Classics in Mathematics Charles B. Morrey, Jr.

Multiple Integrals in the Calculus of Variations

Charles B. Morrey, Jr.

Multiple Integrals in the Calculus of Variations

Reprint of the 1966 Edition

123

Originally published as Vol. 130 of the series Grundlehren der mathematischen Wissenschaften

ISBN 978-3-540-69915-6 DOI 10.1007/978-3-540-69952-1 Classics in Mathematics ISSN 1431-0821

Library of Congress Control Number: 2008932928

e-ISBN 978-3-540-69952-1

Mathematics Subject Classication (2000): 49-xx, 58Exx, 35N15, 46E35, 46E39 2008, 1966 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microlm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specic statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: WMXDesign GmbH, Heidelberg Printed on acid-free paper 987654321 springer.com

Die Grundlehren der mathematischen Wissenschaften

in Ein2;eldarstellungen mit besonderer Beriicksichtigung der Anwendungsgebiete Band 130

Herausgegehen von

J. L. Doob E. Heinz F. Hir^iebruch E. Hopf H. Hopf W. Maak S. MacLane W. Magnus D. Mumford F. K. Schmidt K. Stein

Geschdjtsjuhrende Heramgeher

B. Eckmann und B. L. van der Waerden

Multiple Integrals in the Calculus of Variations

Charles B. Morrey, Jr.

Professor of Mathematics University of California, Berkeley

Springer-Verlag Berlin Heidelberg New York 1966

Geschaftsfiihrende Herausgeber:

Prof. Dr. B. Eckmann

Eidgenossische Technische Hochschule Zurich

Prof. Dr. B. L. van der Waerden

Mathematisches Institut der Universitat Zurich

All right reserved, especially that of translation into foreign languagesi It is also forbidden to reproduce this book, either whole or in part, by photomechanical means (photostat, microfilm and/or microcard or any other means) without written permission from the Publishers (c) by Springer-Verlag, Berlin Heidelberg 1966 Library of Congress Catalog Card Number 66-24 365 Printed in Germany

Title No. 5113

Preface
The principal theme of this book is ''the existence and differentiability of the solutions of variational problems involving multiple integrals/' We shall discuss the corresponding questions for single integrals only very briefly since these have been discussed adequately in every other book on the calculus of variations. Moreover, applications to engineering, physics, etc., are not discussed at all; however, we do discuss mathematical applications to such subjects as the theory of harmonic integrals and the so-called ''^^Neumann" problem (see Chapters 7 and 8). Since the plan of the book is described in Section 1.2 below we shall merely make a few observations here. In order to study the questions mentioned above it is necessary to use some very elementary theorems about convex functions and operators on Banach and Hilbert spaces and some special function spaces, now known as ''SOBOLEV spaces". However, most of the facts which we use concerning these spaces were known before the war when a different terminology was used (see CALKIN and MORREY [5]); b u t we have included some powerful new results due to CALDERON in our exposition in Chapter 3. The definitions of these spaces and some of the proofs have been made simpler b y using the most elementary ideas of distribution theory; however, almost no other use has been made of that theory and no knowledge of t h a t theory is required in order to read this book. Of course we have found it necessary to develop the theory of linear elliptic systems at some length in order to present our desired differentiability results. We found it particularly essential to consider ''weak solutions'' of such systems in which we were often forced to allow discontinuous coefficients; in this connection, we include an exposition of the D E GIORGINASHMOSER results. And we include in Chapter 6 a proof of the analyticity of the solutions (on the interior and at the boundary) of the most general non-linear analytic elliptic system with
general regular (as in AGMON, DOUGLIS, and NIRENBERG) boundary con-

ditions. But we confine ourselves to functions which are analytic, of class C"^, of class C^ or C^ (see 1.2), or in some Sobolev space H^ with m an integer > 0 (except in Chapter 9). These latter spaces have been

vi

Preface

defined for all real w in a domain (or manifold) or on its boundary and have been used by many authors in their studies of linear systems. We have not included a study of these spaces since (i) this book is already sufficiently long, (ii) we took no part in this development, and (iii) these spaces are adequately discussed in other hooks (see A. FRIEDMAN [2] HoRMANDER [1], LiONS [2]) as wcU as in many papers (see 1.8 and papers by LIONS and MAGENES). The research of the author which is reported on in this book has been partially supported for several years by the Office of Naval Research under contract Nonr 222(62) and was partially supported during the year 196162, while the author was in France, by the National Science Foundation under the grant G19782. Berkeley, August I966
CHARLES B . MORREY, JR.

Contents
Chapter 1 Introduction
1.1. 1.2. 1.3. 1.4. 1.5. 1.6. 1.7. 1.8. 1.9. 1.10. 1.11. Introductory remarks T h e p l a n of t h e b o o k : n o t a t i o n V e r y brief h i s t o r i c a l r e m a r k s T h e EuLER e q u a t i o n s O t h e r classical n e c e s s a r y c o n d i t i o n s Classical sufficient c o n d i t i o n s The direct methods Lower semicontinuity Existence T h e differentiability t h e o r y . I n t r o d u c t i o n Differentiability; reduction t o linear equations 1 2 5 7 10 12 15 19 23 26 34

Chapter 2 Semi-classical 2.1. 2.2. 2.3. 2.4. 2.5. 2.6. 2.7. 2.8. results 39 40 41 43 47 48 55 6I

Introduction E l e m e n t a r y p r o p e r t i e s of h a r m o n i c f u n c t i o n s WEYL'S lemma P O I S S O N ' S i n t e g r a l f o r m u l a ; e l e m e n t a r y f u n c t i o n s ; G R E E N ' S functions Potentials Generalized potential t h e o r y ; singular integrals The CALDERON-ZYGMUND inequalities T h e m a x i m u m p r i n c i p l e for a l i n e a r elliptic e q u a t i o n of t h e s e c o n d o r d e r

Chapter 3 T h e s p a c e s H"^ a n d H'^Q 3.1. 3.2. 3.3. 3.4. 3.5. 3.6. 3.7. D e f i n i t i o n s a n d first t h e o r e m s G e n e r a l b o u n d a r y v a l u e s ; t h e s p a c e s ii/^Q (G); w e a k c o n v e r g e n c e . T h e DiRiCHLET p r o b l e m B o u n d a r y values E x a m p l e s ; c o n t i n u i t y ; s o m e SoBOLEV l e m m a s Miscellaneous a d d i t i o n a l r e s u l t s Potentials and quasi-potentials; generalizations . . 62 68 70 72 78 81 86

viii

Contents Chapter 4

4.1. 4.2. 4.3. 4.4.

Existence theorems The lower-semicontinuity theorems of SERRIN Variational problems w i t h / = f{p); the equations (1.10.13) with N = \, Bi = 0, A^" = A<^{p) The borderline cases k = v The general quasi-regular integral Chapter 5 D i f f e r e n t i a b i l i t y of w e a k s o l u t i o n s

90 98 105 112

5.1. 5.2. 5.3. 5.4.

5.5.

Introduction General theory; r > 2 . . Extensions of the DE GIORGI-NASH-MOSER results; v > 2 The case I' = 2
Lp and SCHAUDER estimates

126 128 134 143

149

5.6. 5.7. 5.8. 5.9. 5.10.

5.11.

The equation a-V^w + ^ *Vw + c w Aw = / Analyticity of the solutions of analytic linear equations Analyticity of the solutions of analytic, non-linear, elliptic equations Properties of the extremals; regular cases The extremals in the case 1 < ^ < 2
The theory of LADYZENSKAYA and URAL'TSEVA

157 i64 170 186 191

194

5.12. A class of non-linear equations Chapter 6

203

R e g u l a r i t y t h e o r e m s for t h e s o l u t i o n s of g e n e r a l elliptic systems and boundary value problems 6.1. Introduction 209 6.2. Interior estimates for general elliptic systems 215 6.3. Estimates near the boundary; coerciveness 225 6.4. Weak solutions 242 6.5. The existence theory for the DIRICHLET problem for strongly elliptic systems 251 6.6. The analyticity of the solutions of analytic systems of linear elliptic equations 258 6.7. The analyticity of the solutions of analytic nonlinear elliptic systems 266 6.8. The differentiability of the solutions of non-linear elliptic systems; weak solutions; a perturbation theorem 277 Chapter 7 7A. 7.2. 7.3. 7.4. 7.5. 7.6. y.y. 7.S. A v a r i a t i o n a l m e t h o d in t h e t h e o r y of h a r m o n i c i n t e g r a l s Introduction Fundamentals; the GAFFNEY-GARDING inequality The variational method The decomposition theorem. Final results for compact manifolds without boundary Manifolds with boundary Differentiability at the boundary Potentials, the decomposition theorem Boundary value problems 286 288 293 295 300 305 309 314

Contents Chapter 8

ix

T h e ^ - N E U M A N N p r o b l e m on s t r o n g l y p s e u d o - c o n v e x m a n i f o l d s 8.1. Introduction 316 8.2. Results. Examples. The analytic embedding theorem 320 8.3. Some important formulas 328 8.4. The HiLBERT space results 333 8.5. The local analysis 337 8.6. The smoothness results 341 Chapter 9 I n t r o d u c t i o n to p a r a m e t r i c I n t e g r a l s ; two dimensional problems 9.1. Introduction. Parametric integrals 349 9.2. A lower semi-continuity theorem 354 9.3. Two dimensional problems; introduction; the conformal mapping of surfaces 362
9.4. The problem of PLATEAU 374

9-5-

The general two-dimensional parametric problem Chapter 10

390

10.1. 10.2. 10.310.4. 10.510.6. 10.7. 10.8. Index

T h e h i g h e r d i m e n s i o n a l PLATEAU p r o b l e m s Introduction V surfaces, their boundaries, and their HAUSDORFF measures . . . . The topological results of ADAMS The minimizing sequence; the minimizing set The local topological disc property The REIFENBERG cone inequality The local differentiability Additional results of FEDERER concerning LEBESGUE v-area

400 407 414 421 439 459 474 48O 494 504

Bibliography

Multiple Integrals in the Calculus of Variations

Chapter 1

Introduction
1.1. Introductory remarks The principal theme of these lectures is ''the existence and differentiability of the solutions of variational problems involving multiple integrals/' I shall discuss the corresponding questions for single integrals only very briefly since these have been adequately discussed in every book on the calculus of variations (see, for instance, AKHIEZER [1],
B L I S S [1], BOLZA [1], CARATHEODORY [2], F U N K [1], P A R S [1]. Moreover,

I shall not discuss applications to engineering, physics, etc., at all, although I shall mention some mathematical applications. In general, I shall consider integrals of the form (1.1.1) I [^,G-) = ff[^,z{x),
G

z{x)]dx

where G is a domain, (1.1.2) X = (A;1, . . ., x"), z = (^1, . . ., z^), dx = dx^ . . . dx\ z{x) is a vector function, V^ denotes its gradient which is the set of functions {zioc}, where z^^ denotes dz'^ldx', and f{x, z, p) [p = {pi}) is generally assumed continuous in all its arguments. The integrals
b

f l/l + {dzldx)^ dx and j y [ ( ^ ) ' + ( ^ , ) ' ]

a G

dxidx^

are familiar examples of integrals of the form (1.1.1.) in which iV = 1 in both cases, i' = 1 in the first case, i^ = 2 in the second case and the corresponding functions / are defined respectively by

f{x,z,p) = yi +p^,

f{^,z,p)

=pi+pi

where we have omitted the superscripts on z and p since N =^ \, The second integral is a special case of the Dirichlet integral which is defined in general by (1.1.3) D{z,G) = j\\Jz\^dx,f{x,z,p)
G

^\p\^

=
i,(x

2[pi)'^.

