charts are equivalent. The overlap maps of this atlas are
(1{Il D1{I;±fl(yl ..... y") _ (y' ..... yi-I ±b-(y,)2- ... _(yi_I)2_(yl+I)2_ ... _(y")2.yi+I ..... y ").
Next we show how any manifold is a topological space.
3.1.5 Deftnltlon. Let M be a differentiable manifold. A subset A C M is ""." ffor efJch a E A there is an admissible local chart (U. Ip) such that a E U andOcA. . 3.1.8 propolltlon. The open sets in M define a topology. Proof. Take as basis of the topology the family of finite intersections of chart domains. • tThis defines a Coo manifold. A C' manifold is defined similarly. We have not burdened the text by C"s since the C" case is tbe one th;tf occurs most commonly.
126
MANIFOLDS AND VECTOR BUNDLES
3.1.7 DefInition. A differentiable manifold M is an ,,-manifold when every chart has values in an n-dimensional vector space. Thus for every point a E M there exists an admissible local chart (U, cp) with a E U and cp( U) c R". We write n - dim M. An n-man;fold will mean a Hausdorff, differentiOble n-manifold in this book. A differentiable manifold is a !inite-dimnuional lIIIIIIifolti if its connected components are all n-manifolds (n can vary with the component). A differentiable manifold is called a Hilbert manifoltl if the model space is a Hilbert space. t No assumption on the connectedness of a manifold has been made. In fact. in many applications the manifolds are disconnected (see Exercise 3.1 C). But since manifolds are locally arewise connected. their components are both open and closed.
3.1.8 Exampl..
.
A. given by B. given by
Every discrete topological space S is a O-manifold. the charts being ({ s}. cp,). CP,: $'"-+ O. S E S. Every Banach space is a manifold its differentiable structure being the atlas with the single identity chart. C. S" with a maximal atlas generated by the atlas with two charts described previously makes S" into a n-manifold. The resulting topology is the same as that induced on S" as a subset of R,,+I. D. A set can have more than one differentiable structure. For example. R regarded as a set has the following incompatible charts
They are not compatible since CP2 0 CPII is not differentiable at the origin. Nevertheless. these two resulting structures tum out to be diffeomorphic (Exercise 3.2H). but two structures can be essentially different on more complicated sets (e.g.• S7). E. Essentially the only one-dimensional connected manifolds are R and SI. This means that all others are diffeomorphic to R or SI (diffeomor-
*
t One can form a manifold modeled on any linear space in which one has a theory of differential calculus. For example mathematicians often speak of a "Frechet manifold." a "LCTVS manifold,".etc. We have chosen to stick with Banach manifolds here to avail ourselves of the inverse function theorem. (See Exercise 2.5G). *nat S1 has two nondifCeomorphic differentiable structures is a famous result of J. Milnor (1956). Recently similar phenomena have been found on A 4.
MANIFOLDS
127
.Flgure 3.1.3
s'2 + "handles"
Figure 3.1.4
phic will be precisely defined later). For example. the circle with a knot is diffeomorphic to st. (See Fig. 3.1.3). See Milnor (1965) for proofs. F. A general two-dimensional connected manifold is the sphere with "handles" (see Fig. 3.1.4).t This includes. for example. the torus, whose precise definition will be given in the next section. G. (Grassmann Manifolds). Let E be a Banach space and Consider the set G(E) of all split subspaces of E. For FE G(E). let G denote o~e of its complements. i.e.• E = FeG, and define UG = {H E G(E)IE = HeG}. and 'PF.G: U(;...co L(F. G) by 'PF.dH) = 'If(H.G)o 'If(H. F)
where 'ifF: E
...co
-I.
F. 'lfG: E...co G denote the canonical projections and 'If(H. G)
tTbe classification of two-manifolds is described in Mas.<;ey (1967) and Hirscb (1976).
