Текст
                    145 Isoperimetric Inequalities
This introduction treats the classical isoperimetric inequality in Euclidean space
and contrasting rough inequalities in noncompact Riemannian manifolds. In Eu-
Euclidean space the emphasis is on quantitative precision for very general domains,
and in Riemannian manifolds the emphasis is on qualitative features of the
inequality that provide insight into the coarse geometry at infinity of Riemannian
manifolds.
The treatment in Euclidean space features a number of proofs of the classical
inequality in increasing generality, providing in the process a transition from the
methods of classical differential geometry to those of modern geometric mea-
measure theory; and the treatment in Riemannian manifolds features discretization
techniques and applications to upper bounds of large time heat diffusion in
Riemannian manifolds.
The result is an introduction to the rich tapestry of ideas and techniques of
isoperimetric inequalities, a subject that has beginnings in classical antiquity
and that continues to inspire fresh ideas in geometry and analysis to this very
day - and beyond.
Isaac Chavel is Professor of Mathematics at the City College of The City
University of New York. He received his Ph.D. in mathematics from Yeshiva
University under the direction of Professor Harry E. Rauch. He has published
in international journals in the areas of differential geometry and partial dif-
differential equations, especially the Laplace and heat operators on Riemannian
manifolds. His other books include Eigenvalues in Riemannian Geometry and
Riemannian Geometry: A Modern Introduction.
He has been teaching at the City College of The City University of New York
since 1970, and he has been a member of the doctoral program of The City Uni-
University of New York since 1976. He is a member of the American Mathematical
Society.


CAMBRIDGE TRACTS IN MATHEMATICS General Editors B. BOLLOBAS, W. FULTON, A. KATOK, F. KIRWAN, P. SARNAK 145 Isoperimetric Inequalities
Isaac Chavel The City University of New York Isoperimetric Inequalities Differential Geometric and Analytic Perspectives Cambridge UNIVERSITY PRESS
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE The Pitt Building, Tmmpington Street, Cambridge, United Kingdom CAMBRIDGE UNIVERSITY PRESS The Edinburgh Building, Cambridge CB2 2RU. UK 40 West 20th Street, New York, NY 10011-4211, USA 10 Stamford Road, Oakleigh. VIC 3166, Australia Ruiz de Alarcon 13, 28014 Madrid. Spain Dock House, The Waterfront, Cape Town 8001, South Africa http://www.cambridge.org © Cambridge University Press 2001 This book is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 2001 Printed in the United Stales of America Typeface Times 10/13 pt System UR&.2e [tb] A catalog record for this book is available from the British Library. Library of Congress Cataloging in Publication Data Chavel, Isaac. Isoperimetric inequalities : differential geometric and analytic perspectives / Isaac Chavel. p. cm. - (Cambridge tracts in mathematics ; 145) Includes bibliographical references and index. ISBN 0-521-80267-9 1. Geometry, Differential. 2. Isoperimetric inequalities. 3. Riemannian manifolds. I. Title. II. Series. QA641.C45 2001 5I6.3'6-dc21 2001016177 ISBN 0 521 80267 9 hardback Contents Preface I Introduction 1.1 The Isoperimetric Problem 1.2 The Isoperimetric Inequality in the Plane 1.3 Preliminaries 1.4 Bibliographic Notes II Differential Geometric Methods II. 1 The C2 Uniqueness Theory 11.2 The C1 Isoperimetric Inequality 11.3 Bibliographic Notes III Minkowski Area and Perimeter III. 1 The Hausdorff Metric on Compacta 111.2 Minkowski Area and Steiner Symmetrization 111.3 Application: The Faber-Krahn Inequality 111.4 Perimeter 111.5 Bibliographic Notes IV Hausdorff Measure and Perimeter IV. 1 Hausdorff Measure IV.2 The Area Formula for Lipschitz Maps IV.3 Bibliographic Notes V Isoperimetric Constants V.I Riemannian Geometric Preliminaries V.2 Isoperimetric Constants V.3 Discretizations and Isoperimetric Inequalities V.4 Bibliographic Notes page ix 1 1 4 13 28 30 30 44 51 52 53 58 78 81 99 100 100 110 116 117 118 127 139 155 vn
viii Contents VI Analytic Isoperimetrk Inequalities VI. 1 L2 Sobolev Inequalities V1.2 The Compact Case V1.3 Faber-Krahn Inequalities VI.4 The Federer-Fleming Theorem: The Discrete Case VI.5 Sobolev Inequalities and Discretizations VI.6 Bibliographic Notes VII Laplace and Heat Operators VII. 1 Self-adjoint Operators and Their Semigroups VII.2 TheLaplacian VII.3 The Heat Equation and Its Kernels VII.4 The Action of the Heat Semigroup VII.5 Simplest Examples VII.6 Bibliographic Notes VIII Large Time Heat Diffusion Vin.l The Main Problem VIII.2 The Nash Approach VIII.3 The Varopoulos Approach VII1.4 Coulhon's Modified Sobolev Inequality Vrn.5 The Denouement: Geometric Applications Vm.6 Epilogue: The Faber-Krahn Method Vin.7 Bibliographic Notes Bibliography Author Index Subject Index 157 157 161 165 170 174 183 185 185 192 201 216 222 224 225 226 234 236 240 243 249 253 255 263 265 Preface This book discusses two venues of the isoperimetric inequality: (i) the sharp inequality in Euclidean space, with characterization of equality, and (ii) isoperi- isoperimetric inequalities in Riemannian manifolds, where precise inequalities are un- unavailable but rough inequalities nevertheless yield qualitative global geometric information about the manifolds. In Euclidean space, a variety of proofs are presented, each slightly more ambitious in its application to domains with irregular boundaries. One could easily go directly to the final definitive theorem and proof with little ado, but then one would miss the extraordinary wealth of approaches that exist to study the isoperimetric problem. An idea of the overwhelming variety of attack on this problem can be quickly gleaned from the fundamental treatise of Burago and Zalgaller A988); and I have attempted on the one hand to capture some of that variety, and on the other hand to find a more leisurely studied approach that covers less material but with more detail. In Riemannian manifolds, the treatment is guided by two motifs: (a) the dichotomy between the local Euclidean character of all Riemannian manifolds and the global geometric properties of Riemannian manifolds, this dichotomy pervading the study of nearly all differential geometry, and (b) the discretiza- discretization of Riemannian manifolds possessing bounded geometry (some version of local uniformity). The dichotomy between local and global is expressed, in our context here, as the study of properties of Riemannian manifolds that remain invariant under the replacement of a compact subset of the manifold with an- another of different geometry and topology, as long as the new one fits smoothly in the manifold across the boundary of the deletion of the original compact subset. Thus, we do not seek fine results, in that we study coarse robust invari- invariants that highlight the "geometry at infinity" of the manifold. Our choice of isoperimetric constants will even be invariant with respect to the discretization of the Riemannian manifold. The robust character of these new isoperimetric IX
x Preface constants will then allow us to use this discretization to show how the geometry at infinity influences large time heat diffusion on Riemannian manifolds. Regrettably, there is hardly any discussion of isoperimetric inequalities on compact Riemannian manifolds. That would fill a book - quite different from this one - all by itself. A summary of the chapters goes as follows: Chapter I starts with posing the isoperimetric problem in Euclidean space and gives some elementary arguments toward its solution in the Euclidean plane. These arguments are essentially a warm-up. They are followed by a summary of background definitions and results to be used later in the book. Thus the discussion of the isoperimetric problem, proper, begins in Chapter II. Chapter II starts with uniqueness theory, under the assumption that the bound- boundary of the solution domain is C2. We first show that, if a domain fi with C2 boundary is a solution to the isoperimetric problem for domains with C2 bound- boundaries, S2 must be an open disk. Then we strengthen the result a bit - we show that if a domain is but an extremal for isoperimetric problems, then it must be a disk. Then we consider the existence of a solution to the isoperimetric problem for domains with C1 boundaries. We give M. Gromov's argument that for such domains the disk constitutes a solution to the isoperimetric problem. But only if one restricts oneself to convex domains with C1 boundaries does his argument imply that the disk is the unique solution. Chapter III is the heart of the first half of the book. It expands the isoperimet- isoperimetric problem in that it considers all compacta and assigns the Minkowski area to each compact subset of Euclidean space to describe the size of the boundary. In this setting, using the Blaschke selection theorem and Steiner symmetriza- tion, one shows that the closed disk constitutes a solution to the isoperimetric problem. Since the Minkowski area of a compact domain with C1 boundary is the same as the differential geometric area of the boundary, the result extends the solution of the isoperimetric problem from the C1 category to compacta. Moreover, one can use the traditional calculations to show that the disk is the unique solution to the isoperimetric problem in the C' category. But uniqueness in the more general collection of compacta is too difficult for such elementary arguments. Then, in Chapter III, we recapture Steiner's original intuition that successive symmetrizations could be applied to any compact set to ultimately have it con- converge to a closed disk - in the topology of the Hausdorff metric on compact sets. We use this last argument to prove the isoperimetric inequality for compacta with finite perimeter. The perimeter, as a measure of the area of the boundary, Preface XI seems to be an optimal general setting, since one can not only prove the isoperi- isoperimetric inequality for compacta with finite perimeter, but can also characterize the case of equality. In Chapter IV we introduce Hausdorff measure for subsets of Euclidean space and develop the story sufficiently far to prove that the perimeter of a Lipschitz domain in n -dimensional Euclidean space equals the (n — Yy- dimensional Hausdorff measure of its boundary. The proof involves the area formula, for which we include a proof. Chapter V begins a new view of isoperimetric inequalities, namely, rough inequalities in a Riemannian manifold. The goal of Chapters V-VM is to show how these geometric isoperimetric inequalities influence the qualitative rate of decay, with respect to time, of heat diffusion in Riemannian manifolds. In Chapter V we summarize the basic notions and results concerning isoperi- isoperimetric inequalities in Riemannian manifolds, and in Chapter VI we give their implications for analytic Sobolev inequalities on Riemannian manifolds. Chap- Chapter V consists, almost entirely, of a summary of results from my Riemannian Geometry: A Modem Introduction, and I have included just those proofs that seemed to be important to the discussion here. The discussion of Sobolev in- inequalities in Chapter VI has received extensive treatment in other books, but our interest is restricted to those inequalities required for subsequent applications. Moreover, we have also treated the relation of Sobolev inequalities on Rieman- Riemannian manifolds and their discretizations, one to the other. To my knowledge, this has yet to be treated systematically in book form. Chapter VII introduces the Laplacian and the heat operator on Riemannian manifolds and is devoted to setting the stage for the "main problem" in large time heat diffusion; its formulation and solution are presented in Chapter VHI. The book ends with an introduction to the new arguments of A.A. Grigor'yan, the full possibilities of which have only begun to be realized. I have attempted to strike the right balance between merely summarizing back- background material (of which there is quite a bit) and developing preparatory argu- arguments in the text. Also, although I have summarized the necessary basic defi- definitions and results from Riemannian geometry at the beginning of Chapter V, I occasionally require some of that material in earlier chapters, and I use it as though the reader already knows it. This seems the lesser of two evils, the other evil being to disrupt the flow of the arguments in the first half of the book for an excursus that would have to be repeated in its proper context later. Most of that material is quite elementary and standard, so it should not cause any major problems.
Xll Preface In order to clarify somewhat the relation between material quoted and ma- material presented with proofs, I have referred to every result that either is an exercise or that relies on a treatment outside this book as a proposition, and every result proven in the book as a theorem. This is admittedly quite artificial and obviously gives rise to some strange effects, in that the titles proposition and theorem are often (if not usually) used to indicate the relative significance of the results discussed. That is not the case here. There are bibliographic notes at the end of each chapter. They are intended to give the reader some guidance to the background material, and to give but an introduction to a definitive study of the literature. It is a pleasure to thank the many people with whom I have been associated in the study of geometry since I first came to the City College of CUNY in 1970: first and foremost, Edgar A. Feldman and the other geometers of the City University of New York - J. Dodziuk, L. Karp, B. Randol, R. Sacksteder, and J. Veiling. Also, I have benefited through the years from the friendship and mathematics of I. Benjamini, M. van den Berg, P. Buser, E. B. Davies, J. Eels, D. Elworthy, A. A. Grigor'yan, E. Hsu, W. S. Kendall, F. Morgan, R. Osserman, M. Pinsky and D. Sullivan. But, as is well known, any mistakes herein are all mine. The isoperimetric problem has been a source of mathematical ideas and techniques since its formulation in classical antiquity, and it is still alive and well in its ability to both capture and nourish the mathematical imagination. This book only covers a small portion of the subject; nonetheless, I hope the presentation gives expression to some of its beauty and inspiration. Introduction In this chapter we introduce the subject. We describe the classical isoperimet- isoperimetric problem in Euclidean space of all dimensions, and give some elementary arguments that work in the plane. Only one approach will carry over to higher dimensions, namely, the necessary condition established by classical calcu- calculus of variations, that a domain with C2 boundary provides a solution to the isoperimetric problem only if it is a disk. Then we give a recent proof of the isoperimetric inequality in the plane by P. Topping (which does not include a characterization of equality), and the classical argument of A. Hurwitz to prove the isoperimetric inequality using Fourier series. This is followed by a symme- symmetry and convexity argument in the plane for very general boundaries that proves the isoperimetric inequality, //one assumes in advance that the isoperimetric functional D >-*¦ L2(dD)/A(D) has a minimizer. (So this is a weak version - if the isoperimetric problem has a solution, then the disk is also a solution.) Finally, we present the background necessary for what follows later in our general discussion, valid for all dimensions. The subsections of §1.3 include a proof of H. Rademacher's theorem on the almost everywhere differentiability of Lipschitz functions, and a proof of the general co-area formula for C' mappings of Riemannian manifolds. We obtain the usual co-area formula, as well as an easy consequence: Cauchy's formula for the area of the boundary of a convex subset of R" with C1 boundary. LI The Isoperimetric Problem Given any bounded domain on the real line (that is, an open interval), the discrete measure of its boundary (the endpoints of the interval) is 2. And given any bounded open subset of the line, the discrete measure of its boundary is greater than or equal to 2, with equality if and only if the open set consists of one open interval. This is the statement of the isoperimetric inequality on the line.
2 Introduction In the plane, one has three common formulations of the isoperimetric problem: 1. Consider all bounded domains in R2 with fixed given perimeter, length of the boundary (that is, all domains under consideration are isoperimetric). Find the domain that contains the greatest area. The answer, of course, will be the disk. Note that the specific value of the perimeter in question is of no interest, because all domains of perimeter L\ are mapped by a similarity of R2 to all domains with perimeter L2 for any given values of Lu L2, and the image under the similarity of an area maximizer for L i is an area maximizer for L2. 2. One insists on a common area of all bounded domains under consideration, and asks how to minimize the perimeter. 3. Lastly, one expresses the problem as an analytic inequality, namely, since we know exactly the values of the area of the disk and the length of its boundary, the isoperimetric problem is then expressed as proving the isoperimetric inequality A.1.1) L2>4jr/t, where A denotes the area of the domain under consideration, and L denotes the length of its boundary. The inequality is extremely convenient, in that it remains invariant under similarities of R2, and one has equality if the domain is a disk. One wishes to show that the inequality is always true, with equality if and only if the domain is a disk. One can consider the above for any R", n > 2. The proposed analytic isoperi- isoperimetric inequality then becomes A.1.2) where S2 is any bounded domain in R" and 3SI its boundary, V denotes n- measure and A denotes (n - l)-measure, B" is the unit disk in R", and S" the unit sphere in R". We let u>n denote the n-dimensional volume of B" and cn_i the (n - l)-dimensional surface area of S". It is standard that . 1.3) C.-1 ^-rin/l)' -- n ' where r(jt) denotes the classical gamma function; and A.1.2) now reads as A.1.4) I/n /. 1 The Isoperimetric Problem 3 One wants to prove the inequality and to show that equality is achieved if and only if n is an n-disk. Note that for n = 2 we took in A.1.4) the square root of A.1.1). Remark 1.1.1 Throughout the book, domain will refer to a connected open set. In general, we consider the isoperimetric problem for relatively compact domains when we are working in the differential geometric setting (Chapters I, II, V-VIIIj. Therefore, the disks that realize the solution in R" are open. In Chapters III and IV, where we work in a more general setting, the isoperimetric problem is considered for compacta. In that setting the disks that realize the solution in R" are closed. Remark 1.1.2 We have restricted the isoperimetric problem to domains in R"; but if we could solve this problem, then the isoperimetric problem for open sets consisting of finitely many bounded domains would easily follow from the solution for single domains. Indeed, assume one has the inequality A.1.2) for domains inR". If n = n, u n2 u • • •, where each ft; is a relatively compact domain in R" such that (cl denotes the closure), then Minkowski's inequality implies -Aon). A.1.5) So the inequality extends to the union of domains. Note that equality implies that S2 is a domain. Remark 1.13 Note that for any domain fi in R", its volume is the n-dimensional Lebesgue measure, and if 3ft is C1 then the area of 3ft is given by the standard differential geometric surface area of a smooth hypersurface in R". However, if 9ft is not smooth, then one must propose an area functional defined on a collec- collection of domains such that the area functional will give a working definition of the area of the boundaries of the domains. Besides a number of natural properties [see the discussions in Burago and Zalgaller A988)], one requires that the new definition agree with the differential geometric one when applied to a domain with smooth boundary. Then, with this new collection of domains and definition of the area of their boundaries, one wishes to prove the isoperimeric inequality. Also, one wishes to characterize the case of equality in each of these settings.
4 Introduction Remark 1.1.4 As soon as one expands the problem to the model spaces of constant sectional curvature, that is, to spheres and hyperbolic spaces, one has no self-similarities of the Riemannian spaces in question. And if the disks on the right hand side of A.1.2) are to have radius r, then the right hand side of the inequality in A.1.2) is no longer independent of the value of r. Nonetheless, one still has the isoperimetric inequality in the sense that all domains in question with the same n-volume have the (n — l)-area of their boundaries minimized by disks. For n = 2, the analytic formulation reads as follows: If M — M2, the model space with constant curvature k, then the isoperimetric inequality becomes A.1.6) L2 >4jiA-kA2, with equality if and only if the domain in question is a disk. Of course, one can still consider the isoperimetric problem, whether or not it is to be expressed as an inequality, in the first or second formulation above. Similarly, one can extend the isoperimetric problem and associated inequali- inequalities to surfaces, or, more generally, to Riemannian manifolds. We shall consider such inequalities in Chapter V. Remark 1.1.5 Finally, one can consider a Bonnesen inequality. In R2, such an inequality is of the form L2 - > B > 0, where B is a nonnegative geometric quantity associated with the domain that vanishes if and only if the domain is a disk. L2 The Isoperimetric Inequality in the Plane For any C2 path a> : (a, 0) -* R2 in the plane, the velocity vector field of <a is given by its derivative <a', and acceleration vector field by <a". We assume that to is an immersion, that is, at never vanishes. The infinitesimal element of arc length ds is given by ds = \<o\t)\dt. Given any to e (a, 0), the arc length function of <a based at t0 is given by s(t)= [ \<a\x)\dx. 1.2 The Isoperimetric Inequality in the Plane Let denote the unit tangent vector field along <a, i : R2 -> R2 the rotation of R2 by n/2 radians, and the oriented unit normal vector field along <a. Then one defines the curvature k ofo>by — - ds ~ (indeed, since T is a unit vector field, its derivative must be perpendicular to itself). Then the formula for the curvature, relative to the original path, is given by A.2.1) One can easily show that A.2.2) ~ ds ds The equations A.2.1) and A.2.2) are referred to as the Frenet formulae. One can prove, from A.2.1), that if the curvature k is constant, then o> is an arc on a circle (if not the complete circle). 1.2.1 Uniqueness for Smooth Boundaries As a warm-up, we give the argument from classical calculus of variations. Given the area A, let D vary over relatively compact domains in the plane of area A, with C boundary, and suppose the domain Q,dQ€C2, realizes the minimal boundary length among all such domains D. We claim that ft is a disk. Proof Since Q is relatively compact in R2, there exists a simply connected domain &o such that ?2 = J20 \ {finite disjoint union of closed topological disks). We claim that since ft is a minimizer, then fto = ft. If not, we may add the topological disks to ft, which will increase the area of the domain and decrease
Introduction the length of the boundary, and therefore ft will not be a minimizer. Thus fto = ft, and is bounded by an imbedded circle. Let F : S1 ->¦ R2 e C2 be the imbedding of the boundary of ft. We always assume that the path F is oriented so that u = -N at all points of F, where u is the unit normal exterior vector field along 9 ft. One then considers a 1-parameter family F« : S1 -> R2 of imbeddings v : (—Co, fo) xS'-> R2, such that the variation function v(e, t) given by v(e, t) = T((t) = F(f) + *(e, f)i/(r). *@, t) = 0, is C1. First, 9v 9* Also which implies at Taylor's theorem implies, for the expansion which implies <=o a* at ^(t)\ ""'¦ Therefore, the area element d A in the curvilinear coordinates (t, e) is given by dA = dv dv rfe ds. 1.2 The Isoperimetric Inequality in the Plane 7 For the domain S2( determined by F( we have, for sufficiently small e, A(n«)-A(n)= f da I {<j> + o(l))ds. Jo Jr Therefore, if A(Sl() = A(Si) for all e, then f <f>ds = O. Jr Now let L(e) denote the length of I\. Since T has the shortest length, we have L'@) = 0. Therefore, because dt=f \r'\{l+€K</>+o(€))dt= [ {l Jsi Jr we have 0 = L'@)= f K<pds, I Jr Jr <pds=0 for any such variation of F. Similarly, given any <f> e C1 such that fr <f>ds = 0, there exists a variation v of F such that A(ft<) = A(ft) for all e, and L'@) = /r <4>ds. Then, by assumption, we have k</> ds = 0 V0 e C To show that this implies that k is constant, we argue as follows: Given any ir :S' -> RinC'.set 4>=\lr- I ifrds I I ds . Then fr <j>ds = 0, which implies where L denotes the length of F. Since ^ is arbitrary C, a standard argument then implies that A.2.3) --J- f Kds = 0, L Jr that is, the curvature k is constant. Then, as mentioned, A.2.1) implies that F is a circle.
8 Introduction 1.2.2 Quick Proof Using Complex Variables Theorem 1.2.1 (Isoperimetric Inequality in Rz) Let SI be a relatively com- compact domain, with boundary 3 ?2 e C1 consisting of one component. Then Proof We denote any element of the plane as the complex number z = x + iy, and the area measure as an oriented volume element; so dA =dx Ady = -dz A dz. Then 47M(?2) = if lni lnidzAdz an C ~ z / an =/ / »a JdQ - the second equality follows from the fact that the winding number of 3 ?2 about any point z e ?2 is 1; the last equality follows from Green's theorem - which implies the claim. ¦ 1.2.3 The Method of Fourier Series Lemma 1.2.1 (Wirtinger's Inequality) If f is a C\ L-periodic function on Kand i: f(t)dt=O, then [ \f'\\t)dt > ^L [ \f\\t)dt, o L. Jo with equality if and only if there exist constants a~\ anda\ such that 1.2 The Isoperimetric Inequality in the Plane 9 Proof This is an exercise from Fourier series. The function f(t) admits a Fourier expansion f(t) e2"k"L [ ake2 with ak = y[ f(t)<r2" Similarly, we have fit)'** T he2*"""- with bk = - (L f\t)e-^ik"Ldt. t^ L J The continuity of / implies bo = 0, and the hypothesis implies ao = 0. Inte- Integration by parts implies Parseval's inequality then implies lL\f?dt = JO = L^rTk2\ak\2 which implies the inequality. One has equality if and only if ak = 0 for all |*| > 1. ¦ Theorem 1.12 (Isoperimetric Inequality in R2) If SI is a relatively compact domain in R2, with C1 boundary consisting of one component, then with equality if and only if Si is a disk. Proof If necessary, we translate ?2 to guarantee x ds = 0, x = (jc Jda Let x = jt'ei + *2e2 be the vector field on R2 with base point x = (xl, x2). One now uses the 2-dimensional divergence theorem, namely, for any vector field x »-> ?(x) e R2 with support containing cl ?2, one has A.2.4)
10 Introduction where v denotes the outward unit normal vector field along 3 ?2. One can obtain the formula A.2.4) by converting the traditional Green's theorem A.2.5) ff 1^.-^-1 dA= f Pdx + Qdy, JJa \dx dy } Jia by choosing For the left hand side of (I.2.S) one has 3x dy dx1 For the right hand side of A.2.5) one has Pdx + Qdy = -?2dxl For our vector field x, we have div x = 2 on all ?2. Then the divergence theorem implies 2/t(?2) = f divx<M = / x-uds, Ja Jaa which implies 2/t(?2) = / x.uds = / Ja < / Ja 1/2 = ?l/2C?2)f/ \xfds [Jaa 1/2 - the first inequality is the vector Cauchy-Schwarz inequality, and the second inequality is the integral Cauchy-Schwarz inequality. Parametrize 3fi with respect to arc length. Note that dx ds 1.2 The lsoperimetric Inequality in the Plane ' along 3fi; so Wirtinger's inequality, applied to each coordinate function j x2(s), implies ( / \ [Jan \x\2ds 1/2 1/2 which is the desired inequality. The case of equality follows easily. 1.2.4 A Symmetry-and-Convexity Argument Theorem 1.23 Consider the isoperimetric problem for all relatively cornp^\ domains in R2 with piecewise C1 boundaries, and assume the isoperime!141 problem in R2 has at least one solution, that is, there is a domain ?2 realiz'^ the minimum of the isoperimetric functional D L2(dD) A(D) " Then the open 2-disk also minimizes the isoperimetric functional. Remark 1.2.1 Thus, we are only proving a weak result, namely, the analy1"' isoperimetric inequality A.1.1) is valid if the isoperimetric functional has minimizer. Although the arguments from complex variables and from Fourier given above are also valid when the boundaries are only piecewise C\ here proof requires that the variational problem be defined over the extended of domains. ^ * Proof Let ft be a minimizer. Pick a line II that divides ?2 into two open se fi i, ^2 of equal area. Then z.Cfi n ci no = z.C« n ci n2). If not, and L(dSi n cl Si\) < Z.Cfi D cl n2). then one could replace 0 by P ' consisting of S2i union its reflection in n, which would imply that ?2 is no* minimizer, which is a contradiction. So we may replace ?2 by a domain ?2* tn^ is symmetric with respect to the line n.
12 Introduction Figure 1.2.1: The line through the origin intersecting dil. By carrying out this argument successively, one may assume that there exists a coordinate system in R2 such that ?2 is symmetric with respect to the two coordinate axes. This implies that every line through the origin divides ?2 into two open sets of equal area and, therefore (by the above argument), equal bounding length. Next, we may assume cl ?2 is convex. If not, we may replace it by its convex hull, which increases the area and decreases the bounding length. Since the origin is in ?2, we may think of d ?2 as the image of a piecwise C1 map defined on the unit circle §' in R2. Let w0 € d ?2 have a tangent line, not equal to the line from the origin through w0. (See Figure 1.2.1.) We claim that the line from the origin to w0 must intersect 3 ?2 perpendicular to the tangent line at w0. If not, it will be possible to carry out the construction ?2 i-+ ?2* described in the first paragraph above so that cl ?2* is nonconvex, which is a contradiction. Therefore, ?2 is convex with every point of differentiablity of 3 ?2 having its tangent line orthogonal to the position vector. Let 6 denote the local coordinate on S1, x the position vector on 3 ?2, and n the position vector on S1. Then x= rn at all points of differentiability of 3 ?2, which implies 3x 3r 3n Because n is perpendicular to the tangent space, we conclude that r is differen- tiable, with 3r/30 = 0, at all points of differentiability of 3?2, which implies r is locally constant on the C1 arcs of 3?2. Thus 3 ?2 consists of a finite number of 1.3 Preliminaries 13 circular arcs connected by segments of lines that (the lines, not the segments) pass through the origin. The convexity of ?2 implies 3 ?2 is a circle. ¦ Remark IJ*2 In higher dimensions one has to be careful about the convexity. For example, it is not true that passing to the convex hull of a domain decreases its isoperimetric quotient. Pick a standard imbedded solid torus in R\ where the circle in the jry-plane has radius 1, and the rotated circle perpendicular to the jy-plane has radius € <? 1. Then its isoperimetric quotient is asymptotic to const.€"/3 as € j 0. On the other hand, the convex hull has isoperimetric quotient asymptotic to const.€~2/3 as € i 0. L3 Preliminaries We describe here most of the background required for the rest of these notes. The Riemannian geometric background is discussed in Chapter V. Notation Let X be any space, A a subset of X. We let IA denote the indicator (or characteristic) function of A, that is, Ia(jc) = 1 for x e A, and Ia(jc) = 0 for x e X \ A. The cardinality of A is denoted by card A. Notation In any topological space X we let supp / denote the support of a function / on X. We denote continuous functions of compact support by CC(X). When the target is some vector space V, we write CC(X; V). Notation Whenever discussing a measure space (X, 971, dp), we let || || p denote the associated Lp norm. (In general, the a -algebra 971 is fixed, so we rarely mention it.) If there is doubt as to which space or measure is intended, then we also indicate our choice in the subscript: || ||p,x or || ||p,dM, as required. In general we write || || = || ||2. We also indicate the L2 inner product by ( , ). Unless otherwise indicated, these are spaces of real-valued functions. Proposition 1.3.1 If the sequence (/)) converges to f in Lp(ji) for some p e [1, +00], then (fj)has a subsequence that converges pointwise to f a.e.-[d(i]. Proposition 1.3.2 Let Xbea complete, separable metric space, Y a Hausdorff space, f : X -*¦ Y a continuous map, and n a measure on Y such that every closed subset of Y is ^-measurable. Then f(B) is ^-measurable for every Borel subset B ofX.
14 Introduction Figure 1.3.1: Equation A.3.1) when p = \, dfi(x) = dx. Proposition 133 (Cavalieri's Principle) Let v be a measure on Borel sets in [0, oo), 4> its indefinite integral, given by 4>(t) = v([0, f)) < +00 V t > 0, (fi, E, fi) a measure space, and f a nonnegative ? -measurable function on S2. Then f 4>(f(x))dfi(x)= [ /*({/ > t))dv(t), Ja Jo or equivalently, / dfi(x) / dv(t)= / dv(t) / l{f>,)dpL. Ja Jo Jo Ja In particular, ifdv(t) = ptp~l dt, p > 0, then we have (see Figure 1.3.1) f p f°° p-i f Ja Jo Jq Thus, if p = 1 and fi = Sx then /(*)= [ I{f>t{x)dt. Jo Notation Let (X, d) be a metric space. For any subset A in X, let int A denote the interior of A, ext A = int(X \ A) denote the exterior of A, cl A or the traditional A denote the closure of A, 1.3 Preliminaries 15 dA denote the boundary of A, diam A denote the diameter of A, [A]( = {x € X : d(x. A) < €} denote the €-thickening of A. For any x e X and r > 0, let fl(.x; r) denote the open metric disk and D(jc; r) the closed metric disk in X, with center x and radius r. We denote the metric sphere bounding B(x; r) by S(x; r). We consider n-dimensional Euclidean space R", n > 1, where we denote arbitrary points by x = (xl,...,x"), y = (>>' ,y"),... e R". The inner product is denoted by x, y i-+ x-y. The length of an element x is given by |jc| = y/x-x, and the distance between x and y is given by d(x, y) = \x — y\. In R", we write B"(jt;r) for B(x;r) and D"(jt;r) for D(;t;r). We omit the superscript n when the context is sufficiently clear. We also write S"~'(*;/¦) for S(x\ r), and drop the superscript when no problem will caused by doing so. When x is the origin o we write and = B"(o;r), B" = B"(o; 1), = W(o;r), = ny(o; I), = S"-'(o; 1). Definition Let (X, 971) be a measure space. Then an R*-valued measure A on (X, 9Tt) is a map A : 9JI -*¦ R* such that 00 A(E) = ^2 A(?y) (which converges) for all partitions {Ej} of ?, for all E e SER. (A partition {Ej} of ? is a countable family of subsets in 971 that are pairwise disjoint, and whose union is E.) If A is an Rl-valued measure on X, then its total variation measure, | A|, is defined by |A|(E) = sup ?] |A(E,)|, where the supremum is taken over all partitions of E. One knows that |X\(X) < +00. Definition An Revalued measure A is said to be absolutely continuous with respect to a positive measure fi on (X, 9JI), A «: fi, if A(E) = 0 whenever ME) = 0.
16 Introduction Proposition 1.3.4 (Radon-Nikodym Theorem) IfX<g.ti, then there exists an R*-valuedfunction * in Ll(fi), that is, fx |*| dfi < +00, such that -L for all measurable E. Also, there exists a measurable map v : X -*¦ S* such that dX = i/d\X\. Finally, let fibea positive measure, $ : X -> R* in Z-'(/x), and set X(E) = f Je Then •h Definition Let X be a locally compact Hausdorff topological space, and fi a positive Borel measure on X, that is, its a-algebra contains 33, all Borel subsets of X. We say that fi is locally finite if fi(K) is finite for all compact K; a measurable E is outer regular if = inf{/x(V): E C V, V open}; a measurable E is inner regular if fi(E) = sup{fi(K): K c ?, K compact}; the measure m is regular if every measurable ? is both inner and outer reg- regular. The measure fi is a ftufon measure if it is locally finite and regular. An Revalued measure A is an Revalued Radon measure if its total variation mea- measure I A| is a Radon measure. Proposition L3.5 (Riesz Representation Theorem) Let X be a locally com- compact Hausdorff space, and&a bounded linear functional on CC(X; R*) (relative to the sup norm). Then there exists a unique R* -valued Radon measure XonX such that *(*) = f Jx for all $ : X -> Rk in CC(X;R*). Moreover, 11*11 = |A|(X). 1.3 Preliminaries 17 In R", we denote the standardLebesgue measure by dvn(x). When considering a domain R in R" with C1 boundary 3fi, we denote the Lebesgue measure on Q by dV, and the standard surface area measure on d?2 by dA. For the special case of B" we denote the surface area measure on S"~' by d^in_x. Recall that we use the notation = / d\n, cn_i = / dnn_Y. When n = 2, we replace <M by dL, and d V by rfyl. When discussing an arbitrary fc-dimensional submanifold of an n -dimensional Riemannian manifold, we use dVk to denote the fc-dimensional Riemannian measure on the submanifold. When discussing changes of coordinates on R", we refer to the new coordi- coordinates as Euclidean coordinates if they are obtained from the standard coordinate system by a Euclidean transformation. Proposition 1.3.6 (Lebesgue Density Theorem) For f eL\W)we have lim -L. f (/(,) - /(jto)| d\n(x) = 0 for almost all x<) € R". In particular, for every measurable E in R" with positive measure, we have va(?nB(jc;r)) hm = 1 for almost all x e E. Notation A multi-index a is given by a = (ot\,.. .,an), where aJy j = 1,..., n, is a nonnegative integer, |a| = 01 H \-an. For the multi-index a, D° denotes the differentiation Definition We define a mollifier on R" to be a function j : R" -> [0, 00) € C~(B") satisfying We define, for all € in @, I), Mx) = €-"j(x/€);
18 Introduction and for any locally l} function / on R", that is, / e /-'„,., we define a * /)(*) = f My - x)f(y)d\n(y) = f My)fix + y)dvn(y). Proposition L3.7 (a) Ifu 6 L^., then supp/( * u = [suppu]<. (b) Ifu 6 C°, then jt * u -*¦ u uniformly on compact subsets o/R" as€ I 0. (c) Ifu e L^, then j(*u e C°°, and (Da{j( * u)Xx) = (-1)"" f(Dajf)(y - x)u(y)dvn(y) for any multi-index a. Moreover, ifu e Ck, then for all \a\ < k one has DaU< * «) = U * (Dau), by integration by parts. (d) Ifu e L", then (i) j(*ue L", (ii) ||/€ *n||p < |u||p for all p e [1, oo], and (hi) ||/( * u - u\\p -*¦ 0 for all p e [1, oo), as € ± 0. Remark L3.1 If j(x) = j(-x) for all x e R", then A.3.2) j{Jt * f)g dvn = j f(j< * g) dvn for all /, g e /-'(R"). We shall henceforth assume the symmetry of all molli- fiers, so A.3.2) will always apply. Notation Let F : ?2 -> R* e C, where ?2 is open in R". For each x € ?2, we denote the associated Jacobian linear transformation of F by Jf(x). Definition Let F : ?2 -> R* € C1, where ?2 is open in R\ A point jc e ?2 is called a critical point ofF if its associated Jacobian linear transformation Jf(x) has rank < k. A point y e R* is a critical value if there exists at least one point in its preimage which is a critical point. The complement in R* of all critical values of F consists of the regular values ofF. (This includes points not in the image ofF.) Proposition 1.3.8 (Sard's Theorem) Let F : ?2 -*¦ R* e Cl, where ?2 is open in R", I > max{n — k, 0}. Let A denote the critical points of F. Then n(F(A)) = 0. 1.3 Preliminaries 19 Remark 1-3.2 The cases where we invoke Sard's theorem usually have k = n, in which case t > 1, or k = 1, in which case I > n. Definition Consider the differential operator — defined on some domain D in R". We say that L is elliptic at x if there exists a positive number fj.(x) such that J2 j.k=\ for all p = (p\,..., pn) e R". We say L is elliptic on D if L is elliptic at every point of D. We say L is uniformly elliptic on D if L is elliptic on D and there is a positive constant /xo such that fi(x) > fio for all x e D. Proposition 1-3.9 (Strong Maximum Principle) Let L be uniformly elliptic on the domain D in R", with uniformly bounded coefficients ajk(x), bi(x), and assume u : D -*¦ Re C2 satisfies the differential inequality Lu>0 on all of D. Ifu attains a maximum at a point ofD, then u is constant on D. Furthermore, assume (i) 3D e C1, (ii) u : D -> R e C2(D) n C°(D), (iii) u < M on all of D, (iv) u = M at some point w e 3D, and (v) the ex- exterior normal derivative du/dv is defined at w. Then du A — > 0 3 at w, unless u = M on all of D. We remind the reader that du/di/ at w is defined to be the limit of (grad u)(x)-v, x e D, as x approaches w along the line through w with direction v. 1.3.1 Lipschitz Functions Definition A Lipschitz function on the metric space (X, d) is a function u on X for which there exists a nonnegative number K such that A.3.3) l«(zi) - « , zi) V z,, X.
20 Introduction The Lipschitz constant Lip u is the smallest positive K for which A.3.3) is valid forallzi,Z2 € X. Theorem U.I Let (X,d) be a metric space, A C X.andf : A -> RLipschitz. Then there exists g : X -*¦ R Lipschitz such that g\A — f. Proof Let L = Lip /, the Lipschitz constant of /, and consider g(x) = inf {/(>>) + Ld(x,y)). yeA For all x e X and y, z, e A we have fit) - f(y) < Ld(z, y) < L{d(x, y) + d(x, z)h which implies f(z) - Ld(x, z) <f(y) + Ld(x,y), which implies f(z) - Ld(x, z) < g(x) < f(y) + Ld(x, y), which implies g\A = f. To show that g is Lipschitz, note g(xi) - g(x2) = sup inf f(yi) + Ld(xx, yx) - f(j2) - Ld(x2, >^) < sup L{d(xi, y2) - d(x2, >^)} neA <Ld(xux2). m Theorem 132 (Rademacher Theorem) // / : R™ -* R" is Lipschitz. then f is differentiable a.e.-[d\m]. Proof We may assume n = 1. We may also assume that we know the theorem for m = 1 (since Lipschitz implies / is an absolutely continuous function of one real variable). Fix | e S", and let (D^f)(x) denote the directional derivative of / at x e R" in the direction ? (should it exist). Also, let B( = {x e R" : (D{/)(*) does not exist}. Then \\(B( n {x +1$ : t e R}) = 0 for all x (we know the theorem forn = 1). Therefore (by Fubini's theorem), \n(B() = 0 for every ? e S". In particular, 1.3 Preliminaries 21 by successively picking ? = e, eB (the natural basis of R"), the gradient vector field grad / exists almost everywhere. Next we wish to show that, for each ?, A-3.4) @*/)« = Hgrad/)« for almost every x. Indeed, given any <j> € Cf (R") we have (by change of variables) '-<fi(x)dvn(x) -I t J t which implies (using dominated convergence) ¦f(x)dvn(x). 'iZ7)f which implies A.3.4). Finally, let {?,} be a countable dense set in S", Aa = {x: grad/, D^ f exist &D^f = ?o-grad /}, a=l Then vB(R"\/l) = 0. We now show that / is differentiable on A. For x € A, y = x + t% e R\ we have - fix) - (y - *).(grad /X*) = /(* +1$) - f(x) - f?-(grad f)(x) = /(jr + tt-) - f(x + t$a) + f(x + r&) which implies -$a\\y-x\+ o(\y - x\). which implies the theorem.
22 Introduction Theorem 133 Let u be a Upschitz function on R" with compact support. Subject u to mollification by j( as above. Then \\j( *u- u\\p -*¦ 0, Hgrady, * u - gradi«||p -> 0 for all p e [1, oo) as € I 0. Proof Since u is Lipschitz, the restriction of u to any line is absolutely con- continuous; and thus all partial derivatives exist almost everywhere on R", and are bounded with compact support. Thus, each partial derivative is an element of Lp for all p € [1, oo]. Also, the absolute continuity along lines and the com- compactness of the support of u imply the fundamental theorem of calculus for each partial derivative, which implies Jr. dxk for ft = 1..... n. Then we may use the argument of Proposition I.3.7(c) to show O > One easily obtains the conclusion of the theorem. ¦ Example L3.1 Let K be & convex subset of R" with nonempty interior. (See §1.3.3 below.) If K / R", then K is contained in a half-space H. Let n = 3//, and p : R" -> n the orthogonal projection. Introduce Euclidean coordinates in R" so that n = {*" = 0}. Define, for w = (jc1 jr") e p(K), /(w) = inf{y: (w,y)eK). Then the graph of jc" = /(>v) is the "lower" part of the bounding hypersurface of K, and / is a convex function, that is, f(ku + A - k)v) < kf(u) + A - k)f(v) V u, v e p{K), k € tO, 1]. Given u, v e int p(K), let luv denote the line in n through u, v, and set Then one checks that F(t) - F(x) t - r 1.3 Preliminaries 23 is increasing with respect to both t and r, which implies the right and left hand derivatives of /, F+ and F'_ respectively, exist everywhere, and t — z for all a < A < r < t < B < b, which implies F is locally Lipschitz on (a, b). By Rademacher's theorem, the boundary of AT is a.e.-[Il, d\n-\] differentiable. Remark 133 We recall for the reader the basic definition of Sobolev spaces. Let ?2 be open in R". (The definitions that follow are also the ones that are used when ?2 denotes an open subset of a Riemannian manifold.) By the Meyers- Serrin theorem one has two equivalent characterizations of the Sobolev space IV'•'(«). First one considers the collection H of C°° functions / on ?2 for which both /, grad / € Ll. Then one endows H with the pre-Banach norm ll/lli.i = llgrad/||, and completes //.with respect to the norm || |h,i, to the Banach space W(n) The second characterization goes as follows: We say that / e Z-'(fi) has a weak Ll -derivative if there exists an L'(?2)- vector field ? such that / fdi\ijd\n = - I Jn Ja for all compactly supported C00 vector fields r\ on ?2. The vector field $, should it exist, must be unique, and we denote it by ? = Grad /. Let Si be the Banach space (it is indeed a Banach space) consisting of functions f e Ll possessing weak L'-derivatives, with the norm || ||g given by ll/lla = ll/lli + IIGrad/||,, / e fj. Then, by the Meyers-Serrin theorem, Sj = W'^fi). Clearly, all L'(?2) Lipschitz functions with L'(S2)-derivatives are elements ofW'UO) Notation We let fjc(n) denote the completion of Cf (O) in fj(fi). 1.3.2 Co-area Formula for Smooth Mappings In this section, we use the method of moving frames. One can find treatments of the method in the standard texts. Let Mm, N" be C, r > 1, Riemannian manifolds, wj>«,and4>:Af->A^a C1 map from M to N. We want to give an effective calculation of the volume
24 Introduction disortion of the map, namely, of = v/det ¦,oD»,)^l, where 4>, denotes the associated Jacobian linear map of tangent spaces of M to those of N, and ()** denotes the adjoint map. (When M and AT are Euclidean spaces, we use the notation 7<& for <t>,.) Let {eA), A = 1 wj, be an orthonormal moving frame on M with dual co-frame {<oA}, and {?,}, y = 1,...,«, an orthonormal moving frame on N with dual co-frame [&J). It is well known that the local volume forms on M and N are given by a>x A--- AuT andfl1 A ••• Ad", respectively. There exist functions o> a on M such that where ** denotes the pullback of differential forms on N to M, associated with the map *. If at some x e M we have dimker 4>,|x > m - n (that is, * is a critical point of 4>), then If at some x e M we have dim ker <1>,|X = m - n (that is, x is a regular point of $), then * is a submersion on a neighborhood of x, and we may pick en+i em to be tangent to the fibers of the local fibration about x associated with the submersion, which implies e\,..., en is orthogonal to the fibration. We therefore have 7 = a—n \ m, which implies which implies <fr'@' a ¦•¦ A 6") = (det crJk)(ol A •¦• A(o". One can easily check that det (or-'*) = det *,|(ker Q,I. Therefore, since 0, rank<J>. < n, |det(«I>,|(ker<I>,)-L)|, rank*, = n 1.3 Preliminaries 25 (as one verifes rather easily), we have A.3.5) J*@l A-ACif = ±<t>*(d1 A-- A6n) A (On+1 A--A(Om, depending on whether deUo--'*) is positive or negative, which implies Theorem 1.3.4 (Co-area Formula for Smooth Mappings) Let M, N be C Riemannian manifolds, with m = dim M > dim N = n, r > m - n, and let <t>: M -»¦ N e C. Then for any measurable function f : M -> R that is ev- everywhere nonnegative or is in Ll(M), one has A-3.6) [ fJ*dVn= [ dVn(y)[ (f\<t>-l[y])dVm.n Proof We have, from A-3.5), fJ*dVm [ M dVn {y)f V\*-x 7*-'[y] [y])dVm-,,, where regval * denotes the set of regular values of $ in N. Since the map * is Cr, the critical values of * constitute, at most, a set of measure 0 in AT (by Sard's theorem - it is also valid for manifolds), which implies A.3.6). ¦ Corollary 13.1 (Co-area Formula for Smooth Functions) Let Mm be a Cm Riemannian manifold, and let <P : M -*¦ R e Cm. Then for any measur- measurable function f : M -*¦ R that is everywhere nonnegative or is in Ll(M), one has A.3.7) I f\gnd*\dV = f dvi(y) [ JM 7R 74>-l[y] where grad * denotes the Riemannian gradient vector field of Q on M {see §VIl Corollary 132 Let A/* be a hypersurface in R* given by the graph of a C' function 4>: G -> R, where G is open in R*; so M is given by xk=4>(*1 **-'), (xl i»-')?C. Then the surface area element on M, dA, is given by where gradt_, denotes the gradient of functions on R*.