Another example is the area integral

< < * ' i'-<^^ =Mimj+[m,j+m^p-''"

G Morrey, Multiple Integrals

Introduction

which gives the area of the surface (1.1.5) z^ == z^x'^.x^), {x^,x^)^G, / = 1 , 2 , 3. It is to be noticed t h a t the area integral has the special property t h a t it is invariant under diffeomorphisms (1 1 differentiable mappings, etc.) of the domain G onto other domains. This is the first example of an integral in parametric form. I shall discuss such integrals later (in Chapters 9 and 10). I shall also discuss briefly integrals like t h a t in (1.1.1) but involving derivatives of higher order. And, of course, the variational method has been used in problems which involve a ''functional'' not at all like the integral in (1.1.1); as for example in proving the Riemann mapping theorem where one minimizes sup \f{z) \ among all schlicht functions/(^) defined on the given simply connected region G for which f{zQ) = 0 and f'(zo) = 1 at some given point ZQ in G. We shall consider only problems in which the domain G is fixed; variations in G may be taken care of by transformations of coordinates. We shall usually consider problems involving fixed boundary values; we shall discuss other problems but will not derive the transversality conditions for such problems. 1.2. The plan of the book: notation In this chapter we attempt to present an overall view of the principal theme of the book as stated at the beginning of the preceding section. However, we do not include a discussion of integrals in parametric form; these are discussed at some length in Chapters 9 and 10. The material in this book is not presented in its logical order. A possible logical order would be 1.1 1.5, Chapter 2, Chapter 3, 5.1 5-8, 5.12, Chapter 6, 1.61.9, 4.1, 4.3> 4.4. Then the reader must skip back and forth as required among the material of 1.10, 1.11, 4.2, 5-9, 5-10 and 5-11. Then the remainder of the book may be read substantially in order. Actually, Chapters 7 and 8 could be read immediately after 5-8. We begin by presenting background material including derivations, under restrictive hypotheses, of Euler's equations and the classical necessary conditions of Legendre and Weierstrass. Next, we include a brief and incomplete presentation of the classical so-called ''sufficiency" conditions, including references to other works where a more complete presentation may be found. The second half of this chapter presents a reasonably complete outline of the existence and differentiability theory for the solutions of variational problems. This begins with a brief discussion of the development of the direct methods and of the successively more general classes of "admissible" functions, culminating in the so-called "Sobolev spaces".

1.2. The plan of the book: notation

These are then defined and discussed briefly after which two theorems on lower-semicontinuity are presented. These are not the most general theorems possible b u t are selected for the simplicity of their proofs which, however, assume t h a t the reader is willing to grant the t r u t h of some well-known theorems on the Sobolev spaces. The relevant theorems about these spaces are proved in Chapter 3 and more general lowersemicontinuity and existence theorems are presented in Chapter 4. In Section 1.10 the differentiability results are stated and some preliminary results are proved. In Section 1.11, an outline of the differentiability theory is presented. I t is first shown that the solutions are ''weak solutions" of the Euler equations. The theory of these non-linear equations is reduced to t h a t of linear equations which, initially, m a y have discontinuous coefficients. The theory of these general linear equations is discussed in detail in Chapter 5 However, the higher order differentiability for the solutions of systems of Euler equations required the same methods as are used in studying systems of equations of higher order. Accordingly, we present in Chapter 6 many of the results in the two
recent papers of AGMON, DOUGLIS, and N I R E N B E R G ([1], [2]) concerning

the solutions and weak solutions of such systems. Both the L^^-estimates and the ScHAUDER-type estimates (concerning HOLDER continuity) are presented. We have included sections in both Chapters 5 und 6 proving the analyticity, including analyticity at the boundary, of the solutions of both linear and non-linear analytic elliptic equations and systems; t h e most general ''properly elliptic" systems with "complementing boundary conditions" (see 6.1) are treated. The proof of analyticity in this generality is new. I n Chapter 2 we present well-known facts about harmonic functions and generalized potentials and conclude with proofs of the CALDERON-ZYGMUND inequalities and of the maximum principle for the solutions of second order equations. In Chapters 7 and 8, we present applications of the variational method
to the H O D G E theory of harmonic integrals and to the so-called 5 - N E U -

MANN problem for exterior differential forms on strongly pseudo-convex complex analytic manifolds with boundary. I n Chapter 9, we present a brief discussion of ^^-dimensional parametric problems in general and then discuss the two dimensional Plateau problem in Euclidean space and on a Riemannian manifold. The chapter concludes with the author's
simplified proof of the existence theorem of CESARI [4], DANSKIN, and

SiGALOV [2] for the general two dimensional parametric problem and some incomplete results concerning t h e differentiability of the solutions of such problems. I n Chapter 10, we present the author's simplification of t h e very important resent work of REIFENBERG [1], [2], and [3] concerning t h e higher dimensional PLATEAU problem and the author's extension of these results to varieties on a Riemannian manifold.

Introduction

Notations. For the most part, we use standard notations. G and D will denote domains which are bounded unless otherwise specified. We denote the boundary oiDby dD and its closure by D. We shall often use the notation D C C G to mean t h a t D is compact and D C G. B{xo, R) denotes the ball with center at XQ and radius R. y^ and Fv denote the vmeasure and [v 1)-measure of 5(0,1) and ^5(0,1), respectively. We often denote dB{0,\) by Z. Most of the time (unless otherwise specified) we let Rq be ^-dimensional number space with the usual metric and abbreviate B{0,R) to BR, denote by a the {v 1)-plane x" = 0, and define (1.2.1) R^ ={x\x''>0}, R~={x\x'' < 0} GR^BnORi^, ER^^dBnC^.R^, OR^BROa
G^^BRHR-, Z^ = dBRnRv^

If S is a set in Rq, \S\ denotes its Lebesgue ^-measure; if ;t; is a point, d{x, S) denotes the distance of x from S. We define [a^b] = {x\a'' < x'<,b', ^ = 1 , . . . , v, x^Rv}. In the case of boundary integrals, we often use dx'^ t o denote nxdS where dS is the surface area and nx is the ^-th component of the exterior normal. We say that a function u^ C'^{G) iff (if and only if) u and its partial derivatives of order < ^ are continuous on G and u^ C'^{G) iff u^ C'^{G) and each of its derivatives of order < ^ can be extended to be continuous on G. If 0 < /i < 1, ^ ^ Q(G) (or Q (G)) <^ (i.e. iff) u^ C^{G) (or C^{G)) and all the derivatives of order <n satisfy a HOLDER (LIPSCHITZ ii /J, = 1) condition on each compact subset of G (or on the whole of G as extended). If u^ C^ (G), then A^ {u, G) = sup \x2 xi\-^. \u(x2) u(xi) I for xi and X2^G and Xi ^ x^. A domain G is said to be of class C^ (or Q , 0 < /^ < 1) iff G is bounded and each point PQ oi dG is in a neighborhood n on G which can be mapped by a 1 1 mapping of class C^ (or CJJ), together with its inverse, onto GR U GR for some R in such a way that PQ corresponds to the origin and n U dG corresponds to OR. li U^ ^'^{G)> we denote its derivatives duldx^" by u ,. If ^ ^ C2(G), then V^u denotes the tensor u ,oc^ where oc and /5 run independently from t o r . Likewise V ^ ^ = {^,a/Sy}, etc., and \\/^u\^ 2^\u,ocp\^, etc. If G is also of class C^, then Green's theorem becomes (in our notations) f u ,oc(x) dx = J uUocdS = f udx'^.
G

dG

dG

Sometimes when we wish to consider u as a function of some single x'^, we write x = (%*, x^) and u(x) = u(x^, x^) where x'^ denotes the remaining x^. One dimensional or {v 1)-dimensional integrals are then indicated as might be expected. We often let oc denote a ''multi-index'\ i.e. a vector [oci, . . .,av) in which each oct is a non-negative integer. We

1.3- Very brief historical remarks

then define
d li u

(^,!),c = ^ , f - = (fir-..(i^j^
Using this notation
|a|=m

We shall denote constants by C or Z with or without subscripts. These constants will, perhaps depend on other constants; in this case we may write C = C{h, jbt) if C depends only on h and //, for example. However, even though we may distinguish between different constants in some discussion b y inserting subscripts, there is no guarantee that C2, for example, will always denote the same constant. We sometimes denote the support of u by spt w. We denote by C^iG), Q ( G ) , and C^, {G) the sets of functions in C(G), C^(G), or CJJ(G), respectively, which have support in G (i.e. which vanish on and near dG). But it is handy to say t h a t u has support in GR \J GR <^ U vanishes on and near ER (see 1.2.1); we allow u{x) to be 9^ 0 on GR. 1.3. Very brief historical remarks Problems in the calculus of variations which involve only single integrals (1; = 1) have been discussed at least since the time of the B E R NOULLI'S. Although there was some early consideration of double integrals, it was RiEMANN who aroused great interest in them by proving many interesting results in function theory by assuming DIRICHLET'S principle which may be stated as follows: There is a unique function which minimizes the DIRICHLET integral among all functions of class C^ on a domain G and continuous on G which takes on given values on the boundary dG and, moreover, that function is harmonic on G. RIEMANN'S work was criticized on the grounds that just because the integral was bounded below among the competing functions it didn't follow that the greatest lower bound was taken on in the class of competing functions. In fact an example was given of a (1-dimensional) integral of the type (1.1.1) for which there is no minimizing function and another was given of continuous boundary values on the unit circle such that D[z, G) = + 0 0 for every z as above having those boundary values. The first example is the integral (see COURANT [3])
1

(I.3.I)

I{z,G) = | [ l + ( g ) ' J " ^ ^ . G = (0,1),

0
^(0) = 0 and ^(1) =- 1.

the admissible functions z being those ^ C^ on [0,1] with

Introduction

Obviously I[z, G) > 1 for every such z, I{z,G) has no upper bound and if we define
Zr{x) = | _ ^

+ [1 + 3 (A; r)^l[\ r)2]i/2,

r<.x<.\

we see t h a t I{zr, G) -> 1 as r -> 1-. The second example is the following (see COURANT [3]): I t is now known t h a t Dirichlet^s principle holds for a circle and that each function harmonic on the unit circle has the form

o o
(1.3.2) w{r,0) = ^ + 2 r'^{ancosn0 + bnSinnO), (Un, bn const),

in polar coordinates and that the Dirichlet integral is (1.3-3) D{w,G)=jzZn(al + bl)
n=l

provided this sum converges. But if we define a^ = ^-2 if ^ == ^1 ^ an = bn = 0 otherwise, we see t h a t the series in (1.3.2) converges uniformly b u t that in (1.3-3) reduces to

^2/ A 4
which diverges. DIRICHLET'S principle was established rigorously in certain important cases by HILBERT, LEBESGUE [2] and others shorly after I9OO. That was the beginning of the so-called ''direct methods" of the calculus of variations of which we shall say more later. There was renewed interest in one dimensional problems with the advent of the MORSE theory of the critical points of functional in which M. MORSE generalized his theory of critical points of functions defined on finite-dimensional manifolds [1] to certain functionals defined on infinite-dimensional spaces [2], [3]. H e was able t o obtain the MORSE inequalities between the numbers of possibly ''unstable'' (i.e. critical b u t not minimizing) geodesies (and unstable minmal surfaces) having various indices (see also MORSE and TOMPKINS, [1] [4]). Except for the latter (which could be reduced t o t h e case of curves), MORSE'S theory was applied mainly to one-dimensional problems. However, within the
last two years, SMALE a n d PALAIS a n d SMALE have found a modification

of MORSE'S theory which is applicable t o a wide class of multiple integral problems. Variational methods are beginning to be used in differential geometry.
For example, t h e author a n d Eells (see MORREY and E E L S , MORREY,

1.4. The EuLER equations

[11] and Chapter 7) developed the HODGE theory ([1], [2]) b y variational
methods (HODGE'S original idea [1]). HORMANDER [2], K O H N [1], S P E N CER (KoHN and S P E N C E R ) , and the author (MORREY [19], [20]) have

applied variational techniques to t h e study of the 5-Neumann problem for exterior differential forms on complex analytic manifolds (see Chapter 8; the author encountered this problem in his work on the analytic embedding of real-analytic manifolds (MORREY [13])- Very recently, EELLS and SAMPSON have proved the existence of "harmonic" mappings (i.e. mappings which minimize an intrinsic Dirichlet integral) from one compact manifold into a manifold having negative curvature. Since the inf. of this integral is zero if the dimension of the compact manifold > 2 , they found it necessary to use a gradient line method which led to a non-linear system of parabolic equations which they then solved; the curvature restriction was essential in their work. 1.4. The Euler equations After a number of special problems had been solved, E U L E R deduced in 1744 the first general necessary condition, now known as EULER'S equation, which must be satisfied b y a minimizing or maximizing arc. His derivation, given for the case N = v = 1, proceeds as follows: Suppose t h a t the function z is of class C^ on [a, b] (= G) minimizes (for example) the integral I{z, G) among all similar functions having the same values a t a and h. Then, if C is any function of class C^ on [^, h] which vanishes a t a and h, the function 2: + AC is, for every A, of class C^ on [a, h] and has the same values as ^ at a and h. Thus, if we define

6
(1.4.1) w[X)=I[z+Xl:,G)=jf[x,z{x)
a

+Xl:{x)^z'[x)+XC[x)]dx

(p must take on its minimum for A = 0. If we assume t h a t / i s of class C^ in its arguments, we find b y differentiating (I.4.I) and setting X = 0 t h a t

h
(1.4.2) /{C'(^) 'U[x,z{x),z'{x)] + (:{x)f,[x,z{x),z'{x)]]dx =0

(fv =
The integral in (1.4.2) is called the first variation of the integral / ; it is supposed to vanish for every f of class C^ on [a, 6] which vanishes at a and h. If we now assume that f and z are of class C^ on [a, H] (EULER had no compunctions about this) we can integrate (1.4.2) b y parts to obtain

h
(1.4.3) jC{x)-{fz-~h]dx = 0, U=U[x,z[x),\/z{x)],etc.