128
MANIFOLDS AND VECTOR BUNDLES
- tl'GIH. tI'(H. F) - tI',IH. The deflhit,ion of "'.G makes sense and "F.G is a bijective mapping; indeed. it is easy see that H e UG iff "(H. G): H .... F is an isomorphism (use Banach's i~orphism theorem). In fact "F. G maps H e UG to the continuous linear rna F - G whose graph in F X Gis H. We shall now show that the charts {( • ",. G)} form an atlas for G ( E). Let (UG''',.G) and (UG'''F'.G') be two harts at the points F and F'e G(E) such that UGnUG, .l2J. Suppose He UG nUG, so E = HeG oz HeG'. Thus G and G' and hence F and F' ar~ isomorphic, Isomorphisms between them define an isomorphism T: E .... E such that T(F) ... F' and T(G) ... G'. It is easy to check that
'0
""G· "",'G,(a)'" a· T + (T - I), which is a C"" diffeomorphism of L(F.G') onto L(F.G). Thus G(E) is a C"" Banach manifold locally modeled on L(F,G). Let G,,(E) (reap. G"(E» denote the space of n-dimensional (codimension n) subspaces of E. From the preceding proof we see that G,,(E) and G"(E) are connected components of G(E) and so are also manifolds. The classical Grassmann manifolds are G,,(R"'), where m .. n (n-planes in m space). They are connected n(m - n) manifolds. Furthermore. G,,(R"') is compaCI. To see this. consider the set F..,,,, of orthogonal sets of n unit vectors in R"'. Now F...", is dosed and bounded in R"'·x ... xR'" (n times), so F..,,,, is compact, Also G,,(R"') is the continuous image of F..,,,, under the map (., ......,,} .... the span of ." .....". H. (Projective spaces). Let RP" ... G,(R"+ ') - 'the set of lines in R ". '. Thus from the previous example RP" is a compact connected real n-manifold. Similarly CP" is a compact connected complex n-manifold. There is a projection ,,: SIt - RP" defined by ,,(x) - span {x}. Moreover." is a diffeomorphism restricted to any open hemisphere. Thus any chart for SIt produces one for RP" as well. •
E.erc.... Let S-{(x,y)eR 2 Ixy"'0}. Construct two "charts" by mapping each axis to the real line by (x.O) .... x and (0, y) .... y. What fails in the definition of a manifold? 3.IB Let S ... ]O, I[X]O.I[ C R2 and for each s. 0 < s < I let 'Y..- (s}X]O.I[ and ".: 'Y. .... R,(S.I) .... ,. Does this make S into a one-manifold? , 3.1C Let S'" {(x. y) e R 1 lx 2 - y2 -= I}. Show that the two charts" ±: {(x, y) e SI ± x > O} - R." ±(x, y) = y define a manifold structure on the disconnected set S. 3.1D (Flag manifolds), Let F" denote the set of sequences of nested linear subspaces V, c V2 c··· c V,,_, in R" (or CIt), where dimy; - i. 3.IA
SUSMANIFOLOS, PRODUCTS, AND MAPPINGS
129
Show that F" is a compact manifold and compute its dimension. (Rag manifolds are typified by· F" and come up in the study of symplectic geometry and representations of Lie groups.)
3.2 SUBMANIFOLDS, PRODUCTS, AND MAPPINGS
The idea of a submanifold is analogous to that of a subspace in linear algebra. This is a basic concept used throughout the book. Equally basic is the product of two manifolds, producing a new manifold. Finally, we . introduce the notion of a C' map between manifolds. We are not yet ready to differentiate these mappings; this will be possible after we introduce the tangent bundle in Section 3.3. Let us begin now with suhmanifolds. If M is a manifold and A c M is an open subset of M. the differentiahle structure of M naturally induces one on A. We call A an (lpen submanifold of M. For example G,,(E),G"(E) are open submanifolds of G(E) (see Example 3.1.80).. We would also like to say that S" is a submanifold of R"· " although it is a closed subset. To motivate the general definition we notice that there are charts in R"+ I in which a neighborhood of S" ·becomes part of the subspace R". Figure 3.2.1 illustrates this for n = I. 3.2.1 Deftnltion. A sulnruuti/old of a manifold M is a subset B c M with the property that for each b E B there is an admissible chart (U, cp) in M with bEl; which has the SllbnttlniJoId propnty. namely,
(SM) cp: U .... ExF,andcp(UnB)-cp(U)n(Ex{O}).
