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 () M j is an submanifold in both M j and ~. Show that there exists a unique manifold structure on M for which the M; are open submanifolds. The differentiable structure on M is said to he obtained by the collation of the differentia~le structures of MI' 3.20 Show that the map F ..... FO - {u E r l ulF= O} of G(E) into G( E*) is a COO map. If E = E·· (i.e.• E is reflexive) it restricts to a Coo diffeomorphism of G"(E) onto G,,(E) for all n == 1.2'" . 3.2H Show that the two differentiable structures of R defined in 3.1.80 are diffeomorphic. (Hint: Consider the map .lC ... XII'). 3.2F
open
3.3 VECTOR BUNDLES Roughly speaking, a vector bundle is a manifold with a vector space attached to each point. During the f!irmal definitions we may keep in mind the example of the n-spheres S" c R' . I. The collection of tangent planes to S" (regarded as vector spaces) form a vector bundle. Similarly. the collection of normal lines to Sit form a vector bundle. The definitions will follow the pattern of those for a manifold. Namely. we obtain a vector bundle by smoothly patching together local'vector bundles. The following terminology for vector space products will be usefut
3.3.1 DefInition. Let E and F be vector spaces with U an opell subset of E. We call the Cartesian product U X Fa IoeIII f'«tIIr""". We call U the ".. ¥'Gu. which can be identified with U X{O}. the zem ,ect;"'. For u E U. {u}X F is called the liber over u. which we can endow with the vector space
134
MANIFOLDS AND VECTOR BUNDLES
structUl'e of F. The map .: U X F - U given by .( u, /) = u is called the 0/ U X F, [Thus, for u E U, the fiber over u is .-'(u). Also note that U X F is an open subset of Ex F and so is a local manifold.)
projectiolr
Next, we introduce the idea of a local vector bundle map. The main idea is that such a map must map a fiber linearly to a fiber.
3.3.2 Definition. Let U X F and U' X F' be local vector bundles. A map 'P: U X F - U' X F' is called a C' Ioclll veclor btuuIk map if it has the form 'P(U, f)- ('P.(u).IP2(u)·f) where 'P.: U -U' and'P2: U -> L(F, F') are C. A loeal vector bundle map that ;s a bijection ;s called a local vector bll1lllk i.roInorpIWm. (See Figure 3.3)
A local vector bundle map 'P: U X F -> U' X F' maps the fiber {u} X F into the fiber {'P.(u)}X F' and so restricted is linear. By Banach's isomorphism theorem it follows that a local vector bundle map 'P is a local vector bundle isomorphism iff 'P2(u) is a Banach space isomorphism for every uEU.
BOX 3.3A SMOOTHNESS OF LOCAL VECTOR BUNDLE MAPS
In some examples, to check whether a map 'P is a local vector ' bundle map. one is faced with the rather unpleasant task of verifying that 'P2: U -> L(F, F') is coo. It would be nice to know that smoothness of 'P as a function of two variables suffices. This is the context of the next proposition. We state the result for Coo, but the proof contains a C' result (with an interesting derivative loss).
3.3.3 Propo.atlon. A map 'P: U X F -> U' X F' ;s a Coo local vector bundle map i/l 'P is Coo and is of the form 'P(U, f) = ('P.(u), 'P2(U)· f), where 'P.: U -> U' and 'P2: U -> L(F. F'). Proof. (M. Craioveanu and T. Ratiu [1976]) The evaluation map ev: L(F, F')X F - F'; ev(T, f) = T( f) is clearly bilinear and continuous. First assume 'P is a C' local vector bundle map. Then 'P2: U -> L ( F, F') is COO. Now write 'P2(U)·/- (ev o ('P2 X I)(u, f). By the composite mapping theorem. it follows that 'P2 is COO as a function of two
variables. Thus'P is Coo by 2.4.12 (iii). Conversely, assume ,,(u. f) ... (".(u), 'P2(U)· f) is Coo. Then again by 2.4.12 (iii). ",(u) and "2(U)'/ are Coo as functions of two variables.
VECTOR BUNDLES
.
135
We have to show that '1'2: U - L(E. F') is COO. For this it suffices to prove the following: if h: U X F - F' is C. r ~ I and such that h( u, . ) e L (F, F') for all u e U. then the map h: U - I. ( F. F'). ii( u) h(u,') is C,-I. This will be shown by induction on ·r. Ir r - I we prove continuity of h in a disk around Uo e U in the following way. By continuity of Dh. there exists f > 0 such that for all u e D.(u o} and .,e D.(O).IID,h(u. "~I ~ N for some N '> O. The mean value inequality yields IIh( u.,,)- h( u',")11 ~ Nllu - 11'11
for. all u. u'e D.(uo) and "e D.(O). Thus IIh(u)-h(u')II= sup 11"11 co I
IIh(u,")-h(u'.")II~ Nllu~u11. f
proving that h is continuous. Let r > I and inductively assume that the statement is true for r -I. Let S: L(F, L(E. F'» - L(E. L(F. F'» be the canonical ispmetry: S(T)(e)'/= T( f)·e. We shall prove that
(I)
Dh=So Dlh,
C:
where D,h(u)'v= Dlh(u.v). Thus. if his Dlh is C- I • by induction Dlh is C- 2. and hence by (I) Dh will be C' 2. This will show that his C'- I. In order that relation (I) make sense we need to first show that Dlh(u.·) e L(F. L(E. F'». Since Dlh(u.o)·",'"" lim [h(u + t"')- h(u)]'v/t- lim A"o. 1-0
"-"00
for all v e F. where
A,,'" n( h( u+
*"'
)-h(u») e L(F. F'),
it follows by the uniform boundedness principle (or rather its coroUary 2.2.21) that D.h(u, )o..,e L(F, F). Thus (0...) .... D1h(u• .,),,,, is Hnear continuous in each argument and hence is bilinear continuous (Exercise 2.21). and consequtatly ., .... Dlh(u,o) e L(E. F) is linear and continuous. Relation (1) is proved in the following way. Fix Uo e U and let t and N be positive constants such that 0
IID,h(u,.,)- Dlh(u',"~I" Nllu - u11
.(2)
136
MANIFOLDS AND VECTOR BUNDLES
for all II, II' E Dz.( uo) and v E D.(O). Fix for the moment u' E Dz,( uo), v E D,(O) and apply the mean value inequality to the C-' map g(u) - h(u, v)- D,h(u', v)·u to get IIh(u+ w,v)-h(u,v)-D,h(u',v)'wll -lIg(u + w)- g(u)1I ~ IIwll sup IIDg(u + Iw)1I IEIO.')
-lIwll sup IID.h(u + tw,v)- D.h(u', v)1I IEIO.')
for w E
D.< uo). Letting u' ..... u and taking into account (2) we get IIh(u + w, v)- h(u, v)- D.h(u, v)·wIl4it Nllwll2;
i.e.,
lIiI(u + w)·v- iI(u)·v- [(S 0 D,h )(II),w](v)1I4it Nllwllz for all v E D,(O), and hence
lIiI(1I + w)-il(II)-(S 0 D,h )(II)'wIl4it N llwll2 e
thus proving (I). • Any linear map A E L(E, F) defines a vector bundle map " .. : Ex E ..... E X F by ,,( II,.) - (u, A.). Another example of a vc;ctor bundle map was encountered in Section 2.4. If I: U c E ..... V c F (s C+', then Tf: U X E ..... Vx Fis C' and is given by TI(u,.) = (f(u), DI(u)·.). Using these local notions, we are now ready to define a vector bundle.
FIguN 3.3.1
VECTOR BUNDLES
137
3.3.4 DefInition. Let S be a set. A local buntIk mart of S is a pair ( w, 'P ) where We Sand 'P: W c S ..... U X F is a bijection onto a local bundle U X F.
(U, F depend on 'P.) A vector blllfllk tltltu on S is a family ~ - {(w,. cPl)} 0/ local bundle charts satisfying: (VBI) .. (MAl of 3.1.1) (~ covers S) and, (VB2) For any two local bundle, charts (w,. 'Pi) and (W,. 'Pj) in ~ with W, n JJj ., 12J, 'PI ( W, n JJ:;) is a local vector bundle. and the overlap map 1/Ij l = 'Pj 0 'PI- I restricted to 'PI ( W, n Hj) is a Coo local vector bundle isomorphism. If ~ I and ~2 are two vector bundle atlases on S. we say they are VB -.",iNlnlt i/ ~ I U '!S2 is a vector bundle atlas: A vector "-dk strwtllr. on S is an equivalence class of vectot bundle atlases. A vector""" E is a pair (S, CV), where S is a set and CV is a vector bundle structure on S. A chart in an atlas ofCV is an admissibk vector""" dtiIrt of E. As with manifolds. we often identify E with the underlying set S.
The intuition behind this definition is depicted in Figure 3.3.2.
FIgu... 3.3.2
138
MANIFOLDS AND VECTOR BUNDLES
As in the case of manifolds. if we make a choice of vector bundle atlas
'!6 on S then we obtain a maximal vector bundle atlas by including all charts whose overlap maps with those in '!6 are COO local vector bundle isomorphisms. Hence a particular vcctor bundle atlas suffices to specify a vcctor bundle structure on S. Vector bundles are special types of manifolds. Indeed (VBI) and (VB2) give (MA I) and (MA2) in particular. so '\I'induces a differentiable structure on S. 3.3.5
Definition.
For a vector hundle E
=
(S. '\I') we define the ~el'O
section ( or lHIse) hy
B = {e E Elthere exists (W. 'P) E '\I'with e = 'P I( u.O)}. Hence B is the union of all the zero sections of the local vector bundles (identifying W with a local vector bundle via 'P: W --+ U X F). If (U. fP) E ~ is a vector bundle chart. and bE U with 'P( b) = (u.O). let E,,;. denote the subset fP - I({ u} X F) of S together with the structure of a vector space induced by the bijection 'P.
The next few propositions derive basic properties of vector bundles that are sometimes included in the definition.
3.3.6 Proposition. (i) If b lies in the domains of two local bundle charts 'PI and 'P2' then E",., = E" .• ,. where the eQ!Jality means equality as topological spaces and as vector spaces.
'
For e E E. there is exactly one b E B such that e E E",." for some (and therefore 0/1 ) (U.'P,). (iii) B is a submanifold of E. (iv) The map 'IT. defined by 'IT: E --+ B. 'IT(e) = b [in (ii)l is surjective and COO.
(ii)
Proof.
(i) Suppose 'P,(b)=(ul.O) and 'P2(b)=(u 2 .O). We may assume that the domains of 'PI and 'P2 are identical. for Eh . q• is unchanged if we restrict'P to any local bundle chart containing h. Then /l = 'PI 0 'P! I is a local vector bundle isomorphism. But we have
Hence E",., = E" .• , as sets. and it is easily seen that addition and scalar multiplication in E",., and E".", are identical as are the topologies. For (ii) note that if eEE. fP,(e)=(ul.J.). fP2(e)=(u 2./2)' b;= 'P.'(u,.O). and b2 'Pi '(U2'0). then a(u2./2) os (u,.I,). so CI gives a linear OK
VECTOR BUNDLES
139
isomorphism {u 2}X F2 -+ {ul}X Fl' and therefore CPI(b2 ) = 0(U2'0) = (ul.O) = CPI(b l ). or b2 = bl' To prove (iii) we must verify that for b E B there is an admissible chart with the submanifold property. For such a manifold chart we choose an admissible vector bundle chart (W. cp). b E W. Then cp( W () B) = U x {O) = cp(U)()(E X {O})
Finally. for (iv). it is enough to check that." is C'" using local bundle charts. But this is clear. for such a representative is of the form.(u,. ( U I' 0). That." is onto is clear. •
n ...
The fihers of a vector hundle inherit an intrinsic vector space structure and a topology independent of the charts. but there is no norm that is chart independent. Putting particular norms on fibers is extra structure to he . considered later in the book. Sometimes the phrase Banachable space is used to indicate that the topology comes from a complete norm but we are not nailing down a particular one. The following summarizes the basic properties of a \ector hundle.
3.3.7 Thebrem.
Let E be a vector bundle. The '[.ero section (or btue) B of -> B called the projection that is of class COO. and is surjective (onto). Moreover,for each h E B• .,,-I(h). called the fiber over b. has a Banachahle vector space structure induced by any admissible vector bundle chart. with b the zero element . E. is a submanifold of E and there is a map .,,: E
.
Because of these properties we sometimes write" the vector hundle .,,: E -+ B" instead of "the vector bundle ( E. 'Y )." Fibers art' often denoted hy E" ,.. ." - I( b). If the base B and map." are understood. we just ~ay .. the vector bun~le E."
3.3.8 Example A. Any manifold M is a vector bundle with zero-dimensional fiber. namely M X {O}. B. The cylinder E = S' xR is a vector hundle with w: E -+ B = Sl just the projection to the first factor (J;ig. 3.3.3). This is a trivial vector bundle in the sense that it is a product. C. The MiJbiur band is a vector hundle .,,: E. -+ .\.1 with one-dimensional fiber obtained in the following way. On the product manifold R ,:
140
MANIFOLDS AND VECTOR BUNDLES ~ Fiiler=R
., - projection
;..
h
"
I- base = Sl
1,11 -I
Flgunt3.3.3
=)O,I[XR,UI""SI\{I} and U2 ==SI\{-I} and note that PIVI: Vl ".-I(UI) and p1V2 : ~_".-I(U2) are homeomorphisms and that E= ".-I(UI)U ".-I(U2). Let {(Ul ' 'PI ),(U2 , 'P2)} be an atlas with two charts (or Sl (see 3.1.2). Define ~J: ".-l(~) - R xR by ~j == Xj o( pi Jj)-l and Xj: JjR xR by X/(u, v) - ('Pj(e 2",U).( -1)'+ IV), j == 1,2 and observe that Xj and thUs ~J are homeomorphisms. Since ~2 0 ~II: (R XR)\{(O}xR) -> (R XR)\ ({O}XR) is given by (~2 0 ~II )(x, y) == 'P2 0 'Pll)(x), - y), we see that {( ".-I(UI)' ~l)' (".-I(U2), ~2)} (orms a vector bundle atlas of E.
«
The MObius band can be visualized in R J as shown in Fig. 3.3.4. D. The Grassman" bundles (universal bundles). We now define vector bundles Y,,(E) - G,,(E), y"(E) - G"(E), and y(E) - G(E) which play an important role in the classification of isomorphism classes of vector bundles (see (or example Hirsch [1916]). The definition of p: y,,(E) - G,,(E) is the (ollowing (see Example 3.I.8G (or notations): y,,(E) == {(F, v)IE is an n-dimensional subspace of E and v E E) and p(E, v) == E. The charts (p-I(UG ), ~FG)' where E == EeG, ~FG(H. v) == ('PFG(H). ".(H. F)(v», and ~FG: p- I(UG ) - L( F, G)X F define a vector hundle structure on y.(E) since the
VECTOR BUNDLES
141
Figure 3.3.4
overlap maps arc
(1/I,.O,o1/liJ)(T. f) - (( cpJ.o' ° CP;J)( T). (tr(graph( T), F') ° tr(graph( T). F) -')( f)}. where T e L(F. G)./ e F and graph (T) denotes the graph of Tin E X F; smoothness in T is shown as in 3.1.80, The fiber dimension of this bundle is n. A similar construction holds for G"~E) yielding y"(E); the fiber cadimcnsion in this case is also n. Similarly y(E)-+G(E) is obtained with not necessarily isomorphic fibers at different points of G(E). • As with mappings between manifolds. we can define vector bundle' mappings.
3.3.' Deftnltlon. Let E and E' be two vector. bundles. A map f: E -+ E' is called a C' WdOr , . . . . . " , . , (~) when for each e E E and each admiuible /oetJI bundle chart (V. 1/1) of E' for which f( e) e V there is an admissible local bundle chart (W,cp) with f(W)C V such that the local representative I•• - 1/1 ° / ° cp-,' is a C' local vector bundle mapping (isotnOl'phism). This definition makes sense only for local vector bundle charts and not for all manifold charts. Also, such a W is not guaranteed by the continuity of f. nor does it imply it. However. if we rlfSt check that f is fiber preserving (whicl't it must be) and is continuous. then such an open set W is guaranteed. This fiber-preserving character is made more explicit in the following.
142
MANIFOLDS AND VECTOR BUNDLES
1.3.10 Propoeltlon. Suppose f: E Then: (i) f preserws the zero section:
f
-+
E' is a C' vector bundle map. r ;;;, O.
() B cB'
(ii) f'nduas a IIniqw mapping f.: B commutes:
-+
B' sllCh that the following diagram
f E
E'
",B
B'
1,,'
f. that is. ,,' of-f. 0". (Here. " and,,' are the projection maps.) (iii) A COIl map g: E -+ E' is a vector bundle map iI/there is a Coo map g.: B -+ B' such thai ,,' ° g = B. ° 'IT and g restricted to each fiber is a linear continuous map into a fiber. It is cu.,(omary to say that g: E -+ E' is a vector bundle map over g.: B -+ B.
I'rtHJf. (i) Suppose be B. We must show f(b) E B'. That is. for a vector bundle chart (V. 1#-) and f( b) E V we must show I#-(f( b» = (v. 0). But we have a chart (W, cp) so be W./(W) c V and cp(b) == (u,O). Hence I#-(f(b» - (I#- 0f 0 cp - 1)( U, 0). But this is of the form ( v. 0) since f.,+ is linear on each fiber. For (ii), let f. "" fiB: B ..... B'. With the notations above,
(I#- 0 'IT' 0f)( e) = (I#- 0 f. 0")( e). Also, if f • ." - (ai' az). then U.).,." - al' so f. is C'. One-half of (iii) is clear from (i) and (ii). For the converse we easily see that in local representation, g has the form
g...,,(I1./)- (I#-0gocp-I)(U,
I). '"' (a l (II).a2(u)J).
which defines GI and G z• Since g is linear on fibers, az( u') is linear. Thus, the local representatives of g with respect to admissible local bundle charts are local bundle mappinp by Proposition 3.3.3. • We also note that the composition of two vector bundle mappings is again a vector bundle mapping.
3.3.11 EumpIee. A. To get an impression of how vector bundle maps work. let us show that the cylinder S I X R and the MObius band E are not vector bundle
VECTOR BUNDLES
143
isomorphic. If cP: E .... S· xR were such an isomorphism. then the image of the curve c: [0.1] .... E. C(/)'" (1.1] by cP would never cross the zero section in S· xR. since [I, I] is never zero in all fibers of E; i.e.. the second component of (cp. c)(t) .... 0 for alii E (0.1]. But c(l) == (1.1] ... (0. - I) = -[0.1] = - ('(0) so that the second component of cp • c at I'" 0 and I " I are in absolute value equal and of opposite sign, which. by the intermediate value theorem. implies that the second component of cp • c vanishes somewhere. B. It is shown in differential topology that for any vector bundle E with an n-dimensional base B ami k-dimensional fiher there exists a vector bundle map cp: E .... B xR'. where p ~ k + n, with CPR = I. and which. when .restricted to each fiber. is injective (Hirsch [1976 I). Write cp( e) (w( e), F( e» so F: E .... R ' is linear on fibers. With the aid of this theorem. analogous in spirit to the Whitney embedding theorem. we can construct a 'vector bundle map 4t: E .... y,,(R') by 4t( e) = (F( E~). F( c» where e. E E". Note that 4t R : B .... G,,(RP) maps bEB to the k-plane F(E,,) in R'. Furthermore. note that E is vector bundle isomorphic to the pull-back bundle 4t;(y,,(R (see Exercise 3.3B for the definition of pun-back bundles). It is easy to check that cjI ...... 4t is a bijection. Mappings f: B .... G,,(R') such that f*(y,,(RP» is isomorphic to E are called classifying map., for E; they playa central role in differential topology since they convert the study of vector bundles to homotopy theory (see Hirsch [1976] and Husemoiler [1978]). •
P»
A second generalization of a local C' mapping. f: U c E .... F. g1obalizes not / but rather its graph mapP.ing A,: U .... U x F: u ...... (u. f(u». 3.3.12 Definition. Lei '11': E .... B be a vector bundle. A C' IoctII Netion of 'II' is a C' map E: U .... E. U open in B. such that for each bE U, weE(b» - b. If U'" B. Eis called a C' gIo/HIl sectioll. or simply a C'sectioll of '11'. Let r'(.".) denole the set 0/ all C' sections of '11'. togelher wilh Ihe obvious real (in/inile-dimensiona/) vector space structure.
The condition on Emerely says that E( b) lies in the fiber over b. The C' sections form a linear function space suitable for global linear analysis. As will be shown in later chapters, this general construction includes spices of vector and tensor fields on manifolds. The space of sections of a vector bundle diff~ from the more general class of global C' maps froin one manifold to' another. which is a nonlinear function space. (See. for example. Eells [1958], Palais [1968], Eliasson (1967]. or Ebin and Marsden [1970) for further details.) Submanifolds were defined in the preceding section. 1bere are two analogues for vector bundles.
144
MANIFOLDS AND VECTOR BUNDLES
3.3.13 DetlnlUon. If fT: E -+ B is a vector bundle and M c B a .submanifold, the IYstrkted InuuJIe fTM: EM -+ M is defined by EM = U me MEm' fTM = "1 EM' The restriction fTM: EM -+ M is a vector bundle whose charts are induced by the charts of E in the following way. Let (V, '" I)' where", I: V -+ V' c E' X {O} be a chart of M induced by the chart (V, cP I) of B with the submanifold property, where (,,-I(V), cp), (with cp(e)" (CPI(fT(e», CP2( e», cP: fT-I(V) - V' X F, and V' c E' X E" = E) is a vector bundle chart of E. Then "': fT,W I(V) - V'x F, "'(e) == ("'I( fT(e»,CP2(e» defines a vector bundle chart of EM' It can be easily verified that the overlap maps satisfy (VB2).
For example the restriction of any vector bundle to a chart domain of the base defined by a vector bundle chart gives a bundle isomorphic to a local vector bundle. 3.3.14 DetlnlUon. Let fT: E -+ B be a vector bundle. A subset FeE is a subblllfllk if for each b E B there exists a vector bundle chart (fT - I ( V ), cp) of E, where be V c Band cP: .".-I(V) -+ V' X F, and a split subspace G of F such that cp(.".-I(V)nF)=V'X(G X (O}). It is straightforward to verify that these induced charts form a vector bundle atlas for "IF: F- B. Note that subbundles have the same base as the original vector bundle. Intuitively, the restriction cuts the base keeping the fibers intact, while a subbundle has the same base but smaller fiber, namely Fb == F nEb' Note that a subbundle F is a closed submanifold of E. For example y,,(R") is a subbundle of both y,,(R"+ I) and Yk+ I(R"+ I), the canonical inclusions being given by (F, x) .... (F X {O}, (x, and (F, x) .... (F xA,(x,O», respectively.
0»
3.3.15 ProposlUon. Let.".: E -+ B be a vector bundle and FeE a subbundle. Consider the following equivalence relation on E: e - e' if there exists b E B such that e, e' E E" and e - e' E F". The quotient set E/ - has q unique (lector bundle structure for which the canonical projection p: E .... E/ is a vector bundle map over the identity. This vector bundle is called the fIIOIintt E/F and has fibers (E/F)" = Flo/Flo' Proof. Since FeE is a subbundle there exists a",' '''r bundle chart cP: ,,-I(V) .... V'x F and split subspaces F I, F2 • FIeF, ,lch that cpl'IT-I(V) nF: ('lTIF)-I(U) ·U'X(FI X{O}) is a vector bUll,;., chart for F. The map " induces a unique map i: E/ - -+ B such lhat ito p'" fT. Similarly cp induces a unique map.p: 'IT-I(V)-V'XF such that .p0p=cp, and.p: .. - '(U) ... V' X ({O}X E2 ) is a homeomorphism. It is straightforward to verify
VECTOR BUNDLES
145
that the overlap map of two such .p is a local vector bundle isomorphism. thus giving a vector bundle structure to E/ -. with fiber E"/F,,. for which p: E -+ E/ - is a vector bundle map. From the definition of.p it follows that this structure is unique if p is to be a vector bundle map over the identity. •
3.3.16 Proposition. Let 'IT: E ..... B. p: F -+ B be veclor bundles lind I: E -+ F a vector bundle map over Ihe identity. Let Ib: E" -+ F" be the restriction of f to Ihe fiber over b E B and define the brMI 0/ f by ker( f) = 1J B ker( f,,) and '{he lYUIfle 0/ f by range (f) = U B range (f" l. (i) ker( f) and range( f) are subbundles of E and F respective~y iff for every b E B Ihere are vector bundle charts ( 'IT •( U ). If!) 0/ E and ( p - •( U). '" ) of F such that the local representative of f has the form
"e
"e
where (ii) If E has finite-dimensional fiher. then Ihis conditioll is equimlenl to: The rank of-the linear map f,,: E" -+ F" is I(}ca/~y constant. Proof. (i) It is enough to prove the result for local vector bundles. But there it is trivial since ker(/",.,,) .. = F. and range(f",.,,) .. = G •. (ii) Fix uEU' and put (/",.,,) ..(F) = G •. Then sinceG. is finite dimensional in G. it splits; let G = G.EDG2 • Let F. == ker(/",.,,) .. ; F. has finite codimension in F and hence is also split; let F= F.EDF2 • Then (/",.,,) .. : F2 ..... G. is an isomorphism, and so by 2.5.4 (/",,,,) ... : F2 ..... G. is an isomorphism for u' in a neighborhood of u. Now use 2.5.14 to get the desired bundle charts. •
3.3.17 'Definition. A sequence /
0/
vector bundles maps over the idenlilY
R
,
E ..... F ..... G is said to be exact (al F) if range( f) = ker (1{). II is said 10 he
split exact i/ ker( /). range( g). and range(/) = ker( g) split in each fiber. II is said to he bundle exact i/ il is split exact and ker( / l. range( g). and range( / ) = ker( g) are subbundles.
3.3.18 Proposition. and
Lei E. F. and G be vector bundles over a manifold B /
g
E ..... F ..... G a split exact sequence of smooth bundle maps. Then the ,\'equence is hundle
146
MANIFOLDS AND VECTOR BUNDLES
uact: i.e .• ker I. range 1= ker g. and range g are subbund/es 01 E. F. and G respeclive~v .
Proof. Let. at a point bE B. A = ker I". B = ker ment for B. Then I has the for (symbolically).
B
~
g". and C be a comple-
C
(: .;);
so I: U x A x B ~ U X B X C (U ~pen Iii B). where x. y. w, z are maps on the indicated spaces (depending on bE B). Now w" is an isomorphism by Banach's isomorphism theorem. Extend I" to the map h B: A X B xC ..... A X BxC;
x w
o We desire maps a, b, c, d, e, /. g such that
(~
0) ('
o
h dOh 0
e
0
g
o
1
which can be done by choosing a = I. c = O. h = I. g = 0, b = - xw './ = - w-'g. d= w-',e - I and noting by exactness thaty = xw-'z. The proposition now follows from 3.3.16. • As a special case note that
is split exact when g" is injective and has split range. Here 0 is the trivial bundle over B with zero-dimensional fiber and the first arrow is injection to the zero section. Similarly. taking G = 0 and g the zero map. I
E .....
F~O
is split exact when I" is surjective with split kernel. In both cases range( g) and ker(f) are subbundles by 3.3.18. In 3.3.18 and these cases in particular. we note that if the sequences are split exact at b. then they are also split exact in 8 neighborhood of b by the openness of GL(E, E) in L(E. E).
VECTOR BUNDLES
147
The analogue for vector bundles for the product of two manifolds is the following. 3.3.11 Deflnnlon. If '11': E -+ Band '11": E' - B' tire 111'0 lvctor I>und/~.• the protbtcl bruIIIII 'II' X '11": EX E' -+ B x B' is defim·tJ 1>,. 11/(' following veclor bundle atlas: (('II'-I(U)X'II"-I(U'). cpXtIt)I('II' 1(U).q'). UcB and ('II" -I(U 1)•. tIt). U'c B' are vector bUlldle charts of J-.' tlml H'. respectivel,,}.
It is straightforward to check that the prodtict atlas vcrifies conditions (VBI) and (VB2) of 3.3.4. This section closes with a general construction. special cases of which are used repeatedly in the rest of the book. It allows the tranSfer of vector space constructions into vector bundle constructions. The abstract procedure will become natural in the context of examples given below in 3.3.22 and later in the book. 3.3.20 Detlnnlon. Let I and J be finite sets and consider a pair of families t;. = (E'>A e IUJ' ~~' = (Ei.)A e IUJ of Banachable spaces. Let L(b~.t.;') =
n L{E,. En x n I.(E;. E,)
,E
I
IEJ
and let
i.e .• A/ E L(Ej • E;'). i E I. and A . E L( E;. E;). j E J. An a.v.vignment 0 taking any family t;; to a Banach space ~~~ and any sequence of Iilll'llr maps ( A A ) to a linear continuous map O(A,,) E L(O&. Ob;') Sllli.vb'illg O( 'f,) = 0« B,,) 0 (A,,» = O( B, ) 0 O( A" ) (composition is takell C(mll'tmenlwi.n·) and i.r such that the induced map 0: L(t;;.t;,') -+ L(Ob).O~~·) i.v C'. will he called a temoritll COIfSt",ctiofl 0/ type ( I. J).
'm.
3.3.21
&=
PropoelUon.
Let 0 be a tensorial constructioll of type (I. J) and Let
(E" h e I U J be a family of vector bundles with the same base B.
Then 0& has a unique vector bundle structure over B with SU;" = Ob;" and '11': 0& -+ B sending 0&" to h E B. whose atlas is given by the ('hllrts ('II' - I( U). tit )). where tit: 'II'-.I(U) -+ U'x O(FA) is defined as follow.f. l.el ('fT,,-I( U).,,"). "A: '11',- I( U)
-+
U' X
F'. cpA ( e") = (CPI( 'II'A (e" »).,,~(eA»)
148
MANIFOLDS AND VECTOR BUNDLES
be vector bundle charts on Ek inducing the same manifold chart on B. Define tf(x)- ('P,(".(x».O(tf ..(X»(x» by tf .. (X) == (tf~(.,» where tf~(X) = ('P~(X»-" for i Eland I/I~(x) == ('P;(X» for j E J.
Proof. We need to show that the overlap maps are local vector bundle isomorphisms. We have
( 1/1' 1/1- ')( u. e) = ( 'PI 'PI')( u) and 0 (( 'P~/ 0
0
0 (
'P~) - '( u) ). e ).
the first component of which is trivially Coo. The second component is also Coo since each ql is a vector bundle chart by the composite mapping theorem. and by the fact that 0 is smooth. •
3.3.22 Exampl.. A. Whimey sum. Choose for the tensorial construction the following: J = (2). 1- {l ..... n}. and 0& is the single 8anach space E, X •.. X E". Let O(A,) == A, X •.• X An' The resulting vector bundle is denoted hy E, EB •.• EBEn and is called the Whitney sum. The fiher over he B is just the sum of the component fibers. B. Vector bIIIuIIn 0/ bruulle maps. Let E,. £2 be two vector hundles. Choose for the tensorial construction the following: I. J are one-point sets I- (J},J =- {2}. O(E,. '£2)'" L(E2 • E,). O(A,. A 2 )·S = A, 0 S 0 A2 for S E L(E"E,). The resulting bundle is denoted by L(E2 .E,). The fiber over b E B consists of the linear maps of (E2 h to ( E, h. C. DIIIII brmdIe. This is a particular case of Example 8 for which E = E2 and E, == B XR. The resulting bundle is denoted E*; the fiber over b E B is the dual space E:. D. V«tor""" 0/ muhilinetlr IIIIIpS. Let Eo, E, .... ,E, be vector bundles over the same base. The space of I-multilinear maps (in each fiber) L(E, ..... E,; Eo) is a vector bundle over B by the choice of the following tensorial construction: 1 = {O}. J = {I •.. .• 1}. O( Eo •... ,E,) = L'(E, •...• E,;' Eo).O(Ao.A, ..... A,)·S=AooSo(A, x .. · XA,) for SEL'(E, ..... E,; Eo)·
.'
Exercises
3.3A 3.38
Find an explicit example of a fiber-preserving diffeomorphism between vector bundles that is not a vector bundle isomorphism. (i) Let".: E .... B be a vector bundle and f: B' .... B a smooth map. Define the pull-back bundle /*".: /* E .... B' by /*E = {( e. b')I"'( e) = f( b')}.
/*".( e, b') =
b'
and show that it is a vector bundle OW! R' Show that h: ' /*E .... E. h( e. b') == e is a vector bundle Ill.':
VECTOR BUNDLES
149
(ii) If g: B" ..... B' show that (f 0 g )0.,,: (f 0 g)O E ..... B" is ismorphic to B". (iii) If p: E' ..... B' is a vector bundle and g: E' -+ E is a vector bundle map inducing the map f: B' ..... B on the zero 'sections, then prove there exists a unique vector bundle map. go: E' ..... E inducing the identity on B' and is such that h 0 gO == g. ' (iv) Let a: F ..... B be a vector bundle and u: F -+ E be a v!'Ctor bundle map inducing the identity on B. Show that there exists Ii unique vector bundle map F ..... E inducing the identity on B' and making the diagram
gOr.,,: gOrE . . .
r
ru: r
rF J F
/"u
u
r
rE { ·E
commutative. (v) If .,,: E ..... B. '1/": E' -+ B are vector bundles and if .:1: B -+ B x B is the diagonal map b ...... (b. b). show that £6)/::' .. .:1*( Ex E'). (vi) Let '1/': E ..... B and .,,': E' -+ B' be vector bundles and denote by P.: B X B' ..... Band P2: B X B' ..... B' the projections. Show that Ex E'i!!! pT(E)$p1(E') and that the following sequences are split exact: . 0 ..... E.,... E$E' ..... E'-+O. 0-+ E'-+ E$E' -+ E -+ O. 0-+ pT(E) -+ E X E' ..... p1(E') ..... 0. 0 ..... p1(E') ..... E X E' ..... pT(E) -+ O. 3.3D
(9 (ii)
Show that G,,(R") is a submanifold of G" + I (R"+ I). Denote by ;: G,,(R")- GH I(R"+ I). ;(F)= F XR the canonical inclusion . ' map. If p: R X G,,(R") .... G,,(R") is the trivial bundle show that ;0("1<+ I(R"+ I» i· ""'nJorphic to ",,(R")$(R X G,,(R"».
.!.
3.3D If the sequence 0 ..... E F!. G - 0 is split exact. show that Fi/! E$G. 3.3E Show that the mapping s: E$E -+ E. see. e') = e + (" (fiberwise addition) is a vector bundle mapping over the identity. 3.3F Write down explicitly the charts in examples 3.3.22 given by Theorem 3.3.21. 3.3G Show that any vector bundle over R is trivial. 3.3H (Jel bundles). Let '1/': E -+ B h:" (" vector bundle. Let cP: '1/'- I(U)-+ U' X F. cp(eb) - (cpl(b). q>,I{IW) = U'c E. be a ·Iocal vector
150
U~NIFJjLDS AND VECTOR BUNDLES
1iundle chart for E. where V is open in B. If ~: V -+ E is a local section whose local representative is u .... (u.l,,(u». we can ~",: V' - F tbe principal part of the local representativ~ of €. (i) Let €. 'IJ be local sections defined on V. The pairs (~. bl)' ('IJ. b2 ), where bl' b2 E V are said to have k 1b order contact. k if b l - b2 and Di€",(b l ) - Di'IJ",(bl) for 0 ~ i "k. Show that this defines an equivalence relation independent of the vector bundle chart. Denote by jlt~(b) the equivalence class of (€, b) and call it the k-jet 01 € at b. Let J"E be the set of all
"r.
j"(€)(b).
(ii) Define JIt'IT: JlcE - B by j"~(b) .... b. If fJi: 'IT -I(V) - V'x F is a vector bundle chart of E, show that the map J"fJi: (J"'IT)-I(V) - V' X ei"_oL~(E, F) given by /'~(b) .... (u, ~",(u). D~",(u), ... ,DIe~",(u». where u=fJil(b). is bijective. Show that the coUection of all such J"fJi, when fJi ranges over the vector bundle atlas of E. forms a cr- It vector bundle atlas of JltE, thereby proving that J"'IT: JleE - B is a C'-" vector bundle over B ca1Jed the k-jet bundle of E. (Hint: Use Box 3.3A). (iii) Show that if~: V - E is a local section of 'IT, its k-jet extension j"~: V - JIcE is a C,-It local section of J"'lTo (iv) For I" k define the restriction map rlc/: JA( E) -+ .I ' ( E) by j*(~)(b)""l~(b). Show that rH is a surjective bundle homomorphism whose kernel Kltl(E) splits in each fiber. Show that Kle/(E) is formed by k-jets of sections of E having contact of order "I with the zero section. 3.4 THE TANGENT BUNDLE Recall that for I: VeE - V c F of class cr+ I we define the tangent of I, TI: TV-TV by TV=VXE, TV=VxF. and TI(u,e)= (f(u), DI(u)·e). Hence, TI is a local vector bundle mapping of class cr. Also T( log) - TI 0 Tg. Moreover, for each open set V in some vector space E, let TU: TV - V be the projection (as usual. identify V with the zero section V X (O». Then the diagram TI TV TU
• TV
1
1
TV
V
•V
I is commutative. that is. I
0
TV -
"'.0 • Tf.
THE TANGENT BUNDLE
151
The operation T can now be extended from this local context to the context of differentiable manifolds and mappings. During the definitions it may be helpful to keep in mind the example of the ramily of tangent spaces of the sphere Sit c R If + I. A major advance in differential geometry occurred when it was realized how to define the tangent space to an abstract manirold independent of any embedding in R It. t Several alternative ways to do this can be used according to taste. (See below and Spivak [1979] for further information.)
Coordinate approach. Using transformation properties of vectors under coordinate changes, one defines a tangent vector at m E M to be an equivalence class of triples'(U,cp.e), where cp: U-+ E is a chart and e E E. with two triples identified if they are related by the tangent of the CQrresponding overlap map evaluated at the point corresponding to . 2
mEM. Derivation approach. This approach characterizes a vector by specifying
a map that gives the derivative of a general function in the direction of that vector. 3 The ideal approach. This is a variation of alternative 2. Here T,.,M is the dual of (~!O)/~!I). where ~.!!) is the ideal of functions on Mvanishing up to order j at m. 4 The curves approach. This is ~he method followed here. We abstract the idea that a tangent vector to a surface is the velocity vector of a curve in the surface.
Deflnltlon. Let M be a manifold and m E M. A CIIIW lit m is a C' map c: I .... M from an open interval I c R into' M with 0 E I and c(O) - m. Let c 1 and C2 be curves at m and (U, cp) an admissible chart willi Iff e U. Then we say C1 and C2 are tMgelft lit m willi rap«t 10 cp if and ortIy if. • c. and cp 0 Ci are tangent at 0 (in the sense of Section 2.3- we may ratrict 1M . domain of C1 such that cp 0 c; is defined - see Fig, 3.4.1). 3.4.1
Thus two curves are tangent with respect to cp if they have identical tangent vectors (same direction and speed) in tbe chart cp.
3.4.2 PropoeItIon. Let c, and C2 be two CII1W$ at Iff eM. SIIpJlOH (U" cp,) are admissible charts wilh Iff e~. II -1,2. Then c, __ C2 1ft tangent at Iff with respect to 'I, if and ortIy 1/ tlrq tI1'e tangat lit Iff mp«t to 'fl.
wi'"
tTbe history is DOt completely dear to us, but this idea seems to be primarily due to Riemann. Weyi. IDd Levi-Cmtl aDd was "well kaown" by 1920.
152
MANIFOLDS AND VECTOR BUNDLES
R
1
o
FIgu.. 3.4.1
Note that C I and Cz are tangent at m with respect to CPI iff c i )(0) - D( CPI 0 Cz )(0). By taking restrictions if necessary we may suppose VI = Vz. Hence we have cpz 0 C; = (CP2 0 cpll )o( CPI 0 c;). From the C l composite mapping theorem it follows that D( CP2 0 C I )(0) = D( CP2 0 C 2 )(0). Proof.
D( 'PI
0
•
This proposition guarantees that the tangency of curves at mE M is a' notion that is independent of the chart used. Thus we say CI' C 2 are tangent at m EM if c l ,C2 are tangent at m with respect to cpo for any local chart cP at m. It is evident that tangency at mE M is an equivalence relation among curves at m. An equivalence class of such curves is denoted [clm. where c is a representative of the class.
3.4.3 Definition. For a manifold M and m E M the 'ange,,' spue to M
til
m is the set of equivalence classes of curves at m
T", (M) ... {[ cJ 1ft I c is a curoe at m} For a subset Ac M. let TMIA= UlftEATm(M). We call TM-TMIM the IIIItgeIfI btmdle of M. The mapping TM : TM ..... M defined by TM([cl m ) = m. is the tangellt '-die proi«t- of M.
Now we will work toward a derivation that T M : TM ..... M is a vector bundle. First, however we need to show that in case VeE is open. TV as defined here can be identified with V x E. This will establish consistency with our use of T in Section 2.3.
THE TANGENT BUNDLE
·153
3.4.4 Lemma.
Let U be an open subset of E. and e a euftll' at u E U. Then there is a unique e E E such that the curve c u defined by Cu., ( t ) = u + te (on . some interval I such that c u.,( I) cUI is tangent to c at u. Proof. By definition. Dc(O) is the unique linear map iii L(R. E) such that the curve g: R - E given by g(t) = u + De(O)·t is tangent to c at t = O. If e = Dc(O)·I. then g = cu.,, •
Define a map i: UXE-T(U) by i(u.e)=\cu.,III; the preceding lemma says i is a bijection. Moreover. we can define a local vector bundle structure on' T( U) by means of i. For example. the fihcr over Il E U is ;«u}X E). Then i becomes a local vectur bundle i'somorphism, It will be convenient to define the tangent of a mapping before showing that T M : TM - M is a vector bundle. The idea is simply that the derivative of a map can be characterized by its effect on tangents to' curves; see . Section 2.4.
3.4.6 Lemma. Suppose CI and c2 are curves at mE M and are tangent at m. Let f: M - N be of class C I . Then focI and f 0 c2 are tangent at f(m)EN.
.
From the C I composite mapping theorem it follows that focI and are of class C I . For tangency. let (V.",) be a chart on N" with fern) E V. We must show that ('" ° f ° cl)'(O) = (1//0 f 0 ( 2 ),(0). But", 0 f ° ea =("'ofocp-l)o(cpoCa ). where (U.CP) is a chart on M with f(U)cV. Hence the result follows from the C I composite mapping theorem. • Proof.
f
O
('2
This justifies the following.
3.4.6 Definition.
Iff: M - N is of class C I • we define Tf: TM - TN by Tf([ c]m) = (f 0 e]'(m)
We call Tf the tagent of f·
TI is well defined. for if we choose any other representative from [elm. say C I' then C and C I are tangent at m and hence 1 0 C and f 0 clare tangent at I(rn). That is. If 0 cl,(m)" If 0 c,l'(m)' By construction. the following diagram commutes. Tf TM .... ·---·TN TM
1
1
TN
M
·N
I The basic properties of T are summarized in the following,
154
MANIFOLDS AND VECTOR BUNDLES
3.4.7 Compoalle Mapping Theorem. (i) SUppOSE' f: M -+ Nand g: N ..... K are C r maps of manifolds. Then go f: M ..... K is of class C' and T(Ko f) .. TgoTf· II h: M ..... M Is the Identity map. then Th: TM ..... TM is the identity map. If f: M -+ N is a diffeomorphism. then Tf: TM ..... TN is a bijection and
(ii)
(iii)
(Tf)-I =
Proof.
and
T(r I).
(i) Let (V.IP). (V. I/! ). (W. p) be charts of M. N. K. with Then we have. for the local representatives.
I( V) c
V
g( V) C W.
(g °
n
'f'P
=
P ° g ° f ° IP - 1
=pogo1/l-'o1/lofolP- ' = gtj.p
° f.,tj..
By the composite mapping theorem in Banach spaces this. and hence go f. is of class cr. Moreover. T( g ° f)[ c) m = [g ° f ° c 1g o/(m)
and
(Tg 0 Tf)[c)m
=
Tg((fo C1/(m)
= [go fo C1go/(m).