Introduction

Since (1.4-3) holds for all C as above, it follows that the equation

(1.4.4)

y,=h

must hold. This is Euler's equation for the integral / in this simple case. If we write out (1.4.4) in full, we obtain
(1.4.5) fvv'^" +fvz^' +fpx=fz

which shows that Euler's equation is non-linear and of the second order. I t is, however, linear in z"; equations which are linear in the derivatives of highest order are frequently called q^iasi-linear. The equation evidently becomes singular whenever/^^ = 0. Hence regular variational problems are those for which fpp never vanishes; in that case, it is assumed that fpp > 0 which turns out to make minimum problems more natural than maximum problems. It is clear t h a t this derivation generalizes to the most general integral (1.1.1) provided t h a t / a n d the minimizing (or maximizing, etc.) function z is of class C2 on the closed domain G which has a sufficiently smooth boundary. Then, if z minimizes / among all (vector) functions of class C^ with the same boundary values and C is any such vector which vanishes on the boundary or G, it follows t h a t 2: + AC is a "competing'' or ''admissible" function for each A so t h a t if 99 is defined by (1.4.6) (p{^) =I{z + U,G) then (p'{0) = 0. This leads to the condition that (1-4.7)
Q i=-l l a = l

ll\^C:Jpi

+ C'fz^dx=0
J

for all C as indicated. The integral in (1.4.7) is the first variation of the general integral (1.1.1). Integrating (1.4-7) by parts leads to

Since this is zero for all vectors C, it follows t h a t

(^4-8)

2i^M=f^'' '-''

.N

which is a quasi-linear system of partial differential equations of the second order. In the case A^ = 1, it reduces to

(1.4.9)

"-i

1.4. The EuLER equations

The equation (1.4-9) is evidently singular whenever the quadratic form (1-4.10) 2'^a2.,(^,^>^)^a^^*

in A is degenerate. We notice from (1.4-5) t h a t ii N = v = \ and / depends only on p and the problem is regular, then Euler's equation reduces to / ' = 0. In general, if / depends only onp(== pi), Euler's equation has the form

and every linear vector function is a solution. In particular, if A^ = 1 a n d / = 1^12, Euler's equation is just Laplaces equation
a

In c a s e / = (1 + |_/>|2)i/4 as in the first example in 1.3, we see t h a t

4fvP = {2-p^){^

+^2)-7/4

which is not always > 0. On the other hand fpp > 0 if | ^ | < ]/2 so classical results which we shall discuss later (see 1.6) show that the linear function z{x) = x minimizes the integral among all arcs having \z-(x)\<^. We now revert to equation (1.4-9)- If we take, for instance, A^ = 1, V =^ 2,f = pl pl, then (1.4-9) becomes

which is of hyperbolic type. Moreover, the integral (1.1.1) with this / obviously has no minimum or maximum, whatever boundary values are given for z. Anyhow, it is well known that boundary value problems are not natural for equations of hyperbolic type, li v "> 2 s, greater variety of types may occur, depending on the signature of the quadratic form (1.4.10). A similar objection occurs in all cases except those in which the form (1.4-10) is positive definite or negative definite; we shall restrict ourselves to the case where it is positive definite. In this case Euler's equation is of elliptic type. The choice of this condition o n / is re-enforced by analogy with the case r = 1; in that case fpp > 0 implies the convexity (see 1.8) of f as a function of p for each {x, z) and the non-negative definiteness of the form (1.4-10) is equivalent to the convexity o f / a s a function of ^ i , . . .,pv for each set [x'^, . . ., x"^, z). Our choice is re-enforced further by the classical derivation given in the next section.
* Greek indices are summed from 1 to i; and Latin indices are summed from 1 to N. Hereafter we shall usually employ the summation convention in which repeated indices are summed and summation signs omitted.

10

Introduction

1.5. Other classical necessary conditions Suppose that / is of class C^ in its arguments, t h a t A^ = 1, that z is of class Ci on the closure of G, and t h a t z minimizes I(z, G) among all functions Z of the same class which coincide with z on the boundary and are sufficiently close to z in the C^ norm, i.e. \Z[x) z[x)\ < ( 3 , |VZ(%) - Vz[x)\ <(5, xon G. In classical terminology, we say t h a t z furnishes a le^eak relative minimum to I[z, G). We shall show that this implies the non-negative definiteness of the form (1.4.10) when z = z{x) and p = \/ z(x). We note that our hypotheses imply t h a t for each C of the type above, vanishing on the boundary, the function 99(A), defined by (1.4-6) is of class C^ for |A| < Ao(>0) and has a relative minimum at 2 =: 0. This implies that ^"(0) = f \Za-^{x)
(1.5.1) G ^^'^

C,C,^ + 2 2:b-{x) a , . + c{x) m

^ J

dx^O

a^^[x)=^f^^j,^[x,z{x),Vz{x)],

b^=fj,^^,

c=f^^,

for all ^ as described. The integral in (1.5.1) is called the second variation of the integral (1.1.1). By approximations, it follows that (1.5-1) holds for all LiPSCHiTZ functions C which vanish on the boundary. Now let us select a point xo(^G and a unit vector A, and let us choose new coordinates y related to :^ by a transformation

(1.5.2)

yy = i;dy{x^-x^),
oc

x^-x^^^z^ly''^ ^oc^di
y

where d is a. constant orthogonal matrix so t h a t A is the unit vector in the yi direction, and define co{y)=-a^{y)], ^ ^ ^^ 'hy{y) = b^[x{y)]dy, Then if G' denotes the image of G, (p"[0) = / [ ^ ^ V*5(y)cL>,y ca,5 + 2'hy co'(D,y + 'c(jo^]dy > 0. Now, choose 0 <C h <. H so small t h a t the support of a> G G\ where 'ay'{y)^a^^[x(y)-]dld$, 'c{y) =^-c[x{y)].

lo , otherwise J r2 :== (^2)2 + . . . + (y')2. Then if we divide 99''(0) by the measure of the support of co and then let h and H -> 0 so that hlH -> 0, we conclude t h a t ^ ^^ ^-^ -^ ^ 'all (0) = ^a^ (^^) ^1 ^1 ^ a''^{xo) > a A^ > 0 ^ which is the stated result. This is called the Legendre condition.

1.5.4)

co(y^,...,y'')

=y

f ( A - | y i | ) ( 1 - r / i f ) , if | y i | < / ^ ,

'-^ '^^

' ^

0<r</f\

1.5. Other classical necessary conditions

11

If we repeat this derivation for the case of the general integral (1.1.1), a s s u m i n g / ^ C^ of course, we obtain

^"(0) =I{ZI<CUC$

+ 2 2-bf,C'C\ + CijC'en dx > 0

c^ J- = /^i 2?.

< f (^) == hi 4 [^> ^ (^), V ^ [x)], 6?,. = /g^ pi^, co^(yi, . . ., y") = f^a)(yi, . . ., y"), {i=

Making the change of variables (1.5-2) and (1.5-3) and setting i, . . .,N) where f is an arbitrary constant vector and co is defined by (1.5-4), and letting h and ^ -> 0 as above, we obtain (^5.5) 2 fp^ paxo.z{xo),Vz(xo)]Ao:Apii&'^0 for a l U , f ,

which is known as the Legendre-Hadamard condition (HADAMARD [1]). In this case, we say that the integral (1.1.1) or the i n t e g r a n d / i s regular if the inequality holds in (1.5.5) for all A 9^ 0 and f 9^ 0. It turns out that the system (1.4-8) of Euler's equations is strongly elliptic in the sense
defined by NIRENBERG [2].

Let us suppose, now, t h a t / ^ C^ everywhere and that ^ ^ C^ on G and minimizes I[z, G), as given by (1.1.1), among all such functions with the same boundary values. A simple approximation argument shows t h a t z minimizes / among all LIPSCHITZ functions with the same boundary values. Let us choose XQ^G and a unit vector X, and let us introduce the y coordinates as in (1.5-2) and let us define (using part of the notation of (1.5.4)

I
G

(y^ + h)(p(rh-^l^)

h<y^<0

0<.r<h^l^,

/^i/2(/ji/2 _ yi) .(p{rh-^l^),

(1.5.6) where 99$ C^ on [0,1] with (p{0) = 1 and 99(^) = 0 for Q near 1. Since the first variation vanishes, we have / [/(^, ^ + C/., V ^ + V u ) - fix, z,Vz)~ Cifz^ - ClJpl] dx->0, (1.5.7) fz^ ^ fz^(x, ^, V ^), fpi = fpi{x, z,Vz). We notice first t h a t the integrand is 0(h) (since f^ and \7 Ch are both small) for x^R^ where 0 < yi < ^1^2^ 0 < r < Ai/2, g y setting yi = hrj^,r = Ai/2 ^ in Rl{h < yi < 0, 0 < r < h'^^^), dividing by ^('+i)/2 and letting A -> 0, we obtain
/
Q^l

0 < y i < Ai/2,

otherwise

0 < ^ < ^

/?i/2,

| [ / ( ^ 0 , ^ 0 , i 5 ) ^ a + ' ^ ^ a l * X ^ ) ] -f{X0,Z0,p0)

" > 0.

Z^-<^ ^^ (p{Q)fpi{^o,zo,po)\drj[ i. a J

12

Introduction

We may now choose a sequence {cpn} so q^niQ) -> 1 boundedly. This leads to (1.5.8) f{xo,zo,pio+^-^') -f{^o,zo,po) -Ao.i^fpi{xo,zo,po) > 0

which is t h e Weierstrass condition (see GRAVES). I n case N = 1, (1.5-8) yields the following more familiar form of this condition:

E{x,z,Vz,P)
(1.5.9)

==f(x,z,P)

f{x,z,Vz)

[Poc

z,^)f^^[x,z,\Jz)^0

for all P and all x. The function E(x, z,p,P) here defined is known as the Weierstrass E-function. HESTENES and MACSHANE studied these general integrals in cases
where v 2. H E S T E N E S and E. H O L D E R studied the second variation of

these integrals. DEDECKER studied the first variation of very general problems on manifolds. 1.6. Classical sufficient conditions A detailed account of classical and recent work in this field is given in the recent book by FUNK, pp. 410433) where other references are given. I shall give only a brief introduction to this subject. It is clear that the positiveness of the second variation along a function z guarantees that z furnishes a relative minimum to I[Z, G) among all Z ( = z on dG) in any finite dimensional space. However, if TV = 1, a great deal more can be concluded, namely t h a t z furnishes a strong relative minimum to / , i.e. minimizes I[Z, G) among all Z^ C^{G-) with Z = z on dG ior which \Z{x) z(x)\ <. d for some ^ > 0 regardless of the values of the derivatives. WEIERSTRASS was the first to prove such a theorem b u t his proof was greatly simplified by the use of HILBERT'S invariant integral. Of course, the original proof was for the case N = v = 1; we present briefly an extension to the case N = \, v arbitrary. Suppose G is of class Q , z^ C^ (G), a n d / and fp are of class C^ in their arguments, 0 < ju < i (see 1.2), and suppose that the second variation, as defined in (1.5.1), > 0 for each C$ Q(G) (compact support). By a straightforward approximation, it follows t h a t the second variation is defined for all f ^ Hl^ (G) (see 1.8). If we call the integral (I.5.I) h{z\^]G) we see from the theorems of 1.8 below that 7-2 is lower-semicontinuous with respect to weak convergence in H\Q {G) . Moreover, from the assumed positive definiteness of the form (1.4.10), it follows from the continuity of the a'^P {x) (they ^ Q {G) in fact) t h a t there exist w i > 0 and M l such t h a t (1.6.1) ^^ {x)?ioc?i^ > mi |A|2, J [^'"^ W]^ ^ ^ 1

1.6. Classical sufficient conditions

13

Then, from the SCHWARZ and CAUCHY inequahties, we conclude t h a t there is a i^ such t h a t (1.6.2) h{z\l:\G)>'^ j\\/^\^dx
G

~ K j l:^dx.
G

Since weak convergence in HIQ{G) imphes strong convergence in L2{G) (RELLICH'S theorem, Theorem 3-4.4), it follows t h a t there is a fo in
H\Q{G)
G

(actually C^ (G)) which minimizes 12 among all I^^H\Q[G)

for

which f C^ dx = 1. Since we have assumed /2 > 0 for every f 9^ 0, it follows t h a t (1.6.3) l2(z;C]G) ^AijC^dx^Ai
G

> 0.