figure 3.2.1
130
MANIFOLDS AND VECTOR BUNDLES
An open subset of M is a submanifold in this sense. Here we merely take F = {O}, and use any chart. 3.2.2 ProposHlon. Let B be a submanifold of a manifold M. Then B itself is a manifold with differentiable structure generated by the atlas {( U n B, cP I U n B)
I (U, cp) is an admissible chart in M having property (SM) for B}.
Furthermore the topology on B is the relative topology.
If lft n ~ n B .. f2J. and (lft. CPi) and (~. cp) both have the submanifold property. and if we write CPi = (a,.lli) and cpj=(aj.llj ), where ai: lft-E. ai ~-E.lli: U,-Fand Ilj: ~-F, then a,lu,nB: u,nBcpi(u,)n(Ex{O}) and ajl~nB: ~nB-cpj(U\)n(EX{O}) are bijective. The overlap map (cpjl~ n B)o(CPilU, n Br is given by (e,O) ...... «ai 0 a i- I )(e).O) ... cpji(e,O) and is CO". being the restriction of a CO map. The last statement is a direct consequence of the definition of relative topology and 3.1.2. • Proof.
If M is an n-manifold and B a submanifold of M, the codimension of B in M is by definition codim B = dim M -dim B. Note that open submanifolds are characterized by having codimension zero. In Section 3.5 powerful methods are developed for proving that various subsets are actually submanifolds. based on the implicit function theorem. For now we do a case "by hand."
3.2.3 Example. This example shows that S" eRn + I is a submanifold. To do this. it is enough to observe that the charts in the atlas {(U, t. I/J/)}. i - 1..... n + 1 of S" have the suhmanifold property (see 3.1.4): the 2( n + I) maps x,+: R"+I-+R"·'.X,'(XI, ... ,X"·I)=(xl •...• X,-I.X"' .... ,x"+I, ± 11~11~112) are Coo diffeomorphisms. and hence are charts in an atlas of R"+ I. Since 0<,t I U; ±)(Xl .... • x"+ I) = (XI ..... Xi-l. xi+ I, .. . ,x"+ 1.0). they have the submanifold property for S". • Remark. Any n-manifold can be realized (embedded) as a closed submanifold of R 2 H'. For Whitney's proof. see Hirsch (1976). Next we turn to products. The idea is analogous to the construction of a product vector space E X F from the two constituent spaces E and F. 3.2.4 Definition. Let (S., ~.) and (S2' ~2) be two manifolds. The protblct lItIIIIifoitl (S. X S2'~' X ~2) consists of the set S. X S2 together with the dif-
SUSMANIFOLDS, PRODUCTS, AND MAPPINGS
131
lerentiable structure g, X g2 generated by the atlas {(V, X V2 • 'P, X ~2 )I(lf;. 'Pi) is a chart 01 (S;. gl)}'
That the set in 3.2.4 is an atlas follows from the fact that if 1/1,: V, C H, .... V, C F,; 1/12: V2 C H2 .... V2 C F2 • then 1/1, x 1/12 is a diffeomorphism iff 1/1, and '"",2 are. and in this case (I/I,XI/I2)-'=I/I,IXl/li'. Recall that D(I/I, X 1/12)= DI/I, X D1/I2' It is also clear that the topology on the product manifold is the product topology. Also. if SI' S2 are finite dimensional, the dimension of S, X S2 equals the dimension of S, plus the dimension of S2'
Inductively one defines the product of a finite num~er of manifolds. A simple example of a product manifolti is the n-torus T" = S' X ... X SI (n times). The final topic of this section is that of maps hetween manifolds. The following definition introduces two important ideas: the local representative of a map and what it means for a map to be cr.