26 Introduction Proof Let H be a hyperplane in R*, with normal vector i/; R* 'the hyperplane {xk = 0} in R*, with normal vector ?; and p: H -> R*~' the projection. Then A.3.9) JP = l«"?l- Then apply A.3.6) with / = 1/ Jp, and calculate. ¦ Corollary 133 //fiCCR" w<J domain with C' boundary, v the exterior unit normal vector field along dSi, and, for a given ? e S", p^ : dSl -> ?x is rte projection, then A.3.10) / |i/w Jin Proof Apply A.3.6), with / = 1. ¦ Corollary 13.4 If Si CC R" is a domain with C1 boundary, and for every t e S", p^ : 3 ft -> ?L denotes the projection, then A.3.11) 2u>n-l Proof Integrate A.3.10) over ? e S". Then / in an = 2ACO)c.-2 / cose sin"Odd Jo n - 1 which implies the corollary. Corollary 1.3.5 (Cauchy's Formula) If Si is convex, 30 e Cl. then 1.3 Preliminaries Proof If Si is convex, then card (dSi n p4"'[>>]) = 2 for all >> € P{CS2) = Pt(Si), ? e S", which implies the claim. 27 Corollary 13.6 If Si is convex, dSi e C1, and Qo is open containing Si. C\ then < ACfio)- 1.3.3 Some Geometric Preliminaries Theorem L3.5 Suppose K is compact in R", with the property that, for any ? e S"~', K is symmetric with respect to some hyperplane perpendicular to ?. Then the boundary ofK consists of a union of concentric (n — l)-spheres. Proof Pick new Euclidean coordinates in R" so that K is symmetric with respect to reflection in each of the coordinate planes. Then K is symmetric with respect to the origin. Since any direction has an orthogonal hyperplane of symmetry for K, and since K is already symmetric with respect to the origin, then K must be symmetric with respect to any plane through the origin. Since any half-line from the origin to 3 K can be taken to any other by a reflection in a hyperplane through the origin, each connected component ofdK must be an (n — l)-sphere centered at the origin. ¦ We recall just a few notions about convex sets in R". Definition A set A in R" is convex ifx.yeA implies that Ajc + A - k)y e A for all A. e @,1), that is, for any x and y in A the closed line segment [x, y] in R" joining them is contained in A. Definition A convex linear combination of elements xi,..., xt e R" is the linear combination where the coefficients satisfy * = 1, A.;>OVy. One checks that if A is convex then any convex linear combination of elements of A is a point in A.
28 Introduction Definition Given AcR", define the convex hull of A, conv A, to be the smallest convex set containing A. Proposition 13.10 Given A, then conv A is given by the collection of all finite convex linear combinations of elements of A. If A is compact then conv A is compact. If A is bounded, then conv A is bounded and diam conv A = diam A. IfSis convex then S is convex. IfSis convex compact then S = conv 35. IfCis convex with nonempty interior, then int C is homeomorphic to R". Definition Let A, B be subsets of R", and H a hyperplane. We say A and B are separated by H if A and B lie in different closed half-spaces deter- determined by H. If neither A nor B intersects H, we say H strictly separates A and B. Definition Let A be a subset of R", H a hyperplane, x € A. We say H is a supporting hyperplane of A at x if x € H and H separates [x) and A. Note that x g3A. Proposition 13.11 Let A be closed, with nonempty interior. Then A is convex if and only if every x e 3 A has a supporting hyperplane. Furthermore, if A is closed convex, then A is the intersection of all closed half-spaces containing A. L4 Bibliographic Notes §1.1 For surveys of the isoperimetric inequality in Euclidean space, replete with his- historical remarks and references, see Blaschke A9S6), Hadwiger A9S7), Osserman A978), Talenti A993), and the more recent book by Burago and Zalgaller A988). For further remarks, including some discussion of experimental proofs of the isoperimetric inequal- inequality [one of which goes back to Courant and Robbins A941)], see Dierkes, Hildebrandt, Kurster, and Wohlrab A992, pp. 42(M23). On the issue of the proof of the isoperimetric inequality for general boundaries, the following paragraph of Osserman A978, p. 1188) is worth repeating: First, a general remark. If we start with a relatively smooth boundary, adding "wiggles" to it will have very little effect on the volume enclosed, but will greatly increase the surface area. Thus, one has the somewhat ironic situation that the more irregular the boundary, the stronger will be the isoperimetric inequality, but the harder it is to prove. The fact is, the isoperimetric inequality holds in the greatest generality imaginable, but one needs suitable definitions even to state it. 1.4 Bibliographic Notes 29 The first unified solution of the isoperimetric problem in the model spaces of constant sectional curvature (that is, Euclidean, spherical, and hyperbolic spaces) was given by Schmidt A948, 1949). In Chavel A994) we gave separate proofs in each of the cases, to highlight a variety of methods. The Euclidean argument given there will be given in Theorem II.2.2 below. The isoperimetric argument for the sphere there is a new Ricmannian theoretic argument by Gromov A986), and the hyperbolic space argu- argument follows the symmetrization argument in Figiel, Lindenstrauss, and Milman A977). Different symmetrization arguments will be developed here later. The study of isoperimetric inequalities on surfaces is alive and well. See the discus- discussions in Osserman A978,1979). More recent results can be found (just to give a sample) in Adams and Morgan A999), Benjamini and Cao A996), Howards, Hutchings, and Morgan A999), and Topping A997,1999). Extensive discussion of Bonnessen inequali- inequalities can be found in Osserman A979). §1.2 The complex variables proof follows Topping A997) (the proof was inspired by F. Heiein). The Fourier series argument was first carried out by Hurwitz A901). The name "Wirtinger's inequality" is apparently a misnomer. See the historical remarks in the previous surveys. Theorem 1.2.3 is from Howards, Hutchings, and Morgan A999). Blaschke A956, p. 1), in discussing the idea, refers back to Steiner A881, p. 193ff). The example of Remark 1.2.2 was supplied by an anonymous reader, to whom I extend my thanks. Another line of reasoning was initiated by M. Gage A984), wherein he proved that if a convex curve in the Euclidean plane is deformed along its inward normal at a rate proportional to its curvature at that point, then the isoperimetric ratio L2/A decreases to the limit 4jt, and the convex curves become asymptotically circular - provided that the shrinking curves do not develop singularities such as corners. The subject has developed quite extensively since then (including the discussion whether the singularities exist), both for surfaces and for higher dimensions, and one can find a partially updated bibli- bibliography in Gage A991). Recent results, with applications to geometric inequalities, can be found in Topping A998). Also, see Chou and Zho (scheduled to appear in 2001). §13 For background in analysis, we generally follow Rudin A966). More recent books that are helpful here are Licb and Loss A996) and Ziemer A989). Proposition 13.2 is from Federer A969, p. 69 ff.) For treatments of Sard's theorem, see Hirsch A976) and Narasimhan A968). For the strong maximum principle, see Gilbarg and Trudinger A977) and Protter and Weinberger A984). For Sobolev spaces, see Adams A975) and Gilbarg and Trudinger A977). The general area and co-area formulae of geometric measure theory are discussed in §3.2 of Federer's treatise A969). It seemed helpful to provide the general co-area formula m the C1 category in order to see how much it contains, beyond the usual case of real- valued functions. We present the area formula for Lipschitz mappings of R -*¦ R", « < n, in §IV.2 below. The method of moving frames is discussed in Chavel A994). One can find introductory material on convexity (including the Hahn-Banach theo- *m and the Cauchy formula for areas of boundaries) in Berger's two-volume work on Seometry A987, Chapters 11,12).
II Differential Geometric Methods In this chapter we present the differential geometric arguments. By their very nature, they presuppose a certain smoothness of the boundary. Our method will be to start with assuming that the boundary of the domain in which we are interested is C2, and then weaken the assumption to just C1. We extend the isoperimetric inequality to more general domains in the next chapter. We use classical arguments to show that if a domain provides a solution to the C2 isoperimetric problem, then the domain is a disk. Then we strengthen the result a bit, to show that if the domain is just an extremal of the isoperimetric functional, then it must be a disk (see the definitions below). More specifically, we first introduce the standard local calculations of the differential geometry of hypersurfaces in Euclidean space, and the first variations of volume and area of domains and their boundaries under a 1-parameter family of diffeomorphisms. Then we give Almgren's characterization of the solution to the isoperimetric problem, followed by Alexandrov's characterization of an extremal for the isoperimetric functional. Thus, the C2 theory yields a characterization of the minimum (should it exist) in its category. To prove the isoperimetric inequality for domains with C1 boundary, we present a proof based on Stokes's theorem, following M. Gromov. The only drawback of the method is that, to characterize equality, one requires that the domain be also assumed to be convex. We solve this latter problem with Steiner symmetrization in the next chapter. II.l The C2 Uniqueness Theory II.l.l The Variation of Volume and Area We first review the local calculations of classical differential geometry per- pertaining to an (n — l)-dimensional regular hypersurface F in R". Although we present the definitions of geometric quantities (such as the curvatures, 30 II.l The C2 Uniqueness Theory 31 Riemannian divergence, etc.) relative to specific choices of local coordinates, standard arguments establish the invariance of these geometric quantities rela- relative to changes of coordinates. Assume F is given locally by the C1 mapping x : G -*¦ R" of everywhere maximal rank, where G is an open subset of R". Sox = x(«); and the vectors dx/du1 dx/du"'1 are linearly independent and span the tangent space to F at every x(u). We let n(«) denote a choice of continuous normal unit vector field along F. When con- considering a hypersurface F which is the boundary of a domain in R", the choice for n will always be the exterior normal unit vector field - unless otherwise indicated. The Riemannian metric of F (that is, the first fundamental form) is given locally by the positive definite matrix G(u), where ~(gjk), gjk - a||/gjf. (II.l.l) We also use the notation G~l = (gjk), g = det G. The associated surface area on F is given locally by = 1 n-1. Definition Let Q be a bounded domain in R", with C* boundary, k > 1. We say that Q is a solution to the C* isoperimetric problem if, for any domain D with C* boundary and volume equal to that of Q, we have A(dD) > A(dQ). Let Q be a bounded domain in R", with C* boundary. We say that ft is a Ck extremal of the isoperimetric functional if, for any 1-parameter family of Ck diffeomorphisms *, : R" -> R" satisfying V(*,(?2)) = V(?2) for all t, we have d_ dt = 0. r=O Assume that F is a C2 hypersurface in R" (so the Riemannian metric is C1)- For any tangent vector field dui
32 Differential Geometric Methods along F, we have its Riemannian divergence (see §VII.2.1) given by with an attendent intrinsic Riemannian divergence theorem for (n — 1)- dimensional domains in F with Cl (n - 2)-dimensional boundaries; namely, given an (n — l)-domain A CC F with C1 boundary 3A, and unit normal ex- exterior (F-tangent) vector field u along 3 A, then [ dA The second fundamental form of F in R" is given locally by 32x (II. 1.3) Then that is, (II. 1.4) bjk = •n, j, k = 1, ...,n - 1. 32x _JL( dx\ dn dx - dn dx Jk ~ dujduk'n~ duJ \n'duk) ~dul'~M ~ ~3^*3^*' 3x If C — (Ljk) denotes the matrix of the G-self-adjoint linear transformation associated with B, then C = G~lB. The mean curvature HofV in W is the trace of C, that is, the trace of B relative to G, given by and the Gauss-Kronecker curvature K of V in R" is the determinant of C, that is, the determinant of B relative to G, given by K = det G"'B. We may think of n as a map that associates with every point x g F its normal vector n g S" - commonly referred to as the Gauss map E : F -*¦ Sn~'. The Jacobian linear transformation 0* of the Gauss map is then determined by dx dn III The C2 Uniqueness Theory 33 Then (II. 1.4) implies that the matrix of E, is given by -C, which we write as (II. 1.5) <d, = -G~lB. Example IL1.1 If F is a relatively open subset of S"~'C). the (n - l>sphere of radius r, then for the normal unit vector field along F exterior to B"(r), we have 1 n = -x, r which implies [by (II. 1.4)] therefore, H = — -1) K = rn-\ Theorem IL1.1 LetQ.be a bounded domain in R", with C2 boundary F. Given anyC2 time-dependent vector field X :R"xR->R" onR", let *, : R" -*¦ R" denote the 1 -parameter flow determined by X, that is, <t>, and X are related by d_{ dt Set Then (II. 1.6) = / rj ndA, where n is chosen to be the exterior unit normal vector field along F, and (II. 1.7) -A(*,(D) at = f {divr if - Hij-n}dA = - I Hrj-n dA. where if denotes the tangential component ofi\. Proof First, = ff det U,(x)dyn(x),
34 Differential Geometric Methods where 7<», denotes the Jacobian matrix of *,. Then The formula for differentiating determinants states that for any differentiable matrix function t k> A(t), where A(t) is nonsingular, one has — det A = det A ¦ tr (A~l -- 1. Therefore, II ^ V(*,(«)) = ff (det 7*,) • tr Now so for/ = 0 we have, since *o = id, 3*,* = \ n; dxA the Kronecker delta. Furthermore, d . n c — x & f=0 3, 3** 3/ which implies, for r = 0, r=0 which implies the first equality of (II. 1.6). The second equality follows from the divergence theorem in R\ Assume the surface F is given locally by x = x(u), and set For each fixed t we denote the Riemannian metric tensor on <D,(F) by ay 3y . , , which implies La(*,(D)= f ^¦ at Jr ot H.I The C2 Uniqueness Theory For the derivative of ^det (*_,-*), set H = (h)k), H~[ = (hjk)\ then -VdetW= \ dt 2 w}~' det H ¦ (trW— 35 that is, (II. 1.8) 3 m 3/ Set f = 0; then n(u) = (dy/dt)(u, 0). Write v^ * 3x along F. Then (II. 1.8) implies, for t = 0, -%/detW 1=0 To calculate in more detail, one has n-l on x~~v I ( ox /OX i oq> on —- = 2. 1 ~^—i + *? —:—7 I "•—^n + 0—^. which implies, since n is perpendicular to 3x/3«* for all k = 1,..., n — 1, a. 3x 3») ;.* 3x Ei* 91l v^ ,t / 9x 32x ^^ lk 3x 3n gjKgki—-: + y g} n—• + y <t>gJ—•— ... duJ fr-J 3«* du(duJ *rr? duk duJ ;.*.< j.k.l j.k ,t 3x 3n = divr rf - <pH.
36 Therefore, Differential Geometric Methods d_ di = I (divr nT — Hn*n} dA. Jr Then the intrinsic Riemannian divergence theorem (II. 1.2) implies (since F is closed, with no boundary) the second equality (II. 1.7). ¦ 11.1.2 Uniqueness of a Solution to the C2 Isoperimetric Problem Let SI be a bounded domain in R", with C2 boundary F. Recall that the nor- normal unit vector along F is outward. Thus, if SI were convex, then the second fundamental form B would be negative semidefinite. Theorem II.1.2 Assume that SI is a solution to the C2 isoperimetric problem, with volume of SI equal that of the unit n-disk in R". Then the mean curvature H of T satisfies -H <n-\ on all of V. Proof Consider the isoperimetric functional ACD) where D varies over bounded domains in R" possessing C2 boundary. Let 4>, denote a 1-parameter family of diffeomorphisms of R", with associated time-dependent vector field X = X(x, t) (as in the proof of Theorem II. 1.1), ?(*) = X(x, 0), 17 = ?|F. Then direct calculation implies, by (II.1.6), (II. 1.7), and the fact that Si is an extremal of the isoperimetric functional, °=JtH va which implies ( •¦ ) */(?*)) 1 1=0 f Hrj-ndA f rj-ndA n-1 n A(F) A(F) V(Sl) (both integrals are over F). But since V(Sl) = u>n, and ?2 is a solution of the //./ 77ie C2 Uniqueness Theory isoperimetric problem, then A(F) < c,_i, which implies by A.1.3), 37 (H.1.10) / Hri-ndA J rj-ndA (both integrals are over F). For any u>o e F, let 0 be a nonnegative C°° function compactly supported on a neighborhood of wq in R", and choose X(x, t) = 0(jc)na)o (so this particular vector field is time-independent). Then by picking <p with sufficiently small support about wo one obtains from (II. 1.10) that -H(wq) < n — 1, which is the claim of the theorem. ¦ Remark II.1.1 The argument of the last paragraph, applied to (II. 1.9), implies that H = const, namely H = (n - l)A(F)/n V(?2). Theorem 11.13 Assume Q is a bounded domain in R", with C2 boundary F, and n the exterior normal unit vector field along F. Assume the mean curvature HofY satisfies -H <n-\ along all of Y. Then A(F) > c,,.!, with equality if and only if SI is an n-disk. Proof Start with Q = Q U F, and let C denote the convex hull of ?2, E = 3C the boundary of C. Letiu g E\(E n F), riasupportinghyperplaneofCatw.(SeeFigure0.1.1.) Then there exists an x e Yl n (E n F) (if not, then C would not be the smallest convex set containing SI), which implies (a) FT is the tangent hyperplane to F at x, and (b) the line segment wx is contained in FI. These two imply that FT is the unique supporting hyperplane to C at every point of vux, except possibly at w. If n were not the unique supporting hyperplane of C at w, then E would have a conical singularity at u\ which would imply that C is not the minimal convex set containing Q. Therefore, every point of E has a unique supporting hyperplane, that is, E is everywhere differentiable.
38 Differential Geometric Methods Figure II. 1.1: The supporting hyperplane at x. Since ? is convex, everywhere differentiable, then every point of ? has a unique exterior unit normal vector, and the Gauss map <3 : ? -> S" that takes every point of w e ? to its exterior unit normal vector n(iu) is well defined. The Gauss map is onto; furthermore, the Gauss map restricted to ?\(? fi F) adds no new points to the image in S"~' that are not already in the image of ? n F, namely, <&(?\(?nr))c<8(?nr). Therefore, to calculate the image of the area of the Gauss map, that is, the area of S", we restrict the integral to ? n F. In a coordinate system on F, as described in the calculations above, the Jacobian transformation of <8, <8,, is given in local coordinates by the linear transformation described by —G~lB [see (II. 1.5)]. For points of ? n F we have —B positive semidefinite, which implies, by the arithmetic-geometric mean inequality, _i = f vnr znr -f /snr < A(?nF) < A(F), which implies the inequality. If we have equality, then ? = F. Also, we have equality in the arithmetic- geometric mean inequality, which implies all points are umbilics (that is, G~lB 11.1 The C2 Uniqueness Theory 39 is a scalar matrix at every point of F), which implies — G~' B is the identity. Therefore (II. 1.4) implies there exists a constant vector b such that n = x —b on all of F. Thus which implies F is a unit sphere. 11.1.3 Uniqueness of an Extremal of the C2 Isoperimetric Functional Theorem II.1.4 Let SI be a bounded domain in R", with C2 boundary, that is an extremal of the C2 isoperimetric functional. Then 3 ?2 has constant mean curvature. Proof Use the notation of Theorem II. 1.1, and set 4> = If SI is an extremal for the isoperimetric functional then, for any choice of vector field X for which = const. Vf, (such a flow is called incompressible) we have, by (II. 1.6), / Jr and, by (II. 1.7), / Jr <j>HdA = 0. Thus, / 4>HdA=0 V0 such that [ 4>dA = 0. Jr Jr A standard argument [see the proof of A.2.3)] implies that H — const. ¦ We now write the equation for H as a second order elliptic quasilinear partial differentia] operator of divergence form. We assume the hypersurface F is given in nonparametric form, that is, the equation is described by
40 Differential Geometric Methods over some domain in R"~'. So, in the parametric language, xJ=uJ, ; = 1 n-l, JcB Then (where grad,,,, denotes the gradient in R"), from which one calculates the metric tensor 30 30 which we write as d<p Note that which implies = |gradn_,|2A, = /- lgradB_,|2 Set SJk = Then for the second fundamental form we have ** ** ,c . so B — <t>~lS. Then H = trG"'B implies H = <D-'tr({/ - * In what follows in this section (only), we substitute the notation V for gradn -1 Suppose we are given two fixed hypersurfaces x" = over some domain in R". Set ), IMC2, and let //«,, H+ denote the mean curvatures associated with <p,ir, respectively Similarly, we use the notations G* and G+, A$ and A+,S+ and 5^, to indicat II.1 The C2 Uniqueness Theory the quantities on each of the surfaces. Then direct calculation implies 41 -'t -H+ = *-'tr(/ - - *-')tr(/ - Now which implies ^ Also, tr(A, - >.*=! where V@ + \fr) is written as a row vector. Therefore there exist a vector field B on R" such that where the operator .4 is given by >lu) = *-'tr(/ - We immediately have Lemma II.l.l Considerthe hypersurfaces inR" determined by the functions (f> and \fr, define A and B as above, and assume the two hypersurfaces have identical mean curvature. Consider the second order linear elliptic differential operator on R" given by Luj = Aw +B-Vit). Then the function w = <(> — \jr satisfies the second order linear equation Lw = 0.
42 Differential Geometric Methods Theorem n.1.5 Let Q be a bounded domain in R", with C2 boundary F, and assume F has constant mean curvature. Then His an n-disk in R". Proof settf = n Fix any unit vector { g S", and pick Euclidean coordinates in R" so that inf x" = 0 xeK (where xn is the nth coordinate of x). Then there exists wo € F D {R" x {0}} with tangent plane of F at wo equal to R" x {0}. For each real a let no denote the hyperplane na = R" x {a}, and set Let K~ = K n {xn <a], K+ = KH {xn > a], r~ = rn[xn <a), r+ = rn[xn > a}. rfl K~ = reflection of K~ in no, rfl F~ = reflection of F~ in no. (See Figure II. 1.2.) Then for sufficiently small a > 0, F~ has a nonparametric representation over I~IO, and Set /3 = sup{a0 : rflK~ C AT+ Va < a0} A x R" Figure II. 1.2: The reflection in na. //./ The C2 Uniqueness Theory 43 Figure II. 1.3: Possibility (a). Then one has two possibilities: (a) There exists u>i g rfl FJ n F^", in which case rfl FJ and F^ are tangent at u>,. (See Figure II. 1.3.) (b) There exists w2 € F D FI^ such that the tangent hyperplane to F at w2 is perpendicular to FI^. (See Figure II. 1.4.) In either case (a) or (b) one may introduce Euclidean coordinates y = (>',..., y") in R" about w - which is either wt or u>2 as the case may be - so that r? and rfl F^ are described by r,+ : yn=4>(y)\ TS_t, .,_K rflr, Figure II. 1.4: Possibility (b).
44 satisfying Differential Geometric Methods 0(W) = Vr(W) = 0, V0(w) = = 0. [so w = (w, 0)]. Thus, the hyperplane v" = 0 is the tangent plane of the two hypersurfaces at w. In case (a) we have if and \fr defined on some open set U c R"-' about w, with 0 > V on U, 0(w) = Mw). The strong maximum principle (Proposition 1.3.9), applied to Lemma II. 1.1, then implies 0 = V on a neighborhood of w, and therefore on all of U. Thus K is symmetric with respect to n^. Similarly, in case (b) we have 0 and V defined on an open set U in R"~' with w g dU, and > yjf onU, where vu denotes the exterior normal (with respect to U) unit vector at w. Again, the strong maximum principle implies <p = $ on all of U. Thus K is, again, symmetric with respect to n^. We conclude that given any { 6 S""', then K is symmetric with respect to some hyperplane perpendicular to (. But then Theorem I.3.S implies 3 AT is a finite union of concentric spheres. Since the mean curvature is constant on all of f, one concludes that AT is a closed disk. ¦ IL2 The C1 Isoperimetric Inequality We first establish the equivalence of the geometric isoperimetric inequality on R" to an analytic Ll-Sobolev inequality. Definition Define the isoperimetric constant 3 ofW by 3 = inf- where ?2 varies over bounded domains in R" with C1 boundary. Define the Sobolev constant 6 o/R" by / 11/11, where / varies over C~(R"). 11.2 TheCx Isoperimetric Inequality 45 Theorem IL2.1 (The Federer-Fkming Theorem) The isoperimetric and Sobolev constants are equal, that is, (II.2.1) 3 = 6. Proof Let ?2 be any bounded domain in R" with C1 boundary. For sufficiently small ( > 0 consider the function II, Jte«, \-(l/()d(x,dn), JceR"\n, d(x,dn)<€, 0, jceR"\f2, d(x,dn)>€. Then f( is Lipschitz for every €, and (by Theorem 1.3.3) we may approximate /, by functions <p(J e CfXR") for which 110*.; - /t|l<i/(«-i) -»¦ 0, ||grad0€,y - grad/€||i ^- 0 as j -*¦ oo. One has llgrad/,11, 6 < IIAII, One checks that lim = V(Q). Also, which implies lim / Thus, lgrad/<l = , d(x,dSl)<(, otherwise, lim for all such Q, from which we conclude S<3. It remains to prove the opposite inequality, that is, (n.2.2) forall/eC^R"). \f\n/("-l)dV l-l/n
46 Differential Geometric Methods Lemma IL2.1 Given any<p e C~(R"), then (II.2.3) |grad |0| | = |grad0| almost everywhere on R". Proof On the open set {0 > 0} wehave|0| = 0, and on the open set {0 < 0}we have |0| = -0, and (II.2.3) is certainly valid. Also, {0 = 0} n {grad0 9^ 0} is an (n - l)-submanifoldof R", which implies yn([<p = 0} D {grad0 ^ 0}) = 0. It remains to consider what happens when 0 = 0 and grad 0 = 0. Set, for any e > 0, Then0e -»¦ |0| ase 4,0, and grad 0€ -»¦ ±grad0ase 4,Owhen0 ^ 0(depend- ing on whether 0 is positive or negative at the point in question). When 0 = 0, then grad0c = 0; therefore grad0e -> grad 0 when 0 = 0 and grad0 = 0. In- Integration by parts implies grad0e converges to the weak derivative of |0| in Wu, which is grad |0|. ¦ Proof of (II.2.2) Given / g Cf(R"), let n«) = {x : \f\(x) > t), V{t) = Then the co-area formula (Corollary 1.3.1) implies / \gmdf\dV= HA(\f\-l[t])dt>3 f V(t)l-l/ndt, JR> JO JO and, by Cavalieri's principle (Proposition 1.3.3), f |/r/c-D<fv= » rt Jit- n — l Jo So to prove (II.2.2) it suffices to show 01.2.4) f°°V(t)l-l/ndt > 1-5- /"OOri/c Jo In - Wo To establish (II.2.4) set F(s) = T VC*I/" A, C(j) = (-?— I' t[/(n-l)V(t)dt Jo I»— Wo First, F@) = G@); //.2 r/i« C1 lsoperimetric Inequality second, since V(s) is a decreasing function of j, we have n — 1 f n I1/" f f G'E) = l-il-5_l j r>/<-') 47 Then (II.2.4) follows immediately. ¦ Theorem 11.22 (lsoperimetric Inequality for C1 Surface Area) Let SI be a bounded domain in R", with C1 boundary dQ. Then (II.2.5) Proof By the Federer-Fleming theorem (Theorem II.2.1), it suffices to show that A1.2.6) j Igrad/lrfV > for all / € Ce°°(R"). Set C = [0, 1 ]", the unit n-cube in R". To prove (II.2.6) we first consider, for any nonnegative function \i e C^R"), the maps uJ : R" -> [0, 00) defined by uJ(x) = / d^ / fi(xl xJ-\ &. $'¦+»,.... J-00 JtL-l 7 = 1 n, and the function r\: R" -> C defined by Then 17 = (^' rj") maps R" onto C, and XJ), —'- = [where we set un+\x) = fi(x)], which implies det /„(*) = fj |^(,) = j j^ rtx) dx J where 7, (jc) denotes the Jacobian linear transformation of rj at x.
48 Differential Geometric Methods The best way to appreciate the mapping is to interpret it when ft is a bounded domain in R" with C1 boundary, and fi = Xn, the indicator function of the domain ft [despite the fact that, in this case, fi ? Cj(R")]. Then «'(*) = vfl(ft) :$¦<*¦}) and vB(ft) u2(x) = vfl_i(ft n {? € R" : *' = x1}), 2 _ Vn-l(ft n(|6B":?'= X\ I;2 < X2}) {X) and, more generally, = vB_,+1(ft n « :?' = *»,... , XJ]) (See Figure II.2.1.) If Q is convex, then tj € C'(?2) n C°(?2), and ./tjC*) extends to a continuous matrix function on ft. A special case is ft = B", in which case we denote tj corresponding to Ib» by *. Then <t>\B" : B" -* C is & diffeomorphism, with ~l det 7* = um~l. We return to the general \i € CJ(R"), set x2 length = v\x)n2(x) Figure II.2.1: When fi = Z^j. //.2 77ie C' Isoperimetric Inequality 49 and think of C as a vector field on R" with Kl < 1, as well as a mapping of R" onto B". Then umfi(x) = det I ^ 1 by the arithmetic-geometric mean inequality. To apply the above, given any nonnegative f e Cf, consider Then (II.2.7) implies = f(x)fi(x)l/n which implies A1.2.8) lgrad/|<fV, which implies (II.2.6) for / > 0. Forarbitrary / e C(!,let(,/<)beamollifieronR'1 andsetF, = j, * \f\. Then (II.2.8) implies Let e 4,0; then, by Lemma II.2.1, To consider the case of equality in (II.2.5), we consider, more closely, what happens when fi = In, for the bounded domain ft with C' boundary. But we must^assume that ft is convex to guarantee that the mapping { € C'(ft) n C°(ft), and to guarantee that 7c(jc) extends to a continuous matrix function on ft. First, we repeat the argument for the inequality. For all x € ft, (II.2.7) implies A1.2.9) V(ftI/n
50 Differential Geometric Methods which implies, by the standard divergence theorem, f / which is the original version of the isoperimetric inequality. If we have equality, then we have equality in (II.2.7) and (II.2.9) - which implies <niI1) • >"' "• on all of ?2; and we have equality in (II.2.10) - which implies (II.2.12) on all of 3?2. For convenience we assume V(Sl) = av. so wB/ V(?2) = 1. From (II.2.11) we conclude, sinceC,J = t.J(xl,....x'), that = XJ for all j = 1 n (for / = 1, a' = const.)- On the boundary of ?2 we have (since C|3?2 : 3 ?2 -> S") From (II.2.12) we conclude that there exists a constant X (depending at most on the point of the boundary of ?2) such that for all j = 1 n. From the case j = n we have A. = 1 for all f" ^ 0, which implies N^ * 3«* ^ t—1 dxJ for all points of 3 ?2 for which ?" ^ 0. II. 3 Bibliographic Notes 51 Fix x',..., xn~2. Then ^r = °. which implies that the locus {^r =const.r : t = 1, ...,n — 2}n3?2isacircle. Thus the intersection of any 2-plane with dQ is a translate of a circle on the sphere. Thus the intersection of Q with any 2-plane is a 2-disk of radius less than or equal to 1. Now the map < takes Si onto B". Therefore there must exist at least one 2- plane for which its intersection with SI is a 2-disk of radius equal to 1. Pick two antipodal points m \ and mi on the boundary of this 2-disk of radius equal to 1. Then for any 2-plane through m \ and m2 the intersection of this new 2-plane with ?2 is a 2-disk; since this new 2-disk contains m\ and m2, it must also have radius equal to 1. We conclude that for every 2-plane through mi and tn2 the intersection of the 2-plane with ?2 is a 2-disk with radius equal to 1. But then ?2 itself is an n-disk of radius equal to 1 and center at the midpoint of m \ and mi. IIJ Bibliographic Notes §11.1.1 The calculations of this section are classical, and one can find the necessary background in the standard books. Nowadays, it is not fashionable to do the calculations in local coordinates, but for a hypersurface in R" they are still extremely useful. §11.1.2 The paper of F. Almgren A986), on which this section is based, is far more ambitious than the result presented here. In fact, our formulation constitituted only the heuristic for his results. Almgren starts with a (* - l>dimensional submanifold T (not necessarily a hypersurface) of R\ * < n, and then seeks a * -submanifold Q with boundary r that minimizes *-dimensional measure among all /fc-submanifolds bounded by r. For the pair ($2, T), he then proves the *-dimensional isoperimetric inequality. The first version of this theorem (the isoperimetric inequality for minimal surfaces) goes back to Carleman A921). But Carleman's theorem assumes the existence of the minimal surface, and that it is topologically a disk. Here, the existence must be proven as well. Since it is known that one cannot expect (in sufficiently high dimensions) an everywhere smooth area-minimizing submanifold $2 spanning T, one must broaden the competition of submanifolds fi to more general geometric objects. But then one can include more general r. We also note that methods are sufficiently powerful to yield characterization of equality in the isoperimetric inequality, namely, it is realized if and only if T is a (* - l)-sphere bounding a A-disk Cl. §11.1.3 Our proof of Alexandra's theorem (Theorem II. 1.5) follows his original A962). A subsequent analytic proof was given by Reilly A977). A more recent proof was given in Ros A988); see the presentation in Oprea A997), Chapter 4. Earlier results can be found in Liebmann A900) and Hopf A950). §IL2 Federer-Fleming theorem was first proven independently in Federer and Fleming A960) and Maz'ya A960); and the isoperimetric inequality for C1 surface area is from Gromov A986), based on an idea of Knothe A957).
Ill Minkowski Area and Perimeter In this chapter, we take some first steps in geometric measure theory, namely, we consider general compact sets, irrespective of the smoothness of their bound- boundaries; and our prime method for proving the isoperimetric inequality for them (after we give suitable definitions of the area of the boundary) is by Steiner symmetrization. The symmetrization argument has great appeal to geometric intuition (although its first presentation by Steiner was incomplete), so we work with it in some detail. There are many other symmetrization schemes (see the bibliographic notes at the end of the chapter); but Steiner symmetrization lends itself, especially in the case of Minkowski area, to very simple arguments for the isoperimetric inequality. Moreover, we are able, in the case of C' boundary, to stay within the arguments of Minkowski area and Steiner symmetrization to characterize the disk as the only case of equality. The argument of this last char- characterization seems to be new, albeit unnecessary - because we shall characterize equality in the more general setting of finite perimeter. However, it is a useful preparation for the general argument Also, the C' isoperimetric inequality suf- suffices to prove the Faber-Krahn inequality (§111.3 below) with characterization of equality. There are two types of possible definitions of areas of the boundary dK of an arbitrary compact set K: 1. The area of 3 AT views 9 AT as a subset of R" without, necessarily, any reference to K. Two examples are the (n — 1 )-dimensional Riemannian measure of a C1 hypersurface of R", and the (n - 1 )-dimensional Hausdorff measure of any arbitrary subset F of R" (to be considered below in Chapter IV). In each example, the two definitions are simply applied to dK. 2. The second type of definition of area is a functional defined on the collection of compact subsets of R", which assigns to each compact subset a num- number meant to describe the area of its boundary. The examples we consider 52 ///./ The Hausdorff Metric on Compacta 53 below are Minkowski area and perimeter. Our most general formulation of the isoperimetric inequality for compact sets uses perimeter instead of Minkowski area; and we are able, in the category of finite perimeter, to char- characterize the case of equality. So, for us, this will be the optimal solution to the isoperimetric problem. Practical differences between Minkowski area and perimeter include the following: 1. The perimeter functional is lower semicontinuous with respect to conver- convergence of indicator functions in Ll, whereas no such semicontinuity exists, in general, for the Minkowski area functional with respect to Hausdorff con- convergence of compact sets. Nevertheless, one does have a restricted version of lower semicontinuity for Minkowski area (see the end of the proof of Theorem IH.2.8 below). 2. It is possible for a compact subset of R" to have n-dimensional Lebesgue measure equal to 0 but its boundary to have positive Minkowski area, which is not the case with perimeter. 3. Minkowski area has no natural localization for a relative neighborhood in the boundary (but see a partial attempt in the proof of Theorem III.2.4), whereas perimeter does. 4. Finally, the isoperimetric inequality for perimeter is sharper than the isoperi- isoperimetric inequality for Minkowski area. We note that a third example of the first type of area is the (n - \)-dimensional integral geometric area IGA"~l(F) of a subset F of R", given by Corollary 1.3.4, where pi denotes projection onto the hyperplane ^1. But we do not go into any details here. III.l The Hausdorff Metric on Compacta Definition Let (X, d) be a metric space. For any compact subset K of X, define the circumradius of K, r(K), by r(K) = inf {p > 0: K c D(x; p) for some x e X). Notation (Recall.) D(x;p) denotes the closed n-disk in R" of radius p, and centered at x.
54 Minkowski Area and Perimeter Notation (Recall.) Let (X, d) be a metric space. For any subset A in X, and € > 0, [A]f denotes the (-thickening of A, defined by [A]( = {xeX:d(x,A)<€]. Remark UI.1.1 If A c [B]r and B c [C](, then A c [C]r+(. Also, if K is compact, then k = n [/a. Definition Let X denote the collection of compact subsets of X. Given E and F in X, define their Hausdorff distance, &(E, F), by 8(E, F) = inf {c > 0: ? C [F],, F C [?]<}. One checks that (?, F) )-> 8(E, F) satisfies the axioms of a metric space. Lemma III. 1.1 Let X be an arbitrary metric space. Then the junctions K !-> r(K), K i-> diam K are continuous on X. Iffiisa positive measure on X that is finite on compact subsets, then is upper semicontinuous on X, that is, limsup fi(Ej) for any convergent sequence (Ej) in X. Proof If c > 0, and E, F e X satisfy S(E, F) < c, then, for any x e X and R > 0 for which we have E c D(x; R), we must also have F c D(x; R + c). This implies r(F) </¦(?) + €. By switching the roles of E and F, we obtain which implies the claim for the circumradius. Similarly, if E, F e X, and E c [F]«, then < diam F + 2e. diam E < diam One easily has that 8(E, F) < c implies |diam E - diam F| < 2e. 1111 The Hausdorff Metric on Compact a 55 If Ej -*¦ E, then, given any c > 0, we have Ej C [E]( for sufficiently large j, which implies fi(Ej) < fi([E]() for sufficiently large j, which implies for all c > 0, which implies the claim. ¦ Theorem III.l.l (Blaschke Selection Theorem) Assume that (X, d) has the property that closed and bounded subsets are compact. Then X, the space of compact subsets of X, is complete. Furthermore, if X is compact then X is compact. Proof Let (?,) be a Cauchy sequence in X. Then consider We claim that ?; converges to ? in the Hausdorff metric. The Cauchy hypothesis implies that is bounded for each j, which implies that is compact for each j, which implies that E is nonempty compact. Given e > 0, there exists N( > 0 such that which implies which implies To prove that Ec[Ej](
56 Minkowski Area and Perimeter we argue as follows. If x € Ej, then x € [E& for all i > j, which implies which implies that for each it > j there exists yk e Uo*?i suchthat<f(jt, >*) < c. Since (y*) is a bounded sequence, it has a subsequence converging to some point Jto- Since yt € (J,>* ?, for all I > k, we have xo € cl (J/>t ?, for all * > j, which implies jc0 e ?, and <f (jc, x0) < €, which implies Jc 6 [E](. Therefore X is complete. Assume X is compact. Let (Ej) be a sequence in X, all Ej pairwise distinct. Given c > 0, there exists a cover of X by open metric disks B(x,;c), i = 1,..., M. Then there exists at least one subset {*<, xit) ?{xi,..., for which the cardinality card \ES:ES c\J B(xit;e), Es D B(xit;() # 0 V ik *=i ik I = oo. Then, for each such Es we have Es C U «(*(,;«) C [?,]2<. Set ?,,o = ?(, the original sequence. Assume one has the subsequence (Ei:N) of (?(). Pick the subsequence (Ei;N+i) of (?<;*) by applying the above argument ,;*) with c = l/(N + 1). Set In*, GN+l = *=i Then which implies III. 1 The Hausdorff Metric on Compacta 57 which implies N for all i, j. Set FN = EN;N. Then 4 mm{^V, AT'} which implies (/>) is a Cauchy sequence, which, by completeness, has a limit. ¦ Let (X, d) be a metric space such that closed and bounded subsets of X are compact. Then for each compact K c X there exists jc0 e X such that K c D(xo; r(K)), where r(K) denotes the circumradius of K. We refer to this closed metric disk as a circumdisk ofK, and denote it by DK. In Euclidean space R", Dk is unique. Lemma III.1.2 Assume also that n is a positive measure on X, and that (X, d, fi) is homogeneous with respect to the metric and measure, that is, for every x, y € X there exists a bisection <P '¦ X -*¦ X, preserving the distance d and the measure fi, and taking x to y. Given K € X, define M{K) = {F 6 X: Then there exists E e M(AT) such that r(E) — min r(F). FeMW > 0}. Similarly, given K e X, define fi(K) = {F eX:n(.F)>n(K), diam F < diam K). Then there exists E' e N(AT) such that r(E')= min r(F). N+l' Proof Because we wish to minimize r(F) in both statements of the lemma, it suffices to consider only those F e X for which r(F) < r(K). Furthermore, because of the homogeneity of the metric and measure, it suffices to consider only those F contained in DK, a circumdisk of K. Let H = {FeX:FcZ>*}. Then we already know that H is compact.
58 Minkowski Area and Perimeter In the first statement of the lemma, we let and There exists a sequence (Fj) in Mo(/0 with r(Fj) -> a as ji -»• oo. Since Mo(/ST) C H, there exists a convergent subsequence Ek -»¦?,? e H. By the continuity of F h> r(F), we have r(E) = or. Therefore, it suffices to show E € M(A"). Since fi(Ek) = fi(K) for all k, we have fi(K) < n(E). Given any € > 0, there exists J > 0 such that 7 > 7 implies o= inf r(F)= inf r(F). FeM(JC) FeMo(K) E C [?,]„ [?]* C which implies for all h > 0, which implies Thus, for all €, Ji > 0. First let c -> 0 and then h -* 0. One then has ? e M(/ST), which implies the first claim. In the second statement of the lemma, we let and Since H is compact, and N(K) is closed, we conclude that N0(AT) is compact, and the second claim follows from the continuity of the functions F h> r(F) and F -* diam F. ¦ III.2 Minkowski Area and Steiner Symmetrization Definition Given a compact subset K of R", we define its Minkowski area Mink (tf) by Mink (K) = hm inf A40 = inf r(F) = inf r(F) fN(JC) F<EN(K) III.2 Minkowski Area and Steiner Symmetrization 59 Remark III.2.1 Note that K t-> Mink (AT) is a functional defined on the col- collection of compacts sets K - not on the boundary of K. Remark IIL2.2 When n = 1, then (III.2.1) Mink(tf) = 2 + 2(card {components of R \ K\ - 2). In particular, the Minkowski area of a point x0 is 2, at the same time that v({jc}) O Remark III.2-3 When K is the closure of a domain SI in R", with C1 boundary 3Q, then it is a standard exercise in differential geometry that where A denotes the standard Riemannian (n — l)-dimensional area. Definition Let X = R\ X the collection of compact subsets of R" with Hausdorff metric, and IT a hyperplane in R". For every x € R", let I* denote the line in R" through x that is perpendicular to IT. For every compact subset K of R" define the Steiner symmetrization of K with respect to IT, stn K, by stn K = (J {h-} x lw, wen where r denotes the closed interval in R, lw = [-ow,aw], ow = ^v,(r n K). So Vi(Iw) = y{(ew D K). (See Figure ffl.2.1.) 1 ¦ One verifies that stn K 's compact if K is compact. 2. Also, since (HI.2.3) we have (III.2.4) *n(K) = f v,(r n K)dyn_i(w), vn(stn K) = yn(K) for all K € X. 3- If t is a translation of R", then str(n) (tf) = t(stn (K)). Lemma III.2.1 Let K € X, and assume K ^ ?>*, that is, K is not equal to its circumdisk. Then there exist a finite number of Steiner symmetrizations in
60 Minkowski Area and Perimeter n slnK Figure III.2.1: Steiner symmetrization. hyperplanes U \,... Tlk such that <r(K). Proof The proof of the lemma rests on the following fact: Given a closed n-disk D in R", with boundary (n — l)-sphere 5. Let o denote the center of D, and n a hyperplane through o. For xeO, let {jci, JC2} = txC\S. Then given c > 0, there exists 8 = 8(e, n) > 0 such that D \stn (D Bs(xl;8) U Bs{x2;8), where Bs denotes the spherical cap in 5 with indicated center and radius. (See Figure III.2.2.) That is, if we remove a disk in D of radius t. centered at jc, then after we Steiner symmetrize we are missing spherical caps in S of radius 8 and centered at Jti and x2. If x € 5, then 8 can be chosen independent of the plane PI. Figure m.2.2: Steiner symmetrization for disk with hole. 111.2 Minkowski Area and Steiner Symmetrization 61 We now prove the lemma. We are given that K ^ DK. Let 5 denote the boundary sphere of ?>*. We may assume that K C\S ^ S. For if not, there exists a disk in the interior of DK, and in the complement of K. Then any symmetrization st with respect to a plane through the origin o of DK will produce spherical neighborhoods in 5 contained in Dk\ st K. So we assume K n S ^ 5. Since DK \ K is relatively open, it suffices to produce a finite number of planes FI i,.... FI* for which (III.2.5) (stn, o-..ostn4tf)nS = 0. Given x e S\K, there exists e > 0 such that the spherical cap Bs(x;e) c S\K. Then the compactness of AT n S implies there exist a finite number of points xi xic in K C\ S such that the (relatively) open spherical caps {Bs(X)\e): j - 1 k) cover K D S, that is, k [J Bs(xj;e)DSnK. For each j, consider the hyperplane FI; bisecting the chord xxj . Then this choice of Fly, _/ = 1,..., k will satisfy (III.2.5), which implies the lemma. ¦ Theorem IIIJ5.1 (Isodiametric Inequality) For any K eXwe have fdiam ATI" 2 J ' Thus, the disk minimizes the diameter of a compact set of given volume. Proof We first show that given any F e X, we have (III.2.6) diam stn F < diam F, for any hyperplane FI in R". Pick Euclidean coordinates in Rn where FI is identified by x" = 0, and write R" as R"-1 x R with R" = n, and R = 1°, where o is the origin of R". Let (w, a) and (v, /?) maximize distance in stn ^> that is, d((w, a), (v, /?)) = diam stn F, where w, v e R", a, /? e R. Because d((w, a), (v, B)) = J\w-v\2+(a-B)>, we must have both cases a = — aw, B=av and a = aw, B = —av.
62 Let Minkowski Area and Perimeter = min (w D F, nw = max (w n F, = min t D F, nv = max lv n F. Then which implies so at least one of nw + fv or nv + ?w must be >|or| + |/J|, which implies (III.2.6). To finish the proof, given K, we consider the compact set E' in N(AT) that minimizes the circumradius (Lemma III. 1.2). We claim that D?< = ?', that is, ?' is a disk. If not, then Steiner symmetrizations preserve N(AT) [that is, F e N(/O implies stn F e N(JO V n], and there exist a finite number of them whose composition lowers the circumradius (Lemma III.2.1), which contradicts the definition of E'. But 2r(E') < diam K, which implies J" > which implies the theorem. III.2.1 The Isoperimetric Inequality for Minkowski Area Theorem IIL2.2 (Brunn-Minkowski Inequality) Given K e X, let D de- denote the closed n-disk of the same measure as K, that is, vn(Z>) = \n(K). Then for all ( >0. Proof First note that, by (III.2.1), the theorem is true if n = 1. For general n > 1, we take the same approach as in the proof of the isodia- metric inequality, namely, we first show that (ffl.2.7) vB([stn F]f) < for all F e X, c > 0, and all hyperplanes n in R". 111.2 Minkowski Area and Steiner Symmetrization 63 Pick Euclidean coordinates in R" so that n = R"~' x {0} - so we identify FI with R", let o be the origin of R" , and V = R, as in the proof of Theo- Theorem III.2.1. Let q : R" -*¦ R denote projection onto the jt"-axis= 1°. For any well, let zw : R" -»¦ R" denote the translation by w. To help avoid confusion, weletD(H';c) = ]Dn~1(H';c)denotetheclosed(n - l)-disk in U centered at w and with radius e. Given wen, and F e X, then (FH.2.8) veD(w:t) [We have written the formula in this manner to emphasize that we project f C\F onto the x"-axis to obtain q(lv n F), carry out the 1-dimensional v/e2 - |v — w|2-thickening in the x" -axis, and then translate out to the line lw.] To prove (III.2.8) just note that (h>, a) € [F]^ if and only if there exists (v, /?) e Fsuchthat|H' - v|2 + (P -aJ < c2,whichistrueifandonlyif<fRi(a,q(tv D F)) < y/f2 — \w — v|2, which is the claim. Therefore where /" is the 1-dimensional symmetrization ofivC\F (to pass from the second to the third line we use the 1-dimensional isoperimetric inequality), that is, (IIF.2.9) n [stn F]<) < v,(r n Then (III.2.3) and (IU.2.9) imply (III.2.7). To finish the proof, given K, we consider the compact set E in M( AT) which minimizes the circumradius (Lemma III. 1.2). We claim that DE = E, that is, ? is a disk. If not, then Steiner symmetrizations preserve M(AT), and there exist a finite number of them whose composition lowers the circumradius (Lemma III.2.1), which contradicts the definition of E. ¦ Theorem 111.23 (Isoperimetric Inequality for Minkowski Area) If K is a compact subset o/R", then Steiner symmetrization of K does not increase its
64 Minkowski Area and Perimeter Minkowski area, that is, A11.2.10) Mink (stn AT) < Mink (K) for any hyperplane n. Furthermore, if the closed n-disk D"(/?) has the same measure as K, then A11.2.11) A(S"-'(/?)) = Mink(D"(K)) < Mink(iT). If K is the closure of an open subset ?2 in R" with Cl boundary 3 ?2, then (in.2.12) A($"-l(R))< AC?2), which is expressed analytically as (im.13) _A^]._>nu,V» Proof Indeed, Theorem III.2.2 implies + h)) - for all h > 0, which implies the claim. ¦ The problem with this approach is that while it quickly and elementarily gives the isoperimetric inequality, it cannot characterize the case of equality - unless one can sharpen the Brunn-Minkowski inequality to include an error term that persists to the limit, as h -*¦ 0, and that vanishes if and only if AT is a disk. In what follows, we shall give a variant of this argument. Theorem IIL2.4 Assume SihasC1 boundary, K = ?2 compact. One has equal- equality in (III.2.12) [equivalent^, (III.2.13)] if and only if SI isadisk. Proof First note that we have the theorem for n = 1. One might wish to apply Steiner symmerization directly to prove ACstn ?2) < AC?2), and then characterize the case of equality. But one does not know in advance that if ?2 has C' boundary, Steiner symmetrization will preserve this smoothness. (One can easily construct counterexamples. See Figure m.2.1.) Yet, there are obvious examples where some smoothness is preserved. We show that this is III.2 Minkowski Area and Steiner Symmetrization 65 always true, and the smoothness that is preserved has enough information to provide the characterization of equality in (III.2.12) we are seeking. Namely, we first give a local analogue of (m.2.7). Let n be a hyperplane in R", and coordinatize R" as in the proof of Theorem ffl.2.1. For z e R" and t > 0, let (as above) D(z; e) denote the closed (n - l)-disk in R" centered at z with radius c, and let C(z;c) denote the solid n-cylinder in R" over D(z;c), that is, C(z;c) = D(z;c)xR. For z e R"~', c > 0, and compact K in R", we have V(ATnC(z;c))= f v,(r C\K)dyn.i(w)= V(stn AT nC(z;c)), JD(z;t) We define a "localized" Minkowski area Mink nu;( by (See Figure m.2.3.) Then (HI.2.9) implies (III.2.14) V ([stn K]h nC(z;c)) < V ([K]h nC(z;c)) , Figure m.2.3: {[*]»\*}nC(z;«).