From the theory of 5-25.6, it follows t h a t there is a unique solution f of Jacohi's equation (1.6.4) ^^^i^("^f - ) + iK -o)C = 0

with given smooth boundary values. I t is to be noted t h a t JACOBI'S equation is just the Euler equation (z fixed) corresponding to 12. I t is also the equation of variation of the Euler equation for the original / , i.e. (1.6.5) ^ f = ^ { ^Jpc. [^.^ + eC, Vz + QVCl-fz [same] }^^^.

It follows from Theorems 6.8.5 and 6.S.6 t h a t there is a unique solution of the Euler equation for all sufficiently near (in C^ (dG)) boundary values, in particular for the boundary values z -{- Q, and t h a t z = z (Q) satisfies an ordinary differential equation (1.6.6) p^^Fiz)

in theBanach space (C^^ [G)), where F{z) denotes the solution f of Jacobi's equation (1.6.4) with z = Z{Q) for which C = 1 on dG. We shall show below that this solution f cannot vanish on G for Q sufficiently small; it is sufficient to do this for ^ = 0, when Z{Q) = our solution z, on account of the continuity. So, let fi be this solution. If Ci(^) < 0 anywhere, then the set where this holds is an open set D and fi = 0 on ^ D ( C G). Since Ci is a solution on D, I(Ci, D) ^0 since Ci is minimizing on D. {D m a y not be smooth, b u t see Chapters 3 5)- But if we set ^ = ^i on D and C = ^, otherwise, f ^ HIQ {G) SO (1.6.3) holds and we must have C = 0. Hence Ci{^) > 0 everywhere. Now, suppose Ci(^o) = 0. From Theorem 6.8.7, it follows t h a t we m a y choose R so small that B(xo, R) G G and there is a non-

14

Introduction

vanishing solution co of (1.6.4) on B(xo,R). t h a t V satisfies the equation

Letting fi = coy, we see

and v(xo) = 0. But from the maximum principle as proved by E. H O P F [1] (see 2.8), it follows t h a t v^ cannot have a minimum interior t o G. Accordingly Ci ^ 0 anywhere in G. Therefore it is possible to embed our solution z in s. field of extremals. That is, there is a 1 parameter family Z{X,Q) of solutions of Euler's equation where Z{x, 0) == z{x), our given solution. Z(X,Q)^ C^{GX [^O>^O]) and ^ C^ (G) as a function of x for each Q with ^^ > 0. Consequently there are functions Po: {x, z) on the set F, where (\.6.7) (1.6.8) F: x^ G, Z{x, ~QO) < ^ < Z{x, go), Z,oc{x, Q) = Pa [x, Z{x, Q)]. which act as slope-functions for the field, i.e. By virtue of the facts t h a t Z[x, q) satisfies Euler's equation for each Q, t h a t (1.6.8) holds, and t h a t if {x, z) $ F, then z = Z(x, Q) for a unique Q on [^0, ^o], we find t h a t ^ ' ' ^ f,=fz\x^ z, P{x, z)l etc., (^, z) ^ F. Let us define /*(^, G) = ff*{x,z,Vz)dx, (1.6.10) G' / * {x. z,p) =f^, [X, z, P [x, z)]' \_p. - P [x, z)\ + / [ ^ , z, P {X, z)\. We observe that
Sl = [P- Poc {X, Z)] ' {fp^^ + / p ^ p^ Ppz } Poczfvu

^^''^^^
(1.6.12)

+fz+fv,Po.z; f;^=fvJ^>z,P(x,z)].
/ : [X, z (x), Vz(x)] - ^j;^ [X, z {x), Vz{x)] = 0.

Thus, ii z^ Gi(G) and {x, z{x)) ^F for x^G, we see that Accordingly the integral I*(z, G) has the same value for all such z which have the same boundary values. Moreover, if z{x) = Z[x, q) for some q, then ^ ' I*{z,G)=I{z.G).

This integral 7 * (^, G) is known as Hilbert's invariant integral. Therefore, if z^ Ci(G) and [x, z{x)) ^ F for all x^ G, and z{x) = zo[x) on dG, then (1.6.14) I{z, G) - I(zo, G) = I(z, G) - I*{zo, G) = I(z, G) - I*{z, G) = f E[x, z{x), P{x, z{x)}, Vz{x)] dx

1.7. The direct methods

15

where E(x, z, P, p) is the Weierstrass ^-function defined in (1.5-9). Thus ZQ minimizes I{z, G) among all such z, and hence furnishes a strong relative minimum to / , provided that (1.6.15) E[x, z, P{x, z),p] > 0, [x, z)^r,p arbitrary. This same proof shows t h a t if (1.6.15) holds for all {x, z, p) in some domain Ry where all the (x, z) involved ^ F, then I(zo, G) < I{z, G) for all z for which [x, z[x), \7 z{x)'] ^Rfor all x^G. In the cases v = \, the Jacobi condition is frequently stated in terms of ''conjugate points". A corresponding condition for )^ > 1 is t h a t the JACOBI equation has no non-zero solution which vanishes on the boundary dD oi any sub-domain D G G; D m a y coincide with G or may not be smooth, in which case we say u vanishes ondD <^u^ H\Q [D). The most interesting condition is that there exist a non-vanishing solution a> on G; we have seen above t h a t this is implied by the positivity of the second variation. If we then set ^ = cou, where u = 0 on dG, then the reader may easily verify t h a t I2{^',^',G) = J[w^a'^u,ocU,p
G

u^a)L(o]dx

> 0

for all u(^ H\Q (G), since Leo = 0. In cases where v > 1 and A^ > 1 it is still true (if we continue to assume the same differentiabihty for G , / , and z) t h a t (1.6.2) holds with I2 defined by (1.5.1') even in the general regular case where (1.5-5) holds with the inequality for A 7^ 0 and f 7^ 0. This is proved in 5.2. So it is still true t h a t if /2 > 0 for all f, the Euler equations have a unique solution for sufficiently nearby boundary values. I t is more difficult (but possible) to show t h a t there is an A^-parameter field of extremals and then it turns out t h a t such a field does not lead so easily to an invariant integral. By allowing slope functions P* {x, z) which are not "integrable" (i.e. there may not be zs such that z^^ = P^ {x, z)), W E Y L [1] (see also D E BONDER) developed a comparatively simple field theory and showed the existence of his types of fields under certain conditions. His theory is succesful i f / i s convex in all the p'l. To treat more general
cases, CARATHEODORY [1], BOERNER, and L E PAGE have introduced more

general field theories. The latter two noticed t h a t exterior differential forms were a natural tool to use in forming the analog of Hilbert's invariant integral. However, the sufficient conditions developed so far are rather far from the necessary conditions and many questions remain to be answered. 1.7. The direct methods The necessary and sufficient conditions which we have just discussed have presupposed the existence and differentiability of an ex-

16

Introduction

tremal. I n the cases v = i, this was often proved using the existence theorems for ordinary differential equations. However, until recently, corresponding theorems for partial differential equations were not available so the direct methods were developed t o handle this problem and to obtain results in the large for one dimensional problems.
As has already been said, H I L B E R T [1] and L E B E S G U E [2] h a d solved

the Dirichlet problem b y essentially direct methods. These methods were exploited and popularized b y TONELLI in a series of papers and a book ([1], [2], [3], [4], [5], [7], [8]), and have been and still are being used b y many others. The idea of the direct methods is to show (i) t h a t the integral to be minimized is lower-semicontinuous with respect t o some kind of convergence, (ii) t h a t it is bounded below in the class of ''admissible functions," and (iii) that there is a minimizing sequence (i.e. a sequence {zn] of admissible functions for which I[zny G) tends to its infimum in t h e class) which converges in the sense required to some admissible function. Tonelli applied these methods t o many single integral problems and some double integral problems. I n doing this he found it expedient to use uniform convergence (at least on interior domains) and to allow absolutely continuous functions (satisfying the given boundary conditions) as admissible for one dimensional problems; and he defined what he called absolutely continuous functions of two variables ([6]) to handle certain double integral problems (see the next section). I n the double integral problems (iV == 1, r = 2), he found it expedient to require t h a t f[x, z, p) satisfy conditions such as (1.7.1) (1.7.2) m\p\^ K <f{x, z,p), k>2, m>0, where f{x, z,p)'>0 and f{x, ^, 0) = 0 if k = 2,

in order to obtain equicontinuous minimizing sequences. However, Tonelli was not able t o get a general theorem to cover the case where / satisfies (1.7-1) with 1 < ^ < 2. Moreover, if one considers integrals in which r > 2, one soon finds t h a t one would have to require ^ > i^ in order to ensure that the functions in any minimizing sequence would be equicontinuous, at least on interior domains (see Theorem 3.5-2). To see this, one needs only to notice t h a t the functions loglog(1 + 1/|:v|), 1/|A;|^ are limits of C^ functions Zn in which j\SJZnYdx djnd j \\/Zn\^ dx for k<vl{h+\) are uniformly bounded over the unit ball G. I n the ''borderline case" k = V, it is possible, in case / satisfies the supplementary condition (1.7.2), to replace an arbitrary minimizing sequence of continuous functions b y a minimizing sequence each member of which is monotone in the 0<|A;|<1

1.7. The direct methods

17

sense of Lebesgue (i.e. takes on its max. and min. values on the boundary of each compact sub-domain); the new sequence is equicontinuous on interior domains (see e 4.3)However, even this Lebesgue smoothing process does not work in general for 1 < ^ < r. In order to get a more complete existence theory,
the writer and CALKIN (MORREY [5], [6], [7]) found it expedient to allow

as admissible, functions which are still more general than Tonelli's functions and to allow correspondingly more general types of convergence. The new spaces of functions can now be identified with the Banach spaces HJ (G) (see the next section) (or the Sobolev spaces Wl{G)) which have been and still are being used b y many writers in many different connections (see 1.8). In this way, the writer was able to obtain very general existence theorems. Unfortunately the solution shown to exist was known only to be in one of these general spaces and hence wasn't even known to be continuous, let alone of class C^! So these existence theorems in themselves have only minor interest. However, at the same time,i the writer was able to show in the case v = 2 {N arbitrary) that these very general solutions were, in fact, of class C^ after all provided that / satisfied the conditions (1.10.8) below with ^ = r = 2. A greatly simplified presentation of this old work is to be found in the author's paper [15]; recent developments have permitted further simplifications and extensions which we shall discuss later. In general, it is still not known that the solutions are continuous if / satisfies (1-7.1) with 1 < ^ < r. In the case v = 2, N \ TONELLI [8] showed t h a t the solutions in this case are continuous if / is such that there is a unique minimizing function in t h e small. More recently, SiGALOV ([1], [3]) showed that the solution surfaces always posess conformal maps (possibly with vertical segments) in the case v = 2, N = i. In the case v == 2, N = \, f == f{p) it was proved a long time ago by A. HAAR (see also RADO [2]) that there is a unique minimizing function z which is defined on a strictly convex domain G and which satisfies a Lipschitz condition with constant L, provided that any linear function which coincides with the given boundary values a t three different points on the boundary has slope < Z . This result has recently been generalized
(not quite completely) b y GILBARG and STAMPACCHIA [3]. The author

has completed the extension of HAAR'S results and has extended those
1 This work was completed during the year 1937 38 and the author lectured on it in the seminar of Marston Morse during the spring of 1938 and also reported on this work in an invited lecture to the American Mathematical Society a t its meeting in Pasadena, California, on December 2, 1939 [6]. The necessary theorems about the H^ spaces (called ^ ^ at that time) (see 1.8) were published in the papers referred to above by CALKIN and the author [5]. The remainder of the work was first published in a paper (MORREY (7]) which was released in December 1943," the manuscript had been approved for publication in 1939Morrey, Multiple Integrals 2

18

Introduction

of STAMPACCHIA and GILBARG. These results are presented and proved in 4-2. The results include GILBARG'S existence theorem for equations of the form (I.IO.I3) with N = i, Bi = 0, A"- = A"" {p). The advantages of these theorems are that one can restrict one's self to LIPSCHITZ functions z and no assumption has to be made about how f{x, z, p) behaves as 1^1 - > o o . The convexity assumption for all (x, z,p) is suggested by the conditions in 1.41.6. In the existence theorems mentioned above, the author considered integrals of the form (1.1.1) in which v and N are arbitrary but in w h i c h / is convex in all the pl^; with this convexity assumption, no difficulties were introduced in the proofs by allowing AT > 1. The results have been extended and the old proofs greatly simplified by SERRIN ([1], [2]); we shall present (in 1.8) a simple lower-semicontinuity proof based on some of his ideas and on some ideas of TONELLI. However, for iV > 1, the proper condition would be to assume t h a t (1.5-8) and/or (1.5-5) held for all (x, z, p, A, f). The author has studied these general integrals ([9]) and found that if / satisfies the conditions mV^ - K <.f{x, z,p) <MV^, \M,\M<MVK m>0, \fp{x, z,p)\<. MF^-i, k > 1 F = (1 + 1^12)1/2

then a necessary and sufficient condition that I[z, G) be lower semicontinuous on the space H\ (G) with respect to uniform convergence is t h a t / be quasi-convex as a function oi p. A function/(^), p = {pl^} is quasiconvex if and only if it is continuous and jflPo
G

+ VC()] dx >f{po)

\G\,

C Q ( G ) ;

t h a t is, linear vectors give the absolute minimum to I{z, G) among all z with such boundary values (note that linear functions always satisfy EULER'S equation if f^C^). A necessary condition for quasi-convexity is just (1.5.5)- The author showed t h a t (1.5-5) is sufficient for quasi-convexity if f{p) is of one of the two following forms Wm = <fP'pf K f ) const. N = v+i {2)f{p)=F{Di,...,D^^,).

where F is homogeneous of degree 1 in the Di and Di is the determinant of the submatrix obtained by omitting the ^-th column of the pi_ matrix; or if for each p, there exist alternating constant tensors Af, Aff, . . ., such t h a t

fiP + 71) > / ( / ) ) + AtK

+ ^IKA

+ + AX\\\\7il\..