Hence T( g f) ". Tg Tf. Part (ii) is an immediate consequence of the definition of T. For (iii). f and r 1 are diffeomorphisms with for I the identity on N. while flo I is the identity on M. But then using (i) and (ii). Tf ° Tf I is the identity on TN while Tj loTI is the identity on TM. Thus (iii) follows. • 0
0
Next let us show that in the case of local manifolds. Tf as defined in Section 2.4, which we temporarily denote /'. coincides with Tf as defined here. 3.4.8 Lemma. Let VeE and V c F be local manifolds (open subsets) and f: V ..... V be of closs C 1• Let i: V X E -+ TV be the map defined following 3.4.4. Then the diagram
/' VxE-VxF
illi
T ( V ) - T(V) Tf
commutes. that is. Tf 0 i = i
0 /'.
THE TANGENT BUNDLE
. Proof.
155
For (u. e) E U x E. we have (Tf 0 i)( u.e) "" Tf· [c II ."] ,,
Also. (i 0 f')(u. e) - i(f(u). Df(u)·e) = [C/(u,.D/(II""]/(II" These will be equal provided the curves I .... feu + Ie) 'and I .... f(U)+/(Df(u)'e) are tangent at I = O. But this is clear from the definition of the derivative and the comPosite mapping theorem. • This lemma states that if we identify U x E and T( U) by means of i then we can identify f' and Tf. Thus we will just write Tf and will suppress the identification. One more fact will be useful:
3.4.8 Lemma. If f: U c E - V c F is a diffeomorphism. Ihen Tf: U
X
E
- V X F is a local veclor bundle isomorphism.
Proof. Since Tf(u.e) = (f(u). Df(u)'e), Tf is a local vector bundle mapping. But 'as f is a diffeomorphism. (Tf)-I .". T(F I) is also a local vector ' bundle mapping. and hence Tf is a vector bundle isomorphism. •
For a chart (U.IJI) on a manifold M. we can construct TIJI: TUThen TIJI is a bijection. since IJI is a dirfeomorphism. Hence. on TM we can regard (TU. TIJI) as a local vector bundle chilrt. In the target of TIJI note that we have a special local vector bundle. where the fibers have the same dimension as the base. T( IJI(
U».
3.4.10 Theorem. LeI M be a manifold and fi an Cltlas 4 admissible charls. Then 1Cl.';" {(TU, TIJI)I(U,IJI)E fi) is a veclor bundle alICl,~ of TM called Ihe .
IIIItlllfll Gt/tu.
Proof. Since the union of chart domains of fi is M. the union of the corresponding TU is TM. Thus we must verify VB2 HenCe, suppose we have T~ n T~ .21. Then ~ n ~ • 21 and the overlap map 91 can be formed by restriction of Cff' to IJIj(~n~). '!'1l- we must verify that TIJII o(TlJlj)-1 - T( IJII 01Jl)-1) is a local vector bundle isomorphism. But this is , guaranteed by 3.4.9. •
·.i'
lienee TM has a natural vector bundle structure induced by the differentiable structure of M. If M is n-dimensional, Hausdorff: and second countable, TM will be 2n-dimensional. HausdQrff, and second countable.
156
MANIFOLDS AND VECTOR BUNDLES
We shall now reconcile the bundle projection TM with that for an arbitrary vector bundle. 3.4.11 Propoaltlon. II mE M, then TAi l(m) = TmM is a liber 01 TM and ils base B is difleomorphic 10 M by Ihe map TMIB: B -. M.
Proof. Let (V,,,) ,,(m) -
II.
be a local chart at m E M. with ,,: V -. ,,( V) C E and Then T,,: TMI V -. ,,(V)X E is a natural chart of TM,
by definition of T". and this is exactly T,"M. For the second assertion. TMIB is obviou!ily a bijection. and its local representative with respect 10 T" and" is the natural identification q>( U) x {O} -. q>( U). a local diffeomorphism. .' Thus M is identified with the zero section of TM and TM with the bundle projection onto the zero section. It is also worth noting that the local representative of TM is (" 0 TM 0 T" - 1)( II, e) = II, i.e. just the projection of cp(V)X E to cp(V). Since tangent bundles are vector bundles, all constructions of the previous sections can be applied to them. In particular, if N is a submanifold of M, the restriction (TM)N of TM to N is customarily denoted by TNM.
Let us next develop some of the simplest properties of tangent maps.. First of all, let us check that tangent maps are vector bundle maps. 3.4.12 PropoaItIon. LeI M and N be manifolds, and leI I: M -. N be 01 class cr+ I. 71ren Tf: TM -. TN is a oector bundle mapping 01 class cr. Proof. It is enough to check that TI is a local vector bundle map using the natural atlas. For m E M choose charts (U• .p) and (V, "') on M and N so m F. "./(m)E V and I• .,. - '" 0 I 0 cp- lis of class cr+ I. Then using (TV, Tcp) fOl I M and (TV, T",) for TN. we must verify that (Tf)T•. T." is a local vector bundle map of class cr. But we have (Tf)T•. T." - T",. TI 0 Tcp-l_ T(/• .,,), and TI• .,,(II,e) = (/• .,,(11), DI• .,.(II)·e), which is a local vector bundle map of class cr.• Now that TM has a manifold structure we can form higher tangents. For mappings I: M -. N of class C. we define T'/: T'M.-. T'N inductively to be the tangent of Tr-l/: T'-IM -. T" IN. Induction readily yields the following: Suppose /: M -. Nand g: N -. K are C mappings of manifolds. Then g • / is of class C' and T'( g • /) = T'g 0 T'I. Let us now apply the tangent construction to the manifold TM, forptting for the moment it' 'cctor bundle structure We get the tangent
THE TANGENT BUNDLE
157
bundle of TM. namely Tn/: T(TM)-'TM. In co(lnlinales. if (U,ep) is a chart in M, then (TU, Tep) is a chart of TM and-eTC TV). T(Tep» is a chart of T(TM) and-thus the local representative of Tnl is (Tep 0 Tnl 0 T(Tep--I»: (u,., "I' "1) .... (u,,,). On the other hand, taking the tangent of the map TM : TM ... M. we get TTM: T(TM) -0 TM. thus getting another vector bundle structure on T(TM) in addition to the usual one discussed before. The local representative of TTM is
Applying the commutative diagram for Tf following 3.4.6 with f = TM• we get what is commonly known as the duo/tangent rhomhi(·.
Let us now tum to a discussion of tangent bundles of product manifolds. Here and in what follows. tangent vectors will he denoted by hold race letters such as I)E T",M.
3.4.13 PrOpoeauon. Let MI and M1 be manifold, and P,: MI X M2 --+ Mi' i -1.2 the two canonical projections. The map (Tpl' Tp2): T(MI X M2)-o TMIXTMz defined by (TPI' TP1)(1) = (Tp,(I),TP2(1)) is a vector bundle isomorPhism of the tangent bundle T(M I X M z ) and the product bundle TM,XTMz'
Proof. The local representative of this map is (u l ' u 2 • "I' e 2 ) E
VI X U2 X EI X Ez .... «u •• ".). (UZ'''2»E(UI X EI)X(U1XE1). which clearly is a local
vector bundle isomorphism "sition follO\llfS •
'rio Tp2) covers the identity. the propo-
Since ~he tangent is just a global version of the derivative. statements concerning partial derivatives might be expected to have analogues on manifolds. To effect these analogies. we globalize the definition of partial derivatives. Let MI' M z and N be manifplds and f: MI X M2 -0 N a C'map. For (p.q)E M, X MI' define ip"' M1 ... MI X M2 and i q : MI -. MI X M2 by
i,(y) = (P. y). iq(x) = (x.q).
158
MANIFOLDS AND VECTOR BUNDLES
With these notations the following proposition giving the behavior of T under products is a straightforward verification using the definition and local differential calculus.
3.4.14 Propoeltlon. LeI MI' M 2• N. P be manifolds and 8/: P - Mi' i 1.2. f: MI x M2 - N be C' maps. r ~ I. Identify T( M, x M 2 ) with TMI X TM2. Then T(gl X 82)= T81 X Tg 2 Tf(lIp. Up) = T.t(lI p )+ Td(vq)./or lip E TpMI and Vq E TqM 2 (iii) (lmplic;t F""ct;on Theorem) If Td( p. q) is an isomorphism then there exist open neighborhoods V of p in MI' W of f( p. q) in N and a unique C r muP g: V X W - M2 such that for all (x. w) E V X W (i)
(ii)
f(x. g(x, w» = w. In addition. Tlg(x,w) = -(Td(x.g(x.w») T2g(X.W)'" (Td(x.g(x,w»)
I
o(T,f(x. g(x.w»
I
Throughout this book the tangent bundle and vector bundle constructions on M play a central role and we shall learn a variety of ways of representing them. For the moment we content ourselves with the direct analysis of tangent bundles for some simple manifolds.
3.4.15 Exampl... A. The langent bundle TS I of the circle. Consider the atlas with the four charts {(~ ± .1/1/ )1; = 1.2} of Sl = {(x. y) E R21x 2 + ).2 = I}. (See 3.1.4.) Let us construct the natural atlas for TS I = «(x. y). (U.lJ»E R2 XR21x 2 + y2 = I. «x. ,v). (u. v» - O}. Since
1/-;: VI' ={(x.y)ES ' lx>O}---+l-I.I(.
~;(x.y)=y.
by definition of the tangent we have
J; •. r)+~ (II. v) - (y. v).
T1/I~: TVt -
]-I.l[ xR.
THE TANGENT BUNDLE
159
Proceed in the same way with the other three charts. Thus, for example, 1(x.YI"'i" (u, v) =- (x, u) and hence for x E ]-1,0[,
This gives a complete description of the tangent bundle. But in this case, considerably more can be said. Namely, thinking of Sl as the multiplicative group of complex numbers with modulus I, we shall show that the group ~ 1 operations are Coo: The inversion $ .... $- has local repr~sentatives ("'I 0 ~ I)(X) = - x and the composition
o("'r ).
has local representatives
(here ± can be taken in any order). Thus for each $ E Sl the map Ls(s')., ss', L,: Sl ..... Sl is a diffeomorphism. This enables us to define a map A: TS 1 ..... Sl xR by A(vs )= (s, T1L.-1(v,» (I is the identity element of Sl), which is easily seen to be a vector bundle isomorphism over the identity. Thus TS 1 is a vector bundle isomorphic to SiX R. See Fig. 3.4.2. B. The tangent bundle TT" to the II-tOrus. Again one could start writing charts, but since T" = Sl X ... X Sl (n times) and TS 1 ~ Sl xR. it follows that TT" a; T" XR". C. The tangent bundle TS 2 to the sphere. The previous examples yielded trivial vector bundles. In general this is not the case. The simplest such example is the tangent bundle to the two-sphere which we now
~c-£ s"
Trivial langenl bundle'
Non trivial lanpnt bundle
•
160
MANIFOLDS AND VECTOR BUNDLES
:describe. Choose again the atlas with six charts {(U/. 1/1/ )1; = 1.2,3) of S2 that were given in 3.1.4. Since == {( X l • X 2 • x) >E
1/1:: ut
.
S 21 Xl> o} -+ D 1(0) =
{( x.
y) E R 21 x 2 + Y 2 < I}.
I/I~(Xl,x2,X3) = (X 2 .X 3 ),
we have
.r, + (1 2 3) _ (2 1 2 3) T.(x'.x'.xl)"1 v.v.v - x ,x ,v.v ,
where XlVI + X 2v 2 + X 3V 3 = O. Similarly construct the other five charts. For example one of the twelve overlap maps is
(TI/I3
o
(TI/I:>-I)(X,)'.U,ll)
- (b - x 2 -
y2 • X. -
ux/';':::~:'-y2 - vY/.fl=~2-':"7 ,u). for
x 2 + y2 < I. Y < O.
One way to see that TS 2 is not trivial is to use the topological fact that any vector field on S2 must vanish somewhere. We shall prove this later; see Theorem 7.6.10 in Chapter 7. •
Exercl... 3.4A 3.48
Let N c M be a submanifold. Show that TN is a subbundle of TNM and thus is a submanifold of TM. Let M. N be manifolds and I: M -+ N. (i) Show that (a> lis Coo iff graph (f) = {(m,f(m»
E
M
X
Him E M}
is a Coo submanifold of M x Nand (b)
7;",./''''11( M
x N) a 7;",./(m))(graph (f»e!J7'(m,N for all
mE
M.
(ii) If I is Coo show that the canonical projection of graph ( .r) onto M is a diffeomorphism. (iii) Show that 7;"'.f,,,,))(graph (f» a graph(T",f)
'"' «"... Tlft/(.,..,» I ~...
E
T... M} C T",M
x T"""N.
SUBMERSIONS, IMMERSIONS, AND TRANSVERSALITY
161
Let p: R X S" -. S" and 0: R"+' X S" -+ S" he trivial vector bundles. Show that TS"e(R x S") ~ (R"+' X S").
3.4C
(Hint: Realize p as the vector bundle whose one-dimensional fiber is
3.4D
the normal to the sphere.) (i) Show that there is a map s",: T(TM) -. T(TM) such that s", 0 s'" = identity and .f",
T(TM).
T(TM)
T1"i"\ s'" ft( TAl) TM
commutes. (Hint:
In local natural coordinate charts.'
s"'(u ••••••• 2 )-(u••••••• 2 )..) One calls s'" the canonical involution on M and says that T( TM) is a symmetric rhombic. (ii) Verify that for f: M -+ N of class C 2• T2/ 0 .~\1 = SN 0 T2f. (iii) If X is a vector field on M. that is. a section of T",: TM -+ M. show TX is a section of TT",: T2M -+ TM and X, = SM 0 TX is a section of TT"': T2M -. TM.
Show that T(M. x M2)~ p1(TM.)ep!(TM2 ) wherep;: M. x M 2 -+ M;. i = 1.2 are the canonical projections and p~( TM;) denotes the . pull-back bundle defined in Exercise 3.38. 3.4F (Requires some basic algebraic topology). In algebraic topology. the simply connected universal covering space Sf of a topological space M is constructed together with a projection 71': Sf -+ M which is a local homeomorphism. Show that if M is a manifold. so is Sf and 71' is . a smooth map which is a local diffeomorphism.
3.4E
3.5 SUBMERSIONS, IMMERSIONS, AND TRANSVERSALITY This set:tion brings the powerful implicit function theorem into play. The notions of submersion. immersion. and transversality are geometric ways of stating various hypotheses needed for the inverse function theorem. These ideas are central to large portio"s of calculus on manifolds (lne immediate benefit is easy proofs that various subsets of manifolds are actually submanifolds. Let us begin with a simple consequence of the inverse function theorem.
3.5.1 Theorem. Let M and N be 11Ulnifolds. f: M -. N be of class C. r ;;, I and m E M. Suppose Tf restricted to the fiber over m E M is an
162
MANIFOLDS AND VECTOR BUNDLES
isomorphism. Then f is a C diffeomorphism from some neighh'!rhood of m onto some neighborhood of f(m). Proof. In local charts. the hypothesis reads: (Df",oJ,Xu) is an isomorphism. where ,,( m) "" u. Then the inverse function theorem guarantees that f",oJ, restricted to a small neighborhOod of u is a C diffeomorphism. Composing with chart maps gives the result. •
The implicit function theorem also yields the "local onto" theorem. Let M and N he maniloldr and f: M ~ N he 01 clau C; I. Suppose TI restricted to the liber TnrM is mrjectil'e to T,'m,N. 711£'''
3.5.2 Theorem. ~
r
f is loral/\' 0"'0 lit 111: i.e. there exist Ilei",hhorhoods U of 1/1 and Vof f( m) .ruch that II V: U ~ V is onto: in particular. if T/ i.l' .I'urjectit'e on each tangent space. the" f i.l' an open mapping: (ii) if in addition the kernel ker( Tmf> is split in Tm M there are charts (V. ,,) and ev.~) with me V.j(V)C V.,,: U"" V'x V'.,,(m) = (O.O).~: V .... V' and f",oJ,: U· x V' .... V· is the projection onto the second factor. (i)
Proof. It suffices to prove the results locally. But these follow directly from 2.5.9 and 2.5.13. •
The notions of submersion and immersion correspond to the local surjectivity and injectivity theorems from Section 2.5. Let us first examine submersions. building on the preceding theorem.
3.5.3 Definition. Suppo.re M and N are manifolds with f: M ~ N of c/a.r.s C. r ~ I. A point n E N is cal/ed a reglllllr vallie of I if for etlch m E f I ({ n}). Tmf is surjective with split kernel. Let R f denote the .ret of regular tlalues 01 f: M .... N; note N \f( M) c R feN. If. for each m in a .ret S. Tmf is .fUl'jective with split kernel. we say f is a slIbmersion on S. (Thus n E R / iff I is a submersion on f- I({n». If T",f = O. me M is called a critical point and n = f( m) E N a critical vallie of I.
Note that statement (ii) in 3.S.2 at every point 111 E M is equivalent to f being a submersion. The next result is important when wnsidering level surfaces of maps.
3.5.4 Theorem. Suppose.f: M
r
-+
N is 01 class COY 11111/ "E R f' Then
l(n)=(mlmE M.j(m)=n} is a .mhmanilold of M and T",f- I (I1)=
kerT",/.
r
Proof. If I( n) = 0 the theorem is satisfied. Otherwise. for m E find charts (V. ,,). (V. ~) as described in 3.5.2.
f
I( n)
we
SUBMERSIONS, IMMERSIONS. AND TRANSVERSAL/TY
163
Then it must be shown that the chart (V,,,) has the submanifold = I;~(O) = V' X {O}. This is exactly the subproperty. But ,,(V () I( manifold property. (See Fig. 3.5.1.) Since I •• : V' X V' ..... V' is the projection onto the second factor, where V' C E and V' C F, we have
r n»
T,,(t;~(O») ... TP'- E .... ker(T"/•• ) for u E V', which is the local version of the second statement. • If N. is finite dimensional and E R /' observe that codim (f' I( dim N, as can he easily seen from the second statement of 3.5.4. This makes sense even if M is infinite dimensional. Sard:s theorem. discussed in Appendix E. implies that R r is dense in N.
n
n» ...
3.5.5 Exampl... A. We shall use the preceding theorem to show that S" eRn + I is a submaniCold. Indeed. let I: R"+I ..... R be defined by 1(.1')-11.1'11 2 • Thus S" ... I( I). To show that S" is a submaniCold. it suffices to show that I is a
r
FlguNa.I.,
164
MANIFOLDS AND VECTOR BUNDLES
regular value of f. Suppose f( x) = I. Then identifying TR" + I = R" + I X R" + I. and the fiber over x with elements of the second factor, (T... f)(v) - Df(x)·v-= 2(x.v).
Since x • O. this linear map is not zero, so as the range is one-dimensional. it is surjective. The same argument shows that the unit sphere in Hilbert . space is a submanifold. B. Let Stem. n; k) = {A
E
L(R'" .R")lrank A = k}.
where k..; mine m. n).
We shall prove using the preceding theorem that St(m,n; k}is a submanifold of L(Rm.R") of codimension (m - k)(n - k); this manifold is called the Stiefel manifold and plays an important role in the study of principal fiber bundles. In order to show that St( m. n; k) is a submanifold. we will prove that every point A E St(m. n; k) has an open neighborhood U in L(R"'.R") such that St(m. n; k)nu is a submanifold in L(R"'.R") of the right codimension; since the differentiable structures on intersections given by two such U obviously coincide (being induced from the manifold structure of L(R"',R"». the submanifold structure of St(m, n; k) is obtained by collation (Exercise 3.2F). Let A E St(m, n; k) and choose bases of Rm.R" such that A = [:
with
II
:]
an invertible k X k matrix. The set
U = { [:
~]
Ix is an invertible k
X k
matrix in the foregoing bases}
is clearly open in L(R"'.R"). A matrix [: iff., - tx- Iy = O. Indeed [ -
~]
with x invertible has rank k
t~-I ~]
is invertible and
so rank[!..
Y.,]=rank[Xo
Y].
v-zx-1y ,
SUBMERSIONS,IMMERSIONS, AND TRANSVERSALITY
165
r
The preceding remark shows that 1(0) = St(m, n; k)n U and thus if f is a' submersion, 1(0) is a submanifold of L(Rm,R") of codimension equal to
r
To see that f is a submersion it is enough to remark that for x, y, z fixed, the map v~ v - zx-1y is a diffeomorphism of L(Rm-Ic,R,,-A). • Next we tum to immersions. 3.5.6 Deflriltlon. A C' map f: M ..... N, r ~ I, is called an immers;of'I at m if Tmf is injective with closed split image in ~(m)N. Iff is an immersion at each m, we just say f is an ;mmers;OII. 3.5.7 Theorem. For a C' map f: M ..... N, where equivalent:
r ~
I. the following are
f is an immersion at m. There are charts (V, (JI) and (V, 1f) with mEV, f( U) c V, (JI: U ..... U',~: V ..... V' X V' and (JI (m ) = 0 such that f",oi-: V' ..... V' X V' is the inclusion u-(u,O); (iii) There is a neighborhood V of m such that f( V) is a submanifold in N and f restricted to V is a diffeomorphism of V onto f( V). (i) (ii)
Proof. ' The equivalence of (i) and (ti) is guaranteed by the local immersion Theorem 2.5.12. Assuming (ii), choose V and V given by that theorem to conclude that f(U) is a submanifold in V. But V is open in N and hence , f(U) is a submanifoJd in N proving (iii). The converse is a direct application , of the definition of a submanifold. • It should be noted that the theorem does not imply that f(M) is a submanifold in N. For example f: SI ..... R 2, given in polar coordinates by r = 00528, is easily seen to be an immersion (by computing Tf using the' curve c(9)=OO528) on SI but f(SI) is not a submanifold of R2; any neighborhood of 0 in Rl intersectsf(SI) in a set with "comers" which is not diffeomorphic to an open interval. In such cases we say f is an immersion with self-intersections. See Fig. 3.5.2.
166
MANIFOLDS AND VECTOR BUNDLES
5' --~
FIgu,,'.1.2
In the preceding ellample I is not injective. But even ir I is an injectiVe immersion. I( M) need not be a submanirold or N. as the rollowing ellample shows. Let I: ]0.2.".[ .... R2 be given in polar coordinates by r = sinO (Fig. 3.5.3) Again the problem is at the origin: any neighborhood or zero does not have the relative topology given by N. If I: M .... N is an injective immersion. I(M) is called an immersed submanilold or N.
y
I
~ (
)
o
2".
x
,.....3.5.'
SUBMERSIONS. IMMERSIONS. AND TRANSVERSALITY
167
3.5.8 Definition. An immersion I: M --+ N that is a homeomorphism onto I( M) with the relative topology induced Irom N is called.an embedding. Thus. if I: M --+ N is an embedding. then I( M) is a submanifold of N. The following is an important situation in which an immersion is guaranteed to be an embedding; the proof is a straightforward application of the definition of relative topology.
3.5.9 Proposition. An injective immersion which onto its image is an embedding.
IS
an open or clmed map
The condition "I: M --+ N is closed" may be replaced by the following. Assume that lor each sequence Xn EM such that I(x n ) converges in N. there is a convergent subsequence in M, Indeed, if this hypothesis holds. and A is a closed subset of M. then I(A) is shown to be closed in N in the following way. Let xn EA. and suppose I(x n ) = Yn converges to yEN. Then there eltists a subsequence {xn,} of {x n }. such that x n , ... x. Since A - cl(A). x e A and by continuity of l.y=/(x)E/(A); i.e.• I(A) is closed. In finite dimensiQns, this hypothesis is assured by the condition "I is proper"; i.e .• the inverse image of every compact set is compact. This is clear since in the preceding hypothesis one can choose a compact neighborhood Vof the limit of I(x n ) in N so that for n large enough all xn belong to the compact neighborhood I I( V) in M. The reader should note that wliile both hypotheses in the proposition are necessary. properness of I is only sufficient. In fact there are injective non proper immersions whose image is a submanifold. For example
I: JO,oo[ --+R2,/(t)= (tcos.!.,tsin!) t I . is an injective nonproper immersion which is an open map onto its image so 3.5.9 applies; the submanifold I(JO, ooD is a spiral around the origin. Third, we examine transversality.
3.5.10· Definition. A C' map I: M --+ N, r ~ I, is said to be trrIIISVn'StII to the submanilold P 01 N (denoted Il"I P) il either I( P) = 0, or il lor every
r
m Erl(p),
I. (Tm/)(TmM)+ 1f(m)p = 1f(m)N and 2. the inverse image (Tm/)- 1(1f(nt)P) 01 1f(",)p sp/iu in T",M.
The first condition is purely algebraic; no splitting assumptions arc made on (TmJ)(T",M), nor need the sum be direct. Also note that if j( ~a Hilbert manifold, in particular finite dimensional. then the splitting CClII'lCJi. tion 2. in the; definition is automatically satisfied.
.
168
MANIFOLDS AND VECTOR BUNDLES
3.5.11 Exampl... A. If each point of P is a regular value of f. then Ifrl P since (T",f)(T",M) = 7i(",)N in this case. B. Assume that M and N are finite-dimensional manifolds with dim(P)+dim(M) < dim(N). Then/frl P implies/( M)n P = 0. This is seen by a dimension count: If there would exist a point mEl '( P)n M. then dim(N) = dim«T",f)(T",M) + 7i(m)P) '" dime M)+dim(P) < dim(N) which is absurd. C. Let M = R2. N = R\ P = the (x. y) plane in R\ (l E R and define fa: M -+ N, by fa(x. y) = (x, y. x 2 + y2 + a). Then Ifrl P iff a ... ·0; see Fig. 3.5.4. This example also shows intuitively that if a map is not transversal to a submanifold it can be perturbed very slightly to a transversal map; for an informal discussion of this phenomenon we refer to Box 3.58. .. image of
f.
..
.. a>O
x
a
Figure 3.5.4
An important basic result about transversality is the following.
3.5.12. Theorem. Let f: M -+ N be a COO map and P a submanifold of N. II IfrlP. then rl(p) is a submanifold of M and T",U-I(P»= (T",/)-I(7i(",)P) lor all mEr I( Pl. II P has linite codimension in N. then codim(r I( P» = codim(P). . Proof. Let (V.y,) be a chart at l(mo)E P in N with the submanifold property for P; let
and denote by P2: VI x V2 -+ V2 the canonical projection. Let (V. cp) be a
SUBMERSIONS. IMMERSIONS. AND TRANSVERSALITY
169
,chart at mo in M, such that fJI(m o ) = 0, fJI: U -0 fJlCU) C E and I(U)c V. For mE U n r I(p), Tm(P2 0 tf 0 II U) = P2 0 ~(m)tf 0 Tn'! and TmCtf 0 f) (TmM)+ FI = F l eF2 (by transversality of Ion Pl. Hence
is onto. Its kernel is (T",f) I(~(m)p) since kerp2 = 1'1 and C~(nr)tf) I(FI ) = ~(m)p, and thus it is split in TmM. In other words. 0 is a regular value of P2 0 tf 0 II U: U -0 F2 and thus (P2 0 tf 0 II U) 1(0) = / 1(1' n V) is a submanifold of U. ,and hence of M whose tangent space al m E U equals ker(T",( P2 tf 0/1 U» ~ (T",f) I(~(",)p) by Theorem 3.5.4. Thus I I( P n V) is a submanifold of M for any chart domain V with the submanifold ,property; i.e., I( P) is a submanifold of M. If P has finite codimension then F2 is finite dimensional and thus again by Theorem 3.5.4. 0
r
codim(r I( P») = codimr I( P n V) "" dim(F2) = codim(P}. _ Notice that this theorem reduces to Theorem 3.5.4 if P is a point.
3.5.13 Corollary. Suppose that M. and M2 are submanifv/ds 0/ M.'m E MI n M~. T,.,M. + T",M2 - T",M. and that T",MI n'T",Mz ,tplil,f in T,,,M for all mE M. n M2 (this condition is usua/~~' denoted by MIII'l M z and we say MI and M2 are tl'flltSVerstl/). Then M. n M2 is a submanifold of M and T",( M. n M 2 ) = T",M. n TmM2.t If both M. and M2 have finite codimension in M, then codim(M I n M 2 ) = codim(M.)+codim(M2 ). Proof. The inclusion il: MI-oM, satisfies i.i"tM2• and M 2 • Now apply the previous theorem. _
ii l(M 2 )=M.n
3.5.14 Exampl... A. In R3. the unit sphere M. = {(x. y. z)lx 2 + .1'2 + Z2 = I) intersects' the cylinder M2 = {(x, y, z)lx 2 + .1" II) transversally if 0 < a-I; MI n M2 = 0 if a > 1 and MI n Mz is the union of two circles if 0
B. The twisted ribbon MI in Fig. 3.5.6 does not meet M 2, the xy plane. in a manifold. so M. is not transversal to M 2 • • There is one more notion connected With geomelric ways to slate the implicit' function theorem that generalizes submersions in a different way than transversality. Roughly speaking, instead of requiring that a map /: MI and Mz are said to intersect dean~v when this conclusion holds. Transversality thus implies c1~an intersection.
t
170
MANIFOLDS AND VECTOR BUNDLES
()
a= 1
Figure 3.5.5
Figure 3.5.6
tl>1
SUBMERSIONS, IMMERSIONS, AND TRANSVERSALfTY
171
M -> N have onto tangent map at each point. one asks that its rank he constant.
3.5.15 Definition. A C map f: M -> N. r ~ I IS called a subimmeniOil if for each point m E M there is an open neighborhood V of m. a manifold p. a submer.fion of: V -> p. and an immersion j: P -> N such that II V = j 0 s. Note that submersions and immersions are subimmersions. The order submersion followed by immersion in this definition is very important. because the opposite order would yield nothing. Indeed. iff: M -+ N is any COO mapping. then we can writef = P2" j. wherej: M -> M x N. is given by j( m) - (m. I( m» and P2: M x N -+ N the canonical projection. gearly j is an immersion and P2 a submersion. In other words. any mapping can be written as an immersion followed by a submersion. . The following will connect the notion of subimmersion to "constant' rank."
3.5.16 Proposition. (i) A Coo map f: M -+ N is a .mhimmersion iff for every m E M there exists a chart (V.,!P) at m. where!p: V -> VI X V2 C FI 63E2 • !p(m) = (0.0). and a chart (V. ~) at f(m) where f(V) c V.~: V -+ VI X V2 C FI 63F2 • ~(f(m» = (0.0) such that f" .. (x. y) = (x.O). (ii) If M or N are finite dimensional. f is a suhimmer.vion iff for every mE M the rank of the linear map T.,'!: Tn! M -> T,I "" N i.v constant on connected components of M. Proof. (i) follows directly from 3.5.2. 3.5.7(ii) and tht: composite mapping theorem: (alternatively. one can use 2.5.15). If M or N art: finite dimensional. then necessarily rank (Tn'!) is finite and tllUs hy (i) the local representative of Tmfhas constant rank in a chart at 111: i.c .. the rank of Tntf is constant on connected components of M. thus proving (ii). The converse follows by 2.5.15 and (i). •
3.5.17 Theorem. Suppose I: M
-+ N is Coo. no EN and Tmf is a subimmersion in an open neighborhood of f I( no); (if M or N are finite dimensional this is equivalent to Tntf having constant rank in a neighborhood of each mErl(no)') Then I '(no) is a submanifold of M with Tmll(no)= ker(Tmf)..
Proof.
Iff I("u) = 0 there is nothing to prove. If mE f '( lin) find charts m and (V.~) at f(m)=II n given by 3.5.16(i). Sjnce !p(vn f;..I(O) = {O}X V 2 we see that (V. cp) has the submanifold property
(V.cp) at f I( no» =
172
for
MANIFOLDS AND VECTOR BUNDLES
r I(n o)' In addition. if
U E
Vz• then
which ill the local version of the second statement. • Notice that
if rank(Tmf) is finite. This theorem reduces to 3.5.4 when f is a submersion. The immersion part of the subimmersion f should imply a version of 3.5.1(iii). This is given by the following.
3.5.18 Fibration Theorem. The following are equivalent. (i) The map f: M -. N is a Coo subimmersion (if M or N are finite dimensional. this is equivalent to the rank of Tmf being loca/~v constant). (ii) for each mE M there exists a neighborhood V ofm. a neighborhood Vof f(m). a submanifold. Z of M with mE Z such that feU) is a .~ubmani fold of Nand f induces a diffeomorphism of 1(V)n Z n V onto f(V)nV; (iii) ker Tf is a subbundle of TM (called the tangent bundle to the fiber) and for each mE M, the image of Tmf is closed in ~(m)N. Iff is open (or closed) onto its image. thenf(M) is a suhmanifold ofN.
r
Proof. To show that (i) implies (ii) choose for V. V the chart domains given by 3.5.16(i) and let Z = (JI- I(VI X{O}). Then
i.e.• f(V) is a submanifold of N. In addition. the local expression of f: f(V)n Vis (x.O) ..... (.1'.0). thus proving that so restricted, f is a diffeomorphism. Reading this argument backward shows that (ii) implies (i). . We have T(JI(kerTf n TV) = VI X Vz X{O}X E z• sincef",.y(x, y) = (.1',0). thus showing that (TV, T(JI) has the subbundle property for ker Tf. The same local expression shows that (Tmf)(TmM) is closed in "t(m)M and thus (i) implies (iii). The reverse is proved along the same lines. If fis open (closed) onto its image. thenf( V)n V [resp. int(/(cI(V»)n V) can serve as a chart domain for f( M) with 'Ii, relative topology induced from N. Thusf(M) is a submanifold or ,'V
r I(V)nZ nV -
SUBMERSIONS, IMMERSIONS, AND TRANSVERSALITY
173
In many applications, one checks the closedness of fusing t,he condition following 3.5,9, which, in finite dimensions, amounts to f being proper. If M or N is finite dimensional, then f being a suhimmersion is equivalent to ker Tf being a subhundle of TM. Indeed, 1m T..'/ = TmMlker Tmf and thus dim(Im Tmf> codim(kerTmf> = constant and f is thus a subimmersion by 3.5.16(ii). We have already encountered subimmersions in the study of vector bundles. Namely. the condition in 3.3.16(i) (which ensured that for a vector bundle map f over the identity ker f and 1m fare subhundles) is nothing else but f being a subimmersion, The fibration theorem has important applications in the study of Lie group actions; it can be used to prove that orbits are immersed submanifolds. We cloSe this section with a study of quotient manifolds. 0=
3.5.19 Definition. An equivalence relation R on a manifold M is called reguJar if the quotient space MIR carries a manifold structure such that the canonical projection 71': M -+ MI/!. is a submersion. If R is a regular equivalence relation. then M IRis called the quotient lIUlIIi!old of M by R.
Since submersions are open mappings, 71' and hence the regular equivalence relations R are open. Quotient manifolds are characterized by their effect on mapping~.
3.5.20 Proposition. Let R be a regular equivalence relation on M. (i) A map f: MIR-+ N is C, r ~ I iff f 0 71': M -+ N is C. (ii) Any C' map g: M -+ N compatible with R. i.e. xRy implies g(x) = g(y), defines a unique C map g: MIR -+ N such that go 71' = g. (iii) The manifold structure 0/ MIR is unique if we demand that 71' is a submersion. Proof. (i) If f is C, then so is / 071' by the composite mapping theorem. Conversely, assume f 0" is C'. Since 71' is a su~mersion it can be locally expressed as a projection and thus there exist charts (V, fJI) at m E M and (V, 1/1) at 7T(m)E MIR such that fJI(U) = VI X V2 C E l fBE 2 , 1/I(V) = V2 C E 2 • and 7T",I/-(X, y) = y. Hence if (W. X) is a chart at (f 7T)(m) in N with (f 0 7T)(U) C W. then fXI/- = (f 0 'II')"'I/-I{O} x V2 and thus fXI/- is C. (ti) The mapping g is uniquely determined by go 71' = g. It is C by (i). (iii) Let (MIR)I and (MIRh be two manifold structures on MIR having 71' as a submersion. Apply (ii) for (MIR)I with N=(MIRh and g = 71' to get a unique Coo map hi: ( M I R) I -+ ( M I R h such that h 0 71' = 71'. Since 71' is suIjective, h = identity fhanging the roles of the indices I 0
174
MANIFOLDS AND VECTOR BUNDLES
and 2, shows that the identity mapping induces a COO map of (M/Rh to (M/R) •. Thus, the identity induces a diffeomorphism. • 3.5.21 Corolla". Let M and N be manilolds, Rand Q regular equivalence relations on M and N, respectively, and I: M -- N a C' map. r ~ I, compatible with Rand Q; i.e., il xRy thenl(x)QJ(y). Then I induces a unique C map): M/R -+ N/Q and the diagram
commutes.
Proof. The map) is uniquely determined by '!TN 0I = )0 '!TM' Since '!TN 0I is c.l is C' by 3.5.20. The diagram is obtained by applying the chain rule to "N
0I - )0 '!TM'
•
The manifold M / R might not be Hausdorff. By 1.4. \0 it is Hausdorff iff the graph of R is closed in M X M (R is open since it is regular). For an example of a non-Hausdorrr'tlUQtient manifold see Exercise 3.5H. There is, in fact, a bijective ~ndence between surjective submersions and quotient manifolds. More precisely. we have the following. . 3.5.22 Proposition. Let f: M -+ N be a submersion and let R he the equivalence relation defined hr f; i.e .. xRy ilf I(x) = ltv). Then R i,~ re1{ular. M/R is dilfeomorphic to I( M). and f( M) is open in N.
Proof. As/is a submersion. it is an open mapping. so/<M) is open in N. Moreover, since I is open, so is the equivalence R and thus I induces a homeomorphism of M/R onto I(M) (see the comments following 1.4.9). Put the differentiable structure on M/R that makes the homeomorphism into a diffeomorphism. Then M/ R is a manifold and the projection is clearly a submersion, since I is. • This construction provides a number of examples of quotient manifolds. 3.5.23 EumpIea. A. The base space of any vector bundle is a quotient manifold. (Taite the submersion to be the vector bundle projection.)
SUBMERSIONS, IMMERSIONS, AND TRANSVERSALITY
175
B. The circle S' is a quotient manifold of R defined by the submersion IJ 0-+ e i'; we can then write S' ... R/2"Z. including the differentiable structure. C. Th~ Grassmannian G,,(E) is a quotient manifold in the following way. Let D'" {(x, •...• .1'" )1.1', E E. x, •...• .1'" are linearly independent). D is open in Ex··· X E(k times). Thus one can define the map ,,: D ~G,,(E) by ,,(x, •...• xh)=span (.1', •.•.• .1',,). Using the charts described in Example 3.1.8(G). one finds that " is a submersion. In particular. the projective spaces RP" and CP" are quotient manifolds. D. The Mobius band as explained in 3.3.8C is a quotient manifold. E. Quotient bundles are quotient manifolds (see 3.3.16). ... We close with an important characterization (If regular equivalence relations. 3.5.23 Theorem.
An equivalence relation R on a mani/old M is regular iff
graph(R) is a submanifold of M X M. and (ii) PI: graph(R) .... M. p,(x. y) = x is a submersion. (i)
Proof. First assume that R is regular. Since ,,: M .... M/R is a submersion. so is "X ,,:M X M ~ (M/R)X(M/R) so that graph(R) - (" X '1r) - '(I1 M/ R ). where I1M/R = {(Ix J.lx))l[xJ E M/ R) is the diagonal of M/ R X M/R. is a submanifold of M X M (3.5.12). Let (x. y)E graph(R) and vx E ~!tf. Since '1r is a submersion and ,,(x) = ,,(y), there exists Vv E T,.M such that Tvw( v,.) = T,"( v,); i.e .• (v,. vv) E 1(,. I,,(graph( R» by 3.5.12. But then 1(, .•"p,(v;. v.. ) = v" showing that" is a sU,bmersion. . Conversely, assume (i) and (ii) are satisfied. We construct a chart at [X)E M/R in the following way. First let us show that" is an open mapping. Indeed. if V is open in M. then (V X V )ngraph( R) is open in the submanifold graph( R) and since p,: graph( R) .... M is a submersion it is open (3.5.2) so that V X V)ngraph( R» == ,,-'( w( V» is open; i.e., " is open. Let (x.X)EI1 M • Since graph(R) is a submanifold. there is a chart cp: V X V .... V'x V'c E X E such that cp«V X V)ngraph( R)) = (V'X V')n( F X{O}) .... V". where Ex E = F$G. cp(V X V)= V;' V', V" c F, and V'c G. Since p,: graph( R) .... M is a submersion, its local representative in the chart cp is also a submersion. We still denote it by PI: V" -4 E. Now use the local submersion theorem to change the chart cp inside V" so thal p, becomes projection onto the first factor; i.e.• find an open set of the form V, X VI C V", where cp(x, x) E VI X VI' and a diffeomorphism of an open set in V" onto V, X V, such that in this new chart PI is projection onto VI c E. Here V, cHand E$H'"= F.
p,«
176
MANIFOLDS AND VECTOR BUNDLES
Now for a sufficiently small neighborhood VI
X V'
c HeG;; E (since
Ex E., EeGeH) we have
I. w( '" - I( VI X V') is open in M / R (since w is open); 2. w( '" - I( VI X (O}» = w( '" - I( VI X V'» (by construction); 3. there is a homeomorphism induced between
4.
(again by construction) define x: w( '" - I(VI X (O})) --+ VI by
X( W",-I( v.O»
=
v.
so X is bijective. It is a homeomorphism by 3. The maps X are the charts for M/R. We only need to show that the overlap maps are Coo. But X is essentially detennined from the implicit function theorem and we know frc;>m 2.5.7 that the implicit function is a Coo function of any parameters in the problem. Using this idea, the reader can easily complete the proof. • This theorem essentially occurs in classical works concerning quotients of Lie groups and of manifolds by group actions on them. (See for example Chevalley [1946] and also see Abraham and Marsden [1978. ch. 4] for some applications to mechanics).
BOX 3.5A
LAGRANGE MULTIPLIERS
Let M be a smooth manifold and i: N'-+ M a submanifold of M. i denoting the inclusion mapping. If I: M --+ R, we want to determine necessary and sufficient conditions for n E N to be a critical point of /IN. the restriction of I to N. Since liN = I 0 i. we have T"UIN) = T,.! 0 Tlti; i.e., n E N is a critical point of I iff T,,/l T"N = O. This condition takes a simple form if N happens to be the inverse image of a point under a submersion.
3.5.24 Propo8lt1on. Let g: M --+ P be a smooth submersion. N = g - I( Po) and let I: M --+ R be cr. r ~ I. A point n E N is a critical point 01 liN iff there exists>. E T;.P. called a Lagrange multiplier. such that 1',,1 = >. 0 T"g.
SUBMERSIONS. IMMERSIONS. AND TRANSVERSAL/TY
Proof.
177
First assume such a ). exists. Since
i.e., 0 = (). 0 T"g)1 T"N = T"fl Tn N. Conversely, assume Tnfl TnN = O. By the local form for submersions, there exists a chart (V,cp) at n,cp: V-V, x V,cEX F such that cp( V () N) = {O} X V, and q>( n) = (0.0). and a chart ( V. I/- ) at p()' 1/-: V-VIc.E where g(V)cV.IHpo)=O. and such that g'l't(x. y .>= (l/-ogocp-l)(x.y)=xforall(x.y)EVI X VI' If!'I'=!oq>- : VIXV, -R, we ,have for alii E F, Dd'l'(O, 0)· I = 0 since Tn!1 TnN = O. Thus, letting 11 == Dlf'l'(O, O) E E*,e E E and IE F, we get
Df'l'(O,O)' (e. f) = Il( e) = (11 0 Dg'l'''')(O.O)· (e. f); i.e.,
Df'l'(O.O) = (11 0 Dg'l'''')(O,O). To pull this local calculation back to M and P"let ). = 11 0 1',.1/- E r;op, so composing the foregoing relation with T"cp on the right we get Tnf =). 0 T"g. •
3.5.25 Corollary. LeI g: M .... P be tran.wersal to the submanifold W of P, N = g-I(W), and let f: M .... R be C, r ~ I. LeI E,(n) be a closed complement to T,ln)W in TIf(n)P so T,(n)P = T,(n)WeEIf(n) and let 'IT: T,(n)P .... EIf(n) be the projection. A point n EN is a critical point of fl N iff there exists ). E E~ n) called a Lagnurge mllitiplkr such that Tnf =). 0 'IT 0 T"g. By 3.5.12, there exists a chart (V, cp) at n, with q>( V) = VI X U2 eEl X E 2 , q>(V () N) = (O}X V2' and q>(n) = (0,0), and a chart (V,I/!) at g(n) with I/!(V) = VI X VI C FI X F, I/-(V () W) = (O}X VI,I/!(g(rr» =(0,0), and g(V)CV, such that g'l'",(x,Y)=(l/-ogocp-I)(X,y)= (x, 1/(x, y» for all (x, y) E VI X VI" Let p: EI X F -+ E. be the canonical projection. By the previous proposition applied to the map po g"",: VI X V2 .... VI.(O,O) E V, x V2 is a critical point of fl{O} X V2 iff there exists" E Er such that Df,,(O,O) 0 p 0 Dg'l''''(O.O). Composing this relation on the right with T"cp and letting ). = T,(n)l/!. 'IT = (T,(n)¥- )-IIEIf(n) 0 p 0 TIf(n)l/-: 1~\,,) /' • ElfIn)' we get Tnf =). 0 'IT 0 T"g . Proof.