66 Minkowski Area and Perimeter which implies (m.2.15) Mink n^;«(stn K) < Mink n^*"), Return to K equal to the closure of Q with C1 boundary d&. Let denote the projection map. Fix z e R"~' and c > 0, and assume that D(z;e) consists of regular values of the projection of dQ, p\3Q. Let u denote the exterior unit normal vector field along 3 ?2, and Eh = [v u, u e t e that is, Eh is the curved slab of thickness h determined by starting out at points u € 3 ?2 sitting over D(z; c) and going out h units in the direction uu. Lemma 111.2^ We have the differential geometric estimate (m.2.16) wa[K]h \jr}nc(c«»- V([Eh])\ = O(h2) ash 10. In particular, Minkn.j;<(K) = AC?2 n C(z;c)). Proof (See Figure III.2.4.) Let T = 3f2 n 3C(z;c). Then T is a compact C' (n — 2)-dimensional submanifold of R", with fiber at w € F in the normal bun- bundle of F, spanned by uw (the unit normal exterior vector field along dQ) and nw [the unit normal exterior vector field along 3C(z; c)]. The linear independence of vw and nw is guaranteed by the assumption that p is regular on all of D(z; t). For each w e T, let ,, vw) e @,7r) and * = inf sin dw. wer For any r > 0, [F]r, the tubular neighborhood of F of radius r, satisfies V([F]r) = O(r2) as r 4, 0, since F has codimension 2. For any a > 1, an application of Taylor's formula shows that the symmetric difference {[[K]h \K}n C(z;c)} A Eh of {[K]h \K]D C(z;c) and Eh satisfies \ A Eh C as h i 0, which implies (III.2.16). III.2 Minkowski Area and Steiner Symmetrization 67 Figure m.2.4: ?». Lemma W.23 Assume that D(z; c) consists of regular values of the projection of<M, p\dQ. Let < < ak(w) be C1 Junctions on D(z; c) such that for all w € D(z; €). Define Then forallweD(z,€),and (ni.2.17) AC(stn f2)nC(z;c)) < ACOnC(z;c)) with equality if and only ifk = 1, and SI D C(z;«) ij symmetric, up to transla- translation, with respect to the hyperplane U. n stn #: = [-
68 Minkowski Area and Perimeter Proof Note that w »->• A(w) is C1, and Mink n:z;((stn K) = A(8stn ft n C(z; c)), by Lemma III.2.2. To compare the boundary areas, we have A(8ftnC(z;O) = /" where grad^cr, is the (n - l)-vector(8cr,78wi,... 8cr;/8wn_i), and the same for gradn_,/3;, j = 1 k, and A(8(stn Two applications of the triangle inequality imply ;c)) = 2 f |gradn_, - a,-| (IH.2.18) and k Therefore (in.2.19) if Jk2 + |gradn_,,4|2 dvn_, < A(8« n C(z;OX which implies A11.2.17). We have equality in (III.2.17) if and only if k = 1, and the vectors (gradn_, -cri(w), l)and(gradn_,0i(w), 1) are linearly dependent. Hence (since their last coordinate is equal), = -Pi(w) + const. on all of D(z; e), which implies ft D C(z; e) is symmetric, up to a translation, with respect to the hyperplane n. ¦ Conclusion of the Proof of Theorem m.2.4. Sard's theorem (Proposition 1.3.8) implies that z is a regular value of p|8ft for almost all z e p(dft). Given such a z e p(8ft), there exists e > 0 such that D(z;O consists only of regular 111.2 Minkowski Area and Steiner Symmetrization 69 values, for which (HI.2.17) follows, with its characterization of the case of equality. If we have equality in (III.2.12), then Steiner symmetrization does not strictly decrease the Minkowski area of the boundary. In particular, we have equality in (III.2.15) for any hyperplane n, z e n, and c > 0. Therefore, we have equality in (HI.2.17) for all regular values of p\dQ. Therefore, by continuity, and by applying the argument to every possible hyperplane I~I, we have that every line in R" that intersects ft does so in a segment, that is, ft is convex. Furthermore, ft is symmetric, up to a translation, with respect to any hyperplane FI. By the convexity of ft and Theorem I.3.S, AT is a disk. ¦ III.2.2 Sequences of Steiner Symmetrizations Our proof of the isoperimetric inequality used the fact that Steiner symmetriza- symmetrization did not increase Minkowski area, while it preserved volume. The minimiz- minimizing property of the disk then required an existence argument, using the Blaschke theorem. A more intuitive approach to the minimizing property of the disk is to expect that the more symmetric a compact set, the "closer" it should be to realizing the minimum Minkowski area for fixed given volume. In this section we begin carrying out this approach, namely, we show that starting with any compact subset K of the plane, we may apply to it a sequence of rotations and Steiner symmetrizations so that it will converge, in the Hausdorff metric, to a closed disk. One can then use induction (on the dimension) to prove a similar result for higher dimensions. So we see the Minkowski area decreasing, hopefully, toward the area of the disk. Note that Minkowski area is not necessarily lower semicontinuous, so we cannot (on general principles) pass to the limit to conclude that the area of the disk is less than or equal to all those in the sequence. But we shall be able to prove this result for our particular sequence. We postpone the characterization of equality to the discussion of perimeter. Definition A Borel-measurable function /: R" -»• C vanishes at infinity if f) < +oo V/>0. Definition Let A be a Borel subset of R\ vn(A) < +00. We let A* denote the Schwarz symmetrization of A, that is. A* is the open disk in R" with the same volume as A. Also, define the symmetric decreasing rearrangement of the function IA,J\, by
70 Minkowski Area and Perimeter -*• X Figure III.2.5: Symmetric decreasing reaarangement. For any Borel-measurable /: R" -*¦ C that vanishes at infinity, define the symmetric decreasing rearrangement off, /*, by = r Jo (See Figure III.2.5.) Then we have the following facts (recall Cavalieri's principle (Theorem 1.3.3)): 1. /* > 0. 2. /* is spherically symmetric, that is, \x\ = \y\ implies /*(*) = f*(y). 3. /* is lower semicontinuous (since A* is open), which implies that /• is measurable. 4. {/* > t] = (|/| > t}* for every t > 0; of course, the two sets have the same volume. 5. If *: R+ -*¦ R+ is increasing, then r = ¦<>/•. 6. One has J' 4>o\f\diH = f<t>°fdvn for any <j> = <pi —<fa. such that each <f>j, j = 1,2, is increasing and at least one of <pio\f\ or <fao\f\ is in Ll(Rn). 7. We always have ii/ii, = urn, v P > 1. 8. If/ <g then /•<«*• Theorem 111.15 ///, g are nonnegative on R\ both vanishing at infinity, then (III.2.20) j fgdvn<J f*g*dvn. 111.2 Minkowski Area and Steiner Symmetrization 71 If f = f* and strictly decreasing, then we have equality in (IH.2.20) if and only ifg = g* a.e. Proof By Cavalieri's principle (Theorem 1.3.3), we want to show / ds I dt I Ilf>l}I{g>s}d\n Jo Jo JR" /*OO /*OO f < ds dt I,/>,}'Iu«}*</v»; Jo Jo JR' it suffices to show that lAlBdvn< I Z*ArB, for all A, B of finite volume. Assume vn(A) < vn(B). Then A* c B*, which implies that the right hand side of (HI.2.21) is equal to vn(A*) = vB(A) > vn(A n B), which is the left hand side of (III.2.21). This proves (III.2.20). Now assume equality in (HI.2.20), f = f* strictly decreasing. Then l{g>s)din = j fXlg->s)dvn for almost all 5. It suffices to show that j which would then imply g = g*. Well, since / is spherically symmetric and strictly decreasing, there exists a continuous function r = p(t) such that B"(p(/)) = {/ > V/>0, which implies :=/%„: is continuous for any measurable C. For any given s, pick C = {g > s}. Then Fc(t) < Fc-(t) for all t, by the original inequality (III.2.20), and / Fc(t)dt = / Fc-{t)dt, by assumption. This implies (by continuity) that Fc(t) = Fc(t) for all t, which implies v«({/ > /} n C) = vn({/ > t} n C*) = min (vB(/ > r). vn(C')}. Butthisimplies,foreveryf,thatifC*contains{/ > t]thenCcontains{/ > t], and if {/ > t) contains C* then {/ > t} contains C. Therefore C = C*. up to measure zero, which implies [g > s} is a disk up to measure zero, for every s.
72 Minkowsla Area and Perimeter Corollary III.2.1 For f and g in L2 we have \\r-g*\\2<\\f-gh. A generalization of Corollary III.2.1 goes as follows: Theorem IIL2.6 Let J : R -> [0, oo) be convex, 7@) = 0; and let f,g be nonnegative, vanishing at infinity. Then (m.2.22) f W - g*) dvn < J Jo(f - g) dvn. If J is strictly convex, f = f* strictly decreasing, then equality in (m.2.22) implies g = g* a.e. Proof Write J = J+ + /_, where MO = 0, t < 0, J(t), t > 0, and Then J+ and 7_ are convex. Consider J+, and its right hand derivative; then /+(»)- f J+'(*)ds, Jo which implies i+o(/-«X*)= / J+'(f(*)-s)ds Jgw = /" 7+'(/(x) - iJJ,,^)^)^, Jo which implies I J+o(f-g)Mdvn(x)= f ds f J+'(f(x)-s)Ilg<s](x)dvn(x) Jtt." Jo Jttr < f ds [ J+'(/*(x)- Jo Jr« 111.2 Minkowsla Area and Steiner Symmetrization 73 (the second line follows from Theorem III.2.5). The same argument applies to 7_, which implies the inequality. Now assume J is strictly convex, / = /' strictly decreasing, and equality in (III.2.22). Then we have equality, separately, for J+ and 7_, which implies J J+'(f ~ *)!(,<,) dvn = J J+'if - s)I{g.<s) dvn 2i.c.-[ds], which implies, by equality in (III.2.20), that g = g* a.e. ¦ Corollary m.2.2 For all p > 1 we have (IH.2.23) \\r-g'\\p<\\f-g\\p. Corollary m.23 Let S denote Steiner symmetrization with respect to some hyperplane n in R", and Sf the Steiner symmetrization of any function f, defined by Then (III.2.24) for all fg eLp,p> 1. ,00 Sf(x)= / JS[\f\>t\(x)dt. \\Sf-Sg\\p<\\f-g\\p Proof Use the 1-dimensional Corollary III.2.2 with Fubini's theorem. ¦ Theorem IIL2.7 (Helly Selection Theorem) Given a uniformly bounded col- collection <t> of increasing junctions on [a, b], there exists a sequence in 4> that converges for every x e [a,b]. Proof Let ri, T2,... be an enumeration of the rationals in [a, b]. Then there exists a sequence 01; in 4> such that the sequence ofnumbers0ij(r,)converges, which, in turn, has a convergent subsequence fcjto), and so on. The sequence <t>k,ic converges, as k -*¦ oo, for all rationals in [a, b] to a function <f> defined on the rationals. Extend 0 to the irrationals by = lim where r varies over rationals < x. So <f> is continuous from the left, and increasing.
74 Minkowski Area and Perimeter < \4>k.k(Pz) - <t>k.kiP\)\ + \<t>k.k(p\) - <t>iPi)\ We show that <f>k.k(*o) -*¦ <t>(xo) for every xo at which <f> is continuous. Pick p\ < xo < /t>2, Pj rational. Then <1>k.k(xo) - This implies \<t>k.k(*o) - from which one has the claim. So &,*(*) -*¦ <f>(x) at points of continuity of <f>. Since <p is increasing, its points of discontinuity are countable. But one can use the above diagonal argument to obtain the convergence on these points of discontinuity of <p. ¦ Theorem IIL2? (Isoperimetrk Inequality for Minkowski Length in R2) Let Fbea compact subset of the plane R2. Let stx denote Steiner symmetrization with respect to the x-axis, sty Steiner symmetrization with respect to the y-axis, a an irrational multiple of In, and Ra the rotation ofR2 by a radians. Finally, set Then a subsequence T'k FofT'F converges to the Schwarz symmetrization ofF, F*. in the Hausdorff metric topology. Also, T'F -*¦ F* in Lp, as j; -> oo, that is, Finally, (III.2.25) ||Jr; F - If.II, -*¦ 0 as j -+ oo, V p > I. Mink(F')<Mink(F). Proof Note that (HI.2.23) and (m.2.24) imply (III.2.26) \\T-f-T-g\\p<\\f-g\\p. Certainly, there is a closed disk Do such that F, and all T> ¦ F, are contained in Do. By Blaschke's selection theorem, one has a subsequence T'k-F -*¦ D, where D is a compact subset of Do. We want to show that D is a disk. Set F)k = TikF. First note that, because styStjj, sixFjt = Fh for all jk, we have styst^D = D. 111.2 Minkowski Area and Steiner Symmetrization 75 (Although it is not obvious that Steiner symmetrization is continuous in the Hausdorff metric - both Steiner and Schwarz symmetrization are continuous in Lp - it is nevertheless true that if K; -*¦ K in the Hausdorff metric, and Kj is Steiner-symmetric with respect to some line for all j, then K is Steiner- symmetric with respect to that line.) Next we show that (III.2.27) = D a.e.-[dv2]. Proof: Fix any quadrant Q of R2. Then Q C\ FJt is described by a graph satisfy- satisfying the hypotheses of the Helly selection theorem (except that the functions are decreasing), which implies there exists a subsequence - still denoted by jk - for which the graph of Fjk converges to the graph of D. Lebesgue's dominated convergence theorem then implies (HI.2.28) \\1f, -IdII,-»-0 as *-> oo, for all p > 1. Let g be any spherically symmetric strictly decreasing function, such as g(x, y) = (x, y) 6 R2. Then TJ ¦ g = g and (III.2.26) imply \\g - ItifW I limit (we drop the sub- subscript p for now). Since \\g -IfJI -*¦ \\g -loll, one has \\g-ITj.F\\ -*¦ \\g-XD\\.Aho, \\T(g - ID)|| = \\g - IT.D\\ = lim \\g - lT.Flt II = ll« - IdII = \\g ~ (the last equality follows from the spherical symmetry of g), and \\T(g - ID)|| = list, sty(Rag - I^D)|| = ||stx sty(g - which implies which implies, by equality in (III.2.22) (since g is spherically symmetric strictly decreasing) that sty stx RaD = RaD a.e. -[dv2 ]. In particular, RaD, in addition to D, is symmetric with respect to the x- and y-axes. Let Hx denote reflection in the x-axis. Then RaD = HxRaD = which implies (III.2.27). = R-aD,
76 Minkowski Area and Perimeter Now set then fiF) = 0 on a dense subset of [0,2n). We show that 6 i->- fiF) is contin- continuous when p = 2. It suffices to show that rF):= is continuous. Pick a mollifier y(, c > 0; then is continuous, and \r(@) - r@)\ = = f j( * ID - j( *lD)oR-e)d\n <\\J(*Xd-Id\\\\Ir,d\\ = \\J(*Id-Id\\\\Id\\ -*¦ 0 uniformly in 6, as € I 0, which implies the claim. Therefore ReD — D in L2 for all 0, which implies that D is a disk. Also, since \\ITif - loll, I, we have [by (ffl.2.28)] ||Jr^ - ID\\P 4- 0. It remains to show D = F* and (III.2.25). We certainly have (by Lemma HI.1.1) vn(F) = limsup \n(Fj) < vn(D). Next, for every c > 0, we have vn([Fj\t) I. with respect to j. Since the subse- subsequence Fjk -*¦ D in X, then for every c' < c there exists J such that - C V * > 7, which implies which implies V k > J, vn(D) < We first conclude that vn(D) < vn(F), which implies the two volumes are equal. Therefore D = F*. We also have which implies (III.2.25). 77 111.2 Minkowski Area and Steiner Symmetrization Theorem ITI.2.9 (Isoperimetric Inequality for Minkowski Area in R") Let F be a compact subset ofR", n > 3. Then there exists a sequence Tj consisting of rotations, Steiner symmetriziations, and (n — \)-dimensional Schwarz sym- metrizations, such that Tj-F -*¦ F* in the Hausdorff metric topology, and in Lpforallp> 1. Finally, (HI.2.29) Mink(F*)<Mink(F). Proof Let U denote the rotation of R" that takes en to en_i (and leaves the orthogonal R" pointwise fixed), Fl = e^- = R", S\ = stn denote Steiner symmetrization with respect to II, and Sz denote Schwarz symmetrization in FI; set T = SiSiH. Then the argument of the proof of (III .2.28) implies that there exists a subse- subsequence (kj), Tj=Tk>, such that Zjrf -*¦ ZD (in the various topologies) for some compact D, as ; -+ oo. We also have T> ¦ F ^ D in L'. Since S\ and 52 commute, and TJ F -*¦ Din Lp, we have S2T' ¦ F = and which implies 1 • F = Si • F = -* -F = T'F D, $20=0, that is, D is rotationally symmetric with respect to the x"-axis. Also, given g = e~W\ for example, II* -Ziw,|| = IIK(« -Id)II = U -loll = H« -ZtdII = \\Si$2(g - 2kd)II. that is. which implies [by equality in (III.2.23)] S2HD =
78 Minkowski Area and Perimeter So both D and V.D are rotationally symmetric with respect to the x"-axis, that is, D is rotationally symmetric with respect to both the x"-axis and the x"~'-axis. Let j( be a mollifier, depending only on distance from the origin, that is, spherically symmetric in W, I( = jf *ID, and let p2 = (x1J + ...+(;c'-2J. Then there exist functions / and g of two real variables, such that Set x" = 0. Then which implies "-'J,0) = 2,0) = = It{xl that is, the function It(x) is radial in x, for every c, which implies XdOO is radial in x, which implies D is a disk. The proof of D = F* and (III.2.29) is the same as in the previous theorem. IllJ Application: The Faber-Krahn Inequality We present an application of symmetrization to analysis, more particularly, an application of geometric isoperimetric inequalities to isoperimetric inequalities for eigenvalues. The result goes as follows: Definition Given any open set SI in R", one considers the functional where <p ranges over C^°(Q), and the associated infimum k*(Sl) = inf 0eC«(J We refer to X*(tt) as the fundamental tone of SI. It is well known that k*(Q) is the infimum of the spectrum of the Laplacian -Aonfl, subject to vanishing Dirichlet boundary data (see the discussions in §§VII.2 and VII.4 below). In addition, if Q is a domain with compact closure III. 3 Application: The Faber-Krahn Inequality 79 and C°° boundary, then X*(S2) is an eigenvalue of the Laplacian on SI, that is, there exists a solution <)> e C°°(Sl) to (III.3.1) moreover, <p\Sl # 0 and (d<p/dv)\dSl ^ 0, and the solution <p is unique up to a multiplicative constant. Theorem IIL3.1 Let SI be a bounded domain in R", and let B be the open disk in W satisfying V(Q) = V(B). Then If SI also has C°° boundary, then one has equality if and only if SI is isometric toB. Proof Fix a function / e C?(Sl), and set, for every t > 0, n, = {\f\>t}, r, = an, c {|/| =»}. We use the symmetric decreasing rearrangement (or Schwarz symmetrization) of functions on SI, defined by B, = n; = B"(p(»)), f\x) = f IB,(x)dt. Jo So B, is the disk of radius p(t) whose volume equals that of SI,, and /* is the function that on the sphere S"~'(p(f)) has the value f. Then So to prove the first claim of the theorem, it suffices to show that (IH-3.2) llgrad/||2>||grad/'||2. Set Let (a, fi) be an interval consisting of regular values of /, and for t e (cr, /?) set A(t) := Then 11->- V(t), p(t), A(t) are all C°° on (o, 0) with both V(t) and p(t) strictiy decreasing. On (cr, 0) we have dV n_xdp dt ^ dt -Z— dr , dp/dt
80 Minkowski Area and Perimeter where r denotes distance from the origin. On the other hand, by the co-area formula (Corollary 1.3.1). Finally, set T = sup |/| and r0 = To prove (III.3.2) we have, on the one hand, fo ff On the other hand we have, by the Cauchy-Schwarz inequality, }{^ Igrad/r'rfA}, which implies ff |grad/|2dV = f dtf |grad/|dA JJa Jo Jr, 'Jo A2{t){l lgRld/' = -f A2(tW(t)rl dt Jo > - r A2(sn Jo = -Cn-il —r<it Jo P -1 dt (the last inequality is the isoperimetric), which implies (III.3.2). To consider the case of equality, we assume that ft has a C°° boundary, in which caseX*(ft)hasapositiveeigenfunction/ofXi($2). Of course,/|8ft = 0. The eigenfunction equation (IH.3.1) implies a2/ <0 111.4 Perimeter 81 for at least one of the j e {1,..., n}. Thus, by the implicit function theorem, the set of points for which grad / = 0 is contained in an (n — l)-manifold. In particular, for every f, V(V(t)) — 0, which implies V = V(t) is continuous with respect to t. Now one can argue as above. If X*(ft) = k*(B), then the argument shows that A(dnt)= A(dB,) for all regular values of t. But then for every such t one has Q, isometric to B,, and the strong maximum principle implies that ft = ft@) is isometric to 5@) = B. ¦ III.4 Perimeter III.4.1 Geometric Perimeter Definition We say that a collection of measurable subsets (?,) of K" converges to E in Ll if vn(?;- A E) -+ 0 as ; -> oo, that is, if IEj -*¦ 1e in L'(Rn): lim f\lE)- IE\dvn=0. Definition The geometric perimeter geoper(?) of a measurable subset E of R" is defined by geoper(?) = inf (Iiminf A(dMj)}, where the infimum is taken over all sequences of open subsets Af, of K", with Cx boundary, that converge to E in L1. One easily checks that if Ej -*¦ E in Ll then geoper(?) < Iiminf geoper(?;). that is, ? h> geoper(?) is lower semicontinuous with respect to convergence inL1. Theorem IH.4.1 If K is compact, then K may be approximated in Ll by open subsets with C°° boundary. Furthermore, geoper(Ar)<Mink(Ar). Proof Let f(x) = d(x, K), (j() a mollifier on Rn, and set U = U * /• Then sup |grad/,| = Lip/, < Lip/ = 1.
82 Minkowski Area and Perimeter For any small h > 0, let a — h2. Then there exists c e @, a) such that f(-l[(°,h-a)] CC [K]h\K, and there exists a regular value t € (ct, A - ct ) of /(. By dominated convergence, K may be approximated in L1 by /(~'[@, t)]. Furthermore, by the co-area formula (Corollary 1.3.1), we have / A(frl[t])dt = / |grad/,|dV < vn([K]h \ K), Jo Jf.-<U<>.l>-o)\ which implies there exists a regular value t0 e (a,h — a) such that (h - 2a)A(/r'[«o]) < VnilKl, \ K). Also, /r'fo] = 9ifrl[o, t0)), k c /r'[o, to] c vk\. Given any p > 0, there exists hj -*¦ 0 such that which implies, for €j e @, ay), ct; = li^, and associated /,, j - 2ct, (Mink (K) + p). which implies the theorem. HI.4.2 Functions of Bounded Variation Definition Given an open set SI in R". The collection of functions on SI of bounded variation, BV(ft), consists of those functions / e L'(&) for which sup / fdi\i < J +oo, where ? varies over C,!(Q) vector fields on Q satisfying |?| < 1 on all of SI. We refer to these vector fields as admissible vector fields on SI. If the function / e BV(ft), then the Riesz representation theorem (Proposition I.3.S) implies that there exists a vector-valued Radon measure 6f on SI defined by J ftliv 1II.4 Perimeter 83 for all admissible vector fields ( on SI, and the total variation of 6f satisfies |«/|(n) = sup f fdiv?dvn where the supremum is taken over all admissible (. Example IH.4.1 If / e Cl(Sl) n Z.'(fi), then / e BV(«) and d6f(x) = -(grad f)(x)dvn(x). Indeed, for any admissible ( we have / fdiv?d\n = - I grad/-^dvn Ja Ja Remark m.4.1 The same result (Example III.4.1) is valid in Wll(S2) (see Remark 1.3.3 below). Proposition HI.4.1 If SI' c Q then Ff)\sr = 6(f\Q'). Theorem HI.4.2 Let Si be open in R\ (fj) a sequence of functions in BV(fi) such that fj^f in L^Q). Then f e BV(Q) and \6f\(Sl) <liminf\6fj\(Sl). y-.oo Proof For any admissible ? on SI we have f/div? = = lim I /ydi = lim f < liminf \6fj\(Q) - the second line follows from the compactness of the support of ?, / e L^, and dominated convergence; and the last line uses |(| < 1 on all of SI. ¦ Corollary IH.4.1 Given the open set SI in R", endow the space BV(Sl) with the norm Then B\(Sl) is complete relative to the norm || ||bv-
84 Minkowski Area and Perimeter Proof If (fj) is a Cauchy sequence in BV(fi), then (/)) is a Cauchy sequence in Ll(Q), which implies there exists / 6 Ll(Q) such that /, -»¦ / as j -*¦ oo. Then, by Theorem m.4.2, we have \6(f - fj)\(Si) < liminf \6(fk - fj)\(Q) < liminf \\fk - /y k~*OO K—+OQ which implies the corollary. ¦ Remark IIIA2 Note that for arbitrary / e BV($2) we do not expect to ap- approximate / by C°° functions, namely that there exist (fj) e C°°(n) such *at Wfj ~ /Hi -*¦ 0 and \6(fj - /)h(J2) -*¦ 0, for this would imply that / 6 W '¦' (SI) - which is not necessarily the case (see the case of indicator functions in Example III.4.2 below). Theorem IIIA3 Given the open sets Q in R", A CC «, and f e BV(«), satisfying (III.4.1) |«/|(8A) = 0. i^f G«) fc« a mollifier on Rn, f( = j( * f. Then \6f\(A) = 1im\6f(\(A). Proof One has f( -*¦ f in L'(Q), which implies \6f\(A) < liminf \6f(\(A). For $ admissible on A we have j /(div? = j" (y( * /)div? = f /divO, ¦ C)- Now supp(y» * $) C [supp^ c [Al, which implies I ft*™*. < f /div(y( * O < \6f\([A](). Ja J[aI Therefore \6f(\(A)<\6f\([A](), which implies limsup \6f(\(A) < \6f \(cl A) = \6f\(A) by (III .4.1), which implies the theorem. ¦ 111.4 Perimeter Theorem IH.4.4 /// e BV(R"), (j() a mollifier on R", then = lim \6f( \(Rn). 85 Proof The proof is similar to the proof of Theorem III.4.3. Theorem III.4.5 Let Q be open in R\ Given any f e BV(fi), there exists a sequence of C°° functions (fj) in h\(Q) such that l-}™o f> = /• Proof Given c> 0, there exists an open nocc« such that Let «0 CC fii CC «2 CC • • ¦ be an exhaustion of Q, with chose above, and set Ai = «2. ^ = n<+i\clfi<_1, 1 = 2, 3 (&) a partition of unity of ft subordinate to (A*), (j() a mollifier on R". For each * = 1,2 pick €k in @, c) such that the one we S"PP ju * (ftk) C nk+2 \ cl Hk_2, «_, = 0, f \M * (f<t>k) - f<t>k\ <€/2k, I \U * (/grad^) - /grad&| < c/2*. and set ^One easily sees that /, -> / in L'(O) as e | 0, which implies, by Theorem |«/| < liminf |*/«|. no
86 Minkowski Area and Perimeter Given an admissible ( on ft, we have = E [ /div<**0«. * O) - - the last equality uses the fact that (<?*) is a partition of unity. Now f /div(*,a,*O)<l«/l(«); •/a, A* c ft \ cl Qo f°r aU * > 2, and every point in ft \ cl Qo is covered by at least one of the collection (Ak) and at most two of the collection (At), which implies T f /div (&(]« * O) < 2 f T /div (<fik(U * O) <2|«/|(ft\clfto)<2e, which implies \6f(\(Q)<\6f\(Q) + 3€, which implies the theorem. ¦ 111.4.3 Caccioppoli Sets We consider our most important example of functions of bounded variation - indicator functions of Borel sets in R". Notation For the indicator function of a set E for which Xe has bounded variation, we always write 6E := 6XE. 111.4 Perimeter 87 Example II1.4.2 Given ft, let ? be a domain in K" with C1 boundary, and IE its indicator function, with vn(E Dft) < +oo, that is, Jglft e Lx(Sl). Then it is standard that Jf|ft ^ W1' '(ft). For any admissible ( on ft, we have = f JdEnn where i/? is the exterior unit vector field along BE. If A(dE OQ) < +oo, then Z?jfteBV(ft)and = A(8E n Q). Remark IIL4J For arbitrary measurable ? with Jf|ft e BV(ft), we always have supp 6E C 3 ?. Indeed, if F is open in int (Q \ E), then I? | F = 0 implies / Jfdiv ^ = 0 for all admissible vector fields on F. If G is open in int (ft n ?), then XE\G = 1 implies / J?div? = fG div? = 0 for all admissible vector fields on G, which implies the claim. We now extend Example III.4.2 to Lipschitz domains. Definition A domain ?1 in K" is a Lipschitz domain if for each w e 8ft there exist a neighborhood U = U(w) in R", Euclidean coordinates u, and a Lipschitz function 4>:Un(W~l x {()})-> R such that 9ft n {/ = <t>(u\x) un-\x))). For convenience, we shall write R"~' for R"~' x {0} and « = («',..., «""', u") = (u,un),ueRn-1. Example III.4J Let n be a Lipschitz domain, w e BQ, U, u, and <p as above, and assume <p is differentiable at u(w). [By Rademacher's theorem (Theorem I.I.3.2), <f> is differentiable a.e.-[dvn_i] on U nR".] Then 8ft has a tangent hyperplane at w, with exterior unit vector v given by v_ grad«/>-en where e« denotes the unit vector in the positive direction of the «"-axis [the gra- gradient here is(n - l)-dimensional]. Let (<?,) approximate^ in the Sobolev space Wil(U n R"), 4>j e C°° for all ; (Theorem 1.3.3). Then <f>j -> <f> uniformly
88 Minkowski Area and Perimeter on compacta. Let Aj = {xeU:un(x)><f>j(u{x))), Uj = exterior unit normal vector field along 3 A;. Then / ZA,div? = f ^6Aj, Ju Ju where &la(t) is the 1-dimensional delta function (of t) concentrated at a e R, and Let j -*¦ oo. Then f JA Ju f Therefore, Jn|f/ e BV({/) and which generalizes the classical formula A.3.8). Definition We start with a Borel set E. For any open Si, we define the perimeter ofE in Si, P(?; J2), to be \6E\(il) when finite, and otherwise +oo. Remark III.4.4 ? Borel guarantees that E is |6?|-measurable. Definition We also define P(E) = P(E; R"). ? is called a Caccioppoli subset of R" if P(?; Si) is finite for all bounded open Si in R". A Caccioppoli set in R" is also referred to as a set with locally finite perimeter. Proposition IU.4.2 We always have: (a) E CC Si CC Si, => P(?;Si) < />(?;Si,). (b) P(E{ U ?2; Si) < />(?,; Si) + />(?2; Si), witfi ^Ma/ify ifd(Ex, ?2) > 0. 111.4 Perimeter 89 (cj vfl(?) = 0 => />(?) = 0. ^ vn(?, A ?2) = 0 => />(?,) = P(?2). Remark III.4.5 Note that (c) in the above proposition differs markedly from Minkowski area, where a set might have measure 0 but nonzero Minkowski area (Remark III.2.2). Remark III.4.6 We restate the definition of Caccioppoli set slightly differently. A Borel set ? is a Caccioppoli set in R" if there exists a vector-valued Radon measure /j on R" such that j div i dvn = for all admissible ? on R". Indeed, for any bounded open Si in R" we have /'(?;Si)<|/i|(Si)<+oo, which implies ? is Caccioppoli. The measure ft must be unique; and since fZ-dn= fZ-d6E for all admissible ? on R", we have /j = 6E. Theorem IH.4.6 Let E be a Borel set in R" such that IE e BV(R") (that is, E has finite volume and perimeter), (j() a mollifier on R". Set <p = Xe and (p( = i( * <p. Then (III.4.3) \6E\(A) = lim\6<pt\(A) for all A e R" with \6E\(A) = 0. Also, (III.4.4) Proof This is a direct application of Theorems m.4.3 and III.4.4. ¦ Theorem UI.4.7 (Co-area Formula for Perimeter) Given Si open in R", / e BV(Si), let F, denote the level domain Then (HI.4.5) F, = {x € Si :/(*) 6F,dt, \6f\= P \6F,\dt. J
90 Minkowski Area and Perimeter Remark HI.4.7 Thus, when / is C00, the second equality in (III.4.5) reduces (by Example m.4.1) to the co-area formula for smooth functions (Corollary 1.3.1). Proof For any a e R, and we have Note that r>0> /.oo a= / *« J-oo (t)dt. |a-0|= Then, for /(*), we have If,(x) = 2<-oo./(x))@. which implies that for any admissible ? on Q we have = J dt j {XFl - 1 }div S + j°°dtj Jf,di = r dt f jf,di J-oo J = / i-6F,dt. J-oo \6f\< f \6F,\dt. J Therefore, By Theorem III.4.5, there exists a sequence (/,) e C°° such that fj -»• / in /.'(«) and |fi/;| -»• \6f\. Then FA -XFl\dt = j dtj I/; - [where (Fj), denotes the level domain of /, associated with the value /], which implies (Proposition 1.3.1) there exists a sequence (jk) such that ZiFh), -*¦ If, III.4 Perimeter in L\Q) for almost all /, as k -»• c». Therefore, 91 j \6F,\dt < j liminf \6(Fjt)r\dt \6(Fjk),\dt = lim inf \6fjt | - the next to last equality follows from the co-area formula for C°° functions (Corollary 1.3.1) - which implies the theorem. ¦ Lemma III.4.1 Let E be a Cacdoppoli set in R", (_/<) a mollifier on R", and t e @, 1). Set <p = Ze and <p( = jf * <p, and Then I* - Proof If jc e ?\(?€), then ^(x) = 1 and <p((x) < t, which implies Similarly, if j: e (Ee), \ E then <p(x) = 0 and <p((x) > t, which implies Therefore J \<p(-<p\d\n> J tliEME + 0 ~ > min(/, 1 - /) J ^e,), - lE\dsn, which implies the lemma. ¦ Theorem III.4.8 Let E be a Cacdoppoli set in R". Then there exists a sequence of domains (Et) in R", with C°° boundary, such that L1- lim IEl = Ie, Hm \6Et\ = \6E\. i-+oo y-»oo In particular, the perimeter of E is equal to its geometric perimeter.
92 Minkowski Area and Perimeter Proof Start with <p = XE, the mollifier 0",) on R", <p, = je * tp, and (Ee), as above. Then the co-area formula (Theorem III.4.7) implies \6<pt\= / \6{E(),\dt. Jo For any bounded open SI, Lemma III.4.1 implies that XiE,), -*¦ 1E in LX(Q) as € I 0, for every / e @,1). Therefore, by Theorem III.4.2, for all t, which implies, \6E\ < liminf \6(E(),\ A0 \6E\= I \6E\dt Jo < f liminf \6(E(),\dt Jo <l° < liminf / \6(E(),\dt = lim \6<pe\ = \6E\, - the last equality follows from Theorem III.4.6 - which implies |fi?| = liminf \6{E(),\ tiO for almost all t €@, 1). Now pick e* -*¦ 0 as k -*¦ oo. Then the same argument for E* = E(k shows that \6E\ = liminf \6{Ek),\ for almost all / e @, 1). Sard's theorem (Proposition 1.3.8) implies v, l(Jcritval<0(J =0, where, for any function H, critval H denotes the set of critical values of H. Let / be in the complement of |J* critval <pu in [0,1), that is, / is a regular value of 111.4 Perimeter 93 <p(t, for all k. Then there is a subsequence at = ?*, such that \6E\ = lim \6(Eat),\. t-*oo Then the sequence (Eat), will do the job. Thus geoper(?) < P(E). For any sequence Ee, dEt e C00, with Et-* E in L1, we have liminf A(8?<) = liminf />C?<) > P(E), which implies P(?) < geoper (?), which implies the two are equal. ¦ Remark III.4.8 Note that the sets (Ea, )t converge to E in the Hausdorff metric on compact sets. 111.4.4 The Isoperimetric Inequality for Perimeter Remark III.4.9 Note that Theorem III.4.1 now implies that the isoperimetric inequality (to be proved below) for perimeter is sharper than the isoperimetric inequality for Minkowski area. Theorem III.4.9 Steiner symmetrization does not increase perimeter. Proof Let S denote Steiner symmetrization with respect to some fixed hy- perplane, K a finite disjoint union of compact domains with C°° boundary. Then P(K) = A(8K) = Mink (K) > Mink (SK) > P(SK), by Theorems III.2.3 and III.4.1. For arbitrary Borel E with finite perimeter, let Kt -»• E in L\ P(Ke) -> P(E), as t -* oo , and 8Kt e C°° for all t. Then, by Corollary III.2.3, *,-2seIIi<P*,-IeIIi -^0, which implies Isk, -*¦ Xse in Z.1, which implies P(SE) < liminf P(SKt) < liminf P(Ke) = P(E), which implies the claim.
94 Minkowski Area and Perimeter Theorem ni.4.10 (Isoperimetric Inequality for Perimeter) Let E be com- compact, and E* denote the Schwarz symmetrizjation ofE. Then (III.4.6) P(E') < />(?), with equality if and only if E is an n-disk, except for (at most) a set of d\n-measure equal to 0. Proof (Geometric proof.) By Theorems III.2.8 and III.2.9, there exists a se- sequence Tj consisting of Steiner symmetrizations, (n — l)-dimensional Schwarz symmetrizations, and rotations such that Tj< • E -*¦ E* in V as j -> oo. Then P(E') < liminf P(Tj • E) < P(E), which implies the inequality. (Analytic proof.) Given any / € BVC (where BVC denotes the compactly supported functions in BV), we can approximate / by ft e C™ such that 11/* " /111 -> 0, llgrad/iH, = ||6/*||bv "> II«/IIbv, as k -*¦ oo. We already have Since /* -*¦ f in Z.1, then /* Fatou's lemma. / pointwise a.e.-[dvn], which implies, by " liminf \\fk\\n/{n-n< liminf ||*/*||bv = H*/Ibv- t-»oo k-*oo But this implies the isoperimetric inequality, by picking f = Xe- We now characterize the case of equality in (III.4.6). The reader is referred to the proof of Theorem III.2.4, where it is assumed that the boundary of the minimizer is C1, since we will draw on that argument. Fix E, compact, that satisfies equality in the isoperimetric inequality (III.4.6). Then P(E) = P(SE) for every Steiner symmetrization. Let AT; be a sequence of compact domains with C°° boundary such that Kj -*¦ E, in the Hausdorff metric and in L1, and P(K}) -+ P(E). Fix a hyper- plane n, and let S denote Steiner symmetrization with respect to fl. Since K} -+ E in L\Rn), we have SKj -»• SE in L'(Rn), and, by passing to subsequences if necessary, we may assume \6Kj\ -»• |fi?|, 1*5^1 -»• \6SE\. For each j, define the graph function A) on n associated with SKj, as in the Hl.4 Perimeter proof of Lemma IH.2.3. Then (HI.2.19) implies 95 = P(SKj). which implies P(E)> > II II We claim that P(E) = (III.4.7) lim sup 2 / Jk* + lim sup 2/ Jl + |; lim P(SKj) P(SE). P(SE) implies lim \n-i(kj > j—*oo lgradn_,Al2 gradn_,^|^ l) = 0. Proof Assume there exists e > 0 such that vn_i(A:; > 1) > e for infinitely many j. Now j) > 2 [ Jkj(w? + |gradn_lv4,|(H>J rfv which implies that ,,.,^! > N) < N for sufficiently large j; so we may pick W sufficiently large so that vn_1(|gradn_1.4;|>./V)<| for sufficiently large j. Therefore <N)O {kj > 1}) > 1,
% Minkowski Area and Perimeter for infinitely many j. Because -H |gradn_1v4/|2rfvn_, J{\gnA,-[Ai\<N)r\[kj>\) II * "T iv /2 P(SKj) e > — + const. - 2 2 for infinitely many j, we have a contradiction to P(E) = P(SE), which implies (ni.4.7). Let EL denote the Lebesgue set ofE, that is, ? e EL if $ e ? and lim~^ / IEdvn = l. Then Lebesgue's density theorem (Proposition 1.3.6) implies that vn(? \ EL) = 0. Let ?o denote the closure of El. Then Eq C ?, ?o is compact, and vn(?\?0) = 0. So we replace E by Eo. Then vn(?) = vn(?0) and P(E) = P(E0). Given the hyperplane n = R"~', let p: R" -*¦ R" denote the projection; and given any w e R", let ^^ denote the line through w perpendicular to Rn-1 Let w e p(?o), and set = inf er n ?0, = sup r n ?0. Assume there exists /o € (o(w), ^(w)) such that xo := (w, to) & ?o- Then there exists r > 0 such that !(xo;r) C R"\?o; and there exist t\ e [a(w), to- r] such that xi := (w, »i) e ?o, and r2 e I'o + r, 0(w)] such that X2 ¦= (w, t2) e ?0. Pick e > 0 so that, for any points y\ e B(;ri; e) and y2 e 1(j:2; c), the line joining them must intersect B(*o;»"). Given such an e > 0, there exist yi e B(jri; e) n ?t and n e B(j:2; e) n ?*,, which implies that, for any given 8 > 0, .4 Perimeter 97 Figure 1II.4.1: The hole. there exists e' > 0 such that vn(B(y,;e')n?) > A - (See Figure III.4.1.) Since ATy -»• ? in L1 we have J > 0 such that ; €')) On^) > d - vn(B(y2;O n Kj) > A - > 7. Since the boundaries of Kj dieCx, wehavevn(ATy) = vn(int K})
98 Minkowski Area and Perimeter for all j > J, which implies vn(B(y,;e')flint*,) > A - 2S)vn(B(y,;€% ; «') n int *;) > A - 25)vn(B(y2; e')) for all j > 7, which implies that for every j > J there exist points zi, z2 and jO > 0 such that C B(y,;<r')n B(z2;p) c B(v2;e')n Let llx -Z2 denote the line determined by z i and z2. Then for any line intersecting B(zi; e'/2) and B(z2; e'/2) must intersect B(x0; r), which contradicts (III.4.7). Therefore, C n Eo = l<x(w), p(w)]. Thus, equality in the isoperimetric inequality implies P(SEo) = P(E0) for all symmetrizations S, which implies Eo must be convex. Then for any hyper- hyperplane n, the corresponding functions a(w), -f}(w), w e pn(^o). are convex, and hence locally Lipschitz. Set A(w)= . Since ?n is Lipschitz we have, by Example III.4.3, \6E0\(w, y) = rfvn_i(w) dy (where S1 denotes the Dirac delta function on the line R1), which implies, by (III.2.18), P(E0) = lgradn_, -a|2 > 2 j yi = P(SE0), with equality if and only if ?(w) + o(w) = const, for almost all w e Continuity of the functions a(w) and 0(w) implies /3(w) + a(w) = const, for all w e pn(^o)- Then ?0 is convex and is symmetric, up to a translation, with respect to any hyperplane n, which implies Eo is a disk, by the convexity of Eo and Theorem 1.3.5. ¦ U1.5 Bibliographic Notes 99 IIL5 Bibliographic Notes The fundamental treatise on geometric measure theory is Federer's A969). An elemen- elementary introduction is Morgan A988), and a more advanced treatment is Simon A984). §111.1 The treatment of the Hausdorff metric on compact sets, including Blaschke's theorem, follows BergerA987, Chapter 9). §HL2 Steiner symmetrization can be found in Steiner A838). It initiated a whole school of techniques with applications in geometry and analysis. See the geometric dis- discussion in Burago and Zalgaller A988, Chapter 9), and analytic discussions in the classic P61ya and Szegd A951), and the more recent Kawohl A985), and Baernstein A995). Our use of the phrase Schwarz symmetrization is not universal. Except for Steiner symmetrization, none of the nomenclature seems to be universal; so when reading any presentation, caveat emptor applies. The technique of using sequences of Steiner symmetrizations goes back, at least, to (he important paper of Caratheodory and Study A909). The paper is one of the early -if not the first - rigorous proofs of the isoperimetric inequality by Steiner symmetrization. See the discussion in Blaschke A956, p. 43 ff.) Our treatment follows, rather closely, Lieb and Loss A996, Chapter 3). Deeper results on approximations by sequences of symmetrizations can be found in Almgren and Lieb A989) and Burchard A997). §IIL3 Theorem III.3.1 answers in the affirmative a conjecture of Lord Rayleigh A877, §210). It was first proven in Faber A923), Krahn A925). The argument can be easily carried over to spheres and hyperbolic spaces; namely, as soon as one has metric disks as the solution to the geometric isoperimetric problem in these spaces, one automatically has the disks as solutions to the corresponding eigenvalue isoperimetric problem. See Chavel A984, Chapter IV). §111.4 The discussion of perimeter follows the exceptionally clear discussion of Giusti A984, Chapter 1). Ziemer A989) was also very helpful. The discussion of the isoperimetric inequality for the perimeter of compact sets was influenced by Talenti A993). We did not follow his argument closely, since we were able to use the previous arguments from Minkowski area.