.7^:;;

for all 71. Under some additional conditions the integral is lower-semicontinuous with respect to weak convergence in H\ (G), We discuss these integrals in 4.4- Recently Norman MEYERS has extended the author's

1.8. Lower semicontinuity

19

results to higher order integrals and has improved them somewhat. The proofs are very similar to those of the author for the first order integrals and we shall not present them. Finally, it should be pointed out that the r-dimensional area integral in the non-parametric case with A^ > 1 is not convex in all the ^ j , b u t T is a regular integrand in the general sense; for v = N = 2, the integral is

I(z,G) = JJl/l + (4)' + (4)' + (4)^ + (4)' + (TlJ^J'i^dy.

G

f{x,z,p) = [1 + iPl)^ + (Pl)^ + iPl)^ + (Pl)^ + iPlPl - PlPiW^.

The integrand of a necessarily degenerate (1.5.5) and (1.5.8) with speak about parametric parametric problem is never regular since it is b u t a large class of such integrands do satisfy the equality allowed (whenever/^ C^). We shall problems in Chapter 9-

1.8. Lower semicontinuity We begin with a brief discussion of the spaces of admissible functions which we shall use; a more extended discussion including complete proofs is given in Chapter 3. I t is convenient to call these spaces SOBOLEV spaces; in addition to the brevity of this designation, it is appropriate since he proved some important results concerning these functions [1] and popularized them in his book [2]. However, he was by no means the first person to use these functions. Beppo LEVI was probably the first to use (in (I906) admissible functions which required the use of the Lebesgue integral to express I{z, G); his functions (of two variables) were continuous, absolutely continuous in each variable for almost all values of the other, and their first partial derivatives were in L2. In discussing the area of surfaces, TONELLI [6] introduced his notion of absolutely continuous functions (ACT) in 1926. His definition was identical with that of Levi except that the partial derivatives were required only to be in Li; he used these functions to discuss the double integral problems mentioned above. But meanwhile in I92O, G. C. EVANS {W> [2]) had encountered more general functions, essentially those we now use, in his study of potential theory. RELLICH proved the compactness in L2 of bounded sets in HI. SOBOLEV proved his well known results on theLp-properties of these functions in 1938 [1]. I n 1940, J . W. CALKIN and the author ([5]) proved many of the fundamental properties of these functions stated below. Since the war ARONSZAJN and SMITH have studied these functions in great detail [1]. No doubt many others have studied these functions. Recently these functions have been used by many people
in m a n y different connections (see, for instance, D E N Y , F R I E D R I C H S [1], [2], [3], FuBiNi, J O H N [2], L A X , MORREY [1], [10], MORREY and E E L L S , 2*

20

Introduction

NiKODYM, SHIFFMAN, SIGALOV ([2]); their use is now standard in partial

differential equations (see FRIEDMAN [2], HORMANDER [1], LIONS [1]).

Definition 1.8.1. We say that a function z is of class Hl[G), r > 1, if and only if z is of class Lr[G) and there exist functions pi, . . ., pv, also of class Lr{G), such that (1.8.1) jg[x)Pa,[x)dx=
G

-lg,oc{x)z{x)dx,
G

g^Cl{G),oc=

\,...,v.

Remarks. I t is clear that the functions px are uniquely determined up to null functions and that if z is of class Hj{G) and z'^{x) = z{x) almost everywhere, then ^* is of class Hl{G) and the same functions po^ will do for 2:*. Definition 1.8.2. As in the case of the L^-spaces, the elements of the space Hl[G) are classes of equivalent functions of class Hj(G). We denote the classes of equivalent functions px by z^ and call them the distribution derivatives of the element z. An element z^ space H'^[G) if and only if z and its distribution derivatives u p to order m 1 are successively seen to ^ Hj (G). Remark. Naturally we may regard an element z in H^ (G) as a distribution and then the distribution corresponding to z^ would be the derivative of z in the distribution sense. Definition 1.8.3. 99 is a Friedrichs mollifier if and only if 99$ ^^{Rv), q)(x) > 0, spt 99 (i.e. the support of 99) C ^(0,1), and

f (p(x) dx = \.
B{0,1)

If ^ is locally summable on an open set G, we define its cp-mollified function UQ b y u,{x)= I u{i)(pf{^-x)di, X^G, = {X^G\B(X,Q)CG},
B(X,Q)

The following theorems are almost evident and are proved in Chapter 3 (in fact, Theorem B is evident): Theorem A. The space H^ (G) is a Banach space binder the norm
lUlli _ _

2 k \\v,G \r\\v,G^ ^ ' l\r ,''\\v,G

a=l

IMlo

4- V l l ^ !lo

Theorem B. (a) If u^Lp{G) and UQ denotes its (p-mollified then UQ ^ C'^(Gg) and UQ ->u in Lp(D) for each D CCG.
(b) If u^ ^KQy ^^^^ ^e ,a(^) = {u ,OC)Q(X) for X^GQ SO that UQ ->u

and UQ ^oc ->u ^oc in Lp(D) for each D C C G. (c) The convergence in {a) and {b) holds for almost all x. The lower semicontinuity theorems in this section depend on some well-known theorems on convex functions which we now define:

1.8. Lower semicontinuity

21

Definition 1.8.4. A set S in a linear space is said to be convex ^ the segment P1P2C. S whenever P i and P2^S. A function 99 is said to be convex on the convex set S < > = ^[(1 - A) f i + U2] < (1 - >^) (fih) + M^2), fi, f2 5 , 0 < A < 1. Remark. Evidently cp is convex on the convex set 5 <^ the set of points (f, f) for which f ^ S and C > 99(1) is convex. We now state the characteristic property of convex functions: Lemma 1.8.1. A necessary and sufficient condition that cp he convex on the open convex set S G Rp is that for each ^ in S there exists a linear function ap ^^ -\- b such that (1.8.2) (1.8.3) cp{0=apC^ + b, (p(^)^apS^ + b, ^^S. | , f 6 5. If (p is of class C^ on S, this condition is equivalent to (^,l)-95(|)-9,(a-(|2>_fj.)9,^(^)>0, If Cp is of class C2 on S, this condition is equivalent to for all ^ and all rj. Remark. The function E(^, f) in (1.8.3) is seen to be the WEIERSTRASS ^-f unction. Our first lower-semicontinuity theorem depends on Jensen's inequality which we now state: Lemma 1.8.2 (Jensens's inequality). Suppose cp is convex on Rp, Sis a set, pL is a non-negative bounded measure on S, and the functions ^^ ^ Li(S,pi),p = 1 , . . . , P . Then cp{i\ . . . , F ) < [f^{S)]-^f(p[^Hx),.. .,iP{x)]dii,

(1.8.4)

, i:p =

^ r [i,(S)r^f^Pdpt.

s Remark. I.e. 99 (average) < average of 99. Proof. Choose ap (Lemma 1.8.1) so that ^(C) + cipii^ - C^) < ^(f) for all f and then average over 5. We can now state and prove our first lower semicontinuity theorem; in this general form, it is due to SERRIN [2]: Theorem 1.8.1. Suppose thatf = f[p) is non-negative and convex for all p = {py] and suppose that z and each Zn^ ^\[^) ^'^^ t^^t z^^z (i.e. tends weakly to) z in Li{D) for each DCGG. Then I{z, G) and I[ZnG) are each finite or + 00 and I{z, G) < lim inf I{zn, G)

22

Introduction

Proof. The first conclusion is obvious. Let Z) C C (^; we may suppose t h a t D C Ga for some a^ 0. Let 99 be a moUifier and, iox 0 <i Q <Z a, let ZQ and Znq be the 99-mollified functions (of class C^ on D). From Fatou's lemma, Theorem B, and equation (1.8.4), we conclude that (since/[V-^e(:v)]->/[V^(^)] a.e.)
(1.8.5) I{z,D) < lim inf I[ZQ,D),

Let us now define

e->o
Fn{x) = / [ V Zn{x)] .

F{X) = / [ V Z{X)] ,

Then, from Jensen's inequality with d/j, = q)*(^ x) d^ (see Equation (1.8.4)), we conclude t h a t /[V^eW]<^eW> f[VZne{x)]^FnQ{x), x^D, 0<Q<a. Using this and Theorem B (setting F^ {x) = 0 outside G) we conclude t h a t (1.8.6) I{znQ,D) ^JFnQix)
D

dx<I{zn,

G).

The weak convergence in Li{D') for each D' C. C. G implies that VZnQ converges uniformly on Z) to V ZQ for each Q, 0 <. Q <. a. Thus (1.8.7) I{ze, D) = lim I{znQ, D) < lim inf I{zn, G).

Combining (1.8.5) and [\.^.7)y we conclude t h a t I{z,D) < l i m i n f / ( ^ ^ , G)

TO-* 0 0

from which the result follows easily using the arbitrariness of D. We now give a simple proof of lower semicontinuity for a wide class of integrals b u t using a more restricted type of convergence which is, however, sufficiently general for the existence theory. The hypothesis that the /^^ be continuous can be removed rather easily. More general theorems have been proved b y SERRIN (see Chapter 4). Theorem 1.8.2. Suppose f == f(x, z,p) and the f>pi are continuous with f{x, z, p) ^ 0 for all {x, z, p), suppose f is convex in p for each {x, z), and suppose Znrz in Hl{D) for each D (ZC.G, Then the conclusions of Theorem 1.8.1 hold. Proof. Choose D C G G. The weak convergence in Hl{D') for each D' d a G implies the strong convergence of Zn to z in Li(D) (see Theorem 3.4.4). B y choosing a subsequence, still called Zn, in which I{zn, G) - ^ i t s former lim inf, we m a y assume t h a t Zn{x) -^z{x) a.e. on D. We now suppose I{z, D) < + 0 0 . Then, for each e > 0, there is a compact subset S G D on which z and V ^ (i.e. representatives) are continuous, on which Zn converges uniformly to z, and which is such t h a t {\.8.8) I{z,S) >I\z,D) -8

1.9. Existence

23

(if I(z, D) = + c>o, we may take I{z, S) > M, arbitrary). Then, from Lemma 1.8.1, we conclude t h a t f[x, Zn{x), VZn(x)] > / [ ^ , Zn{x), V z{x)] + fp[x, z(x), V z(x)] ' [V Zn{x) - \7z(x)] +{fp[x,Zn(x),Vz{x)] -fj,[x,z{x),\7z{x)]} [VZn{x) - Vz{x)]. (1.8.9) The weak convergence imphes (see Theorem 3-2.4 (a)) the weak convergence oi VZn to Vz in L\{D) which imphes, in turn, that (1.8.10) Hm (fp[x, z{x), \/z[x)] [\/Zn[x) \/z{x)] dx = 0.

The uniform convergence of ^^^ t o ^ on S, together with the uniform boundedness of the Li norms of V Zn and V z, imphes that (1.8.11) hm f {fj)[x,Zn, Vz] fp[x,z, Vz]} [V^^^ \/z] dx = 0.