="
,,0
•
If P is a Banach space F, then 3.5.24 can be formulated in the following way.
178
MANIFOLDS AND VECTOR BUNDLES
3.5.21 Corollary. Let F be a Banach space. g: M -+ F a smooth submersion. N - g-I(O). and I: M -+ R be cr. r Oil> I. The point n E N is a critical point 01 II N iff there exists lI. E F*. called a Lagrange mIIItipilr. such that n is a critical point 011 - lI. 0 g. Notes
lI. depends not just on II N but also on how I is extended off N. This form of the Lagrange multiplier theorem is extensively used in the calculus of variations to study critical points of functions with constraints: cr. Caratheodory (1965) 3. We leave it to the reader to generalize 3.5.25 in the same spirit. I. 2.
The name lAgrange multiplier was first used in conjunction with the previous corollary in Euclidean spaces. Let V be an open set in R".F ... R'.g_(gl ..... g'): V-+RP a submersion and I: V-+R smooth. Then x eN - g-I(O) is a critical point of fiN. iff there exists p
lI.=
L AieiE (RP)*. i-I
where e l •... • e P is the standard dual basis in R P such that n is a critical point of p
I -lI.
0
g == I
-
L ";gl.
I-I
In classical calculus the real numbers A; are referred to as Lagrange multipliers. Thus. to find a critical point x-(xl •...• x'")eNcR'" of
liN
one solves the system of m + p equations
al
-(x)-
ax'
ag' LP ",-(x) = o. i_I ax}
j=l ....• m
i .. I ..... p
g'(x)=O.
Ar
for the m + p unknowns XI •••.• X'". "1 ..... For example. let N - S2 C Rl and I: R -+R;/(x. y. z)- z. Then II S 2 is the height function on the sphere and we would expect (0.0. ± I) to be the only critical points of II S2; note that I itself has no
SUBMERSIONS. IMMERSIONS. AND TRANSVERSAL/TY
179
critical points. The method of Lagrange multipliers. with g(x. y, z)= x 2 + y2 + Z2 -I, gives 0-2xA-O;
0-2yA - 0; 1-2zA =0;
The only solutions are A = ± ~. x = O. Y = O. z ~ ± I. and indeed these correspond to the maximum and minimum points for / on S2. See an elementary text such as Marsden and Tromba 119R I J for additional examples. For more advanced applications. see Luenberger (1969). The reader will recall from advanced calculus that maximum and minimum tests for a critical point can be given in terms of the Hessian. i.e.• matrix of second derivatives. For constrained problems there is a similar test involving bordered Hessian.v. See Marsden and Tromba [1981. pp. 224-30) for an elementary treatment. We invite the reader to formulate such results intrinsically on manifolds.
BOX 3.5B
TRANSVERSALITY THEOREMS
This box addresses two questions. H /: M -+ N is a C'-map and PeN is a submanifold. what can one say about thc points m at which / is transversal to P? Consider'the set 1"1' ( M. N; ,.) = {l: M -+ N 1/ is of class (" and /1"1 P). What can one say about this set in the set of all (" maps from M to N? The first question is relatively casy.
3.5.27 Proposition. Let f: M
-+
N be C' and!'
The set (m E Mil is transversal to P at m) is ope"
II 11/
.l"IIhlllani/o/d of N. M.
Proof. Assume I is transversal to P at m E M. Choose a submanifold chart (V.q'I) at I(m) E P. q'I: V -+ F, X F2 • q'I( V () 1') = ,.', X{O}. Hence if 'IT: F, x F2 -+ F2 is the canonical projection. V () " = If '( F, x {O}) = ('IT 0 q'I) - '(O). Clearly. 'IT 0 q'I: V () P -+ F2 is a submersion so that by 3.5.4. ker1j(m)('lT o q'l)=Tf (m)P, Thus/is transversal to P at/(m) iff Tf(p)N = ker T'(m)('lToq'l)+ Tf(p,P and (T,.,l) '(Tr(m)!') = ker Tm( 'IT 0 q'I 0 I) is split in Tm M. Since q> 0 'IT is a submersion this is equivalent to 'IT 0q'I 0/ being submersive at mE M (see also Exercise . 2.2E). But as in Proposition E.6 in Appendi){ E. the set where 'IT 0 q'I 0 / is submersive is open in U. hence in M. where U is a chart domain such that/(U)c V. •
180
MANIFOLDS AND VECTOR BUNDLES
The answer to the second question is one of the fundamental theorems in differential topology. Since this-and the techniques involved-would take us too far afield. we shall limit ourselves here to the description of the so-called classical Thom Iransversality Theorem. The reader is referred to Abraham and Rohhin [19671. Hirsch (IQ76J. and Golubitsky and Guillemin (1974) for the proofs and many applications. Guillemin and Pollack (1974) is an excellent introduction to the subject with as little technical machinery as possible. The first task is to give C( M. N) = {f: M -+ Nil of class C} a reasonable topology. The idea is that two C'-maps should be "close" if their graphs and the graphs of the first r derivatives are close and "at infinity" they approximate each other arbitrarily well. This is formalized in the following way. assuming the manifolds in question are finite dimensional and second countable. The Whitney C'-topology is defined for 0" r < 00 in the following manner. Let f1 == {U•• CP.)}.. e A be a locally finite atlas of M; i.e.• every point of M has a neighborhood intersecting only finitely many V,.. Let ~ = {K,.},.e A' K,. c V,. be a family of compact sets in M and ~ = Y,1I)}lIe B' an atlas of N. For every family e = (e,.}ae A of positive numbers. define
«Vp.
"V,.s.x .• (f)'" (g E C'(M. N)lg(K.) C J-I,(." sup
IIDJ( Y,II(., 0
I cP,.)( x)- DJ( Y,1I(ClI 0
0
g 0 cp,.)(x)1I < e. for all
Of
.te".(K.I. j-O ..... ,
The Whitney C'-topology is by definition generated by these as a basis of open sets. If r = 00. (resp. '" (analytic». the COD (resp. C W ) Whitney topology is the initial topology induced by the inclusions COD( M. N )(resp. CW( M. N» .... C'( M. N) for all r. This topology is extremely fine and has rather unpleasant properties. It has uncountably many connected components and is not first countable. thus not me&rizable. It has however a useful property that makes it suitahle to discuss approximations in C'( M. N). namely. C( M. N) with the Whitney C r -topology is a Baire space. Moreover. interesting special classes of mappings have useful properties in C r ( M. N). Thus C' diffeomorphisms. embeddings. dosed embeddings. immersions. submersiOns. and proper maps are dense in the corresponding C' classes 01 maps 01 C'(M. N)/or I < r. r == 1..... 00. Col. Moreover. the same classes 01 C maps are all open in C'(M. N). (r ~ I for diffeomorphisms; homeomorphisms are not open in Co( M. N ).)
E
A} .
SUBM,ERSIONS.IMMERSIONS. AND TRANSVERSALITY
181
This implies in particular that two C' manifolds C' diffeomorphic. I < r. are also C' diffeomorphic. As far as density is concerned. the situation is considerably more complicated. Thus, if dime N ) ;;.>. 2 dim( M), immersions are dense in C(M, N). But embeddings are dense in the C'-proper maps only for dim( N) ;;.>. 2dime M) + l. Since proper maps always exist (see Exercise 5.50), this theorem implies the Whitney embedding theorem: every C' n-manifold is diffeomorphic to a closed submanifold of R 2" + I. Let us return to the most important theorem of this kind. Let 7ft'(M,N;P)-{feC(M,N)lr;;.>.l,f is transveral to P at every point of M}. 3.5.28 Theorem (R. 11Jom). 7ft' (M. N; P) is residual, and hence dense, in C(M. N). If P is closed this set is also open. One can even be more precise regarding the approximation. Let VI' V2 be open in M such that cl( VI) C V2 • Then given f e C( M. N) there exist functions g e C( M, N) arbitrarily close to f. coinciding withfon VI and gi"l P off V2 • This classical transversality theorem has .. parameteric" versions (see Abraham and Robbin (1967» and generalizations to jet bundles (the "jet transversality theorem"), which are very useful in differential topology (Hirsch [1976J, Oolubitsky and Ouillemin (1974».
Exercl... 3.5A Show that the set of elements of L(Rn,R") with determinant I is a submanifold. 3.58 Show that the set O(n) of elements of L(R",R") that are orthogonal (i.e.• QQT = I) is a submanifold. (QT denotes the transpose of Q.) 3.5C (i) Let P c 0(3) be defined by P= {QeO(3)ldetQ= + I.Q=Q T }
Show that P is a two-dimensional compact submanirold of 0(3).
(ii) Define f: RP 2 - 0(3), f(l) = the rotation through 'IT about the line I. Show that f is a diffeomorphism of RP 2 onto P. 3.5D If N is a submanifold of dimension n in an m-manifold M, show that for each x e N there is an open neighborhood V c M with x e V and a submersionf: VcM-R m -", such that NnV=rl(O). (Hint: Choose a submanifold chart and compose it with an appropriate projection.)
182
MANIFOLDS AND VECTOR BUNDLES
3.5E
Show that Rpl is a submanifold of Rp2, which is not the level set of any submersion of Rp2 into Rpl; in fact, there are no such submersions. (Hint: Rpl is one-sided in Rp2, that is, there is no continu()us choice of normal direction to Rpl in RP2.) 3.5F (i) Show that if f: M --+ N is a subimmersion, g: N ...... P an immersion and h: Z --+ M a submersion, then go I 0 h is a subirnmersion. (ii) Show that if /;: M; --+ N;, i = 1,2 are immersions (submersions, subimmersions), then so is II X 12: MI X M2 ...... NI X N2. (iii) Show that the composition of two immersions (submersions) is· again an immersion (submersion). Show that this fails for subimmersions. (iv) Let f: M --+ N and g: N --+ P be C, r ~ I. If g 0 I is an immersion, show that I is an immersion. If g 0 I is a submersion and if I is onto, show that g is a submersion. .
3.SG
Let M be a manifold. R a regular equivalence relation and S another equivalence relation implied by R; i.e.• graph R c graph S. Denote by SjR the equivalence relation induced on MjR. Show that S is regular iff SjR is and in this case establish a diffeomorphism (MjR)j(SjR) ...... MjS. (ii) Let M;. i-I, 2 be manifolds and R, be regular equivalences on Mi' Denote by R the equivalence on MI X M2 defined by RI X R 2 • Show that MjR is diffeomorphic to (MljRI)X (MzlR 2 )· 3.SH (The line with two origins). Let M be the quotient topological space of (R X{O»U(R X{I» by identification of (/,0) with (t,I) for '" O. Show that this is a one-dimensional non-Hausdorff manifold. Find an immersion R --+ M. 3.51
(i)
Let f: M --+ N be C'" and denote by h: TM --+ r(TN) the vector bundle map over the identity uniquely defined by the pull-back.' Prove the following: (i) (ii) (iii)
h
I is an immersion iff 0 ...... TM --+ r(TN) is exact; h I is a submersion iff TM --+ r( TN) --+ 0 is exact; I is a subimmersion iff ker h and range h are subbundles.
3.SJ
Let A be a real nonsingular symmetric n X n matrix and c a nonzero real number. Show that the quadric surface {x E R"I(Ax, x) = c) is an (,. - 1)-submanifold of R". 30SK (Steinefs Roman Surface). Let f: S1 ...... R4 be defined by 1(.1', y, z) = (yz, xz, xy • .t 2 + 2y2 +3z 2 )
SUBMERSIONS,IMMERSIONS, AND TRANSVERSAL/TY
3.5L
3.5M 3.5N 3.50 3.5P 3.5Q 3.5R
3.5S
·183
(i) Show that f( p) = f(q) if and only if p = ± q. (ii) Show that f induces an immersion j: RP 2 -+ R 4. (iii) Let g: RP2 -+ R3 be the first three components of f Show that g is a .. topological" immersion and try to draw the surface g(Rp 2 ) (see Spivak (1979) for the solution). (Covering Maps). Let f: M -+ N be smooth and M compact. dime M) ... dime N) < 00. If n is a regular value of f. ~how (hat I( n) is a finite set {ml ..... m k } and that there exists an open neighborhood V of n in N and disjoint open neighborhoods Vi ..... Uk of ml ..... m A such thatrl(V)=U,U'" UVk • and flU,: L~-+I·.i=I. .... k are' all diffeomorphisms. Show k is constant if M is connected. Let f: M -+ Nand g: N -+ P be smooth maps. such (hat gi"\ V where Vis a submanifold of P. Show thatf/l'\g-I(V) iffg 0 f/l'\ V. Show that an injective irl)mersion f: M -+ N is an embedding iff f( M) is a closed submanifold of an open submanifold of N. Show that the map p: St(n. n; k) -+ Gk(R") defined hy p( A) = range A is a surjective submersion. Show that f: RP" xRP'" -+RP",ntmh given by (.\".y)>-> [xoYo; XOJ'I" .. x,Y,.· ... x"Ym) is an embedding. Show that {(x. Y) ERP" xRpmln ~ m.I:7_ox,Y, = O} is a (m + n -I) manifold. It is usually called a Milnor manifold. (Fiber product of manifolds). Let f: M -+ P and g: N -+ P be ('00 mappings such that
r
r
CHAPTER
4
Vector Fields and Dynamical Systems
This chapter studies vector fields and the dynamical systems, also called flows, they determine. The ensuing chapters will study the related topics of tensors and differential forms. A basic operation introduced in this chapter is the Lie derivative of a function or a vector field. It is introduced in two different ways, algebraically and as the rate of change of the function or vector field along a flow. The Lie derivative formula asserts the equivalence of these two definitions. The Lie derivative is a basic operation used extensively in differential geometry, general relativity, Hamiltonian mechanics, and continuum mechanics. 4.1
VECTOR FIELDS AND FLOWS
This section introduces vector fields and the flows they determine. This topic puts together and globalizes two basic ideas we learn in undergraduate calculus: the study of vector fields on the one hand and differential equations on the other. 4.1.1
DeftntUon. Let M be a manifold. A vector/ield on M is a section of the tangent bundle TM of M. The set of all C' vector jielJ.<:a.'" M IS denoted by ~ '( M) and the C"" vector fields by ~X OO( M) or ~X (M )
Thus a vector field X on a manifold M is a mapping X: M -- TM such that X( m) E TmM flU
for all m E M.
VECTOR FIELDS AND FLOWS
185
In other words, a vector field assigns to each point of M a vector based (i.e., bound) at that point. 4.1.2 Example. Consider the force field determined by Newton's law of gravitation. Here the manifold is R 3 minus the origin and the vector field is
F(x, y, z} =
mMG
-3-r,
-
r
where m represents the mass of a test body, M the mass of the central body, G is the consta!1~~f_gravitation, r is the vector from the origin to (x. y, z), and r _/x 2 + y2 + z2 . See Fig. 4.1.1. •
\
"
I
\
I
-
,/
/
11\
\
,.,. /
-
-,
"-
Figure 4.1.1
The study of dynamical SYSh'lliS. also called flows. may be motivated as follows. Consider a physical system that is capable of assuming various "states" described by points in a set S. For example, S might be R3 X R3 and a state might be the position and momentum (q, p) of a "article. As time passes, the state evolves. If the state is So E S at time to and this changes to S at a later time t, we set
£',lo(SO) =s and call F"I. the evolution operator; it maps a state at time to to what the state would be after time I - 10 has elapsed. "Determinism" is expressed by the law F,.I = identity (sometimes called the Chapman-Kolmogorov law).
186
VECTOR FIELDS AND DYNAMICAL SYSTEMS
By saying the evolution laws are time independent. we mean F,.to depends only on I - 10' Setting F, - F,.o. the preceding law becomes the group properly:
F, 0 F., = F,+T'
\
EO = identity
We call such an F, a flow and F,.to a time-dependenl flow. or as before. an evolution operator. If the system is nonreversible. that is. defined only for' t ;> 10' we speak of a semi-flow. In physics, it is usually not F,. to that is given. but rather the laws of molion. In other words, some differential equations are given that we must solve in order to find the now. These equations of motion have the form ds dl
-
=
X(s)
'
s(O) = So
where X is a (possibly time-dependent) vector field on S. ".1.3 Example. The motion of a particle of mass m under the innuence of the gravitational force field in Example 4.1.2 is determined by Newton's second law:
i.e.• by the ordinary differential equations
Letting q = (x, y. z) denote the position and p these equations become
=
m (dr/dt) the momentum.
~ .. F(,). The state space here is the manifold (R 1 \{0))XR 3, i.e .• the tangent manifold
VECTOR FIELDS AND FLOWS
187
of R 3 \{O}. The right-hand side of the preceding equations define a vector field on this six-dimensional manifold by X(q.p)= «q. p).( p/m. F(q))). In elementary courses in mechanics or differential equations. it is shown how to integrate these equations explicitly. producing trajel:tories. which are planar conic sections. These trajectories comprise the n(.w of the vector field. • Let us now tum to the elaboration of these ideas when a v.ector field X is given on a manifold M. Ir M = V is an open subset of a Banach space E. then a vector field on Visa map X: V .... V X Eofthe form X(x) = (x. V(x». We call V the principal parI of X,. However. having a separate notation for the principal part turns out to be an unnecessary hurden. Thus. in linear spaces we shall write. by abuse of notation. a vector field simply as a map X: V .... E and shall mean the vector field x'" (x. X ( When it is necessary to be careful with the distinction. we shall he. Ir M is a manifold and tp: V eM .... VeE is a chari for M. then a vector field' X on M induces a vector field X on E called the' local represenlalive of X by defining X( x) = Ttp· X( tp - I(.\: n. If E = R" we can identify the principal part of the vector field X with an n-component vector function (XI(x) •...• X"(x». Thus we sometimes just say "the vector field X whose local representative is (XI) = (XI ..... X")." Recall that a curve c at a point m of a manifold M is a C l map from an open interval I of R into M such that 0 E I and c(O) = m. For such a curve we may assign a tangent vector at each point ('(/), I F I. hy c'(/) = TC(/.I).
,l».
iIIt.'"
4.1.4 Definition. LeI M be a manifold and X E <x ( M), An euiw at mE M is a curve c al m such Ihal C'(/) - X(c(t» for each I E I.
0/ X
In case M ... U c E. a curve c(t) is an integral curve of X: V .... E when
t(/)" X(C(/». where t = dc/dl. Ir X is a vector field on a manifold M and X denotes the principal part of its local representative in a chart tp. a curve c on M is an ' integral curve of X when
: (t)-X(l(/».
.
where C" cp. c is the local represenlative of c. If M is an II-manifold and &he local representatives of X and c are (XI ..... X .. ) and «('I ..... C.. ). respectively. then c is an integral curve of, X when the following system of ordinary
188
VECTOR FIELDS AND DYNAMICAL SYSTEMS
differential equations is satisfied: cl(t)= XI(CI(t) •...• c"(t»)
c"(t) "" X"(cl(t) •...• c"(t)).
The reader should chase through the definitions to verify this assertion. These equations are autonomous, corresponding to the fact that X is time independent. If X were time dependent. a t would appear explicitly on the right-hand side of the foregoing equations. As we saw in Example 4.1.3. the preceding system of equations includes equations of higher order (by their usual reduction to first-order systems) and the Hamilton equations of motion as special cases. For a system of ordinary differential equations there are well-known existence and uniqueness theorems. The form that we shall need is the following. 4.1.5 Theorem. Let E be a Banach space and X: VeE ..... E be of class COO. For each Xo E V, there is a curve c: I ..... Vat Xo such that c'(t) = X( c(l» for all tEl. Any two such curves are equal on the intersection of their domains. Moreover, there is a nei~hborhood Vo of Xo E V. a real number a> 0, and a COO mapping F: Vo X I ..... E. where J = ( - a. a) such that the curve c,,(t): I ..... E, defined by c .. {t) = F(u, t) is a curve at u E E satisfying the differential equations c~( t) - X( cII( t» for all tEl. The proof proceeds in several steps. We begin with existence and uniqueness. 4.1.8 Lemma. Let E be a. Banach space. U c E an open set, and X: U c E ..... E a Lipschitz map; that is IIX(x)- X(y~1 CO Kllx - yll for all x. y E V and a constant K. Let Xo E V and suppose the closed ball of radius b.Bb(xo)-(XEElllx-xoll~b} lies in V, and IIX(x~I~M for all xE Bb(x o )' Let toER and leI o:=min(l/k,b/M). Then there is a unique C l curve x(t), t E [to - 0:, to + 0:] such that x(t) E Bb(x o) and {
x'(t} = X(x(t» . x(t o } = Xo
Proof. The conditions x'(t)=X(x(t».x{to}=x o are equivalent to the integral equation x(t}=x o + j'X(x(s»ds.
'0
VECTOR FIELDS AND FLOWS
189
Put xo(t) = Xo and define inductively X"+I(I)=X O + j'X(x,,(s»ds. 10
Clearly x"(t) E B,,(x o ) for all n and IE [10 - a, 10 + a] hy definition of a. We also find by induction that
Thus x,,(t) converges uniformly to a continuous curve x(l). Clearly x(t) satisfies the integral equation and thus is the solution we sought. For uniqueness, let y(t) be another solution. By indul,tion we find that IIx,,(t)- y(tlll ~ MK"II-lol"+ I; thus. letting n - 00 give~ x(1) =~,(t). • Let us observe that exactly the same result holds if X depends explicitly on I and/or on a parameter p, is jointly continuous in (t. p, x), and is Lipschitz in x uniformly in I and p. The next result is a basic estimate used to study the dependence on initial conditions. It is Gronwall's inequality.
4.1.7 Lemma. LeI f. g: [a, b[ - R be conlinuous and nonnegative. Suppose
f(I)~A+ f'f(s)g(s)ds./oraconslantA~O. (J
Then f(t)" Aexp(fg(s)ds)
forall
tE[a,b[.
Proof. First suppOse A> O. Let h(t) = A + J:f(s)g(s) dv; thus h(t) > O. Then h'(t) = f(t)g(t) " h(t)g(t). Thus h'(t)/h(t) ~ g(t). Integration gives h(t)" if exp O. If if = 0, then we get the result by replacing if by E> 0 for every E> 0; thus h and hence f is zero. •
4.1.. Lemma. Let X be as in Lemma 4.1.6. Let F, (x o ) denote the solution (-integral curve) of X'(I)=X(x(t»,x(O)=x o' Then there is a neighborhood V of Xo and a number Il > 0 such Ihal for every y E V there is a unique integral curve x(t) = F,(y) salisfying x'(t) = X(x(t) for IE [- E, II and
190
VECTOR FIELDS AND DYNAMICAL SYSTEMS
x(O) -= y. Moreover. 1IF,(x)- F,( Y)II" eK1'Ilix -
ylI·
Proof. The first part is clear from Lemma 4.1.6. For the second. let /(1) -1IF,(x)- F,(Y~I. aearly
/(1) =- Ilfo'X(~(X»)- X(~(y» dr + x -
so the result foUows from Lemma 4.1.7.
yll" IIx - yll+ Kfo'/{s}
dr.
•
This result shows that F,(x) depends in a continuous. indeed Lipschitz. manner on the initial condition x and is jointly continuous in (t. x). Again. the same result holds if X depends explicitly on 1 and on a parameter p; the evolution operator F!.,o(x) is the unique integral curve x( t) satisfying x'(t) - X(X(I). I, p) and X(lo) = x. Then F,~,,,(x) is jointly continuous in (to. I. P. x). and is Lipschitz in x. uniformly in (to. I. p). The next result shows that F, is C· i~ X is. and completes the proof of 4.1.5.
4.1.8 Lemma. LeI X in Lemma 4.1.6 be 0/ class C k • I ~ k ~ 00. and leI F,(x) be defined as be/ore. Then locally in (t. x). F,(x) is 0/ class C* in x and is Ck+ I in the I-variable. Proof. We deCine "'(I. x) E L(E. E). the continuous linear maps of E to E. to be the solution of the "linearized" or "first variation" equations:
d
dl"'(I.x)-DX(F,(x»)o",(t.x).
'" (0. x) = identity where DX( y): E .... E is the derivative of X taken at the point y. Since the vector field"' ..... DX(F,(x»o '" on L(E, E) (depending explicitly on I and on the parameter x) is Lipschitz in "'. uniformly in (I. x) in a neighborhood of every ('0' x o ). by 4.1.8 it foUows that "'(I. x) is continuous in (I. x) [using. the norm topology on L(E. E»).
VECTOR FIELDS AND FLOWS
We claim that DF,(x) == 1/1(/, x). To show this, lIet 8(t. h) = F,(x F,(x) and write
191
+ h)-
'(I, II )-1/1(/, x )·11 - fo'{X( F,(x + II» - x(f~( x»} ds - fo'[DX(F,(x» 0 1/I(s, x»)·hds· - fo'DX( F,(x»· ['(s, 1I)-1/I(.f. x )·11] tis
+ !'<X(f,(x + h»- X( F,(x»- DX(F,(x»'[F,(x + h)- F,(x)]} tis. Since X is of class C l , given e> 0, there is a I) > 0 such that IIhll < I) implies the second term is dominated in norm by fJellF,(x + h)- F,(x~1 dv, which is, in tum, 'smaller than Aellhll for a positive constant A. By GroDwall's inequality we obtain 118(1,11)-1/1(1, x)'hll" (constant)ellhll. It follows that DF,(x)·1I = 1/I(t,x)·h. Thus both partial derivatives of F,(x) exist and are continuous; therefore, F,(x) is of ~Iass CI. We prove F,(x) is c" by induction on k. Now
d dt F,(x) =
X( F,(x»
so d d . dl dlF,(x) - DX(F,(x»'X(F,(x»
and d dl DF,(x)'" DX( F,(x »·DF,(x).
Since the right-hand sides are Cit -~, so are the solutions by induction.· Thus F itself is Cit. • Again there is an analogous result for the evolution opera~(X) for a time-dependent vector field X( x, I, p), wbK:h may also .. extra parameters p in a Banach space P. II X is ~", then:~tX) is C~ -. •. . variables ~ is C" + I in , and '0' The variable p can be • dealt with .., _pending X to a new vector rleld obtained by appendiDa." ..... . differential equation p. 0; this dermes a vector rleld on BX r ... 44$" may be applied to iL The now on'E X Pisjuat F,(x.p)-(I1{X~;lJ;;'
192
VECTOR FIELDS AND DYNAMICAL SYSTEMS
For another more "modem" proof of 4.1.5 see Box 4.1C. This proof has a technical advantage: it works easily for other types of differentiahility on X on F,. such as Holder or Sobolev differentiability; see Ehin and Marsden (1970] for details. The mapping F gives a locally unique integral curve e. for each u E ~). and for each 1 E I. F, = FI(uo X (t}) maps ~) to some other set. It is convenient to think of each point u being allowed to .. flow for time t" along the integral curve c.. (see Fig. 4.1.2 and our opening motivation). This is a picture of a Uo "nowmg." and the system (Uo• a. F) is a local flow of X. or flow box. The analogous situation on a manifold is given by the following.
Figure 4.1.2
4.1.10 DeflnlUon. Lei M be a manifold and X a r
~
(i)
(ii) (iii)
I. A
cr tleelor field on
M
f10w box of X al me M is a triple (Uo• a. F). where
Uo eM is open. me Uo. and a E R. where a> 0 or a = + 00; F: Uo X I" -+ M is of class cr. where la = (- a. a); for each u E Uo. eu : I" curve of X al u;
-+
M defined by eu(t) = F(u./) is an integral
(iv) if F,: Uo -+ M is defined by F,(u) = F(u.n. Ihen for open. and F, is a C' diffeomorphism onto its image.
1E
la. F,(Uo) is
Before proving the existence of a flow box. it is convenient first to establish the following. which concerns uniqueness.
4.1.11
Prop08lUon. Suppose c. and c 2 are two integral curves of X at
me M. Then c. = c 2 on the intersection of their domains.
VECTOR FIELDS AND FLOWS
193
Proof. This does not follow at once from 4.1.5 for c, and C2 may lie in different charts. (Indeed. if the manifold is not Hausdorff. Exercise 4. t L shows that this proposition is false.) Suppose c,: I, - AI and c;: 12 . - M. Let I = I, n 12, and let K = (III E I and c,(t) = C,2(1»; K is closed since M is Hausdorff. We will now show that K is open. From 4.1.5. K contains some neighborhood of O. For IE K consider cl and c~. where c'(s) = c( I + of). Then c: and c~ are integral curves at c.(t) = c2(1). Again by 4.1.5 they agree on some neighborhood of O. Thus some neighborhood of I lies in K. and so K is open. Since I is connected, K"" I. • The next proposition gives elementary properties of now boxes.
4.1.12 Prop08It1on. Suppose (Uo• a, F) is a Iriple .,alis/i'ing (i), (ii), lind (iii) 0/2.1.3. Then lor I, s, and I + s E la' we have and
Fo is Ihe identUv map.
Moreover, i/ U, - F,(Uo) and U, n Uo *0, Ihen F,I U., n Uo: U _/ n Uo Uo n U, is a diffeomorphism and ils inverse is F _,I UO n u,.
Proof.
F,+.(u)" COl(1 + s), where cu is the integral curve defined by Fat u. But d(t) = F,(F,(u» = F,(cu(s» is the integral curve through cues) and /(/) -'C,,(I + s) is also an integral curve at cu(s). Hence by 4.1.11 we have F,( F,(u» - C.,( 1+ s) .. F,+.( u). For F, •• = F. 0 F,. merely note that F, • • = F,+,-F,oF,. Since C.,(/) is a curve at u, c,,(O)=u, so f;, is the identity. Finally, the Jast statement is a consequence of F, 0 F / = F .. , 0 F, = identity. Note, however, that F,(Uo)n Uo .. 0 can occur. •
Now we are ready to state and prove the existence and uniqueness of flow boxes. (E%i8~ _ U.~ of Flo .., Bo.'(es). C' veclor field on a ""''Ii/old M. For each m E M there is a flow m. Suppose (Uo• a. F), (Uo• a', F) are two flow boxes at m E M. F are equal on (UO n Uo)x(/.. n I ... ).
4.1.13 Theorem
Lei X be a box 0/ X at Then F and .
Proof (Uniqueness ). Again we emphasize that this does not follow at once
from 4.1.5, for Uo;Uo need not be chart domains. However, for each u E uon Uo we have FI{u}X I .. F'l{u}X I, where I -I.. n I .... This follows from 4.1.11 and 4.1.10 (iii). Hence F= F' on (UonUo)x I. (Exislence). Let (U, cp) be a chart in M with m E U. It is enough to establish the result in cp( U) by means of the local representation. That is, let (Uci. a, F') be a flow box of i. the local representative of X, at cp(m) as.
194
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Biven by 4. J.5. with
and let
F: UoX!" ..... M; Since F is continuous, there is abE (0, a) c R and ~, c Uo open. with mE Vo. such that F(Vo x I h ) c Uo. We contend that h. F) is a now box at m (where F is understood as the restriction of F to x/,.). Parts (i) and (ii) of 4.1.10 follow by construction and (iii) follows from the n:marks following 4.1.4 on the local representation. To prove (iv). note Iha. for t E Ih' F, has a C' inverse, namely. F _I as V, () Uo = v,. It follows thai /';1 is open. And. since F, and F I are both of class C. F, is a C' diffeomorphism. •
(v.,. v.,
v.,)
As usual. there is an analogous result for time- (or parameter-) dependent vector fields. The following result. called the "straightening out theorem." shows that near a point m that is not a critical point. that is. X( m) "" O. the now can be modified by a change of variables so that the integral curves oecome straight lines.
••1.14 Theorem.
Let X he a vector field on a manifold M and suppose at
me M, X( m) ... O. Then there is a local chart (U, cp) with m E U such that
(i) cp(U) = V x IcE X R, VeE. open, and 1 = (- a, a) c R. a > 0; (ii) cp-ll{v}X I: I ..... M is an integral curve of X at cp-I(V,O),for aI/v E V; (iii) the local representative K has the form K(y, t) = (y,t; 0,1).
Proof. Since the result is local. by taking any initial coordinate chart. it suffices to prove the result in E. We can arrange things so that we are working near 0 E E and X(O) = (0.1) E E = E1eR where EI is a complement to the span of X(O). Letting (~,. h, F) be a now box for X at 0 where Uo = VO X( - e. e) and Vo is open in E •. define
But Dfo{O.O) = Identity since
aF,(O,o)
at
I,-0 -
X(O)
=
(0,1)
and
Fo = Identity.
VECTOR FIELDS AND FLOWS
195
By the inverse mapping th~rem there are open neighhorhoods V X 10 c Vo X 1b and U- f(V X la) of (0,0) such thatfX I,,: V X 10 -+ U is a difrcomorphism. Then, I: U -+ V X la can serve as chart for (i). Notice that c ... /I{y}X I: 1-+ U is the integral curve of X through (y.O) for all y E V. thus p'roving (ii). Finally. the expression of the vector field X in this is D,I(y./)·X(f(y.I)I=D/ l(c(t)I·c'(I)= local chart given by (f- 1 0 c)'( I) = (0. I). since (f- 1 0 c)( I) = (y. I). thus proving (iii). •
",IV
,I
Now we turn our attention from local nows to global considerations. These ideas center on considering the now of a vector rield as a whole. extended as far as possible in the I-variable.
4.1.15 Definition. Given a manifold M and a tle('tor field X on M. let "i)x c M X R be the .ret of (m. I) E M X R such Ihat there is an integral ('urtJe
c: 1-+ M of X at m with (E I. The vector field X is complete if6fJx = M xR. Also. a poinl mE M is called (J complfle. where (J = +, -. or ±. if ('ilx n({m}XR) contains all (m. t) for I > 0,1 < 0,' or IE R. respeclively. Lei T': (resp. T;) denole Ihe sup (resp. in/) of Ihe limes of exiSlence of Ihe inlegral curves through m: T': (resp. T';;) is called Ihe positiw (".,me) life,u.ofm.
Thus. X is complete iff each integral curve can be extended so that its domain becomes ( - 00.(0): i.e.• T': - 00 and T; - - 00 for all m E M.
4.1.11 Exempt... A. For M - R 2. let X be the constant vector field. whose principal part is (0,1). Then X is complete since the integral curve of X through (x, y) is' .... (x.y+t). B. On M = R 2 \{O}. the same vector field is not complete since the integral curve of X through (0. - I) cannot be extended beyond' = I; in fact as I -+ I thiS. integral curve tends to (0.0). Thus ~; _I) = I. while ~o: -1)-00.
e.
On R consider the vector field X(x) = I + x 2 • This is not complete since the integral curve c with c(O) - 0 is c( (I) - tan (I and thus it cannot be continuously extended beyond - .,,/2 and .,,/2; i.e.• To t = ± .,,/2. •
4.1.17
P~ltIon.
Let M be a manifold and X
E
'X'( M), r ~ l. Then
(i) 6j)x::> M x {O}; (ii) 6fJx is open in M X R; (iii) Ihere is a unique C' mapping Fx: oDx -+ M such ,hal Ihe mapping I .... Fx(m,/) is an inlegral curve al m. for all mE M. (iv) for(m.t)E 6j)x.lhe pair (Fx(m. I).S)E oTJx'iff(m. (+ s)E "i)x: in this CtUe Fx (";, t + s) - Fx(Fx(m, I). s).
196
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Parts (i) and (ti) follow at once from the flow box existence theorem. and (iii) by the uniqueness of integral curves: "Dx is the union of all now box domains Uo x I" and Fx is the unique extension of the now box maps to UDx . Thus (iv) is a reformulation of the first part of 4.1.12. • Proof.
Thus. if X is complete, (M. 00. Fx) is a flow box. 4.1.18 DeflnlUon. Let M be a manifold and X E <X( M), r ~ 1. Then the mapping Fx is called the ;"tegl'Gl of X. and the curoe t ..... Fx (m, t) is called the mtIXimtII ;"tegrwl CfIrW of X at m. In case X is complete. Fx is called the /10M' ofX.
Thus. if X is complete with now F. then the set {F,I t E R} is a group of diffeomorphisms on M. sometimes called a one - parameter group of diffeomorphisms.t For incomplete flows. (iv) says that F, 0 F. = F,+ .• wherever it is defined. Note that F,(m) is defined for t E ]T'; • T,;! (. The reader should write out similar definitions for the time-dependent case and note that the lifetimes depend on the starting time to' Next we describe some basic criteria that ensures completeness of a vector field. 4.1.19 ProposlUon. Let X be C. where r ~ 1. Let c(t) be a maximal integral curoe of X such that for every finite open interoal ]a. b( in the domain 14(0)' T.·;O)( of c. c(]a. b() lies in a compact subset of M. Then C' is defined for alitER. Proof. It suffices to show that a E I. bEl. where I is the interval of definition of c. Let ttl E ]a, b(. ttl .... b. By compactness we can assume some subsequence c(l,,) converges. say. to a point x in M. Since the domain of the flow is open. it contains a neighborhood of (x,O). So there are e> 0 and T > 0 such that integral curves starting at points (such as c(l",) for large k I closer than f! to x persist for a time longer than T. This serves to extend c to a time greater than b, so bel since c is maximal. Similarly, a E I. •
The support of a vector field X on a manifold M is the closure of the set (1ft E MIX(m). O}.
4.1.20 Corollary. A C vector field with compact support on a manifold M is complete. In particular. a C' vector field on a compact manifold is complete. Since F" - (F,)" (the ntb power), the notation F' is sometimes convenient and is used where we use F,.
t
VECTORFIELOSANOFLOWS
197
Completeness corresponds to well-defined dynamics persisting eternally. In some circumstances (shock waves in fluids and solids. singularities in general relativity. etc.) one has to live with incompleteness or overcome it in some other way. Because of its importance we give two additional criteria. In the first result we use the notation X[ fl == df· X for the derivative of f in the direction X. Here 1: E - Rand df stands for the derivative map. In standard coordinates on R".
af af ) df(x) ... ( - , •...• ax" ax
and
X(f] ==
n / af LX -,' ax
/ _,
4.1.21 Proposition. Suppose X is a c k vector field on E. k ~ I. and f: E - R is a C' proper map (that is. the inverse images of compact sefs are compact). Suppose there are constants K. L IX(f](m)1 ~ Klf(m)1
~
0 such that
+ L forall mE E.
Then the flow of X is complele. Proof. that
From the chain rule we have (a/at)f(F,(m»== X[fJ(F,(m», so
f{F,(m»)- f(m) ;'l'X(f](~(m» dT. o Applying the hypothesis and Gronwall's inequality we see that 1f(F,(m»1 is bounded on any finite I-interval. so as f is proper, F,(m) lies in a compact set. Hence 4.1.19 applies. • Note that the same result holds if we replace "properness" by "inverse images of compact sets are bounded" and assume X has a uniform existence time on each bounded set. This version is useful in many infinite dimensional examples. The next theorem gives a completeness criterion in terms of the norm' on a Banach space.
4.1.22 ProposlUon. Let E be a Banach space and X a C' vector field on E. r ~ 1. Let" be any integral curve of X. Assume IIX( ,,(t)~1 is bounded on , finite t-intervals. Then the flow of X is complete. Proof. have
Suppose IIX(,,(t)~1 < A for tEla, b[ and let
"''(In)-''(I,,,)I~j'''lfJ'(t)ldl= I"
In
~ h. For
tn
< I", we
j'm Il X(,,(t»IIc/I< AI/",-I.I. 'n
198
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Hence G( t n) is a Cauchy sequence and therefore. converges. Now argue as in 4.1.19. • The following is a typical application. 4.1.23 IXllmple.
Let E be a Banach space. Suppose X(x) "" .... ·x + B(x).
where .... is a bounded linear operator of E to E and B is sublinear; i.e.• B: E .... E is C' with r;;. I and satisfies IIB(x~11Iit Kllxll+ L for constants K and. L. We shall show that X is complete by using 4.1.22 (In RIO. 4.1.21 can also be used withf(x) -lIxIl 2 ). Let x(t) be an integral curve of X. Then x(t) == x(O)+
fo'( .... ·x(s)+ B(x(s») dx.
Hence
IIx(t)lIlIitllx(O)II+ 1'(11 .... 11+ K)lIx(s)lItb + Lt. D
By Gronwall's inequality.
IIx(t)II'" (LT+llx(O)U)eUIAII+KI'. Hence x(t) and so X(x(t» remain bounded on bounded t-intervals.
•
BOX "-1A PRODUCT FORMULAS
A result of some importance in both theoretical and numerical work concerns writing a flow in terms of iterates of a known mapping. Let X E ~(M) with flow F, (maximally extended). Let K,(x) he a given map defined in some open set of [0. co[ X M containing {O} X M and taking values in M. and assume that (i) 'Ko(x)-x
(ii) K.(x) is C' in the "algorithm."
I
with derivative continuous in (e, .t). We call K '
Let X be a C' vector field. r> I. Assume that the algorithm K,(x) is ,....,., with X in the sense that X(x)-
a
a,K,(x)J.-o
VECTOR FIELDS AND FLOWS
199
Then, if (I, x) is in Ihe domain of F,(x), K,i,,(x) is defined for n sufficienlly large and converges 10 F, (x) as n --> 00 . Conversely.. if K,i,,(x) is defined and converges for 0 Et lEtT, then (T. x) is in Ihe domain of F and the limit is F,(x).
In the following proof the standard notation O( x ). x E R is used for any continuous function in a neighborhood of the origin such thaI O( x)/ x is bounded. Recall from Section 2.1 that o( x) denotes a. continuous function in a neighborhood of the origin satisfying limx_oo(x)/x - O. Proof. First, we prove that convergence holds locally. We begin by showing that for any x o, the iterates K,i,,(x o ) are defined if I is sufficiently small. Indeed, on a neighborhood of x o' K,(x) == x + O(e), so if K!/j(x) is defined for x in a neighborhood of x(). for j = 1,00" n -I, then
+ ... +(K,/,,(x)-x) ==O(t/n}+ .. · +O(I/n) =
OCt}
This is small, independent of n for I sufficiently small; so. inductively, K,i,,(x) is defined and remains in a neighborhood of -"0 for x near xo' Let P be a local Lipschitz constant for X so that 1IF,(x)- F,(yXI ~ e~ltlllx - yll. Now write
200
VECTOR FIELDS AND DYNAMICAL SYSTEMS
wherey.. " K,~,,(x). Thus 1IF,(x)-K,i,,(x)lI.;;
E" elll,l(n-k)lnllF,ln(Yn~kl)-K,/nCVn
~ 1)11
.. -I
..;nelll'lo(l/n)-O
as n ..... oo
since F.(y)- K,(y) - o(e) by the consistency hypothesis. Now suppose F,(x) is defined for 0.;; I .;; T. We shall show K,i,,(x) converges to F,(x). By the foregoing proof and compactness. if N is large enough, F,/N -lim n _ oa K,inN uniformly on a neighborhood of the curve t>-+ F,(x). Thus, for 0.;; I';; T,
By uniformity in t. Fr(x)= limKL(x). I
~oo
Conversely. let K,in{x) converge to a curve c(t).O~/~T. Let S = (t IF,( x) is defined and c( I) = F,( x )}. From the local result. S is a nonempty open set. Let E S.l k - t. Thus F,,(x) converges to c(t). so by local existence theory. F,(x) is defined, and by continuity, F,(x)= C(/). Hence S= [0. T) and the proof is complete. •
'4
4.1.25 Corollary. LeI X. Y E 'X,(M) wilhflows F, and G,. Let S, be + Y. Then for x E M.
the flow of X
The left-hand side is defined iff the right-hand side is. This follows from 4.1.24 by setting K,(x) - (F. 0 G,)(x). for example. for n X n matrices A and B. 4.1.25 yields the classical formula e(A+8)= lim (e A1 "e 8/,,)". n-oo
To see this. define for any n X n matrix C a vector field Xc E 'X,(R") by XC< x) - Cx. Since Xc is linear in C and has now F,(.r) - e'c.r. the formula follows from 4.1.25 by letting I = I. The topic of this box will continue in Box 4.20. The foregoing proofs were inspired by Nelson (1969) and Chorin et al. (1978).