IV Hausdorff Measure and Perimeter In this chapter we introduce the most general construction of measures in a metric space, Hausdorff measure. As emphasized earlier, it provides a measure on the boundary of any domain, in terms of that set itself (not as a boundary), irrespective of regularity of the set. We show that, in Euclidean space, Hausdorff measure in the top dimension leads to the usual Lebesgue measure, and on smooth Riemannian manifolds it leads to the usual Riemannian theoretic volume measure. Finally, we show that if ft is a Lipschitz domain in R", then the (n — l)-dimensional Hausdorff measure on the boundary of Q, 8Q, coincides with the perimeter measure of Q. Our proof requires the area formula for Lipschitz maps, which we prove here as well. It was tempting to include in this chapter the structure of the boundary of domains with locally finite perimeter, and the relation of perimeter to Hausdorff measure. But we would have had hardly anything to add to the excellent treatment in Chapters 2-4 of Giusti A984); so we left the matter to the reader to explore there. IV.l Hausdorff Measure Definition (Recall.) GivenasetX,aa-a/?e6nJ<SJnXisacollectionofsubsets of X satisfying: (i) <p,XeS; (ii) ifAe<SthenX\/l€S; (iii) if(Aj) e S then (J^ Aj e S for any countable sequence of subsets of X. A measure on Sis a function ft: S -*¦ [0, +oo] satisfying: (i) /i@) = 0, (ii) (i is countably additive, that is, fi(\J} Aj) — Y.j MCA;) for countable pair- wise disjoint (Aj) € «S. 100 IV.l Hausdorff Measure 101 Sometimes it is awkward to specify at the outset the a-algebra S. Instead, using Carathe'odory's criterion,, one starts with an outer measure on the full collection of subsets of X, 2X, aind creates the a-algebra and measure from the outer measure. Namely, Definition A function 4>: 2X ->* [0, +oo] is an outer measure if it satisfies: for any (i) #0) = 0; (ii) <f> is monotone, that is, if A c B then <p(A) < 4>{B); (iii) 4> is countably subadditiwe, that is, #((J, Aj) < countable sequence of subsets of X. One then calls a subset E of X ^-measurable if <f>(A)= for all A e 2X. The collection olf ^-measurable subsets of X, S+, is known to be a ct-algebra, and the restriction n = <p\S,p of 4> to S$ is a measure. Definition An outer measure (i iis called regular if to each Ac X there exists a ^-measurable B D A such that n(B) = n{A). (Note that this definition of regularity is for outer measures.;) Note that if fi is regular, then any sequence (Aj) in X satisfying Aj C Aj+i,, for all j, also satisfies If X is a topological space with oiuter measure ft, we say that fi is Borel regular if for each Ac X there exists a Borel set B D A such that fi(B) = fi(A). If (X, d) is a metric space withi an outer measure ft, we say that /a is a metric outer measure if d(A, B)>0 H(A UB) = fi(A) + fx(B). Theorem IV.I.I (Carath^odory's Theorem) If ft is a metric outer measure °n X, then all Borel sets are fi-mieasurable. Proof It suffices to show for all E c X satisfying fi(E) < +oo, and for all closed subsets C. Let Cj = {x€X:: d(x, C) < \/j) = [C]Uj
102 Hausdorff Measure and Perimeter for j = 1,2 Then E\C D E\Cj, which implies fi(E) > ix((E\Cj) U (E n C)) = n(E\Cj) + m(? n C). Let 7 -»• oo; we shall show that fi(E\Cj) -> (i(E\C). To this end, set Then E\C = (?\Ci) U |J /?; = (?\C*) U 0 Rj, which implies fi(E\Ck) < a(E\C) < Now the series JV fi(Rj) converges; indeed, ;=* 7=1 Therefore ?/>* 0 as Jfc -»• c», which implies the theorem. Definitioii Let X be a metric space, 7" a family of subsets of X, and (:P -*¦ [0, +oo]. A covering (?;) in J7 of a subset A in X is called a S-cover of A if diam?j < S for all j. Assume that (X, 7", ?) satisfy: 1. for every S > 0, X has a 5-cover; 2. for every S > 0, there exists ?ef such that f (?) < 5. Define the S-Hausdorff (outer) measure fa on X by where (?;) varies over all 5-covers of A by sets in J7, and the Hausdorff (outer) measure rj/ on X by = lim IV.l Ham Theorem IV.1.2 We have (a) ^@) = 0; (b) rjr(A) is well-defined, because fa (c) A (-> ^f(A) is an outer measure. Proof We only have to check subadd is a 5-cover of B, then (Ej, F}) is a S which implies which implies Theorem IV.l3 The Hausdorffmeas fore, rjr is a Borel measure. Proof Given A, B with d(A, B) > 0, ofAUB. Then any given Ej car which implies U S) > One now easily has rJf(AUB)=: Theorem IV.1.4 If the members ofT c Proof GivenA c X,pickopen(l/*)-c satisfying
Hausdorff Measure and Perimeter 104 and set k=\ j=\ One checks that B is a Borel set containing A, and rj/(A) = Definition Let X be a metric space, 0 < s < +oo, T = 2X all subsets of X, and f(E) = (diamf)', with the conventions 0° = 1 and (diam 0Y = 0. The resulting S-Hausdorff and Hausdorff measures are denoted by 7i's,7is, respectively, and referred to as the standard S-Hausdorff and the standard Hausdorff measures on X. When no explicit mention of T and f is made, we are always speaking of the standard Hausdorff measures. Note that W°(A) = card A. Proposition IV.1.1 If X is a separable metric space, we obtain the same mea- measure W if we pick = {EcX:E closed}. = Hn (because diam conv A = diam A for all A), we may add to the list: = {E C X : E = convex}. Proposition FV.1.2 Let X bea separable metric space. The following are equi- equivalent: (a) H'(A) = 0; (b) H\(A) = 0 for some & > 0; (c) given € > 0, there exists a cover (Ej) of A such that < €. Theorem IV.1.5 For 0 < s <t< +oo, A C X we have (a) H'(A) < +oo =>¦ H'(A) = 0. (b) H'{A) > 0 => H'(A) = +oo. TV. 1 Hausdorff Measure Proof Let (?,) be a 5-cover of A. Then H'S(A) < ? (diam Ej)' < B'~s ? (diam Ej)', which implies which implies the theorem. Definition The Hausdorff dimension of A in X is given by dim A = sup {s: H'(A) > 0} = sup {s : K'(A) = +00} = inf{/: 7i'(A) < +00} = inf {/: H'(A) = 0}. Note that Also, ACB dim A < dim B. dim M Aj = sup dim Ay. J 105 IV.1.1 Example. Euclidean Space Remark IV.1.1 By Proposition IV. 1.2, if a Lebesgue-measurable set A in R" has Lebesgue measure 0, then W(A) = 0. Remark IV. 1.2 If X = R\ T a Euclidean transformation of R\ then If A. >0then = k"H'(A). Lemma IV.1.1 Given any bounded open U C R\ and S > 0, there exists a pairwise disjoint family of closed disks (D;) all contained in U such that diam D, < S V ;, yn(U \UjDj) = 0. |*roof Let U = \Jk Ck, where Ck are closed n-cubes with pairwise disjoint interiors, and diam Ck < 8 for all k. Pick closed disks D; to satisfy Dj c Cj, diamDy > -sideC, V j.
106 HausdorffMeasure and Perimeter Then the radius r(Dj) of Dj satisfies side Cj r(Dj) > 4 ' vn(C,) > vn(D,) > -?i and which implies Therefore there exists N\ > 0 such that Now set and find (in the same manner) pairwise disjoint DNl+l DNi c U\ satisfy- satisfying One continues in the same manner, to obtain the lemma. Theorem IV.1.6 For all Lebesgue-measurable A C R" we have Proof Let (?/) be a 5-cover of A. Then IV.l Hausdorff Measure by the isodiametric inequality (Theorem III.2.1), which impli vn(A) < ^W(A). 107 es For the opposite inequality: given any e > 0, there exists a sequence of open /»-cubes Ij such that diam/,• < S, vn(A) < (one first covers A with open U such that vn({/) < vn(i4) + e/3, then fills U with closed n-cubes of diameter < S, and then thickens them slightly). Now to each Ij we have pairwise disjoint closed disks (D,,*) with vn(/; \ U^ ";.*) = 0. which implies, by Remark IV. 1.1, Therefore MJ 2" . (the next to last inequality follows from the fact that each D,,* covers itself and has diameter < 8), which implies the theorem. ¦ IV.1.2 Example. Riemannian Manifolds Theorem IV.1.7 UtXbea separable, locally compact topological space that admits an exhaustion by compact sets, and let (X,dx)and(X, d2) be two metric structures on X, whose metric topologies coincide with the original topology, and which satisfy lim m >-«0 d\{x,y)
108 Hausdorff Measure and Perimeter locally uniformly. Then Hausdorff measure is independent of the choice oj metric. Proof Let K be compact in X. Given e > 0, there exists p > 0 such that d1(x,y)<p =» l-€<d^lll<l+€ di(x,y) for all x, y e K. Then any 5-cover of K with respect to d\ is a A + eM-cover of /if with respect to d2, from which we imply (with obvious notation), which implies Hs(K;d2)<H*(K;dt) for all compact K. By switching the roles of d\ and di we obtain the opposite inequality, which implies equality for all compact K. One easily obtains the equality for general subsets of X. ¦ Theorem IV.1.8 Let M be a Riemannian manifold, with associated Rieman- Riemannian measure dV. Let dH"'M denote the n-dimensional Hausdorff measure associated with the Riemannian distance in M. Then (IV. 1.1) on all of M. dV = -^ Proof (For the basic definitions and facts about Riemannian manifolds, see the summary in § V. 1 below.) Let M be a Riemannian manifold, x: U -*¦ R" a chart on M. Then the Riemannian measure on M is given in the chart by dV = y/gdxl---dx", where the local metric {, ) is given by the positive definite symmetric matrix gij = <3/3x', d/dxj), g = det (gij). On any compact subset of U, the Riemannian measure dV and the local Euclidean measure dx = dx1 • • - dx" are equivalent, in the sense that each is absolutely continuous with respect to the other. Fix xo e U, and determine Riemann normal coordinates based at xq. Then 8'j(x) -*¦ &U' ti* Kronecker delta, as x -*¦ xo. In particular, y/~g(x) -*¦ 1 as x -*¦ xo [(V.1.6) below]. Moreover, for any p > 0, there exists a sufficiently small ro > 0 such that B(xo\ ro) C U and such that, for any vector ? in the IV1 Hausdorff Measure 109 tangent bundle of B(xq; r0), the Riemannian length |?| of ? and the Euclidean length 11A1 of ? induced by Riemann normal coordinates based at x0 satisfy Furthermore, ro can be picked sufficiently small so that the geodesic path reali- realizing the distance between two points of B(x0; ro) is completely contained i B(xo; ro). Therefore, in , y) (with obvious notation) for all x, y e B(xo; ro), which impli lim We first conclude ies c, y) A + p)'sHsV < HsM < A 4-Y on 5(jco; ro). Therefore, dHnM is equivalent to dHnK>, which is equivalent to dV. Therefore there exists a positive function 0 on B(xo; ro) such that dH"-M = Since ^/g(x) -> 1 as x -*¦ xo, we have 2" But xo is arbitrary in M, and d V and dHnM are independent of the choice of any local coordinate system. We therefore have (IV. 1.1) on all of M. ¦ Corollary IV.1.1 Let X be a k-dimensional submanifold of the n-dimensional Riemannian manifold Y. Then for any measurable A C X we have, by an application of Theorems IV. 1.7 and IV. 1.8, Vk(A) = ~ CX)= ~ c Y) *vhere d Vj denotes the (k-dimensional) Riemannian measure of the Riemannian Manifold X, endowed with its induced Riemannian metric.
110 Hausdorff Measure and Perimeter IV.2 The Area Formula for Lipschitz Maps Theorem FV.2.1 Let Xbea separable metric space, with family of subsets containing all Borel sets, and monotone and countably subadditive $:!F [0, oo]. Let \/f be the Hausdorff measure determined by ?. Then = sup V {(«), bTh for all if/ -measurable A in X, where H varies over all partitions of A by Borel sets. Also, if H\, Hi,. ¦. is a sequence of partitions of A by Borel sets satisfying lim sup diam B = 0, then lim V f(«) = 1r(A). Proof Let S be a Borel set. If (Ej) is a <5-cover of S by Borel sets, then < 5Z; <(?¦>). which implies t(S) < fo(S) for all 5 > 0, which implies <(S) < V Borel 5, which implies BeH BeH for all Borel partitions H of A (the equality is where we use the Vr-measurability of A). Set Sj — sup diam B, Sj -> 0. Given e > 0, there exists jo such that 7 > jo implies by the very definition of the measure \foy. Therefore we also have which implies the theorem. IV.2 The A reea Formula for Lipschitz Maps 111 Theorem FV.2.2 Let X be a sseparable metric space, (Y, fi)a measure space, and assume the map f: X -*¦ Y takes every Borel set B in X to a measurable set in Y. Let ? be the collectioon of Borel sets in X, define the outer measure = n{f{B)) on T, and let ty be the inducced Hausdorff measure on X. Then for any y(r- measurable set A in X we havee = [f aad(ADf-l[y])dn(y). JYY Proof Pick a sequence of Boreel partitions H\, H2,... of A such that Sj = sup diam B, lim Si = 0, and such that, for every j, each 1 Borel set in //, is the union of some subfamily of Hj+\. Then for any y 6 Y wve have as j —*¦ 00, which implies +{A) = Unm = linm = linm ¦/~>00° ((B) BeHj / = f icud(Anf-l[y])drty)- Theorem IV12J Let X be a coomplete, separable metric space, Y a metric space, and f: X -*¦ Y Lipschitz.; Then nm(f(A)) < I card (A in f'^y}) dHm{y) < (Lip f)nHm(A) for all Borel sets A in X. Proof By Proposition 1.3.2, thee image of every Borel set B in X is Hm- measurable in Y, We may then ddefine, as above,
112 Hausdorff Measure and Perimeter for any Borel set B in X, and let V denote the induced Hausdorff measure on X. Then for a sequence of Borel partitions Hi, H2,... of A such that Sj = &upBeHj diam B, lim Sj = 0, and for every j, each Borel set in Hj is the union of some subfamily of Hj+i, we have = lim j—>00 < (Lip/)m lim = (Lip/)m?T(A), that is, +(A) < (Lip f)mHm(A) for all Borel sets A, which implies dHm(y) < f cud (A n /- ) ^K for any Borel set A in X. Proposition IV.2.1 (a) Assume L and a are linear transformations qfRm jucA f/wf for all ?. Then |deti-| < |dets|. ^ Any //near transformation T: Rm -»¦ R" may be factored as T = hog, when geGLQC), h a O(Rm;R") [where GUW) denotes the general linear group ofRm, and ?>(RW;R") denotes the orthogonal linear mappings ofRm to R"]. Lemma IV.2.1 Given a continuous map f: Rm -> R", m < n, consider Uj = {x: J/(x) exists and is univalent], where J/(x) denotes the Jacobian linear transformation of f at x. Then, for any real number k > \,Uf has a countable covering G by Borel sets such that, for every Borel set E e G, IV.2 The Area Formula for Upschitz Maps (i) f\E is univalent, (U) there exists a linear automorphism *: R« _» R-» such tha( -') < A, LipCo(/|?)-i) < < |det Jf(x)\ < 8 " to its image Jf(xXRm) in R". Proof Pick f > 0 such that A~'+e< 1 <A-e, and fix a countable dense subset <S of GL(Rm) Therefore, if x, , e Z(a, A)> |y _ x| < - JO| - 1 ]S(y - t0 Jf(x) = hog, where? « e GL^R), A Then e O(r;R") = \8d)\ Since <S is dense in GL(R-"). we rmay pick , e «S such that y j 3 for this choice of a,, satisfies (i) above. Furthermore, by the existence of
114 Hausdorff Measure and Perimeter J/(x), there exists a positive integer k such that \f(y) - fix) - Jf(x)(y - x)\ < -4-1? - x\ II8 II for all y e B(x; \/k), which implies I/OO - fix) - Jf(x)(y - x)\ < e\8(y - x)\ for all y e B(x; l/k), which implies that x e Z(s; k). ¦ Theorem IV.2.4 (Area Formula) Let f: Rm -»• R" fee Lipschitz, m <n, and A a Lebesgue measurable subset o/Rm. Then (IV.2.1) /* |det //(*)! d\m = f card(/l n /"' JA JK" Proof If vm(i4) = 0 then, certainly, both integrals vanish (the right hand side by Theorem IV.2.3). First assume that A c Uj, and choose the covering G of A given by the lemma. Pick a Borel partition H of A such that each B in H is contained in some E in G. Then one has a map B i->- a e 5, for which k-mHm(8(B)) = A-m|deta|vm(fl) < / |detJf\d\m JB < Am|deta|vm(B) that is. We also have which implies \-7mnm{f{B)) < f \dtt Jf\d\m < A^W JB If we sum B over the Borel partition H, then Theorem IV.2.2 implies A f card(Anf-l[y])dHm(y)< f \tetJf\d\m Jtt" JA ^k2 f card(Anf-1[y])dnm(y). Jtt" Since k > 1 is arbitrary, we have (IV.2.1) when A c U/. k-mHm(8(B))< f \detJf\d\m<kmHm(8(B)). JB < Hm{f{B)) < kmH(s(B)), 1V.2 The Area Formula for Lipschitz Maps 115 Now assume A C [x: dimker Jf(x) > 0} = (x: det Jf(x) = 0). For any e > 0 write f = P°g, where p:R"xr^r, p(y,z) = y. Then jc e /i implies which implies that 5 and Jg(x) are univalent. Also, we have II-WII < Lip/ +6.f since ker J>(x) # {0}. Since /,,(*) is univalent, we have from the first half of me proor that = ? |det Jg\ dym < e(Lip/ + Let e -> 0. Then wm(/M)) < liim sup / [det 7, | </vm = 0 which implies the theorem. ^f P"* the Lipschitz hypersiurface given by on Then
116 Hausdorff Measure and Perimeter Proof One directly calculates .A* to be given by 3d> The discussion of the perimeter measure of 3fi is found in Example II1.4.3. IVJ Bibliographic Notes There are many excellent treatments of Hausdorff measure, especially in dynamics. However, our emphasis is toward its employment in geometric measure theory, for which we have already referred the reader (in the previous chapter) to Federer A969), Morgan A988), Simon A984). §IVJ Our treatment follows Federer A969, pp. 241 -244). Isoperimetric Constants In this chapter we change our venue: we move from Euclidean space to the broadercollection of Riemannian manifolds of bounded geometry. Thus we skip over the intermediate levels of generalization, namely, the sphere and hyperbolic space of all dimensions, symmetric spaces, and homogeneous spaces. Even in these intermediate spaces, one gives away the abelian translation group of isometries and the homotheties that play such an important role in Euclidean space. Therefore one requires different methods from those we used earlier. In the more general setting of this half of the book, we also have to change our point of view with regard to the questions that we ask. In a general Riemannian manifold, the chances of finding the domain of mini- minimum boundary area, given the volume of the domain in advance, are essentially nil except maybe in some very special cases. Furthermore, different choices of the prescribed volume may change the whole character of the problem. Thus, from the analytic perspective, we shall not be able to ask for the precise infimum of the functional D A(dD) (where n is the dimension of the manifold); nor is it obvious that this is the correct analytic functional to study. Rather, we shall ask the following: Find v e [ 1, oo] such that the isoperimet- isoperimetric functional D A(8D) ) is bounded away from 0. (For v = 1 the functional is D t-*- A(dD), and for v = oo the functional is D t-*- A(dD)/V(D).] And it will suffice to let D vary over relatively compact domains with smooth (that is, C°°) boundary. 117
118 Isoperimetric Constants We shall have to adjust even this formulation of the problem to take ac- account of the difference between the local Euclidean character of Riemannian manifolds and their varied global behavior. This distinction will allow us to con- consider discrete isoperimetric inequalties on graphs and, using the hypothesis of bounded geometry, will allow us to compare global isoperimetric properties of a Riemannian manifold with those of its discretizations (especially, Theorem V.3.1). In subsequent chapters, we shall continue this theme, employing it in the study of large time heat diffusion in Riemamannian manifolds. But more of that later. In this chapter, we mostly present just a summary of the basic background, nearly all of which is considered in Chavel A994). Detailed proofs are given for results not covered there, or for results that are at the heart of the point of view presented here. V.I Riemannian Geometric Preliminaries In this section we describe the universe we inhabit for the rest of the book- Riemannian manifolds. The necessary definitions and results are given that enable us to discuss the type of isoperimetric inequalities in which we are in- interested. Whenever we refer to a manifold, unless otherwise noted, we only have in mind the interior. Also, unless otherwise noted, it is C°°, Hausdorff, with countable base, and connected. The boundary, when it exists, is usually mentioned sep- separately. (We do not assume that the boundary is connected.) Thus a compact manifold has no boundary. We are given an n-dimensional manifold M,n > 1. For any point p e M, we denote the tangent space to M at p by Mp; and we denote the full tangent bundle by TM, with natural projection n : TM -> M. If x: U -»¦ R" is a chart on M, we denote the natural frame field associated with x by 01,..., 3n), dj = —j. At each point p of U, the frame is the natural basis of Mp. A Riemannian manifold is a manifold with an inner product on each of its tangent spaces, denoted by {, >; the assignment of inner products, referred to as a Riemannian metric on M, is C00 in the sense that if X, Y are C°° vector fields on M, then the function {X, Y) is a C°° real-valued function on M. For any f e TM the length of ?, |?|, is defined by |?| = <$, i V.I Riemannian Geometric Preliminaries 119 Given Riemannian manifolds (M, g), (N, h), where g and h denote the re- respective Riemannian metrics, one naturally considers the product Riemannian structure on the product manifold M x N,as follows: For x e M, y e N the tangent space (M x AO(x,y) is canonically isomorphic to MX(B Ny. For vectors ?, r) 6 Mx, f,ne Ny, we define the inner product of ? © ? and r\ © v by , v). © f, ij © v) = Also, given a Riemannian manifold (A/, g) and an imbedded submanifold N in M, then N carries a natural Riemannian metric g\N obtained by restricting the inner product on tangent spaces of M to the tangent subspaces of W in M. Let x: U -*¦ R" be a chart on the fixed Riemannian manifold M. Then for each p e U, the matrix G(p) given by (V.I.I) = (gij(p)), gij(p) = <%,„ 3;,,), is positive definite symmetric, and the functions gy :U -*¦ R, i, j = 1,..., n, are C°° on U. We also use the notation (V.I.2)  = (g}k), g = det G. Let Dl denote the continuous, piecewise C1 paths in M. For any path o>: fa, P] -*¦ M e ?>', a>' denotes the velocity vector of a> at values oft where w is C1, and the length of a>, l(a>), is defined by /(»)=/ \<o'(t)\dt. Ja For M connected (our usual assumption), p,q e M, we define the distance between p andq, d(p, q), by d(p, q) = inf ?(a>), where o> ranges over all paths (a:[a, fi] -*¦ M 6 ?>' satisfying <o(a) = p, &{(i) = q. Then (/defines a distance metric on M. A path}/: (a, /?) -»¦ M e Dl, \y'\ = const., is called a geodesic if for every t e (or, ft) there exists an e > 0 •such that for all [oro, fio] C(t -e,t + e). That is, geodesies locally minimize distance. It is known that any geodesic y must in fact be C°°. Furthermore, it is known that given any ? 6 TM, there exists one and only one maximal geodesic y( : (-or*, ft) -*¦ M, at, fr 6 @, +oo], satisfying ^@) = n{i-) and ^'@) = ?. Of course, \y'\ = |?| on all of (-or*, fa). We write I( = (-af,fy).
120 lsoperimetric Constants Let TM denote the subset of T M consisting of those ? e TM forwhich 1 e I(. Define the exponential map, exp : TM -*¦ M, by exp ? = Then for every t e R, ? 6 TA/, for which f? e TM, we have exp tl; = Notation For every p e M we let expp = exp|(TA/ n A/,,). For every p e M and r > 0, we let B(p;r) denote the open n-disk in Mp centered at the origin of radius r. Recall that, for every p e M and r > 0, B(p; r) denotes the metric disk in M centered at p of radius r. For any p e M there exists an e > 0 such that expp is defined on B(p; 6) and is a diffeomorphism of B(p; e) onto the metric disk B(p; e). Furthermore, if for any r > 0 we have expp defined on all of B(p; r), then exp, B(p;r) = B(p;r); one just cannot guarantee that exp | B(p;r) is a diffeomorphism. The injectiv- ity radius of p, inj p, is defined to be the supremum of all r > 0 for which expp | B(p; r) is a diffeomorphism. By our first remark, inj p > 0 for every p. The injectivity radius ofM, inj M, is defined to be the infimum of inj p, where p varies over all of M. The Riemannian metric also has a local convexity property, namely, given any p e M, there exists ro e @, inj p] such that for any two points u,v € B(p\ro) there is a unique length-minimizing geodesic in M joining u to v, and it is completely contained in B(p; ro). Let M be a Riemannian manifold. We say that M is geodesically complete if for every ? e TM, the geodesic y$ is defined on all of R, that is, if exp is defined on all of TM. Proposition V. 1.1 If M is geodesically complete, then every closed and bound- bounded subset is compact. As a consequence, M is a complete metric space. IfM is a complete metric space, then M is geodesically complete. IfM is complete, in either of the two meanings, then any two points of M can be joined by a minimal geodesic. V.I.I Connections and Curvature We must first consider the Levi-Civita connection on M. Of course, differen- differentiation of functions on a manifold is determined by the differentiable structure V.I Riemannian Geometric Preliminaries 121 alone. But the differentiation of vector fields is not uniquely determined. A con- connection is a rule for differentiating vector fields on a manifold. Let P°°(TM) denote C°° vector fields on M, that is, the C°° sections in TM. Then a con- connection on Mis a map V:TM x r°°(T M) -^ TM, which we write as V$Y instead of V(?, Y), with the following properties: First we require that be in the same tangent space as ?, and that for aJeR,p€^?-16 Y 6 r°°(TM), Second, we require that for p 6 M,$ e MP,Y,YU Y2 6 r°°(TM)>f e C°°{M), we shall have v*(y, + y2) = vty, + vty2, vt(/y) = ($/)y,, + /0»)vty. Finally we require that V be C°° in the following sense: if X, Y e r°°(rAf) then Vxy e r°°(rA/). A connection is a local operator in the following sense: Vt Y is uniquely deter- determined by the values of Y in a neighborhood of p = n (?); in fact, it is determined by the restriction of Y to a path in a neighborhood of p that passes through p and which has velocity vector ? at p. If M is a Riemannian manifold, then there exists a unique connection V (henceforth called the Levi-Civita connection) for which (V.I.3) (VI.4) for all differentiable vector fields X, Y,Z e T(TM). We shall always work with this connection. M is a Riemannian manifold with Levi-Civita connection V. The curvature tensor, R, of M is defined by R(X, Y)Z = VYVXZ - - VlY.X]Z, L *'/' Z'W e r(™)- One has that R(X, Y)Z at the point p depends «nly on the values of X, Y, Z at „, and is trilinear on Mp. More generally, if «(«*, WUZ) = a/?e*(X, y)z. Of course, and(V 1.3) implies R(Y, X)Z = 0; R(X, Y)Z + R(Z, X)Y + R(Y, Z)X = 0.
122 Isoperimetric Constants One also has {R(X, Y)Z, W) - {R(Z, W)X, Y) = 0, {R(X, Y)Z, W) + {R(X, Y)W, Z) = 0. Since the curvature tensor vanishes identically for dim M = 1, all discussions concerning the curvature tensor will assume dim M > 2. To define the sectional curvature, one defines, for p e M, ?, r? e Mp, f, if). . if). where If |, r; are linearly independent tangent vectors in A/,,, then is well defined and only depends on the 2-dimensional subspace determined by ? and r}. We refer to /C(f, 77) as the sectional curvature of the 2-section determined by ?, r?. We note that if G2 is the complete collection of all 2-dimensional spaces tangent to M, then G2 can be provided a C°° structure in a natural manner, and K.: G2 -*¦ R will then be C°°. For p e M, the Aicci tensor Ric: Afy, x Afy, -> R is defined by Ric (?, if) = trace (f m In particular, we have for any orthonormal basis of Mp, [e\,..., en} Thus Ric is a symmetric bilinear form on Mp. To calculate its associated quadratic form (referred to as Ricci curvature), pick [et,..., en} so that en = So, for any unit vector ?, Ric (|, |)/(n - 1) is the average sectional curvature of all 2-sections containing ?. The Riemann and Ricci tensors measure how the Riemannian metric and the Riemannian measure, respectively, of M differ at the infinitesimal level from the Euclidean metric and measure (see §V.1.2 below). It is expressed as V.I Riemannian Geometric Preliminaries 123 follows: Fix p e M and U = B(p; inj p), U = B(p; inj p). Then every choice of orthonormal basis [e\ en} of Mp determines a chart n: U -*¦ R", referred to as Riemann normal coordinates, given by for q e U, that is, for v = v e U we have j e U we have n'(expv) = vJ. Then for = SJk - \{R(v, ej)v, ek) + O{\v\\ (V.I.6) det^(exp v)) = 1 - |Ric(i>, v) + O(\v\3) as u -»¦ 0. If Af 1 and M2 are Riemannian manifolds, we say they are isometric if there exists a diffeomorphism <p : M\ -*¦ M% such that 10.1 b = If li for all I e TA/| (the subscripts indicate in which Riemannian manifold the length is being evaluated), where <p* denotes the Jacobian map from TM\ to TM2. The map <p is referred to as an isometry. Let M be a Riemannian manifold of dimension > 2, AC the Riemann sectional curvature of 2-dimensional spaces tangent to M. We say that M has constant sectional curvature k,k € R,if JC(<r) = *• for all 2-sections <r. One has the three standard model spaces of simply connected complete Riemannian manifolds of constant sectional curvature. For k = 0 one has Euclidean space R", with its standard Riemannian metric. Of course, the geodesies are the straight lines in R". When k > 0, the n-sphere Sn(l/y/ic) in Rn+1 of radius \jyfic has con- constant sectional curvature k. The geodesies ofSn(l/y/ic) are given by the "great circles," the intersection with Sn(l/V*) of 2-planes in Rn+I that pass through the center ofSn(l/y/ic). When k < 0, we have hyperbolic space given by the ball model, namely, on the n-disk B"A /y/—ic) we define the Riemannian metric 2 _ ~{ ) then the sectional curvature is constant equal to k < 0. One easily sees that, in this model, the geodesies emanating from the origin are given by straight lines emanating from origin, and their length to the boundary Sn~'(l/V-^) is infinite. The rest of the geodesies consist of the intersection with B" A /^J^ic) of those circles in R" that orthogonally intersect the boundary sphere By the Hopf-Killing theorem, a complete simply connected Riemannian manifold of constant sectional curvature k is uniquely determined up to
124 hoperimetric Constants isometry. In particular such a space is isometric to the appropriate model among those discussed above. We always refer to the model space as the model space MK of constant sectional curvature k. For convenience, we define the function [ I sin k > 0, k = 0, k < 0. Of course, <p(r) = SK(r) satisfies 0, <p@) = 0, <p'@) = 1. Let M, be the model space of constant sectional curvature k. It is standard that for any x e M, unit vector ? e Mx, and r > 0 (that is, when k < 0 - otherwise, we have to restrict r e @, n/y/ic])y we have (V.I.7) ds2(exp r|) = dr2 where \di-\2 denotes the Riemannian metric on the (n — l)-dimensional unit tangent sphere (henceforth denoted by) Sx in Mx. V.I.2 Volumes of Disks and Areas of Spheres We now consider the Riemannian measure dV on M. Let x: U -*¦ R" be a chart on My with the Riemannian metric given in local coordinates by (V.I. 1) and (V.1.2). Then the local measure dV = V*°x~' dvn(\) = yfgdx1 ¦dx" [where g is given by (V.1.2)] is well defined, that is, the integral /(/;«/)= I (fVg)ox-1 dvn(x) depends only on / and U - not on the particular choice of chart x. One then de- defines a global Riemannian measure on M, also denoted by dV, using a partition of unity. i Definition Given any x e M, we let Dx denote the largest starlike (relative to j the origin) neighborhood of the origin of Mx on which exp is a diffeomorphism.' For any ? e Sx, we let e Dx V/e [0, r)}. V.I Riemannian Geometric Preliminariy es So 125 For any x e M, r > 0> w ^ ^ ^ ^ ^ those elements « for which r« e D,. which ri * we write as wnte the Riemannian measure on Dx as where rfMx($) denotes the Riemannian measure on S, induced by the Euclidean Lebesgue measure on Mx. We also let V(x; r) denote the volume of B(x- r) S(x-T^ M' ^ 1Ct ^^ den°te *" /OVVer are° «the •Xjc, r), with center jc and radius r, that is, and we let *(,; r) denote ;r):=/" mefrfc jpAere S(jr; r)> the (n - l)-dimensional Hausdorff measure of S(x;r). Proposition V.I 2 One has V(x;r)= ['nix;! Jo ;t)dt and V(x;r+€)-V(X;r) limsup __^ < %{x.r)<A(x;r) for all r > 0. Finally, A(x;r) = 2l(jc; r) for almost all r. K/. J Curvature and Volume Recall^ c^ denotes the („ - l^dimensional Riemannian measure of the rim"/1 Unit SpherC S""'in R"- F«^ a real constant, and M the TJi7z?rp]cX? Riemannian manifoid °f constant -«^ -- above, let rfy, = ^fr; I) rfrrf,.,«) denote the Riemannian
126 Isoperimetric Constants measure on M» in geodesic spherical coordinates described above, and VK(r) the volume of the metric disk in MK of radius r. Then ), VK(r) = c._, f SKn~ Jo t)dt. Bishop's comparison theorem states that if M is an n-dimensional Riemannian manifold, x e M such that B(x; R) c TM [that is, the exponential map is defined on all of B(x, R)], and for all ? e A/,, q e B(x; R), then (V.I.9) v/• which implies (V.I.10) y/g(r,$)<SKn-l(r\ for all r e @, min {R, c(?)}), and (V.I.I 1) V(*;r) < V,(r) for all r e [0, R]. In particular, (V.I.9) implies that the function that is, is nonincreasing with respect tor on@, min {R, c(?)}). Gromov's refine- refinement of Bishop's comparison theorem states that, under the same hypotheses, the function (VI. 13) r»V(x;r)/VK(r)l, that is, is nonincreasing with respect to r on [0, R]. V.I.4 LiouviUe's Theorem On the unit tangent bundle of M, SAf, one has the geodesic flow, <t>t, defined by that is, <!>,? is the velocity vector of the geodesic y$@ = exp f? at the point ). Since y^'(t) has the same length as ?, <!>, indeed maps the unit tangent V.2 Isoperimetric Constants 127 bundle to itself. The kinematic density is the measure dfi on SAf giv / Fdf*= f dV(x) f JSM JM JS. for functions F on SAf. given by Js, U3 (Liouvffle's Theorem) Themeasure dponSAi is invariant respect to the geodesic flow. sonant V.2 Isoperimetric Constants In this section we define the basic apparatus of isoperimetric constants that we use in the sequel. We mention some of the most elementary examples that guide the intuition, and we discuss the necessity of introducing a modified apparatus of constants in order to deal with the dichotomy between local and global considerations in the geometry of Riemannian manifolds. Probably the most important result for our point of view here is Theorem V.2.4, which proves the invariance of positivity of modified isoperimetric constants under compact perturbations of a complete Riemannian manifold. V.2.1 Definitions and Examples Definition Let M be an n-dimensional Riemannian manifold, n > 2. Denote by V its n-dimensional Riemannian measure, and by A the Riemannian measure associated with (n - 1 )-dimensional submanifolds of Af. For each v > 1 define the v-isoperimetric constant ofM, 3V(M), to be the infimum where il varies over open submanifolds of M possessing compact closure and C°° boundary. For v = oo define Cheeger's constant, 3X(M)> by above. where n ranges over open submanifolds of M described Remark V.2.1 If M is compact, then, by considering M\B(x;e) for small « > 0, one can easily show that 3V(M) = 0 for all v. Therefore, we shall restrict our interest to the case of noncompact manifolds. One can adjust the definitions
128 Isoperimetric Constants of isoperimetric constants for the compact case and obtain a rich theory. See our comments in § VI.2 below. Also, see the remarks in the Bibliographic Notes to Chapter VI (§VI.6). Definition Let <f>: M -*¦ N be a diffeomorphism between Riemannian mani- manifolds M, N with respective metric tensors g, h. We say that ^ is a quasi-isometry if there exists a constant c > 1 such that for all ? e TM, the tangent bundle of M. Two metric tensors g,h on a Riemannian manifold M are quasi-isometric if the identity map \dM : (A/, g) -*¦ (M, h) is a quasi-isometry, that is, if there exists a constant c > 1 such that Remark V.2.2 The property 3V(M) > 0, v e [1, oo], is invariant under quasi- isometry. Example V.2.1 The first, R", n > 1, as discussed in previous chapters. Example V.2.2 The hyperbolic spaces, M», k < 0. Then the Cheeger constant 3<»(M*) = (» — 1 )¦>/-*, where n = dirnM,. The proof goes as follows: Let M be arbitrary Riemannian, x e A/, and r, ? geodesic spherical coordi- coordinates on Dx. Let dr denote the radial vectorfield on Dx, given by Then it is standard that the Riemannian divergence of dr (see §VII.2.1 below) is given by When M = M* then div3r(exp r%) = ;L, _^ • V.2 Isoperimetric Constants 129 When *: < 0, then c(?) = +00 V ^ 6 Sx - so D, = M,. Therefore any relatively compact Q CC D,. Furthermore, diva, >(n- By the Riemannian divergence theorem, A(dto) > / {dr, v)dA = if di\drdV > (n - l^/^ where v denotes the exterior unit normal vectorfield along 3 ?2. The lower bound > (n - is sharp, because one has A(x;r) (n - ^ic asr -*¦ +00. As mentioned in Remark 1.1.4, if M = Mj, the simply connected Rieman- Riemannian 2-manifold of constant Gauss curvature equal to k, then the isoperimetric inequality is A.1.6): L1 > An A - kA2, where A denotes the 2-dimensional measure of the domain, and L the 1-dimensional measure of the boundary, with equality if and only if the do- domain in question is a metric disk. When k < 0, the inequality A.1.6) is sharp in two senses, because A.1.6) implies both the inequalities L/Al/2 > y/4Jr, LI A > yf^ic. The first is sharp for geodesic disks of radius r as r X 0, and the second is sharp for geodesic disks of radius r asr ^ +00. Thus, one has distinct isoperimetric behavior when discussing the local prop- properties of the manifold versus the global ones. Indeed, one expects all local isoperimetric behavior to be Euclidean in character [see (V. 1.5)], and the global behavior to be quite diverse. Remark \23 The inequality 3V(M) > 0 is only possible for n < v < 00. Indeed, let v < n, and consider small geodesic disks B(x; e), with center x e M and radius e > 0. Then for the isoperimetric quotient of B(x; e) we have as e I 0; thus 3V(M) = 0 whenever v < n. So it seems at first glance that one only has a discussion of isoperimetric constants for v > n = dim Af.
130 Isoperimetric Constants Definition For each v > 1 and p > 0 define the v-modified isoperimetric con- constant ofM, 3v,p(A/), by = inf where ft varies over open submanifolds of M possessing compact closure and C °° boundary and inradius greater than p, that is, ?2 that contain a closed metric disk of radius p. Definition For v = 1 and p > 0 define the I-modified isoperimetric constant ofM, 3,,P(A/), by where ?2 varies over open submanifolds of M, described above, with inradius greater than p. For v = co and p > 0 define the modified Cheeger constant, 3<»,p(A/). by . , A(dQ) 18f where ?2 ranges over open submanifolds of M, described above, with inradius greater than p. Remark V.2.4 Given v e [1, co], then the existence of p > 0 for which 3v,p(A/) > 0 is invariant under quasi-isometry. Example V.23 Consider the Riemannian product M = Mo xRk, where Mo is an (n - &)-dimensional compact Riemannian manifold. Then h(M) = 0, as noted in Remark V.2.3. Yet, for extremely large domains - for example, metric disks of large radius - one expects the volume of these domains and the area of their boundaries to reflect /^-dimensional space. We shall see (Example V.3.6 below) that 3V,P{M) > 0 for any p > 0 for all v e [1, *]. Example V.2.4 The 2-dimensional jungle gym JG2 is constructed by (a) con- considering the integer lattice Z3 in R3, (b) connecting points, for which precisely one of their coordinates differ by 1 and the other two coordinates are equal, by a line segment parallel to the coordinate axis, (c) considering the surface consisting of all points with distance from the 1-dimensional network to be precisely equal to e, for some given small e > 0, and finally (d) smoothing out V.2 Isoperimetric Constants 131 the corners in a bounded periodic (that is, Z3-invariant) fashion. We shall see (Example V.3.7 below) that 33(JG2) > 0. The verification of the positivity of the isoperimetric constants in the last two examples requires discrete isoperimetric inequalities for graphs, which we con- consider later. V.2.2 Bask Results Remark \2? It is an easy consequence of Minkowski's inequality that for domains ft;, j = 1,..., Af, in M, and any * > 1, we have From this Yau A975) has shown that in the definition of 3V(M), v e A, co], it suffices to let ?2 range over open submanifolds of M that are connected. Definition For each v > 1, define the Sobolev constant 6V(M) ofM by where / varies over C~(A/). For each p > 0, let C™p(M) consist of those compactly supported Lipschitz functions 4>onM for which (a) there exists an x € M such that the preimage of max |^|, n*, satisfies fy 2 D(x;p), and (b) $ \ M\fy e C°°. For each v > 1 and p > 0, define the modified Sobolev constant of M, 6V,P(M), by (V-21) 6v.fi(M) = inf where / ranges over CC°°O(A/). Remark V.2.6 Even for GV(M) we may (by Theorem 1.3.3) allow / to vary over compactly supported Lipschitz functions on M. Theorem V.2.1 (The Federer-Fleming Theorem) The isoperimetric and Sobolev constants are equal, that is. (V.2.2) MM) = 6V(M),
132 Isoperimetric Constants and the modified isoperimetric and modified Sobolev constants are equal, thai is. (V.2.3) 3v.fi(M) = 6v,fi(M). Proof The proof of (V.2.2) is the same as the proof of Theorem II.2.1, with the substitution of v for n. The argument for (V.2.3) is a slight modification of the proof of Theorem II.2.1. ¦ Theorem V.2.2 Suppose, for a given v e A, oo), we have 3V,P(M) > 0. Then for the area and volume of metric spheres and disks we have (V.2.4) A(x;r)>3v,p(M)V(x;r)i- for all x e M and r > p. For the volume of metric disks we have (V.2.3) V'(x;r)>3v,p(M)V(x;rI- (where the prime denotes differentiation with respect to r)for almost all r > p. In particular, (V.2.6) liminf V(jc; r)r~v > 0. rfoo Proof The point, of course, is even when r > inj x. Assume r > p. For e > 0 define the function rf: [0, oo) -> [0, 1] by (a) t((s) = 1 when s e [0, r], (b) r((s) = (r + e- s)/e when s 6 [r, r + e], and (c) x( (s) = 0 when s > r + e. Also define the Lipschitz function f(:M^> Rby My) = r((d(x, y)). Then Now let € I 0. Then which implies the first claim. V.2 Isoperimetric Constants 133 The second claim follows from the fact that A(x; r) = V'(x\r) for almost all r. Now integrate (V.2.5) to obtain (V.2.6). ¦ A similar argument, using the corresponding Federer-Fleming theorem for v = oo, shows Theorem X23 Suppose 3OO(M) > 0. Then for the volume of metric disks we have for all x e M (V.2.7) for almost all r > 0. In particular, (V.2.8) V'(x;r)>300(M)V(x;r) lim inf V(x; rfoo > 0. Proposition VJ.1 If3v.fi(M) > Ofor some v 6 [1, oo), P > 0, then V(x; p + 6) is uniformly bounded from below for all x with d(x, dM) > p + e, for any 6 > 0. /Ifco, I^+^M) > 0 for all (x in [1, v). Proposition V.2.2 Let SI be a domain in M with compact closure and C°° boundary. Then 3OO(&) > 0. Theorem V.2.4 Let D be a relatively compact domain in the Riemannian complete M with C°° boundary V, and D' an n-dimensional Riemannian manifold with compact closure and C°° boundary F, such that M' given by M' = {M\D} U D' is C°° Riemannian. If3v,p(M) > Ofor given v e [1, oo) and p>0, then there exists p' > 0 such that 3V^(M') > 0. Proof (See Figure V.2.1.) Suppose we are given any a 6 @, oo). The number « will be fixed throughout the argument, although its estimated value will be determined as we go along. Since 3ViP(M) > 0, there exists R > p such that for all x e M. Let 8' denote the diameter of D'. We pick p' = 2p + R + 8'. Let E = M \ D, and suppose we are given ft' c m\ with compact closure and C°° boundary, and J2' 2 B(y'; p').
134 Isoperimetric Constants Figure V.2.1: The compact perturbation. (i) We first assume that d{y\ D') < R. Then B(y'; p') 2 D', which implie SO' c E. Set J2 = (O' \ D') U D. Then 3ft = 3ST, and J2 contains a disl in M with inradius p, which implies We obtain ' n ' n that is, (V.2.9) iP() (ii) We now assume d(y', D') > R. Then B(y';«) c E, which implies (V.2.10) V{Sl')x~xlv > V(y'; «)I-|/p > <M(r). (ii-a) Assume first that V(Q' n ?) > i V(O'). Then n ?) ' n ?)) - a(D D because /? > p. So 135 VT2 Isoperimetric Constants Therefore we pick a at the very outset to also be greater than or equal to 3,.p(A/)/22-'^. We obtain (V.2.11) AidST) > pv.p(M)/22-l/v)V(n')l-l/v. (ii-b) The final situation to consider is, therefore, d(y', D') > R and V(n' C\iy)>{ V(SI') We still have (V210); then , threfore, d(y V(n' C\iy)>{ V(SI'). We still have (V.2.10); then > A(da n d') 'WiSi1 HD')-- V(Q')l-l/v a ^2 I a "I -\/v (the third inequality uses Proposition V.2.2). In addition to the above, pick, at the outset, a sufficiently large so that (V.2 1 Then (V.2.9HV2.12) will imply the theorem. Example VJ.5 The proofs of Proposition V.2.1 and Theorem V.2.4 break down when v — oo - unless we postulate the existence of po > 0 for which (V.2.13) v(x; po) > const. > 0 for all x (the constant independent of x). Moreover, one has a counterexample for Proposition V.2.1. Consider the Riemannian metric on M = R1 x S1 given by Certainly there is no po > 0 for which (V.2.13) is satisfied for all at. At the same time, one easily shows 3,,p(A/) = 0 for all v e [l,oo), p > 0.