Hence, from (1.8.81.8.11), we conclude that I{z, D) s <, lim (f[x, Zn{x), \/z{x)] dx < lim inf I{zn, G). The theorem follows easily. 1.9. Existence I f / satisfies the condition (\.7A) with ^ > 1, an existence theorem can easily be deduced from the lower-semicontinuity theorems of 1.8 and Theorems 3-4.4 and 3-4-5- The following simple lemma enables us to prove easily a more general existence theorem (Theorem 1.9-1 below) Lemma 1.9.1. Suppose that fQ{p) is continuous and that (1.9.1)
39^00

lim|^i-i/o(i^) = + 0 0 .

Then, for each M, there exists a function cp such that cplg) '^ 0 for ^ >- 0 and (P{Q) -^0 as Q -^0 such that (1.9.2)
e

l\p{x)\dx<cp{\e\)*

for every measurable e C. G and every vector p [x) satisfying (1.9.3)

G

p^[p[x)]dx^M.

Proof. We define (p{Q) as the sup of the left member of (1.9.2) for all ^such that 1^1 < ^ and_^ satisfying (1.9-3)- If 9^(^) does not - > 0 a s ^ - ^ 0 , H an 0 > 0 and sequences {cn} and {pn} such that \en\ - > 0 and r \pn{x) \ dx > Q. We define ip{G) = inf \p\~^fo{p): for \p\ > a,

* We often use 1^1 to denote the measure of the set e.

24

Introduction

Clearly ^(cr) -> + oo as c - ^ + c>o. For each n, let gn be the subset of en where \pn{oc)\ > cr^ = eol2\en\. Obviously j
engn

\pn(x)dx

<an-\en\

= ol2^

Consequently, since/o(_^) > | ^ | 'y^(o') for | ^ | > (X, we obtain

(1.9.4) / \pn{x)\dx > 0/2,
gn

f Mpni^)] dx > tp{an)'ol2.

gn

Since or^ -> + 00 and ^'(o'^) -> + "^^^^ (1.9.4) contradicts (1.9.3)We now state and prove our principal existence theorem. Theorem 1.9.1. We suppose that (i) / and the fpi are continuous in their arguments; (ii) / is convex in p for each {x, z); (iii) there is a function /o satisyfing the conditions of Lemma 1.9.1 such that f{x, z, p) >fo(p) for all (x, z, p); (iv) F * is a (non-empty) family of vector functions which is compact with respect to weak convergence in H\[G)\ (v) F is a family, closed under weak convergence in H\[G), such that each z in F coincides on dG with a fimction z"^ in F"^ (i.e. z ~ z'^^B.\^{G), see 3-2); (vi) ^(-2^0, G) < + ^>o for some ZQ^ F; (vii) G is bounded. Then I{z, G) takes on its minimum for some z in F. Remarks. Since we have not made any assumptions on G other than boundedness and since the admissible functions are not continuous, the most convenient way to specify the boundary values of a function is t o state t h a t z ~ z"^^ ^io(Q ^^^ some given ,^*. Thus the family i^* defines, so to speak, the class of boundary values being allowed. Of course, F * could consist of a single function ^*. Proof of the theorem. Let {zn} be a minimizing sequence; we may assume t h a t I{zny G) </(2:o, G) = M. Using (iii) and Lemma 1.9.1, we conclude that the set functions f\\/Zn{x)\ dx are uniformly bounded
e

and uniformly absolutely continuous. From (v), we conclude for each n that there is a 2:* ^ F * such that w^ =" Zn z*^ HIQ(G). From (iv) we conclude that there is a subsequence, still called {n }, such that z* -v some ^* in -F*. From Theorem 3.2.4 (a) with ^ = 1, it follows that the set functions { \^z^\ dx are uniformly bounded and uniformly absolutely

continuous so that the same is true for {wn}. From Theorem 3.2.4 and the corollary to Theorem 3-2.1, it follows that there is a further subsequence, still called {^}, such t h a t Wn7W^R\^{G). Then, of course Zn7Z=:^z'^ + wvi\K\((G) and the theorem follows from Theorem 1.8.2. Actually our hypotheses do not imply that f{x, z,p) > 0; however, it is clear that /Q takes on its minimum, so that / is bounded below and there is no loss in generality (since G is bounded) in assuming t h a t / > 0. Remark 1. I f / s a t i s f i e s (1.7.1) with ^ > 1, it obviously satisfies (iii) above.

1.9. Existence

25

Remark 2. In the case v = i, one concludes immediately the uniform absolute continuity of the (unique in this case) absolutely continuous representatives Zn and hence their equicontinuity. The weak convergence implies uniform convergence, the limit function being obviously absolutely continuous. Remark 3. In Theorem 1.9.1, F could be, for example, the set of all z such t h a t z z"^^ HIQ {G), for some 2:* in -F* (i^* assumed given), and z satisfies a system of equations of the form (1.9.5) < [x, z {x)] zi,{x) + bij [x, z{x)] z3 {x) + Ci [X, z {x)-\ = 0, i= i,. . . , P ,

where the af^, bij, and Ci are everywhere bounded and continuous. For, suppose t h a t {zn} is any sequence in i^ ^^^ -^ ^ in Hl{G) and let {zr} be an arbitrary sub-sequence of {zn}. From Theorem 3-4-4 (applied to each smooth domain Z) C C G), it follows that Zs(x) -^z{x) almost everywhere for some subsequence {zs} of {zr}. liv^,i = 1, . . ., P , are arbitrary bounded functions, then <, [x, Zs{x)]vi(x), bij [x, Zs(x)]v^x), Ci[x, Zs{x)] v^{x) converge almost everywhere and boundedly to their limits. By replacing 2: in (1.9.5) by Zs, multiplying by ^'* and summing, integrating the result, and passing to the limit using the weak convergence, we find that z^ F. Remark 4. If 2 is the vector function defined by zi{x) = Z)= w^{x), 0 < |(%| < Mi 1, ^ = 1, . . ., N, " the integral I{z, G) is equal to an integral J{w, G) of a variational problem involving derivatives of order '<mi of wK Then Theorem I.9.I f o r i implies a corresponding theorem for / . In fact the existence theorem of / . GEL'MAN can be deduced immediately in this way. In order to obtain more meaningful boundary value problems, it is convenient to restrict ourselves to Lipschitz domains G. Using the general compactness theorems of 3-4 we can prove the following boundedness lemma (by contradiction): Lemma 1.9.2 (cf. Theorem 3-6.4). Suppose G is Lipschitz, a is an open subset of G, and r is an open subset of dG. Then there are constants Ci{G, a,r,v) and C2{G,r,r,v) (depending only on the quantities indicated) such that I z||J, < Cx {\\Vz\\U + \\ z\\\J-\ I f^HI(G),

Remarks. This lemma implies that if F is any family of functions z in Hl[G) for which the Lr{G) norms ||V2:||J^^ of the first derivatives are bounded, then the Lr norms of the functions will be bounded if their Li-norms either over a fixed open subset or of G or a fixed open subset T

26

Introduction

of ^G are bounded. The family F* could then be replaced by a compact (in Lr(dG)) family of functions really defined only on dG. As an example of the use of our general existence theorem in connection with Lemma 1.9.2, we mention the problem of Plateau for a surface of least area which has a given arc F, with end points on a manifold M, as part of its boundary, the rest of its boundary being required to lie on M, On account of conformal mapping (see Chapter 9), it is sufficient to minimize the Dirichlet integral among all z^ Hl[B{0,i)] whose boundary values along d'^B(0,\) (the closed upper semicircle) are continuous and give a 1 1 continuous parametric representation of F, in which the point (0,1) corresponds to a fixed point of F, and whose (possibly discontinuous) values on d''B{0,\) lie on ikf (almost everywhere). I t turns out (see 9-3) t h a t the mappings from d+B(0,\) to F are equicontinuous for any minimizing sequence. Since they are uniformly bounded || z^ \\1^B s^^e uniformly bounded and a subsequence, still called {zn} converges weakly in Hl[B(0,\)] and uniformly along d+B{0,\) to a minimizing function z. However, z (harmonic and conformal inside 5(0,1) m a y not be continuous along d-B(0,\) if M is allowed to have edges, as has been shown in an example of Courant ([2], [3]. However see, L E W Y [2]). Remarks. The theorems of this section and the preceding one can be carried over to integrals involving derivatives of higher order of the form I{z, G) = Jf[x, L[z), M{zy\ dx
G

where M[z) denotes all the derivatives D' z^ of highest order where | ^ | == nii and L[z) denotes all those derivatives where 0 < | ^ | < w^-. I t is assumed t h a t / i s continuous in its arguments and convex and differentiable in the set of arguments M{z). There are essentially no differences in the proofs. Of course each vector z in F (Theorem 1.9-1) would be such that z^ - ^*'*$ ^foHQ foi" some ^* in i^*, /o =:/o[M(^)], etc. FICHERA has.observed t h a t lower-semicontinuity theorems for such integrals can be obtained when M{z) consists only of some combinations of the derivatives of highest order and L(z) consists only of some of those of lower order. 1.10. The differentiability theory. Introduction Of perhaps greater interest than the existence theory is the theory of the differentiability of the solutions. In this chapter, we shall confine ourselves to the non-parametric case; we shall discuss the parametric case in Chapters 9 and 10. The first result about differentiabihty was that of LICHTENSTEIN who proved in 1912 [1] t h a t a solution z of class C" of a regular double integral problem {v = 2, N = 1) in which / is analytic is of class C "

1.10. The differentiability theoy. Introduction

27

and hence analytic b y the famous theorem of S. BERNSTEIN ([1]). The same conclusion was shown to hold (a) if z is of class C^ with Holder continuous derivatives by E. Hopf in 1929 ([2]) and (b) if z is Lipschitz by thewriter in 1938 (MORREY [4]). Using the latter result, it follows that the Lipschitz solutions (mentioned above) obtained by A. Haar for the c a s e / = / ( ^ ) , r = 2, N ^= 1, G strictly convex, are analytic i f / i s . But except for problems whose Euler equations are linear and certain integrals where / is of the form f{x, z, p) = a(x, z) \p\^ + 2Z h^{^, ^)pcc + c {%, z){v = 2,N= 1)

treated by HIRSCHFELD there were no other results in which the solutions which were shown to exist were shown to be differentiable until the work of the writer ([6], [7] see footnote on p. 17) in which the entre existence and differentiability program was carried through for essentially the class of problems in w h i c h / satisfies the conditions (1.10.8) below with V = 2, N arbitrary, and k = 2; although N was allowed to be > 1, the convexity hypothesis on / was retained, that is <Mi Later ([8]), the writer applied these results to prove the differentiability of his solutions of the problem of Plateau on a Riemann manifold. A greatly simplified version of this old work is to be found in [15], recent developments have permitted still further simplifications and some extensions. The methods used in this differentiability theory with i^ = 2 would not generalize to larger values of v and it was not until the recent results
of D E GIORGI [1] and N A S H [3], as simplified still more by MOSER t h a t

the existence and differentiability program could be carried through for problems in which v '> 2, The methods of NASH and D E GIORGI were entirely different; NASH obtained his results as a by-product of his work on parabolic equations and confined himself to bounded solutions whereas D E GIORGI dealt only with elliptic equations b u t allowed solutions in L^. Using MOSER'S simplification of D E GIORGI'S work, a student E. R. BULEY was able in the spring of I960 to obtain the results stated in Theorem 1.10.4 (ii) for the cases where / satisfies (1.10.7) or (1.10.7') and the author was able to obtain those in (ii) (the (1.10.8) cases), (iii), and (iv). At about the same time, LADYZENSKAYA and URAL'TSEVA ([1], [2], [3]) obtained their results stated in Theorem 1.10.4 (v) which include the author's results. These results are discussed further below and most of the details are proved in Chapter 5. An unfortunate feature of the D E GIORGI-NASH results is that they have been proved only for iV = 1.