VECTOR FIELDS AND FLOWS
201
BOX 4.18 INVARIANT SETS If X is a smooth vector field on a manifold M and N c M is a submanifold. the now of X will leave N invariant (as a set) ifr X is tangent to N. If N is not a submanifold (e.g.• N is an open subset together with a non-smooth boundary) the situation is not so simple; however. ·for this there is a nice criterion going hack to Nagumo [1942]. Our proof follows Brezis (1970).
4.1.21 Theorem. Let X be a locally Upschitz vector field on an open set U c E. where E is a Bana-:h space. Let G cUbe relatively closed and set d(x. G) = infOlx - Yilly E G}. The following are equivalent: Iimlr~o(d(x+hX(x).G)/h)=O locally uniformly in xEG (or just pointwise if E == R n); and . (ii) if x(t) is an integral curve of X starting in G. then xU) E G for all t'~ 0 in the domain of x(·).
(i)
Note that x(t) need not lie in G for t.s; 0; so G is only + invariant. (We remark that if X is only continuous the the~rem fails.) We give the proof assuming E = R n for simplicity. Proof. Assume (ii) holds. Setting x(t) = F,(x), where F, is the now of X. we get
d(x+hX(x).G).s;lIx(h)-x-hX(x)II=lhl II x(h) h - x -X(x) II
from which (i) foIlows. Now assume (i). It suffices to show x(t)E G for small t. Near x'" 'x(O). say on a ball of radius r. we have
and
We can ~ssume 1IF,(x)-xll
202
VECTOR FIELDS AND DYNAMICAL SYSTEMS
needed here.) Thus,IIY, - xII < r. For small h,IIF"y, - xII < r, so
cp(1 + h)" 1IF,+,,(x)- F"y,II+IIF,,(y,)- y, - hX(y,)1I + d(y, + hX(y,).G) "eK"IIY, - F,xll+IIF,,(y,)- y, - hX(y,)II+ d(y, + hX(y,). G) or
cp(1 + hl- cp(l) " ( e K: -I )CP(I)+ II
FIr<'l~)- Y,
- X( y,)11
+ d(y, + hX(y,).G)/h. Hence lim sup cp(l+h)-cp(l) .r;;Kcp(l) "lO h As in Gronwall's inequality, we may conclude that
so cp(t) = O. • 4.1.27 Example. Let X he a smooth vector field on R n. let g: R n --0 R he smooth. and let ~ E R he a regular vullie for g. so g I( ~) is a submanifold. Let G = f{ I( 1- 00. ")) and suppose that on g 1(,,). (X. grad g) ,,0. Then G is + invariant under F, as may be seen by using 4.1.26. This result has been generalized to the case where fJG might not be smooth by Bony (1969). See also Redheffer (1972) and Martin (1973). Related references are Yorke (1967). Hartman (1972). and Crandall (1972).
BOX 4.1C A SECOND PROOF OF THE EXISTENCE AND UNIQUENESS OF FLOW BOXES We shall now give an alternative "modern" proof of Theorems 4.1.5 and 4.1.13, namely. if X E 'X'( M). k ~ I. then for each mE M there exists a unique c' now box at m. The basic idea is due to Robbin (1961) altbouah similar alternative proofs were simultaneously
VECTOR FIELDS AND FLOWS
, 203
discovered by Abraham and Pugh [unpublished) and Marsden [1968b, p. 36S). The present exposition follows Robbin (1968) and Ebin arid Marsden [1'970).
SUp 1. Existence and Uniqueness of Integral Curves for Vector fields.
c·
Proof. Working in a local chart, we may assume that X: Dr(O) ...... E, where Dr(O) the open disk at the origin of radius r in the Banach space E. Let U = Dr/ 2 (O), 1 = [ - I, I) and define
41: R x CJ(I, U) ...... C°(l, E) by
4>(s,y)(t)
=
dy di(t)-sX(y(t».
where Ci(l, E) is the Banach space of Ci-maps of 1 into E, endowed with the II'II;-norm (see Box 2.4.B), qU, E) = (f E CiU, E) 1/(0) = O} is a closed subspace of CiU. E) and Cj(l, U) .... (f E CjU, E)I/U) C U} is open in CjU, E). We first show that 41 is a C·-map. The map dldt: CJu, E) -+ C°(l, E) is clearly linear and is continuous since IIdldtll ~ I. Moreover, if dyldt = on I, then y is constant and since yeO) = 0, it follows y =- 0; i.e., d I dt is injective. Also, given 8 E C°(l, E), yet) = Iri8(s) ds defines an element of CJ(l. E) with dYldt = 8, i.e.• dldt is a Banach space isomorphism from CJu, E) to C°(l. E). From these remarks and the O-Iemma (2.4.I'S), it follows that 4> is a C· map. Moreover, Dy4>(O,O) = dldt is an isomorphism of CJU. E) to Co( 1. E). Since 4)(0,0) = o. by the implicit function theorem there is an E> such that 4>(E. y) = 0 has a unique solution y.(t) in CJ(l,U). The unique integral curve sought is y(t)=y,(lIE), -f~
°
°
t
~f• •
The same argument also works in the time-dependent case. It also shows that y varies continuously with X.
Step 1. Proof.
setting
The local flow of a C" vector field X is C k . First, suppose k = I. Modify the definition of
+: R x U x CJ(I, U) -+ COU, E),
c)l
in step I by
+(s, x. y )(t) = y(t)-sX(x + y(t ».
°+
As in step I.
is a C· map and Dy+(O,O.O) is an isomorphism, so can be locally ~lved for y giving a map H,: U-+ cJ(I.U).e> 0. The local flow is F(x.t)=x+ H,(x)(tlr), as in step I. By' 2.4.17 (differentiability of the evaluation map), F is C·. +( E. x, y) ..
204
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Next we prove the result for k ~ 2. Considet the Banach space F-Ck-'(cl(U),E) and the map "'x:F .... F; 1J ..... Xo1J. This map is C' by the O-lemma (remarks following 2.4.18). Regarding "'x as a vector field on F, it has a unique C' integral curve 1J, with 1Jo ... identity. by Step I. This integral curve is the local flow of X and is C k - . since it lies in F. Since k ~ 2, 11, is at least C' and so one sees that D1J, =", satisfies .,Idt ... DX( 11;)''',. so by Step I again, II, lies in C k - '. Hence 11, is Cit. • The following is a useful alternative argument for proving the result for k - I from that for k ~ 2. For k -I. let X" .... X in C', where X" are C 2 • By the above, the flows of X" are C 2 and by Step I. converge uniformly i.e. in Co, to the now of X. From the equations for DlI,", we likewise see that DTI," converges uniformly to the solution of dM,ldt,. DX(TI,)·",. "0" identity. It foUows by elementary analysis (see Exercise 2.41 or Marsden [1974a, p. 109]) that TI, is C· and DTI, = II,. • This proof works with minor modifications on manifolds with vector fields and flows of Sobolev class H S or Holder class C k ~"; see Ebin and Marsden (1970) and Bourguignon and Brezis [1974]. In fact the foregoing proof works in any function spaces for which the O-Iemma can be proved. Abstract axioms guaranteeing this are given in Palais [1968].
Exercl... Find an explicit formula for the flow F,: R 2 .... R2 of the harmonic oscillator equation .f+ ",2X "" O. '" ERa constant. 4.lB Show that if (UO. a. F) is a flow box for X. then (Uo• a. F _) is a flow box for - X. where F _Cu. t) == F(u. - t) and (- X)(m) = -(X(m». 4.1 C Show that the integral curves of a C' vector field X on an n-manifold can be defined locally in the neighborhood of a point where X is nonzero by n equations ,,;(m, t) = cj == constant. i .. l •... ,n. Such a system of equations is called a local complete system 0/ integrals. (Hint: Use the straightening-out theorem.) 4.1 D Prove the foUowing generalization of Gronwall's inequality. Suppose vet) ~ 0 satisfies v(t)" c + fJ Ip(s)lv(s) tis, c ~ O. Then 4.1A
v(t) < cexP({IP(s)1 tis ). 1 I",.
this to generalize 4.1.23
,- -iilX.
In
allow A to be a time-dependent
VECTOR FIELDS AND FLOWS
4.1H
205 '
now
(VariatiON of COIUltmlllomtIIIa). Let F, - e'x be the of a linear vector field X on E. Show that the solution of the equation
.t- X(x)+ I(x) with initial condition Xo satisfies tbe integral equation
4.1 F
Let F( m. I) be a efta mal>ping of R x M to M such that F, •• - F, 0 F, and Fo - identity (wbere F,(m) - F(m, Show that there is a unique efta vector field X wbose now is F.
I».
4.10 Let a(l) be an integral curve of a vector field X and let g: M .... R. Let T(I) satisfy T'(I)= g(a(T(I))). Then show t .... CJ(T(t» is an inte&ral curve of gX. Show by example that even if X is complete. gX need not be. 4.1H
(i) (Gradienl Flows) letf: R" -+R be e l and let X= (iJf/ax l ••••• ilf/ ax") be the gradient of f. Let F be the now of X. Show that f(F,(x» ~ f(~(x» if I ~ s. (ii) Use (i) to find a vector field X on R" such that X(O) = o. X'(O) = 0, yet 0 is globally attracting; i.e.. every integral curve converges to 0 as I -+ 00.
4.11
let X= y 2 a/ax and Y=x 2 il/ay. Show that X and Yare complete on R 2 but X + Y is not. (ii) Prove the following theorem: Let H be a Hilbert space and let X and Y be locally Lipschitz vector fields tbat satisfy the following: (a) X and Yare bounded and Lipschitz on bounded sets; (b) there is a constant fJ ~ 0 sucb tbat (i)
(Y(x). x) < fJllxlf' (c)
for all
x EX
there is a locally Lipschitz monotone increasing function C(I) > O. I ~ 0 such that JOOdx/c(x)'" + 00 arid if X(I) is an integral curve of X.
!lIx(t)1I < c(lIx(t)lI) Then X. Y and X + Yare positively complete. NOle: One may assume IIX(xo~I'" c(lIxolI) in (c) instead of (d/dl~I%(t~1 < cOI%(I~I>. (Hinl. Find a differential inequality
206
VECTOR FIELDS AND DYNAMICAL SYSTEMS
for t
4.1J
+ fo'C(f(S»ds.
IE
[0.11
Then /(/)" r(/) for IE [0, T[.) Prove the following result on the convergence 0/ flows: Let X .. be locally Lipschitz vector fields on M for 0 in some' topological space. Suppose the Lipschitz constants of X" are locally bounded as 0"'" 00 and X .. ..... X .... locally uniformly. Let (.( I) be an integral curve of X .... 0 " 1 < T and e > O. Then the integral curves C.(/) of X. with c..(O) ... c(O) are defined for IE [0. T - e) for 0 sufficiently close to 00 and C..(/) ..... c(/) uniformly in 1 E [0. T - e) as 0"'" Go. If the flows are complete, P,'''''' F, locally uniformly. (The vector fields may be time dependent if the estimates are locally I-uniform.) (Hin/: Show that
IIc.. (/)-c(/)II" Kl'lIc.( .,)-c( ")lId., +l'IIX.(c( .,)- XII (c( "»lId., o 0 • and conclude from Gronwall's inequality that c.( I) ..... c""( t) for since the second term ..... O. This estimate shows that C.. (/) exists as long as c(/) does on any compact subinterval of [0. T[.) 4.1K Prove that the C flow of a C+' vector field is a C' function of the vector field by utilizing Box 4.1C. (Caulion. It is known that the C" flow of a e" vector field cannot be a C· function of the vector field; see Ebin and Marsden (1970] for the explanation and further references). 4.1 L (Nonunique inlf!gral curves on non-Hausdorff mani/olds). Let M be the line with two origins (see Exercise 3.5H) and consider the vector field X: M ..... TM defined by X[x. i] = x. i = 1.2: here Ix. i] denotes a point of the quotient manifold M. Show that through every point other than (0.0] and [0.1], there are exactly two integral curves of X. Show that X is complete. (Hin/: The two distinct integral curves pass respectively through (0,0) and [0,1).) 0 ..... 00
4.1 M Give another proof of Theorem 4.1.5 using Exercise 2.5J.·
· VECTOR FIELDS AS DIFFERENTIAL OPERA TORS
207
4.2 VECTOR FIELDS AS DIFFERENTIAL OPERATORS
In the previous section vector fields were studied from the point of view of dynamics-the flows they generate. Before continuing the development of dynamics, we shall treat some of the algebraic aspects of vector fields. The specific goal of the section is the development of the Lie derivative of functions and vector fields and its relationship with flows. One important feature is the behavior of the constructions under mappings. The operations should be as natural or covariant as possible when subjected to a mapping. Let. us begin then with a discussion of the action of mappings on functions and vector fields. First. recall some notation. Let' C(M. F) denote the space of C maps f: M - F, where F.is a Banach space. and let ' X '( M) denote the space of C vector fields on M. Both are vector spaces with the obvious operations of addition and scalar multiplication. For brevity we write '!i'(M)=C"'(M.R), '!i"(M)=C(M.R) and ~X(M)= ~OO(M). Note that '!i"( M) has an algebrastructwe; that is. forf. g E IS'(M) . the product fg defined by (fg)(m) = f(m)g(m) obeys the usual algebraic properties of a product such as fg = gf and f( g + h) = fg + fh. 4.2.1 Definition.. (i) Let cp: M - N be a C' mapping of manifnld.~ and f E IS'(N). Define the puIl-lNIek off by cp by cp.f =
f
0
cp E IS' ( M).
If cp is a C' diffeomorphism and X E CX'( M). the puS"- forwtlrd of cp is defined by
(ii)
X
~y
CP.X= TIP, Xo cp-I E CX'(N). 0
The reason for the names is shown in Figure 4.2.1. Consider local charts (V,X). x: V-V'cE on M and (V.I{I). I{I: V -+ V'c Fon N, and let (TX 0 X 0 X-I)(u) = (u, X(u». where X: V' -+ E is the local representative of X. Then from the chain rule and the definition of push-forward. the local representative of CP. X is
where v = (I{I cp X-I)(u). The different point of evaluation on each side of the equation corresponds to the necessity of having cp- I in the definition. If M and N are finite dimensional, Xi are local coordinates on M and Xi local coordinates on N. the preceding formula gives the components of CP.X by 0
0
(cp.X)i(X) where X- cp(x).
=
iJcpi . (x)X'(x) iJx
-I
208
VECTOR FIELDS AND DYNAMICAL SYSTEMS
or- =
pull back
Figure 4.2.1
We can interchange "pull-back" and "push-forward" by changing qJ to I, i.e., defining qJ* (resp. qJ*) by qJ* ... (qJ-I)*(resp. qJ* = (qJ-I ).). Thus the push-forward of a function f on M is qJ. f = f 0 qJ - 1 and the pull-back of a vector field Yon N is qJ·Y-(TqJ)-loY0qJ. Notice that qJ must be a diffeomorphism in order that the pull-back and push-forward operations make sense, the only exception being pull-back of functions. Thus vector fields can only be pulled back and pushed forward by diffeomorphisms. ' However, even when qJ is not a diffeomorphism we can talk about qJ-related vector fields as follows.
qJ-
4.2.2 DeflnlUon. Let qJ: M -+ N be a C' mapping of manifolds. The vector fi~1ds XE'X,-I(M) and YE~X'-I(N) are called cp-n/l"ed, denoted
"
X-
Y-
Y,
if TqJ ° X = yo qJ.
Note that if qJ is a diffeomorphism and X and Yare qJ-related, then In general however, X can be qJ-related to more than one vector
qJ. X.
VECTOR FIELDS AS DIFFERENTIAL OPERA TORS
209
field on N. cp-relatedness means that the following diagram commutes: TfP TM---'TN X t fP I Y M N
4.2.3 PrOposition. (i) Pull-back and push-for....ard are linear maps. and . 'I' fP*(fg) = (fP*f)( fP*g), fP.(fg) = (fP*f)( fP*g)· Moreover. if X, - Y,. i = 1.2,
'"
and a, b e R. then aX, + bX2 - aY, + bY2 • (ii) For fP: M -+ Nand 1/1: N -+ P, we have ( 1/1 0 fP )* = fP* 0 1/1* and ( 1/1 0 fP ). = 1/1* 0 fP *.
'"
'"
IfXe~Y..(M), Ye'X(N),Ze'X(P),X- YandY-Z.thenX ",. -
'" Z.
In, this proposition it is understood that all maps are diffeomorphisms with the exception of the pull-back of functions and the relatedness of vector fields. Proof.
'" Y"
X, -
(i)
This consists of straightforward verifications. Por example. if
i =-1,2, then
TfP o(aX,
+ bX2 ) ... aTfP 0 X. + bTfP 0 X2 - aY. 0 fP +
bY2 0 fP. i.e., aX. + bX2 ! aY. + bY2 • (ii) These relations on functions are simple consequences of the definition, and the ones on 'X.(P) and 'X.(M) are proved in the following way using the chain rule: T( 1/1 0 fP) 0
x ... TI/I 0 TfP
0
X = TI/I 0 yo fP
=
Z 01/1 0 fP.
•
In this development we can replace
4.2.4 ProposlUon. Let fP: M -+ N be a C'-mapping of manifolds and X e <X'(M) and Y e 'X.'(N). Let F,x and F,Y denote the flows of X and Y. respectively. Then X - Y iff fP 0 F,x = F,Y 0 fP. In particular. if fP is a diffeomorphism, then Y == fP * X iff the flow of Y is fP 0 F,x 0 fP - I.
'"
Proof. Taking the time derivative of the relation (fP F,x)(m) = (F,Y 0 fP)(m), for m eM, using the chain rule and definition of the flow, we 0
210
VECTOR FIELDS AND DYNAMICAL SYSTEMS
get Ttp ( (Ttp
0
aF.X(m») . aF. Y I
a,
"" -t,- (tp( m»,
i.e.
X 0 F,X)( m ) - (y 0 F, Yo tp )( m) .. (y 0 tp a F,X)( m ),
which is equivalent to Ttp a X = yo tp. Conversely, if this relation is satisfied, let C(I) - F,x(m) denote the integral curve of X through me M. Then d( tp. C)(I) _ Ttp( «(I) ) ... Ttp( X(C(I»)
dl
dl
== Y« tp a C )(1»
says that tp. C is the integral curve of Y through tp( c(O» = tp( m). By uniqueness' of integral curves, we get (tp • F,X)( m) = (tp a c)( I) = F, Y( tp( m»•
•
We call tp. F,. tp-I the push-forward of F, by tp since it is the natural way to construct a diffeomorphism on N out of one on M. See Fig. 4.2.2. Thus, 4.2.4 says that the flow of the push-forward of a veclor field is Ihe push forward of ils flow. Next we define how vector fields operate on functions. This is done by means of the directional derivative. Let f: M - R, so Tf: TM - TR == R xR. Recall that a tangent vector to R at a base point). E R is a pair ()., ,,), the number" being the principal part. Thus we can write Tf acting on a vector., E T", M in the form
Tf·"= (J(m),df(m)·.,). F,
"o/'r . q'
I
conjugation ---......
c-inlqnl curve of X
'" • c - integral curve
or",.x
F..... 4.2.2
VECTOR FIELDS AS DIFFERENTIAL OPERA TORS
211
This defines, for each m EM the element df(m) E T.:M. Thus dfis a Section of -r M. -a covector field, or one-form. 4.2.5 Definition. The C01.'ector field df: M..,. T* M defined called the differential of f.
thi.~
way is
For F-valued functions. f: M -+ F, where F is a Banach space. a similar definition gives df(m) E L(TmM, F) and we speak of df as an F-valued ' one-form.
Clearly if f is C'. then df is C'- I. Let us now work out df in local charts for f E ~if( M). If q1: U c M -+ VeE is a local chart for M. then the local representative of f is the map j: V -+ R defined hy i = f q1 I. The local representative of Tf is the tangent lJlap for local manifold~: 0
Tj(x, v) = (j(x), Dj(x)·v).
Thus'the local representative of df is the derivative of the local representative of f. In particular. if M is finite dimensional and local coordinates are denoted (Xl, . .. ,x"), then the l~al components of df are
( df),
=
aax~ .
The introduction of df leads to the following. 4.2.6 DefinlUon. Lei fE'!t'(M) and Xe'X' '(M).r;21:l. Define the directiolull or Lie derivative, 0/ f along X by
Lxf(m) == X(f](m) =df(m)·X(m). for any me M. Denote by X[f]=df(X) the map m EM· .. X[f)(m)E R. If fis F-valued. the same definition is used, but now X[f] is F-vahied. The local representative of X[f] in a chart.is given by the real valued function x ...... Dj(x)'X(x), where j ,and X are the local representatives of I and X. In particular, if M is finite dimensional then we have
X(f] == LxI =
t
i-I
af X'.
ax'
Evidently if I is C' and X is C,-I then Xlf) is
C'~ '.
212
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Let us observe that from the chain rule, d( f qJ) = df TqJ where qJ: r ~ l. For real-valued functions, Leibniz' rule gives 0
0
N ..... M is a C' map of manifolds,
dUg) - fdg
+ gdf·
(If f is F-valued, g is G-valued and B: F X G ..... H is a continuous bilinear map of Banach spaces, this generalizes to d(B(f, g» = B(df, g)+ B(f, dg).)
4.2.7 Propolltlon. (i) Suppose qJ: M ..... N is a diffeomorphism. Then Lx is natural with respect to push- forward by qJ. That is, for each f L".x( qJ.f) = qJ.Lxf; or the following diagram commutes: qJ •
\'t'( M)
E
'!f(M),
• \1"( N)
1L".x
Lxl ~(M)
qJ.
~(N)
(ii) Lx is IfIItIIrtII witll rnpect to rntmiOlU. That is, for V open in M andfE~(M), Lx1u(fl V )"" (Lxf)IV; or, if IV: ~(M)--+~(V) denotes
restriction to V, the following diagram commutes: IV ~(M)-~(V)
Lx!
!Lxlu
CJ(V)
CJ(M) IV
Proof.
For (i), if n E N then
L".x( qJ.f)(n)'" d(t 0 cp-I)( qJ.X)(n) = d(t 0 cp-I )(n)· (TqJ 0 X 0 qJ-l)(n) = df{ qJ-l(n»)-( X
0
qJ-l)(n) = qJ.(Lxf)(n)
Then (ii) follows from the fact that d(fl V) = (df)1 V, which is clear from the definition of d. • This proposition is readily generalized to F-valued C' functions. Since qJ. = (cp - I)., the Lie derivative is also natural with respect to pull-back by cpo This has a generalization to qJ-rl'lated vector fields as follows.
VECTOR FIELDS AS DIFFERENTIAL OPERA TORS
213
4.2.1 Proposition. Let cp: M ..... N be a C' map. X E 'X,-I( M) and
"
Y e ~'-I(N). II X - Y. then Lx(cp*f) - cp*LI'I for all / E C(N. F); i.e .• the lollowing diagram commute,,:
".
C'(N.f)---· C'(M.F)
Ly
I ".
C-I(N.F)
Proof.
For m
E
l
Lx
• C-I(M, F)
M,
Lx( cp·f)(m) = dU 0 cp)(m )·X(m) = df(cp(m»'(T",cp( X( m») = df( cp(m»· Y( cp(m» = df( Y)( cp(m»' = (cp·Lrf)(m) . • Next we show that Lx satisfies the Leibniz rule.
4.2.9 Propoaltlon. (i) The mapping Lx: C( M, F) ..... C- I ( M, F) is a tkriwltion. That is Lx is R-linear and for / E C'(M, F), g E C(M.G) and B: F X G ..... H a bilinear map. Lx(BU,
g» = B(Lx/. g)+ BU. Lx1d.
For real-valued functions, LxUg) = gLxf + fLxB. '(ii) If c is a constant function, Lxc'" O. Proof. (i) This follows from the product rule and the definition of Lx/. , Part (ii) results from the definiiion. •
,The connection between Lx f and the flow of X is as follows.
4.2.10 Propoaltlon. Let/ E C'(M, F), X E
ex' 'I(M) and suppose X has
a flow F,. Then Proof. By the chain rule, the definition of the differential of a function and the flow of a vector field.
d d dF.(m) dt(Fif)(m) = dtU°F,)(m)=df(F,(m». 'dt = df( F,(m »·X( F,(m» = df( X)( F,( m» == (Lxf)( F,( m»
= (FiLxf)( m).
•
214
VECTOR FIELDS AND DYNAMICAL SYSTEMS
As an application of the Lie derivative, we consider a partial differential equation on R" + 1 of the form
(P) {
al
.
n'
i
af
a/(X./)=i~IX(X)aXi(X,/)
f(x,O) = g(x)
for given smooth functions Xi(x), i = 1, . .. ,n, g(x), and a scalar unknown f(x, I).
4.2.11
Proposition. Suppose X = ( X l , ••• , X n) has a complete flow F,.
Then f(x,t)=g(F,(x)) is a solution of the foregoing problem (P). (See Exercise 4.2C for uniqueness.) Proof.
Thus one can solve this scalar equation by computing the orbits of X and pushing (or "dragging along") the graph of g by the flow of X; see Fig. level curves of g • F,
Flgurw 4.2.3
VECTOR FIELDS AS DIFFERENTIAL OPERATORS
215
4.2.3. These trajectories of X are called characteristics of (P). (As we shall see below. the vector field X in (P) can be time dependent.) 4.2.12 Example. Solve the partial differential equation af =(x+ y )( af _ af), at ax ay f(x. y.o) = x 2 + y2.
The vector field X(x. y) = (x + y, - x - .1') has a complete flow F,(x. y) = «x + y)t + x. -(x + y)t + y). so that the solution of the previous partial differential equation is given by Solution.
Now we tum to the question of using the Lie derivative to characterize vector fields. We will prove that any derivation on functions uniquely defines a vector field. Because of this, derivations can he (and often are) used to define vector fields. (See the introduction to Section 3.4.) In the proof we shall need to localize things in a smooth way, hence the following lemma of general utility is proved first. 4.2.13 Lemma. Let E be a C', Banach space. i.e" one whose norm is C' on E\{O}, r;;, I. Let VI be an open ball of radius r l ahout Xo and V2 an open ball of radius r2 • r l < r2 • Then there is a C' function II: E -+ R such that II is one on VI and zero outside V2 •
We call h a bump function. Later we will prove more generally that on a manifold M. if VI and V2 are two' open sets with cJ(VI ) c V2 • there is an he 'J'(M) such that h is one on VI and is zero outside V2 • Proof. By a scaling and translation. we can assume that VI and V2 are balls of radii 1 and 3 and centered at the origin. Let 8: R -+ R be given by 8(x) = {exp( -1/(I-l x I2
o.
»).
(See the remarks folJowing 2.4.15.) Now set
8 1(s ) =
t1-
8(t) dt
0000
-- 00
8( t) dt
Ixl
216
VECTOR FIELDS AND DYNAMICAL SYSTEMS
so ',(s) is a Coo function, 0 if s < -1, and 1 if s > 1. Let
so
'2 is a Coo function that is 1 if s < 1 and 0 if s > 3. Finally, let
The nonn on a real Hilbert space is Coo away from the origin. The order of differentiability of the norms of some concrete Banach spaces is also known; see Bonic and Frampton (I966J and Yamamuro (1974).
4.2.14 II",
Carol'' ' '
Let M be a manilold modeled on a C' Banach space. II
e r:,M, then there is an I e ~'( M) such that d/(m) =
11m'
.
Proof. If M .. E, SO TmE g; E, let I(x) = II",(X), a linear function on E. Then dl is constant and equals II",. The general case can be reduced to E using a local chart and a bump function as foUows. Let cp: U ..... U' c E be a local chart at m with cp( m) = 0 and such that U' contains the ball of radius 3. Let a", be the local representative of II", and let h be a bump function 1 on the ball of radius 1 and zero outside the ball of radius 2. Let i( x) = am (x) and let
1_ {(hJ)oCP, 0,
on U on M\U
It is easily verified that I is C' and d/( m) -
II",.
•
4.2.15 Propoeltlon. (i) Let M be a manifold modeled on a C' Banach space. The collection 01 operators Lx lor X e CX'(M), defined on C'+ '(M, F) and laking values in C'(M, F) lorms a real vector space and IJ(M) module with (/Lx )(g) -/(LxS), and is isomorphic to CX(M) as a real vector space and as an ~(M) module. In particular, Lx =0 iff X=O; and LfX = /Lx. (ii) Let M be any manilold.11 LxI = olor aliI e C'(U, F),lorall open subsets U 01 M, then X = O.
I'rrIof. (i) Consider the map a: X .... Lx. It is obviously R and
IJ(M)
linear; i.e.
Lx,+/x,'" Lx, + /Lx, To show that it is one-to-one, we must show that Lx = 0 implies X ... O. But if Lx/(m)'" 0, then 4j(m)X(m) == 0 for all f. Hence, 1I",(X(m»'" 0 for all II", e r:,( M) by 4.2.14. Thus X( m) .. 0 by the Hahn- Banach theorem.
VECTOR FIELDS AS DIFFERENTIAL OPERA TORS
217
(il) This has an identical proof with the only exception that one works in a local chart, so it is not necessary to extend a linear functional to the entire manifold M as in 4.2.14. Thus the condition on the differentiability of the norm of the model space of M can be dropped. •
Now we are ready to give the main result on derivations.
4.2.18 Theorem. (i) If M is finite dimensional, the collection of all derivations on iJ(M) is a real vector space isomorphic to ~x'(M). In particular, for each derivation 8 there is a unique X E ~ ( M) such that 8 = Lx. (ii) Let M be a manifold modeled on a Coo Banach space E i.e. E has a Coo norm away from the origin. The collection of all (R-linear) derivations defined on Coo(M, F) (for all Banach spaces F), furms a real vector .fpace isomorphic to ~(M).
We prove (ii) first. Let 8 be a derivation. We wish to construct X such that 8, = Lx. First of all, we note that 8 is a loca/ operator; that is, if he Coo(M, F) vanishes on a neighborhood V of m, then 8(h)(m)= o. Indeed, let g be a bump function equal to one on a neighborhood of m and zero outside V. Thus h = (1- g)h and so Proof.
8(h)(m)
=
8(1- g)(m)·h(m)+8(h)(m)(l- gem»~ = O.
(I)
If U is an open set in M, and f E Coo(U, F) define (8/uXf)(m) = 8( gf)( m), where g is a bump function equal to one on a neighborhood of m and zero outside U. By the prevlOi.1~ remark, (81 UXf)(m) is independent of g, so 81 U is well defined. For convenience we write 8 = 81 U. Let (U, 'P) be a chart on M, m E U, and f E C oo ( M, F) where 'P: U -+ U' c E; we can write, for x E U' and a '"" 'P( m), ('P.f)(x)= ('P.f)(a)+1 1 aa ('P.f)[a+t(x-a)]dt 0 t
,
= ('P.f)(a)+ foID('P.f)[a+t(x-a)].(x-a)dt.
This formula holds in some neighborhood 'P(V) of a. Hence for u E V we have (2) feu) = f(m)+ g(uH'P(u)-a), whereg E Coo(V, L(E, Applying 8 to (2) gives
F» is given by g(u) = fJD('P.f){a + t('P(u)~ a)]dt.
, 8f(m) = g(m)· (8!p )(m) = D( 'P.f)(a)· (8'P)(m)
(3)
218
VECTOR FIELDS AND DYNAMICAL SYSTEMS
since 8 was given globally, (3) is independent of the chart. Now define X on U by its local representative
X.(x) - (x,e(4p)(u». where x - cp( u) E U'. It follows that XI U is independent of the chart cp and hence X E 'X(M). Then, for IE C«J(M, F), the local representative of LxI is
Hence Lx = 8. Finally, uniqueness follows from 4.2.14. The vector derivative property was used only in establishing (I) and (3). Thus, if M is finite dimensional and e is a derivation on ~ ( M), we have as before
I(u) - I(m)+ g(u)'(4p(u)-a) -/(m)+
"
L (c,I(u)-a')8,(u). ,-,
where g, E Cj(V) and
0= (0', ... ,0"). Hence (3) becomes
8/(m) =
L" gi(m)8(4pi)(m)
i-'
and this is again independent of the chart. Now define X on U by its local representative
(x, 8( 4p')( u), ... ,e( 4p")( u» and proceed as before. • There is a difficulty with this proof for derivations mapping C .. , to C. Indeed in (2), 8 is only C if I is C+', so e need not be defined on g.
VECTOR FIELDS AS DIFFERENTIAL OPERATORS
219
The result is. however. still true in this case. hut rcquires a different argument (see Box 4.2C). For finite-dimensional manifolds. the preceding theorem provides a local basis for vector fields. If (U. "'). "': U - VcR" is a chart on M defining the· coordinate functions Xi: U --+ R. define n derivations a/ax' on ~(U) by
These derivations are linearly independent with coefficients in
(Lli~)(XJ)=f'=O
Eli-;=O. then ax
ax
i_I
~'f(U).
for if
forall j=I ..... n.
since (a/axi)x i = 8/. By 4.2.16, (a/ax i ) can be identified with vector fields on U. Moreover, if X E 'X( M) has components Xl •...• X" in the chart ",. then
Lxl=X(fJ=
E
Xi
i-I
ali ax
=
(E, _ X'~)/. ax I
i.c.
X=
EX'~. ax'
i_I
Thus the vector fields ( a/ ax I). i - I •...• n form· a local basis for the vector fields on M. It should be mentioned however that a global basis of 'X( M). i.e.• n vector fields. XI ..... X n E 'X(M) that are linearly independent over ~(M) and span 'X(M). does not exist in general. Manifolds that do admit such a global basis for ~ (M) are called parallelizable. It is straightforward to sho~ that a finite-dimensional manifold is parallelizable iff its tangent bundle is trivial. For example. it is shown in differential topology that S3 is parallelizable but S2 is not (see Hirsch (1976». This completes the discussion of the Lie derivative of functions. Turning to the Lie derivative of vector fields. let us begin with the following.
4.2.17 Propoaltlon.
II X
and Yare C' vector fields on M. then [~x. Ly) H I( M. F) to C,-I(M. F).
... Lx Ly-:- Ly Lx is a derivation mapping C 0
0
Proof. More generally, let 81 and 6z be two derivations mapping C'· I to C and C to C- I • Clearly [8 1, 6z1 = 81 06z - 6z 08 1 is linear and maps C+ I to C- I . Also. if IE C'+ I (M. F), g E C+ I (M,G). and BE L(F,G; H),
220
VECTOR FIELDS AND DYNAMICAL SYSTEMS
then
(1 •• 8z](B(f. g» ... (I. o8z)(B(f. g»-(8z o8.)(B(f. g» - 1.{B(8z(f), g)+ BU. 8z(g))}- 8z{B(8.(f). g)+ BU.I. (g))} ==
B(I.(8z(f». g)+ B(8z(f), 8.(g»+ B(I.(f). 8z( g» + BU. 1.(8z( g»)- B( 8z(8.(f». g) - B(8.(f). 8z(g» - B(8z(f),I.(g»- BU.~(8.(g)))
=
B([8 •. ~](f). g)+
BU. [8 •. ~]( g».
•
Because of 4.2.16 the following definition can he given. 4.2.18 Definition. Let M be a manifold modeled on a COO Banach space M). Then (X, Y I is the unique vector field such that L [X. Y J = and X, Y E ( Lx, L y). This vector field is also denoted Lx Yond is called the Lk tkrivatit¥ 0/ Y willi ~$pect to X, or the Lk brtJcht 0/ X tuUI Y.
ex ""(
Even though this definition is useful for Hilbert manifolds (in particular for finite-dimensional manifolds). it excludes consideration of C' vector fields on Banach manifolds modeled on nonsmooth Banach spaces, such as L' function spaces for p not even. We shall, however, establish an equivalent definition, which makes sense on any Banach manifold and works for C' vector fields. This alternative definition is based on the following result. 4.2.18 Theorem. Let M be as in 4.2.18 and X, Y have (local) flow F,. Then
E
'X,(M) and let X
d dt (F,"'Y) = F,"'(LxY) (at those points where
Proof.
F, is defined).
If t = 0 this formula becomes
ddt
I,-0
F,"'Y= Lx Y
(4)
VECTOR FIELDS AS DIFFERENTIAL OPERA TORS
221
Assuming (4).
d (P,'Y)- .1dF:+.Y=P,' i d.1_ dI ~ ~
.-0
I.-0J;Y=p,"LxY'
Thus Jhe formula in the theorem is equivalent to (4). which is proved in the following way. Both sides of (4) are clearly vector derivations. In view of 4.2.16, it suffices then to prove that both sides are equal when acting on an arbitrary function f e COO( M. F}. Now
: (p'"Y)[f](m)l._o = =
:
I.-o{ df(m).(TF,(m,F._,
0
yo F,)(m)}
:1._op'"(Y[F!,f])(m).
Using 4.2.10 and Leibniz' rule, this becomes
X[Y[f)](m)- Y[X(fJ1(m) ... [X, Y][f](m) . • Note that in the preceding proof we derived the following fact of general utility: if.,: M ..... N is a' diffeomorphism and Y E ~. (N). then for
f e6J(M),
(5)
Since the formula for LxY in Eq. (4) does not use the fact that the norm of E is Coo away from the origin. we can state the following defmition of the Lie derivative on any Banach manifold M .
••2.20 Allermdlve DefInition of Lie bnIcket. If X,Ye~'(M), r;;.1 and X has flow F" the C- I vector field LxY - IX, Y] on M defined by [X,Y)- ddl I
,-0(p,'Y)
is called the Lk deriwltiw of Y with respect to X, or the Lk brrlebt of X andY. . From the point of view of this more general definition, 4.2.19 can be rephrased as follows .
••2.21 ~. Let X, Y e 6X(M), 140; r. Then IX, Y]- LxY is the uniqUe C- I vector field on M satisfying [X. Y][f] == X[Y[fJ] - Y[X(f]] for all f e C+ I(U, F), where U is open in M.
222
VECTOR FIELDS AND DYNAMICAL SYSTEMS
The derivation approach suggests that if X, Y E 6.X.'(M) then IX, Y] might only be C,-l, since [X, Y] maps C+ 1 functions to C- I functions, and differentiates them twice. However 4.2.20 (and the coordinate expression below) show that [X, Y] is in fact c- I. We can now derive the basic properties of the Lie bracket.
4.2.22 Propoeltlon. The bracket IX, Y] on 'X(M), together with the real v«tor space structure of 'X ( M), form a Lk lIIgebra. That is, (i) (,] is R bilinear: (ii) [X, X) - 0 for all X E 'X( M): (iii) [X,[Y, Z)]+(Y,[Z, X))+[Z.[X, Y)) = 0 for all X. Y, Z ( JlICObi ilkntily).
E ~X(M)
The proof is straightforward, applying the brackets in question to an arbitrary function. Unlike 'X(M), the space 'X'(M) is not a Lie algebra since [X. YJE 'X,-I(M) for X. Y E 'X'(M). Note that (i) and (ii) imply that [X. Y] = - [Y. X). for
[X + Y. X+ YJ -0 "" [X. X)+[X. Y)+ [yo X)+ I Y. Y). Also, (iii) may be written in the following suggestive way: Lx[Y. Z) = [LxY, Z) + [y, LxZ):
i.e., Lx is a Lie bracket derivation. Strictly speaking we should not use the same symbol Lx for both definitions of Lxf and LxY. However. the meaning is generally clear from the context. The analog of 4.2.7 on the vector field level is the following.
4.2.23 PropoalUon. (i)
Let ep: M
--+
N be a diffeomorphism and X
E
'X ( M). Then Lx: 'X ( M) --+ ~ ( M) is natural wit" IYspen to p_-1ol'WtU'd by ep. That is. L".xep.Y" ep.LxY, or [ep.X, ep.Y) = ep.[X. Y), or the following diagram commutes: 'J'.
CX(M)--• CX(N)
I
LxI
L".x
CX(M)-- ~(N)
••
VECTOR FIELDS AS DIFFERENTIAL OPERATORS
223
(ii) Lx is """"'" witII rnp«t 10 rntrictiolu. ThaI is. for U c M open, we have [XI U, fIUJ- IX, fHU; or Ihe following diagram commule.f: IV ~(M)----' ~(U)
Lxl ~(M)
Proof.
any Z
we get
!L xlu ~(U)
IV
For (i),letf E "(V), vbe open in N. and tp(m) = n E V. By (5). for E ~x'(M)
fro~
«IP.Z }[f])(n)
= Z [f
° IP)( m).
4.2.21
(IP.[X, f])[f]{n) - [X, f](f°IP]{m)
... X[f(f ° IP ]](m)- f[X(f ° IP ]](m) = X [( IP.Y Hf] ° IP ](m)- f [( IP. X
)(f] ° IP](m)
... (IP.X)[( IP. f)[ f]]( n) - (IP. Y)[ (IP.X)[ f]]( n) ... [IP.X.IP.Y)[f)(n).
Thus IP.[X, fJ- [IP.X,IP.fJ by 4.2.15(ii). (ii}"rollows from the fact that d(1I U) ... d/I u. • let us now compute the local expression for [X. fJ. Let IP: U ... VeE be a chart on M and let the local representatives of X and f be X and Y respectively, so X, Y: V ... E. By 4.2.22, the local representative of IX, fJ is
[X, fJ. Thus,
[X, t][JJ(x) - X[y[JJ](x)- Y[X [JJ](x) ... D( y[j])(x )·X(x)- D( X[i])( x )·Y(x) .. Now f[lJ<x)- D](x)'Y(x) and its derivative may be computed by the product rule. The terms involving the second derivative of ] cancel by symmetry of D 2J(x} and so we are left with
D](x)· (DY{x).·X(x)- DX(x)· Y(x )}.
224
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Thus the local representative of (X. Y} is
DY·X-DX·Y. If M is n-dimensional and the chart '" gives local coordinates (xl •...• x") then this calculation gives the components of IX. Yl as
(6) Part (i) of 4.2.22 has an important generalization to ",-related vector fields. For this. however. we need first the following preparatory proposition.
4.2.24 Propoeltlon. Let "': M -+ N be a C' map of C' manifolds. X E CX,-I(M). and X' E CX,-I(N). Then X - X' iff (X1f)) ° '" - Xlf ° "'lfor all
f
"
E C§'I(V). where V is open in N.
Proof. By dermition, «X11l)o ",)(m) -
d!(",(m»·X'(",(m». By the chain
rule. X(f ° ",)(m) ... d(fo ",)(m)·X(m) .. df( ",(m»·T",,,,(X(m».
If X! X'. then T", ° X - X' ° '" and we have the desired equality. Conversely. if X[f ° "'] - (X'lll)° '" for all f E '!f1(V), and all V open in N, then choosing V to be a chart domain and f the puD-back to V of linear functionals on the model space of N. we conclude that II,,· (X' ° '" )( m) - II,,· (T", ° X)(m), where n = ",(m). (or all II" E Using the Hahn-Banach theorem. we conclude that (X'o ",)(m) = (T", 0 X)(m). for all mE M. • , It is to be noted that under dirferentiability assumptions on the norm of N (as in 4.2.16), the condition "for a1lf E '!f1(V) and all V eN" can be replaced by "for all f E 6J1(N)" by using bump function', This holds in particular for Hilbert (and hence for finite-dimensional) n.illifolds.
r:N.
4.2.25 PropoeHIon. Let "': M -+ N be a C'map of manifolds. X. Y E CX·-I(M). and X',Y'E~:-I(N). If X!.X, and Y!Y'. then
IX. Y] !·[X'. Y1.
·VECTOR FIELDS AS DIFFERENTIAL OPERATORS
225
Proof.' By 4.1.14 it suffices to show that ([X'. Yll1) ° cp = lX. YII cp] for aU I e ~I(Y). where V is open in N. We have 0
([ X'. Y'][/» ° cp'" X'[Y'[/]]· cp - Y'[X'[ I]] cp 0
- X[(Y'[/). cp)- Y[( X'[/). cp) ==
X[Yl/ 0cp)) - Y[xl/ 0.,,)]
==
[X. Y][/·.,,] . •
The analog of 4.2.9 is the following.
4.2.21 Propoeltlon. For every X e ~ (M). Lx ;s a derivation on (Cft( M). ~(M». That is. Lx is R linear on each. and Lx(/ Y) - (I.x/)Y + I(LxY).
' For g E coo(U, E), where U is open in M, we have
Proof.
[X,/y][g] - Lx{Lfyg)- LfyLxB - LxULyg)- fLyLxB - (Lxf}Lyg + fLxLyg- fLyLxB. so [X,/y) - (Lxf}Y + /[X, Y] by 4.2.IS(ii). • Commutation of vector fields is characterized by their nows in the followin& way.