136 Isoperimetric Constants V.2.3 Bounded Geometry Proposition X23 (a) (Croke's inequality) Let M bean arbitrary Riemannian manifold. Given any o 6 M, p > 0, such that exp0 is defined on D(o\ p), then for r < - min I inf inj x, p \ 2 [xeB(o,fi) r\ we have (V.2.14) > const. (where const.,, is a constant depending only on n = dim M) for all J2 c B(o;r), which implies (V.2.15) V(o;r) >const.nr" for all r < i min {infx€B(<,:p) injx, p\. (b) (Buser's inequality) Let M be a complete n-dimensional Riemannian manifold, with (V.2.16) *:<0, foralll- 6 TM. Then there exists a positive constant c(n,K,r) depending on n, k, andr, such that for any given x 6 M ,r > 0, and a dividing C°° hypersurface T inB(x;r) withT imbeddedin D(x;r)andB(x\r) \ T = fti U ft2. where J2i, J?2 on open in B(,x;r), we have (V.2.17) min{V(ni), V(ft2)} < c(n,K,r)A(r). Moreover, for any fixed r0 > 0 we have (V.2.18) c(n,k, r) < c(n, *:, ro)r Vre@, ro]. Remark V.2.7 Croke's inequalities (V.2.14) and (V.2.1S) are statements that the local isoperimetric behavior of an arbitrary Riemannian manifold is indeed Euclidean. Definition A Riemannian manifold M has bounded geometry if the Ricci cur- curvature of M is bounded uniformly from below [as in (V.2.16)], and if the injectivity radius of M is bounded uniformly away from 0 on all of M. Theorem \2S If M is Riemannian complete with bounded geometry, then > 0 for every p > 0. Proof Set V.2 Isoperimetric Constants 8 = min{/0, inj A/}. 137 Assume we are given Q as above, containing B(x; p) for some x e M. One easily has the existence of z e M for which 3 J2 D B(z; 8/2) divides B(z; 8/2) into two open subsets for which the smaller volume is greater than or equal to V(z;8/2)/3. Then Buser's inequality (V.2.17) implies A(dS2 n B(z;8/2)) > const.V(z' / \ and Croke's inequality (V.2.15) implies V(z;8/2)>cn8\ which bounds A(dQ) away from 0. ¦ Here is an alternate characterization of modified isoperimetric constants. Theorem V.2.6 Let M bea complete Riemannian manifold with bounded geo- geometry. Then3Vifi(M) > Oforsomev > 1, p > 0 if andonly ifthere exists vq > 0 such that (V.2.19) A(dii) > const. for all domains ft with compact closure and C°° boundary. Proof Assume that (V.2.19) is valid for all ft. Since M has bounded geometry, then for any r0 > 0 there exists vo > 0 such that V(x; r0) > i>o for all x in M. Then 3v,ro(M) > 0. ¦ To prove the converse, we first prove Lemma V.2.1 Let M be Riemannian complete with bounded geometry, and set R = j inj M. Then there exists a positive constant c such that for any domain ft in M with compact closure and C°° boundary, satisfying V(QnB(x;R))< l-V(x;R) for all x in ft (so ft is uniformly thin), we have
138 Isoperimetric Constants V.3 Discretizations and Isoperimetric Inequalities Proof Given such an J2, then for each x e J2 there exists r{x) e @, R) such Proof There exists xo such that that 139 V(x;r(*)). _ Let & _ J2 u B(x0; /?). Then J2' has inradius greater than or equal to R, which Pick a finite collection [xt, r,}, r, = r(xt), such that {?(*,; r,)} are pairwise imphes disjoint and {B(xf;3r,)} cover J2. Then Buser's inequality (V.2.18) and Croke's _ Am) > - ,Wom-i/,. > - wo,i-i/v inequality (V.2.15)imply l J - J**K(") >-V*V(J2)' 1/v. Also, ' ; R), 2const. > const.r,"-', which implies '> R) < const./?" < const./i(aJ2 n B(x0; R)) < const.ACJ2) (the second inequality follows from the same one in the proof of the previous lemma), which implies > const. > const. J^ V(x,; r,I < const.ACJ2), which implies the lemma. ~l/" > const, ;r.) ,_1/n ConclusionoftheProofofTheorem V.2.6 We are assuming that 3, P(M) > 0 for sorae „ > { ^ p > Q ^ ^ = ^^ ^ R% R J ^ ^ Then V(J2> ^ "o implies the hypotheses of Lemma V.2.1, which implies > const. \J2 V(x,;3r,) > const. V(n)l-l/n A(dV) > const. Assume V(Q) >vo.Uv>n, then, by Theorem V.3.2 below, [the second inequality is (V.I.I 1); the third line in Minkowski's inequality; anfl A®O) ~ C°nSt" V&)l~l/v the fourth line follows from (V.I.13)]. [°r a" «• which implies the theorem in this case. Therefore assume v < n If in tntSfiCS ** hyPOthCSeS Of Lemma V22' ^n we ¦« d«ne- Otherwise we are Lemma V.2.2 L«/ M be Riemannian complete with bounded geometry, am Then C&SC ^ Le"lma V21> that n is ""°'formly *in" but has large volume. assume 3v,fi(M) > Ofor some v > 1, p > 0. Let R = \inj M. There exists positive constant d such that, ifQ is a domain in M with compact closure cam /1(an) > const.V(SlI-1'" > const V(m1~l/vvnl/''-l/m = C°° boundary for which there exists x' eO satisfying wh . . * Whlch ""Plies the theorem. Here VJ Dlscretl2attons and Isoperimetric Inequalities is no D^ti"^1106 diSCrete aPProximations t<> Riemannian manifolds. There P^nse here to fine approximation; rather, we highlight what one might J2 ii «o/ uniformly thin), then MO) > • V(a)«-"-.
140 Isoperimetric Constants V. 3 Discretizations and Isoperimetric Inequalities 141 refer to as the coarse geometry of the manifold. First we discuss graph structures in their own right, and an apparatus of measures and isoperimetric constants associated with graphs. And then we discuss them as discretizations of smooth Riemannian manifolds. The most important result is Theorem V.3.1, which allows one to verify the positivity of an isoperimetric constant in one category, by checking in the other. V.3.1 Volume Growth and Graphs Definition Let X and Y be metric spaces with a bijection <(>: X -*¦ Y. We say that # is a quasi-isometry if there exists a constant c > 1 such that e-'dOc,,x2) < d(<t>{xx),<j>(x2)) < cd(xux2) for all x\,X2 inX. Let Q be a countable set such that for each ? e Q we have a finite nonempty subset N(?) C5\{?|, of cardinality m(?), each element of which is referred to as a neighbor oft-,v/'tih\he property that tj e N(?)ifandonlyif? e N(r?)forall i-,T) 6 Q. Then one determines a graph structure G by postulating the existence of precisely one oriented edge from any ? to each of its neighbors, the elements of N(?). We refer to m(?) as the valence ofG at ?. We say that the graph G has bounded geometry if the valence function m(?) is bounded uniformly from above on all of Q. In such a case we let m = max {m(?): ? e Q). A sequence of points (?o, ...,?*) is a combinatorial path of length k if ?; e N(?;_ i) for all j = 1 *. The graph G is called connected if any two points are connected by a combinatorial path. Note that m(?) > 1 for all ? if G is con- connected. For any two vertices ? and rj in the connected graph G, one defines their distance d(?, r?) to be the infimum of the length of all combinatorial paths connecting ? to r}. We also refer to d as the combinatorial metric. We set the notation for the respective metric "disks": for any ? e Q. Example VJ.1 Let F be a finitely generated group, with generator set A. Then consider every element of F to be written as a word of minimum length in the generators in A and their inverses. For the graph structure, given any y eF, we let N(y) = y ¦ {AU A~x] be the neighbors of y. Then the combinatorial metric of the graph structure may be realized by the word metric as follows: Let A := (y\ yk] be a given set of generators of F. With every y e F we associate the word norm ofy,\y\^, defined to be the minimum length of y as a word in the given set of generators A and their inverses. Note that IkU>0, with \0y\a<\P\a for all /J, y in F. The word metric on T is then given by Thus the group theoretic word metric is equal to the graph theoretic combina- combinatorial metric. It is common to refer to this graph as the Cayley graph of F. Note that if we use a different set of generators B := {yj\ .... y*}, then we have the metrics induced by A and B quasi-isometric to each other, namely, Example VJ.2 Let M be a complete Riemannian manifold, and let F be any finitely generated subgroup of isometries M acting freely and properly discon- tinuously on M, such that M/ F is compact. For each x e M let || • \\x denote the displacement norm on F, given by \\Yh=d(x,yx) for all y € F, where d denotes distance in M. Then again we have IML>0, with Wy\\x< for all fi, y in F. Given a set of generators A of F, one has the existence of a constants > 1 such that oTx\y\a< WyWx <a\Y\A for all yeP, which implies that the induced metrics are quasi-isometric. Furthermore, the map ^: F -> M, given by satisfies ,-i , Y) < , y) forall p, Y in r - so * is a quasi-isometry (at the metric level) of r to its image
142 Isoperimetric Constants Definition Let X and Y be metric spaces with map <f>: X -*¦ Y. We say that 0 is a rough isometry if there exist constants a > 1, b > 0, and e > 0 such that <ad(xux2) + b for all x\, x2 in X, and <j> is (-full, that is, jreX Note that the definition of rough isometry does not require that ^ be a bijection, just a map; in fact, the map <j> need not be continuous. Proposition VJ.1 If<j>:Xx -> X2 and \/r:X2^> X^are rough isometries, then so is iff o<f>. If 4>\X -*Y is a rough isometry, then there exists <f>~ : Y -*¦ X a rough isometry, for which both d(<(>~ o#(x), x)andd{$ o#~(y), y) are uni- uniformly bounded on X and Y, respectively Any two spaces of finite diameter are roughly isometric. If X and Y are roughly isometric, then X and Y x K are roughly isometric, for any compact metric space K. Return to the graph G. On the collection of vertices Q we have two natural measures. The first is simply the counting measure di; thus, for any subset K. of Q we have t(AC) = card AC. The second is what we call the volume measure dV on Q, defined by Of course, when G has bounded geometry, the two measures are commensurate in the sense that the Radon-Nikodym derivative of dM with respect to di is uniformly bounded away from 0 and +oo. Because in what follows we generally discuss graphs of bounded geometry, and we are only interested in qualitative estimates on volumes, we shall work with the counting measure di - even when we announce the results in terms of the volume measure d\l. Now denote the collection of oriented edges of the connected graph G by Qe. The oriented edge from ? to tj will be denoted by [?, r?]; and when we wish to consider the unoriented edge connecting ? and rj, we denote it by [? ~ r?]. Any finite subset AC in Q determines a finite subgraph K of G, for which one can describe a variety of suitable definitions for its boundary. Our definition will be that the boundary ofK, 3K, will be the subset of Q, consisting of those oriented edges that connect points of AC to the complement of AC in Q. We define the area measure dA on Qe to be the counting measure for the oriented edges. V.3 Discretizations and Isoperimetric Inequalities 143 Thus, for any finite subset of vertices, the area of its boundary will be equal to the number of edges in the boundary. Another definition of the boundary of a finite subgraph K of G is given by Thus, by this definition, dK is a subset of vertices in the complement of K; its area is defined to be its cardinality. When G has bounded geometry, the two choices of area functions K ¦-> ACK), as functions on the collection of subgraphs K of G, are commensurate each with respect to the other in the sense that the quotient of the two functions is bounded uniformly away from 0 and oo. Therefore, when G has bounded geometry, we will work with the counting measure for the volume of K, and the second definition of dK as a subset of G with counting measure for its area - despite the fact that the theorems are formulated with respect to the original notions of volume, boundary, and area. Definition We say that the Riemannian manifold M has exponential volume growth if In V(x; r) > const.r for sufficiently large r; otherwise we say that M has subexponential volume growth. Also, we say that M has polynomial volume growth if there exists * > 0 such that V(x; r) < const./-* for sufficiently large r 0 Similarly, we say that the graph G has exponential volume growth if In V(?;r) > const.r for sufficiently larger; otherwise we say that Ghassubex- ponential volume growth. Also, we say that G has polynomial volume growth if there exists k > 0 such that V(?; r) < const.r* for sufficiently large r > 0. Example X33 (Example V.3.1 continued.) When one is given a finitely gener- generated group P with k generators A = [yi,... yK), then for the counting function one has where y is any element of F (because all elements of V have 2* elements in each neighborhood). Moreover (Milnor, 1968), the limit of n^Wl/x, as A. -> oo, exists.
144 Isoperimetric Constants Furthermore, when M covers a compact Riemannian manifold with deck transformation group f, then there exist constants a > 1, b > 0, and C > 1 such that c~lnA(a~lk - b) < V(x; k) < Cn^(aA + b) for every x e M. Proposition V.3.2 Let G, F be connected, roughly isometric graphs, both with bounded geometry. Then G has polynomial (has exponential) volume growth if and only i/F has polynomial (has exponential) volume growth. Definition For any v > 1 define the v-isoperimetric constant of G by where K varies over finite subgraphs of G. Proposition X33 // G, F are roughly isometric graphs, both with bounded geometry, then \V(G) > 0 if and only if\v(F) > 0. Example X3.4 (Example V.3.3 continued.) Considera finitely generated group F, with fixed collection of generators A. We (henceforth) denote the combina- combinatorial disk centered at the identity e of radius k by 0(k). On the graph associated with F we work with the counting measure (and in this example we denote it by V); so when discussing the volume of combinatorial disks, we are working with the counting function. We define its "inverse" <t(A.), A. > 0, by *(A.) = rain {k 6 Z+ : V(/t) > A.}. That is, V(*(A.) - 1) < A. < V(*(A.)). Then for any finite subgraph K of F, 1 \/(K) (V3 1) A,^wr^ ^ * "\"-/ The proof goes as follows: Definition In any graph, not necessarily a group, given any function / on the vertices, one defines the differential off, 2)/, on the edges by »/([*, y]) = f(y) - fix), and one integrates 2)/ relative to dA (see §VI.5 below). For any function <f>: F -> R on the group F of finite support, and nonnegative integer k, define the mean value function fa on F by = J^ E V.3 Discretizations and Isoperimetric Inequalities 145 So <f>k(x) is the mean value of 4> on the combinatorial disk centered at x with radius k. For k = \,y varies over AU A~\ which implies which implies E * yeAKJA-[ For any* > 1, let y = (set zo = e) zt, I < k, where zj \<f>(xy) - <f>(x)\ < which implies Next, for every k, the triangle inequality implies {|*l > A-} C {|^ - ^| > k/2) U which implies V({|*| > A.}) < V({|^ - <f>k\ > k/2}) Also, one certainly has so if we pick *o to be the smallest * for which 1. One has < k/2, that is,
Isoperimetric Constants 146 that is, Pick <j> = TK, and A. = 1. Then one obtains (V.3.1). In particular, (V.3.2) V(*) > const.*" V* => ACK) > const.V(K)'-V' for all K of finite cardinality. So in the case of a group, a lower bound for volume growth alone is equivalent to the corresponding isoperimetric inequality. For v = oo the argument yields a slightly weaker result, namely, const.V(K) V(/t) > const.*00-* V * for all K of finite cardinality. ACK) > const, log V(*) + const. Example V33 If G is a Lie group with left-invariant Riemannian metric and bi-invariant Riemannian (Haar) measure, then one can formulate and prove, in similar manner, a corresponding result, that volume growth uniformly bounded from below implies an isoperimetric inequality. V.3.2 Discretizations Definition Let M be a Riemannian manifold. A subset Q of M is said to be (-separated, e > 0, if the distance between any two distinct points of Q is greater than or equal to e. Remark V.3.1 If Q is e-separated, then one always has only a finite number of elements of Q in B{x;r) when the exponential map is defined on B(x;r + e/2) [where, as above, B(x; p) denotes the disk of radius p centered at the origin of the tangent space at x, Mx]. Indeed, 2 \J B(x;r where the union on the right hand side is disjoint union. Therefore > card \Q n B(x; r)} inf V(r?;e/2). egnB(xs) But the compactness of B(x; r) implies inf,€j;n«(.r;r) ^(r?; e/2) > 0, which im- implies an upper bound on card {Q D B(x; r)). So the real question is to obtain a uniform upper bound for card Q n B(x; r) independent of x and Q. V.3 Discretizations and Isoperimetric Inequalities 147 Lemma V.3.1 Let M be complete, with Ricci curvature bounded from below as in (V.2.16) on all ofTM, and Q an (-separated subset ofM. Then for allx 6 M and r > 0. Proof Since card Q n B(x; r) is finite, there exists ^egfl B(x; r) such that = inf Therefore, the Bishop-Gromov volume comparison theorem (V.I. 13) implies Definition Let M be a Riemannian manifold. A discretization ofM is a graph G, determined by an €-separated subset Q of M, for which there exists p > 0 such that M = U*€ff &(%' P)- Then € is called the separation, and p the cov- covering, radius of the discretization. The graph structure G is determined by the collection of neighbors of $, N($) := \Q n fl($; 3/>)} \ \%\, for each ? e Q. Remark \32 Note that card N(?) > 1, so G is connected. Remark VJJ Lemma V.3.1 implies that when the Ricci curvature is bounded from below as in (V.2.16), then for the graph G we have 1 + m(?) < VKFp + */2)/ VK((/2) := Nf,.p for all ? e G - so G has bounded geometry. Proposition V3.4 Let M be Riemannian and G a discretization of M. Then there exist a > 1 and b > Ofor which (V3.3) a-'rf(*.,fc)<d(?i,&)<arf(*l,fc)+*> for all ?,, ?2 in Q.Thus, M is roughly isometric to any of its discretizations, and any two of its discretizations are roughly isometric. Lemma V.3.2 Let Mbea complete Riemannian manifold, with Ricci curvature bounded from below as in (V.2.16), and assume there exist positive constants ro and Vq such that V(x;ro)>Vo
148 Isoperimetric Constants for all x G M. Then for any r > 0 one has a positive constant const., such that V(x;r) > const., for all x e M. Proof Ifr > r0, then simply use Vo. If r < r0, then simply note that the Bishop- Gromov theorem (V. 1.13) implies VK(r0) which implies the claim. Proposition V3.5 Let Mbea complete Riemannian manifold, with Ricci cur- curvature bounded from below as in (V.2.16). Then for any discretization G ofM, G has polynomial (has exponential) volume growth only if (if) M has polynomial (has exponential) volume growth. If, on the other hand, there exist positive constants r$ and Vo such that V(jc;ro)> Vo for all x e M, then for any discretization G of M, G has polynomial (has exponential) volume growth if (only if) M has polynomial (has exponential) volume growth. V.3.3 Isoperimetry and Discretizations This is the main theorem of the chapter. Theorem V3.1 Let M be a complete Riemannian manifold with bounded ge- geometry. Then for any v > 1 we have 2VtP(M) > 0 if and only if\v(G) > Ofor any discretization G ofM. Proof First, given 0ViP(M) > 0. By Proposition V.3.4, we may work with any discretization. Therefore we consider a discretization G of M with separation constant e > 0 and covering radius R = p. To show that IV(G) > 0, it suffices to prove the existence of positive constants such that given any K. c Q we may find ii c M of inradius > p for which (V.3.4) and (V.3.5) A(dQ)< const, card SIC, V(Q) > const, card K. V.3 Discretizations and Isoperimetric Inequalities We proceed as follows: Given a finite subset K., set Then 149 V(t-;R) < where M^r is an upper bound (depending on e, R, and the lower bound of the Ricci curvature) of the maximum number of (-separated points in a disk of radius R (see Remark V.3.3). So for VR := inf V(x; R) > 0 jreitf [the positivity of Vr follows from Croke's inequality (V.2.15)], we have which implies (V.3.5). For the upper bound of A(dSi) we note that 3QC (J S($;R). Indeed, if x G dQ, then d(x, ?) > R for all ? € ?, and there exists ?0 G ? such that x G S($o\ R). But there must exist ?' e Q such that d(x, ?') < R, which implies ?' ^ K. Then d(?0, H') < 2/?. which implies |b e N(?')- So ?0 e 9(^ \ IC), which is the claim. Therefore, A(d?l) < AK(R)cardd(g\K:) < mAK(R)caiddK:, which implies (V.3.4). So we have the "only if claim of the theorem. For the "if claim, we again note that we may work with any discretization. Therefore assume that we are given the graph G, for which \V(G) > 0, with covering radius R = p < inj M/2. Suppose we are given ?2, with compact closure, C°° boundary, and inradius greater than p. Set fCo := {* € Q: V(Q n ««;/>)) > V($; p)/2), Kx := {? € ^: 0 < V(fi n So both ACq and ACi are contained in [il]p (ihe set of points with distance from
150 Isoperimetric Constants Q less than or equal to p). Then for at least one of j = 0, 1 we have (V.3.6) <v(nn|J «(*; V If (V.3.6) is valid for j = 1, then we have directly from Buser's inequality (V.2.18) V(O) n fi(?;p)) < const.M«,pAOQ) < const. [without any hypothesis on IV(G)], which implies A(d?i) > const.V(Q) = const.V(Q)l/vV(QI-|/v > const. since Si contains a disk of radius p, which, by Croke's inequality (V.2.1S), has volume uniformly bounded from below. So we must consider the case when (V.3.6) is valid only for j = 0. First, Therefore it suffices to give a lower bound of A(dSi) by a multiple of card dfCo - the multiple independent of the choice of fCo. To this end, define H c [Si]P by H := {x e M: V(x;p)/2 = V(Q n B(x;p))). For each % e 3ACo there exists t) G N(^), rj e ACo; we have, of course, </(?, t)) < 3p. By definition, D B(r)\p)) > V(r)-p)/2, V(Q D which implies the minimizing geodesic connecting ? to r) contains an element ? e H, which implies dfCo ^ [H\$p, which implies V.3 Discretizations and Isoperimetric Inequalities Let Q be a maximal 2p-separated subset of H. Thus u which implies 151 = 2M«,pconst. ? V(S2 n D - the third inequality uses the Bishop-Gromov comparison theorem (V.I. 13); the following equality follows from the definition of H 3 Q; and the fourth inequality uses Buser's inequality (V.2.18). ¦ Example V3.6 (Example V.2.3 continued.) Consider the Riemannian pro- product M = Mo x R*. where Afo is an (« - t)-dimensional compact Riemannian manifold. Then Proposition V.2.1 and Theorem V.3.1 imply that 3V,P(M) > 0 for any p > 0 for all v e [1, Jfc]. Indeed, both M and R* are discretized by Z*. Since V(x;r) < const./-* for all r > 0, then Jfc is the maximum of all v for which > 0 for some p > 0. Theorem Y3J2 Let M have bounded geometry. Then, for any v >n,we have 3V{M) > 0 if and only i/lv(G) > Ofor any discretization G of M. Proof The "only if is precisely as above. So we assume that IV(G) > 0. Suppose we are given Q, with compact closure, C°° boundary, and no assumption on the inradius. As above, we set ACo := {? 6 Q: V(S2 n B(t-; p)) > Vtf;p)/2}, Ki := {? 6 G: 0 < V(€l n Again, both ACq and /Ci are contained in [?2]p, and for at least one of j = 0, 1
152 we have (V.3.6): Isoperimetric Constants V(O) < V If (V.3.6) is valid for j = 0, then we may argue as above; for the only place we invoked the hypothesis of the inradius uniformly bounded away from 0 was when (V.3.6) is valid only for j = 1. We therefore adjust the argument for (V.3.6) valid only for j = 1. ¦ Lemma \33 There exists a constant jv > 0 such that AtfSi D BQ;p)) > JvV(Si n BQ;p))l-l/v, p < ^ for all I-eld. Conclusion of the Proof of Theorem VJ.2. Assume the lemma is valid. Then A.1.5) implies I\v/(v-l) J2 V(QnBG;p))l-l'v\ Therefore, (V.3.6) and the lemma imply V(O) <v(an J BG-. < const.AOO)|p/(l'-|), which implies the theorem. So it remains to prove the lemma. Proof of Lemma \33. Let D = Si n B($;p); then Therefore v > n implies V3 Discretizations and Isoperimetric Inequalities 153 Figure V.3.1: C(. so it suffices to prove V(D)i-i/» < COnst.ACDn Since p < inj A//2 we have, by Croke's inequality (V.2.14), A(dD n = const{A(dDC\ So we want to show A(dD n Consider geodesic spherical coordinates centered at ?. Let subset of S| for which B($;p)}. denote the (See Figure V.3.1.) For each d e Cf, let ct@) = sup{f > 0: exp td Note that if oF) < p, then the geodesic segment from exp oFN to exp p6 is contained in D. The Bishop-Gromov theorem (V.I. 13) implies
154 which implies V(D) > ^^- f AkKP) Jc Isoperimetric Constants which implies, by Buser's inequality (V.2.18), A(dDnS(t-;p))< < consl.A(dD D fl(?; p)) = const. A(dDr\B($;p)), A(dD which implies the claim. ¦ Example VJ.7 (Example V.2.4 continued.) Let JG2 denote the 2-dimensional jungle gym in R3. Then Theorem V.3.2 implies that 33<JG2) > 0. Indeed, JG2 and R3 are simultaneously discretized by Z3. Example V3.8 Given a noncompact M that covers of a compact Riemannian manifold Mo, the deck transformation group F of the covering is a discretiza- discretization (see Example V.3.2) of M. Then, by the proof of Proposition V.3.5 (see Kanai, 1985), a volume growth of order > rk in M implies a volume growth in F of order > ifc, which, by Example V.3.4, implies a ifc-isoperimetric inequality in T, which implies a (modified) ifc-isoperimetric inequality in M. V.3.4 Isoperimetric Inequalities on Products For future reference we note: Given the Riemannian manifolds Mj,j = 1,2, we consider the respective vol- volume and area measures associated with the standard product metric on M, x M2. Given the graphs GU), j = 1,2, we consider the graph GA) x GB) with vertices and oriented edges [(&, ?2). (m. %)] where either (a) ?, ~ $2 and ^1 = 172, or 0>) ?1 = Hi and fi ~ fa. with corresponding volume and area measures on the vertices and edges. Given the Riemannian manifolds Mj, j = 1,2, of bounded geometry, with respective discretizations Gw, then the inclusion of the discretization (GA) x GB))o in Mi x M2 is a rough isometry. V4 Bibliographic Notes 155 Proposition VJ.6 (a) Let G0), j = 1,2, be two graphs, both with uniformly bounded valence functions m.V\$), and assume IV/(GO)) > 0, Vj > I, j = l,2. Then I,,+^(GA) x GB)) > 0. (b) Let Mj, j = 1,2, be Riemannian manifolds of respective dimension nj. Let vj > »j, j = 1,2, and assume that Then x M2) > 0. (c) LetMj, j = 1, 2, be Riemannian manifolds as in (b). Letvj > 1 andp > 0, and assume that IVl,P(Mj)>0, y = 1,2. Then (V.3.7) Ivi+v2 ^(Mt x M2) > 0. If, in addition, both Mi and M2 have bounded geometry, and then (V.3.8) IVl+vl(Mi x M2) > 0. V.4 Bibliographic Notes Nearly all the material of this chapter is detailed, with references, in Chavel A994). We chose proofs from among the material there to emphasize our major themes, isoperime- isoperimetric inequalities and discretizations. An very nice introductory survey is Burstall's article in the 1998 Edinburgh Lectures (Davies and Safarov, 1999, pp. 1-29). §V.l Proposition V.I .1 is known as the Hopf-Rinow theorem; see Hopf and Rinow A931). § V.Z3 Propositions V.2.2 and V2.3(a) were first proved in Croke A980), and Proposi- Proposition V2.3(b) in Buser A982). Theorem V.2.6 was communicated to me by A. Grigor'yan; sec also his Grigor'yan A994b). §VJ.l The proof in Example V.3.4 is from Coulhon and Saloff-Coste A993).
156 Isoperimetric Constants §V3.2 Proposition V.3.4 was first proved in Kanai A985), with the hypothesis of Ricci curvature bounded from below to obtain the upper bound. Subsequently, this hypothesis was shown in Holopainen A994) to be unnecessary. Theorem V.3.2 was first proved in Kanai A985). §V3.4 An elegant proof of the product theorem. Proposition V.3.6, for the discrete case is given in §3 of Varopoulos A985). (It is surely valid in the continuous case.) Also, see Grigor'yan A985). VI Analytic Isoperimetric Inequalities In this chapter we explore, in greater detail, the applications of isoperimet- isoperimetric inequalities to Sobolev inequalities, a phenomenon only hinted at in the Federer-Fleming theorem (Theorem II.2.1) and the Faber-Krahn inequality (Theorem III.3.1). The application of isoperimetric inequalities to analysis is a rich subject in its own right, although we only develop here what we need for our particular study of the heat equation in later chapters. We also present the analogue of these arguments in the discrete case, and discuss the equivalence of Sobolev inequalities on Riemannian manifolds of bounded geometry, and the corresponding Sobolev inequalities on their discretizations. VI.l L2 Sobolev Inequalities Let M be an n-dimensional Riemannian manifold. Definition Let / be a C1 function on M. We define the gradient of f to be the vector field on M satisfying <grad /, ?) = df^) = $f Vt-eTM, where df denotes the differential of / on M. For C1 functions /, h on M we have grad (/ + h) = grad / + grad/t, grad fh = /grad h + /igrad /. If x: U -* R" is a chart on M, then, with the usual notation gij = (8,, dj), G = (jy). G~l = (gij), 157
158 we have Analytic Isoperimetric Inequalities d j.k dxJ Note that the Riemannian metric on M naturally induces a metric on fibers of TM*, namely, if a and p are covectors in the same fiber, then (a, 0) = (9~l(u), d~l@)), where d: TM ->¦ TM * is the bundle isomorphism defined by {0(?)}(»J) = (?> 1) for any ?. f e rw tot belong to the same fiber. One verifies, rather easily, that for any differentiable function on M we have = |grad/| on all of M. Lemma VL1.1 If the isoperimetric constant 3V(M) is positive for some given v > 2, then for any function <j> in Cf. Proof Of course, we have 3V = 6V > 0, by the Federer-Fleming theorem. Therefore, Ilgrad/H, > 3J/L/(,-n for all / e C~. Given any 0 e Cf\ let / = \<f>\p, p = 2(v - l)/(v - 2). Then Igrad/| = pM"-1 Igrad |*| I = pW'1 |grad^| (the last equality valid a.e.-dV - see Lemma II.2.1), which implies, by the Cauchy-Schwarz inequality, that L |grad/|rfV < which easily implies the claim. ¦ Remark VL1.1 We refer to (VI. 1.1) as the Nirenberg-Sobolev inequality. Lemma VI.1.2 If3v(M) is positive for some given v > 2, then there exists a positive const. „ such that (VI.1.2) forallfeC?. VI. I L2 Sobolev Inequalities 159 Proof Start with the previous lemma for v > 2. The first step is to show that &v > 0 for some v > 2 implies f d>2+4/vdV Indeed, if v > 2, set / = \<f>\2, p = v/(v - 2), and g = \4>\*/v, q = v/2, and use H6lder's inequality to obtain (VI.1.3) from (VI.1.1). On the other hand, if v = 2, one has directly II02||2 < const.||grad@2)||, < const.||4>||21|grad0||2. Square both sides to obtain (VI. 1.3). To obtain (VI. 1.2), simply apply Holder inequality to / = |^|4/(v+4), p = (v + 4)/4, and g = 2 (v + 4)/v. s q = Remark VI.1.2 We refer to (VI.1.2) as the Nash-Sobolev inequality. Theorem VI.1.1 The Nash-Sobolev inequality is equivalent to the Nirenberg- NirenbergSobolev inequality. Proof The proof of Lemma VI. 1.2 shows that the Nirenberg-Sobolev inequal- inequality implies the Nash-Sobolev inequality. So it remains to consider the converse. Suppose we are given a positive constant C such that \\4>hl+Vv < for all <f> e Cf, for some v > 2. Set 2v Then v + 2' v + 2 Given any / e C» define, for every k e Z, [0. / < 2*. /* = min{(/ - 2*)+, 2*} = | / - 2*. 2* < / < 2*+1 I 2*. 2*+' < /
160 Analytic Isoperimetric Inequalities V A k . Figure VI. 1.1: /*. (see Figure VI. 1.1) and Then which implies 4/K r , nl+2/x = L-,A2] 1+2/v that is. Let <C2{2*V(/>2*)}4/"/" lgrad/|2, < C2{2*V(/ > 2*)}4/v / |grad/| ak = 2k"V(f>2k), bk= t |grad/|2; 'ft then (VI.1.4) implies V/.2 T/i« Compact Case 161 which implies JteZ |2(l-«) which implies the theorem. ¦ Theorem VI.1.2 (Cheeger's Inequality) If3x(M) > 0, then for any function (f> in Cf we have In particular, X(M) = inf spec - A (see §VII.l and §VII.2.2) satisfies MM) > ]-3oo(MJ. 4 Proof The inequality for grad / is simply an adjustment of the argument for (VI. 1.1) to the case v = oo. To obtain the Cheeger inequality for k(M) consider thecase# = /2,forany/ e C™, and then use the Cauchy-Schwarz inequality, followed by the characterization of eigenvalues (VII. 1.2) and (VII.2.4). ¦ Remark VI.13 When our manifold M is a relatively compact domain in some larger Riemannian manifold, and M has C°° boundary, then our Sobolev con- constant 6V, and the inequalities that follow from it, correspond to vanishing Dirichlet data on the boundary dM. Therefore, in the Cheeger inequality in Theorem VI.1.2, \{M) is the lowest Dirichlet eigenvalue of the Laplacian on M. For vanishing Neumann boundary data, we use the definitions of the Sobolev constants sv of the compact case (below), in which case A.((Af) is the lowest nonvanishing Neumann eigenvalue of the Laplacian on M. VI.2 The Compact Case Here compact means compact without boundary, what was once called a closed manifold. Let M be compact Riemannian, n > 1 the dimension of M. Then, as men- mentioned in Remark V.2.1, all isoperimetric constants vanish. Alternatively, by
162 Analytic Isoperimetric Inequalities considering the function / = 1 on M one has that all the Sobolev constants of M vanish. Nevertheless, one can adjust the definitions as follows (here one only needs, for the geometry, the isoperimetric dimensions n and oo): Definition Define the isoperimetric constant in(Af) by in(M) = inf where F varies over compact (n — l)-dimensional C°° submanifolds of M that divide M into two disjoint open submanifolds Qi, Q2 of M. Define Cheeger's constant ioo(Af) by where F varies over compact (n — l)-submanifolds of M described above. Remark VI.2.1 One has here, as in Remark V.2.5, that it suffices to assume that the open submanifolds J2i and fi2 are connected. Definition For v e A,00] define the Sobolev constant ofM, &V(M), by (with the obvious interpretation for v = 00), where a varies over R, and / over C°°. The analogue of the Federer-Fleming theorem (Theorem II.2.1) reads as follows: (VI.2.1) in(M) < and Cheeger's inequality reads as where ki(M) denotes the lowest nonzero eigenvalue of M. Definition Given any function on M, we define its mean value fM by '><¦¦= Remark VI.2.2 For any s > 0, the ao that realizes infa || / - a \\,+i is charac- characterized by {sgn(/-ao)}l/-aol* = O. VI.2 The Compact Case 163 Forv = ^wehavein^ \\f - a || 2 realized by the mean value fM of /over M. For v = 00, we have infa \\f - a\\x realized by that a0 for which ao}) = Definition For M compact, we define the alternate Sobolev constant s'v(M) by inf ^ . vetl.oo], ' 11/ — /mIIv/(v-1) where / varies over C°°(Af). Remark VL23 We comment on the relation between i^, Soo, s^. Clearly, Suppose we are given a disconnection of M = Q, U Q2 U F, where Q,, Q2 are domains in Af and F is a compact (n - l)-submanifold of Af. Let F€, e > 0, denote all points in Af whose distance from F is less than e, and assume V(Q,) < V(Q2). Let /, be the function on M which is equal to 1 on J2, \ F,, equal to -1 on J22 \ T(, and linear across F on F(. Then a standard argument shows that lim / and /« -a\dV > |1 -aUV(Q,)-const.a(} + |l +a\{V(Q2) - > {|1 -o| + |l +a|}{V(Q1)-constA}, where St -* 0 as e I 0, which implies »nf / \f(-a\dV > ( which implies ««, < A(V)/ V(Q,); so Soo < ioo Now given any function / e C°°(A/), pick the constant 0 so that, for ?2. = {/ > 0}, n2 = {/ < fi],
Analytic Isoperimetric Inequalities 164 we have (see Remark VI.2.2 above); and let D, = {q eQ Then V(D,) < V(M)/2 for all f > 0, which implies, by the co-area formula A.3.7) and Cavalieri's principle A.3.3), / |grad(/-0)|</V= / A(dD,)dt n, Jo V(D,)dt = ioo f \f- Jn, 0\dV. The same argument yields which implies n2 >\x f \f- Jn2 0\dV, / |grad/|<fV = f \gml(f-P)\dV n Jo which implies s<x> > i<»; so \f-a\dV, On the other hand, if we work with fM instead of 0, then we are only guaranteed that at least one of {/ — /«> 0}, {/ — /«< 0} has volume < V(M)/2. The same argument then yields In sum, we have (VI.2.2) *«> — Remark VL2.4 If M is not closed, but has compact closure and C°° boundary, then the above apparatus of isoperimetric and Sobolev constants is still well de- defined, with the above inequalities. We only have to note that A.i (M) in Cheeger's inequality is now replaced by fii(M), the first nonzero Neumann eigenvalue of M. VI.3 Faber-Krahn Inequalities 165 VL3 Faber-Krahn Inequalities One may easily check that, for M = R", n > 1, one may express the Faber- Krahn inequality, Theorem III.3.1, as the analytic inequality for relatively compact SI in R" with C°° boundary, where CB depends only on the dimension n, with equality if and only if Q is an n-disk in R" [in particular, CR = k(B) V(BJ/n, where B denotes an n-disk in R"]. The bottom of the spectrum, A.(?2), is discussed in VII. 1 and VII.2.2. For general Riemannian manifolds, we do not even hope for precise geometric or eigenvalue inequalities. Nonetheless, we can consider qualitative versions of the Faber-Krahn inequality, in the spirit of the geometric isoperimetric inequality. Lemma VI3.1 Let v > 2. Then (a) The Nirenberg-Sobolev inequality ||gradM||2 > a\\uhvi(V-2h for all u e Cf, implies the Faber-Krahn inequality for relatively compact domains Si in M with C°° boundary, (b) The NashSobolev inequality llgrad«|h>0||«||21+2/lK||r2/\ for all u G C*. implies the Faber-Krahn inequality HSi) > p2V(S2)-2/v for relatively compact domains S2 in M with C°° boundary. Proof Let u > 0 be an eigenfunction of k(Sl), || u \\ = 1 (see Proposition VII.4.1). Then Also, which implies ' = f U2 < j f M2"A"-2) (w—2)/v 2/k which implies the claim.
166 Analytic Isoperimetric Inequalities Again, u > 0 be an eigenfunction of k(Si), \\u\\ = 1. Then A. > 02||M||r4/\ and / m < \\u\\2V(SiI'2 = V(Q), which implies the claim. ¦ Remark VI3.1 We refer to the Faber-Krahn inequality of this lemma as a Faber-Krahn inequality ofv-Euclidean type. Theorem VI3.1 The Riemannian manifold M satisfies a Faber-Krahn in- inequality ofv-Euclidean type, v > 2, if and only if it satisfies the corresponding Nirenberg- and NashSobolev inequalities. Proof We already know the "if part of the theorem, so we only consider "only if." Let 2v * ~ v-2' and, for any relatively compact Si in M with smooth boundary, set av(Si) = k(Si)V(SiJ/v, fiv(Si) = fof It suffices to show that there exists const. > 0 such that (VI.3.1) av(Si) < const.fiv(Si) for all Si. Proposition VI3.1 There exists a bounded u G C°°(Si) n f)c(Si) such that. (i) u > 0, (ii) fQ uq dV = 1, (iii) A u = -/zv Remark VI3.2 For the definition of f)c see Remark 1.3.3 and the discussion in §VU.2.2. Lemma VI3.2 Assume M satisfies the Faber-Krahn inequality ofv-Euclidean type. Let u satisfy Proposition VI.3.1. Then there exists const.v, > 0 such that V({u > ||m||oo -'}) > const., I — ,./,_,) v' e [0, llMlUinregval., where regval,, denotes the set of regular values ofu, av = av(Si), fiv = fiv(Si). Proof Set L = Si, = {u > L-t). Then VI.3 Faber-Krahn Inequalities 167 > avV(Si,)-2'v, and K(Sit) < fQi(u-L+t? Now [u - L + tf > j [u-L+t}2> (t/2JV(Sil/2), which implies which implies Therefore, which implies 1/2 'IV(Q,/2)J ' = V(O,)(lP+4V2lP. ( a»q [ */» at Also, = const.,.
168 Analytic lsoperimetric Inequalities Next, fix o e Si that realizes Hgradulloo = max |gradu| := <5. Then therefore. which implies as I -*¦ oo. Thus, Sd(x,o)<— => xeSit/2>, lim V (Sil/2,f«v+4»' = 1, which implies the claim of the lemma. ¦ Conclusion of the Proof of Theorem VL3.1 Given u of Proposition VI.3.1, then, by Cavalieri's principle (Proposition 1.3.3), 1 = f uqdV = f rL = / qV(u >t)t"-ldt Jo fL = / qV(u > L - t)(L - /)«"' dt Jo "/2 - -(«-l)v/2+v/2+4 = const. „ = const. „ I — since -(q - \)v/2 + v/2 + q = 0. Thus Faber-Krahn implies Nirenberg-Sobolev, which implies Nash- Sobolev. ¦ All the analytic isoperimetric inequalities considered thus far require v > n (check!), but a simple way to deal with v < n goes as follows: Definition Let M be an arbitrary Riemannian manifold. Given a positive in- increasing function g( v), v > 0. We say that a domain ?2 in M satisfies a geometric VI.3 Faber-Krahn Inequalities 169 g-isoperimetric inequality if A(dD) > g(V(D)) for all domains D CCSi. Given a positive decreasing function A(v), v > 0. We say that a domain Si in M satisfies an eigenvalue A-isoperimetric inequality if X(D) > A(V(D)) for all D CC Si. Remark VUJ Thus, A-isoperimetric inequalities are generalizations of Faber-Krahn inequalities of Euclidean type. Theorem VL3.2 Suppose Si satisfies a geometric g-isoperimetric inequal- inequality, with g(v)/v a decreasing Junction ofv. Then Si satisfies an eigenvalue K-isoperimetric inequality with Proof This is the usual proof of Cheeger's inequality. For any u e C^°(D) apply the co-area formula to u2. We define we and Then = {xeD:\u(x)\2>z], V(t) = V(D(r)), A(z) = 2||«||2||grad«||2 > / |grad(«2)|rfV JD f°° = / A(z)dz Jo g(V(z))dz f0 V(T) V(z)dz V(z)dT g(V(D)) V(D) \\u\h\
170 Analytic lsoperimetric Inequalities which implies, by (VII. 1.2) and (VII.2.4), which implies the theorem. 4 V V(D) ) Example VI J.I Let M be a complete Riemannian manifold with bounded geometry, and assume 3V,P(M) > 0 for some v > 1, p > 0. Then, by Theorem V.2.6, M satisfies a geometric g-isoperimetric inequality, with g(y) given by g(v) = const. v < v0, v > v0. Then Theorem VI.3.2 implies that M satisfies a eigenvalue A-isoperimetric inequality, where !V~2/", V < Vn, 2/v const.Vov~z/l\ v > v0. Example V133, Similarly, let M be a complete Riemannian manifold with bounded geometry and positive Cheeger constant, that is, 3oo(AO > 0. Then, by the argument of Theorem V.2.6, M satisfies a geometric g-isoperimetric inequality, with g(v) given by = const. v < v0, Then Theorem VI.3.2 implies that M satisfies a eigenvalue A-isoperimetricj inequality, where fy-2/n v <; v 2/n vo/", v > v0. We shall consider a variant of the Nirenberg-Sobolev inequality in §Vm.4g VI.4 The Federer-Fleming Theorem: The Discrete Case We now formulate and prove the discrete Federer-Reming theorem. But firstj Definition (Recall.) To every function / on a graph G we associate its di ential, defined on Qe the oriented edges of G, by VI.4 The Federer-Fleming Theorem: The Discrete Case 171 Theorem VI.4.1 (Discrete Co-area Formula) Let f: Q -* [0, +oo) have finite support. Denote the collection of its values by f(Q) := {0 = fa < ft\ < ¦ ¦ • < PnY To each i'60,...,JV associate K, the subgraph ofG determined by the vertices by the vertices Then Proof First note mat 3K, = {tt. „]: m > A > /(|?)}; so and 1 ')¦ for any such [f, „] 6 aK, we have which implies E ? E E E E E . f(n)<Pt 1=1 which is the claim.
172 Proof First, Analytic Isoperimetric Inequalities (jc - x"-y" _Pxp -l px p-\ so we may choose c-i = p. Next, fix any a > 1. If y < x < ay, then y)P ~l ~ Ba)P~1' (jc - y)(x + yy-1 ~ and if ay < jc.then xp - yp > (I - a~p)xp, x-y<x, x + y <(l+a~l)x, which implies XP-yP l-Ct-p (x - y)(x So we may pick = min i — c\ = min This proves the lemma. ¦ Lemma VI.4.2 Let a, be a decreasing sequence ofnonnegative numbers, and Pj an increasing sequence of nonnegative numbers, with A) = 0. Then for any p > 1 we have IN \P N Y.ttJllP<fii - ft-i)| ? const., ?>,(?/ - Pj-t") 7=1 I 7-1 for all N = 1,2 Proof UN -I, then both the right and left hand sides of (VI.4.1) are equal to a\P\p. So (VI.4.1) is valid for N = \. For the induction step we have .7=1 \N+l N l - fiN) N+l VI.4 The Federer-Fleming Theorem: The Discrete Case 173 i" - Pit") Therefore, if for any N we have then we also have {N+\ (N Eaj'/pwj-Pj- y-i 1 E•>(/>/ 1 •'*i N+l which is the claim. U ** S°b0leV COnSUmt S where /ranges over functions on Q with finite support, and the L> norms are ST "d 'A on a "d a" ^^(We have «¦ indicatOT fancti0° of «. = 2A0K), which implies8,@) < 2A(aK)/{V(K))(-i)/»fcrdl«chclioicesofK which "nphes 8,@) < 2IV(G). Therefore, if SlP(G) > 0, then |r(G) > 0. To show 2
174 Analytic Isoperimetric Inequalities opposite direction, we first note that since \a — b\>\\a\ — \b\\ for all real a and b; therefore it suffices to consider the case where / is nonnegative. Assume that I,,(G) > 0. Then for / nonnegative we have / |D/| dA = 2 J2 AOK.-XA - ft-,) JQ, ,=1 N > 2IV(G)? MK,-)}0"-"/^/* - ft-i) ? > const.lv(G) \(v~l)/v = const. I V(G) ||/1| „/(„_!) - the third line follows from the argument of Lemma VI.4.2. Therefore, I,,(G) > 0 implies SV(G) > 0. ¦ Remark VI.4.1 The results of §VI.l above all follow from the positivity of the Ll Sobolev constant, and integral inequalities. Therefore, since the Federer- Fleming theorem holds for graphs (as far as positivity of the isoperimetric and Sobolev constants are concerned), we have the corresponding versions of the above results for graphs. VI.5 Sobolev Inequalities and Discretizations Given a graph G with bounded geometry, let K be a finite subset of Q. Then for any function / on Q, and any s > 0, we have that is, the two expressions are equivalent in that their quotient is contained in a compact subset of @, +oo), this subset independent of the choice of K and /. For convenience define In what follows, we will work with s = 2, so the notation will not get out of hand. And whenever we write |3?/|(§) we mean v|3?/h(f )¦ One can then work out the details for any other fixed value of s, for example s = 1. Let X be a graph with bounded geometry, Y a metric space, and <p: X ->• Y a rough isometry with rough isometry Constantsa > l,b > O.Thencard0~'[)']is VI. 5 Sobolev Inequalities and Discretizations 175 uniformly bounded as y varies over Y. Indeed, if <j>(x\) = <f>(x2) then a~ld(x\, x{) — b < 0, which implies that d(jtj, xi) is uniformly bounded, in- independent of y. The bounded geometry hypothesis now implies the claim. Similarly, if AT is a graph with bounded geometry, Y a metric space, and <p: X -*¦ Y a rough isometry with isometry constants a > 1, b > 0, then for every R > 0,y e F, the number of disks B(<p(x); /?)thatcoverjisboundeduniformly from above independent of y. Indeed, if y e B{<p(x\); R) n B(<j>(x2); R) then d@(jti), 0C*2)) < 2R, which again implies the claim by the rough isometry hypothesis on <j> and the bounded geometry hypothesis on X. Lemma VI.5.1 Let Xbea graph with bounded geometry, Y either a connected graph or a complete Riemannian manifold, and <j>: X -*¦ Y a rough isometry. Let f: X -* R andg: Y -*¦ R be nonnegative functions satisfying (VI.5.1) f{x) < const. / g JBWx);p) for some p > 0, for all x € X, where & stands for /J or B depending on whether Y is a graph or a Riemannian manifold. Then (VI.5.2) / / < const. / Jpu.R) yB(, g- Proof Recall that the constants a and b are from the definition of the rough isometry. Because <Pi0(x; R)) C. BD>ix);aR + b), we only have to count the number of times any y in B@(jc); a R + b + p) contributes to the integral on the right in (VI.5.2). Well, all multiple counting of<f>(x) comes from #"'[#(*)], which has cardinality uniformly bounded from above. For the multiple contribution from v within p of 4>(X), we only have to note that Biy; p) n <p(X) has cardinality uniformly bounded from above for all y 6 Y. ¦ Notation When we have a map from 0: X -> Y, and a function f:Y-*R, we denote by </>*/ the function defined on X by the usual <p*fix) = f(<p(x)). Lemma VI.5.2 Suppose X and Y are graphs, X with bounded geometry, <P ¦ X ~+ Y a rough isometry. Let f :Y -*¦ R. Then I \4>*f\2 < const JpU.R) I l®@*/)l2 < const. / Jfi(x:tt) J/H for sufficiently large R, independent ofx.