28

Introduction

We now show how to prove differentiabihty in the case v = 1, A^ arbitrary. We do not attempt to prove the most general theorem. Theorem 1.10.1. Suppose that v= \, that f^C^ everywhere, that f satisfies (1.7-1) with ^ >> 1, and that there is a positive function Mi{R) such that (1.10.1) | ^ ( ^ , ^ , / > ) | , | / , ( ^ , ^ , ^ ) | < M i ( i ^ ) ( | + |:^|^)for Suppose z minimizes I{z, G) (G = {a, b))\ among all A.C. functions with the same boundary values. Then z^ C\a, hi and satisfies Euler's equations. Proof. Let f be any Lipschitz vector which vanishes at a and h. Since / satisfies (1.7.1), it follows t h a t z is A.C. and \\z'\\l < oo. Since C is bounded and z and f are absolutely continuous, we see from (1.10.1) t h a t 'C*W -h'ix. z{x) + A f W , z-{x) + <C'W] + CW -fAx, ^ + A f, / + A f ] is dominated for | A | < 1 by the fixed summable function 2^.^1(7^) [1+2^-1 (g^+|/(^)|^)] where ^ is a bound for | I^'(x) \ and | C(^) | and R is one for A;^ _|_ | ^ _|_ ^^|2 for IAI < 1; here we used the inequality [a + h)^ < 2^-i(a^ + h^), ^ > 1. Thus the function (p{X) defined in (1.4.6) is of class C for \X\ < | and (p'ip) = 0 so that (1.4.7) holds. Now (1.4.7) is of the form (1.10.2) / C C M ^ + I:iBi)dx
a

= 0, C^-(a) = f^(6) = 0

where Ai and Bi are summable. Let 71^ be an arbitrary bounded measurable vector such that (1.10.3) j 7i^{x)dx = 0
a

Any Lipschitz f vanishing at a and h is of the form (1.10.4) f^y) = jn^

a

{x)dx =

J7i^{x)dx
y

Breaking the integral (1.10.2) into the integrals of its two terms, replacing xhy y m the second, then substituting (1.10.4) for C, interchanging orders of integration, and recombining, we conclude from (1.10.2) t h a t (1.10.5) 17t^x)\Ai(x) -.lBi{y)dy\ dx = 0.

Since this holds for all TT satisfying (1.10.3), it follows from a well-known lemma that (1.10.6) fpi [x, z{x), z'{x)] = Ai{x) == J Bi{y) dy -f- Q ( Q = const.) a.e..

1.10. The differentiability theory. Introduction

29

in which the right member is A.C. Since / is regular, the equations (1.10.6) can be solved for the 'z^ in terms of x, z{x), and the integral. Thus 'z^ is equivalent to an A.C. function. Thus 'z^ is easily seen to be A.C. Hence both sides of (1.10.6) are A.C and the right side ^ C^. Consequently z^ C^ so z^ C^ and EULER'S equations follow as usual. For r > 1, the theory is not so simple of course. Rather than trying to present the most general conditions under which each part of the demonstration can be carried through, we state now two sets of conditions on the integral function / under which differentiability results have been obtained. Common Condition, f ^ CI in its arguments or f and fp ^ CJJ~^ for some n > } and some fji with 0 <i ju <, 1. The use of Holder conditions in connection with elliptic differential equations is very common. The desirability of their use arises in potential theory (see Chapter 2). Besides the common condition above, we require / to satisfy one of the following sets of conditions for all {x, z, p): imV^ K <f{x,z,p) \\fv\' (\A0.7)l\fpz\' ^MV^, < Mf F2A:-2 V^~^\7t\^, <Mi, ect. H' = I{<)'^ \p\^)^l^. + Ifvccl' + \fz\' + \U\^ + \fzz\'<MlV^^-^

\mi F^-2 I T |2 < Zfplp'p T [o < M < M , k> \fpp\'=I{fpipir> (1.10.7') imV^ K <f{x,

{^' ^> P)<^h^^l

\, F == (1 + | ^ | 2 + |j^|2)i/2, 0 <mi \fpz\'=I{fpiz^)'> z, p) < MV^,

Same as (1.10.7) b u t / = /(:v,^), F = (1 + |/.|2 + |/.^|2 + | ^ | 2 < M ( i ? ) F 2 ^

(1.10.8)

\\fp? + \fpz\^ + \fpx\''^M[R)V^^-^ Ui(i?)F^-2l7r|2<2'Aj^2,^:7rj,4<Mi(J^)FA^-2|7T|2, 0<m<M, 0 < mi{R) < Mi{R), \x\^ + \z\^ <R^ F =- (1 + \p\^)^l^

k>v,

Remarks. We notice that (1.10.8) reduces to (1.10.7') in case / does not depend on z (except for the i^-condition which is somewhat meaningless if z is not present since our domains G are bounded). To see the difference between (1.10.7) and (1.10.8), we notce that the function/defined by f(x,z,p) = [\ + affix, z)pip^fl^ satisfies the conditions (1.10.8) but not (1.10.7) if the aff^ CI or Q - ^ if ^ > 3 and the quadratic form is positive definite and bounded above and below in the obvious way.

30

Introduction

The first step in the differentiabihty theory is to note continuity properties of the minimizing functions as follows: Theorem 1. 10. 2. In all cases if k'>v the minimizing functions are Holder continuous on interior domains and at any boundary points in the neighborhood of which d G is Lipschitz. If k v, the minimizing functions are Holder continuous on interior domains and (1.10.9) I / V^dxY'^^C'dlVWD^irlRy satisfies m\p\^ ^f{x,z,p) <M\p\^ iiB(xo,r)c:B{xo,R)(ZG.

If k =^ V = 2, and f (1.10.8*)

G is bounded by a finite number of disjoint Jordan curves, z*^Hl(G), and 2:* is continuous on G, then any minimizing function with z ^ * $ H\Q[G) is continuous on G. li k == v, G is Lipschitz, the boundary values are continuous, and f satisfies the supplementary conditions (1.10.10) f{oc,z,p) > 0 , f{x,z,p^, . . .,^^-1,0,^^+1, . . . , ^ ^ ) = 0, r = 1, . . ., N{p^ = {Pi}, i fixed) then any minimizing function with those boundary values is continuous on G. Proof. If ^ > r and I{z, G) is finite, it follows from the first assumptions in (1.10.7) and (1.10.8) that the H\[G) norm of the minimizing function is finite. That z is continuous on the interior follows from the corollary to Theorem 3-5.1. The continuity on the boundary is obtained by first ''flattening out'' a piece of the boundary by a bi-Lipschitz map (see the definition in 1.2) and then reflecting the function. In the case v = 2 = k where / satisfies (1.10.8*), the minimizing vector z satisfies mD[z, B{xoR)] < I[z, B{xo, R)] < I[H, B[xo, R)] <MD[H, (1.10.11) B[xo, R)]

where H is the harmonic function coinciding with z on dB(xo, R). If we let (p{r) = D[z, B(xo, r)], take polar coordinates about XQ, expand z in a Fourier series (1.10.12) z = ' ^
^

+ Z [an[r) cosnO + bn{r) smnB],

n

and use the formula (1.3.3) for D[H, BR], we find that
R

<K7c2J[aliR)
n

+bl{R)]^K-R-q,'{R), \jK.

K =

Mlm

from which (1.10.9) follows w i t h / / =

1.10. The differentiability theory. Introduction

31

The condition (1.10.11) holds, obviously, with B{xo,R) replaced by any other sub-domain of G, the resulting condition being invariant under conformal mappings. The two-dimensional results on continuity at the boundary involve this idea and are carried out in 4-3- The proof above can be generalized to the general case k = v; the harmonic function H is replaced by the function H{r, d) =z + {rlRY[z{R, Q) ~ z], 0 < r < 7?,

where [r, 6) are spherical coordinates, 6 denotes a point on dB{0,i), and z is the average of z(r, 0) over dB{0,\). The remaining results are proved in 4.3. We must prove following interesting theorem: Theorem 1.10. 3. Iffis of class C^ and satisfies (1.10.7) or (1.10./') for some ^ > 2, then I(z) = I{z, G) is of class C^ over H\[G). If, instead, f satisfies (1.10.8) for some ^ > 2, then I(z) is of class C^ over the space *Hl(G) = H\[G) n C^{G), the norm in this space being

max 1^(^)1 + ||V^||?.GxeG

If we merely have 1 < ^ < 2, then I(z) is of class C^ over the corresponding space. In either caes, if zo minimizes I{z) among all z with given boundary values, then the first variation, i.e. the first differential Ii(zo,C), is zero, where Ii{z, C) is defined 5y (1.47). Proof. We shall prove only the first statement; the others are proved similarly. For almost all x, we may write fix, z + l:,p + n)= f{x, z, p) + C 'fz^ + <f^^ ^ + \ \fz^ z^ ^'^' +

+ ; C IK fff = / (1 - ^) {fvlT>\l^, z + tl:,p + tn-\-

fp^pi^ix, z, p)} dt

and fj- and Sij are given by similar Hl{G), then/j,[Zn] ->fpM and/^[Zn\ fvvi^n\ ->fpp[z]> fpz[zn] -^fpz[z], s = kj{k 2), and the efj^(zs), etc.,

formulas. It is clear that if Zn -^2: in -^fz W strongly mLr^,r = kl{k~ \), and fzz[Zn] -^fzz[^] strongly in Ls, - ^ 0 strongly in Ls. Thus we see that

i{z + 0= l(z) + ii(z,C) + Ihiz, 0 + R(z, C )

1 To see this, let {q} be any subsequence of {n}. Then there is a subsequence {s} of {q} such that Zs{x) -> z{x) and V ^s{^) -> V ^(^) almost everywhere so t h a t / p M - > / 3 ? M for almost all x. We have [/^[-^s]!^ and | / z M | ^ dominated by C ' Ml V^{x) and the strong convergence of jSg to z in H'^{G) implies the uniform absolute continuity of the set functions j V\{x) dx. A similar argument holds for
fpp> ij> e t c .

32

Introduction

where / , / i , and I^ are continuous in both variables and

\R{z,t:)\<s{z.i:)-m\^
where e[z\ t) ->0 uniformly as ||C|| - ^ 0 for ||2:' ^|| < some h > 0. Definition. Any vector z ^ Hl[G) for which Ii[z, C) = 0 for all f as in Theorem 1.10.3 is called an extremal. We can now state our principal results: Theorem 1.10.4. We assume that f is regular and satisfies the "common condition.'' Then: (i) If N = \ and z is Lipschitz and is an extremal, then z^ C^ [D) for each D C C G. (Recall the existence theorem w i t h / = f{p) and G strictly convex mentioned at the end of 1.7). (ii) / / / satisfies (1.10.7) or (1.10.7') with N = \, v arbitrary, and ^ > 2 or if f satisfies (1.10.8) with N == \ and k ^ v, then every extremal Q {D) for all Dec G. (iii) If f satisfies (1.10.8) with iV = 1, r arbitrary, and k = v or with N arbitrary and k == v = 2, then any extremal which ^ C^^ [D) for D G G G and which satisfies the Dirichlet growth condition (1.10.9) ^CJJ(Z)) for DCCG. (iv) / / / satisfies (1.10.7) or (1.10.7') with N = i, v arbitrary, and 1 < ^ < 2, and if z"^^ HI [G), then there exists an extremal z^ C^ {D) for all D G G G and which minimizes I{Z, G) among all Z such that Z ^* ff*o(G). (v) If f satisfies (1.10.8) with N = \,v arbitrary, and ^ > 1, then any boimded extremal ^ CJJ [D) for all D G G G. (vi) In all cases above, the extremal ^ C (D) for D G G G if / ^ C^ and is analytic on G iff is analytic. Remark 1. The restriction N = \ limits the applicability of the results very seriously. The removal of this restriction would not only increase the applicability of our results in such fields as Differential Geometry and Topology b u t is essential for the proof of the differentiability of the solutions of variational problems of higher order. The removal of this restriction involves the extension of the De Giorgi-Nash results to systems of equations. Remark 2. If f{x, z, p) is quadratic in z and p, the Euler equations are linear and more detailed results are available in this case. The results in (iii) with N arbitrary and k = v = 2 are those obtained before the war (193738) b y the author and referred to earlier. The results in (ii)(iv) other than those just mentioned, were obtained early in I960 in conjunction with a student (E. R. Buley). The remarkable results in (v) were obtained at about the same time by
LADYZENSKAYA and URAL'TSEVA ([1], [2], [3]). Their results include

1.10. The differentiability theory. Introduction

33

those of ours in (ii)(iv) above which refer to the cases (1.10.8) with N = \ and ^ > v since we have seen in Theorem 1.10.2 above that the solutions are continuous in these cases. The theory of LADYZENSKAYA and URAL'TSEVA is discussed briefly in 5 'I 'I; they have treated these and other details in their recent book [3]. However, the results in (ii) and (iv) in the cases (1.10.7) and (1.10.7') are of interest because the continuity of the solutions doesn't follow from the existence theory in the cases where 1 < ^ < r. Remarks. The functions z which we are calling extremals are also called weak solutions of the Euler equations