4.2.27 PropoeIUon. Let X, Ye ~r(M), r .. I, and let flows. The loIlowlng are equioalent.
F"G, denote their
(i) lX, y) ... 0; (ii) FiY-,Y;
G:
(iii) X -' X; (iv) F, • G. - G. ° F,. (In (ii)-(ro), equoIity is undentood, as usual, where the expressions are
defined.)
F, • G• .. G, 0 F, iff G. - F, • G. • r, I. which by 4.2.4 is equivalent to y ... FiY; i.e., (iv) is equivalent to (ii). Similarly (iv) is equivalent to (iii). If FiY-Y, then
Proof.
[X,Y)=
dtdl ,-0F,*Y""O.
226
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Conversely. if IX. YJ - LxY "" O. then·
so that p,'Y is constant in t. For t = O. however. its value is Y. so that p,'Y - Y and we have thus showed that (i) and (ii) are equivalent. Similarly . (i) and (iii) are equivalent. • Just as in 4.2.10, the formula for the Lie derivative involving the flow can be used to solve special types of first-order linear n x n systems of partial differential equations. Consider the first-order system: .
f. (XJ(K)aY'(K.t)_YJ(K.t)aXi(K.t») ax' ax}
aaYi(K.t)", (P,,) {
t
J_ I
yl(K.O) "" gi(K) for given smooth functions Xi( K). gi( K) and scalar unknowns yi( K. t ). i = 1•..•• n. where K -= (Xl •. .•• x"). 4.2.28 Proposition. Suppose X= (XI •.... X") hav a complete /1011' F,. Then letting Y == (yl •...• Y") and G = (gl ..... g,,).
y= p,'G is a solution of the foregoing problem (P,,). (See Exercise 4.2C for uniqueness).
ay
at = =
d
dt p,'G = P,'[ x. G) = [F,* X. p,'G)
[X. Y)
since P,' X = X and Y = p,'G. The expression in the problem (P,,) is the ith component of IX. Y). • 4.2.21 Example. ayl
-
at
ay2
Solve the system of partial differential equations: =
ayl
ayl
ay2
ay2
(x+ y ) - -(x+ y) _ _ yl_y2 ax ay
-=(x+y)--(x+ .-)_+yl+y2 at ax . ay yl(X. y.O)
=
x
y2(X. y.O) = y2.
VECTOR FIELDS AND DIFFERENTIAL OPERA TORS
227
The vector field X(x, y) = (x + y, - x - y) has the complete flow F,(x, y) "" «x + y)1 + x, -(x + y)t + y), so that the solution is given by Y(x. y, t) ... F:(x, y2); i.e., { .
yl( x, y, t) ...
«x + y)t + x)( 1- t) - I [y - (x + y)t]2
y2(x, y, I) = t«x
+ Y)I + x)+(t + I)[y - (.t + y)tY. •
In later chapters we will need a flow type formula for the Lie derivative of a time-dependenl vector field. In Section 4.1 we discussed the existence and uniqueness of solutions of a time-dependenl Vl'l'lor field. Let us formalize and recall the basic facts.
4.2.30 DaflnlUon. A C tiIM-depent/Dft vector field i.f a C map X: RxM ..... TM such that X(t,m)ETmM for all (t,m)E:RXM; i.e:, X,E ~.'(M),
lor
where X,(m) = X(t, m). The time-de",_nt flo ... or ft:JOhIlw"opelYl-
F, .• of!l is defined by the requirement thai I .... F, ..,(m) be Ihe integral
curoe of X slarling at m altime t ""' s; i.e.,
d dl F, .• (m) = X(/, ~ .• (m»
and
F. .• (m) = m.
By uniqueness of integral curves we have F, .• F,,, = F,., (replacing the flow property F,+. = F, F.), and F,., = identity. It is customary to ~rite F, - F,.o· If X is time independent, F, .• " F,- •. In general F:X," X,. However, the basic Lie derivative formulae still hold, 0
0
4.2.31 Theorem. Lei X, E ~'(M), r., I for each t and suppose X(/, m) is continuous in (I, m). Then F,.• is of class C' and for f E C+ '(M, F), and Y E 'X'(M). we have
(i) d dlFi..Y= Fi..([X,. Y)) = Fi..(Lx,Y).
(ii)
Proof. The rarst part was proved in Section 4.1. For functions, the proof is a repeat of 4.2.10:
~(Fi..f)(m)- ~(fo F, .• )(m)=df(F, .• (m»· dF,'d!m) - d/(F,.,(m»·X,( F, .• (m» == (Lx,/)( F,.Am»= Fi..(Lx./)(m).
228
VECTOR FIELDS Afl/D DYNAMICAL SYSTEMS
For vector fields, note that from (5), (~.Y)[f1 = ~.(Y[F;,/])
(7)
since F.., = F,~". The result (ii) will be shown to follow from (i), (7). and the next lemma.
4.2.32 Lemma. d
dtF;,/= - X,[~,/]. Proof. Differentiating ferential equation:
F.., F, .• = 0
identity in t. we get the backward dif-
d
dt F.., =
-
TF.., 0 X,.
Thus
d
dt F:"f(m) - - dfU:.,(m »·TF.,,( X,(m»
... - d(f 0 F..,)·X,(m) = - X,(f 0 F..,](m).
9
Thus from (7) and (i). d
dt (~.Y)[f] -= ~.( X, [Y[F:,,/]])- ~.(Y[X,[~,fJ]). By 4.2.21 and (7). this equals (~.[X" Y)[f). • If f and Yare time dependent, then (i) and (ii) read
(a
d -P. dt , ...1 =P. ,.. -atf+ L x/ )
(8)
and
d
dt~'Y= ~s
(ay at + [X,. Y,] ).
(9)
Unlike time-independent vector fields. we generally have
Time-dependent vector fields on M can be made into time-independent ones on a bigger manifold. Let , E ~ (R x M) denote the vector field given
VECTOR FIELDS AND DIFFERENTIAL OPERATORS
229
Figure 4.2.4
by '(s, m) -
(~,
1),Om) e 1(•. m)(R X ~);; T,R X TmM. Define the .suspen-
sion of X by Xe 'X,(R X M) where XU, m) - «t, 1), X(t, m» and observe that X - , + X. Since b: I .... M is an integral curve of X at miff b'( I) ... X(t, b(t» and b(O) ... m. a curve c: J .... R X M is an integral curve of X at (0, m) iff c(t) - (t. b(I». Indeed, if c(1) == (a(t), b(/» then C(/) is an integral curve of X iff C'(I)" (a'(t), b'(/» = X(C(I»; that is a'(t) == I and b'(/) - X(a(/), b(/». Since a(O) - 0. we get aCt) - t. These observatians are
summarized in the following (see Fig. 4.2.4.)
,
4.2.33 P~ltIon. Lei X be a C-time-dependent veclor field on M wilh evolution oPerator F"., The flow F, of the suspension Xe CX'(R X M) is given by F,(s. m) = (t + s. F,+.,.(m». • Proof. In the preceding notations. b(l) = F"o(m). c(t) = F,(O, m) = (I. F"o( m», and so the statement is proved for s = O. In general. note that ~(s. m) = F,+s(O. Fo,.(m» since 1 - F, .. • (0. i;,.•(m» is the integral curve of X. which at 1=0 passes through F,(O.m)=(s.U:.nol·;,.,)(m»=(s.m). Thus F,(s. m) = F,+$(O. Fn. ,(m» = (I + s. (F,+"II ° 1;,. • )(m» = (t + .f. F,+s .• (m». •
BOX 4.2A
AUTOMORPHISMS OF FUNCTION ALGEBRAS
The property of flows corresponding to the derivation property of vector fields is that they are algebra preserving:
P,'(fg) ... (P,'f)( P,'g).
230
VECTOR FIELDS AND DYNAMICAL SYSTEMS
In fact it is obvious that every diffeomorphism induces an algebra automorphism of ~ ( M). The following theorem shows the converse. t 4.2.34 Theorem. Let M be a paracompact second-countable finitedimensional manifold. Let a: ~(M) - ~(M) be an invertible linear mapping that satisfies a(fg)=a(f)a(g) lor all I. g E ~(M). Then there is a unique Coo diffeomorphism fJ!: M - M such that a(f)=fOfJ!.
Remark. A. There is a similar theorem for paracompact second-countable Banach manifolds; here we assume that there are invertible linear maps a.T: COO( M. F) - C oo ( M. F) for each Banach space F such that for any bilinear continuous map B: F x G - H we have
for I ECoo(M. F) and gECOO(M.G). The conclusion is the same: there is a Coo diffeomorphism fJ!: M - M such that
for all F and all I E C oo ( M. F). Alternative to assuming this for all F. one can take F = R and assume that M is modeled on a Banach space that has a norm that is Coo away from the origin. We shall make some additional remarks on the infinite-dimensional case in the course of the proof. B. Some of the ideas about partitions of unity are needed in the proof. Although the present proof is self-contained. the reader may wish to consult Section 5.6 simultaneously. C. In Chapter 5 we shall see that finite-dimensional paracompact manifolds are metrizable. so by \.6.14 they are automatically second countable. Proof. (Uniqueness). We shall first construct a Coo function x: M - R which takes on the values I and 0 at two given points mi' m2 E M. m, tWe tIumk Alan Weinstein for suggesting this theorem. We think this result must be known in the literature. but we could not locate a specific reference except for an analogous result of Mackey (1962) for measurable functions and measurable auto-
morphisms.
VECTOR FIELDS AND DIFFERENTIAL OPERA TORS
23.1
"" m 2 • Choose a chart (U. cp) and mi' such that m 2 1i:. U and such that cp( U) is a ball of radius r, about the origin in E. cp( m, ) = O. Let V C U be the inverse image by cp of the ball of radius r2 < r, and let 8: E -+ R be a Coo-bump function as in Lemma 4.2.13. Then the function x: M -+ R given by .
80 CP.
x= { 0,
U M\U
on on
is clearly Coo and x(m,) = I. x(m 2 ) = O. Now let us assume that cp./ == If·/ for two different diffeomorphisms "'.If of M. for all/ E tj(M). Then there exists a point me M such that cp(m)-If(m) and thus we can find XE~'t<M) such that (X 0 cp)(m) = I,(X 0 If)(m) = 0 contradicting cp·X = If·X. Hence cp == If. The proof of existence is based on the following key lemma. 4.2.35 Lemma. Let M be a (/inite-dimensiona/) parucompact mani" /old and p: 'fJ( M) -+ R be a nonzero algebra homomorphism. Then there is a unique point m E M such that PU) /(m'>. =0
Proof. (Following suggestions of H. Bercovici). Uniqueness is clear. as before. since for m, "" m 2 there exists a bump function / E tj(M) satisfying/(m,) == O. /(m 2 ) = I. To show existence. note first that P(l) = 1. Indeed P(I) = P(12) = P( I )P( 1) so that either P( I) = 0 or P( I) = 1. But P< I) = 0 would imply P is identically zero since PU) = P(l·/) = P(I)·PU). contrary to our hypothesis. Therefore we must have P( I) = I and thus P( c) = c for cER. For mEM, let Ann(m) = {/E tj(M)I/(m) = O}. SecOnd, we claim that it is enough to show that there is an mE M such that kerp={/E~T(M>lPU)=O}==Ann(m). Clearly. if ~U)-/(m) for some m, then kerp = Ann(m). Conversely, if this holds for ~ mE M and / It:. kerp, let c,. PU) and note that / - cE kerPAnn(m), 'SO/(m) == c and thus P
232
VECTOR FIELDS AND DYNAMICAL SYSTEMS
and hence/ e g + / c J; i.e.• / -IJ(M). Similarly. the ideal Ann(m) is maximal since the quotient IJ(M)/Ann(m) is isomorphic to R. Let us assume that k.er fl - Ann( m) for every m e M. By maximality. neither can be included in the other. and hence for every
m e M there exists a relatively compact open neighborhood U", of m and /", e kerll such that /",1 U", > O. Let V", be an open neighborhood of the closure. c1(U",). Since M is paracompact. we can assume that {V",lm eM} is locally finite. Since M is second countable. M can be covered by {V",V eN}. Let ~ = /m and let Xj be bump functions equal to 1 on cl(U",) and v~ishing in M\Vm • If an < IlIn 2sup(x,,(m)/,,2(m)lm eM}). the function I = l::"_la nXnln2 is C"" (since the sum is finite in a neighborhood of every point). I> 0 on M and the series defining I is uniformly convergent. being majorized by E:'_ln-2. If we can show that II can be taken inside the sum. then lIe I) = o. This construction then produces I e ker11./ > 0 and hence 1'"' (I/!)/ e kerll; i.e.• kerll = IJ(M). contradicting the hypothesis fl .. O. To show that II can be taken inside the series sign. it suffices to prove the following "g-estimate:" for any g e IJ(M), III(g)1 '" sup(lg(m~lm eM}. Once this is done, then
~ supl r, a"x .. /"2 " -I
II-
0
as n -+ 00 by uniform convergence and boundedness of all functions involved. Thus 11(/) = E':_lfl(a"x"I,,2). To prove the g-estimate. let " > sup(lg(m~lm e M} so that" ± g "" 0 on M; i.e.• "± g are both invertible functions on M. Since II is an algebra homomorphism. 0-11(" ± g) -" ± lI(g). Thus. ± fl(g) """ for all sup(lg(m~lm eM}. Hence III(g)1 '" sup(lg(m~lm eM}. •
,,>
Aemn. For infinite-dimensional manifolds. the proof of the lemma is almost identical with the following changes: we work with fl: COIl ( M. F) -+ F. absolute values are replaced by norms. second countability is in the hypothesis and the neighborhoods V", are chosen in such a way that 1",1 Vm is a bounded function (which is possible by continuity of 11ft)'
VECTOR FIELDS AND DIFFERENTIAL OPERATOR8
233
Proof of em'... iII·f 71. For each mE M, define the algebra • R by /J",(f) - a(f Xm). Since a is inhomomorphism /J",: . vertible. a(l) .. 0 and !oil i~'; a(1) - a(12) - a( 1)a(1). we have a(1 ).- I. Thus /J",'." 0 for all m E M. By 4.2.35 there exists a unique point, which we call .,,(m)EM such that /J",(f)=/(.,,(m»-(.,,*/Xm). This dermes a map.,,: M - M such that a(f)- rp*/for alii E 'J(M). Since /I is an automorphism, " is bijective and since a(f) - .,,*1 E ~(M),a-I(f) - .".1 E 'J(M) for all I E 'J(M), both rp, rp- I are C'" (take for I any coordinate function multiplied by a bump function to show that in every chart the local representatives of cpo rp - 1 are C"') .
•
The proof of existence in the infinite-dimensional case proceeds in a similar way.
BOX 4.28 THE METHOD OF CHARACTERISTICS
The method used to solve problem (P) also enables one to solve first-order quasi-linear partial differential equations in R". Unlike 4.2.1 J. !he solution will be implicit, not explicit. The equation under consideration in R" is
.;. " I) al. -I... Xi( x 1 ,. ..• x. i-I
I
ax"
y( x 1•...• x. " I) ,
(Q)
where/=/(xl, ... ,x") is the unknown function and Xi,Y.i-J .... ,n are C' real-valued functions on R II + I, r ~ I. As initial condition one takes an (n -I)-dimensional submanifold r in R"+ 1 that is nowhere tangent to the vector field
Thus, if r is given parametrically by X ' ... X i (t l , ... ,t"_I),i==1 ..... n
and
1-=1(11 ..... 1"-1).
234
VECTOR FIELDS AND DYNAMICAL SYSTEMS
this requirement means that the matrix
x· ax·
at;
X" ax" a/.
al art
ax· al,,_.
ax" al"_.
al al"_.
y
has rank n. It is customary to require that the determinant obtained by deleting the last column be '* O. for then, as we shall see. the implicit function theorem gives the solution. The function / is found in the following way. Consider F,. the flow of the vector field r a/ax; + ya / 1 in R" + • and let S be the manifold obtained by sweeping out r by F,. That is. S'" U{F,(f)11 E R}. The condition that xia / axi + ya/al never be tangent to r ensures us that the manifold r is "dragged along" by the flow F, to produce a manifold of dimension n. If S is described by / .. /(x· •...• x") then 1 is the solution of the partial differential equation. Indeed. the tangent space to S contains the vector E7_.X 1a/8xl + Y8/a/; i.e.• this vector is perpendicular to the normal (81/ax· ..... al/ax".-I) to the surface/=/(x· ..... x") and thus (Q) is satisfied. We work parametrically and write the components of F, as Xi=XiU ...... I"_ •• /). i=I ..... n and /=1(1 ...... 1,,-1./). Assuming that
r:.7_. r:.7_.
a
x· ax·
o.
at; ax· al"_.
X" ax" a/. ax" al"_ •
ax· al ax· a/. ax· ar" _.
ax" al ax" a/. ax" ar"_.
one can locally invert to give 1= r( x· .... ,x"), I, = r,( x·, .... x"), i = I ..... n. Substitution into / yields / = /(x· ..... x"). The fundamental assumption in this construction is that the vector field E':_lxla/aXi + ya/al is never tangent to the (n - 1)manifold r. The method breaks down if one uses manifolds r. which are tangent to this vector field at some point. The reason is that at such a point. no complete information about the derivative of 1 in a
VECTOR FIELDS AND DIFFERENTIAL OPERA TORS
235
complementary (n - 1)-dimensional subspace to the characteristic is known. 4.2.38
Examples.
with initial condition one-manifold
Consider the equation in R 2 given by
r=
{(x. y. f)lx ... s. y'" !S2 - S. f = s}. On· this
f
s s- 1
=-1*0
so that the vector field a/ax + fa/ay +3a/afis never tangent to r. Its flow is F,(x. y. f) = (t + x.(3/2)t 2 + fl + y.3t + f) so that the manifold swept out by r along F, is given by x(t. s) = t + s. y(t. s) = (3/2)/ 2 + sl + (I/2)S2 - s. f(/. s) = 3t + s. Eliminating I. s we get
The solution is defined only for I - 2X2 + 4x + 4.1' ~ O.
•
. Another interesting phenomenon occurs when S can no longer be described in terms of the graph of f: e.g.• S "folds over." This corresponds to the formation of shock waves. Further information can be found in Chorin and Marsden (1979), Lax (1973). Guillemin and Sternberg (1977), John (1975), and Smoller (1982).
4.2C, DERIVATIONS ON
C" -1
FUNCTIONS
4.2.37 Theorem (A. Blasst). Let M be a C A , 2 finite-dimensional manifold: where k ~ O. The colleclion of all R linear derivations 8 from ':t h + I ( M ) to 'ff" ( M) is isomorphic to 'X" ( M) as a real t'ector space. The proof of 4.2.16 sho~s that there exists a unique C" vector field X such that 81~k+2(M)= Lxl~"+2( M). So all we have to do is show that 8 and Lx agree on C"+ I functions. Replacing 8 with e- Lx. Proof.
t
Private communication.
236
VECTOR FIELDS AND DYNAMICAL SYSTEMS
we can assume that 8 annihilates al1 C k + 2 functions and we want to show that it also annihilates all C k + 1 functions. As in the proof of 4.2.16, it suffices to work in a chart, so we can assume without loss of generality that M-R" . .For a Cl<+ 1 function / on R" we shall prove that (8/)(0) = 0 which implies that (8/)( p) = 0 for all pER n by translation of the argument of f. Replacing / by the difference between / and it~ Taylor polynomial or order k + I about 0, we can assume /(0) = 0, and the first k + I derivatives vanish at 0, since 8 annihilates any polynomial. We shall prove that / = g + h, where g and h are two Ck+ 1 functions, satisfying gl U - 0 and h IV = 0, where U, V are open sets such that OEcI(U)ncl(V). Then, since 8 is a local operator, 8gIU=O and 8hlV- 0, whence by continuity 8glcl(U) - 0 and 8hlcl(V) = O. Hence (8/)(0) ... (8g)(O) + (8h)(O) - 0 and the theorem will be proved.
4.2.38 Lemma. Let cp: S,,-I c R" -+ R be a Coo function and denote by.,,: R"\{O} -+ S,,-I, .,,(x) = x/llxll the radial projection. Then lor any positive integer r D'( cp 0.,,)( x) = for some C"" functlon.p: S,,-I
-+
(.p 0.,,)( X Vllxll'
L~(R";R). In particular Dk(cp
0
'II")(x)
- O(lIxlI-') as IIxll- O.
Proof. For r - 0 choose cp =.p. For r -I, note that I V D'II"(x)·v= - -DII·II(x)·v+-
IIxll2
II xII ,
and so
But the mapping i(x) = (1/lIxll)DII·II(x) satisfies i(lx) = i(x) for all t > 0 so that it is uniquely determined by 1= iIS,,-I. Hence D( cp • .,,)(x) - (I/II,I"II)(.p o.,,)(x), where .p( y) = Dcp( y)(I y». y E s"-'. Now proceed by induction. • Returning to the proof of the theorem. let I be as before (i.e .. of class C k and Di/(O)'" 0, 0 "'" i "'" k + 1) ·,nd let cp.'" be as in the lemma. From Taylor's formula with! we see that Di/(x) = o(II.I"IIk+ 1-').0"," i "'" k + I. as x -+ 0 ; the product rule and
-ie
+,
VECTOR FIELDS AND DIFFERENTIAL OPERATORS
237
the lemma, DI(f'(fJJO"»(x)=
I:
O(UXUk+I-J)O(Uxlrl)=O(nXUHI-i)
j+l-i
,,»,
so that DIU' (fJJ 0 0 " i " k + I can be continuously extended to 0, by making them vanish at O. Therefore f· (g 0 , , ) is C h I for all R ". Now choose the CfXJ function fJJ in the lemma to be zero on an open set 0, and equal to I on an open set O2 of S"-I,O, noz -=121. Then the continuous extension g of f'(fJJ 0 , , ) to R" is zero on U=,,-'(Ol) and agrees withfon V-".-I(OZ)' Let h-f-g and thus f is the sum of two C k + I functions, each of which vanishes in an open set having 0 in its closure. This completes the proof. • , In infinite dimensions, the proof would require the norm of the model space to be Coo away from the origin and the function'" in the lemma bounded with all derivatives bounded on the unit sphere. Unfortunately, this does not seem feasible under realistic hypotheses. The foregoing proof is related to the method of "blowing-up" a singularity-see for example Takens [1974] and Buchner, Marsden, and Schecter [1982]. There are also difficulties with this method in infinite dimensions in other problems, such as the Morse lemma (see Golubitsky and Marsden [1983] and Buchner, Marsden, and Schecter [ 1983».
4.2.39 Corollary. If M is a C k + I finite-dimensional manifold, the only derivation from Ijk( M) to Ijk( M), where I " k < 00, is zero.
.
Proof. By the theorem, such a derivation is given by a' C k - I vector field X. If X-O, then for somefeljk(M),X[f) is only C k - ' but not Cle. •
We do not know of an example of a derivation not given by a vector field, when the hypotheses of 4.2.37 fail.
BOX 4.2D PRODUCT FORMULAS FOR THE LIE BRACKET
Tltis box is a continuation of Box 4.lA and gives the flow of the Lie bracket [X, Y] in terms of the flows of the vector fields X, Y e 'X(M).
238
VECTOR FIELDS AND DYNAMICAL SYSTEMS
".2.40 Propolltlon. LeI X, Y E ~(M) have flows F, and G,. If B, denotes the flow of(X, Y), then for x EM,
Proof. Let
K.(x) - (G- i
0
F _.j 0 G.j 0 F.j)(x), e> O. The formula
rollows from 4.1.24 if we show that
aae K.(x)1
.-0
=
lX, Y](x)
ror all x E M. This in tum is equivalent to
..!.K:fl
ae
.-0
=
[X. Y}[f]
for any f E C!f(M). By the Lie derivative formula,
By the chain rule. the limit of this as e - 0 + is half the a/ a.fE -derivative of the parenthesis at e = O. Again by the Lie derivative formula, this equals a sum of 16 terms. which reduces to IX. Y][f]. • For example. for n X n matrices ..4 and B, 4.2.40 yields the classical formula e IA .-]= lim (e- A/-!ie--/ ii e A/-!ie·/.,4i)".
"-00 where the commutator is given by (..4. B)-..4B - B..4. To see this, define for any n X n matrix C a vector field Xc E ~(R") by Xc(x) = Cx. Thus Xc has flow F,(x) == e'cx . Note that Xc is linear in C and satisfies (XA • X.)- - XIA ._] as is easily verified. Thus the flow of IX., XA ) is e'IA._] and the formula now follows from 4.2.40. The results or 4.1.25 and 4.2.40 had their historical origins in Lie group theory, where they are known. by the name of exponential /omtUIas. The converse of 4.1.25, namely expressing e,Ae,· as an exponential of some matrix for sufficiently small t is the famous Baker-Campbell-Hausdorff formula (see e.g. Varadarajan (1974, sec.
VECTOR FIELDS AND DIFFERENTIAL OPERATORS
239
2.15). The formulas in 4.1.25 and 4.2.40 have certain generalizations to unbounded operators and are called TrOller product formula., after Trotter (1958). See Chorln et. al. (1978) for further information. Exercl...
4.2A
(i) On R2. let X(x. y) == (x. y; y. - x). Find the flow of X. (ii) Solve the following for f(t. x, y):
af -y~-X af at
ax
£ly
if f(O. x. y) ... ysin x. 4.2B (i) Let X and Y be vectors fields on M with complete flows F, and G,• respectively. If [X. Y) = 0, show that X + Y has flow H, "" F, 0 G,. Is this true if X and Yare time dependent? (ii) Show that if [X, Y) = 0 for all Y E <:\ ( M). then X = O. ( Hint: From the local formula conclude first that X is constant: then take for Y linear vector fields and apply the Hahn..:.Banach theorem. In infinite dimensions, assume the conditions hold locally or that the model spaces are ex». 4.2C Show that. under suitable hypotheses, the solution f(x, t) = g( F,(x» of problem (P) given in 4.2.11 is unique. (Jlint: Consider the function E(t) = fRol/,(x, t)- 12(X' 1)1 2 dx, where II and 12 are two solutions. Show that dE/dt ~ aE for a suitahle constant a and conclude by Gronwall's inequality that E = O. The "suitable hypotheses" are conditions that enable integration by parts tQ be performed in the computation of dE/dt.) Adapt this proof to get uniqueness of the solution in 4.2.28. 4.2D Let X. Y E 'X( M) have flows F, and G" respectively. Show that
[X'Y]-=/~tl s
1•• -0
(Hint: The flow of p,'Y is s ..... F, 0 G.
4.2E 4.2F
(F,oG,oF , ), 0
F _ I')
Show that SO(n) is paralleIizable. See Exercise 3.5S for a proof that SO(n) is a manifold. (Hint: SO(n) is a group.) Solve the following system of partial differential equations. ayl
-
at
ay2
ayl
ayl
ax
ay
yl- y2
ay2
ay2
I
= (x + y)-.- +(4x -2y)- -
Tt = (x + y)
ax +(4x -2y) ay -4Y +2Y
2 •
240
VECTOR FIELDS AND DYNAMICAL SYSTEMS
with initial conditions yl(X. y.O) = x
+y
(Hint: The flow of (x + y.4x -2y) is
4.20 Consider the following equation for I(x, t) in dit1er1{ence lorm:
ill + ~(H(f)) =0 ilt ilx where H is a given function of f. Show that the characteristics are given by .k - H'( I). What does the transversality condition discussed in Box 4.2B become in this case? 4.2H Let M. N be manifolds with N modeled on a Banach space which has a C" norm away from the origin. Show that cp: M -+ N is C" iff f 0 cp: M -.R is C· for all/ E '!f"(N). 4.21 Develop a product formula like that in Box 4.1A for the flow of X + Y for time-dependent vector fields. (Hint: You will have to consider time-ordered products.)
4.3 AN INTRODUCTION TO THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS We have seen quite a bit of theoretical development concerning the interplay between the two aspects of vector fields. namely as differential operators and as ordinary differential equations. It is appropriate now to look a little more closely at what flows look like geometrically. Much of the work in this section holds for infinite-dimensional as well as finite-dimensional manifolds. The reader who knows or is willing to learn some spectral theory from functional analysis can make the generalization. This section is intended to link up the theory of this hook with I.'ourses in ordinary differential equations that the reader may have taken. The section will be most beneficial if it is read with such a course in mind. We begin by introducing some of the most basic terminology regarding the stability of fixed points.
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
241 '
4.3.1 Definition. Let X be a C l vector field on 1II1 n-man;fold M. A point mo is called a cmictll po;"t (also called a s;"plar point or an ('qui/ibri,,", poiItt) of X if X( mo)" O. The Iilteari'{.lItion of X at a critical point mo is the linear map X'( mo): T",.M - T",.M defined by X'(m o )'''= dd (TF,(mo)·.,)1 t /-0 where F, is the flow of X. The eigenvalues of X'( m 0) are called the clu"uctemtic exponents of X at mo.
Some remarks will clarify this definition. First of all. F, leaves mo fixed: F,(mo)=m o' since c(t)=m o is the unique integral curve through mo' Conversely. it is obvious that if F,( mo) = mo for all I. then is a critical point. Thus TF,(mo) is a linear map of TmoM to itself and so its t derivative at O. producing another linear map of Tm.M to itself. makes sense. '
mo
Computationally. this definition is not so convenient. The following is more useful. 4.3.2 PropoalUon. Let mo be a critical point of X (/lid let (V.,,) he a chart on M with cp(m o ) - Xo e R". Let x -= (xl •...• x") denote cOlJrdinateJ in R" and XI(XI ..... X") ..... X"(x· ..... x") the component,l' of Ihe local repr.esentalive oJ X. Then the matrix of X'( mol in these coordinates is
Proof.
This follows from the equations d
XI( F,(x» = dt F,'(x) after differentiating in x and setting x equations in Lemma 4.1.9) •
= XO'
t = 0, (Sec also the linearized
'
Another. equivalent procedure for linearizing. which generalizes to sections of arbitrary vector bundles. is outlined in E~ercise 4.3A. The name "characteristic exponent" arises as follows. We have the linear differential equation
242
VECTOR FIELDS AND DYNAMICAL SYSTEMS
and so TWlo F.I = e,x·(m.) •
Here the exponential is defined. for example. by a power series. The actual computation of these exponentials is learned in differential equations courses, using the Jordan canonical form. (See Hirsch-Smale (1974). for instance.) In particular. if ~I"" '~n are the characteristic exponents of X at mo. the eigenvalues of T....F, are e'''' ..... e'''·. The characteristic exponents will be related to the following notion of stability of a critical point.
4.3.3 DeflnlUon. Let mo be a critical point of X. Then mo is sl'" (or Lillpulrov stabk) if for any neiKhborhood U 01 mo. there is a neighborhood V of mo such that if m E V. then m is + complete and F,(m) e U for all t ~ O. [See FiK. 4.3.I(a).) (ii) mo is IUY"'PIOIically stable if there is a neighborhood V of mo such that if me V. then m is + complete. F,(V)C F,.(V) if t > sand (i)
lim F,(V) 1-
+ 00
=
{mo}
[i.e .• for any neighborhood U of mo. there is a T such that F,( V) c U if t ~ T). [See Fig. 4.3.l(b).)
It is obvious that asymptotic stability implies stability. The harmonic oscillator x = - x giving a flow in the plane shows that stability need not imply asymptotic stability [Fig. 4.3.I(c»). The following result of Liapunov gives a basic criterion for stability.
.v
x
(a) Stable
(b) Asymptotically stable
Figura 4.3.1
(c) Harmonic oscillator
THe QUALITATIVe THeORY OF DYNAMICAL SYSTeMS
243
4.3.4 Theorem. Suppose X is C' and mo is a critical point of X. Assume the characteristic exponents of mo have strictly negative real parts. Then mo is asymj)lotically stable. [In a similar way, if Re(l£i) > 0, mo is asymptotically unstable, i.e., asymptotically stable as t -+ - 00.) Proof. We can assume M = E = R n and that mo = O. Let - E> r = max(Re;\., ..... Re;\.n). Then we claim there is a norm 11'11 on E in which
lIe,x'(OIIl
~
e-·' .
If X'(O) is diagonalizable (e.g.• has distinct eigenvalues) a~ a complex.matrix. this is easy, for we can let II II be thp. sup norm associated with a basis of eigenvectms. If X'(O) is not diagonalizable. we can approximate it by one and get the same conclusion. (If the reader is familiar with the spectral radius formula. choose IlIe m , x '(OI(x)1I1 IIxll = sup ,"I • m;o 0 e /;0
0
where 111'111 is any norm.) Write A = X'(O) = DX(O). From the local existence theory, there is an r-ball about 0 for which the time of existence is uniform if the initial condition Xo lies in this ball. Let R(x) = X(x)- DX(O)·x.
Findr2 ~ r such that IIxll ~ r2 implies IIR(x)!1 ~ allxll. where a = E/2. Let B be the open rd2 ball about O. We shall show that if Xo E B, then the integral curve starting at Xo remains in Band -+ 0 exponentially as t -+ + 00. This will prove the result. Let x(t) be the integral curve of X starting at Xo' Suppose x( t) remains in B for 0 ~ t < T. The equation x(t}=x o + [X(x(s»ds
o
=xo+ ll.fx(s)+R(x(s))]d~
o
gives, by the variation of constants formula (Exercise 4.1 E).
and so
244
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Letting/(t) ... e"lIx(I~I, I( I)
"
II Xo II + a 10 I/( S ) tis,
and so, by Gronwall's inequality,
I( I) " IIxolle"'. Thus Hence x(l) E B,O" 1 < T, so as in 4.1.19, x(l) may be indefinitely extended, in 1 and the foregoing estimate holds. • A critical point mo is called hyperbolic or elemenlary if none of its characteristic exponents has zero real part. A generalization of Liapunov's theorem shows that near a hyperbolic critical point the flow looks like that of its linearization. (See Hartman (1973, Ch. 9] and Nelson (1969, Ch. 3), for proofs and discussions.) In the plane, the possible hyperbolic flows near a critical point are shown in Fig. 4.3.2. (Remember that for real systems, the characteristic exponents occur in complex conjugate pairs.) In higher dimensions, we have similar portraits for hyperbolic critical points. Here is the idea. For a critical point, mo, let In(mo) denote the insel, . defined by
and similarly, the
011'8.' is defined by
In the case of a linear system,
.r= Ax, where A has no eigenvalue on the imaginary axis (so the ongm is a hyperbolic critical point), In(O) is the generalized eigenspace of the eigenvalues with negative real parts, while Qut(O) is the generalized eigenspace corresponding to the eigenvalues with positive real parts. Clearly, these are complementary subspaces. The dimension of the inset, S( A ,0) = dimln(O) is called the slability index of the critical point. Now for a nonlinear system
m= K(m),
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
245
1',
(e)
(d)
Figure 4.3.2
suppose mo is a hyperbolic critical point. The linearization theorem alluded to above slates that the phase portrait or X near mo is topologically conjugate to the phase portrait or the linear system
.t= Ax near the origin, where A = X'(mo), the linearized vector field at mo. This means there is a homeomorphism or the two domains, preserving the oriented trajectories of the respective flows. Thus in this nonlinear hyperbolic case, the inset and outset are CO submanifolds. Another important thtw:'m of dynamical systems theory, the stahle manifold theorem (Smale [1 '" • says that in addition, these are smooth injectively immersffi ~lIb rnanli .. lds, intersecting transversally at the critical point mo. See I \
246
VECTOR FIELDS AND DYNAMICAL SYSTEMS
•
\
FlguI'84.3.3
for an illustration showing part of the inset and outset near the critical point. It follows from these important results that there are (up to topological conjugacy) only a few essentially different phase portraits, near hyperbolic critical points. These are classified by the dimension of their insets. called the stability index.
s(X. mol = dimln(mo) as in the linear case. The word index comes up in this context with another meaning. If m is a critical point of a vector field X. the topological index of m is + I if the number of eigenvalues (counting multiplicities) with negative real part is· even and is - I if it is odd. Let this index be denoted I( X. m). so that J( X. m) = (- nSfX • m). The Poincare-Hopf index theorem states that if M is ' compact and X only has (isolated) hyperbolic critical points. then
L
J(X, m) = X(M)
mis a critical point or X
where X(M) is the Euler characteristic of M. (For isolated nonhyperbolic
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
247
critical points the index is also defined but requires the development of degree theory for its definition-a kind of generalized winding number; see Box 7.5A). For a proof and discussion, see Guillemin and Pollack [1974, p. 133).
We now illustrate these basic concepts about critical points with some classical applications.
4.3.5 Exampl.. A. The simple pendulum with linear damping is defined by the second-order equation
x + eX + k sin x =
0
(c>O).
This is equivalent to the planar dynamical system v { x= v=-cv-ksinx The phase portrait is shown in Fig. 4.3.4. The stable focus at the origin
Figure 4.3.4
248
VECTOR FIELDS AND DYNAMICAL SYSTEMS
represents the motionless, hanging pendulum. The saddle point at (k7T,O) corresponds to the motionless bob, balanced at the top of its swing. B. Another classical equation models the buckling column (see Stoker [1950, ch. 3, sec. 10)):
or equivalently. the planar dynamical system
x-v co a,x a 3 x x-------m m m 3
with the phase portrait shown in Fig. 4.3.5. This has two stable foci on the horizontal axis, denoted m, and m2' corresponding to the column buckling. (due to a heavy weight on the top) to either side. The saddle point at the origin corresponds to the unstable equilibrium of the straight, unbuckled column.
~-== Figure 4.3.5
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
249
::: ::::::::::::. :a:::::::::::::::::: :::::::::::::: ................................................. ... ........... .................... ............. .
................... ............. . ,
..... , ..... .
, • • • • • • • • • • 1 •.
:::::::::.::::::::::
,Figure 4.3••
:Note that in this ph:!, ,ait, some initial conditions tend toward one stable focus. while some tend toward the other. The two tendencies are divided by the curve, In(O,O), the inset of the saddle at the origin. This is called the ~eparatrix, as it separates the domain into the two disjoint open sets, In(mo) and In(m,). The stable foci are called ell/ractors, and their insets are called their basins. See Fig. 4.3.6 for the special case x + x - x + Xl = O. This is a special case of a general theory, which is increasingly important in dynamical systems applications. The attractors are regarded as the principal features of the phase portrait; the size of their basins measures the probability of observing the attractor, and the separatrices help find them. A Another basic ingredient in the qualitative theory is the notion of a closed orbit (also called a limit cycle). 4.3.8 DeflnlUon. An orMr
/I
vector field X is called cltn«l when
y( t) is not a fixed point and Iltere u U T > 0 such thaI y( I + T) = y( t) for all t. The of y, In(y), is the set of points me M such that F,(m) - y as t - + 00 (i.e., the distance between F,(m) and the (compact) set {y(t)I0.,.;; t
wet
250
VECTOR FIELDS AND DYNAMICAL SYSTEMS
4iI T} tends to zero as t ..... 00.) Likewise, the 0IItHt, Out(y). is the set of points tending to y as t ..... - 00.
4.3.7 Example. One of the earliest occurrences of an attractive closed orbit in an important application is found in Baron Rayleigh's model for the violin string (see Rayleigh (1887, vol. I, sec. 68a)),
or equivalently,
with the phase portrait shown in Fig. 4.3.7. This has an unstable focus at the origin, with an attractive closed orbit around it. That is, the closed orbit y is a limiting set for every point in its basin (or inset) In(y). which is an open set of the domain. In fact the entire plane (excepting the origin) comprises the basin of this closed orbit. Thus every trajectory tends asymptotically to the limit cycle y and winds around closer and closer to it. Meanwhile this closed orbit is a periodic function of time, in the sense of Definition 4.3.6. Thus the eventual (asymptotic) behavior of every trajectory (other than the unstable constant trajectory at the origin) is periodic; it is an oscillation.
Figure 4.3.7
THE QUALITA TlVE THEORY OF DYNAMICAL SYSTEMS.
251
This picture thus models the sustained oscillation of the violin string. under the innu~nce of the moving bow. Related systems occur in electrical engineering under the name Van der Pol equation. • We now proceed toward the analog of Liapunov's theorem for the stability of closed orbits. To do this we need to introduce Poincare maps and characteristic multipliers. 4.3.8 DeflnlUon. Let X be a vector field on a manifold M. A local trYIIUWnaI :rection of X at m E M is Cl suhmanifold SCM of codimensioll one· with me S and for all s E S. X(s) is not {·ontained ill 7: .'''. Let X be a vector field on a manifold M with integral F: L'i\ c M X R --+ M. y a closed orbit of X with period T, and Sa llJ!:al trllnlTl'r.ml section of X at me y. A Poincare ""'" of y is a mapping 8: Wo -+ WI where:
Wo. WI C S are open neighborhoods of me S. and 8 is a diffeomorphism; (PM 2) there is a continuoUs function 6: Wo --+ R such that for all s E Woo (s. T - 6(s» E 6i\. and 8(s) '"" F(s. T- 6(s»; and finally. (PM 3) ift E]O. T - 6(s)[. then F(s. t)~ Wo(see Fig. 4.3.8).
(PM I)
Figure 4.3.8
252
VECTOR FIELDS AND DYNAMICAL SYSTEMS
4.3.. Theorem (Emtmee IIIIIllllliqwlleu 0/ PoinCtlN ""'I's), (i) If X is a vector field on M, and y is a closed orbit of X, then there exists a Poincare ' mop ofy, (ii) If 9: Wo .... WI is a Poincare map of y (in a local transversal section Sat m e y) and 9' also (in S' at m' e y), then 9 and 9' are locally conjugate. That 18, there are open neighborhoods W1 0/ me S, Wi of m' e S', and a diffeomorphism H: W1 .... Wi, such that W2 c Wo n WI' Wi C Wo n WI and the diagram
commutes.
Proof.
(i) At any point m e y we have X( m) .. O. so there 'exists a now box chart (U, cp) at m with cp(U) ... V x Ie R,,-I xR, Then S X{o}) is a local transversal section at m. If F: 6j)x c M x R .... M is the integral of X. 6j)x is open, so we may suppose U x [ - ", ., ) c OJ)x' where., is the period of y. As ~(m) ... m e U and ~ is a homeomorphism, Uo = ~- I( U)n U is an open neighborhood of m e M with ~(uo) cU. Let Wo'" S n Uo and Wz ~(Wo)' Then Wz is a local transversal section at me M and F.: Wo .... W2 is a diffeomorphism (see Fig. 4.3.9). Now if U1 - ~(Uo)' then we may regard Uo and U2 as open submanifolds of the vector bundle V x R (by identification using,,) and then F.:
_,,-lev
s IJ
FIgu... 4.3.'
u
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
253
Vo ..... V 2 is a diffeomorphism mapping fibers. into fibers. as !p identifies orbits with fibers. and F. preserves orbits. Thus W2 is a section of an open subbundle. More precisely. if 'IT: V X j ..... V and p: V x·1 -I> I are the projection maps. then the composite mapping
has a tangent that is an isomorphism at each point, and so by the inverse mapping theorem. it is a diffeomorphism onto an open submanifold. Let WI be its image. and 8 the composite mapping. We now show that 8: Wo ..... WI is a Poincare map. Obviously (PM I) is satisfied. For (PM 2). we identify V and V x I by means of !p to simplify notations. Then 'IT: W2 ..... WI is a diffeomorphism. and its inverse ('lT1 W1 ) - I: WI -I> W2 C WI xR is a section corresponding to a smooth function 0: WI ..... R. In fact. 0 is defined implicitly by
=
('lTF.(wo),PF.(wo)-I)(lIh»
=
('lTF.( wo),O)
=
8( wo).
Finally, (PM 3) i~ nhvious as (V.!P) is a flow box. (ii) The prOl . ,Irdensome because of the notational complexity in the definition of local conjugacv. so we will be satisfied to prove this uniqueness under additional simplIl)'ing hypOtheses that lead to global conjugacy (identified by italics). The general case will be left to the reader. We consider first the special case m "" m'. Then we choose a flow box chart (V.!P) at m. and assume SUS' c V. and that Sand S' intersect each . orbit arc in Vat most once, and that they intersect exactly the same sets of orbits. (These three conditions may always be obtained by shrinking Sand S'.) Then let W2 = S, W:i = S'. and H: W2 -I> W; the bijection given by the orbits in V. As in (i). this is easily seen to be a diffeomorphism. and Ho8=8'oH. Finally, suppose m .- m'. Then Fu( m) = m' for some a e )0. T(. and as 6D x is open there is a neighborhood V of m such that V X {a} C "Dx ' Then
254
VECTOR FIELDS AND DYNAMICAL SYSTEMS
F,,(U () S) -. s" is a local transversal section of X at m' E y. and H = Fa effects a conjugacy between a and a" - Fa 0 a 0 Fa- , on S". By the preceding paragraph. a" and a' are locally co~ugate. but conjugacy is an equivalence relation. This completes the argument. • If 'Y is a closed orbit of X E 'X (M) and mE y. the behavior of nearby orbits is given by a Poincare map a on a local transversal section Sal m. Clearly Tma E L(T",S. TmS) is a linear approximation 10 8 al m. By uniqueness of a up to local conjugacy. Tm·8' is similar to Tm8. for any other Poincare map 8' on a local transversal section at m' E y. Therefore. ' the eigenvalues of T",8 are independent of m E 'Y and the particular section S atm.