176 Analytic hoperimetric Inequalities VI.5 Sobolev Inequalities and Discretizations Proof We only prove the second claim. We want to estimate |2>(^*/)l2 from above, for each x e X. Namely, x'eN(jr) Set a + b := L. Then d(jc, jc') = 1 =>• d@(x), <f>(x')) < L, which implies |2W'/)I2(*) < const. / |2>/|2. Now use Lemma VI.5.1. ¦ Lemma VI.53 Let X be a graph with bounded geometry, and <t>: X -* X a rough isometryfor which d(jc, 4>(jc)) is uniformly bounded from above. Then [ \f -4>*f\2< const. [ |2>/|2 Jp(x.R) Jp(x\cona.R) for sufficiently large R > 0, independent ofx. Proof Because d(jc, <t>(x)) < K for some K > 0, for all x, we have |/(jc) - f(<P(x))\2 < const, f |2>/|2, which implies the claim. Lemma VL5.4 Let$:X -*¦ Ybe a rough isometrybetween graphs ofbounded geometry, and f:Y-+R a function on Y. Then, for sufficiently large R, [ I/I2 < const, f |2)/|2 + const, f \4>*{f)\2, Jpiy.R) J0(y;const.«) ./0(*(y);const.«) where r/r denotes a rough inverse of<j>. Proof Recall that every rough isometry <p has a "rough" inverse, that is, therft exists a rough isometry ir : Y -*¦ X such that are uniformly bounded on X and Y respectively. Now for any y e Y,v/e hav* \f\\y) < const.H/ - D>oirr(f)\2 + |(^o^r(/)|2}(y) < const. [ |2)/|2 + const.|@o^r(/)|2(y), Jp(y.K) 177 which implies / I/I2 < const. / |2)/|2 + const, f |(^*o^' < const. which implies the proposition. const. |</>'(/)|2 Lemma Vlii (Kanai, 1985.) Let M be a complete Riemannian manifold with Ricci curvature bounded uniformly from below. Then for any r > 0 there exists a constant depending on r such that, for any u G C°°(D(o; r)). JB(o |gradM|rfV > const.r |« - uB{or)\dV for all o e M, where ub(O\t) denotes the mean value ofu over B(o; r). Proof Write B for B(o; r). Assume uB = 0 and V({x e B:u(x) > 0}) < V(B)/2. For t > 0 set D, := {jc e B : u > t). Then, by Buser's inequality (Proposition V.2.6(b)), we have for regular values of u A(dD,)> const.r which implies, by the co-area formula. ¦r A(dD,)dt > const., / V(D,)dt Jo = const., const which is the claim. UdV - I \u\dV, Jd Given the complete Riemannian manifold M and a discretization G of M, with bounded geometry, for which the separation radius is e and covering radius s P, wlth e < p. Recall that t, g N(f) if rf(?, r,) < 3p. We now investigate the
178 Analytic Isoperimetric Inequalities simultaneous validity of the various Sobolev inequalities on M and G. Of course, we need maps between the function spaces on M and G. VI.5.I The Discretization T> of Functions on M Definition Given a smooth function F: M -*¦ R, its discretization f = VF': G -+ R is defined by >:i) Assume V(x;3p) > const. > Oforallx e M.Thenforanyp e [l,oo) we have 1 < const JBtf-Jp) \F\"dV, which implies (by Lemma VI.5.1) f \VF\pdV < const. / So one has ||Z>|| ,,_>,, < const, for all p e [1, oo]. :ii) Similarly, one has < const . / JBiS.3 \F\dV < const. I \F\dV; . I JU so HPIIi-k* < const. (D:iii) Assume M has Ricci curvature bounded from below, and assume V(jc; 3p) > const. > 0 for all x e A/. Then for i? e N(§) we have JB$; ¦;3p)nB(r,;3p)) Since fl(?;3p)n B(i;;3p) 2 fi(midptf ,;p), which implies 3p) n B(j)\ 3p)) > const., we have - (X>FXi?)| < const. I [ \F(x) - (PF)(?)| dV(x) U B(S;3pY\B(rr,}p) + f |F(jr)-(PFXij)|rfV(*)| < const. I / |F(jc)-i + [ \F{x)-(VFKn)\dV(x) JB(n;lp) dV(x). VI.5 Sobolev Inequalities and Discretizations 179 < const. I / |gradF\dV + f |gradF\dVI < const. / |gradF|rfV, JBdfip) (we used Lemma VI.5.5 in the next to last inequality), that is, \<PFYM) - (VFXn)\ < const. / |grad F\dV JB(liAp) for all ij ? N(f), for all ? e Q, which implies |©(PF)|2(f) < const, f |grad F|2 rfV for all ? 6 ?. Then Lemma VI.5.1 implies / \X)(VF)\2 dV < const. / Igrad F\2 dV for all f e ?; letting /? -* +oo, we have / |S>(PF)|2</V< const. /" Igrad F\2dV. JG JM VI.5.2 The Smoothing S of Functions on Q Definition Fix a function ^: [0, +oo) -* [0, 1 ] e Cf([0, 2p)) such that \H@, p] = 1. With ? e ? associate the function ^ : M -* [0,1] defined by and define So the collection {^ }^6!; is a partition of unity on M subordinate to the locally finite cover {fi(?; p))^eg of M. Then for each f:Q-*R define its smoothing F = 5/: M -* R by :i) Then ^cS 1 and < 1.
180 Analytic Isoperimetric Inequalities Also, assume the Ricci curvature is bounded from below. Then one has from the Hdlder inequality that s E = E (eg which implies I \Sf\»dV< f JB(o;R) JB{o\R) = E < const. < const. JfH.ru, :conjt./?+const.) [the third line follows from the Bishop comparison theorem (V.1.11)], that is, f \Sf\"dV < const. / \f\"tN, where rjo is a vertex in Q within Riemannian distance p of o. In particular, we also have [ \Sf\"dV < const. / \f\pdV, Jm Jg that is, HSH^p < const, for all p > 1. E:ii) Furthermore, assume V(x\p) > const. > 0 for all x G Af.If / > Othen f JM (Sf)"dV - E [by the Bishop-Gromov theorem (V.I. 13)], that is, for all/>0. E:iii) Now consider > const. V1.5 Sobolev Inequalities and Discretizations Given x, there exists ij, e Q D fi(jr; p), which implies 181 Therefore KgradSf)\(x) < const. which implies for any x e fi(/?; p),»? e Q, Kgrad 5/)|2U) < const. ]T |/(f) - /(i?)|2 = const.|2)/|2(i7). ?efl(i);3p) One now obtains, by the argument of Lemma VI.5.1, / |grad5/|2</V < const. / \T>f\2dV, HgradS/lh < const.||2)/||2. VI.S.3 First Smooth, Then Discretize: T>S (VS:\) No assumption on M and G. We have = E (X>S:ii) Assume M has Ricci fdV. curvature bounded from below, and V{x\ p) > et / b i f rm below, and V{x\ const, for all x e M. Let / be a nonnegative function on G; then BM/)(?) = —^—- / SfdV > const. / SfdV> const. /(?), which implies ||/||p < const.p||P5/||p for all p > 1.
182 Analytic hoperimetric Inequalities VI.5.4 How Smoothing Followed by Discretization Differs from the Identity: idg — VS Note that = ^^ f V(S,3p)Jbu3P) which implies Since G has bounded geometry, we have Therefore, / 1/ - PS/I2 </V < const, f |2>/|2 </V VI.5.5 How Discretization Followed by Smoothing Differs from the Identity: idM — ST> Again, we require both Ricci curvature bounded uniformly from below, and V(x; p) > const, for all x e M. For a smooth function F: M -* R we have {F(x) - F(y)} dV(y), which implies |F - SVF\(x) < const. [ |F(jr) - F(y)| dV(y). JB<.x?p) The Bishop comparison theorem (V.I. 10) implies |F - SVF\(x) < const. / dpx(v) I Igrad F|(exp sv)ds, Js. Jo which implies / \F-SVF\2dV < const. / dp(v) I IgradF|2(exp sv)ds, Jb(o;R) JSB&;R) Jo Vl.6 Bibliographic Notes 183 where SB(o;R) denotes the unit tangent bundle over B{o\R) and dfj. the Liouville measure on SB(o; R) (see V. 1.4). Therefore, / |F-«SPF|2</V < const. / dpL{v) I Igrad F|2(exp sv)ds Jb(o;R) JSB(o-.R) Jo = const. / ds I |grad F|2Gro4>,(v))^(v), Jo JSB(o;R) where 7r : SM -*¦ M denotes the natural projection, and <t>, the geodesic flow. By Liouville's theorem on the invariance of the Liouville measure under the action of the geodesic flow (Proposition V. 1.3), we have f \F-SVF\2dV < const. / ds I Igrad F|2Gro4>,(v))^(v) Jb(o:r) Jo Jsb(o,r) = const. / ds I Igrad F\2(jz(v))dn(v) Jo /¦I(S«(o;«)) < const. / Igrad F\2(n(v))dn(v) JSB(o;R+Sp) = const. / Igrad F\2dV. JB(o;R+ip) To summarize, f \F-SVF\2dV < const. / |gradF|2rfV, Jb(o;R) JB(o\R+ip) f \F - SVF\2dV < const. / |gradF|2rfV. M Jm VI.6 Bibliographic Notes Surveys of applications of isoperimetric inequalities in analysis can be found in Bandle A980), Kawohl A985), Mossino A984), Payne A967), and the classic and Pdlya and Szegfi A951). Also, see the recent survey (Hebey, 1999). §VI.l The Nirenberg-Sobolev inequality can be found in Nirenberg A959, p. 14). For Lemma VI. 1.2, see Moser A964. p. 116), Cheng and Li A981), and Nash A958). For Theorem VI. 1.1 see Bakry, Coulhon. Ledoux, and Saloff-Coste A995). One might wonder whether the Nirenberg- and Nash-Sobolev inequalities imply the Federer-Fleming inequality, that is, whether the L2 Sobolev inequalities imply the the L' Sobolev inequality. Counterexamples are presented in Coulhon and Ledoux A994). Theorem VI. 1.2 is from Cheeger A970). §VL2 Details for the analogue of the Federer-Fleming theorem (Federer and Fleming, 1959) in the compact case can be found in Chavel A984, p. 111).
184 Analytic lsoperimetric Inequalities A more delicate isoperimetric function for the compact case is inspired by M. Gromov's proof of the isoperimetric inequality for spheres (Gromov, 1986). One lets I(M, /?), 0 € @,1), denote the infimum of A(dSi) among all Si satisfying V(S1) = PV(M). See the disucussion in B6rard A986, Chapter IV). §VL3 As mentioned in V.4, one has a Faber-Krahn theorem in the model spaces of constant sectional curvature. Similarly, one has generalizations of the Faber-Krahn argument in Berard A986, Chapter IV). Theorem VI.3.1 is from Canon A996). Proposition VI.3.1 is from Aubin A982, p. 116 ff.). The general Faber-Krahn inequalities were first treated in Grigor'yan A994c). §VL4 To my knowledge, the earliest proofs of the Federer-Fleming theorem for graphs are in Varopoulos A985), where he considers the result there for Cayley graphs, and in Dodziuk A984), where he considers the discrete Cheeger inequality. One can consider both Riemannian manifolds and graphs with weight functions associated to their measures. See the discussion in Chavel and Feldman A991) and Coulhon and Saloff-Coste A995). §VL5 The discussion of discretization and smoothing of functions follows the treatments of Kanai A986a, 1986b) and Coulhon A992). VII Laplace and Heat Operators Here, we introduce the Laplace operator on Riemannian manifolds and its asso- associated heat diffusion, and prepare for the study of how geometric isoperimetric inequalities on Riemannian manifolds are reflected in the properties of large time diffusion. We present the necessary definitions and background results, most of which can be found in Chavel A984). The ones we discuss in some detail are either to correct an argument, or to fill in matters not discussed, there. The manner in which the heat diffusion expresses the geometry of the manifold will be presented in the next chapter. VTI.l Self-adjoint Operators and Their Semigroups Definition Let H be a Hilbert space with inner product (,). Recall that a linear operator T: V(T) -> H on the subspace V(T) in H is an extension of the linear operator of S: V(S) -> H ifV(S) c V(T) and T\D{S) = S. We write ScT. The linear operator 7 : V(T) -y U is closed if every sequence (xk) in T>(T) satisfying x, Txk x,yeH, must also satisfy x G V{T), y = Tx. A linear operator 7: V(T) -* H is dosable if it has a closed extension. When 7 is dosable, then it has a minimal closed extension T, called its minimal extension or, for short, its closure. Definition Assume T : V(T) -*¦ H is a linear operator with dense domain, that is, T>(T) = H. The domain of the adjoint 7* of T will consist of those x G H 185
186 Laplace and Heat Operators for which there exists an element x* efi such that (jc, Ty) = (jc*, y) for all y e TKT). For such jc we define T*x = x*. The operator 7 is called symmetric if Gjc, y) = (jc, Ty) for all x,y e V(T). In general, 7* is a closed operator. When 7 is symmetric, then T* is an extension of 7, so 7 is closable. Then 7 is the smallest closed symmetric extension of 7. Definition We say that 7 is self-adjoint if 7 = T*. We say that 7 is essentially self-adjoint if 7 has a unique self-adjoint extension, in which case 7 = 7*. One knows that if 7,, 72 are self-adjoint, and X>G,) c V(T2), then 7, = 72. Definition Let 7 be a symmetric operator on H with dense domain V(T), for which there exists a real number e such that (VH.1.1) (*, Tx) > e(jc, x) VjceZ>G). We then say that 7 is semibounded from below. We say 7 is nonnegative if e >0. Let 7 be semibounded from below, with e as given in (VII. 1.1); then the bilinear form on Z>G) defined by is positive definite, and hence defines an inner product. Complete V(T) to DT relative to the inner product (, )r- Then it is known that DT may be realized as a subspace of H, so T>(T) QDTC.H. Definition Let 7 be semibounded from below and symmetric, with dense domain. Define the Friedrichs extension JJr of 7 to be T*\Dt. That is, the domain of JJr, 2)r, is given by and 5r is given by CTX,y) = (x,Ty) Vjc e QT,y G V(T). Therefore, V{T) c3)rCDrcK, Also, 5r is self-adjoint, and Er^,3') = U. Ty) VxeQT.yeDT. The full result is: VII. 1 Self-adjoint Operators and Their Semigroups 187 Proposition VII.1.1 Every semibounded from below symmetric operator with dense domain has at least one self-adjoint extension, its Friedrichs exten- extension. Moreover, 3> is the unique self-adjoint extension ofT whose domain is contained in DT. VII.1.1 The Spectrum of Self-adjoint Operators Definition Let Ti be a Hilbert space with inner product (, ), 7: Z>G) ->«a linear operator on H. A complex number A. is in the resolvent set ofT if XI — T maps X>G) one to one onto all of 7i, and if (A./ - 7)~' is bounded. In this case we refer to (A./ - 7) as the resolvent ofT at k. The spectrum ofT, spec 7, is the complement of the resolvent set of 7. The spectrum is always a closed subset of the complex numbers. When 7 is self-adjoint (and this is the case we always study), spec 7 is contained in the real axis. For 7 self-adjoint we have the spectral theorem, formulated as follows: Recall that an orthogonal projection is characterized as a self-adjoint trans- transformation PofH for which P2 = P. In particular, , <(>) = {P2<t>, <t>) = (P4>, P<t>) > 0 for all <j>. A spectral family is a family {?*: A. e R} of orthogonal projections of Ji satisfying 1- ?ji<^ [that is, (Ex4>, <t>) < (E^, <f>) V^] when A. < fi; 2. Ea+o = Ex; 3. Ex -*¦ 0 as A. ->¦ -oo, and Ex -* I as A. ->¦ +oo (all convergence here is in the operator norm). Then the spectral theorem states that any self-adjoint transformation TofH possesses a uniquely determined spectral family [Ex: A. e R} for which the representation f° = J-o is valid. That is, for any 4> e V(J) we have J—
188 Laplace and Heat Operators where (dEx<f>, <t>) is now a Lebesgue-Stieljes measure (with respect to X) on R, supported on spec T. The domain of T, TKT), consists of those 4> e 7i for which +00. Furthermore, for any Borel function /(X), defined &.e.-[dEk] and finite, we have r = r J- This last equation is a theorem if one takes the functional calculus of self- adjoint operators as already well defined, and is a definition if one wishes to first develop the functional calculus of self-adjoint operators from the spectral family formulation of the spectral theorem. In particular, l|7>ll2 = (T24>, <f>) = (°° X2 {dEx<t>, <t>). J—oo Moreover, if / and g are two such functions satisfying TKfiT)g{T)) = V(g(T)) n V(fg(T)), then f(.T)g(T)= f f(X)g(X)dEk. J-oo Definition Within spec T we distinguish a variety of subsets. First, a number X is in the point spectrum if X is an eigenvalue of T, that is, there exists a nontrivial element fofTi for which Tf = X/. Next, assume T is self-adjoint. Then spec T c R. The discrete spectrum ofT consists of those X € spec T for which there exists t > 0 such that dim(?i+< - EX-t)(H) < oo. So X is an eigenvalue of T of finite multiplicity, and is an isolated element of spec 7". We refer to the complement of the discrete spectrum as the essential spectrum. Therefore, the essential spectrum consists of those X € spec T such that for all € > 0. VII. I Self-adjoint Operators and Their Semigroups 189 Thus X € spec T precisely when there exists a sequence [<f>n} Q T, \\4>n|| = 1 V n, such that (X — T)<pn -*¦ 0 as n -*¦ oo. And X is in the essential spectrum precisely when the sequence can be chosen to be orthonormal. For any X 6 spec T, the sequence {4>n} is referred to as a sequence of normalized approximate eigenfunctions ofX. The spectral theorem implies that if Xo = inf spec T > —oo, then > f Therefore, T is semibounded from below, with best constant € > X<>. Con- Conversely, if T is semibounded from below with constant t, X is an element of spec T, and {<&,) is a sequence of normalized approximate eigenfunctions of X, then 0 = lim ((X/ - Tyt>H, 4>H) = X - lim (T<f>n, <pH)<k- c, so the spectrum is bounded from below, with € < Xo. We conclude that T is semibounded from below if and only if the spectrum is bounded from below, in which case we have (VII. 1.2) infspecT= inf ^ #*> II*II2 Note that the argument only involves those <t> € V(T), but the quadratic form 4> t-+ {T<(>, <t>) is defined on V((T - t)l/2), and the infimum, above, does not change if we allow 4> to vary over V((T - Ol/2) 3 V(J -() = V(T). VII.1.2 Quadratic Forms In defining the Friedrichs extension, above, we started with a semibounded symmetric operator 7 on a dense domain V(T), with which we associated a quadratic form (, )r- We then used the quadratic form (, )r to define a self- adjoint extension $T of the original T. Moreover, the domain of the quadratic form is precisely V((T - Ol/2) [where e is given by (VII.1.1)]. Here we note that one can start with the quadratic form as the fundamental object.
190 Laplace and Heat Operators Definition Let V denote a dense domain in a Hilbert space H. Then a sesquilin- earform Q' on V is a map Q':PxC->C such that 1 • G'C*. y) is linear in x, 2. Q'(x,y) = Q'(y,x). We say Q' is semibounded from below if there exists (eR such that (Vn.1.3) Q'(x,x) > c\\x\\2 for all x 6 T>. We say Q' is nonnegative if € > 0. Given Q' semibounded from below, with t given by (VII. 1.3). Complete V with respect to the inner product = <2'(*. y) + (i - €X*. y)- Assume the resulting space may be identified with a closed domain Dq 5 V in %. Denote the new inner product on Dq by (, )q, and define on Dq the quadratic form Determine the Friedrichs operator $q by sothedomainSgof^econsistsofthosejc € Dq for which there exists x* e H such that (*•, y) = G(x, y) for all y e DQ. Then we define 5e* = x". One knows that 5C is self-adjoint, and that Dq = TH^e'72)- VI1.1.3 1-Parameter Semigroups Definition Given a Banach space B, a family {T,:t> 0} of bounded linear op- operators on B is said to be a C° 1 -parameter semigroup {.semigroup, for short) if: 2. if 0 < s, t < <x>, then T,TS = T,+s; 3. the map (t, /)>-»• T,f from [0, co) x B to B is jointly continuous. We say that the semigroup T, is contractive if \T, || < 1 for all t > 0. Given A) and B) in the definition of a semigroup, to verify C) of the definition it suffices to verify that UjnT,/-/ for all/eB. VII. I Self-adjoint Operators and Their Semigroups 191 Definition With every semigroup T, we associate its (infinitesmal) generator, Z, defined by the domain of Z, D(Z), being those / 6 B for which the limit exists. Proposition VII.13. (a) T>(Z) is a dense linear subspace ofB, and 7",(D(Z)) c V(Z)forall t > 0. Moreover, T,Z = ZT, on all ofV(Z), for all t > 0. (b) Iff e V(Z), then F(t) =T,fis C1 on [0, +co)( and F'(t) = ZF{t). (c) Furthermore, Z isa closedlinearoperator, withV(Z) completewith respect to the norm Moreover, T, acts as a semigroup on V(Z)for this norm, (d) Conversely to (b), if a path *(/) in V(Z) satisfies on some interval [0, a), then *(/) = 7,@@)) Vfe[0,a). Thus, the infinitesmal generator uniquely determines the semigroup, (e) A densely defined operator Z on the Banach space B is the infinitesmal generator of a contractive semigroup if and only if all X > 0 lie in the resolvent set of Z and iiu - zr1 ii <x-' for all X >0. Proposition VII.13 (a) If B is a Hilbert space, and T, is a self-adjoint semigroup acting on B, that is, T, is self-adjoint for each t > 0, then Z is also self-adjoint. IfZ is
192 Laplace and Heat Operators nonnegative self-adjoint, then the spectral theorem implies that —Z is the infinitesmal generator of a self-adjoint contraction semigroup, (b) If Z is a symmetric operator with dense domain V in the Hilbert space B, and for every f €V there exists € = €(/) > Osuch that the heat equation (VII. 1.4) F'(t) = -ZF(t) has a solution satisfying F@) = / and F(t) 6 V for all t € [0, «(/)], then Z is essentially self-adjoint on V, and the solution to (VII. 1.4) subject to the given conditions is unique. VII.2 The Laplacian VII.2.1 The Laplacian Acting on C2 Functions Let M be an n-dimensional Riemannian manifold. If x :U -*¦ R" is a chart on Af, then we have g,, = C,,3;>. G = (*,-). G =(*"), g = detG>0. Recall that, for any C function / on Af, we have (grad /, |) = df{^) = |/ V | e TM. For C functions /, h on Af we have grad (/ + A) = grad / + grad h, grad fh = f gradh + hgrad/. If x:(/ -»• R" is a chart on Af, then Definition For any C vector field X on Af we define the divergence ofX with respect to the Riemannian metric, div X, by (Recall, V denotes the Levi-Civita connection of the Riemannian metric.) For the C1 function / and vector fields X, Y on M we have div fX = (grad /, X> + /div X. Vll.2 The Laplacian 193 If x: U -> R" is a chart on Af, and then Given x e M, and r € [0, c(|)), ^ eS, geodesic spherical coordinates on M about x, with the Riemannian measure in spherical coordinates (where d\ix denotes the standard measure on S*), then for the radial vector field B/dr we have 9 One verifies that if X has compact support on M then we have the Riemannian divergence theorem: divXdV=0. In particular, if / is a function and X is a C1 vector field on M, at least one of which has compact support, then (VII.2.1) f /div XdV = - j (grad /, X) dV. Definition Let / be a C2 function on M. Then we define the Laplacian of /, A/, by A/ = div grad /. Thus, in a chart x: U -»• R", Furthermore, for C2 functions / and h on M we have A) = A/ +AA, div/gradA = f Ah + (grad f, gradh)
194 Laplace and Heat Operators (this last formula only requires that f e Cl,h e C2), which implies A /A = /AA + 2(grad /, grad A> + A A/. One has Green's formulae: Let /: M ->• R 6 C2(M), A: Af ->• R e C'(Af), with at least one of them compactly supported. Then A A/ + (grad A, grad/>}dV = 0. u If both / and A are C2, then f {AA/-/AA}</V = 0. Ju Let M be oriented, ft a domain in M with C°° boundary 3ft, and v the outward unit vector field along 3ft that is pointwise orthogonal to 3ft (there is only one such vector field). Then for any compactly supported C vector field X on M we have // divXdV= / (X, v)dA. JJa Jaa The corresponding Green's formulae are: Given M, ft, and v as just described, and given / 6 C2(M), A e C\M), at least one of them compactly supported. Then (VII.2.2) ff {AA/ + (gradA, grad/)} dV = f A(v, grad/> dA. JJa Jen If both / and A are C2, then (VH.2.3) ff [hAf-fAh}dV= f (A(v,grad/> - /(v, gradA)}<M. J Ja Jsa Notation We generally write 3//3v for (v, grad />. VH.2.2 The Laplacian as an Operator on L1 Until now, we have only considered the pointwise action of the Laplacian on functions which are C2. We now wish to view the Laplacian as an operator on j the Hilbert space L2 = L2(M, dV). For any two functions /, g in L2 we have the inner product and norm (/.*)= / fgdV, \\ff= f fdV. Ju Ju VII.2 The Laplacian 195 We may also speak of L2 vector fields, and the associated inner product and norm norm (X,Y)= f (X,Y)dV, \\X\\2= f lx\2dV. JM Jm Recall that, for any subset A of M, [A]r denotes the set of all points in M with distance from A less than or equal to r. We now consider the action of the Laplacian A on L2. Recall that Green's formula states that if / € C2, h e C, one of which is compactly supported, men (VH.2.4) (-A /, h) = (grad /, grad A). Let Ac = A|C~. Then — Ac is a nonnegative symmetric operator on the dense subspace Cf of I2, with associated quadratic form Q'c = D[<p, yjr] := (grad <p, grad ^), 0, $ e Cc°°. We refer to D[, ] as the Dirichlet energy integral. Complete Cf to Sjc relative to the metric on C?°, and define the Friedrichs extension $c — Ja, of Ac on the domain ©c = DAf. So Cf c T>c C f)c- Henceforth, unless otherwise noted, the self- adjoint Laplacian on L2 will be the Friedrichs extension ffc. We will refer to $c as the Friedrichs extension of the Laplacian on M. The identification of Sjc goes as follows: Let H denote the collection of C°° functions f on M for which both /, grad/ 6 L2. Endow H with the inner product (/, g)H = D[f, g) + (/, g), f,g€H, and complete H, with respect to the inner product (,)//, to the Hilbert space fj. Then Sjc is the closure of Cf in Sj. For the elements of Sj themselves, we consider weak derivatives. Namely, Definition We say that / e L2 has a weak derivative if there exists an L2 vector field X such that (/, div Y) = -(X, Y) for all compactly supported C°° vector fields Y on M. The vector field X, should it exist, must be unique, and we denote it by X = Grad / [see (VII.2.1) above].
196 Laplace and Heat Operators By the Meyers-Serrin theorem, f) is the Hilbert space consisting of functions f e L2 possessing weak derivatives, with the inner product (, )fl given by (/. g)r> = (Grad /, Grad g) + (/, g), f,gef). Until now, M was an arbitrary Riemannian manifold. Assume, now, that M has C°° boundary F and compact closure. Then we may consider the Laplacian acting on functions with prescribed boundary conditions, chosen so that the action is symmetric, from which we then determine the new Friednchs extension for the domain of functions in question. The simplest is: Definition (Vanishing Dirichlet Boundary Conditions) Let Pdir = {/ e C°°(A7): f\r = 0}, p Then —Adi, is nonnegative symmetric, and has associated quadratic form and inner product given by for all 0, \ff 6 Ddir. Complete the inner product (, )<«, to the subspace D&, and thereby determine the Friedrichs extension $& of - Adu, on 33^, the Dirichlet Laplacian on M. Lemma VIL2.1 Let M have C°° boundary F and compact closure. Then given any a, S > 0, there exists a Junction p e Cf(M) such that the support of p has distance from F greater than S, that is, supp p C M \ [F]j, and Proof Since M is compact, it has a constant c and a finite covering [U,:i = 1 ,...,€} by charts, for which the Riemannian metric in any of the charts satisfies U J for all I = (^ |"). Then mere exists a constant K such that for every i = 1,..., I, and every C°° function F compactly supported on Ut we have (VII.2.5) ||F||«2 + llgradFH*2 < *{||F||R.2 + HgradF||R.2}, where the subscripts R" and M refer to the respective L2 spaces. Fix the cover 14 = {U,:i = l,...,l], with subordinate partition of unity {<p,). VI1.2 The Laplacian Now consider the function q{ : M -» R defined by 197 I = j d(q, D/& - 2, qe Then Subject each <p,es to mollification by j(t, as in §1.3; namely, for the index « = («i €t), define Pip) = Qi.eiP) = ^>2 (/«, *(<P,Qs)ox,-l)(x,(p)) i=i i /. = ? / A,(y. -x.(P))((*>«^)ox,-|) i=i J Then one can easily use Theorem 1.3.3 and (VII.2.5) to prove the lemma. ¦ Lemma VTL2.2 Let M have C00 boundary V and compact closure, and let f:~M^>Rbe Lipschitz on ~M, with /|F = 0. Then, given any € > 0, there exist S > 0 and a function <j> € C™(M) such that the support of<j> has distance from F greater than h, that is, supp# C M \ [F]4, and Corollary VIL2.1 The Dirichlet Laplacian gfa on M coincides with the Friedrichs extension 3v of Ac on M. In particular, their associated quadratic forms coincide, possessing the common domain f)c. Proof of Lemma VII^-2. Let p: M -» [0, oo) 6 Cf satisfy the properties of Lemma VII.2.1, and let <p = fp. Then ||/ - <f,\\2 = f f\\- pJ dV < consL/S, by a standard calculation in Fermi coordinates based on F. Similarly, grad(/ - <t>) = grad/(I - p) = A - p)grad/ + /grad(l - p).
198 Laplace and Heat Operators which implies lgrad(/ - <t>)\\p) < 2{|1 - p|2|grad/|2 + |/|2|gradp|2}(p) < const., [the d(p, VJ in the numerator of the second term of the parentheses follows from / vanishing on F]. But then calculation in Fermi coordinates, based on f, implies ||grad/-grad0||2= f |grad/(l - p)\2dV < const/5, which implies the proposition. ¦ Remark VIL2.1 We shall show in Theorem VII.4.2 below mat the Laplacian with vanishing Dirichlet boundary condition is essentially self-adjoint. Definition A function / e L2 on M has A/ acting as an L2 distribution on M if mere exists g e L2 such that *,*) = (/.A*) We write A/ for g. Theorem VII.2.1 If M is a complete Riemannian manifold, then Sjc = S). Given any function f e C°° for which /, A / € L2, we also have grad / € L2 and (VII.2.6) (-A/,/) = (grad/, grad/). Finally, Ac = A|C?° is essentially self-adjoint, so the Friedrichs extension & of Ac is the unique self-adjoint extension of Ac. Corollary VII.2.2 Let M be a complete Riemannian manifold. Then M has no L2 nonconstant harmonic function, that is, any function f in L2(M) satisfying A/ = 0 on all ofM must be constant. Moreover, given any X < 0, there is no solution u 6 L2 to the equation An + An = 0 except for u = 0 identically on M. Proof of Theorem VII.2.1. Assume / e f){M) n C°°(Af). Fix o e M, and for each €, R > 0, consider the function fR = ftf>R, where 4>r e Cf(B{o\ R + 3)) Vll.2 The Laplacian 199 satisfies tpR\B(o;R)=l, 4>r, Igrad^l < 1+t; such a function 4>R is constructed in Lemma VII.2.1. Then 11/ - Ml2 = / /2<1 - 4>kJ dV < const. / f2 dV -+ 0 as R -*¦ <x>, since / e L2. Similarly, j -4>R\2\gP>df\2}dV lgrad/|2}dV < if f M\B(o:K) < const . / Jkr\B(<r,R) asR -*¦ oo. Therefore/is aipproximatedinij by functions in fyc. Since Sj n C00 is dense in f) (by definition), we obtain fyc = fy. Assume we are given the function / e C°° such that /, A / e L2. Then which implies ^{Igrad/I2 + /A./} = -2J 4>Rf (grad**, grad /> < \ J 4>R2\&ad f\2 + 2 f f2\grad4>R\2, which implies 5 /" z JB B(o;R) lgrad/|2<^ < const. / /2 + |/A/| < +00
200 Laplace and Heat Operators for all R. This implies grad / 6 L2. But f 0*2{|grad/|2 + /A/} = -2 j <t>Rf (grad**,grad/) = -2 / ^/(grad^.grad/) ./i»(o:fl+3)\i»(o;fl) < const. / I/Ilgrad/|, ./i»(o:fl+3)\i»(o:fl) As /? -> oo we have /" ^{Igrad/I2 + /A/} -> y |grad/|2 + /A /, |/||grad/| -> 0, which implies (VH.2.6). For the essential self-adjointness of Ac we consider L2 to be a complex Hilbert space with inner product Ju fgdV, and with the Laplacian defined by operating on real and imaginary parts of a complex-valued function. It is standard that Ac is essentially self-adjoint if and only if ker{Ac±j}* = 0, that is, the subspace of functions / e L2 that satisfy is trivial. So assume {Ac ± j }•/ = 0. Then (A ± /)/ = 0 as an L2 distribution. Elliptic regularity then implies that / € C°°, and A / in the classical sense is the L2-distributional Laplacian of /. Therefore A / ± // = 0 on all of M. Multiply both sides of the equation by / and integrate over M. Then - I jAfdV= [ |grad/|2dV Ju Ju for all / e L2 n C°° for which A / 6 L2 [the purpose of this comment is to VII. 3 The Heat Equation and Its Kernels 201 prove that (A/, /) is real]; therefore A / + if = 0 implies if I/I2= f Igrad/I2, Ju Ju which implies / = 0. A similar argument applies to a solution of A / — if = 0, and the theorem is proven. ¦ VIIJ The Heat Equation and Its Kernels We are given a fixed Riemannian manifold M of dimension n > 1, with associ- associated Laplacian A acting on C2 functions on M. The heat operator L associated with the Laplacian acts on functions / = f(x, t) in C°(M x @, oo)), and is given by where / is C2 in the space variable x e M, and C in the time variable / e @, oo). The homogeneous heat equation, or heat equation, is given by Lu = 0, that is. —, dt and the inhomogeneous heat equation is given by Lu @, oo) ->¦ R € C°. = — F, where F:Mx The physical interpretation of the equation is given by considering the Riemannian manifold as a homogeneous isotropic medium (with conveniently normalized physical constants so that, for example, heat and temperature can be considered one and the same), and u(x, t) is the temperature of x 6 M at time /. Then the inhomogeneous heat equation describes the evolution of the temperature distribution, with F(x, t) the instantaneous rate with respect to time at which (by some external source) heat is supplied to, or withdrawn from, x at time r. The argument is predicated on Newtonian heat conduction, namely, given any domain n in M with compact closure and C°° boundary, the instan- instantaneous change of the total temperature in fi with respect to time is equal to the total change due to the supply of heat to, or withdrawal of heat from, ft, added to the spatial rate of change of the heat distribution across the boundary of ft, 9ft. That is, Jaa Sv (w,t)dA(w).
202 Laplace and Heat Operators By Green's formula we have which implies IL for all such domains S2 in M. This then implies the inhomogeneous heat equation on all of M. Notation Henceforth, if there is no comment otherwise, given any function /: M x @, +00) -> R, we denote the spatial component of the gradient of / by grad /. Definition A heat kernel is a continuous function p:MxMx@,oo) such mat for every y e M the function u(x, t) = p(x, y, t) is a solution to the heat equation with initial data lim u(x, t) = &y(x), no the delta function concentrated at y, namely, lim no f for every bounded continuous 0: M -*¦ R. Intuitively, p(, y,) is the solution of the heat equation resulting from an ini- initial temperature distribution having total temperature equal to 1 concentrated at y. We use the linearity of the heat equation (otherwise known as superposition of solutions) as follows: If at time t = 0 the initial temperature distribution was concentrated at y with total temperature a, then the solution to the heat equation will be u = ap(, y,). If at time / = 0 the initial temperature distribution was concentrated at the points y and z with total temperatures a and fi, respectively, then the solution to the heat equation will be u = ap( ,y,) + fip(, z,). There- Therefore if at time t = 0 we are given the temperature distribution /(>), then by summing "spatially" the contribution of each point to the initial data, we obtain with no u(x,t)= [ p(x,y,t)f(y)dV(y), Jm , t) = lim / p(x, y, t)f(y)dV(y) = f(x). noJu VII.3 The Heat Equation and Its Kernels 203 If, in addition, heat or refrigeration is supplied to M, as described by F(x, t) above, then the contribution of f(, r), r e @, /), to the temperature distribution at time / is given by p(x,y,t-z)F(y,T)dV(y), / since we think of the function F(, r) as initial data for heat diffusion starting at time r and lasting for t — t time units. So the total contribution to the temperature distribution, at time /, by F is given by summing "temporally," namely, [ dr [ p{x,y,t-T)F{y,x)dV(y). o Ju Therefore, we expect the solution to the inhomogeneous heat equation with initial temperature distribution / and external source F to be given by «(*,»)= f p(x,y,t)f(y)dV(y)+ f dr f p(x,y,t - r)F(y, r)dV(y). Jm Jo Jm Remark VIL3.1 Note that we have argued that p(x, y, t) is symmetric with respect to Jt and y as 11 0, since p(x, ,t)-> Sx, intuitively, and p(, y, t) -*¦ Sy, by hypothesis, as / ! 0. We shall see shortly that in all cases that we study, have the symmetry p{x, y, t) = p(y, x, t) on all of M x M x @, +00). y we VII.3.I The Heat Kernel of Euclidean Space In order to calculate the heat kernel on R", we first derive a candidate, using the Fourier transform. Then we check that our candidate is indeed legitimate. Given a function/:R" -»• C in L'(R"), define its n-dimensional Fourier trans- transform, /, by tL" -ixi dvH(x), where The basic properties of the Fourier transform are: 2. g(x) = f(x - y) => g(?) = «-'>•«?«); {*"-*/} (?) = ?» - y).