3.2.5 DeftnlUon. Suppose I: M .... N. where M and N are manilolds (that is, I maps the underlying set 01 M into that 01 N). We say I is of dtw C r illor each x in M and admissible chart (V. 1/1) 01 N with I (x) E V, there is a chart (V, 'P) 01 M with x E V and I( V) c V and the IDetll re".."tlltiw 01 I. I"", - '" 0 I 0 ~ - ' . is 01 class C' (see Fig. 3.2.2). M
N
f
v
FIgu.. 3.2.2
183
MANIFOLDS AND VECTOR BUNDLES
For r - 0, this is consis~nt with the definition of continuity of I, regarded as a map between topoLogical spaces (with the manifold topologies). If I is continuous, the requirement I(U) c V can always be satisfied, with a suitable choice of U. The followina is a useful criterion for a map to be C'.
3.2.8 ~1tIon. Let I: M -- N be a continuous map 01 manilolds. Then is C' iff the local representatives 01 I relative to a collection 01 charts which cover M and N are cr.
I
Proof. Assume that the local representatives of I relative to a collection of charts covering M and N are c. If (U, 'P) and (U, 'P') are charts in M and (V.",).(V.",/) are charts in N such that I,,~ is then the composite mapping theorem and condition (MA2) of 3.1.1 show that I".~. == ( ",' 0 '" - I) 0 I •• 0 ('P' 0 'P - I) is also Moreover. if q>" and "''' are restrictions of q> and '" to open subsets of U and V. then I" ..~ .. is also C. Finally. note that if I is C on open submanifolds of M. then it is C on their union. That I is C' now follows from the fact that any chart of M can be obtained from the given collection by change of diffeomorphism. restrictions. and/or unions of domains. all three operations preserving the C' character of I. The converse is a direct consequence of the definition. •
cr.
cr.
Any C' map from (open subsets of) E to F in the usual sense is C in the sense of Definition 3.2.5. Other examples of coo maps are the antipodal map x .... - x of Sit and the translation map on T" by (8 1......8,,) given by (e l "
..... e l '.) ....
(e l ("
... ,,) •• ••• e ll
'.+'.».
3.2.7 Propoeltlon. II I: M -- Nand g: N -- Pare C maps. then so is
go f. Proof. This follows directly from the previous proposition and the composite mapping theorem. •
3.2.8 DefInition. A map I: M -- N. where M and N are manilolds. is called a C dif/eomorpIIum ill is 01 class C. is a bijection. and I: N -- M is 01 class C'. II a diffeomorphism exists between two manilolds, they are called dif/eomorpIIk.
r
I t follows from 3.2.7 that the set Diff'( M) of C diffeomorphisms of M to M forms a group under composition. This very large and complex group will be encountered again several times in the book.
VECTOR BUNDLES
133
Ex...... 3.2A Sbow that (i) if (U. If) is a chart of M and I/!: 'J'( U) .... Y c F is a diffeomorphism. then (U. I/! 091) is an admissible chart of M and (ii) admissible local charts are diffeomorphisms. 3.2B A C' diffeomorphism that is also a C' map is a C diffeomorphism. (Hillt: Use the comments after the proof of 2.5.2.) 3.2C Show that if N; c M, are submanifolds. i = 1•...• 11 then NI X •.• x N" is a submanifold of MI X ••• x M". 3.20 Show tbat every submaoifold N is locally closed in M; i.e.• every point n e N has a neighborhood U in M such that N () U is closed in U.
3.2E
Show that /;: Mi .... N;. i == I •...• n are all C' iff II X ... X fit: MI X .... X Mit .... NI X ... X Nit is C'.
Let M be a set and {Mi},EI a covering of M. each Mi being a manifold. Assume that for every pair of indices (i. i). MI ()