".3.10 DeftnlUon. If'Y is a closed orbit of X E 'X(M). the characteristic y are the eigenvalues of T",8. for any Poincare map e at
llUlltiplie" of X at any mE y.
Another linear approximation to the flow near y is given by Tm FT E L(T",M. T",M) if mE y and 'I' is the period of y. Note that fi(X(m» = X(m). so TlftF. always has an eigenvalue I corresponding to the eigenvector X( m). The (n - I) remaining eigenvalues (if dim( M) = n) are in fact the . characteristic multipliers of X at y.
Propoeltlon. If y is a closed orbit of X E 'X ( M) of period 'I' and cy is the set of characteristic multipliers of X at y. then cy U{l} is the set of eigenvalues of TmF.. for any m E y.
".3.11
Proof. We can work in a chart modeled on R n and assume m = O. Let V be the span of X(m) so R n = TmSe V. Write the flow F,(x. y)'" (F,'(x. y). f;2(X. y». By definition. we have
and
Thus the matrix of T", F. is of the form
where A .. D,F.2(m). From this it follows that the spectrum of TmF. is' {I)Ucy'
•
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
255
If the characteristic exponents of an equilibrium point lie (strictly) in the left half-plane. we know from Liapunov's theorem that the equilibrium is stable. For closed orbits we introduce stability by means of the following definition. 4.3.12 DeflnlUon. Let X be a vector field on a manifold M and y a closed orbit of X. An orbit F,(m o ) is said to. w_ tOwtlrd y if mo is + complete and for an}' localtran.rver,ral S to X at me y there is a to su('h that 'F,Il( mo) E S and successive applications of the Poincare map yield a sequence of points that converges to m. See Fig. 4.3,10. '
Figure 4.3.10
4.3.13 Theorem. Let y be a closed orbit of X E ',,\ ( M) and let the characteri~tic multipliers of y lie strictly inside the unit circle. Then there is a neighborhood U of y such that for any mo E U, the orbit through ""0 winds toward y. The proof of 4.3.13 relies on the following lemma.
4.3.14 Lemma. Let f; S ..... S be a smooth map on a manifold S with f( so) = So for some so' Let the spectrum of lie inside the unit circle. Then there is a neighborhood U of So such that if s e U, fer) E U and I"(s)"'" So as n ..... IX>, where I" - f 0 f 0 • • • 0 fen times).
T,J
The lemma is proved in the same way as 4.3.4. The Iinear'transformation T,.freplaces e ....(11 in the argument. Compitctness of S· then aUows the argument to be extended from the Poin<:ari: section to a neighborhood of the closed orbit.
256
VECTOR FIELDS AND DYNAMICAL SYSTEMS
4.3.15 DeflnlUon. 1/ X E ~(M) and y is a closed orbit 0/ X, y is called "yper6oIic i/ none 0/ the characteristic multipliers 0/ X at y has modulus I. The local qualitative behavior near a hyperbolic closed orbit is especially simple. For example, they are isolated (see Abraham-Robbin (1961, Ch.5]). The local qualitative behavior near an hyperbolic closed orbit, y, may be visualized with the aid of the Poincare map. S: Wo c S .... WI C S, as shown in Fig. 4.3.8. The qualitative behavior of this map. under iterations. determines the asymptotic behavior of the trajectories near y. Let m E y be the base point of the section, and s E S. Then In(y), the inset of y, intersects S in the inset of m under the iteration of S. That is, s E In( y) if the trajectory F,(s) winds toward y, and this is equivalent to saying that SACS) tends to mo as k .... + 00. The inset and outset of mo E S are classified by linear algebra, as there is an analogue of the linearization theorem for maps at hyperbolic critical points. The linearization theorem lor maps says that there is a CO coordinate chart on S, in which the local representative of S is a linear map. Recall that in the hyperbolic case. the spectrum of this linear isomorphism avoids the unit circle. The eigenvalues inside the unit circle determine the generalized eigenspace of contraction. This is the inset of mo E S. under the iterates of S. The eigenvalues outside the unit circle similarly determine the outset of mo E S. Although this argument provides only local CO submanifolds, the global stable manifold theorem improves this: the inset and outset of a fixed point of a diffeomorphism are smooth, injectively immersed subrnanifolds meeting transversally at mo. Returning to closed orbits, the inset and outset of y C M may be visualized by choosing a section S", at every point me y. The inset and outset of y in m intersect each section S", in submanifolds of Sm. meeting transversally at me y. In fact, In(y) is a cylinder overy, that is, • bundle of injectively immersed disks. So, likewise, is Out( y). And these two cylinders intersect transversally in y, as shown in Fig. 4.3.11. These bundles need not be trivial. Another argument is sometimes used to study the inset and outset of a closed orbit, in place of the Poincare section technique described before, and is originally due to Smale (1961). The flow F, leaves the closed orbit invariant. A special coordinate chart may be found in a neighborhood of '1. The neighborhood is a disk bundle over y, and the flow.. is a bundle map. On each fiber, F, is a linear map of the form Z,e R " wht.:re Z, is a constant, and R is a linear map. Thus, if s "'" (m, f) is a point in the chart, the local representative of F, is given by the expression F,(m, J)= (m"Z,e R , . / )
called the Floquet normal/orm. This is the linearization theorem lor closed
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
257
Figure 4.3.11
orbits. A related result. the F/oquellheorem. eliminates the dependence of Z, on I. by making a further (time-dependent) change of coordinates (see Hartman [1973. ch. 4. sec. 6). or Abraham and Robbin (1967)). Finally. linear algebra applied to the linear map R in the exponent of the Floquet normal form. establishes the CO structure of the inset and outset of y. To get an overall picture of a dynamical system in which all critical elements (critical points and closed orbits) are hyperbolic. we try to draw or visualize the insets and outsets of each. Those'with open insets are attraclors. and their open insets are their basins. The domain is divided into basins by the separatrices. which includes the insets of all the nonattractive (saddle-type) critical elements (and possibly other. more complicated limit' sets, called chaotic atlractors. not described here.) We conclude with an example of sufficient complexity, which has been at the center of dynamical system theory for over a century. 4.3.18 Example. The simple pendulum equation (4.3.5) may be "simplified" by approximating sin x by two terms of its MacLaurin expansion. The resulting system is a model for a nonlinear spring with Imear damping.
x=v.
258
VECTOR FIELDS AND DYNAMICAL SYSTEMS
Adding a periodic forcing term, we have
x-v, k
b- - cv - kx + 3xl + FooswI.
This time-dependent system in the plane is transformed into an autonomous system in a solid ring by adding an angular variable proportional to the time, , - wI. Thus
x= v;
v- - cv -
kx
+
iXl +
Fcos(J;
IJ= w. Although this was introduced by Baron Rayleigh to study the resonance of tuning forks, piano strings, and so on, in his classic book Theory of Sound (1877J, this system is generally named after Duffing who obtained the first important results (see Duffing [1908] and Stoker (1950».
Figure 4.3.12
THE QUALITATIVE THEORY OF DYNAMICAL SYSTEMS
259
Depending on the values of the three parameters (c; k, F) various phase portraits are obtained. One of these is shown in Fig. 4.3.12, adapted from the experiments of Hayashi (1964]. There are three closed orbits: two' attracting, one of saddle type. The inset of the saddle is a cylinder topologically, but the whole cylinder revolves around the saddle-type closed orbit. This cylinder is the separatrix between the two basins. For other parameter values the dynamics can be chaotic (see for example. Holmes (1979a,b] and Ueda (1980». Exercl...
4.3A"
(i)
Let E -> M be a vector bundle and mn EM an element of the zero section. Show that TmoE is isomorphic to TmnM E9 Em" in a natural, chart independent way. (ii) If E: M -> E is a section of E, and E(m o) = 0, define E'(mo): TmoM -> Emo to be the projection of TE< m n ) to Em". Write out E'( mo) relative to coordinates. (iii)
Show that if X is a vector field. then X'( mo) defined this way coincides with Definition 4.3.1.
4.3B
Prove that the equation 1J+2klJ- qsin(J = O(q > 0, k > 0) has a saddle'point at (J = 0, IJ= O.
4.3C
Consider the differential equations ;= ar' - br. 6= I using polar coordinates in the plane. (i) Determine those a, b for which this system has an attractive periodic orbit. (ii) Calculate the eigenvalues of this system at the origin for various a, b.
4.3D Floquet's theorem in differential equations deals with the fundamental matrix solution of a linear homogeneous system x= A(t)x, where A is a T-periodic matrix. The fundamental solution G(t) satisfies G= AG and G(O) = I. Floquet's theorem states that G(t) has the form P(t)e·' where P is I-periodic and B is constant. Look up the proof in Hale (1969) or Hartman (1973). (ii) Relate this result to the stability analysis of periodic orbits, as outlined in the text. (i)
4.3E
Let X E 'X (M), cp: M -> N be a diffeomorphism and Y = cp. X. Show that (i) mE M is a critical point of X iff cp(m) is a critical point of Y and the characteristic exponents are the same for each. (ii) y c M is a closed orbit of X iff cp( y) is a closed orbit of Yand their characteristic multipliers are the same.
260
4.3F
VECTOR FIELDS AND DYNAMICAL SYSTEMS
The energy for a symmetric heavy top is
pJ
+ J + MglcosO where I, J> 0, b, Po/!' and Mgl> 0 are constants. The dynamics of the top is described by the differential equations 0= aHlflO. p, = - aHlaO. (i) Show that 0 = O. p, = 0 is a saddle point if 0 < Po/! < 2{Mgli (a slow top) (ii) Verify that cosO = 1- ysech2({1Fi 12). where y = 2- b 21P and P= 2MgljI, describes both the outset and inset of this saddle point. (This is called a homoclinic orbit.) (iii) Is8 ... 0, p, =- 0 stable if Po/! > 2..jMgll? (Hint: Use the fact that H is constant along the trajectories.) 4.4 FAOBENIUS' THEOREM AND FOLIATIONS
The three main pillars supporting differential topology and calculus on manifolds are the implicit function theorem, the existence theorem for ordinary differential equations, and Frobenius' theorem. which we discuss brieny here. First some definitions:
*
4.4.1 Definition. Let M be a mani/old and let E c TM be a subbundle 0/ its langent bundle. (So E is a vector bundle over M and is a submani/old 0/ TM by means 0/ a vector bundle chart 0/ E.) We call E a dutributiorl (or a p/IIM /kill) on M.
We say E is invoIutiw if/or any two vector /ields X and Y defined on \open sets 0/ M and which take values in E,[X. YJ takes values in E as well. (ii) We say E is iII,egrabk if for any mo E M there is a (local) submanifold N eM. called a (local) iIItegrvl numi/oill of E at mo containing mo whose tangent bundle is exactly E restricted to N. (i)
The situation is shown in Fig. 4.4.1.
*According to Lawson (1977). the theorem or Frobenius is due to A. Clebsch and F. Deahna.
FROBENIUS' THEOREM AND FOLIATIONS
281
E"'II
Figure 4.4.1
4.4.~
Examples A. Any subbundle E of TM with one-dimensional fibers is involutive; E is also integrable. and this is seen in the following way. Using local bundle charts for TM at mo e M with the lIubbundle property for E, we can find in an open neighborhood of mo. a vector field that never vanishes and has values in E. Its local integral curves through mo have as their tangent bundles, E restricted to these curves. The situation is nicer if the'vector field can be found globally and has no zeros. Then through any point of the manifold there is exactly one maximal integral curve of the vector field. and this integral curve never reduces to a point. We shall return to this crxample later. B. LCt I: M ~ N be a submersion and consider the bundle ker TI c TM. This bundle is involutive since for any X. Y e <X( M). where X, Y take values in ker TI, TIl X, Y) == 0 by' 4.2.25. The bundle is also integrable since for any mo e M, the submanifoldr l(f(mo» has the restriction of ker TI to itself as its tangent bundle (see Section 3.5). C. Let T" be the n-dimensional torus, n ~ 2. Let I ~ k ~ nand consider E-«vl .... ,v,,)eTT"lvHI == .. • = v" = O}. This distribution is involutive and integrable; the integral manifold through (t I ' " .• In) is T k x (/HI'· ... /,,). D. E =
TM is involutive and integrable; the integral submanifold through any point is M itself. E. An example of a noninvolutive distribution is as follows. Let M == S0(3), the rotation group (see Exercise 3.5S). The tangent space at
262
VECTOR FIELDS AND DYNAMICAL SYSTEMS
I - identity consists of the 3 X 3 skew symmetric matrices. Let
E,- {AeT,SO(3)A-
[~
o o -p
a two-dimensional subspace. For Q E S0(3), let
Then
is a distribution but is not involutive. In fact. one computes that the two vector fields with p "'" I, q = 0 and p = O. q = 1 have a bracket that does not lie in E. (Some additional background in the theory of Lie groups makes this example natural). • Frobenius' theorem asserts that the two conditions in 4.4. J are equivalent.
4.4.3 Frobenlu.' Theorem. only if it is integrable.
A subbundle E of TM is involutive if and
Proof. The" ir' part is obvious since the bracket of two vector fields on a. manifold N is again a vector field on N. If E has one-dimensional fibers, Frobenius' tneorem amounts to the existence theorem for differential equations. The proof of the "only ir' part actually proceeds along these lines. By choosing a vector bundle chart, one is reduced to this local situation: E is a model space for the fibers of E, F is a complementary space, and U x VeE x F is an open neighborhood of (0,0), so U x V is a local model for M. We have a map
f: UXV-L(E,F)
such that the fiber of E over (x, y) is E(J<,Y)"'" {(".j(x. Y)''')I"E E) c E x F,
and we can assume we are working near (0.0) and f(O, 0) = identity.
FROBENIUS' THEOREM AND FOLIA TlONS
263
The condition of involutivity is eaSy to work out. One computes the bracket of two vector fields of the form
X(x. y) - ( .. (x. y)./(x. y)· ..
e... y».
Y(x. y) = (b(x. y)./(x. y)·b( .... y». Using the local formula for [X, Y). one finds that E is integrable iff at a poini (x. y)
D/·( .. ./· .. )·b= D/·(b./·b),"
(I)
for all II, bEE. Consider these two time-dependent vector fields on U X V:
X/ex. y) = (0. I( IX. y )·x) and
Y,(x. y) = (x. 1(lx. y )·tx). These are chosen with the form of E(JC •.V) in mind. so Yo is simple. but especially so that
dY, [ X/,.Y, ] + dt = O.
(2)
Condition (2) is easily verified (it does not use (I)). From it, we gel d
dt P,'Y, - 0 where F, is the flow of X" Fo = identity. Since X,(O,O) .. 0, we can assume F, is defined for 0 ... I ... 1. Thus
Since YI(x, y) = (x, I(x. y)x) and Yo(X, y) - (X,O). we ~pect FI to be a maps E(JC •.v) to E X{O} and be local diffeomorphism of U X V such that the required coordinate change to complete the proof. . Let N= FI(U X (Yo}), We shall complete the proof by showing that if (x, y) EN, then
r.
Clearly, since the first component of X, is zero,
F,(x, y). - (x, cp(t, x, y»
2tU
VECTOR FIELDS AND DYNAMICAL SYSTEMS
for a map .,(1, x, y) e F. A typical tangent vector to F,(U X{Yo}) is clearly (." D...,(I, x, Yo)·.,),
.,e E.
Thus the proof is complete if we establish this identity:
0 ...,(1, x, Yo)'" ·'/('X, .,(1, x, Yo»·." which forI-I shows that the tanaent space to F,(U X (Yo}) at (x,.,(x, Yo,I» equals E( ... y,' Indeed, they are equal at 1- 0 and
a
a, .,(1, x, Yo) = I(IX,.,(I, x, Yo»'x so
a
at 0 ...,(1, x, Yo)'" -I0.l( lx, .,(1, x, Yo»''''x
+ 0d(IX,.,(I, x, Yo»' (0...,(1, x, Yo)''')'x + 1(IX,.,(t,x, Yo»''''
Also,
(3)
a
a, tl( tx, .,(1, x, Yo»'" "" I( tx, .,(1, x, Yo»'"
+ to.l(tx,.,(t, x, Yo»'x'" + tDd(IX,.,(t, x, Yo»' (J(tx, .,(t, x, Yo»'x ) . ., (4) For x, Yo' ., fIXed, let
+(t) - O".,(t, x, Yo ) . ., - II( lx, .,(1, x, Yo»·., so that IjI(O) = O. Subtracting (3) and (4) and using (1) we find thai
Hence by Gronwall's inequality, or by uniqueness of solutions of differential equations, +(1) - 0, and the result follows. •
The method of using the time-one map of a time-dependent flow 10 provide the appropriate coordinate change is useful in a number of situations. Sometimes these are called "Lie transforms." We shall use this
FROBENIUS~
THEOREM AND FOLIATIONS
265
method again in Chapter 6 to prove the Poincare lemma and in Chapter 8 to prove the Darboux theorem. For a direct proof of the Frobenius' theorem from the implicit function theorem using manifolds of maps in the spirit of Box 4.IC. see Penot (1970). The Frobenius theorem is intimately connected to the global concept of loliations. Roughly speakirig, the local manifolds N obtained berore can be glued together to form a "nicely stacked" family of submanifolds filling out M (see Fig. 4.4.1 or Example 4.4.2A).
y
--+---x
figure 4.4.2
4.4.4 DeflnlUon. Let M be a manifold and 4» = {eo}.. E If a partition of M into disjoint connected sets. 4» is called a /oliation if each point 0/ M has a chart (V, cp), cp: V -+ V' X V' c E $ F sueh that for each the connected V' X{c~}, components (V e .. )' of V f .. are given by cp«V where c~ E F are constants for each II E.A and fl. Sueh charts are called /oIitItItJ (or distinpUW) by ell. The dimsuion (respectively, etJtIinwIuion ) 0/ .tile /oliation 4» is the dimension of E (respectively, F). See Fig. 4.4.2. .
n
e. n e.>') -
n
en
Note that each leaf is a connected immersed submanifold. In general, this immersion is not an embedding; i.e., the induced topology on eo from M does not necessarily coincide with the topology of ea (the leaf ea may accumulate on itself, for example). A differentiable structure on ea is induced by the foliated charts in the following manner. If (V, cp), cp: V -+ V' X V' c E$F is a foliated chart on M, and x: E$F -+ E is the canonical. projection, then X 0 cp restricted to (U defines a chart of
n eo>'
e...
4.4.5 Ex....pI.. A. The trivial /oliation of a manifold M has only one leaf. M itself. It has codimension zero. If M is finite dimensional both M and the leaf have
266
VECTOR FIELDS AND DYNAMICAL SYSTEMS
the same dimension. Conversely, On a finite-dimensional manifold M a foliation of dimension equal to dim(M) is the trivial foliation. B. The discrete foliation of a manifold M is the only zero-dimensional foliation; its leaves are all points of M. If M is finite dimensional, the dimension of M is the codimension of this foliation. C. A vector field X that never vanishes on M determines a foliation; its leaves are the maximal integral curves of the vector field X. The fact that this is a foliation is the straightening out theorem (see Section 4.1). D. Let f: M - N be a submersion. It defines a foliation on M (of codimension equal to dim( N) if dim( N) is finite) by the collection of all connected components of I( n) when n varies throughout N. The fact that this is a foliation is Theorem 3.5.5. In particular EeF is foliated hy the family {E x {f}}, e F· •
r
We next define the tangent bundle to a foliation.
4.4.6 ProposlUon. Let M be a manifold and ell = {e.. }.. e A a foliation on M. The set T( M, ell) = u "e A U me "Tmf. .. is a subbundle of TM called the ""'gmt InuuIk to the foliation. The quotient bundle JI(eII) = TMjT( M, ell), is called the "OrmaJ InuuIk to the foliation ell. Proof. Let (U, fJI), fJI: U - U' x V' - E e F be a foliated chart. Since TufJI(T): .. ) = E X (O) for every u E U () we have TfJI(TU () T( M, = (U' x V') x (E x {O». Thus the standard tangent bundle charts induced by foliated charts of M have the subbundle property and naturally induce vector bundle charts by mapping v'" E T",(M, ell) to (fJI(m), T",fJI(v",» E (U' X V')x(E x (O». •
ell»
e.. ,
Now we can reformulate Frobenius' theorem in terms of foliations.
4.4.7 Theorem. lent: (i)
(ii) (iii)
Let E be a subbundle of TM. The following are equiva-
there exists a foliation ell on M such that E E is integrable; E is involutive.
=
T(M,eII);
Proof. The equivalence of (ii) and (iii) was proved in 4.4.3. Let (i) hold. Working with a foliated chart, E is integrable by 4.4.{j, the integral submanifolds being the leaves of ell. Thus (ii) holds. Finally, we need to show that (ii) implies (i). Consider on M the family of (local) integral manifolds of E. each equipped with its own submanifold topology. It is straightforward to verify that the family of finite intersections of open subsets of these local integral
FROBEN/US' THEOREM AND FOLIATIONS
267
submanifolds defines a topology on M, finer in general than the original one. Let {E .. }ae A be its connected components. Then, denoting by (fa n V)fj, the connected components of Ea n V in V, we have by definition that E(une.)fl=(E restricted to (Vne,,)fj) equals ~«Vne .. )fj). Let (.,.- I( V), y,), V c M be a vector bundle chart of TM with the subbundle property for E, and let!p: V -+ VI be the induced chart on the hase. This means, shrinking V if necessary, that!p: V -+ V' X V' c E(fJF and y,(.,.- I( V) nE}=(V'XV')X(EX{O}). Thus !p«Vne .. )fj)=u'X{c~} and so M is foliated by 'the family ell == (Ea}ae A' Because E(u() I'.}" = T«V n C,,)fj). we also have T(M,eII) = E . • There is an important global topological condition that integrable subbundles must satisfy that was discovered by Bou 11970]. The result, called the Bott vanishing theorem can be found, along with related results. hy readers with background in algebraic topology, in Lawson [1977]. Let R denote the following equivalence relation in a manifold M with a given foliation ell: xRy if x, y belong to the same leaf of ell, It is of interest to know whether M / R is a manifold. Foliations for which R is a regular equivalence relation are called regular foliations. (See Section 3.5 for a discussion of regular equivalence' relations.) The following is a useful criterion.
4.4.8 Propoaltlon. Let ell be a foliation on a manifold M and R the equivalence relation in M determined by ell. R is regular iff for every' me M there exists a local submanifold Em of M such that Em intersects every leaf in at most one point (or nowhere) and TmEmal Tnr( M, ell) = Tm M. (Sometimes Em is called a slice or a local cross-section for the foliation,)
Proof. Assume that R is regular and let ",: M -+ M/ R be the ca!lonical projection. For Em choose the submanifold using the following construction. Since ", is a submersion, in appropriate charts (V,!p). (V, y,). where !p: V -+ U' x V' and y,: V -+ V'. the local representative of",. "'",>/-: U' X V' -+ V' is the projeCtion onto the second factor (3.5.5) and every leaf",- I( v) C V. v e V is represented in these charts as V' x {v,}, where (,' = y,( v). Thus if Em = y,- I({O}X V'), we see that Em satisfies the two required conditions. The converse is obtained by reversing the argument. •
To get a feeling for the foregoing condition we will study the linear flow on the torus.
4.4.9 Example. On the two-torus T 2 consider the glohal now F: R X T 2 -+ T2 defined by F(t.(sl' S2»'" (s le 2,"', S2e2",a,) for a fixed number lie
268
VECTOR FIELDS AND DYNAMICAL SYSTEMS
[0.1[. By 4.4.5C this defines a foliation of T2. If a E Q. notice that every integral curve is closed and that all integral curves have the same periqd. The condition of the previous theorem is easily verified and we conclude that in this case the equivalence relation R is regular; T 2/ R = s I. If a is irrational the situation. however. is completely different. Let cp(t) '= (e 2w ;/. e 2w ;a/) denote the integral curve through (I. I). The following argument shows that cl( cp(R» = T 2; that is. cp(R) is den.ve in T 2. Let p(e 2"'''.e 2'''Y)E T2; then for all mE Z.
where y = ax + z. It suffices to show that C = {e 2 ,,;m,. E Slim E Z} is dense in SI because then there is a sequence m, E Z such that e 2";m,,. converges to e 2 .. ;t. Hence. cp(x + mil) converges to p. If for each k E Z+ we divide SI into k arcs of length 2f1/k. then. because {e 2..;"'·ES l lm-I.2..... k+.I} are distinct. for some 1 < n, < m, < k + 1. e 2";m,,. and e 2f1 ;ro,,. belong to the same arc. Therefore. le 2,,;m.,. - e1 ..;ro'-1 < 2f1/k. which implies le 2",P.- -II < 2f1/k. where Pte - m" - n". Because
U {e 2 ..tas E Slls E (jp".(j + I)p,,]) = S... jEZ.
every arc of length less than 2'IT/k contains some e 2,,;jP •• which proves c1(C)=SI. Thus any submanifold Em.mET2. not coinciding with the integral curve through m will have to intersect cp(R) infinitely many times; the condition in the previous theorem is violated and so R is not regular. • Even though the concept of foliations seems encompassing enough to study "nice" partitions of a manifold into submanifolds. there are important situations when foliations are unsatisfactory because they are not regular or the leaves jump in dimension from point to point. Consider for example. R2 as a union of circles. Here the origin is a singularity. As this example suggests. one would like to relax the condition that M / R be a manifold. provided that M/R turns out to be a union of manifolds that fit "nicely" together. Stratifications. a concept generalizing foliations. turn out to be the natural tool to describe the topology of orbit spaces of comp~ct Lie group actions (see. for instance. Bredon [1972]. Burghelea. Albu. and Ratiu [1975]. Fischer [1970]. and Bourguignon [1975]). We shall limit ourselves here only to the definition in the finite-dimensional case and some simple remarks. 4.4.10 Definition. ut M be a locally compact topological space. A
/ktI,itM of M is a partition of M into manifolds {M";.,, , called
s,ra';lI,ra,,,.
FROBENIUS' THEOREM AND FOLIATIONS
269
-YlllisfYing the following conditions: (S I)
(S2)
(S3) (S4)
M. are manifolds of constant dimension; they are .whmanifoldv of M if M is it.~elf a manifold; . ' The fami(v {M;} of connected components of all the M is a locally finite partition of M; i.e.,for every m E M. there exists an open nei!{hborhood U of m in M intersecting only finile(v many Mff. If M; ncl( Mff) .. 0 for (a. a)" (b, Pl. then M: eMf and dime MQ)'
From the definition it follows that if
M; ncl( Mil)
T
0 and if mE
M:;
c c1(Mff) has an open neighborhood U in the topology of M;; such that U c M: ncl(Mf). then necessarily M;; eMf and thus dim(Mu ) < dime M~). To see this, it is enough to note that the given hypothesis makes M;; ncl( Mf)
M:.
open, in Since it is also closed (by definition of the relative topology) and M: is connected. it must equal M: itself, whence M(~ c c1( Mf) and by (S3), M:c Mf and dim(M
Exercl... 4.4A
4.48
4.4C
Let M be a manifold such that TM = E1al ... alEp • where E, is involutive. Show that t~ere are local charts ( U, q> ). q>: U ~ E I e ... eE, such that E, is described by the equations ~j - O. j "" i. In R4 consider the family of surfaces given by .\: + y2 + Z2 -t 2 = const. Show that these surfaces are the leaves of a foliation that is not regular. What part of R4 should be thrown out so as to obtain a regular foliation? Let f: M ..... N be a C'" map' and • a foliation on N. The map f is said to be transversal to ., denoted f7f'l. if for every me M, T",f(T",M) +1{(m)(N,.)=1{(m)N and (Tmf)-I(1{<",)(N,.» splits in TmM. Show that if (f.. }.. e ... are the leaves of., the connected components of f-I(f .. ) are leaves of a foliation (denoted by f*(.» on M, and if • has finite codimension in N, then so does the foliation/*(.) on M and the two codimensions coincide.
270
VECTOR FIELDS AND DYNAMICAL SYSTEMS
4.4D (Bourbaki (197ID. Let M be a manifold and denote by M' the manifold with underlying set M but with a different differentiable structure. Show that the collection of connected components of M' defines a foliation of M iff for every me M. there exists an open set U in M. m e U. a manifold N. and a submersion p: U -+ N such that the submanifold p-I(n) of U is open in M' for all n E N. (Hint: For the "if' part use 3.3.5 and for the "only if" part use Exercise 3.2F to define a manifold structure on the leaves; the charts of the second structure are (U () a )lI-+ U'.)
e
CHAPTER
5
Tensors
In the previous chapter we studied vector fields and functions on manifolds. Now these objects are generalized to tensor fields. which are sections of special vector bundles that generalize the tangent bundle. This study is continued in the next chapter when we discuss differential forms. which are tensors with special symmetry properties. One of the objectives of this chapter is to extend the pull-back and Lie derivative operations from functions and vector fields to tensor fields. ' 5.1
TENSORS ON LINEAR SPACES
Prepatory to putting tensors on manifolds. we must first understand them on vector spaces. This subject is an extension of linear algebra sometimes called .. multilinear algebra." Ultimately our constructions win be done on each fiber of a vector bundle producing a new vector bundle. The case of the tangent bundle win be the most important. As in Chapter 2. E. F •... denote Banach spaces and Lk(E••..•• Ek: F) denotes the vector space of continuous k-multilinear maps of E. x ... X Ek to F. The special case L(E.R) is denoted E*. the dual space of E. If ' I- (~ ••...• ~,,) is an ordered basis of E. there is a unique ordered _basis of E*. the dUal basis ,. ... (~I •...• ~"). such that ~J(e,)=8I;. where 81;=1 if j = i and 0 otherwise. Furthermore. for each ~ e E.
271
272
TENSORS
and for each a e E·, n
a=
E a(~/)~;' i-I
Employing the summation convention whereby summation is implied when an index is repeated on upper and lower levels, these expressions become ~ _ ~i(~)~1 and a - a(~I)~;' . We may map E into E" -= L(E·,R) by associating with each ~ e E,~*·e E*·, given by ~··(a)= a(~) for allae E*. If E has finite dimension, then E·· and E have the same dimension and as the map ~ .... ~.* is one to one, it is an isomorphism. In infinite dimensions. recall that E is reflexive when ~ ..... ~.* is an isomorphism. .
5.1.1 DeflnlUon. For a vector space E we put 1'.'(E)=L'-I··(E· •...• E*. E •...• E; R) (r copies of E· and s copies of E). Elements of 1'.'(E) are called 'MSOI'S ... Eo COII''''''''''''''t of order r and of order s; or simply. 01 typ. (~). Given '. e 1'.:'(E) and'2 e 1'.:'(E) the '~lISor protblc, of'l and'2 i.' the tensor "~'2 e r.:'++.7(E) defined by
COWl"'"
'1~'2( PI •...• P',. yl •...• y". II" ··.1.,. RI.··· .R,,)
='1 (pi, .... p". II ....•1., )'2 (yl •... , Y". '1.··· .R.,) where pl. yj e E*
and~"j
e E.
Replacing R by a space F gives r.'(E; F), the F-valued tensors of typl." The tensor product now requires a bilinear form Oil the value space for its definition. For R-valued tensors•• is associative. bilinear and continuous; it is not commutative. We also have the special cases Td(E)'" E**. Tlo(E)- E*. T2o(E)- L(E; E*). and TII(E)- L(E; E··) and make the convention that Too( E; F) ... F. (~).
5.1.2 Proposition. Let {e;};e I be an algebraic basis of E and {~j}jeJ an algebraic basis of E*; i.e .• every element of E (resp. E·) is a finite linear combination of elements of {e;}/e I (resp. (ej}jeJ)' Then an algebraic basis of r.'(E) is given by
In particular, if dim(E) = n, then r.'(E) has the structure of an n,+s-dimensional vector space.
TENSORS ON LINEAR SPACES
273
We must show that the elements. ® ... ® •.I, ®.iI® ... ®.J. of . 'I T,'(E) are linearly independent and span T,'(E). Suppose a finite sum ti""/j ....•/ ® .. , ®.j ®.it® ... ®.J, = O. Then apply this to (."" .... . • k"./:,.~·,,~/,> using tite identification .,(.') = .i(.,) to give t ", .. A",." = n, Next, we easily check that for I e T,'( E) we have a finite sum I == I( ." •... , Proo,' ,,-
.1,• 5j.9 • ••• ,5j, • ). ® ... ® •.I. ®.J.® . , . ®.i, • 51.
..
_
The coefficients t l ""1 1""j, ... ,(.;" ••••• '·.eit .... ,e,.) are called the components of I relative to i. .
5.1.3 Exampl.. A. If' is a (~) tensor on'E then' has components
tij = I(e j • •, ) . an n x n matrix. This is the usual way of associating a bilinear form with a matrix. For instance. in R 2 the bilinear form
,(x, y) = Ax,y, + Bx,yz + CX 2 Y, + D:(~r, (where x = (x,. x 2 ) and J' = Cv,. Y2» is associated to the 2 x 2 matrix
(~ ~). B. If' is a (~) tensor on E., it makes sense to say that' is symmetric; i.e.• t(., •• 2) = '(.2'.')' This is equivalent to saying that the matrix ,'/ is symmetric, Symmetric (~) tensors I can be recovered from their quadratic form Q(e) = 1( ••• ) by 1(., •• 2) = HQ(., + .2)- Q(., -e2»)' the polarIzation identi~v. If E = R 2 and t has the matrix
(~ ~)
(n
then Q(.l') .... Ax~ + 2Bx,x 2 + Cx~. Symmetric tensors are thus closely related to quadratic forms and arise. for example. in mechanics as moment of inertia tensors and stress tensors. C. In general. a symmetric (~) tensor is defined by the condition
t( a' , .... a') = I( a"(1) ..... a"e,) for all permutations of (J of {I ..... s}. and all elements a' ..... a· e E·. On~ may associate to I a homogeneous polynomial of degree k. P( a) ... ,( a •. , .. a) and as in the case s == 2. P and, determine each other. A similar definition holds for (~) tensors. It is clear that a tensor is symmetric iff all its components in an arbitrary basis are symmetric.
274
TENSORS
D. An inner product (.) on E is a symmetric (~) tensor. Its matrix is often written gij - (_/'_j)' Thus gij is symmetric and positive definite. The inverse matrix is written gij. Eo The space L"(E, ..... E,,; F) is isometric to L"(E,,(,) ..... Eo (,,); F1 for any permutation a of {I •...• k}. the isometry being given by A ..... A, A('.. (1) ..... ' ..(k» - A(,' ... ·.'k)· Thus if , E 1'.'(E; F). the tensor' can be regarded in (,;,) ways as an (r + s )-multilinear F-valued map. For example. if t E Tl( E). the standard way is to regard it as a 3-linear map t: E* x E* x E -+ R. There are two more ways to interpret this map. however. namely as E* x E x E* -+ R and as E x E* x E* -+ R. In finite dimensions. where one writes the tensors in components. this distinction is important and is reflected in the index positions. Thus the three different tensors described above are written
F. In classical mechanics one encounters the notion of a dyadic (cc. Goldstein (1980». A dyadic is the formal sum of a finite number of dyads; a dyad being a pair of vectors ',.'2 ERl written in a specific order in the form "'2' The action of a dyad on a pair of vectors. called the double dot product of two dyads is defined by
','2: ","2 = ("'''')('2'''2)' where . stands for the usual dot product in Rl. In this way dyads and dyadics are nothing else but (~) type tensors on R3; i.e.• ','2 E T20(R 3), by identifying (R 3)* with R3. G. Higher order tensors arise in elasticity and Riemannian geometry. In elasticity the stress tensor is a symmetric 2-tensor and the elasticity tensor is a fourth-order tensor (see Marsden and Hughes [1983]). In Riemannian geometry the metric tensor is a symmetric 2-tensor and the curvature tensor is a fourth-order tensor (see Spivak [1979]). A The classical operations of tensor algebra are the interior product and the contraction. These are defined as follows. The inlerior product of a vector vEE (resp. form P E E*) with a tensor t E 1'.'( E; F) is a ~,) (resp. (' • ') type F-valued tensor defined by
L
(iot)( P' ..... P'. v, •.... v,_,) =
1(14' ... .• P'. v. v, .... ,v,_ ,)
(i")( P' ... ·.P'-'. v, ..... v,) "" '(14.14' .... ,14'- '. v" .... v.).
TENSORS ON LINEAR SPACES
275
Clearly;,,: 1'.'(E; F) -+ 1'.'-I(E; F). ;-: 1'.'(E; F) -+ 1'.' I( E; F) are linear. continuous maps. as are v ..... ·;".p ..... ,'fS. If F ... R and dim(E)=n these operations take the following form in components. If e A Crespo e k ) denotes the kth basis (resp. dual basis) element of E. we have
By 5.1.2 these formulas and linearity enable us tn compute anv interior product. The contraction of the k th j::Ontravariant with the Ith covariant index. or for short. the (k./)-conlraclion. is a family of linear continuous maps C/: 1'.'(E) -+ 1'.'_-II(E) defined for any pair of natural numbers r. s ~ I by
eN VI® ... ®V,®pl® ... ®p') -P'(Vk}V I ® .... ®t3k ® .. · ®V,®pl® ... ®9'®", ®P'. means that the symbol underneath is omitted. If dim(E) = nand E E* denote a basis and the dual basis. the expression for the contraction operation in components takes the form where
e; E E. e j
The Kronecker delta is the tensor 3 E TII(E) defined by 8(a, e) = aCe). If E is finite dimensional. 3 corresponds to the identity I E L(E; E) under the canonico:al isomorphism TII(E) .. L(E; E). where we have identified E** with E. Relative to any basis. the components of 3 are the usual Kronecker symbols 8j; i.e., 3 = 8je;®e j • Suppose E is a finite:dimensional real inner product space with a basis el, ... ,e" and corresponding dual basis el ..... e" in E*. Using the inner product. with matrix denoted by gil' we get isomorphisms b: E -+ E* ; e ...... (e . . )
and its inverse . 0f The matnx·
b'IS gij; I.e., •
#:E*-+E.
276
TENSORS
. and of • is gil; i.e.•
where e i and (Xj are the components of e and Cl. respectively. We call b the index lowering operator and I the index raising operator. These operators can be applied to tensors to produce new ones. For example if t is a (~) tensor we can define an associated tensor I of type ( : ) by
I(e. Cl) ... ,(e. cxI). The components are
i/ =
gJktilc
(as usual. sum on k).
In the classical literature one writes t/ for gJktik' and this is indeed a convenient notation in calculations. However. contrary to the impression one may get from the classical theory of Cartesian tensors, , and I are different tensors. Confusion over this point can lead to trouble. 5.1.4 Exam.,.... Let E be a finite-dimensional real vector space with basis el ..... e" and dual basis el •...• e". A. If t E T I2(E) and e = eie,. then
Thus. the components of ill' are ePt k',. The interior product of the same tensor with Cl- Clpe P takes the form
B.
If' E Tl(E). the (2. I)-contraction is given by
C. An important particular example of contraction is the trace of a ( : ) tensor. Namely. if , E Til ( E). then
tr(,} = CII(t)
= ti,.
TENSORS ON LINEAR SPACES
277 '
D. The components of the tensor associated to g hy raising the second index are gil = gikgj/< = gJkgk; = 6 J,. E. Let e ®eJ ®e k ®e'®e"/ • t E '7'3(E) '2 • t = ,;jk 1m, Then t has quite a few associated tensors. depending on which index is lowered or raised. For example
and so on. F. The positioning of the indices in the components of associated tensors is important. For example if t E T2o( E). we saw earlier that 1/' = gJk',k' However Ii; = gJkl k ,. which is in general different from Ii' when tis not symmetric. For example, if ~ = R3 with gij = 6/, and the nine components of , in the standard basis are '12 = I. 121 = - 1.1" = 0 for all other pairs i. j. then l,i = Iii' Iii = 'ji' so that 112 = 112 = I while 121 = 121 = - I. G. A convenient abuse of language often occurs in the physics literature when considering traces of tensors that are not of type ( :) but which can be brought to this type by lowering and raising indices. i.e.• tensors of type (~) and (~). The trace of a (~) tensor is thus defined to be the trace of the associated (:) tensor; i.e.• if t = lije;®ej • then tr(t) ... Iii'" giklik. The question naturally arises whether we get the same answer by lowering the first index instead of the second. i.e., if we consider Ii;' By symmetry of gii we have 111 - gkllik,.. 1/. so that the definition of the trace is independent of which index is lowered. Similarly. if IE T20(E). trCt) = 1/; - g,kl'k ... (kk' In particular tr(l') == gi 1 == glkg;k - dime E). '" Now we tum to the effect of linear transformations on dual spaces. If cpEL(E.F), the Iranspose of cpo denoted cp*EL(F*.E*) is defined by cp*(p)·e = p(cp(e». where PE F* and e E E. Let us analyze the matrices of cp and cp*. As customary in linear algebra, ve<;tors in a given basis are represented by a column whose entries are the components of the vector. Let", E L(E. F) and let e= (el ..... en ) and i=(/I ...../m) be ordered bases of E and F respectively. Put cp(e;) =
278
TENSORS
A"; fa. (We use a diffetent dummy index for the F-index to avoid confusion). This defines the matrix of 4p; A = (A";). Thus. for" = t"e, E E the components of 4p(v) are given by cp(v)" = A";";. Hence. thinking of" and cp(v) as column vectors, this formula shows that cp(,,) is computed hy multip(l'illg v on the kft by A. the matrix of 4p. as in elementary linear algebra. Thus the upper index is the row index, while the lower index is the column index. Consequently. 4p(~I) represents the ith column of the matrix of 4p. Let us now investigate the matrix of 4p* E L ( r. E*). If rand are the dual ordered bases of i and J respectively. then cp*(r)·~;=- r'cp(~I)"" A";6"" - A GI, i.e.• 4p*( AG;~; and thus the ith component of cp*( Is A";. Hence 4p.( is the ath row of A. Consequently the matrix 0/ cp* is the transpose of the matrix 0/ cpo If P- p" E r then cp*( P) - p"cp*( PQAQ;~I. which says that the ith component of 4p·(P) equals cp·(P); = /l"A";. Thinking of elements in the dual as rows whose entries are their components in the dual basis. the foregoing calculation says that cp*( P) is computed by multiplying P on the rigllt by A. the matrix of cpo again in agreement with linear algebra. Now we tum to the effect of linear transformation on tensors. We start with an induced map that acts "forward" like 4p.
r
r) -
r)
r
r·A"d,,r) r) -
5.1.5 Definition. If 4p E L(E. F) is an isomorphism, define the push-lor'ttffIrfI of 4p, T,'4p = 4p. E L(T/(E). T,'(F» by 4p.'(p', ...• P',
I" ... ,/,) = t(4p*(P'), ... ,4p*(P'),cp-'( I,),· .. ,cp-'( /,»)
where IE T,'(E), P' •.... P' E
r. and I"
...•/, E F.
We leave the verification that CP. is continuous to the reader. Note that T{'4p - (4p - ') •• which maps" forward" like cpo If E and F are finite dimensional. then 1'0'(£) - £. To'(F) - F and we identify 4p with To'4p. The next proposition asserts that the push-forward operation is compatible with compositions and the tensor product. 5.1.8
P~ltIon.
Let cp: E .... F and IjI: F-+ G be isomorphisms. Then
(i) (+04p)._IjI.04p •. (ii) (iii)
(iv)
II i: E
-+
£ is
th~
identity. then so is i.: T,'(E) .... T,'(E);
cp.: T,'(E)-+ T,'(F) is an isomorphism, and(4p.)-' = (cp-'). If I, E T,:'(E) and '2 E T,:Z( E). then 4p.(I,®t 2 ) = CP.(t, )®4p.(t2)
TENSORS ON LINEAR SPACES
Proof.
279
For (i),
".(tp.t)(yl, ... ,Y'.81' ...• 8.) .