204 Laplace and Heat Operators 3. {/(*x)f($) = A-?(|A); {/*gftf) = /(?)?(*),wheref*gdenotes the convolution (normalized here to be) (/ * gXx) = Bjt)-" 4. One also has the specific example S. Consider the collection S of functions on R" that, with all their partial derivatives of all orders, have rapid decrease on R". For functions / in 5 we have The Riemann-Lebesgue lemma implies, with these last two formulae, mat the Fourier transforms maps S into S. 6. Finally, one has the Fourier inversion formula: fix) = B*r"/2 f ?«y*-* <*Ў Now, consider a solution u(x, t) to the heat equation on R", with initial data 4>(x). Since this is a formal calculation in search of a candidate solution, we treat all functions as though they were in S, with respect to the space variable for every t. Let v(%, t) be the Fourier transform of u(x, t) in the space variables, that is. f JtL' u(x,t)e-ix(dvn(x). Then jp : [ Au(x,t)e-ix*dvH(x) Jk> ,t)e-ixidvH(x) that is, W/. J 77» //ca/ Equation and Its Kernels dv 205 For any fixed ?, view the equation as an ordinary differential equation in t, from which one has To evaluate c(?), let / = 0. Then one obtains c(?) = ^(?), the Fourier transform of the initial data of 0. We conclude If one now inverts the Fourier transform, to recapture u(x, t), one obtains Our candidate, therefore, for the heat kernel on R" is One can directly verify that E is C°° on R" x R" x @, oo) and is a solution to the heat equation on R" in x and /. We now check that lim E(x, y, t) = 8y(x). First start with any nonnegative integrable function p: R" -*¦ R with /p(x) dvH(x)=l, and set p((x) = €-"p(x/(). Then /pf (x)dvn(x) = 1, and it is standard that for any bounded continuous function f(x) on R" we have lim / p((x-yWx)dvn(x)=Ay) <io J [see Proposition l(b), where we have j( compactly supported on R"]. So, for any fixed y, Hm p((x -y) = Sy(x). In our situation, we have p(x) = 7i-»/2e-W\ E(x, y, t) = p1j(.x - y), which implies the claim; so E is a heat kernel on R". Two important properties, which we see repeatedly, are the symmetry and positivity of the heat kernel,
206 Laplace and Heat Operators namely, (VII.3.1) Eix,y,t) = Eiy,x,t)>0, for all x, y e M and t > 0. For a function fix) on R\ we have the following results for solutions to the initial value problem for the homogeneous heat equation: (VII.3.2) Proposition VII.3.1 (a) If fix) is bounded continuous, then the function (VII.3.3) uix,t)= f Eix,y,t)f(y)dvniy) is a solution of (VII.3.2). (b) We always have (VH.3.4) / Wx,y,t)dvHiy)=\ for all (x, t) 6 R" x @, <x>). Iff is a function in L\ and u(x, t) is given by (Vn.3.3), then (VII.3.5) uix,t)dvnix)= I fix)dyHix) R- JV for all t >0. (c) If f e C° is supported on the compact set K, and uix, t) is given by (VII.3.3), then Wix,t)\ < (&n)-*l*e-***>KVM \j \fiy)\dvHiy)\ . (d) If fix) is continuous on R", vanishing at infinity, that is, lim fix) = 0, |jr|-»oo and uix, t) is given by (VII.3.3), then lim uix,t) = 0, lim (gradii)(x, t) = 0. |jf|->00 l*|->0O (e) If fix) is bounded continuous on R", then from among all functions vix, t) on R" x @, oo) satisfying \vix, t)\ < const., VII. 3 The Heat Equation and Its Kernels 207 where the constant is independent of\x, t), thefunctionuix, t)givenbyiVU3.3) is the unique solution to the initial value problem (VII.3.2). For the initial-value problem for the inhomogeneous heat equation we have Proposition VIL3.2 Given F: R" x [0, oo) -»• R e C1 such that supp F C K x [0, oo) for some compact K in R", then the function uix, t) = j^ Eix, y, t)fiy) dyHiy) + jf' dx ? ?(*, ,,, - T)F(y, x) dyn satisfies 8u (VII.3.6) Am - — = -Fix, t), \imuix, t) = fix), at t\o VII.3.2 Preliminary Principles Proposition \U33 (DuhameTs Principle) Let M be a Riemannian manifold with compact closure and C°° boundary ipossibly empty), and let u, v: M x @, t) -*¦ R e C be C2 with respect to the space variable in M. Then for any [a, 0] C @, t) we have jj Wiz, t - 0)viz, P) -11B, / - a)viz, a))dViz) = / dx f[ {Luiz, t - x)viz, x) - uiz, t - r)Lviz, x)} dViz) dx / \—r-iw,t-x)viw,x) dv 1 , t - x)-—(iy, r)} dAiw). dvw I Proposition VIL3.4 (Strong Maximum Principle) Let u be a bounded con- continuous function on M x [0, T] that is C2 on M x @, T) and that satisfies A—->0 on M x @, T). If there exists ix0, t0) in M x @, T] - note the half open interval - such that time «(*o. 'o) = sup ii, «x[0J|
208 then Laplace and Heat Operators u\M x [0, (ol = u(xo, t0). Furthermore,ifMhasC°°boundary,u e C°(M x [0, T]),and(w, t\) e dM x @, T] satisfies then u(w,t\)= sup u, Mx[0,T] du —(u). Remark VII3.2 If one is given An — du/dt < 0, then one has a correspond- corresponding minimum principle. For solutions of the heat equation, both principles are valid. VII.3.3 Properties and Types of Heat Kernels Definition Given a Riemannian manifold with heat kernel p(x, y, t), we say that p(x, y, t) is symmetric if p(x, y, t) = p(y,x,t) on all of M x M x @, +oo). We say that p(x, y, t) is positive if p(x, y,t)>0 on all of M x M x @, +oo). We say that p(x, y, t) satisfies the conservation of heat property, or that P(x, y, t) is stochastically complete, if / P(x,y, Jm t)dV(x)=l for all (y, t) e M x @, +00). For M noncompact, we say that p(x, y, t) satisfies the Feller property if lim p(x,y,t) = JT-KJO for all (y, t) e M x @, +00). Remark VII.3.3 If p is a heat kernel satisfying the symmetry property, then . /) = (AypX*. y, t). Definition Let M be a compact Riemannian manifold with no boundary, a closed manifold. Then any heat kernel on such a manifold will be referred to as a closed heat kernel. VII.3 The Heat Equation and Its Kernels 209 Definition Let M be a Riemannian manifold with compact closure and C°° (nonempty) boundary. Then a Dirichlet heat kernel q: M x M x @, 00) 6 C^ is^heat kernel on M that can be extended to a continuous function on M x M x @, 00) such that q(,y,t)\dM = 0 for all (y, t) e. M x @, 00). Thus the vanishing Dirichlet boundary data will correspond to absolute refrigeration of the boundary. Theorem VH3.1 Let p be either a closed or a Dirichlet heat kernel on a Riemannian manifold M. Then p is symmetric in the two space variables in M. One has uniqueness of heat kernels in both cases, that is, a closed Riemannian manifold has at most one heat kernel, and a Riemannian manifold with compact closure and C°° boundary has at most one Dirichlet heat kernel. Proof Fix one of the two cases. Let p\, p2 be two kernels on M, set u(z, r) = pi(z, x, r), v(z, r) = p^z, y, r), and apply Duhamel's principle. Then the boundary integrals in Duhamel's prin- principle (should the boundary be nonempty) vanish. Let a 4. 0, 0 f /. Then one obtains P2(x,y, 0 = pi(y,x,t). If we apply the argument to any heat kernel p\ = P2 = p, then we have P(x, y, t) = p(y, x, t) for all x, y e M. So any heat kernel is symmetric in the space variables. But given arbitrary heat kernels p\, pi, we may apply the symmetry to p\ and thereby obtain p\ = pi. ¦ Theorem VII.3.2 Let pbea closed heat kernel on a Riemannian manifold M. Then p satisfies the conservation of heat property. Proof Simply note that ? [ p(x,y,t)dV(x)= [ ^p(x,y,t)dV(x) ot Ju Jm °t which implies the claim. -L , y, t)dV(x) = 0,
210 Laplace and Heat Operators Given any Riemannian manifold M, set We do not presume that ? will actually be a heat kernel on M, but it should be some sort of approximate heat kernel. The question is in what sense, and with what validity. The answer, for our purposes, goes as follows: Proposition VII3.5 Assume M isa closed Riemannian manifold. Then M has a closed heat kernel, which we know is unique. Similarly, assume M is a Riemannian manifold with C°° boundary and com- compact closure. Then M has a Dirichlet heat kernel, which we know is unique. In either case, the heat kernel is locally Euclidean in the sense that, for any compact K in M, P(x, y, i (VII.3.7) lim no ?(x,y,t) uniformly for all x, y e K satisfying = 1 + OK{d(x, y)) d{x, y) < mm j - j. Theorem VII33 Assume p is a closed or Dirichlet heat kernel on a Riemannian manifold M. Then p is always positive. Proof From (VII.3.7) we know that the heat kernel p(x, y, t) assumes positive values, and moreover is unbounded from above on M x M x @, T] for any T >0. If we are given the closed heat kernel p, and there exists (xo. yo, 'o) in M x M x @, +oo) such that p(*o. yoi 'o) < 0, then, for fixed yo, u(x, t) = P(x, yo, t) has a nonpositive minimum value on M x @, /0], which implies by the strong minimum principle that there exist t\ e @, t0) and a nonpositive con- constant S such that u(x, t) = S on all of M x @, t\ ], which implies a contradiction. If we are given the Dirichlet heat kernel p, then for any fixed >o in M, the function u(x, t) = p{x, yo, t) vanishes on the boundary 3M. Therefore, if u is not positive everywhere on M, then there there exist (xo, t0) in M x @, +oo) such that p{xo, yo, 'o) < 0. The argument is then the same as above. ¦ Remark VIIJ.4 Let p be a Dirichlet heat kernel on M. Since p(x, y, t) > 0, then dp ?(wyt)<0 V1I.3 The Heat Equation and Its Kernels 211 on all of 3 M x M x @, +oo). In particular, ? f p(x,y,t)dV(x)<0, ot Ju which implies (VH.3.8) f p(x,y,t)dV(x)<l, on all of M x @, +oo). Definition Let M be an arbitrary noncompact Riemannian manifold. We say that the positive heat kernel p is the minimal positive heat kernel if, for any other positive heat kernel P(x, y, /) on M, we have p < P on all of M x M x @, +oo). Certainly, the minimal positive heat kernel is unique, by definition. For the existence of the minimal positive heat kernel one starts with domains Ci in M with compact closure and C°° boundary, and associates with each ?1 its Dirichlet heat kernel qn- An application of the maximum principle implies that for domains J2i cc Q2 we have on ?2i x &! x @, +00). The minimal positive heat kernel of M, p(x, y, t), is given by P = sup qQ, a where we set qa to be identically equal to 0 on the complement of $2 x $2 x @, +oo)inAf x M x @, +00). For any exhaustion $2, f M by relatively com- compact subsets of M, we have the monotone convergence of qu, t P uniformly on compact subsets of M as j -> 00. All derivatives of qu, converge to those of p uniformly on compact subsets of M as j -> 00. Then p is positive on all of M x M x @, +00), it is the minimal positive heat kernel of M, it is symmetric, and it satisfies (VII.3.9) P(x,y,t)dV(x)< for all (y, t) e M x @, +00). So, for 0 bounded continuous, the function u (x, t) given by «(*,»)= ? p(x,y,tL>(y)dV{y)
212 Laplace and Heat Operators satisfies the heat equation, with lim u(x, t) = <f>(x) no for all x e M. Definition Given a Riemannian manifold M, we say that M satisfies the unique- uniqueness property for the heat equation if for any bounded continuous function / on M there is at most one bounded solution to the heat equation on M satisfying lim u{x, t) = f(x). Remark VII3.5 Of course, if M satisfies the uniqueness property, then M has a unique heat kernel. Remark VII3.6 Assume M satisfies the uniqueness property. Since for any bounded continuous f(x) the function (VH.3.10) u(x,t) = f p(x,y,t)f{y)dV(y) Jm is a solution to the initial value problem for initial data /, we conclude that every solution to the initial value problem for the heat equation on M with bounded continuous initial data / admits the representation (VII.3.10). In particular, the solution u(x, t) = 1 admits the representation 1= I p(x,y,t)dV(y), Jm which implies M is stochastically complete. Theorem VIL3.4 Let M be a compact Riemannian manifold. Then M satisfies the uniqueness property. Also, ifM has compact closure and C°° boundary, then given the boundary data u(w,t) = U(w,t) V(iy,/)e3Af x(O.oo) for some function U(w, t), t > 0, M satisfies the uniqueness property. Proof Let u(x, t) be a solution to the heat equation on M, in either of the two cases. VII.3 The Heat Equation and Its Kernels Then 213 (uAu)(x,t)dV(x) = - ff \gndu\2(x,t)dV(x)+ f (up-)(w,t)dA(w) Therefore, if we have two solutions v\, V2 to the same initial-boundary value problem, namely, vj (x,0) = V2(x, 0) = 4>(x) VxzM and (if there are nonempty boundary data) v\(w,t) = vz(w,t) V(iy,/)e 8M x@,oo), then the solution u = v\ — V2 has both vanishing initial and boundary data. But its L2 integral does not increase with respect to time. So it stays constantly equal to 0. So vi =V2. Therefore, solutions to the initial-boundary value problems are unique. ¦ Remark VIL3.7 If M has C°° boundary and compact closure, then the Dirichlet heat kernel of M is equal to the minimal positive heat kernel of M, that is, p —qu- Proposition VIL3.6 Assume M is Riemannian complete, with Ricci curva- curvature bounded from below. Then M has the uniqueness property and the Feller property. curva- curvaCorollary VDL33 Assume M is Riemannian complete, with Ricci curvature bounded from below. Then M is stochastically complete. Remark VIL3.8 If M does not satisfy the uniqueness property, then the coun- counterexample to uniqueness cannot come from the minimal positive heat kernel
214 Laplace and Heat Operators and solutions of the form (VII.3.10), because (VII.3.9) implies there is only one solution of the initial value problem with initial data / of the form (VII.3.10). Indeed, (VII.3.10) and (VII.3.9) imply L P(x,y, So / = 0 implies u = 0, which is uniqueness. Note the remark applies to any positive heat kernel P on M satisfying the inequality (VII.3.9) for all (x, t) e M x @, oo). On the other hand, if the minimal positive heat kernel p of M is stochastically complete, then p is the unique heat kernel among all positive heat kernels on M satisfying (VII.3.9). Indeed, if P is a positive heat kernel on M satisfying (VII.3.9), then P > p on all of M x M x @, oo), which implies 1= I p{x,y,t)dV(y)< I P(x,y,t)dV(y)<l, JM JM which implies the claim. Definition Let M be a Riemannian manifold that satisfies the uniqueness prop- property for the heat equation. Then to every bounded continuous function 0 on M one assigns the heat flow t v-* H,<f> of 0, defined by = u(x, t), where u(x, t) is the solution of the initial value problem with initial data <j>. Theorem VII3.5 Let M be a Riemannian manifold that satisfies the unique- uniqueness property for the heat equation. Then the heat flow H, is given by = I p(x,y,t)<t>(y)dV(y), Ju where p is the unique heat kernel ofM. Furthermore, the heat flow H, satisfies the semigroup property, namely, +s = H,oH, forallt,s > 0. Proof Given a bounded continuous function /, both u(x, t) = Ht+,f(x), v(x, r) = Hl0H,f(x) are solutions to the heat equation. Both have the same initial values u(x, 0) = v(x,0) = H,f(x), which is bounded continuous. The uniqueness then VII.3 The Heat Equation and Its Kernels impliesthat^, T, = v(x, r)foral](,; r) ? „ x (Q 215 Theorem VII3.6 Let M be a Riemannian manifold with compact closure and C°° boundary, and let Q, denote the heat flow on M subject to the i ' constraint that the solution vanishes ondM x @, oo). Then Q, is given by (Q,4>)(x)= f q{x,y,tL>{y)dV(y), JM where q is the Dirichlet heat kernel ofM. Furthermore, the heat flow Q, satisfies the semigroup property. Proof The same as the previous theorem. ¦ Definition Given a Riemannian manifold M with heat kernel p. For any / > 0, the heat kernel defines an integral operator P, by (P,f)(x)= f p(x,y,t)f(y)dV(y) JM for any bounded continuous function / on M. Then foralljce M. Theorem VLL3.7 Let pbea closed, Dirichlet, or minimal positive heat kernel on the Riemannian manifold M. Then p satisfies .11) p(x,y,t+s)= f p(x,z,t)p(z,y,s)dV(z) JM for all x, y andt,s. Proof In the first two cases, M satisfies the uniqueness property, which implies the semigroup property. For the minimal positive heat kernel, we first note that the Dirichlet heat kernel q of a domain 12 in M satisfies (VII.3.11). Next, for an arbitrary noncompact M, and an exhaustion $2y of M by domains with compact closure and C°° boundary possessing Dirichlet heat kernels qjy one obtains (VII.3.11) for p by passing to the limit of (VII.3.11) for qj. ¦
216 Laplace and Heat Operators VII.4 The Action of the Heat Semigroup Theorem VII.4.1 Let Mbea Riemannian manifold with minimal positive heat kernel p (this includes closed and Dirichlet heat kernels). We denote by C the space of functions Cf(M). Then for each positive k, the action of P, on C extends to a contractive, continuous semigroup on Lk(M). Furthermore, P, is positive self-adjoint on L2(M). Thus, any eigenvalue of P, must be positive, with C°° eigenfunction. Proof First, (VII.3.9) and the symmetry of p imply (VH.4.1) f p(x,y,t)dV(y)<l Jm for all (x, t)e M x @, +00). For / e C we have \\P,fh= f dV(x)\f p(x,y,t)f(y)dV(y) Jm \Jm < f dV(x)f P(x,y,t)\f(y)\dV(y) Jm Jm = I \f(y)\dV(y) [ p(x,y,t)dV(x) JM JM < / \f(y)\dV(y) Jm = ll/lli, that is, HP,/Hi < ||/|h; in particular, ||/»,||i_»i < 1. Similarly, Haider's inequality and (VII.4.1) imply 11^/11**= / 1/ JM \JM p(x,y,t)f(y)dV(y) dV(x) < [ dV(x) f p(x,y,t)\f(y)\kdV(y) Jm Jm < I \f(y)\kdV(y), Jm j thatis,||P,/|U<||/||tforalU6[l,oo),/eC. ¦ Because the subspace C is dense in L\ for each k e [1, oo), we may extend the bounded operator P,\C -* C°° to a bounded operator on Lk satisfying < I for all k e [1, oo] (the case k = oo is in Remark VII.3.8). VII.4 The Action of the Heat Semigroup 217 Next, given k e [1, oo), we approximate / e Lk by <f> e C. Then P,f ~f = P,{f -<t>) + P,4>-4> + 4>-f\ which implies limsup \\P,f - f\\k < limsup {||P,{/ - ^}|U + \\P,4> -<t>h + 11^ - fh) no no = O(\\f-4>\\k) as <p -»¦ / in Lk. Thus P,f -*¦ f in Lk for all k e [1, oo). Therefore P, is continuous at / = 0. The semigroup property remains valid on Lk for all k. Because P, is bounded, with symmetric kernel, P, is self-adjoint on L2. Next, consider P, acting on L2, t fixed. Then (P,f, f) = (Pt/2P,/2f, f) = (P,/2f, P,/2f) = \\P,/2ff > 0 (we mean L2 norm unless otherwise indicated). If there exists / such that P,f = 0, then the above shows that P//2/ = 0. One uses the continuity of the semigroup to show that this implies that / = 0. Therefore the quadratic form associated with P, (as a self-adjoint operator) is positive definite, and all eigenvalues of P, must be positive. Moreover, any eigenfunction of P, must beC00. _ As discussed in §VII.1.3, the semigroup P,:Ll -*¦ Ll has generator A* defined by ,/ lm, uo / with domain T>(K() consisting of functions for which the limit on the right hand side exists in Ll. Then T>(K() is dense in Ll, and T>(Et) is complete with respect to the norm lfte = Wfh + IIA,/||,. Notation We shall write A for A2. Henceforth we shall fix P,: L2 -*¦ L2. Since P, is self-adjoint for all / > 0, we have A is self-adjoint. It remains to identify A. We know that it is a self-adjoint extension of Ar, the Laplacian acting on C™.
218 Laplace and Heat Operators Theorem VII.4.2 Assume M has compact closure with C°° boundary, Q, the heat semigroup associated with the Dirichlet heat kernel ofM. Then A is equal to the Friedrichs extension 5c of Ac, the Laplacian acting on C™(M). Proof Let Z>dir consist of those functions / 6 C°°(M), which satisfy f\dM = 0, so PdSr determines the Laplacian Adir associated with vanishing Dirichlet boundary data, with Friedrichs extension 5dir = 5c (see Corollary VII.2.1). Since Q,{V&1) 9 T>At for all / > 0, Proposition VII. 1.3 implies that A*, is essentially self-adjoint. So, when M has compact closure and C°° boundary, 5C is the unique self-adjoint extension of A^r- In particular, the Friedrichs extension 5c = A. ¦ Remark VII.4.1 It is worth noting that once we have the self-adjoint Dirichlet Laplacian realized as the Friedrichs extension of the function space Cc3o(Af), we have removed any mention of the boundary, its regularity, and the associated boundary data from the characterization of the self-adjoint extension. There- Therefore, for an arbitrary Riemannian manifold M one may refer to the Friedrichs extension of C™(M) as the Dirichlet Laplacian. So one is always speaking of a specific self-adjoint extension of the Laplacian. The issue of genuinely identi- identifying the domain of the Dirichlet Laplacian is then a separate question, should one be interested in it, or should one require the specific information. Then the details of the geometry of M and its boundary would come to the fore. Assume M is a complete Riemannian manifold. Then, by Theorem VII.2.1, A is the unique self-adjoint extension of Ac. In particular, A agrees with the Friedrichs extension 5c Theorem VII.43 Let p denote the minimal postive heat kernel of an arbitrary Riemannian manifold M, with associated semigroup P,. Then (Vn.4.2) Hgrad Pt<p\\2 = -(ft*. A/>,0) forall4>eC™,t > 0. Proof (Here, A still denotes the pointwise Laplace operator.) Start with an exhaustion of M, Dj f M, consisting of relatively compact domains with C°° boundary, and their attendant Dirichlet heat kernels qj = qD, t P- Associate with each qj its semigroup QJ. Then for any function <f> e Cc°°(Af), we have QJ,<t> -> P,<p, grad Q{<t> -*¦ grad as j VII.4 The Action of the Heat Semigroup oo, uniformly on compact subsets of M. This implies 219 = -(Qi<t>, A<f>) [the fourth line follows from qj f p, and the fifth line follows from Proposition VII.1.2(a)]. For any compact K in M and e > 0 there exists j0 such that for all j > jo we have f JK < f Ju which implies / Igrad P,4>\2 dV < liminf / Igrad QJ4>\2 dV In particular, letting e -J. 0 and K f M, we obtain So grad P,4> e L2. Therefore, given 4> e Cf, we have P,<p e Sj = I>(A1/2), the domain of the quadratic form associated with A. Next, we note that the sequence QJcp is a Cauchy sequence in f). Indeed, we certainly have QJ,4> a Cauchy sequence in L2. As above, we also have fgrad QJ,<p - grad Qfyf = -(QJ,4> - = -(QJ,<t> - ^¦0 , AQj<t> , QJ,A<t> as j, k -*¦ oo. So Qjtfr is Cauchy in fy. Therefore, there exists * e fy such that QJ,<p -* * in Sj as j -*¦ oo. Then * has an L2 weak derivative, that is, there exists an L? vector field «C on M such that (¦, X) = -(*, div X)
220 Laplace and Heat Operators for all C°° vector fields X on M with compact support. But then for such a vector field X we have = lim -@/0,divX) j-*oo ' = Urn (grad QJ,4>, X) j-*oo = lim (*. X), j-*oo which implies * = grad P,4>, which implies QJ,4> -*¦ P,<f> in Sj. Therefore ||grad/>,0||2 = lim^ ||grad QJ,4>f = -(/>,*, AP,4>), which conludes the proof of the theorem. ¦ Theorem VII.4.4 Let M be an arbitrary Riemannian manifold, yc the Friedrichs extension of Ac, and p the minimal positive heat kernel, with asso- associated semigroup P, and infinitesmal generator A. Then Proof Let T, be the 1-parameter semigroup generated by fo, tf> e Cf(M), and set Then v is a solution of the heat equation with vanishing initial data. We want to show that v vanishes on all of M x @, oo). The spectral theorem, Lebesgue's dominated convergence theorem, and the estimate < 2, 0 < x « 1, combine to imply that we may differentiate / v2(x, t) dV(x), under the integral sign, with respect to /. Then (see the proof of Theorem VII.3.4) = f Av(x,t)v(x,t)dV(x). By Theorem VII.4.3, for each / > 0, P,<f> is in the domain of &, and therefore VII.4 The Action of the Heat Semigroup 221 v is in the domain of 3>, which implies 2diJ v2(x<*)dV(x)= I Av(x,t)v(x,t)dV(x) = - j |gradt;|2(jt.O<*V(*)<0, which implies the claim. ¦ Notation We henceforth write, when there is no confusion, A for A = $r. Thus, if {Ex.: A. > 0) is a spectral family for A, then f°° A = -/ XdEk, Jo with associated quadratic form D[f, f] = (-A/, /) = r\ (dExf, f) Jo possessing domain D((-AI/2) 2 2>(A); and /•OO P, = e~udEx. Jo Theorem VII.4.5 Let pbea closed or Dirichlet heat kernel on a Riemannian manifold M. Then —A has discrete spectrum, and p(x, y, t) has the absolutely, uniformly convergent Sturm—Liouville expansion (VII.4.3) p(x, y,t) = oo ^2 7=1 where {0i, fa,...) is a complete orthonormal basis of L2(M) consisting of eigenfunctions of—A, with the eigenvalue oftpj equal to X,. Also, <pj e C°° for all j. Proof Indeed, under these hypotheses, for every fixed /, p(x, y, t) is a Hilbert- Schmidt kernel (that is, symmetric and square-integrable on M x M), which implies the discreteness of the spectrum and the expansion (VII.4.3) in L2(M x M). The stronger statement of convergence of (VII.4.3) follows from Mercer's theorem (Riesz and Nagy, 1955, p. 245). ¦ Remark VII.4.2 Let p be the heat kernel of a compact Riemannian manifold M. The constant functions clearly are annihilated by A, so \\ — 0. On the other hand, if A<f> = 0, then 0 = (—A <j>, <p) = ||grad <f>\\2, which implies <p = const.
222 Laplace and Heat Operators This implies X2 > O.ln particular, 1 which implies (using L2-L°° estimates of eigenfunctions, as in Li A980) for example) So the heat is evenly distributed throughout a compact manifold as / f +oo. Notation For any Riemannian manifold M we let = infspec - A. Proposition VII.4.1 Assume X(M) has an L2 eigenfimction. Then the eigen- function never vanishes, and the eigenspace of\(M) is I-dimensional. Proposition VH.4.2 Given two domains ft) c JJ2 '" a Riemannian manifold M, let —A] and — A2 denote their respective Laplacians; then A(S2i) > \(Sl2)- When fli CC Sl2 then the inequality is strict. In particular, A.(?2) > \(M)for any relatively compact ?2 in M. VIL5 Simplest Examples Example VIIJ.l If M = R", then ds2 = = dxi-dxn, as discussed in §VII.3.1. Example VIL5.2 Let M = H" = M_!, the n -dimensional hyperbolic space of constant sectional curvature —1, and x e M. Then any point y e M is de- determined by its distance r from x, together with the unit tangent vector ? at x of the unique geodesic from x to y. So (r; ?) determine geodesic spherical coordinates on all of M, relative to which where |</?|2 denotes the Riemannian metric on Sz (the unit tangent sphere to VII.5 Simplest Examples 223 M at x) with its standard metric. The volume element is given by dV = sinki^ where d(ix(%) denotes the (n - l)-volume element of Sr. The Laplacian i given by is ;|) = i| + („ - i)coth r ^ to where Af denotes the Laplacian on Sx, and fr denotes the restriction of / to the metric sphere in M centered at x with radius r, viewed as a function on S,. The simplest explicit formula for the heat kernel is when n = 3, namely, on H3 the heat kernel is given by p(x y ,)=_L p(x'y'l) D 2 Sinh d(x, y) Example VIU3 If M = M\ x M2 is the Riemannian product of M\ and Mi, then for any chart x = (x(, X2) on M, where xj is a chart on Mj,j = \, 2, have for the matrix G* of the product Riemannian metric we C* = where Gj denotes the matrix of the Riemannian metric on Mj. For F: M -* R given by M1 x M2, we have (with obvious notation). For the heat kernel we have p((x]tx2), (y\,y2), t) = p\(xuy\,t)P2(x2, y2, t). Example VIL5.4 If p is a closed or Dirichlet heat kernel on M, then by Theorem VII.4.5, p(x, y, t) has the Sturm-Liouville eigenvalue-eigenfunction expansion (VII.5.1) p(x, y, t) = 2 e-x''*j(x»j(y). where {Xi < k2 < • ¦ ¦ | +00} denotes the spectrum of—Aon M (it is discrete), with eigenvalues repeated according to their multiplicity, and i<f>j :j = 1,...}
224 Laplace and Heat Operators is a complete orthonormal basis of L2 such that <t>j is an eigenfunction of Xj for each j. Example VIL5.5 The 2-dimensional jungle gym JG2 as described in Example V.2.4. We shall see that for sufficiently large time one has P(x,y,t) < VII.6 Bibliographic Notes General references for the material in this chapter are Davies A980, 1989, 1995) and Chavel A984,1994). [Formula C8) of Chavel A984, p. 150) is only valid when k < 0.] See also the articles of Chavel, Davies, and Grigor'yan in the 1998 Edinburgh Lectures (Davies and Safaro. pp. 30-94,14-225). §VII.l Additional background for this section may be found in Yosida A978) and Reed and Simon A975). §VIL2 The Meyers-Serrin theorem is proved in Adams A975), Aubin A982), and Gilbarg and Tnidinger A977). Theorem VII.2.1 was first proved in Gaffney A954); our proof follows Karp A984). Corollary VII.2.2 is from Yau A976). Elliptic regularity is treated in Gilbarg and Tnidinger A977, p. 176). §VIIJ One can find a proof of the strong maximum principle in Plotter and Weinberger A984, §111.3.4) for the case where M is diffeomeorphic to a domain in Euclidean space. A standard continuation argument then extends the theorem to arbi- arbitrary Riemannian manifolds. The proof of Theorem VH.3.3 corrects an incorrect proof given in Chavel A984, p. 139). The uniqueness property and stochastic completeness for Ricci curvature bounded from below (Proposition VII.3.6) was first proved in Dodziuk A983), and the Feller prop- property in Yau A978). The last two results are valid in much greater generality; for stochastic completeness, see Azencott A974), Davies A992), Gaffney A959), Grigor'yan A987), Hasminkii A960), Ichihara A982, 1984, 1986), Karp and Li (unpublished), Oshima A989), Pang A988,1996), Takeda A989,1991), Varopoulos A983). An extensive sur- survey of these and other results can be found in Grigor'yan A999). For the Feller property, see Azencott A974), Davies A992), Karp and Li (unpublished), Pang A988, 1996), Gaffney A959). Hasminkii A960). §VIL4 Theorem VII.4.4 is proved in Dodziuk A983). Proposition VII.4.1 is from Sullivan A987). §VIL5 Explicit calculations for the heat kernel on a hyperbolic space of any dimen- dimension can be found in Davies and Mandouvalos A987) and Grigor'yan and Noguchi A998). For the heat kernel on Lie groups, see the articles of Arede A985) and Fegan A976, 1983). For the heat kernel on homogeneous spaces, see Benabdallah A973). For the special case of spheres, see Fischer, Jungster, and Williams A984) and Ndumu A986). vra Large Time Heat Diffusion In this chapter we present the main result of this half of the book - Theorem Vni.5.4 - that, in a complete Riemannian manifold of bounded geometry, a positive modified isoperimetric constant 3V,P implies an upper bound for the minimal positive heat kernel p of the type P(x, y, t) < const./-y/2 V / > To > 0, or, equivalently, (Vm.0.1) p 't+oo for all x, y e M. So the result states that if M is "v-dimensional" in the sense of isoperimetric inequalities, then it is at least "u-dimensional" in the sense of heat diffusion. The method of proof is to pass from the geometric inequality 3vp > 0 to a modified Nirenberg-Sobolev inequality, which is then shown to be equivalent to the large time upper bound (VIII.0.1). The transition from 3V<P > 0 to the modified Nirenberg-Sobolev inequality is achieved by shifting the problem to any discretization of the manifold (Theorem VIII.5.2). A by-product of this argument is the invariance of (VIII.0.1) under compact perturbations of M, of the type discussed in Theorem V.3.1. The final section of the chapter closes the book with an alternate proof of Theorem VUI.5.4 (see Example VI.3.1 and Example VIII.6.1 below) using Grigor'yan's replacement of the Nirenberg-Sobolev inequalities with Faber- Krahn isoperimetric inequalities as the primary tool for establishing upper bounds on the heat kernel. The full possibilities of this method seem to go well beyond the result presented here; so the best may be yet to come. We assume, throughout this chapter, that M is a Riemannian manifold and p its minimal positive heat kernel, with attendant semigroup P, acting on L2(M), which is contractive, continuous, and self-adjoint. The infinitesimal generator 225
226 Large Time Heat Diffusion of the semigroup is the Friedrichs extension of Ac, the Laplacian acting on C™(M) considered as a dense subspace of L2(M). The semigroup P, also has an action on Lk, k > 1, which is contractive and continuous. Since Lk O L2 is dense in Lk, we shall be able to study the norms of P,: Lk n L2 -> Ll n L2 as though they were mappings from Lk to Le, with their associated norms UP, «*,«. VIII.l The Main Problem Proposition VIII.1.1 Assume x e M such that the exponential map exp is defined on the closed disk B(jc; R) in Mx. Then for d(x, w) < R/2, 0 < r < d(x, w) we have (VIII.1.1) p(x,w,s) < const.-^ s-n/2 + sr-(*+2) ,-{<Hx,w)-2rJ/4s where 3n(B(z; p)) denotes the n-dimensional isoperimetric constant ofB(z\ p). Recall that, for any Riemannian manifold M, we set ?(x, y, t) = Theorem VIII.1.1 Assume M is a complete Riemannian manifold. Then for every compact K CC M there exist positive constants {depending on K) such that (VIII. 1.2) limsup P(x, y, t) ?{x, y, t) -1 < const.d(jt, y) is valid for all x, y e K. Proof We already know the result for M compact, so we assume that M is complete and noncompact. Given x and y in M, let Q be any domain in M, with C°° boundary and com- compact closure, containing x, y, and let q denote the Dirichlet heat kernel of ?2. Then p(x, y,t) > q(x, y, t) by the maximum principle, which implies by VIII.l The Main Problem 227 Proposition VII.3.5 that To prove li™oUPf^-1+C°nSt^>) we argue as follows: Let Si and q be as above. Then Duhamel's principle implies P(x,y,t)-q(x,y,t) da = -/ ds I p(x,w,s)-^-(w,y,t-s)dA(w) Jo Jan Bvw ' <| sup p(x,w,s)\ f ds f -~{w,y,t-s)dA(w < sup p(x,w,s). But one has, for any r e @, d(x, w)/2), the upper bound (VIII.1.1). Therefore pick Si so that d(x, 3fi) is very large compared to d(x, y), fix r very small, and let t i 0. Then one has, for some a e @,1), .. P(x, y, t) q(x,y,t) limsup -—-— < limsup msup -—-— no ?(x,y,t) ?(x,y,t) + lim sup '10 ?(x, y, t) < 1 + const.d(jc, y), which is the claim. Remark VIII.1.1 One can see (VIII. 1.2) explicitly in both the example of H3 (Example VII.5.2) and the product metrics (Example VII.5.3), in the sense that if one has this asymptotic behavior on each of the factors, then one also has it on the product. Remark VIII.1.2 The theorem remains valid even if M is not complete, but one must formulate it differently. Namely, given any SI cc M, x e Si, the inequality (VIII.1.2) will be valid for y sufficiently close to x.
228 Large Time Heat Diffusion For large time considerations the behavior of the heat diffusion takes into ac- account the large scale structure of the manifold, and then anything can happen. Hyperbolic space displays an example of when the heat kernel decays expo- exponentially with respect to time, and Euclidean space when the rate of decay with respect to time is given by the dimension of the manifold. For M compact, one has p(x, y, t) -> 1/ V(M) as t f +00. Therefore, for M = S""* x R*. the heat kernel decays at the rate t~k/2, slower than that indicated by the dimension n of M. Finally, we have examples of polynomial rate of decay that are faster than that indicated by n, the dimension of M, namely, the jungle gym of Example VH.5.5. Main Problem Study the large time decay of the heat kernel, as a consequence of geometric hypotheses on M. Except for Theorem VIII. 1.4 (which is included for motivation), we shall restrict ourselves to the study of upper bounds. In general, upper bounds on the heat kernel are studied in two stages: The first is referred to as on-diagonal upper bounds, namely, upper bounds for p(x,x,t). By the semigroup property, the Cauchy-Schwarz inequality, and the symmetry property we have P(x,y,t)= [ P(x,z,t/2)p(z,y,t/2)dV(z) Jm Ul 1/2 I r 1 1/2 p\x,z,t/2)dV(z)\ \J P\z,y,tl2)dV{z)\ that is, (Vm.1.3) P(x, y, I) < y/p(x,x,t)y/p(y,y,t). Therefore, as soon as we have any upper bound valid for p{x, x, t) for all x, we automatically have an upper bound on p(x, y,t). But this upper bound could not contain a Gaussian term of the form Then the second stage is to produce new arguments to sharpen the upper bounds based solely on the on-diagonal upper bounds. The introduction of the Gaussian correction to the upper bounds is also referred to as the off-diagonal correction. We shall not pursue this second stage here. It has recently been definitively established that the correction in upper bounds does not depend on the specific VIII. 1 The Main Problem 229 ^T^l*!.?™™* "r. * **""*** of the heat equation that kernel automatically imply off-diagonal VHI.1.1 General Considerations Theorem Vm.l2 For on-diagonal upper bounds have (VUL1-4) 8UPJ>(*.*.0=llfi|||-oo. Proof First, given / e L\ we have \P,f\(x) - \J p(x, y, t)f{y)dV(y) which implies tPtf\{x) < sup p(x,y, x.y x.y , jc, 1). Therefore, || P, ||,_«, < sup, p(x, x, t). For the opposite inequality recall that an approximate identity at x e M is a family of nonnegative functions 6h h>0, satisfying Vht L 4>hdv = 1 0, Hm f <t>hdV =0 for all c > 0. Then 4>h -> Sxt the delta function concentrated at x, as h 1 0 For such an approximate identity 0A at x we have I P(x, y, t)<ph(y)dV(y) = pl<t>h{x) < IIP,IIl^oc||0*III = II fill l-oo. *• ™ Aat MM) demotes ihe inlimum of the specttum of the Lapteia, -a
230 Large Time Heat Diffusion Theorem VIII.1J For all x e M we have that t \-+ ek(M)l p(x,x,t) is a decreasing function oft. Proof Let D be a relatively compact domain in M with C°° boundary and Dirichlet heat kernel q. Then the Sturm-Liouville eigenvalue-eigenfunction expansion of q, and X(D) > X(M) combine to imply e^U)lq(x, x, t) is a de- decreasing function of t. Now pick an exhaustion of M, Dtr f M as j f +00, by domains that are relatively compact in M and that possess C°° boundary. Let qj denote the Dirichlet heat kernel of Dr Since qj \ p, the lemma follows immediately. Corollary VHI.1.1 The norm || P, || i-^x, decreases with respect to time. Proposition VIII.1.2 Let X = X(M). For all x, y in M we have the existence of the limit (VIH.1.5) for which we have the following alternative: Either T vanishes identically on all of M x M, in which case k possesses no L2 eigenfunctions; or T is strictly positive on all ofMxM, in which case X possesses a positive normalized L2 eigenfunction 4> (normalized in the sense that its L2 norm is equal to I) for which (VIII. 1.6) lim /t+oo locally uniformly on all ofMxM. The simplest example of the case T = 0 is R", n > 1, discussed above — just note that A. = 0, so e* p(x, y, t) = p(x, y, t) -*¦ 0 as t \ +oo. When M is noncompact with compact closure and smooth boundary, then one always has T strictly positive, and (Vm.1.6) follows from the Sturm-Liouville expansion (VII.5.1)ofp. We note some easy consequences of the proposition. Corollary VIII.1.2 We always have In p(x,y,t) (VIII. 1.7) lim VIII. I The Main Problem 231 locally uniformly on M x M; and when M has finite volume V, we have (VIII. 1.8) locally uniformly on M x M. Proof We wish to show lim If the limit function T is positive, then the result is obvious. So we are only concerned with the situation where T is identically equal to 0. For any domain Din M, let q& denote the Dirichlet heat kernel of D, and XD the lowest Dirichlet eigenvalue of D. Then, of course, we have eup(x, y, t) We let t t +oo. Then A - Xd < limmf /t+oo f Now let D t M. We conclude that , Ine"p(x,y, t) 0 < liminf HK } . t+ / But since T — 0, we have In e*'p(x, y,t) < 0 for large r, which implies (x,y,t) lim sup /t+oo < 0, which implies (VIII.1.7). When M has finite volume then X = 0 with normalized L2-eigenfunction = 1 I<JV(M} (for all x), which implies (VIII. 1.8). ¦ Corollary Vm.1.3 For any M we have (Vni.1.9) lim p(x, y, t) = 0 /t+oo if and only ifM has infinite volume. Proof If M has finite volume, then (VIII. 1.8) implies that lim p, as t f +°o. is nonzero. If, on the other hand, M has infinite volume, then (a) for X > 0 simply use (VIII. 1.5); and (b) for X = 0, if T were positive, we would have the
232 Large Time Heat Diffusion existence of an L2 harmonic function on M, which is impossible by Corollary vn.2.2. ¦ Corollary Vffl.1.4 Suppose M noncompact is a covering of a compact Riemannian manifold. Then T is identically equal to zero. Consequently, if the covering is nonamenable - by Brooks A981), X > 0 - then p tends to 0 faster than e~kl. Proof If A. = 0, one uses the above corollary, because M has infinite volume. If A. > 0 and T > 0, then, as mentioned above, the L2 eigenspace of X, which is nontrivial, is 1-dimensional (by Proposition VII.4.1). But this is impossible, by the invariance of the eigenspace under the action of the deck transformation group. ¦ Remark VIU.l 3 A fundamental distinction emerges between the cases X = 0 and the cases X > 0. If one subjects the complete Riemannian manifold M to a compact perturbation, then in the case of bounded geometry (that is, Ricci curvature bounded from below and positive injectivity radius) X = 0 remains invariant under the perturbation. Moreover (Theorem VIII.5.2 below) the rate of polynomial decay of the heat kernel also remains invariant under the per- perturbation. However, in the case X > 0 the value of X does not necessarily remain invariant under the perturbation. Simply consider H", ^-dimensional hyperbolic space of constant sectional curvature -1, with X = (n — 1 J/4, and subject it to a compact perturbation in which it has a flat disk with lowest Dirichlet eigenvalue strictly less than (n - lJ/4. Then the new Riemannian metric has A. < (n - lJ/4 with appropriately slower heat kernel decay for large time. Our interest in what follows will usually involve polynomial decay with respect to time. So we will be interested in the case X = 0. VIII.1.2 Volume Growth Considerations The most elementary geometric intuition is that, because the total amount of heat in the space cannot increase with respect to time, the larger the space, the quicker its heat kernel decays. Indeed, if M is compact, then \/V(M) is the limit if the heat kernel as t \ +oo. In general, the intuition is valid, if not sufficiently precise. Most, if not all, of our geometric results will aim toward a deeper understanding of the effect of volume growth of a Riemmanian manifold on heat kernel decay. VIII. 1 The Main Problem 233 Theorem Vm.1.4 Giveno eMandfi> 0, for which one has the heat kernel lower bound (Vm.1.10) p(o, x, t) > con&l.rvl2e-dHo'x)'fi' forallx eMandt>T>0, then one has the volume growth upper bound (Vm.1.11) V(o;r)<comt.r" for sufficiently large r > 0. Proof The heat kernel lower bound implies p(o,x, t)dV(x) > which implies the volume upper bound. y(o;>/?), Remark VIIL1.4 There are results for a partial converse (Coulhon and Gngor'yan, 1997): If (Vm.1.11) is valid for a given o € M and sufficiently large r, then (VIII.1.12) p(o, o, t) > const.(f In t)~v/2 for large t. Examples exist to show the estimate is sharp. Also, one can improve (VIII. 1.12) to P(x,x,t) > const.r~v/2 assuming one is also given the upper bound p(x,x,t) < const./~v/2 for large /. [See earlier versions in Benjamini, Chavel, and Feldman A996)]. Exampk Vffl.1.1 Does a volume growth lower bound imply a heat kernel upper bound? No. Endow M = R2 with the Riemannian metric ds2 = 4>{y){dx2 + dy2), where 0 < 4>(-y) = Hy) e C°°, = y-2 for Then one easily sees that M has exponential volume growth. But general con- considerations show, since the metric is conformal to the Euclidean plane (this is
234 Large Time Heat Diffusion just for the 2-dimensional case), that f< p(x,x,t)dt = +00, which precludes a rate of decay of the form / v/2, v > 2. We shall need the following heat kernel upper bound in the sequel. Proposition VIII. U Let M be complete with Ricci curvature bounded from below by the constant -K, K > 0. Then for any a > 1, one has the upper bound (VIII. 1.13) where the constant C(e) -> +oo as e -> 0. Or Remark VHL1.5 An application of Holder's inequality implies that for any k, I and respective conjugates k', I' we have II fill*-* = llfilr-*'. Because P, is self-adjoint for all t, we have We shall use this repeatedly throughout our arguments. VfflJ The Nash Approach Theorem VDL2.1 Assume there exists v > 0 such that the NashSobolev inequality (VIH.2.1) llgrad/lb > const.||/||2'+2/v||/||r2/v is valid for all f e Cf. Then (VIII.2.2) ||/»f||^00<const./-l'/2, for all t >0. Proof Set«(/) = || fi/||22. We then obtain l-u\t) = (p,f, ytP,f\ = (P,f, AP,f) = -Hgrad < -const.«(/I+2/" u'(t) VI11.2 The Nash Approach 235 by Theorem VII.4.3. Then ^«'(/) = -||gradP,/||22 that is, which one integrates to obtain «(/) < const.||/||i2/~l>/2. So ||fi||i_»2 < const./-v/4, which implies, by duality (Vm. 1.14), \\P,h-*oo < const./-"/4, which implies (Vin.2.2). ¦ Remark VIII21 One can pass from «(/) < const-ll/lli2/-11/2 to (VIII.2.2) without using duality (m.2.5). Namely, for any / € Cf we have that is, (w. y. t)fiy)dV{y)\ dV(w) < const.Il/H,2/-"/2, constll/H,2/-"/2 > I \j p(w,y,t)f(y)dV(y)\ dV(w) = j \j p(yv,y,t)f(y)dV(y)\ x j / p(w,z,t)f(z)dV(z)\ dV(w). Given any x e M, let 4>h be an approximation of the identity concentrated at x, and let f = 4>h- Then, letting h I 0, we obtain const./-"/2 > / p(w, x, which implies the theorem. t)p(w,x, t)dV(w) = p(x,x,2t), Theorem VIII.2.2 Conversely, assume (VIII.2.2) for all t > O.Then(VIll.2.l) is valid on C*. Proof Let 0 € C*. Then the spectral theorem impl ies -/Hgrad0||22,
Large Time Heal Diffusion 236 which implies < r||grad0||22 + const.rl>/2||0||i2 for all / > 0. Now minimize the right hand side with respect to t. The minimum is achieved at t0 = const.||0|li2/||grad0||22, which implies the Nash-Sobolev inequality (VIH.2.1). ¦ Remark VTIL2.2 Here is a proof without using the spectral theorem. One simply has d and A ? — ||grad/>,0||22 = --(P,<t>, A/>,0) = -2\\AP,if>\\2 < 0, at at which implies t >-*¦ ||grad Pt<Ph2 is nonincreasing, and 1Mb2 = \\P,<t>h2+ [ ||grad/>,0||2</r. Jo The Riesz-Thorin interpolation theorem (see Proposition Vm.3.1 immediately below) implies, from (Vm.2.2), || P, || 1-2 < const./""/4 for all t > 0. Therefore and we minimize as above. VIU J The Varopoulos Approach We first state some background results. Proposition VIII J.I (Riesz-Thorin Interpolation Theorem) LetT.L2 L2 satisfy where po, Qo, Pi,Qi € [0, 00]. Fork € @, 1), set 1 _ 1 - X X 1 _ 1 -X ± Pa Po Pi' Qk <?o Qi V11I.3 The Varopoulos Approach 237 Then Proposition VIII J.2 (Maximal Theorem) Let Tt: L2 ^ L2 be a self-adjoint contraction semigroup. To every function f e L2 associate the function f*(x) = sup \T,f\(x). Then Wf'Hp < const.,11/11, forallfeLPnL2,pe [1, +00]. Definition Let T, be a self-adjoint contraction semigroup with generator - A (so A is nonnegative). We define (where {Ek} is a spectral family for A) with domain consisting of those 4> e L2 for which / Jo The definition is motivated by the fact that / e~*V-' dt = Jo for any given (i e @, 1). Theorem VIII J.1 Let M be a Riemannian manifold, A = -A. Then if and only ifM has infinite volume, if and only ifp(x, y, t) -* 0 as t f +oo. Proof We have i f°° r°° "(M) Jo Jo [°° -l Jo ^ € dt ro = / Jo* dEk = I -Eo,
238 Large Time Heat Diffusion where ?o is the eigenspace of A. = 0, either the line of constant functions (Propo- (Proposition V1I.4.1) or the origin in L2. If M has infinite volume, then L2 has no nontrivial constant function. If M has finite volume, then the constant functions are all in L2. ¦ Theorem VIII J.2 Let v > 2. Then (VIII.2.2) is valid for allt > 0 if and only if (Vm.3.1) ll/|l2v/(.-2) < const.||grad/||2 for all f 6 Cf, equivalently, if and only if (Vm.3.2) ll(-A)-1/2/||2l)/(l)_2) < const.||/||2 forallfeV((-A)-1/2). Proof Certainly, given (VITJ.3.1) on C* we have (VIII.2.1) by the proof Lemma VI. 1.2, which implies (VIII.2.2) for all t > 0. For the converse, set A = -A, and assume (VIII.2.2) on L2. By Theorem VIII.3.1 we have the representation A-u2f=_l n tw-iPifdt Vn Jo for all / e 2?((-A)-|/2), which implies, for any T > 0, \A~l/2f\(x) < (sup |. l/>0 1 f j r^WP.h^d Now use the Riesz-Thorin interpolation theorem, with Po = 1, qo = oo. P\ — q\ - oo, A. = 1/2. Then Pi. =2, qx = oo, which implies llfllh-oo < Mn'^Afi* < const.r-"/4, which implies V7//.5 77j« Varopoulos Approach 27-1/2 y/Jt 239 Minimize the right hand side with respect to T. Then the minimum is realized at which implies which implies \A'll2f\{x) < cv\nl-2'v{x)\\fh2l / M We therefore have (VIII.3.2). ¦ Remark Vm J.I We could have easily used Theorems VI. 1.1 and VIII.2.2 to prove the above theorem. But the method is useful for other considerations, as well. For example, the proof of the theorem also implies a short time result: Theorem VfflJJ Let v > 2. Then (VIII.2.2) is valid for all t € @, 1] if and only if (Vm.3.3) ll/ll2v/(v-2) < const.{||grad/||2 for all f e Cf, equivalently, if and only if (VIII.3.4) ||(-A)-I/2/||2v/A)_2) < const.{||/||2 + forallfeV((-A)-l<2). Proof Assume || P, || i^.^ < ct~v/2 for all t e @,1]. Consider the semigroup Then, by Theorem VIII. 1.1, ||/»,||i_»oo decreases with respect to time, which implies T, satisfies || T, || i-oo < ct~v/2 for all t > 0. Now the infinitesimal gen- generator of T, is A - /, which implies (VIII.3.3) by the previous theorem. The converse is similar. ¦
240 Large Time Heat Diffusion A direct consequence of the Li-Yau upper bounds (Proposition VIII. 1.3) and Crake's inequality (V.2.15) is Proposition VIIIJJ lfM is n-dimensional Riemannian complete with bound- bounded geometry, then < const./-n/2 for all t €@, 1]. In particular, when n>2,we have (VIU.3.5) forallf^Cf. -2) < const.{||grad/||2 VDL4 Coulhon's Modified Sobolev Inequality What if we are only given (Vm.2.2) for all t > 1? Theorem VIU.4.1 If (VIII.2.2) is valid for all t > I, then for each f e 2?((-A)-|/2) we have, instead of (Vm.3.2), where g e L2 and h € L2v^v~2\ with High < COnSt.il/H2, ||*||2v/(v-2) < COnSt.il/H2. Proof First, set A = — A, and write Certainly, = r Jo which implies (using the proof of Theorem VIII.3.2), \\PiA-1/2f\\2vnv-2)< COnSt.il/H2; so we pick h = PiA~l/2 and g = (/ - Pi)A~l/2. Now VIU.4 Coulhon 's Modified Sobolev Inequality 241 and P,A*'2= f This implies, for / e T>(P,Al/2), \\P,Al/2ff < s-1 j (ks)e-2^ (dEkf, /) < that is, which implies What about a converse? We first have < const. Theorem VHI.4.2 (Extrapolation Theorem) Given 1 < a < 0 < +00 f which ll^/lla-/} < const./ "* for all t > 1. Then ll/'/lli-oo < const./"', where&=e{a-1-0-l}-i,forallt> 1. Proof Fix 6 such that or Of so 0, 1 - d e @, 1). First note that, for / > 2, WPtfh = IIW//2/IU < const./"'IIP;/2/||a < const./"*||P,/2/||i < const./"'\\f\U6\\Pl/2f\\0l-e < const./"'||P,/2l|iV~9ll/lli - the third line is Holder's inequality, applied to , P=l/ad,
242 Large Tone Heat Diffusion - which implies II/>, II 1^0 < const./-«||/>;/21|, V"*, which, in turn, yields te/e IIP, IIi-*, < const. {(t/2Y"> ||P,l21|,_,}'-" . So we wish to solve the inequality a(t) < const{o(t/2))l-e, t > 2. Well, we have for all t > 1. Therefore, if we vary r e [1,2] and A = 0, 1,2,.... we obtain o(t) < const., / > 1, which implies P, || !_»A < const. for all/ > 1. Next, duality then implies const./ </9 -</e for all / > 1, where 0' = 0/@ - 1) denotes the conjugate of 0. Now duality also implies \\P,\\p'^a-< const./"' for all / > 1. Then for 0 given by 0' a' ' we have (using Holder's inequality applied to \P,f\ = \P,f\e\P,f\l~e) (VIII.4.3) for all / > 2. So for all / > 2. Set e' = eA - 0). Then the argument for (VIII.4.2) also implies which implies which implies the theorem. const./-<'/e+<'/e), VII1.5 The Denouement: Geometric Applications 243 Theorem VIII.4.3 Let M have infinite volume. Suppose we are given (VUI.4.4) (-A)-|/2: L2 -* L2 + LlvKv~2\ is bounded inthe sense ofTheoremVlUAX and Pi: Ll -> L00 is also bounded. Then (VIII.2.2) is valid for all / > 1. Proof We first have by the Riesz-Thorin interpolation theorem [with a = 2, and 0 = 2v/(v - 2)] that Pi: L2 ->• L2v/lv~2) is bounded, which implies Pi(-A)-l/2: L2 -+ L2v/iv~2) is bounded, which implies < const./~l/2 by (VUI.4.1), which implies, by the extrapolation theorem [with a = 2 0 = 2v/(v-2),e = l/2)], II P,+i II i^oo < const./""/2 for all / > 0. _ VIII.5 The Denouement: Geometric Applications Let G be a graph with vertices Q. Recall that for every ? € <7, we denote its collection of neighbors by N(?), and its valence by m(?) = card N(^). The collection of edges of G is denoted by ?,, with oriented edges denoted by [^, n] where n € N(?). The measure of functions on ? is given by dV(lj) = m(%)di(%\ and of functions on Qe by </A([?, n]) = di([tf, n]), where di always denotes counting measure. Lemma VIIL5.1 For any function f ong we have ' K.nUG, Proof We have /«)- r m-\ E m+na. which is the claim.