== tp.t( ".(y1) •...• ".(y,).,,-1(81 ) ..... "-1(8.))
where yl •...• y' E G·. 81 ..... 8. E G. and t E T.'( E), We have used the fact that the transposes and inverses satisfy (" ° tp ). = tp. ° ". and (" ° tp) - I = tp - I ° ,,- I, which the reader can easily check. Part (ii) is an. immediate consequence of the definition and the fact that i· = ; and ;- I =;. Finally, for (iii) we have tp.o(tp-I). =;., the identity on T.'(F), by (i) and (ii). Similarly, (tp-I).0tp. = i. the identity on T.'(E). Hence (iii) follows. Finally (iv) is a straightforward consequence of the definitions. . , Since (tp-I). maps "backward" it is called the pull·back of tp and is denoted tp.. The next proposition gives a connection with component notation.
5.1.7 ProposlUon. Let tp E L(E. F) be an isomorphism of finite dimensional vector spaces. Let A"i denote the matrix of tp in the ordered ba.fes e of E and J of F, i.e. tp(.,) = A"J". Denote by (B i,,) tire matrix of cp- I. i.e. tp' I( /,,) _ B i"./. Then (B',,) is the intlerse matrix of ( A",): B i" A", = 6'/, Let t e T,'(E) with components t~' "Il'" 'J. relatille tn I and q ~: '1;'( F) with .('omponents q"'" '~'''' '''''. relative to f. Then the components /If cp. t relative to / and of tp., relative to I are given respectively by
·Proof. We have .i = tp-I(tp(.i»'" tp-I(A"J,,) = A"/tp-I( I,,) == A~IBi..1' whence Bi"A"/ = 6i / for all i, j. Similarly. one shows that A"/B i" - 3"". so
280
TENSORS
- t{ AU,
.',
,~.....
I
AU,,It;. .i, Bl, h5," • ,
I
I
••••• Bi •
• h--,· I
•
)
To prove the second relation. we need the matrix of (cp ')* e I. ( E*. We have
so that (cp -. )*(e') = case. •
Biu
r).
r. Now proceed with the proof as in the previous
Note that the matrix of (cp 1)* e L (E*, r ) is the transpose of the inverse of the matrix of ._ The assumption that cp be an isomorphism for •• to exist is quite restrictive but clearly cannot be weakened. However, one might ask if instead of "push-forward," the "pull-back" operation is considered, this restrictive assumption can be dropped. This is possible when working with covariant tensors, even when cp E L(E, F) is arbitrary.
5.1.8 Definition. If. E L(E, F) (not necessarily an isomorphism). define the pull-back cp*e L(T.,o(F), r.°(E» by cp*,( e l , ... ,e.) "" ,( cp( e.), ... ,cp( e.
n,
The next proposition asserts that cp* is compatible with compositions and the tensor product. Its proof is almost identical to that of 5.1.6 and is left as an exercise for the reader.
5.1.9 Propo.ltlon. Let cp e L (E; F ),1jI e L (F; G). Then (1jI 0.)* = cp*o 1jI*. (ii) If;: E .... E is the identity, then so is ;* e 1.( 0 ( E). 0 ( E». (iii) If cp is an isomorphism. then so is cp* and cp* = (cp • )*. (iv) If'l E r.~( F) and t2 e r.~( F), ,hen cp°(t.®t 2) = (cpo,. )®( cp*1 2 ). (i)
r.
r.
TENSORS ON LINEAR SPACES
281
'Finally the components of 1f*1 are given by the following. 5.1.10 propotltlon. LeI E and F be finile-dimensionl!' vec!or sPlfces and. cp E L(E, F). For ordered bases i= (e l , .. . ,en ) of E and 1= (/I .... ./m) of F. suppose that cp( e i ) = AQ'/Q' and leI t E T.0 ( F) have components I",h, relative 10 j. Then Ihe componenls of cp*t relalive 10 i are given hI' ( m*l) .... ,,"'j, - , h",,", Ah,J,
PI'fJOf.
(1f*1)".,,'"
···A",
I,'
(If·,)(eft ... ·.e,,)
=
t{ If(e,,) •...• If(e,,))
=
t{ Ah'J,/", ..... A\,/",)
=
I{ Ih, .....I", )A'\" . A\
=
'h
LA"'· .. A'" j. • 11
I'· "',
Examples. A. On R 2 with the standard basis {e l • e 2 }. let t E + 2e l ®e 2 - e2®e l + 3e2®e2 and let If E 1_
To2(R2) be given by 1= el®e
r~
:].
since A has an inverse matrix given by B = [ ~ I -; I ]. According to 5.1.7. t~e components of i = CP*I relative to the standard hasls of R 2 are given by iii=Biu 8 j "I U " ,
with 811=1,821=812=-1
and
8?2=2
so that ,-12 _ 8 118 21,11 + 8 118 22,12 + 8 128 21,21 + 812822,22
- t· ( -\ ).\ + I· 2· 2 + ( - 1)( - I)· ( - I) + ( - 1)·2·3 , =-4
= (-1)·1·1+( -IH -1)·2+2·1·( - 1)+2·( - 1)·3 = ,-11 -
=
-7
.:·,n l /
I
+ 8 118 12 ,12 + 8 12 8 11,21 + R12R12t22
1 . I . I + I . ( - 1)·2 + ( - I)· J.( - I) + ( - I ) -( - 1)3
=4
282
TENSORS
and
- (-1)·( -1)·1 +( -1)·2·2+2·( -1).( -1)+2·2·3 -II.
Thus, cp., - 4el~el -4el~e2 -7e2~el + Ile2~e2. B. Let ''''el~e2-2e2~e2ETII(R2) and consider the same map cp E L(R2. R2) as in part A above. We could compute the components of CP.' relative to the standard basis of R 2 using the formula in 5.1.7 as before. An alternative way is to proceed directly using S.I.6(iv). that is. the fact that CP. is compatible with tensor products. Thus
C.
Let t - -
2el~e2
E T2o(R 2 ) and cp E L(R\ R 2) be given hy the
matrix A - [~ ~ I ~]. Again we shal\ compute cp., E T2o(~3) by using the fact that cp. is compatible with tensor products and that the matrix of •• E L(R2. R1) is the transpose of A. Recall that cp·(e i ) is the ith row, since matrices act on the right on covectors. Denote by /1. /2, /1 the standard dual basis of R1. Then cp·(e l ) ... /1 + 2/1 and ••(e 2 ) _ - /2 + /3. so that cp.(,) ... -2 •• (e l )®cp.(e 2) = -2( /1 +2/3)~( _/2 + /3) = _2/1~/2 -2/1~/3 +4/3~/2 _4/3~/3 . •
Exercl...
S.IA Compute the interior product of the tensor' = el~el~e2 + 3e2~e2~ e l with e'" - e l +2e2 and u= 2e l + e 2. What are the (1,1) and (2.1) contractions of ,., S.IB Compute all associated tensors of , ... e lee 2 ee 2 + 2e2~el~e2 e2eelee' with respect to the standard metric of R2.
TENSOR BUNDLES AND TENSOR FIELDS
5.lC
Let
1=
283
2e l ®e l - e2®e l + 3e l ®e 2 and cp E L(R 2 .R 2 ). '" E L(R3.R2)
[!
be given by the matrices I :]. [~ ~ -;1]. Compute: . tr(I). cp*(I). ",*(1). trcp*(I). tr",*(t). cp*'. and all associated tensors of t. cp*(t). ",*(t) and cp*(1) with respect to the corresponding standard inner products in R2 and R). 5.ID Let dim(E)-n and dim(E)=m. Show that T,'(E: E) is an mn'+'· dimensional real vector space by exhibiting a basis. 5.2
TENSOR BUNDLES AND TENSOR FIELDS
We now extend the tensor algebra to local vector bundles. and then to vector bun.dles. For VeE (open) recall that V X E is a local vector ·bundle. Then V X 1','(E) is also a local vector bundle in view of 5.1.2. Suppose cP: V X E -. V' X E' is a local vector bundle mapping and is an isomorphism on each fiber; that is. CPu = cpl{u}X EE L(E. E') is an isomorphism. Also. let CPo denote the restriction of cP to the zero section. Then cp induces a mapping of the local tensor bundles as follows. 5.2.1 Definition. If cp: V X E -. v' X F' is a /0('(1/ vector bundle mapping such that for each u E V. CPu is an isomorphism. let If *: V X T.'( E) -co V' X T.'( F') be defined by
where t
E
1','( E).
Before proceeding. we shall pause to recall some useful facts concerning linear isomorphisms. 5.2.2 ProposlUon. Let GL(E. E) denote the set of linear isomorphisms from E to E. Then GL(E. E) C L(E. E) is open.
This was proved in 2.5.4. Let us also recall 2.5.5. 5.2.3 ProposlUon. Let (i': L(E. E) -. L(E-. E*); cp ..... cp* 'and GL(E. E) -co GL(E. E); cp ..... cp-I. Then (£ and ~ are of class coo and
Smoothness of (£ is clear since it is linear.
~:
284
TENSORS
5.2.4 ProposlUon. If cP: V x F --+ v' x F' is a local vector bundle map and CPu is an isomorphism for all u E V. then cp.: V X T,' ( F ) --+ V' X T,'( F') is a local vector bundle map and ( CPu). = ( cp.) u is em isomorphism for aI/liE V. MoreOt'er. if cP is a local vector bundle isomorphism thell so is CP •• Proof. That CP. is an isomorphism on fibers follows from 5.1.6 (iii) and the last assertion follows from the former. By 5.2.1 we need only establish that (CPu). = (cp.)u is of class C". Now. CPu is a smooth function of u. and. by 5.2.3. CP: and cP'; I are smooth functions of u. The Cartesian product of smooth functions is easily seen to be smooth and (CPu). is a multilinear mapping on a Cartesian product of smooth functions (this is not linearity in cp). Hence from the product rule ( CPu). is smooth. •
This smoothness can be verified also for finite-dimensional bundles by using the standard bases in the tensor spaces as local bundle charts and proving that the components (cp.t)i, .. \ "-j, are COO functions. We have the following commutative diagram. which says that CP. preseroes fibers: '1'. V X T,'( F) - - V' x T,'( F')
wI
1w'
v'
V
5.2.5 Definition. Let w: E --+ B be a vector bundle with E" = w . I(b) the fiber over bEB. Define T,'(E)=U"eBT,'(E,,) and 'IT;: T,'(E)--+B by .,,;(e)'" b if e E T,'(E,,). Furthermore,for a subset A of B. we put 1','( E)IA'" U /t8 ... T,'(E,,).lf"": E'--+ B' is another vector bundle and (cp. CPo): E --+ E' is a veclor bundle mopping with cp" = cpl E" an isomorphism for all b E B. leI CP.: 1','( E) --+ 1','( E') be defined by cp.1 1','( E,,) = (cp,,) •. Now suppose that (EI V, cp) is an admissible local bundle chart of w, where V c B is an open set. Then the mapping cp.IIT,'( E)I V] is obviously a bijection onto a local bundle. and thus is a local bundle chart. Further, (cp.),,;" (cp,,). is a linear isomorphism. so this chart preserves the linear structure of each fiber, which in this case is given in advance. We shall call such a chart a nalural chart of 1','( E). .
5.2.8 Theorem. If w: E --+ B is a veclor bundle. then the set of all natural charts of w:: 1','( E) --+ B is a vector bundle atlas. Proof. Axiom (VB 1) ping natural charts.
j,
(Ihvious. For md "' •. I
I •. n '\
'IUppose we have two overlap-
. 'i",,·ity. let them have the same
TENSOR BUNDLES AND TENSOR FIELDS
285
domain. Then a = '" 0 fP - I is a local vector bundle isomorphism. and hy 5.1.6. "'.~(fP.)-1 = a •• a local vector bundle isomorphism by 5.2.4 • This atlas of natural charts. the natural atlas of "'.r. generates a vector bundle structure. and it is easily seen that the resulting vector bundle is Hausdorff. and all fibers are isomorphic Banadlahlc: spaces. Hereafter. -r,'( E) will denote all of thi.f structure. 5.2.7 ProposlUon. If f: E -+ E' is a vector bundle map that is an isomorphism on each fiber. then f.: -r,'( E) -+ -r,'( E') is also a vector bundle map that is an isomorphism on each fiber. Proof. Let (U. fP) be an admissible vector bundle chart of E. and let (V. ",). he one of E' so that I( U ) c V and f"o/- = '" 0 f 0 fP I is a local vector bundle mapping. Then using the natural atlas. we see that (/ .)"' •. ~. = (/"'~,).. •
5.2.8 Proposition. Suppose I: E -+ E' and g: E' ..... E" are vector bundle maps that are isomorphisms on each fi~er. Then .WI is 1{ f. and 0
(i) (g 0 f). = g. 0 f •. (ii) If i: E -+ E is the identity. then i.: -r,'( E) -+ T:( E) is the identity. (iii) If f: E ...... E' is a vector bundle isomorphism. then so i.f I and (/.)
1=
(/- I) •.
Proof. For (i) we ell8mine representatives of (1{ 0 f). and g. 0 f •. These representatives are the same in view of 5.1.6. Part (ii) is clear from the definition. and (iii) follows from (i) and (ii) by the same method as in 5.1.6 .
We now specialize to the important case where "': E vector bundle of a manifold.
-+
•
B is the tangent
5.2.9 Definition. Let M be a manifoid and 1".11: TM -+ M its tangent bundle. We call r.'(M) = -r,'( TM) the ~clo" b"ndle 0/ tensors contravariant 0'" r IIlIII covariant orde,. s; or simp~)' 0/ type (.~). A Iso Tt( M) is called the cotangent bundle 0/ M and is denoted by 1":': PM --+ M.
Since E can be embedded in E**, we may thereby embed TM as a closed submanifold in TOI( M). If M is finite dimensional we may identify TM = Tol( M). The zero section of r.'( M) may be identified with M. Recall that a section of Ii bundle assigns to each base point b a vector in the fiber over b. In the case of -r,'( M) these vectors are called tensors. Also. the addition and scalar multiplication of sections takes place within . each fiber. The COO sections of "': E -+ B were denoted r OO ( "'). or r OO ( E).
186
TENSORS
Jlecall that '5'( M) denotes the set of mappings from M into A that are of
class COO (the standard local manifold structure being used on A) together ~th its structure as a ring; namely. f + g. cf. fg for f. g E ~(M). c E A are given by (f + g)(x) =- f(x)+ g(x). (cf)(x) = c(f(x» and (fg)(x) = f(x)· g(x). Finally. recall that a vector field on M is an element of ':X (M) = r"'(TM).
5J.10 DefinlUon. A lellSor field of type (~) on a manifold M is a Coo section of T,'( M). We denote by '5.'( M) the set f"'(T,r( M» together with its (infinite-dimensional) real vector space structure. A coveclor field or dille,,",titIf OIfI-Io"" is an element of *( M) = 5 10 ( M ). IffE'5'(M) and IE5.'(M). let fl: M-+T,r(M): m ...... f(m)t(m). If obo XjE 'X(M). i-I ... .• s. and a i E *(M). j = I •... ,r, let
ex
ex
'(al •...• a'. XI ..... X.): M .... A; m ...... I(m}{ al(m), .... X.(m»). (falso t' E 5/( M). let I®I': M .... T,'++/(M); m ...... l(m)®I'(m). 5.2.11 ProposlUon. I(CXI .....
X,) E
With f. t. Xj' a i • and I' as in 5.2.10. fl E '5:( M).
~(M), and I®I' E ~1:.:~·.
proof. The differentiability is evident in each case from the product rule in l(lCal representation. •
Note that for the tangent bundle TM. a natural chart is obtained by taking TIJI. where IJI is an admissible chart of M. This in turn induces a chart (f,,), .. T,'IJI for T,rM. We shall call these the natural charts of T,'M. Now we turn to the expression of tensor fields in local coordinates. p-ecallthat a/ax; - (TIJI)-I(e i ), forlJl: U .... U'c An a chart on M, is a basis of~(U). The vector field a/ax; corresponds to the derivationf ...... ilf/ax'. Since dx i( a/ax}) = axil ax] = 6j, we see that dx j is the dual basis of l at every point of U, i.e., that dx i = 1JI*(e i ). where el ..... e n is the dual ))8Sis to el .... •e n • Let
a;ax
i, a/ax" , .•. , il/ax I .) E 6r(U) t i""i,jl"'j. .... I(dx i ,, ... ,dx ., J .
fben applying 5.16(iv) at every point yields
",him is the coordinate expression of an G) tensor field. To discuss the behavior of these components relative to a change of c;oordinates. assume that Xl: U ..... R. i-I ... . ,n is a different coordinate
TENSOR BUNDLES AND TENSOR FIELDS,
287
system. We can write a/ax/-a/a/ax i , since hoth are bases of 'X(U), Applying both sides to x" yields a/=axi/ax i i,e., (a/ax i ) (axi/ax/Xa/ax i ). Thus the dx i as dual bases change with the inverse of the Jacobian matrix ax i / axi; i.e., dx i = (ax i/ ax) ) dx'. Hence writing t in both coordinate systems and isolating equal terms gives the following change of coordinate formula for the components:
_""
ax '" , 'axl,
ax Ie, axh ax" a.t l,
(Ix it
I ,'" '/ ... / = - - ... - - . - - ... --t" iJ.~/.
'j,
i,"'i.
This formula is known as the len,wriality aiterion. A set of n'U functions li' .. ··/l, ...},/or each coordinate syslem on Ihe open sel U of M locally define an (:) tensor field iff a change of coordinates has the aforemenlioned effect on them. This statement is clear since at every point· it assures that li''''/l''''j,(u) are the components of an (~) tensor in T"U and conversely. The algebraic operations on tensors, such as contraction, inner products and traces, all carry over, fiberwise, to tensor fields. For example, if . 1m E TII(T",M) is the Kronecker delta, then I: M .... TII(M); m ..... 1m is obviously C"", and I E ~II(M) is called the Kronecker delta. Similarly, a tensor, field of type (~) or (~) is called symmetric, if it is symme,tric at every point. A basic example of a symmetric covariant tensor field is the following.
a
5.2.12 Definition. A weak pseudo-RierntlMitur IMtric on manifold Mis E T20(M) Ihal is symmetric and weakly nondegenerate, i.e., such that al each m EM, ,( m )( 0"" wm) = 0 for all "'m E Tm M implies q" - O. A' (strong) pseudo-R;'lIUIlfIIitur ""'ric is a 2-tensor that. in addition is (strongly) ~rtlteforall me M; i.e., the map om .... '(m)(o"'.·) isa (Banachable space) isomorphism of TmM onto T!M. A weak (re,fp. strong) pseudoRiemannian metric is called weak (resp. strong) Riema",," if in addition ,(mX om' "... ) > 0 for all q" E T.M, "m" O. .
alensor g
A strong Riemannian manifold is necessarily modeled on a Hilbertaable space; i.e., the model space has an equivalent norm arising from an inner product. Also, for fmite-dimensional manifolds weiak and strong metrics coincide: indeed T.M and T:,M have the same dimension and so a one-to-one map of T",M to T!M must be an isomorphism. It is possible to . have weak metrics on a Banach' or Hilbert manifold that are not strong. For example. the L2 inner product on M - CO aO, I),R) is a weak metric that is not .trong. For a similar Hilbert space example, see Exercise S.2e. Any Hilbert space is a Riemannian manifold with constant metric equal to the inner product. A symmetric bilinear (weakly) nondegenerate
286
TENSORS
Recall that iff( M) denotes the set of mappings from Minto R that are of class Coo (the standard local manifold structure being used on R) together with its structure as a ring; namely, I + g. cl. Ig for I, g E ~(M), c E Rare given by (f + g)(x) ... I(x)+ g(x). (cf)(x) = c(f(x» and (fg)(x) = I(x)' g(x). Finally, recall that a vector lield on M is an element of ~X(M)'" foo(TM).
5.2.10 DefInition. A tftUor lield 01 type (~) on a manilold M is a Coo section 01 T,'( M). We denote by ~T.'( M) the set f""( T,'( M» together with its (infinite-dimensional) real vector space structure. A covector lield or di/fefWItitll OM-form ;s an element 01 <X *( M) = '5 10 ( M ). II IE iff(M) and t E '5.'(M), let It: M -- T:(M): m ...... I(m)t(m). II also Xi E <X(M), i-I, .. . ,s, and a i E <X *(M). j = I •... . r. let
I( a l , ... ,a', XI , .. . ,X.): M II also t' E '5/( M), let tet': M
5.2.11
Propoaltlon.
I( a l •.. .• X,)
E
-- R; m ...... t(m)( a l ( m ) •.. .• X.(m»). -+
T,'++/(M): m ...... t(m)el'(m).
With I. t. Xi' a i • and t' as in 5.2.10. It E ~j/( M).
~(M). and tet' E~. ::~'.
The dirrerentiability is evident in each case from the product rule in local representation. •
Proof.
Note that for the tangent hundle TM. a natural chart is obtained by taking TfJ!. where fJ! is an admissible chart of M. This in turn induces a chart (TfJ!)." T,'fJ! for T,'M. We shall call these the natural charts of T,'M. Now we tum to the expression of tensor fields in local coordinates. Recall that a/ax i - (TfJ!)-I(ei ). for fJ!: U -oU'c R W a chart on M. is a basis of <X(U). The vector field a/ax i corresponds to the derivation I""" al/ax/. Since t/x i ( aI axi) = axil axi = 8 ii • we see that dxi is the dual basis of aI ax; at every point of U. i.e.• that t/x i = fJ!*( e i ). where e I .... • e lt is the dual basis to el .... ,ew' Let
Then applying 5.I6(iv) at every point yields
which is the coordinate expression of an (;) tensor field. To discuss the behavior of these components relative to a change of coordinates, assume that Xi: U -0 R. i ... I •...• n is a different coordinate
TENSOR BUNDLES AND TENSOR FIELDS,
287
system. We can write alax ' - a/alax j , since hoth are bases of ~.(U). Applying both sides to Xk yields a/ = ax ilax; i.e., (alax l ) ... (axilaxi)(alax i ). Thus the dx; as dual bases change with the inverse of the Jacobian matrix ax j I ax;; i.e., dx i = (axil ax ') dx '. Hence writing t in both coordinate systems and isolating equal terms gives the following change of coordinate formula for the components:
"11
. 'I.
This formula is known as the tensoriality criterilln. A set of n'+J junctions ti""ij, ...,.!or each coordinate system on the open set U of M locally define an (:) tensor field iff a change of coordinates has the aforementioned effect on them. This statement is clear since at every point it assures that ti""\"'j,(u) are the components of an (;) tensor in T,p and conversely. The algebraic operations on tensors, such as contraction, inner products and traces, all carry over, fiberwise, to tensor fields. For example, if . 1m E TII(T",M) is the Kronecker delta, then I: M - TII(M); m ...... I", is obviously Coo, and I E 5 11( M) is called the Kronecker delta. Similarly, a tensor, field of type (~) or (~) is called symmetric. if it is symme.tric at every point. A basic example of a symmetric covariant tensor field is thc= following.
a
5.2.12 Definition. A weak psellllo-RiI"""",iIut metric on manifold M is tensor g E T20(M) that is symmetric and weakly nondegenerate, i.e., such that at each m EM, g( m )( """ HI",) = 0 for all HI", E T", M implies "", = o. A' (strong) pseudo- Rientlllllfilut metric is a 2-tensor that. in addition is (strongly) IfDIfIIqeMfYlte for all m EM; i. e .• the map "m ..... I( m )( f'",. ' ) is a ( Banachable space) isomorphism of T",M onto T:,M. A weak (resp. strong) pS('IJdoRiemannian metric is called weak (resp. strong) Riemannilut if in addition I( m)( "",. "",) > 0 for all ,,'" E T", M. "", .. o.
ci
A strong Riemannian manifold is necessarily modeled on a Hilbertizable space; i.e., the model space has an equivalent norm arising from an inner product. Also. for finite-dimensional manifolds weak and strong metrics coincide: indeed T",M and T:,M have the same dimension and so a one-to-one map of T",M to T:,M must be an isomorphism. It is possible to , have weak metrics on a Banach' or Hilbert manifold that are not strong. For example. the L 2 inner product on M = CO «(0, I J, R) is a weak metric that is not ~trong. For a similar Hilbert space example, see Exercise S.2C. Any Hilbert space is a Riemannian manifold with constant metric equal to the inner product. A symmetric bilinear (weakly) nondegenerate
288
TENSORS
two-form on any Banach space provides an example of a (weak) pseudoRiemannian constant metric. The most famous pseudo-Riemannian manifold that occurs in the theory of special relativity is R4 with the MinkiJwski pseudo -Riemannian metric
.
g(x)( v. w) =
Vlw l
+ V 2 W 2 + V 3 w3 -
V4 W4•
where x. v. K' E R4. As in the algebraic context of Section 5.1. pseudo-Riemannian metrics (and for that matter any strongly nondegenerate bilinear tensor) can be used to define associated tensors. Thus the maps I. b become vector bundle isomorphisms over the identity~: TM ... T* M.': T* M ... TM • • - b I. where v!. = g( m)( vm .·). In particular. they induce isomorphisms of the spaces of sections b: ~"\ (M) ... ~~ *( M). #: ~l: *( M) -+ ~X (M). In finite dimensions this becomes the operation of raising and lowering indices. Thus formulas like the ones in Example 5.1.4E should be read pointwise in this context. There is. however. a particular index raising operation that requires special attention. 5.2.13 DetlnHlon. LeI M he a pseudo-Riemannian n-manifold with metric g. For f E ~(M). grad f = (df)# E ~l: (M) is called the grrulimt off. To find the expression of grad f in local coordinates, note that if gij
then for
a
X_XI_. ax'
we have
Xb( Y) - g( X. Y) - X'Yie(
a~i' a~j ).
g(
=
and
a
Y=Y'-. ax'
a~;' a~j ) .. Xiyjg,}.
i.e.
If II E ~X *( M j has the coordinate expression II = a; dxi. we have CJ!. = a;gi}a I ax) where gij is the inverse of the matrix gij' Thus for 11:= df. the
local expression of the gradient is gradf = gijJ.L
~
ax} iJx'
If M
=
or (gradfr
=
gi} af. ax}
R n with standard Euclidean metric g" = S", this formula becomes g.rad f =
af ~:
il:t' il:t'
i.e..
g.rad f =
(~ . . . .. : ~n ). il.\
(l.'
the familiar expression of the gradient from vector calculus.
TENSOR SUNDLES AND TENSOR FIELDS
289
Now we turn to the effect of mappings and diffeomorphisms on tensors. 5.2.14 DefinlUon. If fJi: M ~ N i.f a diffeomorphi.fm and t E T,'( M), let fJi.t - (TfJi). ° to fJi ',the push-forward of t by fJi. If t E T,'( N). the puN-back of t by fJi is given by fJi*' = (fJi - ').t. 5.2.15 PropoalUon. If fJi: M ~ N is a diffeomorphism. and t E '?i.'( M). then (i) fJi*t E '?i,'(N); (ii) fJi.: '?i,'(M) ~ '?i,'(N) is a linear isomorphism. (iii) (fJi. y,). - fJi*o y,.; and (iv) fJi.(t®t') - fJi*'®fJi*t'. where t E ~'i:(M) and t'E~/(M).
Proof. (i) The differentiability is evident from the composite mapping theorem. together with 5.2.4. The other three statements are proved fiberwise. where they are consequences of 5.1.6. • . As in the algebraic context. the pull-back of covariant tensors is defined even for maps that are not diffeomorphisms. Globalizing 5.1.8 we . get the following. If fJi: M ~ Nand t E ~'i.o(N). then fJi*l, the puN-back
5.2.16 Definition. of I by fJi is defined by
for
mE
M. v" .... v,
E
TmM.
The next proposition is similar the proof of 5.1.9.
!(l
5.2.15 and is pfllved by globalizing .
5.2.17 ProposlUon. If fJi: M ~ N is Coo and I E '?is o( N), then (i) fJi*1 E '?i,o(M), (ii) fJi*: '?i,o(N) ~ '?i,o(M) is a linear map; (iii) (y, ofJi)* = fJi*o y,* for y,: N ~ P; (iv) if cp is a diffeomorphism ,hen cp* is an isomorphism with inverse CP •• and (v) if 'I E 5s~( N). '2 E '!j,~(N). fJi*( ',®t 2 ) = (fJi*', )®( fJi*/ 2 ).
For finite-dimensional manifolds the coordinate expressions of the pull-back and push-forward can be read directly from 5.1.7 and 5.1.10. taking .into account that Tcp is given locally by the J&cobian matrix. This yields the following. . 5.2.18 PropoSition. LeI M and N he finite-dimensional mallifold~. cp: M ~ N a C' map and denote by .~.I = cpl(X I..... x P ) the local expression of cp relative to charts where p = dime M) and j = I .... , n = dime N).
290
TENSORS
(i)
If 'E~/(M) and cp is a diffeomorphism. the coordinates of the cp • tare
push~forward
If' E ':i/(N) and cp is a diffeomorphism, the coordinates of the pull-back cp.' are
(ii) If' E '.'i, o( N) and 11': M ~ N is arhitrarr. the ('()(}rdinates of the pull-hack cp.' are
Notice the similarity between the formulas for coordinate change and pun-back. The situation is similar to the passive and active interpretation of similarity transformations PAP-I in linear algebra. Of course it is important not to confuse the two. 5.2.19 Exampl... A. Let cp: R2--+R2 be defined by cp(x.y)=(x+ 2y. y) and lett = 3x (a/ax)®dy +(a/ay)®dy E '.'i11(R 2 ). Then the matrix of cp. on vector fields is (
;~ ) = [~ ~ ]
and on forms is
(;;:) =
[~
t]- [6 -2] I . I
=
In other words. cp.(
a: ) :x. =
cp.(dx) = dx - 2 dy. and
CP.(~v)=dy.
Noting that cp -I(X. y) = (x - 2y. y). we gel
CP.' = 3( x - 2y )cp.( :x )®cp.( ~r) + cp.( ;~ ) ®q>.( eld =
a
(aax iil) II ®dl" + - a ®d)'. ®dl" = (3x -6r +2)ly' . ax' ay
3(.t-2r)- ®dy+ 2- + .
ax
TENSOR BUNDLES AND TENSOR FIELDS
291
B. With the same mapping and tensor. we compute ",*1. Since
",*(..!.) =..!., ax ax ",*(dx) =dx +2dy, and ",*(dy)
=
dy.
we have
a ®dry + (- a iJ = 3(x, + 2.1')2 - + - a) ®dry = (3x +61' -2)-'®dv + - a ®d~'. ax ax ay . iJx" a.I' , C. Let "': R J ..... R 2 • ",(x.y.z)=(2x+z,xyz) and 1:=(u+2v)du® dU+(U)2du®dv E T2o(R 2 ). Since ",*(du) = 2dx + dz and ",* (00) = yzdx + xzdy + xydz, we have ",*, = (2x
+ z +2xyz)(2 dx + dz)®(l dx + dz)
+(2x + z)2(2dx =
+ dZ)®(yzdx + xzdy + xydz)
2[4x +2z +4xyz +(2x + Z)2 yz] dx®dx +2(2x + z ('adx®dy
+2[2x + z +2xyz +(2x + z )2xy] dx®dz + xz(2x + z)2 dz®dy + [2x
+ [4x + 2z +4xyz + (2x + z )2] dx®d:(
+ z +2xyz + xy(2x + z )2] dz®dz.
D. If "': M - N represenis the deformation of an elastic body and. is a Riemannian metric on N, tben C - fJI*1l is called the Cauchy - Green tensor: in coordinates
Thus, C
~asures
how fJI deforms lengths and angles.
•
Finally, we describe an alternative approach to tensor fields. Suppose '?t(M) is defined as before, and ~(M) either our way or some equivalent way. With the "scalar multiplica\ion"
IX
becomes an ~(M)-module. That is, ~ (M) is essentially a vector space over ~(M), but the "scalars" ~(M) form only a commutative ring with identity,
292
TENSORS
rather than a field. as
Ilf may not exist. even if f '* O. We may thus define
L"f(M'(~X (M).'S( M» = P( M)
the
~1 ( M)
linear mappings. and similarly '1:;( M) == L;~~)( P( M ) •...• ~(M); 5( M»
the -:t( M) multilinear mappings. From 5.2.10. we have a natural mapping '!r,'( M) ~ '1:;( M) which is lj( M) linear.
5.2.20 ProposlUon. Let M be a finite-dimensional manifold or be modeled on a Banach space with norm C'JO aw~y from the origin. Then '?T:( M) is isomorphic to the lj(M) multilinear maps from .P( M) X ... X ~x ( M) into lj(M). regarded as lj(M) modules or as real vector spaces. In particular. ~ *( M) is isomorphic to P( M). Proof. Consider the map '?T,'( M) ~ L,,( M,(.(*( M), ...• ~X (M); \1 ( M» given by l(al •...• a'.XI •...• X.)(m)=/(m)(al(m) •...• X.(m». This map is clearly -:t( M) linear. To show it is an isomorphism. given such a multilinear map I. define I by I( m)( a l ( m)•...• XJ ( m» = I( a l •...• X,)( m). To show this is well defined we must show that. for each 11, E Tm( M). there is an X E '.\ ( M) such that X( m) ... q.,. and similarly for dual vectors. Let (V. cp) be a chart at m and let T",cp(q.,) == (cp(m), G,). Define Y E ~X.(V') by Y(u) == (u'. q;) on a neighborhood VI of cp(m). Extend Y to V' so Y is zero outside V2 • where cl(VI)c V2 • cI(V2 )c V'. by means of a bump function. Define X by X" - Y on V. and X - 0 outside V. Then X( m) "" q.,. The construction is similar for dual vectors. Also. I(m) so defined is Coo; indeed. using the chart CPt the local representative of , is Coo by Box 3.3A since I induces a Coo map M X T,'( M) ~ R (by the composite function theorem). which is (r + s)linear at every m E M. If M is finite dimensional this last step of the proof can be considerably simplified as follows. In the chart cp with coordinates (x' •.... x").,=ti,·\ .. ,.(illilx")®··. ®illilx"®dx"®'" ®dx" and all' components of , are C'" by hypothesis. •
The preceding proposition can be clearly generalized to the C A situation. One can get around the use of a smooth norm on the model space if one assumes that the multilinear maps are localizable. that is. are defined on ~*(V)X ... x 'X.(V) with values in 'ff(V) for any open set U in a way compatible with restriction to V. We shall take this point of view in the next section. The direct sum '!reM) of the '?T:< M). including '?Too( M) = lj( M). is a real vector space with ® -product. called I" . ' , . " . , J!f<,ebra of M. and if cp: M ~ N is a diffeomorphism. CP.: '?T (M)" i'" .llgebra isomorphism.
THE LIE DERIVA TlVE: ALGEBRAIC APPROACH
293
The construction of ~1 ( M) and the properties discussed in this section can be generalized to vector bundle valued (~) tensors (resp. tensor fields), i.e., elements (resp. sections) of L(T*Me ... eT*MeTMe ... eTM, E), the vector bundle of vector bundle maps from T*Me ... eTM (with r factors of T* M and s factors of TM) to the vector bundle E, which cover the identity map of the base M. Exercl...
S.2A
Lei cp: R 2 \{O} -oR 2 \{O} be defined by cp(x. y) = OX2 + tv 2 • - xy)' and let t = xiJ / iJx@dx@dy + yiJ / fly@dl'@dy. Show that cp is a diffeomorphism and compute CP.t. cp*t. Endow R 2 with the standard and Riemannian metric. Compute the associated tensors of t. cp., as well as their (1. I) and (1.2) contractions. What is the trace of the interior product of t with iJ / + xiJ / fir? Let cp: R 2 -oR\ cp(x.y)=(y.x.y+x 2 ) he thc deformation of an elastic body. Compute the Cauchy-Green tensor and its trace. Let H be the set of real sequences {a n },.-I.2. such that IIll,,1I 2 = Ln2a~ <00. Show that H is a Hilhert spal·C. Show that g(a.h)= La~bn is a weak Riemannian metric on H that i~ nnt a strong metric.
CP.'.
ax
S.28 S.2C
5.3 THE LIE DERIVATIVE: ALGEBRAIC APPROACH This section extends the Lie derivative Lx from vector fields and functions to the full tensor algebra. We shall do so in two ways. This section does this algebraically and in the next section. it is done in terms of the flow of X. The two approaches will be shown to be equivalent. . We shall demand certain properties of Lx such as: if t is a tensor field of type (~). so is Lx', and Lx should be a derivation for tensor products and contractions. . First of all. how should he defined on covector fields? If Y is a vector field and Cl is a covector f' i ! 1hen the contraction Cl' Y is a function, so Lx (Cl' Y) and Lx Yare alread 'HIed. (See Section 4.2). Jiowever. if we would like to have the derivation pwperty for contractions. namely
"x
this forces us to define LxCl by
(LxCl)'Y= Lx(Cl·Y)-Cl·(LxY) for all vector fields Y. Since this defines an ~'f(M)-linear map. LxCl is a well-defined covector field. The extension to general tensors now proceeds inductively in the same spirit.
294
TENSORS
5.3.1 DefInition. A diflnwrlilll operator on the full tensor algebra '5 ( M) of a manifold M is a collection (O;( U)} of maps of '5.'( U) into itself for each r and s ~ 0 and each open set U C M. any of which we denote merely 0 ( the r, s and U are to be inferred from the context). such that (DOl)
0 is a lemor Iknvat;on i.e. 0 is R-linear and if
IE '5:( M). Cl I•...• Cl, E ~X *( M) and Xl •...• X .• E ~X (M). then
O(I(ClI, ... ,Cl,. XI •...• X.))
=
(OI)(ClI •...• Cl,. XI.···.X.)
,
+
$
E I(ClI,···.OCl/ •...• Cl"
XI.···,X.)
+
E I(ClI •...• Cl,. XI.···,OX" •...• X.,).
k-I
j-I
We sometimes refer to this by saying that 0 COIIUnIIln with COIIlrtlC-
tHwu. (D02)
0 is local. or is IIIltIIral with rnpect 10 rntrict;OIU. That is. for U eVe M open sets. and t E ~'(V) (Ot)IU= O(tIU) E c:r:(U)
or the following diagram commutes:
'5/(V) -
IV
01
~'(U)
10 , ~T:(U)
5/( V) IV
Note that we do not demand that 0 be natural with respect to push-forward by difreomorphisms. The reason is that it is not needed for the following unique extension theorem, and indeed. the latter can be used to extend the covariant derivative. which is not natural with respect to diffeomorphisms. .
5.3.2 Theorem. Suppose for each open set U c M we have maps Eu: Iff( U) ..... Iff( U) and Fu: <.X. ( U ) ..... X ( U). which are (R linear) tensor derivations and natural with respect to restrictions. That is
f. g E '!leU);
THE LIE DERivA TIVE: ALGEBRAIC APPROACH
295
For/ E 6J(M). Eu(/I V) = (EM!) I V; Fu(/®X)- (Eu!)®X + /®FuX; (iv) for X E 'X(M).Fu(XI V) = (FMX)I V.
(ti) (iii)
Then there is a unique differential operator D on ~ ( M) that coincides· with EIJ on ~T( V) and with Fu on ~l: (V). Proof. Since D must be a tensor derivation. define D .on 'X *( V) by (Da)·X=D(a·X)-a·(DX)=Eu(a·X)-a·FuX for all XE~t(V). Thi~ shows that D exists and is unique on 'X *( V) (by the Hahn- Banach theorem). The rest of the theorem follows by requesting (DOl). namely. define Du on ~'(V) by
(Dut)(a, ..... a,. X, ..... X.) = EIJ(t(a, ..... a,. X, ..... X,)
,
- L
,-, t(a, ..... DaJ ..... a,.X, ..... X~) ..
- L
t(a, ..... a,. X, ..... FuX" .... ,X.).
k-I
If V is any open subset of V, by (ii) and (iv) it follows that Dv(tl V)(Dut)W. This enables us to define D on 6J(M) hy (Dt)(m) = (Dut)(m), . where V is any open subset of M containing m. Since D/. is unique. so is D. and so (D02) is satisfied by the construction of D. •
5.3.3 Corollary.
We have D(t,®t 2 ) = Dt,®t 2 + t,®Dt 2
(i) and
D8 = O. where 8 is Kronecker's delta.
(ii)
Proof. (i) is a direct application of (DOl). For (ii) let aE 'X*(U)'and X E 'X(V) where V is an arbitrary chart domain. Then
(D8)(a. X) == D(8(a.X»-8(Da, X)-8(a, DX) =
D(a·X)-Da·X-a·DX-O.
Again the Hahn-Banach theorem assures that Da = 0 on V, and thus by (D02), Da ==
o. •
296
TENSORS
Taking Ev and Fv to be LXI v we see that the hypotheses of theorem 5.3.2 are satisfied. Hence we can define a differential operator as follows.
5.3.4 Definition. If X E ~'X ( M). we let Lx be the unique differential operator on '!i( M). called the Lie tkrivative wit" IYS~ct to X. such that Lx coincides with Lx as given in 4.2.6 and 4.2.20. 5.3.5 Propoaltlon. Let cp: M -. N be a diffeomorphism and X a vector field on M. Then Lx is ""tUM Moit" res~ct to pusIt-/OrwtlN by cp; that is, L •• xcp.t .. cp.Lxt for t E '!i,'( M). or the following diagram commutes; ~T:( M)
Lx
'1'.
~'i:( N)
!
5:( M)
1Lrp.x
"'.
.
~:(
N)
Proof. For an open set V eM define D: ~1:(U) -. '~:(V) by Dt = cp·L",.xl v( CP.t), where we use the same symbol cp for cpl V. By naturality on '5"(V) and <.X(V). D coincides with LXIV on '5"(V) and 'X.(V). Next. we show that D is a differential operator. For (001). we use the fact that
which follows from the definitions. Then for X. XI'" .• X, E a l ..... a, E '.\ *( V). D( t( a l ..... a,. Xl" .. • X.))
=
':X (V) and
°
cp*Lrp.x (cp.( t( 1" ... a,. XI'" .. X.»)
, +
L
(cp.t)(cp.al.· ... Lrp.xcp.o, ..... cp.a,.·j\I ... ·.cp.X.)
i-I J
+
E (cp.t)(cp.al ..... cp.a,.cp.XI ..... Lrp.xcp.XA ..... cp.X.)]. k -I
THE LIE DERIVATIVE: ALGEBRAIC APPROACH
297
By (DOl) for Lx. this becomes
, (01)(°1 •...• 0,. XI' ...• X.)+ ~ 1(°1•...• 0°, .... ,0" Xl ... ·'~.• ) ) -I .f
+ ~ t(al, ...• a"
XI'· .. ·OXk .. · .. X.)
A -I
since cp. - (cp I). by 5.2.14. For (D02), let
IE
'.'i, r( M) and write
[by (1)02) for Lx] =O(tIU)
The result now follows by 5.3.2. • The reader can check, by using the same reasoning, that a differential .operator that is natural with respect to diffeomorphisms on functions and vector fields is natural on all tensors. Lt;t us now compute the local formula for Lxt where I is a ,tensor field of type (~). Let cp: U c M ~ VeE he a local chart and let X and 1 he the principal parts of the local representatives, CP.X and cp.1 respectively. Thus X: V ..... E and;: V ..... r,'( t:). Recall from Section 4.2 that the local formulas . for the Lie derivatives of functions and vector fields are: (Lxff(x)
=
Dj(x)·X(x)
(I) ,
and (LxY nx) = DY(x )·X(x)- DX( \). Y( r)
(2)
In finite dimensions these become Lxf=
x,!..L ax'
and (2')
298
TENSORS
Let us first find the local expression for Lxa where a is a one-form. By 5.3.5. the local reprc:.sentative of Lxa is CI>.(Lxa) = L.".xCl>*a, which we write as Lj&' where X and &. are the principal parts of the local representatives. so X: V ..... E and &.: V ..... E*. Let vEE be fixed and regarded as a constant vector field. Then as Lx is a tensor derivation, Li(&"V)'" (Li&)'''+ &·(Liv).
By (I) and (2) this becomes D(&· v)·X- (Lia)'v - &. (DX· v).
Thus (Li&')'V= (D&"X)·v+&.·(DX·v).
In the expression (D&"X)'v, D&.·X means the derivative of &. in the direction X; the resulting element of E* is then applied to v. Thus we can write Li&'= D&.·X+ &.·DX.
(3)
In finite dimensions, the corresponding coordinate expression is ( L a) v' x,
=
aa;. ax'
ax'. 'ax;'
-XlV' + a.--v'·
i.e.,
.aa;
ax}
( L a).=X'-+ax, axi , ax;
(3')
Now let t be of type (~), so i: V ..... L(E*, ... ,E*, E, ... ,E; R). Let a l , ... ,a' be (constant) elements of E* and VI" ..• v, (constant) elements of E. Then again by the derivation property.
Lx (1( a l •...• a'. VI'"
.• V,)] = (Lil)' (a l •...• a'. VI'" .• V.)
, +
~-( a I •...• L ; a."I.···.v, ' ) LJt ia •...• i-I J
+
E ~al •...• a'.vl.···.Li"i.· ..• V,). i-I
Now using the local formula (I )-(3) for the Lie derivatives of functions.
299 '
THE LIE DERIVATIVE: ALGEBRAIC APPROACH
vector fields. and one-forms. we get
(DLf)· (ex l•...• ex'. Vi •••• ,Vs ) = (Lii)( ex l •...• ex'. t'I ..... 1', )
,
+ 1: l(ex l..... ex;·1)X ..... ex'.v1..... 0.) i-I
.r
+
E ~exl •...• ex'.VI ..... -
DX·o; ..... O,).
i-I
Therefore.
,
- E l(ex l..... exi ·DX..... ex'.v 1..... b,) i-I
s
+
E i(exl ..... ex'.I'I ....• DX.O, ..... V,). i-I
In components. this reads
(L
x' );'''';'
i,"'i, =
X.
a t i,"·i';, ax.
"'j,
_ axi,I t 1i''''i,i "'}' _
ax
"
ax t , "m}" ".; + --. m
aX"
".,"
"
(
. d'Ices) aII upper In
. d'Ices) + ( aIIIower In
(4)
We deduced the component formulas for Lx' in the case of a finit~ dimensional manifold as corollaries of the general Banach manifold formulas. Because of their importance. we shall deduce them &gain in a different manner. without appealing to 5.3.5. Let
. . . -~ a ... '=1"""' ,
II'''}.
axi,
a·
"t:r',( U )
~_~dxJI ••• ~dxJ'E~1
axi,
where U is a chart domain on M. If X=
x"a/ax'.
•
.
the tensor derivation
300
TENSORS
property can be used to compute Lx" For this we recall that
and that
by the general formula for the bracket components. The formula for
Lx(dx') is found in the following way. The relation
8\ = dx*(a/ax i )
implies by (001) that
- (L.(dx'») (- a) + dx' (ax' - -. - a ). ax' ax' ax' Thus
(LX(dx"»)(~) -dx'( ax' ~) _ ax'.
ax'
ax' ax'
ax'
io Lx(dx')-(ax'/ax')dx'.
Now ODe simply applies (DOl) and collecta terms to pt the same local formula for Lx' found in (4). Note espcciaIly that
LI
hi
(.L)_I ax J
and
L..!...(dxJ)-I,
foraD i,j.
la'
U.I~
A. Let
,- x :, edxedy + , Compute
L.,.
:, edyedy and X -
:x + x :"
THE LIE DERIVATIVE: ALGEBRAIC APPROACH
301
Solution. Method 1. Note that
LJtI-LiI •• _a_I-I•. ~I+L.at
(i)
. III'
lit
i/o
if ..•
Li-I=L~{X~®dX®dY+ )'~®dY®d)'} a" i/x ay i1y
(ii)
= L:!.(X~ ax fJ y
®dX®dv)+L~ ()'~ ®dY®dr) . I/x . ay "
=
11_ ®dx®ti)' +0 ,Ir '
Now note Lxi-
1/1'
~ =O.L.i- ~ = - L~(X~) = fJy '.I' ax 'ax ay
L,,~dx ar
= O. and
-{I' ~a)' +X.O}
=
-~. ay
L,,!_dy = dx. I/r
Thus (iii)
L i! , = L '.1.1'
=
.iI '.1..
{x~ ®dx®dl' + y~ ®dy®dv} fJy . ay "
(0+0+0+ x :v ®dX®dX) a ®dy®dy +0+ y-a a ®dx®d)' + )'-~ a ) + ( x-a Y y - . oy ®dy®dx .
Thus. substituting (ii) and (iii) into (i). we find
a a a, + x-a y ®dy®dy + y-a )' ®dx®dy + .v-a y ®dy®dx
a
a
= (y + I) fJy ®dx®dy + x ay ®dx®d.\
a
a
+ x ay ®dy®~v + y ay ®dy®dx. Method 2. Osing component notation. I is a tensor of type (~) 'whose nonzero components are t 2 '2 = x and t 222 = y. The components of X are
302
TENSORS
x' -I and X 2 ""' x. Thus, by the component formula (4), ; ,ax; ; ax m (L xl ) ;jk'" X" - Ia e t jle - t jle --, + t ",Ie - - . ax ax ax'
;
+ t jp
ax p Ie' ax
The nonzero components are
(LXI)\, = 0-0+0+ x = x; (LXI)222'" x -0+0+0 = x; (LXI)\, = 0-0+0+ y = y. and hence
a
a
LxI = (y + I) ay ®dx®dy + x iJy ®dx®dx iJ
+ x ay
®dy®dy
+y
iJ ay ®dy®dx.
The two methods thus give the same answer. It is useful to understand both methods since they both occur in the literature, and depending on' the circumstances, one may be easier than the other. B. In Riemannian geometry vector fields X satisfying Lx g = 0 are called Killing vector fields; their geometric significance will become clear in the next section. For now, let us compute the system of equations that the components of a Killing vector field must satisfy by the method described in the previous example. If X = xla I ax;, and g == g;jdx; ®dxi, then
Note that Lxg is still a symmetric (~) tensor, as it must be. Hence X is a Killing vector field iff its components satisfy in every chart the following
THE LIE DEfflVA TlVE: ALGEBRAIC APPROACH
303
system of n partial differential equations, called Killing '.~ equations
ag;j ax" ax" x"--" + g"j--' + g;,,--. = o. ax ax' ax' C. In the theory of elasticity, if. represents the displacement vector field, the expression L.g is called the slrain tensor. As we shall see in the next section, this is related to the Cauchy-Green tensor C = .,,*g by linearization of the deformation .". p. Let us show that Lx does not necessarily commute with the formation of. associated tensors; e.g.~ that (Lx'),," (/'xl),J' where I'" Ija/ax'edx' e TII(M) and 1-I;jdx'edx J is the associated tensor with components IiI = g;"I·r We have from (4)
.
(L I)' x j
al;
ax;
ax·
x·_J ax" - t·j-ax"- + I'Ie -ax-j •
and so
But also from (4)
ax' ax" x"-al;j I t + I'j-' + ,;,,--. ax ax' ax' • a (gill ,j )+ g,le' " ax' + gill.-, ax A ..... x -. ax ax' ax'
(Lxf)/j'"
j-'
il , "a/~ " ax' +g"I.-, ax".. =x"ag -"I ax j + X w"-ax" +g,,,1 jax' ax'
Thus. in order to have equality it is necessary and sufficient that il , " ax' X "ag -"lj+g,,,1 j-I
ax
"ax' =0 ax +gjfl j--" ax
for all pairs of indices (i, j), which is a nontrivial system of n 2 linear partial differential equations for g;)" If X is a Killing vector field. then
ax' ax' ,agu, g,,, axl + gIl ax" .. - x ax .
304
TENSORS
which substituted in the preceding equation, gives zero. The converse statement is proved along the same lines. In other words, a necessary and sufficient condition that Lx commute with the formation of associated tensors is that X be a Killing vector field for the pseudo-Riemannian metric g. •
As usual, the development of Lx extends from tensor fields to F-valued tensor fields. Exercl...
S.3A Let 1- xya/axedx + yiJ/ayedx + a/axedy e TII(R ~). If cp: R2 -. R Z is defined by cp(x, y) = (x + y. xy). show that cp is a diffeomorphism and compute tr(I), cp.'. cp.l. LxI. Lxcp·'. L.,..xl and tr(L •• xcp.l) for X ... ya lay + x 2 alay. S.38 Verify explicitly that Lx'*(Lxt)' where' denotes the associated tensor with both indices lowered. for X and I in Exercise 5.3A. 5.3C Compute the coordinate expressions for the Killing equations in R 3 in standard. cylindrical, and spherical coordinates. What are the Killing vector fields in R "? 5.4 THE LIE DERIVATIVE: DYNAMIC APPROACH
.
We now turn to the dynamic interpretation of the Lie derivative. In Section 4.2 it was shown that Lx acting on an element of ~'t( M) or '.'x (M). respectively, is nothing but the time derivative at zero of that element of ~if(M) or ~(M) dragged along by the flow of X. The same situation holds for general' tensor fields. Given I e '!f, r ( M) and X E ~X ( M). we can find a curve at I( m) in the fiber over III hy using the flow of X. The derivativ~ of this curve is the Lie derivative. (In spirit. the flow plays the same role for Lie derivation that parallel translation plays in covariant differentiation.) More precisely, for m E M and a vector field X on M let (U. a, F) be a now box at m. For each A e III - ]- a, a[ we can form the diffeomorphism F" ... FI U x{A}: U -. U" == F,,(U). For I E ~,'( M), define the map T(m): I" -. T:T..,(M): A... Fr('1 U,,)(m). Then the A-derivative at zero of T(m) should equal Lx'(m)(see Fig 5.4.1). Note: We use A for the time variable to avoid notational confusion with the tensor'.
5.4.1 Theorem. Let XE~·Xk(M).tE'!T.'(M) be of class C~ and F" be the flow of X. Then
THE LIE DERIVA TlVE: DYNAMIC APPROACH
305
integral curve of X M
FIgu ... 5.4.1
Proof.
It suffices to show that
d~ I 1\
~
_ ()
Ftt = Lx t .
Indeed. if this is proved then
d~ F:t = 1\
I
dd F:+,.t 1-',.-0
=
I
dd F,.*Ftt 1-',.-0
For smoothness of T(m), we have T(m)(~) = (n~ I(m».t( FA(m». Hence we need verify only that (TF'; I(m». is a smollth function of ~. Consider F: U XI" - M. which is smooth in (u. ~). Then TF: TV x ( I x R) - M is .also smooth. For smoothness of TF". note Tf~ - TFI TF X{~}X {O}. Then from 5.2.4, we see that (TF'; 1)* is smooth in ~. Since Fo is the identity, it is clear that T(m) is a curve at t(m). Now define 9x :
(cp*t)( ~*(lI, ...• CP*(I'. cP* Xl •...• cp* X,) = cp*( t{ (II •.... a'. Xl •... • X,)) for cP a diffeomorphism. Hence 9x is a differential operator. It remains to
306
TENSORS
show that ex coincides with Lx on <5(M) and ~ ( M). For f X E 'X(M). we have
~'J ( M). and
E
and
by 4.2.10 and 4.2.19. respectively. Thus by 5.3.2 and 5.3.4. 0xl
=
1... 1 for all
t E <:i(M).
•
This theorem can also be verified in finite dimensions by a straightforward coordinate computation. See Exercise 5.4A. The identity in this theorem relating flows and Lie derivatives is often taken as the definition of the Lie derivative (see Exercise 5.4C). 5.4.2 Corollary. If t ThaI is. t = Ftt.
E
<:i (M). L xt = 0 iff t is constant along the flow of
x.
As an application of 5.4.1. let us generalize the naturality of Lx with respect to diffeomorphisms. As remarked in section 5.2. the pull-back of covariant tensor fields makes sense even when the mapping is not a diffeomorphism. It is thus natural to ask whether there is some analogue of 5.3.5 for pull-backs with no invertibility assumption on the mapping ",. Of course. the best one can hope for, since vector fields can be operated upon only by diffeomorphisms. is to replace the pair X. '" _ X by a pair of !p-related vector fields. 5.4.3
Propo.ltlon.
Let "':
M ..... N
be
Coo. X
E
'X (M). Y
E
~(N). X!: Yand tE <:is o(N). Then ",-(Lyt) = Lx",-t. Proof. Recall from 4.2.4 that X!: Y iff Gil. ° '" = '" ° FA' where F/I. and Gil. are the flows of X and Y. respectively. Thus by 5.4.1.
Lx",·t=.!!..1 Ft",·t=.!!..1 (",oFA)·t= ~I (Gil. ° ",)·t d" 11.-0 d" 11.-0 d" /1.-0
.!!..I d"
11.-0
",·G:t = ",-
.!!..I d"
G:t = ",.( Lyt). • /1.-0
As for functions and vector fields. the Lie derivative can be generalized to include time-dependent vector fields.
THE LIE DERIVATIVE: DYNAMIC APPROACH
307
5.4.4 Theorem. Let XA E ~X k (M), k ~ I for each;\, E R and suppose that X(;\.. m) is continuous in ;\.. m. Then if FA.,. is the evolutiol/ operator for XA• we have
where t
E
Proof.
~'1"( M) is of dass C k •
As before. it is enough to prove this at;\.
= JL
where FA• A = identity.
for then
Now as in 5.4.1, ex.' = (d/d;\')f"A.,.tI A _,. is a differential operator that coincides with Lx. on ~'t and ~'\ by 4.2.31. Thus by 5.3.2, ex. = Lx. on all. tensors. • We caution again, as in Section 4.2. that it is not generally true for time-dependent vector fields that the right hand-side in 5.4.4 is Lx.f"A.,.t. Let us generalize the relationship between Lie derivatives and' flows one more step. Call a smooth map t: R X M -+ T.'( M) satisfying tA(m) = t( >'. m) E (Tm M)~ a time-dependent tensor field. Theorem 5.4.4 generalizes to this context as follows.
5.4.5 Theorem. Let tA and XA be C k time-dependent tensor and vec·tor fields. k ~ I and denote by FA.,. the et'olution operator (If X,. Then
Proof.
By the product rule for derivatives and 5.4.4 we get d -d i F! t-d~ A-A. A.,. A - d"A
I
A-A"
Pt
A.,.
An
+F!
A•. ,.
dtA _.d>'
I
A -Au
308
TENSORS
5.4.8 Examples A. If g is a pseudo-Riemannian metric on M. the .Killing equations are Lxg ... 0 (see Example 5.3.68). 8y 5.4.2 this says that F!g = g. where F" is the now of X. i.e., that the now of X consists of isometries. B. In elasticity the vanishing of the strain tensor means. by Example A. that the body moves as a rigid body.
5.4A
Verify Theorem 5.4.1 by a coordinate computation as follows. Let F,,(x) = (yIp,. x), .. . ,y"(A. x» so that ilyi/ ilA = Xi( y) and iI y'/ ax' satisfy the variational equation
Then write
(.._ ).,......,....!II -a,·' . . . a,·· - - ... --I' • fix ' fix' fI, , fI, . 1~,
I
!II
I
iJxll
iJx)'
!II.
!II
•
i"';
.Jt- ..J.
DilfeRDtiate this in A at A - 0 aDd obtain the coordinate cxprasioo (4) 01 SectioG S.3 for Lx" S.4B Carry out abe proof outlined in Exercise S.4A for time-depeodent wc:tor fields. S.4C Star1iDa with Theon:m 5.4.1 as the definition of L II ,. dJeck ....t Lx .Ii.... (DOl). (1)02) ud the propertieI (i)-(iv) 01 S.3.1S.4D Let C be. contraction operator mappina 5'.'(M) to ~~i '(JI). Use both 5.4.1 ud (DOl) to show that L II ( 0) - C(LII')' S.4E
F.xteod theorem 5.4.1 to F-valued teDsors.
S.4F
(lNfomttl/iOft LmtmG). Two tensor fields " aDd " +'2 of the same type OIl • compact manifold AI are called ~I if then is • diffeomorphism .,: AI ... M such that "-', - " + 'a. Assume that for o.. A .. )•• wc:tor facld X). on M can be found that satisfies LxJl1
Pro\Ie that I, and
+ '2)+1 2 • 0
'I + '2 are equivalent as follows: Let 'Pl be the now
of Xl with 4fIo -Id. Now calculate :", 'P%(',
+ A'2)'
NOI.: the method in this argument is caUed the method of U. trtI1&r/omu. It has been used already in the proof of Frobenius
PARTITIONS OF UNITY
309
theorem (section 4.4) and we shall see it again in 6.4.14 and 8.1.2. The method is also used in the theory of normal forms (cf. Takens [1914] and Guckenheimer and Holmes [1983]).
5.5 PARTITIONS OF UNITY A partition of unity is a technical device that is often used to piece smooth local tensor rields together to form a smooth global tensor field. Partitions of unity will be especially useful for studying integration. but in this section they are used to study when a manifold admits a Riemannian metric.
5.5.1 Definition. If t is a tensor field on a manifold M. the carrier of t is the set of m E M for which t( m) * O. and is denoted carr t. The .fuppon of t. denoted suppt. equals the closure of carr t. Also. we s~v t has compact support if suppt is compact in M. An open set U eM is called a C' carrier if there ex/st.I' an f E ~'f'( M). such that f ~ 0 and U = carr f. A collection of subsets {Ca} of a manifold M (or. more generalll'. a topological space) is called toeaUy linite if for each 111 E M. there is a neighborhood U of m such that U () Ca = 0 eX('ept for finitc/I' many indices a: 5.5.2 ,Definition. A partition 01 lIIIity on a manifold M i.t a collection
{(1I,. gil). where (i)
(ii) g; E '!F( M). gj( m) ~ 0 for all m EM. and supp g, c 1I, for all i; (iii) For each mE M. Ljg;(m) = I. [By (i). this is a finite ovum.) If if = {( Va. CPa)} is an atlas on M. a partition 01 unity ,fubordinate to ti'is a partition of unity {( 1I,. g,)} such that each open set 1I, is a subset of some chart domain Val;)' If for any atlas if there exists a partition of unity subordinate to it. we say M admits part;tions olllll;ty.
Occasionally one works with C k partitions of unity. They are defined in the same way except g, are only required to be C* rather than C'''. Partitions of unity are used for a basic patching constnlction. which may be explained as follows.
5.5.3 Proposition. Let M be a manifold with an atlas ti = {( VB' CPa)} SO CPa: Va -+ V~ c E is a chart. Let ta be a C k tensor field. k ~ I. of fixed type (~) defined on V~ for each a. and assume that there' exists a partition of unity
310
TENSORS
{( lI;, gi)} subordinate to (i. Let t be defined by t(m) == Lg/CP:(I)ta(/)(m), /
a /lnite .fum at each me M. Then t is a
e· tensor field of type (;) on M.
Proof. Since {lI;} is locally finite, the sum at every point is a finite sum, and thus t(m) is a type (~) tensor for every mE M. Also, t is e* since the local representative of' in the chart (VO(i),CPa(i) is Lj(gioCP.;(J)tO(j)' the summation taken over all indices j such that Valil n Va(j) '* 12l; by local finiteness the number of these j is finite. •
It should be remarked that t is not unique; it clearly depends on the choices of the indices a(i) such that lI; c VO(i) and on the functions gi' As we shall see later, under suitable hypotheses. one can always construct partitions of unity; again the construction is not unique. The same construction (and proof) can be used to patch together local sections of a vector bundle into a global section when the base is a manifold admitting partitions of unity subordinate to any open covering. To discuss the existence of partitions of unity and consequences thereof, we need some topological preliminaries.
5.5.4 DeftnlUon. Let S be a topological space. A covering {Va} of S is called a ",iMrrtettt of a covering {V,} if for every VA there is a V, sudI that Va c V,. A topological space is called paracomptld if every open covering of S has a locally finite refinement of open sets, and S is Hausdorff. 5.5.5 Propoaltlon. Second-countable, locally compact Hausdorff spaces are paracompact.
By second countability and local compactness of S, there exists a sequence 0., ... ,0", ... , of open sets with c1(0,,) compact and u nENO" ""' S. Let V" - 0. U .•• U 0", n -1,2,'" and put U,- V,. Since (V,,) is an open covering of S and c1(V.) is compact, cl(U,)c v,. u ... U v, .. Put V2 ... v" U ... U V, ; then c1(V2 ) is compact. Proceed inductively to show that S is the countable union of open sets V" such that c1(U,,) is compact and c1(V")cV,,n If Wo is a covering of S by open sets, and K,,=c1(V,,)\V,,_., then we can cover K" by a finite number of open sets, each of which is contained in some W" n V,,+.' and is disjoint from c1(V,,_2)' The union of such collections yields the desired refinement of {Wa}' • Proof.
Another class of paracompact spaces are the metrizabJe spaces (see 5.5.15 in Box 5.5A). In particular, Banach spaces are paracompact.
PARTITIONS OF UNITY
311
5.5.8 Propoaltlon. Every paracompact space is normal. Proof. W~ first show that if A is closed and u E S\A. there are disjoint neiahborhoods of u and A (relularity). For each v E A leI UU ' be disjoint neiahborhoods of u and v. Let W" be a locally finite refinement of the covering Vv ' S\A. and V - U W". the union over those a with W" n A "" 0. A neighborhood Uo of u meets. a finite number of Wa' Let U denote the intersection of Uo and the corresponding UV ' Then V and U are the required neighborhoods. The case for two closed sets proceeds somewhat similarly, so we leave the details for the reader. •
v..
Later \:Ve shall give general theorems on the existence of partitions of unity. However, there is a simple case that is commonly used, so we present it first.
5.5.7 Theorem. Let M he a second-countable Then M admits partitions 01 unity.
(Hau.~dorlf)
n-manilold.
Proof. The proof of 5.5.5 shows the following. Let M be an n-manifold and {W,,} be an open covering. Then there is a locally finite refinement consisting of charts CV;, cp;) such that cp, (V.) is the disk of radius 3. and such that cp/-·(D.(O» cover M, where D.(O) is the unit disk, centered at the origin in the model space. Now let (£. be an atlas on M and let {( cp;)} be a locally finite refinement with these properties. From 4.2.13, there is a nonzero function hi E <j(M) whose support lies in V. and hI;" O. Let
v..
h;(u) g;(u) = L;hi(u)
(the sum is finite). These are the required functions.
•
If (V.. ) is an open covering of M, we can always find an atlaS
cl' ... {( u,. CPi)} such that (u,) is a refinement of (V.. ) since the atlases generate the topology. Thus. if M admits partitions of unity. we can find partitions of unity subordinate to any open covering. . , The case of CO.partitions of unity differs drastically from the smooth case. Since we are primarily interested in this latter case. we su1llllUtrize the topologicai· situation, without giving the proofs: If S is a Hausdorff space, the following are equivalent:
S is normal. (ii) For any two closed nonempty disjoint sets A, B there is a continuous function I: S -+ (0, I) such that I( A) - 0 and 'f( B) - I (Urysohn's lemma). (i)
312
TENSORS
(iii) For any closed set A c S and continuous function f: A --+ rlI. h), there is a continuous extension j: S --+ [a. h) of f (Tietze extension theorem). A Hawdorff space is paracompact iff it admits a CO partition of unity subordinate to any open covering.
Indeed, if {( U;. g;)} is a continuous partition of unity subordinate to the given open covering {Va}' then by definition, {U;} is an open locally finite refinement. The converse-the ·existence of partitions of unity-is the hard part; the proof of this and of the equivalences of (i) with (ii) and (iii) can be found for instance in Kelley (1975) and Choquet (1969: sec. 6). These results are important for the rich supply of continuous functions they provide; We shall not use these topological theorems in the rest of the book, but we do want their smooth versions on manifolds. Note that if M is a manifold admitting partitions of unity subordinate to any open covering. then M is paracompact, and thus normal by 5.5.6. This already enables us to generalize (ii) and (iii) to the smooth (or C k ) situation.
5.5.8 Propoeltlon. Let M be a manifold admitting smooth (or C") partitions of unity. If A and B are closed disjoint sets then. there exists a smooth (or Cit) /unction /: M ... [0.1] such that f(A) - 0 and /(B) -I.
Proof. As we saw. the condition on M implies that M is normal and thus there is an atlas {(Va' CPa)} such that V n A '* 0 implies Va n 8 = 0. Let R
(II;. g;} be a subordinate C k partition on unity andf =
LS;. where the sum is over those i for which It; n B '* 0. Then f is C k • is one on 8 and zero on A .
•
Next we prove an extension theorem, which is a smooth analogue of (ii).
5.5.9 PropoalUon. Let M be a manifold admitting partitions of unityj and let '11': E ..... M be a vector bundle with base M. Suppose 0: A ... E is a C" section defined on the closed set A (i.e .• every point a E A hils a neighborhood Va and a C" section G,,: V" ..... E extending G). Then G can be extended to a C" global section i: M ..... E. In particular, iff: A ..... F is a k function defined on the closed set A, where F is a Banach space. then there is a Cit extension j: M ..... F; iffis bounded by a constant R, i.e .• llf(a~I" Rfor all a E A, then so is j.
c
Consider the open covering {V". M\A la E A} of M. with Va given by the definition of smoothness on the closed set A. Let {(U;, g;)} be a C" partition of unity subordinate to this open covering and define G;: U; ..... E, by Gj ... G" IU; for all U; contained in some Va and OJ 0 on all U; disjoint Proof.
=
PARTITIONS OF UNITY
313
from all Va. a EA. Then g,o,: If; -+ E is a C A section over If; and since supp(g;o;)csupP(g;)Clf;, it can be extended in a C A manner to M by putting it equal to zero on M\lf;. Thus 8;0,: M -+ E is a C"-section of 'If: E -+ M and hence a= 1:;g;o; is a C k section; note that the sum is finite in a neighborhood of every point mE M. Finally, if a EA.
a( a ) =
L 8, ( a ) 0, ( a ) = ( L g; ( a ) ) °(a ) = °(tl ) • ,
"
i.e.•
alA =
0.
The second part of the theorem is a particular case of the one just proved by considering the trivial bundle M x F -+ M and the section defined by o(m)= (m.J(m». Finally. the bounded ness statement follows from the given construction. keepIng in mind that all the g, have values in [0.1]. •
°
Before discussing general questions on the existence of partitions of unity on Banach manifolds. we will discuss the existence of Riemannian metrics. Recall that a Riemannian metric on a Hausdorff manifold M is a tensor field g E 5 2o( M) such that for all m E M, g( m) is symmetric and positive definite. Our goal is to find topological conditions on an n-manifold that are necessary and sufficient to ensure the existence of Riemannian metrics. The proof of the necessary conditions will be simplified by first showing that any Riemannian manifold is a metric space. For this, define form,nEM, d(m. n) = inf{l( y )Iy: [0.11-+ M is a c l curve with y(O)
Here I(y) is the length of
tlJ~-
=
m. y( I) = n}:
curve y. defined by ,
where Y(t) = dy/dt is the tangent vector at y(t) to the curve y and == [gY(I)(y(t), 1'(1»)1/2 is its length .
. IIY(t~1
. 5.5.10 Proposition. d is a metric on each connected ('I)mponent oj M whose metric topology is the original topology oj M. Proof. Clearly d(m. m) = O. d(m. n) = d(n, m). and d(m. p)" d(m"n)+ d(n, p) by definition. Next we will verify that d(m. n) > 0 whenever m * n .. Let m E V c M where (U, qJ) is a chart and suppose qJ( V) = u'. c E. Then for any. u E V. g(uXv. V)I/2, defined for vE TuM. is a norm on TuM.
314
TENSORS
This is equivalent to the norm on E, under the linear isomorphism TmfJ!. Thus, if f is the local expression for g, then f( u') defines an inner product on E, yielding equivalent norms for all u' E U'. Using continuity of g and choosing U' to be an open disk in E, we can conclude that the norms f( u')1/2 and f( m')1/2, where m' -= fJ!( m) satisfy: af( m,)1/2 ~ f( U,)1/2 ~ bf(m,)1/2 for all u' E U', where a and b are positive constants. Thus, if 71: [0, I] ..... U' is a C l curve, then
~a
f,(m')(
,,(I). ,,(/»1/2 dl
I
I
~a,(m') ( 10 ,,(/)dl.fo ,,(t)dl
)1/2
~ ai( m')( 71(1 )-1/(0).1/( I )-1/(0» 1/2
Here we have used the following property of the Bochner integral: II[f(/) dtll
~ t//f(t)//dt.
valid for any norm on E (see the remarks following 2.2.7). Now let y: [0. I) ..... M be a C l curve. yeO) = m. y(l) = n. m E U.( u. ' O. The equivalence of the original topology of M and of the metric topology defined by d is clear if one notices that they are equivalent in every chart domain U, which in turn is implied by their equivalence in cp( U). • Notice that the preceding proposition holds in infinite dimensions. 5.5.11 PropoaIUon. A connected Hausdorff n-manifold admits a Riemannian metric if and only if it is second countable. Hence for Hausdorff n-manifolds (not necessarily connected) paracompaclness and metrizabilily are equivalent.
PARTITIONS OF UNITY
315
Proof. If M is second countable, it admits partitions of unity by 5.5.7. Then the patching construction 5.5.3 gives a Riemannian metric on M hy choosing in every chart the standard inner product in R". Conversely, assume M is Riemannian. By 5.5.10 it is a metric space, which is locally compact and first countable. being locally homeomorphic to R". By \.6.14, it is second countable. •
The main theorem on the existence of partitions of unity in the general case is as follows.
5.5.12 Theorem. Any second-countable or paraCtlmptll"l manifold moJt!/ed on a.separable Banach space with a C A norm away from 'he origin tldmils C k partitions of unity. In particular paracompact (or second ("OItntahie) manifolds modeled o~ separable Hilbert spaces admit eX) parlitions flf unity. 5.5.13 Corollary. Paracompact «()r second countable) manifolds modeled in separable real Hilbert spaces admit Riemannian metrics. Theorem 5.5.12 will be proved in the following two boxes. There are Hausdorff nonparacompact n-manifolds. These manifolds are necessarily nonmetrizable and do not admit partitions of unity. The standard example of a one-dimensional nonparacompact Hausdorff manifold is the "long line." In dimensions 2 and 3 such manifolds are constructed from the Priifer manifolds. Since nonparacompact manifolds occur rarely in applications, we refer the reader to Spivak [1979, vol I, Appendix A] for the aforementioned examples. .
BOX S.U PARTITIONS OF UNITY: REDUCTION TO THE LOCAL CASE
Let us begin with some topological preliminaries. Let S be a paracompact space. If {UII} is an open covering of S. it can be refined to a locally finite open covering (WIl ). The first lemma below will show that we can shrink this covering further to get another one, {Va} such that cl( Va) C Wa with the same indexing set. A technical device used in the proof is the concept 'of a well-ordered set. An ordered set A in which any two elements can be ' compared is well-ordered if every subset has a smallest element. The axiom qf choice is equivalent to Zermelo's lemma: "Every set can be well-ordered." See Appendix A.
5.5.14 Lemma (Sltrinlcing i.emma). Let S be a normal space and {Wa)"eA a locally finite open covering of S. Then there exists a locally
316
TENSORS
finite open refinement {Va}a. A (with the same indexinR set) ,\'uch that cI(Ve)C We'
Proof. Well-order the indexing set A and call its smallest element a o. The set Co = S\ V .. > "OW.. is closed. so by normality there exists an open set v..o such that Co C c1( Va.) C Wa.. If Vy is defined for all y < a. put Ca - S\( V y < aVy V V y > eWy») and by normality find Va such that C.. C c1(V.. )C W... The collection {V.. }.. eA' is the desired locally finite refinement of (W.. }.. e A' provided we can show it covers S. Let s E S. By local finiteness of {W,,}. of he\ongs only to a finite collection W.. ,..... W.... If fJ,., max(al ..... a n ) then clearly s E Wy for all y > fJ. so that if in addition sEVy for all y < fJ. then s E Cp C Vp, i.e. s E Vp. •
5.5.15 Lemma (A. H. Stone). compact.
Every pseudometric space is para-
Proof. Let {V.. }.. EA be an open covering of the pseudometric space S with distance function d. Put Vn."-={xEV,,ld(x.S\Vn)~1/2n}. By the triangle inequality we have d(U,•. ". S\U. I I.,,)? 1/2 n -1/2" I I = 1/2 n t I. Well-order the indexing set A and let u,~ n = U•. ,,\ V II < II V. f l . p . If y,6 E A. we have Vn~y C S\V. t I,3' if Y < 6, or u,~" c S\ Vn+I.y. if 8
Let us now tum to the question of the existence of partitions of unity subordinate to any open covering. 5.5.16 Proposition (R. Palais). Let M be a paracompact manifold modeled on the Banach space E. The following are equivalent. M admits C" partitions of unity. Any open covering of M admits a locally finite refinement by C" carriers; (iii) For any open sets 1, 2 , such that c1(01) cO2 , there exists a C" carrier V such that 0 1 eVe 02' (iv) Every chart domain of M admits C" partitions of unity subordinate to any open covering. (i) (ii)
°°
PARTITIONS OF UNITY (v)
317
E admits C· partitions 01 unity subordinate to any open covering olE.
Proof. ((i) implies (ii». If {( ll" g,)} is a C"-partitioll of unity suhordinate to an open covering, then clearly carr K, form a locally finite refinement of the covering by C k carriers. «ii) implies (iii». Let {V,,}ae A be a locally finite refinement of the open covering (02 ' S\c1( O.)} by C" carriers and denote hy 10 E CS'* (M), the function for which carr 10. = Va' Let B = {a E A IVa C 0 2 }' Put V = nile BVfJ./ = E"e sill and remark that 0. eVe O2 , carr I = V. «iii) implies (iv». Let U he any chart domain (If M. Then V is diffeomorphic to an open set in E which is a metric space. so is paracompact by Stone's theore~ 5.5.15. Let {Un} he an arbitrary open covering of V and {VfJ} be a locally finite refinement. By the shrinking lemma we may assume that c1( VfJ) c V. Again by the shrinking lemma refine further to a locally finite covering {WfJ} such that c1( W,I ) C "~. But by (iii) there ex.ists a C·-carrier Oil such that W" c 0" C VfJ. and so {Oil} is a locally finite refinement of {VfJ} by CA-carriers. whose correspqnding functions we denote by IfJ. Thus I = E"I" is a C k map and {(VfJ./fJ/f)} is a C" partition of unity subordinate to {Un}' ((iv) implies (v». Consider now any open covering {Vn}n' • of E and let (V. q» be an arbitrary chart of M. Refine first the covering of E by taking the intersections of all its elements with all translates of q>(V). Since E is paracompact. refine again to a locally finite open covering {Vp}. The inverse images by translations and cr of these open sets are subsets of V, hence chart domains, and thus by (iv) they admit partitions of unity subordinate to any coveri~g. Thus every Yp admits a C" partition of unity subordinate to any open covering, for example to {VIl n Vala E A}; call it {gf}. Then g = Ei.llgf is a C" map and the double-index.ed set of functions gf/g forms a C" partition of unity of E. «v) implies (iii». If E admits C" partitions of unity subordinate to any open covering. then so does every open subset by the (already proved) implication of (i) =0 (iii) applied to M = E. which is paracompact by 5.5.13. Thus if (V, q» is a chart on M. V admits partitions of unity, since q>( V) does. Finally, we show (iii) implies (i). Choosing a locally finite atlas, this proof repeats the one given in the last part of (iv) ,. (v). • As an application of this proposition we get the following.
5.5.17 Proposition. filions of unity.
El.'erl' paracompact n-manilold admits C<XJ-par-
318
TENSORS
1'roof.
By 5.5.16(ii) and (v) it suffices to show that every open set in
R" is a Coo carrier. Any open set U is a countable union of open disks D1• By 4.2.13, D/- carr/;, for some Coo function /;: R" -+ R. Put M, - sup(JID"/;(%~11 %E R", k.; i} and let 00
/- E /;/2 IM/. I-I
By Exercise 2.4J / is a Coo function for which carr/ ... U clearly holds.
•
In particular, second-countable n-manifolds admit partitions of unity, recovering 5.5.7. .
BOX AlB PARTInONS OF UNITY: THE LOCAL CASE Theorem 5.5.15 reduces the problem of the existence of partitions of unity to the local one, namely finding partitions of unity in Banach spaces. This problem has heen studied by Bonic and Frampton (1966) for separable Banach spaces.
5.5.1' PropoalUon CR. /lottie tntd J. FlYIIIIfJtOtl/I966J). separable Banach space. The /oIlowing are equivalent.
Lei E be a
(i) Any open set 0/ E is a C k carrier. (ii) E admits CIt partitions 0/ unity subordinate to any open· covering
(iii)
o/E; There exists a bounded nonempty CAt carrier in E.
Proof. By 5.5.16, (i) and (ii) are equivalent since E is paracompact by 5.5.15. It remains to be shown that (iii) implies (i), since clearly (ii) implies (iii). This proceeds in several steps. First, we show that any neighborhood contains a Cit carrier. Let U be any open set and let carr / c Dr(O) be the bounded carrier given by (iii), / E C"(E), /;;. O. Let e E U, fix eo E carr/, and choose r> 0 such that D.(e) c U. Define g E Ck(E), g ~Oby
K>O,
where K remains to be determined from the condition that carr g c +lIeoIO/r, this
D,(e). An easy computation shows that if K> (r
PARTITIONS OF UNITY
319
inclusion is verified. Since e E carr g. carr g is an open neighborhood of e. Second. we show that any open set can be covered by a countable locally finite family of C" carriers. By the first step. the open set V can be covered by a family of C" carriers. By Lindelors theorem. 1.1.6. V == U "V" where Vn is a C" carrier. the union being over the positive integers. We need to find a refinement of this covering by C" carriers .. Let/"E'C"(E) be such that carr In = V". Define Vn={eEEl/n(e» O. h(e) < lin for all i < n}. Oearly VI = VI and inductively Vn = v,. n n'; < ,,!; I(]_ oo.l/n[). By the composite function theorem the inverse image of a C" carrier is a C" carrier. so that /, I(]_ oo.l/n() is a C" carrier. since ]00.1/ n[ is a ~" carrier in R (see the proof of 5.5.17). Finite intersections of C" carriers is a C" carrier (just take the product of the functions in question) so that Vn is also a C~ carrier. Clearly Vii c v". We shall prove that {Vn } is a locally finite open covering of U. Let. e V. If. e Vn for all n. then clearly e e VI = VI' If not, then there exists a.smallest no such that e E v" .. Then /,(e) = () fori < no and thus e e V" - (e e EI/".(e) > 0, h(e) < Iino for all i < no}. Thus {Vn} covers V. This open covering is also locally finite for if e E V" and N is such that I(e) > liN. then the neighborhood{u E VII.,(e) > 1/ N) has empty intersections with all Vm for m > N. Third we show that the open set V is a C"-carrier. By the second step. V - U "V". with Vn a locally finite open covering of V by C" carriers. Then I = Enl" is C". i(e) ~ 0 for all e E E and carr 1= V. • The separability assumption was used only in showing that (iii) implies (i). There is no general theorem known to us for nonseparable Banach spaces. Also. it is not known in general whether Banach spaces admit bounded C" carriers. for Ie ~ I. However. we have the following. 5.~.1' Propo8ltlon. II the Banach space E has a norm C" away Irom its origin. k ~ I. then E has bounded C"-carriers.
Proof. By 4.2.13 there exists cp: R ..... R.C"" with compact support and equal to one in a neighborhood of the origin. If 11'11: E\{O} ..... R is C". k ;a. I, then cp °11'11: E\{O} ..... R is a nonzero map which is C". has bounded support 11'11- 1 (suppcp), and can be extended in a C~ manner toE. • Theorem 5.5.12 now follows from 5.5.19,5.5.18. and 5.5.16. The situation with regard to Banach subspaces and submaru.foJds is clarified in the following proposition, whose proof is an immediate consequence of 5.5.18 and 5.5.16.
320
TENSORS
5.5.20 ProposlUon. (i) If E i.f a Banach space admitting C A partitions of unity then so docs any closed suhspace. (ii) If a manifold admits C k partitions of unity subordinate to any open covering, then so does any submanifold. We shall not develop this discussion of partitions of unity on Banach manifolds any further, but we shall end by quoting a few theorems that show how intimately connected partitions of unity are with the topology of the model space. By 5.5.18 and 5.5.19. for separable Banach spaces one is interested whether the norm is C k away from the origin. Restrepo [1964] has shown that a separable' Banach space has a C I norm away from the origin if and only if the dual is separable. Bonie and Reis [1966] and Sundaresan [1967] have shown that if the norms on E and E· are differentiable on E\{O} and E·\{O}. respectively, then E is reflexive. for E a real Banach space (not necessarily separable). Moreover. E is a Hilbert space if and only if the norms on E and E· are twice differentiable away from the origin. This result has been strengthened by Leonard and Sundaresan [1973]. who show that a real Banach space is isometric to a Hilbert space if and only if the norm is C 2 away from the origin and the second derivative of e ..... UeU 2/2 is bounded by I on the unit sphere; see Rao [1972] for a related result. Palais [1965b] has shown that any paracompact Banach manifold admits Lipschitz partitions of unity. Because of the importance of the differentiability class of the norm in Banach spaces there has been considerahle work in the direction of determining the exact differentiabilit\ . l:J~s of concrete function spaces. Thus Bonic and Frampton [1966] haH: shown that the canonical norms on the spaces LP(R).IP(R). p ~ I. p < 00 are COO away from the origin if p is even. C p - I with DOl-liP-I) Lipschitz, if p is odd. and C[pi with Dlp)(II'UP) HOld"r continuous of order p -[pl. if p is not an integer. The space Co of sequences of real numbers convergent to zero has an equivalent norm that is Coo away from the' origin. a result due to Kuiper. Using this result, Frampton and Tromba [1972] show that the A-spaces (closures of Coo in the HOlder spaces) admit a Coo norm away from the origin. The standard norm on the Banach space of continuous real valued functions on [0. I] is nowhere differentiable. Moreover. since CO([O.I).R) is separable with nonseparable dual. it is impossible to find an equi\ .• lent norm that is differentiable away from the origin. To our knowledge it is still an open problem whether CO([O.I).R) admits C k partitions of unity for k;;d. Finally. the only results known to us for nonseparable Hilbert spaces are those of Wells [1971]. [1973]. who has proved that nonsep-
PARTITIONS OF UNITY
321
arable Hilbert space admits ("2 partitions cif unity. The techniques used in the proof. however. do not seem to indicate a general way to approach this problem.
Exercl... 5.5A
5.5B
Show that any closed set Fin R" is the inverse image of 0 hy II C' real-valued positive function on Rn. Generalize this to any n-manifold. (Hint: proof of 5.5.11 for R"\F.) In a paracompact topological space. an open suhset need not he paracompact. Show that (i) If every open subset of a paracompact space is paracompact. then any subspace is paracompact. (ii) Every open submanifold of a paracompact manifold is parac:ompICt. (HiIU: Uk chart domains to coru:lude metIUabilily.)
Let ..: £ - M be a vector bundle. £' c £ a 5ubbundle and assume III . admits C· partitions of unity subontiDate to any open coverin&. Show that £' splits in E, i.e.. there exists • subbundle £11 such that E - E'eE". (H;III: The result is trivial for local bundles. Construc:t for Cvay element of a locally fmite covering (ll;) a vector 'bundIe map /, wbose kernel is tt, .. complement of E', u,. for u" II» a C· partitioa of unity, put; <...;&;/, and show that E - E'eter/.) S.SD Let ..: E - III be a vector bundle OWl" the base M that admits C· partitions of unity and with the fibers 01 E modeled OIl a Hilbert space. Show that E admi~ a C· bundle metric, i.e., a C· map ,: M -Tz°( E) that is symmetric stronaJy aondeaenerate and positive definite at every point n • . ;. N. S.SC
«
Let E - III be a line bundle OWl" the manifold III admittina· C· partitions of unity subordinate to any open coveriDa- Sbow that Ex E is trivia) (HIIII: Ex E - L(E·, E) and coastnact.1ocal bale that am be extended.) S.SF Assume M admits C· partitioos 01 unity. Show that any submanifold . of III diffeomorphic to S' is the intepaJ CUIW of a C· wctor Ileld OIl M. 5.50 Let /II be a connected paracompact manifold. Show that tIMn exists a C· proper mappinl/: M - R·. (Hlltt: /II is secoad countable, beina Riemannian. Show the statement for Ie -1. where 1and is a countable partition of unity.) 5.5H Let /II be a connected paracompact ,...manifoJd and X e ~(/II). Show that there exists II E ~(M).II > 0 such that Y -AX ii ~ S.SE
<.,)
t::..i.,;.
322
TENSORS
plete. (Hint: With! as in 5.50 put h = e C~'lfl" so that I YUIi ~ 1. Hence (f 0 c)(Ja. h[) is bounded for any integral curve (' of Y and Ja. b[ in the domain of c.) 5.51 Show that every compact n-rnanifold embeds in some Rk for k big enouah in the following way. II {(lJ,.cp/)}/.' •... N is a finite atlas with cp/(lJ,) the ball of radius 2 in R", let X e COG(R"). X -Ion the ball of radius 1 and X - 0