244 Large Time Heat Diffusion Definition We define the Laplacian of the function f:Q-+R, denoted by AG/.by So Ac / is a function on Q. Then for functions / and h we have - f /AG h d\ = - K.I)] {/(I?)- = 1 / (Vf 2JQ. iDh)dA, that is, - f JQ [ Q. (X)f X>h)dk\ and = 2 / / We conclude Theorem VIIL5.1 Ac »'* self-adjoint nonpositive bounded operator on L2(Q, dV), with heat semigroup P, = exp t AG, wi/A associated heat kernel p(?, ^, r). Remark VIII.5.1 Note that we are using two definitions of |D/|2. In §VI.5 we defined |S/|2 as a function on Q, the vertices of G, namely, forthevertex^ e ^.HerelD/^isviewedasafunctionontheorientededges^,. of G. But as noted at the very beginning of §VI.5, the definitions are equivalent. VIII. 5 The Denouement: Geometric Applications 245 So for all arguments involving Sobolev inequalities, we may work with either definition. Example VIII.5.1 Consider the integer lattice Z* in R*. The graph structure on Z* is given by: n € N(?) if and only if \n - ?| = 1 in R*. Then the heat kernel p on the associated graph is given by where Iv(z) = e~'v*/2Jv(iz) denotes the modified Bessel function. Proof We only consider the case k = 1. The higher dimensional case is similar. Let / = /(n), n € Z. The Laplacian here is given by For u = u(n, t), where n ranges over the integers and t ranges over positive time, consider the heat equation 3« = —, t > 0, at (Vin.5.2) with initial values given by where <p(n) is a given function on the integers. Consider the function on the circle, having time t as parameter, then (VIII.5.2) becomes which implies u(n,t)eine; W —@,0 = {cos 8- l}f/@,O. dt U(8. t) = e-"- One now uses the integral representation 4>(n)e ine Jn(z) = _L [
Large Time Heat Diffusion 246 to show for the solution to the initial value problem, which implies the claim. Note that 1,@ = B^rr1/V{l - 0,A"')}, t t +00, which implies, on Z*, IIP/ IIi^x. ~ const./-*/2, t t +oo. Remark VHI.5.2 The Varopoulos and Coulhon criteria apply as well to graphs with heat diffusion with continuous time, and with "heat diffusion" with discrete time, namely, random walks on graphs. Theorem VHI.5.2 Let M be Riemannian complete with bounded geometry, and let v > 1. Then (VIII.4.4) is satisfied on M if and only if the discrete Nirenberg-Sobolev inequality (VIII.3.1) is valid on every discretization ofM. In particular, the heat semigroup P, ofM satisfies (VIII.5.3) IIP, II 1^00 = if and only if (VUI.5.4) IIP, II.^oc = for the heat semigroup P, of any discretization G ofM. Also, the estimate (Vm.5.3) remains invariant under compact perturbations ofM. Proof First, because M has bounded geometry, one knows, by Proposition VIII.3.3, that Pi: L1 -> L°° is bounded. Next, assume that v > 2. Let G be a discretization of M, and assume the Nirenberg-Sobolev inequality (VIII.3.1) is valid on G. Consider the associated discretization operator V and smoothing operator S on functions defined on M and G, respectively. Given any smooth function F: M -*¦ R, consider * = (- Am)~1/2F, and write = {<t>- SV<t>) + SV<t>. Then (see §VI.5.5) -SV<t>||2 <const.||grad<t>||2 = const.||F||2, V1I1.5 The Denouement: Geometric Applications 247 and(byP:iiiof§VI.5.1) ll<5P*||2i,/(v_2) < COnSt.||Z><t>||2v/(v-2) < const.||X>T><t>\\2 < const.||grad*||2 = const. || F || 2. Therefore, if one starts 'with the discrete Nirenberg-Sobolev inequality (VIII.3.1) on G, one also has the inequality (Vin.4.4) on any discretization ofM. Conversely, assume (VHI.4.4) on M; and note that SV: Ll (M) -> L°°(M) is bounded, which implies by Proposition VIII.3.1 (the Riesz-Thorin interpolation theorem), that ST>: L\M) -> L2v is bounded, which implies [with (VIII.4.4)] SV(-Am)-1'2 : L\M) - is bounded. Therefore, given <f>: Q -*¦ [0, +oo) then [by E:ii) of §VI.5.2 and (P5:ii)of§VI.5.3] v-2) < const.\\SVS4>h*K»-2) = const, llgrad S4>\h < const. || which is the discrete Nirenberg-Sobolev inequality on G for nonnegative func- functions. For arbitrary functions, one has II XWI Ib < and now can easily obtain the result for general <p. By Theorem VIII.3.2 (applied to graphs) and Theorems VIII.4.1 and VTII.4.3, we obtain the equivalence of the large time upper bounds on the heat semigroups of M and G when v > 2. Now assume 1 < v < 2, amd let Mi denote the Cartesian product of M with R2. Then the upper bounds for the heat kernel and the Coulhon criteria are to be considered for v, = v + 2, for which the theorem is valid. For (x, y) € M x R2 := M\ (with obvious notations) we have p(x, x, t) = const.r2/2pi((jc, y), (x, y), t)
248 Large Time Heat Diffusion and for (?, n) e G x Z2 we also have p(?, ?, r) ~ consU2/2Pl((?, 17), (?, 17),») as t1 +00. So P{x,x,t)<rv'2 One can now use the above argument for vj to obtain the result for v. Finally, the estimate (VIII.5.3) remains valid under compact perturbation of a manifold of bounded geometry since any discretization of M remains a discretization under compact perturbation. ¦ Theorem VIIL53 Let M satisfy 3V(M) > 0 for some v > 1. Then (VIII.2.2) is valid for all t > 0. Proof Let v > 2. By Theorems VI. 1.1 and VI. 1.2, the hypothesis implies the Nash-Sobolev inequality (VIII.2.1), which implies, by Theorem vm.2.1, the upper bound (VIII.2.2) for all t > 0. For v 6A,2), one argues as in the similar case of the above theorem. ¦ Corollary VIIL5.1 For the 2-dimensional jungle gym in R3 we have WPth^co <const.r3/2 Vr>0. For the 3-dimensional Heisenberg group with left-invariant Riemannian metric, we have I|/J,lli-O0<const.r4/2 Vr>0. Proof Indeed, ZI3 is positive for the 2-dimensional jungle gym in R3, since one has a simultaneous discretization of R3 and the jungle gym. Also, 34 is positive for the 3-dimensional Heisenberg group (see Pansu A982)). ¦ Theorem VIII.5.4 If M is Riemannian complete with bounded geometry, and 3ViP(M) > Ofor some v > 1 and p > 0, then we have the upper bound (VIII.5.3). In particular, for every complete Riemannian manifold with bounded geom- geometry we have (VIII.5.5) ||/J,lli^0O = O(r1/2), t VIII.6 Epilogue: The Faber-Krahn Method 249 Proof First assume that v > 2. Then, by Theorem V.3.1, for any discretization G of M we have IV(G) > 0, which implies (VIII.5.4) by the discrete version of Theorem VIII.3.2. But this then implies (VIII.5.3). If v < 2, then consider M\ = M x R2. This implies %+2.p(M\) > 0 by Proposition V.3.6, which implies Iv+2(Gi) > 0 for the discretization Gi = G x Z2 of Mx (where G is a discretization of M), which implies (VIII.5.4), which implies (VIII.5.3) for A/,. But this then implies (VIII.5.3) for M. ¦ VIII.6 Epilogue: The Faber-Krahn Method Recall (§VI.3) that, given a positive increasing function g(v), v > 0, we say that a domain 12 in a Riemannian manifold M satisfies a geometric g-isoperimetric inequality if ACD) > g(V(D)) for all D CC 12. Also, given a positive decreasing function A(v), v > 0, we say that a do- domain 12 in a Riemannian manifold M satisfies an eigenvalue A-isoperimetric inequality if k(D) > A(V(D)) (where X denotes the lowest Dirichlet eigenvalue) for all D CC 12. Proposition VI.3.2 stated that if 12 satisfies a geometric g-isoperimetric in- inequality, with g(v)/v a decreasing function of v, then 12 satisfies an eigenvalue A-isoperimetric inequality with Definition Given a positive decreasing function A(v) on [0, +00), define the function €(f) by /«(') dv vA(v) So we must assume that v >-> {v A(v)} is integrable near v = 0. Equivalently, ?@ is defined as the solution to the initial value problem €' = ?A(C), C@) = 0. Lemma VIII.6.I Suppose ?2? satisfies the eigenvalue A-isoperimetric inequal- inequality, u e Cf°°A2)( u > 0. Then Jor any 8 e @,1) we have ||grad«|||22 ^ n CXA /2||«||,2>
250 Large Time Heat Diffusion Proof For any z > 0 we have / u2 -2zu+z2 = I (u-zJ Ja Ja -L+L < f (u-zJ+ f z2 <f (u-zJ+fz2, Ju>r Ja which implies which implies ' -2t\\u\U= f u2-2ru< [ (u-zJ, Ja Ju>z |grad«|2 > I |grad«|2 >X({u>z])f (u-zJ > A(V({u > z})) f (u-zJ. Now which impbes V({ U> Z})< - I U < - I U, t Ju>x * Ja - rJ > A(Rm|i/t) {||<f||22 - Pickr = ^||«||22/2||M||i,andtheclaimfollows. ¦ Theorem VIII.6.1 If M satisfies the eigenvalue A-isoperimetric inequality, then const. (VIII.6.2) for all t >0. fl 1-00 < C(const.r) Proof Fix y e M, and define u(x,t) = p(x,y,t); VIII.6 Epilogue: The Faber-Krahn Method then fM u(x, t)dV(x) < 1. Define 3@= / u2(x,t)dV(x). Ju Then 3'(O = -2||grad«||22<-2(l- by the previous lemma. Therefore, " } da <jAB/8o) which impbes, letting t0 I 0, da <-2A - 8)(t - t0). f Jm aAB/Sa) Now substitute v = 2/Sa; then > 2A - S)t. r which implies which is the theorem. 'o vA(v) 3@ < > 2A - S)t, S<SB(l - S)t)' 251 Example VIII.6.1 (Example VI.3.1 continued.) Let M satisfy a geometric g- isoperimetric inequality, with g(v) given by where avol~u" = ^vo1^. Then Theorem VI.3.2 impbes that M satisfies a eigenvalue A-isoperimetric inequality, where v~2/\ v < vo, V~2/v, V > Vo, Therefore, one has the existence of to > 0 such that 't*<2, t>to,
252 Large Time Heat Diffusion which implies, by Theorem VIII.6.1, that f,-/l/2 t < tn that is, we have an alternate proof of Theorem VIII.5.4 above. Example VIII.6.2 (Example VI.3.2 continued.) Let M satisfy a geometric g- isoperimetric inequality, with g(v) given by v > v0> where avo1'" = 0. Proposition VI.3.2 implies that M satisfies a eigenvalue A-isoperimetric inequality, where v<v0, 7] = 02/4, V > V0, Therefore, one has the existence oft0>0 such that or',"/', ,<*, which implies, by Theorem VIII.6.1, that which corresponds qualitatively to hyperbolic space [note that 17 is not neces- necessarily k(M) — which is the sharp result, by Proposition VIII. 1.2]. Definition We say that a function /: @, ex?) -> @, 00) has at most polynomial decay if there exists a > 0 such that for all / > 0,0 e [1, 2] we have fifit) > af(t). An example would be a function / satisfying f'(t)/f(t) > -N/t for all t. Theorem VIII.6.2 Given M, consider the functions A(v) and C(r) related by (VIII.6.1), and assume that ?'@/ ?(t)hasat most polynomial decay. If we are also given that for all t > 0, then > const.A(V(n)/*) VIII.7 Bibliographic Notes 253 for all k = 1, 2,..., where A.*(ft) denotes the k th Dirichlet eigenvalue ofth* domain ft. Proof Fix ft. From the Sturm-Liouville eigenvalue-eigenfiinction expansion we have where q is the Dirichlet heat kernel of ft, and {0i, fa,...} is a complete or. thonormal basis of L2(ft) consisting of eigenfunctions of - A«ur with the eigen- eigenvalue of 4>j equal to kj. This implies *«-**'< -k<'= [ q(x,x,t)dV(x)< f p(x,x,t)dV(x)<V(n)/<?(t), ,=, Jo Jo which implies tkk > In for allt > 0. Pick t such that ?(t) = V(ft)/Jt, and pick t = 2t. Then 1 f rf 1 = - { — In <S(t)\ 2 I at J, forsome^ e @, 1) f+e 2 <?(t) which implies the theorem. VIII.7 Bibliographic Notes §VIII.l Proposition VIII. 1.1 is from Cheeger, Gromov. and Taylor A982); also see Chavel A984, §VIII.4). Theorem VIII. 1.2 was first proved in Chavel and Karp A991), and Corollary VIII. 1.2 was communicated privately by P. Li. (See also Li 1986a.) Example VIII. 1.1 is from Varopoulos A989). The arguments we use in subsequent sections all revolve around semigroup theory, in one way or another, namely, the isoperimetric inequalities lead to Sobolev
254 Large Time Heat Diffusion inequalities, which give upper bounds on \\P,\\ /.<_>/,*, for good choices of I and k. The arguments can also yield off-diagonal corrections - see Davies A987, 1989), Fabes and Stroock A986), and Carlen, Kusuoka, and Stroock A987). Davies's argu- arguments A987,1989) emphasize log Sobolev inequalities, instead of the Sobolev inequal- inequalities of the type considered here. A different method, using parabolic Harnack principles for Riemannian manifolds, also gives strong results, including lower bounds on the heat kernel. The early funda- fundamental paper is Li and Yau A986b). In fact. Proposition VIII. 1.3 is from that paper, but our use of the result gave no hint of its ultimate power. More recent fundamental papers in parabolic Harnack inequalities are Grig or'yan A991) and Saloff-Coste A992), as well as the very recent Yau A994,1995) and Bakry and Qian A999). For arguments not relying on global uniformity of Harnack inequalities, see Davies A997). Still another method for producing on-diagonal upper bounds is the use of the isoperi- metric function of a manifold. It has been used to date on compact manifolds (see § V.4); one can find a heat kernel comparison theorem, based on comparison of isoperimet- ric functions, in Beiard and Gallot A983). See also the discussion in Berard A986, Chapter V). The discussion there includes a Schwarz symmetrization argument for the heat kernel, in the spirit of the original Faber-Krahn argument. See the original version of this argument in Bandle A980, Chapter IV). Finally, see an abstract version of the Schwarz symmetrization argument for the heat kernel, based on Kato's inequality, in Besson's appendix to Berard A986). §VIIL2 Theorem VIII.2.1 is from Nash A958), and Theorem VIII.2.2 from Carlen, Kusuoka, and Stroock A987). §Vm_3 Proposition VIII.3.1, the Riesz-Thorin interpolation theorem, can be found in Reed and Simon A980, Vol. I); Proposition VIII.3.2, the maximal theorem, in Stein A970). Theorem VIII.3.2 is from Varopoulos A986). §VUI.4 Theorem Vm.4.2 is fromCoulhon A990). §VIIL5 Lemma VIII.5.1 is from Dodziuk A984), and Theorem VHI.5.2 from CoulhonA992). Theorem VIII.5.4 was first proved in Chavel and Feldman A991), although we preferred here to give the proof of Coulhon A992). The estimate (VIII.5.5) gives the best rate of decay possible, for such generality. The first such result was proved by Varopoulos, in Varopoulos A984), with exponent — j + € for any e > 0. Our result confirms Varopoulos's conjecture, there, that one can sharpen the result to e = 0. §VIII.6 This section is from Grigor'yan A994c). See also Grigor'yan's article in the 1998 Edinburgh Lectures (Davies and Safarov, 1999). The methods of Grigor'yan have been developed to give very strong results on off-diagonal corrections - strictly as consequences of on-diagonal upper bounds alone (see Grigor'yan, 1994a). All of our results on upper bounds were unobtainable as consequences of lower bounds on volume growth alone (recall Example VIII. 1.1). More recently. Barlow, Coulhon, and Grigor'yan B000) have shown that merely postulating V(x;r) > const.A* as r f +00 only implies a rate of decay given by p < /-*/<*+" as t f +00. Furthermore, the result is sharp! Bibliography C. Adams & F. Morgan A999). Isoperimetric curves on hyperbolic surfaces. Proc. Amer.Math. Soc. 129, 1347-1356. R.A. ADAMS A975). Sobolev Spaces. New York: Academic Press. A.D. ALEXANDROV A962). A characteristic property of spheres. Ann. Math. Pura AppL 58,303-315. F. ALMCREN A986). Optimal isoperimetric inequalities. Indiana Univ. Math. J. 35, 451-547. F. ALMCREN & E.H. Lieb A989). Continuity and discontinuity of the spherically decreasing rearrangement. J. Amer. Math. Soc. 2,683-773. T. AREDE A985). Manifolds for which the heat kernel is given interms of geodesic lengths. Lett. Math. Phys. 9,121-131. T. AUBIN A982). Nonlinear Analysis on Manifolds. Monge-Ampire Equations. New York: Springer-Veriag. R. AZENCOTT A974). Behaviour of diffusion semigroups at infinity. Bull. Soc. Math. France 102,193-240. A. Baernstein A995). A unified approach to symmetrization. Symp. Mat. 35, 47-91. D. Bakry, T. Coulhon, M. Ledoux, & L. Saloff-Coste A995). Sobolev inequalities in disguise. Indiana Univ. Math. J. 44, 1033-1074. D. Bakry & Z. Qian A999). Harnack inequalities on a manifold with positive or negative Ricci curvature. Rev. Mat. Iberoamer. 15,143-179. C. BANDLE A980). Isoperimetric Inequalities and Applications. Boston: Pitman. M. Barlow, T. Coulhon, & A.A. Grigor'yan B000). Manifolds and graphs with slow heat kernel decay. Preprint. N. Benabdallah A973). Noyau de diffusion sur les espaces homogenes compacts. Bull. Soc. Math. France 101,265-283. I. Benjamimi & J. CaO A996). A new isoperimetric theorem for surfaces of variable curvature. Duke Math. J. 85, 359-396. I. Benjamini, I. Chavel, & E.A. Feldman A996). Heat kernel lower bounds on Riemannian manifolds using the old ideas of Nash. Proc. London Math. Soc. 72, 215-240. PH. Berard A986). Spectral Geometry: Direct and Inverse Problems, Lecture Notes Math. 1207. Berlin: Springer-Veriag. P.H. Berard & S. Gallot A983). Ine'galite's isope'rime'trique pour 1'equation de la chaleur et application a 1'estimation de quelques invariants. Siminaire Goulaouic-Meyer-Schwartz XV A983-1984). M. BERGER A987). Geometry. Berlin: Springer-Veriag. 2 volumes. W. Blaschke A956). Kreis und Kugel. Berlin: de Gruyter, 2nd edn. 255
256 Bibliography R. Brooks A981). The fundamental group and the spectrum of the Laplacian. Comment. Mat. Helv. 56, 581-598. Yu.D. BURAGO & V.A. ZALGALLER A980). Geometric Inequalities. Berlin: Springer-Verlag, 1988. Original Russian edition: Geometricheskie neravenstva, Leningrad. A. BURCHARD A997). Steiner symmetrization is continuous in W1- p. Geom. & Fcnl. Anal. 7,823-860. P. Buser A982). A note on the isopcrimetric constant. Ann. Sci. ?c. Norm. Sup., Paris 15,213-230. C. CARATHEODORY & E. STUDY A909). Zwei Beweise des Satzes dass der Kreis unter alle Figuren gleichen Umgangs den grOssten Inhalt hat. Math. Ann. 68, 133-144. T. CARLEMAN A921). Zur theorie der Minimalflachen. Math. Z 9,154-160. E.A. CARLEN, S. Kusuoka, & D.W. STROOCK A987). Upper bounds for symmetric Markov transition densities. Ann. Inst. H. Poincari Prob. Stat. 23,245-287. G. CARRON A996). In6galites isopenmetriques de Faber-Krahn et consequences. In Actes de la table ronde de giomitrie dufffrentielle en I 'honneur de Marcel Berber, Collection SMF S6 minaires et Congres 1,203-232. I. Chavel A978). On A. Hurwitz' method in isopcrimetric inequalities. Proc. Amer. Math. Soc. 71,275-279. I. CHAVEL A984). Eigenvalues in Riemannian Geometry. New York: Academic Press. I. Chavel A994). Riemannian Geometry: a modem introduction. Cambridge: Cambridge Univ. Press. I. Chavel & E.A. Feldman A991). Modified isoperimetric constants, and large time heat diffusion in Riemannian manifolds. Duke Math. J. 64,473-499. I. Chavel & L. Karp A991). Large time behavior of the heat kernel: the parabolic A. -potential alternative. Comment. Math. Helv. 66,541-556. J. CHEEGER A970). A lower bound for the smallest eigenvalue of the Laplacian. In Problems in Analysis, R. Gunning ed., 195-199. Princeton Univ. Press. J. CHEEGER, M. GROMOV, & M. Taylor A982). Finite propagation speed, kernel estimates for functions of the Laplace operator, and the geometry of complete Riemannian manifolds. J. Diff. Geom. 17,15-54. S.Y. Cheng & P. Li A981). Heat kernel estimates and lower bounds of eigenvalues. Comment. Mat. Helv. 56, 327-338. K.S. Chou & X.P. Zhu B001). 77k Curve Shortening Problem. Chapman & Hall. To appear. T. COULHON A990). Dimension a l'infini d'un semigroup analytique. Bull. Soc. Math. France 114,485-500. T. COULHON A992). Noyau de chaleur et discretisation d'une varied riemanniene. Israel J. Math. 80,289-300. T. COULHON & A.A. Grigor'yan A997). On-diagonal lower bounds for heat kernels on non-compact manifolds and Markov chains. Duke Math. J. 89,133—199. T. COULHON & M. LEDOUX A994). Isoplrimftrie, decroisaance du noyau de la chaleur et transformations de Riesz: un contra-example. Ark. Math. 32,63-77. T. COULHON & L. Salopf-Coste A993). Isopenmettie pour les groupes et les vari&es. Rev. Mat. Iberoamer. 9, 293-314. T. COULHON & L. SalOFF-Coste A995). Varietes riemanniennes isoperimetriques a la l'infini. Rev. Mat. Iberoamer. 3,687-726. R. COURANT & H. ROBBINS A941). What« Mathematics? Oxford: Oxford Univ. Press. C.B. CrOKE A980). Some isoperimetric inequalities and eigenvalue estimates. Ann. Sci. tc. Norm. Sup. Paris 13, 419-435. Bibliography 257 E.B. Davies A980). One-Parameter Semigroups. London: Academic Press E.B. Davies A987). Explicit constants for Gaussian upper bounds on heat kernels Amer. J. Math. 109,319-334. E.B. Davies A989). Heat Kernels and Spectral Theory. Cambridge: Cambridge Univ. Press. E.B. Davies A992). Heat kernel bounds, conservation of probability and the Feller property. J. Analyse Math. 58,99-119. E.B. Davies A995). Spectral Theory and Differential Operators. Cambridge- Cambridge Univ. Press. E.B. Davies A997). Non-gaussian aspects of heat kernels behaviour. J. London Math. Soc. 55,105-125. E.B. Davies & N. Mandouvalos A987). Heat kernel bounds on manifolds with cusps. J. Fcnl. Anal 75, 311-322. E.B. Davies & Y. Safarov, eds. A999). Spectral Theory and Geometry. Cambridge- Cambridge Univ. Press. U. DiERKES, S. HILDEBRANDT, A. KiJSTER, & O. Wohlrab A992). Minimal Surfaces. Berlin: Springer-Verlag. J. DODZIUK A983). Maximum principle for parabolic inequalities and heat on open manifolds. Indiana Univ. Math. J. 32,703-716. J. DODZIUK A984). Difference equations, isoperimetric inequality, and transience of certain random walks. Trans. Amer. Math. Soc. 284, 787-794. C. Faber A923). Beweiss, dass unter alien homogenen Membrane von gleicher Flachc und gleicher Spannung die kreisfftrmige die tiefsten Grundton gibt. Sitzungber. BayerAkad. Wiss., Matk-Phys., Munich, 169-172. E.B. FABES & D.W. STROOCK A986). A new proof of Moser's parabolic inequality via the old ideas of Nash. Arch. Rat. Meek Anal. 96, 327-338. H. FEDERER A969). Geometric Measure Theory. New York: Springer-Verlag H. FEDERER & W.H. FLEMING A960). Normal integral currents. Ann. Math. 72, H.D. Fegan A976). The heat equation on a compact Lie group. Trans. Amer. Math. Soc. 246,339-357. H.D. Fegan A983). The fundamental solution of the heat equation on a compact Lie group. J. Diff. Geom. 18,659-668. T. FiGffiL, J. Undenstrauss, & V.D. Milman A977). The dimension of almost spherical sections of convex bodies. Ada Math. 139, 53-94. H.R. FISCHER, JJ. Jungster, & F.L. WILLIAMS A984). The heat kernel on the 2-sphere. Adv. in Math. 54, 226-232. M.P. GAPFNEY A954). A special Stokes' theorem for complete Riemannian manifolds. Ann. Math. 60, 140-145. M.P. Gaffney A959). The conservation property of the heat equation on complete Riemannian manifolds. Comm. PureAppl. Math. 12,1-11. M. Gage A984). Curve shortening makes convex curves circular. Invent. Math. 76, 357-364. M. Gage A991). References for geometric evolution equations. Dept. of Math Univ of Rochester. D. GlLBARG & N.S. TRUDINGER A977). Elliptic Partial Differential Equations of Second Order. Berlin: Springer-Verlag. E. Giusn A984). Minimal Surfaces and Functions of Bounded Variation Boston- Birkh&user. A.A. GRIGOR'YAN A985). Isoperimetric inequalities for Riemannian products (in Russian). Mat. Zametki 38. Eng. Transl. Math. Notes 38, 849-854 MR: 87g53O62.
258 Bibliography A.A. GRIGOR'YAN A986). On stochastically complete manifolds. Soviet Math. DokL 290. AMS TransL 34 A987), 310-313. A.A. GRIGOR'YAN A991). The heat equation on noncompact Riemannian manifolds. Mat. Sb. 182. AMS TransL 72 A992), 47-77. A.A. GRIGOR'YAN A994a). Gaussian upper bounds for the heat kernel and for its derivatives on a Riemannian manifold. In Classical and Modem Potential Theory and Applications, K. GowriSankaran et al., eds., 237-2S2. Kluwer Academic. A.A. GRIGOR'YAN A994b). Heat kernel on a manifold with a local Harnack inequality. Comm. Anal. Geom. 2,111-138. A.A. GRIGOR'YAN A994c). Heat kernel upper bounds on a non-compact manifold. Rev. Mat. Iberoamer. 10, 395-452. A.A. GRIGOR'YAN A999). Analytic and geometric background of recurrence and non-explosion of the Brownian motion on Riemannian manifolds. Bull. Amer. Math. Soc. 36, 135-249. A.A. GRIGOR'YAN & M. NOGUCM A998). The heat kernel on hyperbolic space. J. London Math. Soc. 30,643-650. M. GROMOV A986). Isoperimetric inequalities in Riemannian manifolds. In Asymptotic Theory of Finite Dimensional Normed Spaces, Lecture Notes Math. 1200, Appendix I, 114-129. Berlin: Springer-Veriag. H. Hadwiger A957). Vorlesungen uberlnhalt, Oberfi&che, und Isoperimetrie. Berlin: Springer-Veriag. R.Z. Hasminkd A960). Ergodic properties of recurrent diffusion processes and stabilisation of the solution of the Cauchy problem for parabolic equations. Theor. Prob. Anal. 5, 179-196. E. HEBEY A999). Nonlinear Analysis on Manifolds: Sobolev Spaces and Inequalities. Inst. Math. Sci. New York: Courant. M.W. HlRSCH A976). Differential Topology. New York: Springer-Veriag. I. HOLOPAINEN A994). Rough isometries and p -harmonic functions with finite Dirichlet integral. Rev. Math. Iberoamer. 217,459-477. H. HOPF A950). Cber Flachen mit einer Relation zwischen den Hauptkrummungen. Mam. Nachr. 4, A950-1), 232-249. H. HOPF & W. RINOW A931). tiberden Begriff der vollstindigen differential- geometrischen Flachen. Comment. Mat. Helv. 3, 209-225. H. Howards, M. Hutchings, & F. Morgan A999). The isoperimetric problem on surfaces. Amer. Math. Monthly 106,430-439. A. HuRwrrz A901). Sur le probleme des isoperim&res. C. R. Acad. Sci. Paris 132, 401-403. K. ICHIHARA A982). Curvature, geodesies, and the Brownian motion on a Riemannain manifold ii. Nagoya Math. J. 87,115-125. K. ICHIHARA A984). Explosion problems for symmetric diffusion processes. Proc. Japan Acad. 60, 243-245. K. Ichihara A986). Explosion problems for symmetric diffusion processes. Trans. Amer. Math. Soc. 298, 515-536. M. Kanai A985). Rough isometries, and combinatorial approximations of geometries of noncompact Riemannian manifolds. J. Math. Soc. Japan 37, 391-413. M. Kanai A986a). Analytic inequalities, and rough isometries between noncompact Riemannian manifolds. In Curvature and Topology of Riemannian Manifolds, K. Shohama, T. Sakai, & T. SunaDa, cds., Lecture Notes Math. 1201, 122-137. Berlin: Springer-Veriag. M. Kanai A986b). Rough isometries and the parabolicity of Riemannian manifolds. J. Math. Soc. Japan 38,227-238. Bibliography 259 L. Karp A984). Noncompact manifolds with purely continuous spectrum. Mich. Math. J. 31, 339-347. L. KARP & P. Li. The heat equation on complete Riemannian manifolds. Preprint. B. Kawohl A985). Rearrangements and Convexity of Level Sets in PDE, Lecture Notes Math. 1150. Berlin: Springer-Veriag. H. KNOTHE A957). Contributions to the theory of convex bodies. Mich. Math. J. 4, 39-52. E. KRAHN A925). Uber cine von Rayleigh formulierte Minimaleigenschafte des Kreises. Math. Ann. 94,97-100. P. Li A980). On the Sobolev constant and the p -spectrum of a compact Riemannian manifold. Ann. Sci. tc. Norm. Sup. Paris 13,419-435. P. Li A986). Large time behavior of the heat equation on complete manifolds with nonnegative Ricci curvature. Ann. Math. 124,1-21. P. Li & S.T. Yau A986). On the parabolic kernel of the Schrodinger operator. Ada Math. 156, 153-201. E.H. LiEB & M. LOSS A996). Analysis. Providence, RI: American Math. Soc. H. LlEBMANN A900). Ober die Verbiegung der glossenen Flachen positiver Krummung. Math. Ann. 53,81-112. V.G. Maz'ya A960). Classes of domains and embedding theorems for functional spaces (in Russian). Dokl. Acad. Nauk. SSSR133,527-530. Engl. transl., Soviet Math. Dokl. 1 A961), 882-885. J. MlLNOR A968). A note on curvature and the fundamental group. J. Dig. Geom. 2, 1-7. F. MORGAN A988). Geometric Measure Theory: A Beginner's Guide. Boston: Academic Press. J. MOSER A964). A Harnack inequality for parabolic differential equations. Comm. PureAppl. Math. 17, 101-134. J. MOSSINO A984). Inigalitis isopirimitriques et application en physique. Paris: Herman. R. NaRASIMHAN A968). Analysis on Real and Complex Manifolds. Amsterdam: North Holland. J. Nash A958). Continuity of solutions of parabolic and elliptic equations. Amer. J.Math. 80, 931-954. M.N. NDUMU A986). An elementary formula for the Dirichlet heat kernel on Riemannian manifolds. In From Local Times to Global Geometry, Control and Physics, K.D. ElwORTHY, ed., n -Pitman Res. Notes Math. 150,320-328. Essex: Longman Scientific & Technical. L. Nirenberg A959). On elliptic partial differential equations. Ann. Scuola Norm. Sup. Pisa 13,115-162. J. OPREA A997). Differential Geometry and its Applications. Upper Saddle River, NJ: Prentice Hall. Y. Oshima A992). On conservativeness and recurrence criteria of Markov processes. Preprint. Potential Analysis 1, 115-131. R. OssERMAN A978). The isoperimetric inequality. Bull. Amer. Math. Soc. 84, 1182-1238. R. OSSERMAN A979). Bonnesen-style inequalities. Amer. Math. Monthly 86, 1-29. M. Pang A988). L' properties of two classes of singular second order elliptic operators. J. London Math. Soc. 38,525-543. M. Pang A996). L' and L2 spectral properties of a class of singular second order elliptic operators with measurable coefficients on R". J. Diff. Eq. 129, 1-17.
260 Bibliography P. PanSU A982). Une inegalit? isoperimetrique sur le groupc de Heisenberg. C. R. AceuL Sci. Paris 295,127-130. L.E. PAYNE A967). Isopcrimetric inequalities and their applications. SIAMRev. 9, 453-488. G. P6lya & G. SzbgO A951). Isoperimetric Inequalities in Mathematical Physics, Ann. Math. Studies 27. Princeton, NJ: Princeton Univ. Press. M.H. Protter & H.F. Weinberger A984). Maximum Principles in Differential Equations. New York: Springer-Verlag, 2nd edn. (Lord) J.W.S. RAYLEIOH A877). The Theory of Sound. New York: Macmillan. Reprinted, New York: Dover A945). M. REED & B. SIMON A975). Functional Analysis (Methods of Mathematical Physics, Vol. II). New York: Academic Press. M. REED & B. SIMON A980). Functional Analysis (Methods of Mathematical Physics, Vol. I). New York: Academic Press, 2nd edn. R.C. REILLY A977). Applications of the Hessian operator in a Riemannian manifold. Indiana Univ. Math. J. 26,459-472. F. Rffisz & B. Sz. Nagy A955). Functional Analysis. New York: Frederick Ungar. A. Ros A988). Compact surfaces with constant scalar curvature and a congruence theorem. J. Diff Geom. 27,215-220. W. RUDIN A966). Real and Complex Analysis. New York: McGraw-Hill. L. SalOFF-Coste A992). A note on Poincani, Sobolev, and Hamack inequalities. Duke Math. J. 65, International Mathematics Research Notices, 27-38. E. SCHMIDT A948,1949). Der Brunn-Minkowskische Satz und sein Spiegel-theorem sowie die isoperimetrische Eigenschaft der Kugel in dereuklidischen und nichteuklidischen Geometrie I, II. Math. Nachr. 1, 81-157; 2, 171-244. L.M. SIMON A984). Lectures on Geometric Measure Theory. Canberra, Australia: Centre Math. Anal., Australian Nat. Univ. E.M. Stein A970). Topics in Harmonic Analaysis, Ann. Math. Studies 63. Princeton, NJ: Princeton Univ. Press. J. STE1NER A838). Einfache Beweise der isoperimetrische Haupts&tze. J. Reine Angew. Math. 18,281-296. Reprinted, Gesammelte Werke. Bronx, NY: Chelsea A971) (reprint of 1881-1882 edn.), Vol. 0, 75-91. J. STEINER A881). Uber Maximum und Minimum. In Gesammelte Werke. Bronx, NY: Chelsea A971) (reprint of 1881-1882 edn.). D. Sullivan A987). Related aspects of positivity in Riemannian geometry. J. Diff. Geom. 25, 327-351. M. Takeda A989). On a martingale method for symmetric diffusion processes and its applications. Osaka Math. J. 26,605-623. M. Takeda A991). On the conservativeness of the Brownian motion on Riemannian manifolds. Bull. London Math. Soc. 23,86-88. G. Talenti A993). The standard isoperimetric theorem. In Handbook of Convex Geometry, Vol. A, P.M. Gruber & J.M. Wills, eds., 73-123. Amsterdam: North Holland. P. TOPPING A997). The optimal constant in Wente's L"° estimate. Comment. Mat. Helv. 139,316-328. P. Topping A998). Mean curvature flow and geometric inequalities. J. Reine Angew. Math. 503,47-61. P. Topping A999). The isoperimetric inequality on a surface. Manuscripta Math. 100, 23-33. N.TH. VAROPOULOS A983). Potential theory and diffusion on Riemannian manifolds. In Conf. on Harmonic Analysis in Honor of A. Zygmund, W. BECKNER et al., eds. Wads worth. Bibliography 261 N.TH. VAROPOULOS A984). Brownian motion and random walks on manifolds Ann Inst. Fourier Grenoble 34,243-269. N.Th. VAROPOULOS A985). Isoperimetric inequalities and Markov chains J Fail Anal. 63,240-260. ' " N.Th. VAROPOULOS A986). Hardy-Littlewood theory for semigroups. J. Fcnl. Anal. 66,406-431. N.TH. VAROPOULOS A989). Small time Gaussian estimates of heat diffusion kernels Part i: The semigroup technique. Bull. Soc. Math. France 113. 253-277. S.T. Yau A975). Isoperimetric constants and the first eigenvalue of a compact manifold. Ann. Sci. tc. Norm. Sup. Paris 8,487-507. S.T. Yau A976). Some function-theoretic properties of complete Riemannian manifolds, and their applications to geometry. Indiana Univ. Math. J. 25, 659-670. Also cf. Ibid. 31 A982), 307. S.T. Yau A978). On the heat kernel of a complete Riemannian manifold. J. Math. PuresAppL 57,191-201. S.T. Yau A994). On the Hamack inequalities of partial differential equations. Comm. Anal, and Geom. 2,431-450. S.T. Yau A995). Harnack inequality for non-self-adjoint evolution equations. Math. Res. Lett. 2, 387-399. K. YOSIDA A978). Functional Analysis. Berlin: Springer-Verlag, 5th edn. W.P. Ziemer A989). Weakly Differentiable Functions. New York: Springer-Verlag.
Author Index Adams, C, 29 Adams. R.A.. 29. 224 Alexandrov,A.D.,51 Almgren,F..51,99 Arede, T., 224 Aubin.T, 184 Aubin, T., 224 Azencott, R., 224 Baemstein, A., 99 Bakry.D., 183,254 Bandle, C, 183. 254 Barlow, M., 254 Bendabdallah, A.. 224 Benjamini, I., 29, 233 Berard, PH., 184,254 Berger. M., 29,99 Besson, G., 254 Blaschke, W., 28 Brooks, R., 232 Burago, Y, 3.28,99 Burchard, A., 99 Buser.P., 155 Cao, J., 29 Caratheodory, C, 99 Carleman, X, 51 Carlen, E.A., 254 Carron. G.. 184 Chavel, I., 29,99,155, 183-185, 224, 233, 253,254 Cheeger.J., 183,253 Cheng, S.Y., 183 Coulhon.T. 155, 183, 184,233, 254 Courant, R., 28 Croke,C, 155 Davies, E.B., 224. 254 Dierkes, U., 28 DodziulcJ.. 184,224.254 Faber, C, 99 Fabes, E., 254 Federer. H., 29,51,99, 116.183 Fegan. H.D., 224 Feldman, E.A.. 184.233,254 Figiel, T. 29 Fischer, H.R., 224 Fleming. W.H., 51, 183 Gaffney, M.P., 224 Gage. M., 29 Gallot, S., 254 Gilbarg. D., 29,224 Giusti, E, 99. 100 Grigor'yan, A.A., 155. 156. 184,224. 233,254 Gromov, M., 29.51. 184, 253 Hadwiger. H., 28 Hasminkii, R.Z., 224 Hildebrandt, S.. 28 Hirsch, M.W., 29 Holopainen, I., 156 Hopf,H.,51.155 Howards. H.. 29 Hurwitz, A.. 29 Hutchings, M., 29 lchihara, K.. 224 Jungster, J J., 224 Kanai.M., 154, 156. 177, 184 Karp, L., 224,253 Kawohl, B., 99, 183 Knothe, H, 51 Krahn. E., 99 KUrster. A.. 28 Kusuoka. S., 254 Ledoux,M., 183 U. P., 183,222,224,253,254 263
264 Author Index Ueb, E.H.. 29.99 Liebmann. H., 31 Lindenstrauss, J., 29 Loss. M.. 29.99 Mandouvalos, N., 224 Maz'ya.V.G.,51 Milman. V.D.. 29 Milnor.J., 143 Morgan, E, 29,99. 116 Moser, J., 183 Mossino, J.. 183 Nagy,B.Sz..221 Narasimhan, R., 29 Nash. J., 183,254 Ndumu, M.N., 224 Nirenbcrg, L., 183 Noguchi, M., 224 Oprea.J.,51 Oshima, Y., 224 Osserman. O.. 28,29 Osserman, R., 29 Pang, M.. 224 Pansu. P.. 248 Payne, L.E., 183 Polya.G.,99,183 Proner, M.H., 29,224 Qian. Z.. 254 Rayleigh. Lord. 99 Reed, M., 224,254 ReiUy.R.C.,51 Riesz, E, 221 Rinow, W.. 155 Robbins. H., 28 Ros, A., 51 Rudin, W., 29 Saloff-Coste, L., 155,183. 184.254 Schmidt, E. 29 Simon, B., 224,254 Simon, L., 99,116 Stein. E.M., 254 Steiner, J.. 99 Stroock, D.W., 254 Study. E., 99 Sullivan, D., 224 SzegO.G.,99,183 Takeda, M., 224 Talenti, G.. 28.99 Taylor, M.. 253 Topping. P.. 29 Trudinger, N.S., 29, 224 Varopoulos, N.Th., 156, 184,224, 253,254 Weinberger, H.E, 29,224 Williams, F.L., 224 Wohlrab, O.. 28 Yau, S.T, 224 Yau, ST., 131,224.254 Yosida, K.. 224 Zalgaller, V.A., 3.28.99 Ziemer. W.P., 29.99 Subject Index i-cover, 102 f-thickening, 15.54 immeasurability, 101 o -algebra, 100 approximate identity, 229 arc length function, 4 area of boundaries in graphs, 142 integral geometric, 53 of metric sphere in Riemannian manifold, 125 lower. 125 Minkowski, 58 surface, 2, 3, 17.25,31 boundaries in graphs, 142 area of. 142 boundary conditions, vanishing Dirichlet, 196 bounded geometry for graphs, 140 for Riemannian manifolds, 136 Cheeger's constant, 127 < in compact case, 162 modified. 130 circumdisk, 57 circumradius, 53 combinatorial length of. 140 metric, 140 path. 140 complete, geodesically, 120 connection. 121 Levi-Civita, 121 constant alternate Sobolev in compact case. 163 Cheeger's, 127 in compact case, 162 modified. 130 isoperimetric of compact Riemannian manifold. 162 of graphs. 144 modified, of Riemannian manifold, 130 of R\ 44,45 of Riemannian manifold, 127 Lipschitz.20 Sobolev in compact case, 162 of graphs, 173 ofR\44.45 of Riemannian manifold, 131 convergence of sets in Z.1, 81 convex function, 22 hull of. 28 linear combination, 27 set, 27 coordinates Euclidean. 17 geodesic spherical. 125 Riemann normal, 123 countable additivity, 100 countably subadditive, 101 curvature Gauss, 32 mean, 32.36,44 of path in R2, 5 Ricci, 122 sectional, 122 constant, 123 tensor, 121 cut locus, domain inside, of a point. 125 differential operator elliptic, 19 uniformly, 19 Dirichlet energy integral, 195 discretization of Riemannian manifold, 147 covering radius of, 147 separation radius of, 147 265
266 Subject Index Subject Index 267 distance combinatorial, between points, 140 Hausdorff, between compact sets. 54 Riemannian, between points, 119 domain, 3 Lipschitz, 87 exponential map, 120 formula Cauchy's, for area of boundaries of convex domains, 26 co-area for perimeter, 89 for smooth functions, 25 for smooth mappings, 25 Fourier transform, 203 Frenet formulae, 5 function of bounded variation, 82 characteristic, 13 differential of. on graphs, 144, 170 discretization of, 178 gradient of, on Riemannian manifold, 157 indicator, 13 Laplacian of, on Riemannian manifold, 193 acting as L1 distribution, 198 Friedrichs extension of, 195 Lipschitz, 19 mean value, 162 polynomial decay, 252 smoothing of. 179 weak derivative of, 23 Z.2,195 functional isoperimetric, 36 fundamental form first, 31 second. 32 fundamental tone. 78 Gauss map, 32 geodesic, 119 completeness. 120 flow, 126 kiuematic density of, 127 Lkxiville's theorem, 127 spherical coordinates, 125 graph, 140 with bounded geometry, 140 Cayley, of group, 141 combinatorial path in, 140 distance between points, 140 Hausdorff dimension, 105 distance between compact sets, 54 measure, 102 standard, 104 heat equation homogeneous, 201 inhomogeneous. 201 uniqueness property. 212 heat flow, 214 heat kernel, 202 closed, 208 conservation of heat property of, 208 Dirichlet, 209 Feller property of, 208 minimal positive, 211 positive, 208 stochastic completeness of, 208 symmetric. 208 heat operator associated with the Laplaciai 201 hyperplane separation by, 28 strict, 28 supporting, 28 incompressible flow, 39 inequality arithmetic-geometric mean, 38,49 Bonnesen, 4 Brunn-Minkowski, 62 Buser's, 136 Crake's, 136 eigenvalue A-isoperimetric, 169 Faber-Krahn, 79 of Euclidean type, 166 geometric g-isoperimetric, 168 isodiametric, 61 isoperimetric, 2-4 for C1 surface area, 47,51 eigenvalue A-, 249 geometric g-. 249 for Minkowski area. 63 for Minkowski area in R", 77 for Minkowski length in R2.74 in model space, 4 for perimeter, 94 inR2.8,9 Wirtinger-s, 8. 29 injectivity radius of point, 120 Riemannian manifold, 120 inradius, 130 isodiametric inequality, 61 isometry quasi-, of metric spaces, 140 quasi-, of Riemannian manifolds, 128 of Riemannian manifolds, 123 rough, of metric spaces, 142 isoperimetric constant of compact Riemannian manifold, 162 of graphs, 144 modified, of Riemannian manifold, 130 of R\ 44,45 of Riemannian manifold, 127 function, 254 functional. 1,11.36. 117 C* external of the. 31 uniqueness of C2 extremal of, 39 inequality forC1 surface area. 47.51 eigenvalue A-, 249 geometric g-, 249 for Minkowski area. 63 for Minkowski area in R", 77 for Minkowski length in R2, 74 in model space, 4 for perimeter, 94 in R2.8,9 problem, 2 in model space, 4,29 solution to the C*,31 uniqueness of C2 solution to. 36 Laplacian on Riemannian manifold, 193 acting as L2 distribution. 198 Dirichlet, 196 Friedrichs extension of. 195 Lebesgue set, 96 length of path in Riemannian manifold. 119 vector in Riemannian metric, 118 linear operator adjoint of, 185 closable. 185 closed, 185 closure of. 185 extension of, 185 minimal, 185 resolvent of, 187 resolvent set of. 187 self-adjoint, 186 approximate eigenfuctions of, 189 essentially. 186 spectral family of, 187 semibounded from below, 186 Friedrichs extension of. 186 spectrum of, 187 discrete, 188 essential, 188 point, 188 symmetric, 186 nonnegative, 186 Lipschitz constant, 20 domain. 87 function. 19 manifold closed. 161 compact, 161 measure, 100 R* -valued, 15 absolutely continuous, 15 Radon, 16 total variation of, 15 counting, 142 Hausdorff, 102 5-, 102 standard, 104 standard J-, 104 inner regular, 16 kinematic, of geodesic flow, 127 locally finite. 16 outer, 101 Borel regular, 101 metric, 101 regular, 16, 101 Radon. 16 regular, 16 volume, on graphs, 142 metric combinatorial, 140 Riemannian, 118 word, on group, 141 model space. 124 isoperimetric inequality in, 4 isoperimetric problem in. 4,29 mollifier, 17 natural basis of tangent spaces, 118 norm on group displacement. 141 word. 141 orthogonal projection, 187 perimeter. 88 geometric, 81 set with locally finite, 88 point, critical, 18 principle Cavalieri. 14 Duhamel's, 207 strong maximum elliptic. 19 parabolic. 207 product Riemannian structure, 118
268 Subject Index quasi-isometry of metric spaces, 140 of Riemannian manifolds, 128 Riemann normal coordinates, 123 Riemannian distance between points, 119 manifold, 118 discretization of. 147 length of path in. 119 metric. 118 length of vector in, 118 rough isometry of metric spaces, 142 semigroup, one-parameter, 190 contractive, 190 generator (infinitesmal), 191 domain of, 191 sesquilinear form, 190 nonnegative, 190 semibounded from below, 190 set Cacciopoli, 88 t -separated, 146 partition of, IS Sobolev constant alternate, 180 in compact case. 162 of graphs. 173 ofR\44,45 of Riemannian manifold, 131 Sobolev spaces, 23 surface area, 2, 3, 17,25,31 symmetric decreasing rearrangement of functions, 70 symmetrization Schwarz, 69 Steiner, 59 tensor curvature, 121 Ricci, 122 theorem Alexandrov's, 42,51 Bishop's comparison, 126 Bishop-Gromov comparison, 126 Blaschke selection, 55 CaratModory's. 101 Coulhon extrapolation, 241 divergence Riemannian, 32,193 2-dimensional, 9 Federer-Fleming, 45,51 in Riemannian manifolds, 131 Helly selection, 73 Lebesgue density, 17 Liouville's, 127 maximal, 237 Mercer's, 221 Mcyers-Serrin, 196.224 Rademacher, 20 Radon-Nikodym, 16 Riesz representation, 16 Ricsz-Thorin interpolation, 236 Sard. 18 spectral, 187 value critical, 18 regular, 18 vanishes at infinity, function that, 69 vector field admissible, 82 divergence of, on Riemannian manifold, 192 radial, 128 volume growth. 143