I t was noticed b y LADYZENSKAYA and URAL'TSEVA t h a t identical proofs

could be used to prove the differentiability of weak solutions of equations of the form (1.10.13) -^^A^ = Bi, A^ = Anx,z,p), Bi = Bi{x,z,p);

t h a t is, solutions z of the equations (1.10.14) f(CiA^

G

+ C'Bi)dx

=0

for the f in Theorem 1.10.3. The equations (1.10.14) are obtained from (1.10.13) formally b y multiplying (1.10.13) by f, integrating, and then integrating b y parts; and, of course, the equations (1.10.13) are obtained from (1.10.14) formally by following the derivation of Euler's equations given above. The assumptions on the Af and Bf corresponding to (1.10.7), (1.10.7'), and (1.10.8) are: \A\^ + \A^\^ + | 5 | 2 + \B^\2 < Mf F2*-2, + M P | 2 + \B,\^ + | 5 p | 2 < Mf F2^-4, \A,\^

(1.10.7")

mlV^-^\n\^^A^{x,z,p)7li7r$, wi > 0, ^ > 1, F = (1 + |^|2 + |^|2)i/2 B = 0,

(1.10.7'") The same as (1.10.7") with A = A{x,p), V = {\ + 1:^12)1/2, k>\, \B^\^<Ml{R)V^^ \B\^ + \B,\^ +

(1.10.8")

mi{R)V^-^\7i\^<A^pi

71^71^^, 1^2,1 < M l (i?)F^-2

mi{R) > 0 , F = (1 + |^p)i/2. In all cases, it is assumed that the A^ and 5 ^ ^ CJ if n = 2 or the ^ ? ^^""^ and Bi^ CJJ~^ iin > } . And, of course, it is not assumed that
Morrey, Multiple I n t e g r a l s 3

34

Introduction

Theorem 1.10.4. With these assumptions, the results (ii), (iii), (v), and (vi) hold if the assumption that z be an extremal he replaced hy that that it he a weak solution in H\{G) of (1.10.13) cind satisfy the additional conditions mentioned in the cases (iii) and (v).
LADYZENSKAYA and URAL'TSEVA ([1], [2], [3]) and GILBARG and

others have obtained existence theorems for equations of the type

(1.10.13). B u t in the last two or three years, V I S I K , MINTY, B R O W D E R ([3])

and LERAY-LIONS {[6^)'\ have developed an existence theory for nonlinear equations in Banach spaces which covers a wide class of equations including many of higher order and all equations of the type (1.10.14) in which the Af and Bi satisfy (1.10.7") or (1.10.7'")- Their theory yields solutions in H\{G) (or corresponding spaces in the higher order cases) rather than classical solutions and therefore takes the place of the existence theory rather than the differentiability theory. However, combined with our regularity results we now have an existence theorem for any equation of the form (1.10.13) which satisfies the (1.10.7") or (1.10.7"') conditions with N = \. The relevant abstract theorem of Leray and Lions is stated and proved in 5.12 where this existence theorem is proved. 1.11. Differentiability; reduction to linear equations In this section we wish to give some indication as to how one goes about proving the results on differentiability which we stated in Theorem 1.10.4. We begin b y applying a difference quotient process to the equations (1.10.14) in which we regard the solution as known. We shall assume first that the A^ and Bi satisfy (1.10.7") with ^ > 2; we shall indicate the modifications for the case (1.10.8"). To do this we choose an integer y, 1 < y < i^, let ^y be the unit vector in the x^ direction and define ^l (x) = h~^ \p {x-h (1.11.1) Cy) - C* {x)], zi {x) = h~^ [z^ [x + h ey) - z^ {x)], 0 < \h\< a

where C bas support in a domain D' (Z C Ga. If we replace f by ^h in (1.10.14), make a change of coordinates x in the terms containing C{x hCy) or V C(^ ^ ^v)y we obtain the equation (1.11.2) AAl =A'^[x-{fh-HCU^^f + C'ABi]dx = 0, S7z(x)]

hey, z(x + hcy), \7z(x + he-y)] Af[x, z(x),

and A Bi is given b y a similar formula. Now for each fixed h,0 <\h\ <, a, the two terms in AAf and ABi are defined and measurable for almost all X and for such x can be expressed in terms of the difference quotients A z^jh and their derivatives using the integral form of the Theorem of the

1.11. Differentiability; reduction to linear equations

35

Mean. If this is done (1.11.2) becomes

(1.11.3)
where

=0
1

An(x) = | [ 1 + l^(^) + tAz\^ + \p(x) +

tAp\^]-^+^^^dt p(x) + tAp]dt

'^h(x)a^f^{x)=lA^pi^[x

+ they,
1

z(x)+tAz,

^hh?i,^^lAf,idt,
0
1

AneiyPn=JAt,ydt
0
+ \p(x) + M ^ | 2 ] ( A : - I ) / 2 ^ ^
1

(1.11.4)

AnPh = l [ \ + \z{x) + tAz\^

^hO^ij = J Bipi^dt,

Ahdjiij

= J

Bizjdt

Anfl^Pn^

^Bi^vdt,

p{x)=\/z(x)

Az = z{x + h Cy) z{x), Ap = p(x -{- h ey) p(x). In this case {k > 2), we see from (1.10.7") that the functions A^ etc., are all measurable, the coefficients aji, hn, cn, dji, en, and fji are all uniformly bounded (independently of h) and (1.11.5) ^ i | ^ | 2 < < f , - ( ^ ) 7 i t 4 ' Zi^hiA^W^^l' An[x)^\ for all h with 0 <\h\<a. Next, we let C^ = rj w^, w = rj zl where rj = i on D which we assume c. D'^, r] = \ 2a~^ d(x, D) for 0 ^d{x, D) < ajl, and r] == 0 otherwise. Then we have CU = rj(ze;% + n,a4), iqzy = w\^ - YJJ Z{. If these C* are substituted into (1.11.3), one obtains JAn{{w\, + 77,a4) KIM^ - Vji) + ^tii^' + rieUPu\ +

(1.11.6) + wi[cl,^(w{, - ri,.z{) + dmow^ + rifl,Pn\)dx = 0. Using (1.11.5), the Schwarz inequality, and the Cauchy inequality ( 2 | ^ 6 | < a 2 + -i&^ e>0), (1.11.6) yields (1.11.7) ( An\Vw\^dx<^C j Ah[\Vri\^zn\^ + \wf + YJ^P^dx.

From the definition oi rj, we then conclude that (1.11.8) jAn\VZh\^dx^ClAh[{a-^+

D

i)\zh\^

Pl]dx.
3*

D'

36

Introduction

But, from Lemma 3.4.2 and Theorem 3.6.8, it follows that A^ ->V^-^ in Lr(D'), Zfi ->z^y in L]c[D') and P;^ - ^ F in L]c{D') so that the right side of (1.11.8) is bounded independently of h. A straightforward argument, given in 5-9, shows t h a t we may lei h ^0 (through a subsequence) to obtain the following theorem: Theorem 1.11.1. Suppose that z ^ Hl{G) and is a weak solution of (1.10.14), where the Af and Bi satisfy (I.IO.7") or (1.10.7'") with ^ > 2. Then z and U = V^l^^ Hl(D) for each D (Z d G and the vectors py satisfy

(1 .11. 9) (1 .11. 10) / Vfc--^vp \^dx <.Ca D where (1 .11. 11)

-V V^dx ,
D

DCD C C G , C .

p^[^' z{x), /-W]. Vk- -2 b'^. =. ^tzi lefy = Af,y, F*-2 cfi = Bij^ Vfc-^dij = Bui, V^-^fr = Bi^y.
Aoc

-^[x)al v^- fw

, v^-

In the case (1.10.8") with k>:v, we can prove in essentially the same manner Theorem 1.11.1' which is the same as Theorem 1.11.1 except that h'^j, c^ and f\ must he replaced respectively by bf^ * V, cf^ V, and fl V fand dij must be replaced by V^dfj and we must assume that z satisfies 1.10.9). Moreover we conclude also that (1.11.12) IVf^+^dx<Ca-^
D
D'

fVf^dx.

The proof in this case is more difficult but is simplified b y using a lemma
of LADYZENSKAYA and URAL'TSEVA [2] (see Lemma 5.9.1). There is no

result corresponding to (iv) of Theorem 1.10.4 for the variational problem. In the case of the variational problem we have the following result: Theorem 1.11.1''. Suppose that f satisfies (1.10.7) or (1.10.7'), with 1 < ^ < 2, and z"^ ^ Hl{G). Then there is a z which minimizes I(Z, G) among all Z such that Z -- z"^^ ^K^) <^^^ which is such that its derivatives ^ HI (D) and satisfy (1.11.9) and (1.11.10) on each D d G Ga and the function U = F^/2^ HI [D) for each such D. The difficulty in the proof of this arises from the fact that Ah always < 1 and this causes trouble in the difference-quotient procedure. To get around this we minimize I[z, G) with D{z, G) < K. For each K, we obtain certain bounds independent of h and so we may let /^ - ^ 0. Then we may allow i^ ->c>o. The special argument used in this proof is presented in 5.10. In all cases, the next theorem to be proved is the following: Theorem 1.11.2. Suppose that N = \ and that f or the A'^ and B and z satisfy the hypotheses of Theorems 1.11.1, 1.11.1', or 1.11.1". Then the

1.11. Differentiability; reduction to linear equations

37

function (1.11.13)

W = U^ satisfies an inequality of the form / K , a K ^ l ^ , ^ H- B^PW) + ^{C^^PW^o. + DP^W)]dx ifC^Lip,{G), C{x)^OonG, ^ 0

for some Ay 1 < A < 2, where the a*^ satisfy (1.11.5) (^ = ^), the coefficients B, C and D are hounded and measurable, and P^Lp (G) and satisfies (1.11.14) j f PUxY" <.ZrQ, DCCG

where Z depends on A, G, and

\\P\\^^Q.

This is proved in 5.9. The idea is to set ^ =.yjV-'py, y)^Lipc(G), xp[x) > 0 on G,

in equations (1.11.9) (technical lemmas allow this). I t turns out to be possible to choose e small enough so t h a t W = V^~^ = U^ satisfies (1.11.13). In the cases (1.10.7), etc., P = 1 whereas P = F in the cases (1.10.8). T h e inequality (1.11.14) follows in the cases (1.10.8), etc., from the Holder inequality in case k ^ v and from Theorem 1.10.2 in the case k = V. The idea of studying the solutions of inequalities like (1.11.13)
is basic to MOSER'S simplification [1] of the D E GIORGI-NASH theory. This

is presented in 5-3 and 5-4. But b y combining the results of Theorems 1.11.1, 1.11.1', 1.11.1", and 1.11.2, we conclude t h a t the function U satisfies the hypotheses of Theorem 5.3-'I and hence U and therefore z and all the py are bounded on each domain D C. G G. Then the equations (1.11.9) take the form (5.3.21). I t follows from the theorems in 5.3 that the py are Holdercontinuous on interior domains. Then the equations (1.11.9) take the form (5.2.2) with Holder-continuous coefficients and it follows from Theorem 5.5.3 t h a t the derivatives of the py are Holder-continuous on interior domains from which it follows that z^ Cl(G). The results under (v) of Ladyzenskaya and Ural'tseva for bounded weak solutions, when 1 < ^ < r, are somewhat more difficult and are presented in 5-11. But as soon as the first derivatives are seen to be bounded on interior domains, the proof that z^ C^ in the case ;^ = 2 is the same as that sketched above. Once the solution {N 1) z is known to $ Cl(D) for each D G G G, it then follows that z satisfies the equation (1.10.13) which becomes (1.11-15) a'^^x, z, Vz)z^oc^ = g{^> ^ V^) > , a^^ [Xy z,p)=l (A;^ + AD, g = B-A^p.~

A%.

where in t h e case of a variational problem, this is Euler's equation a n d

38

Introduction

If/satisfies the "common condition" with ^ > 3, i.e. i f / a n d / ^ ^ Cl~^, in the variational case, or if ^ ^ CJJ"^ and B ^ C'^~^ in general, it follows t h a t the a'^^ and g^ C^^ and the differentiability of z stated in Theorem 1.10.4 follows from a repeated application of Theorem 5-6.3 as follows: Since z^ Cl{D) for each Z) C C G, it follows t h a t a""^ and g^ Cl(D) for such D so t h a t z^ Cl(D) for such D by Theorem 5-6.3. Then a^"^ and g^ Cl{D) for such D and hence z^ C"*(Z)) for such D. The result follows by induction. The C^ result follows b u t the analyticity requires a separate proof. The analyticity proof for this case {N = 1) is presented in 5.8 where references are given. It is to be noticed t h a t the first theorem above which does not hold for systems is Theorem 1.11.2 (see the remark after the proof of Theorem 1.11.2 in 5.9)- This enables us to show t h a t solutions are Lipschitz on interior domains. But even if this could be proved, the D E GIORGI-NASHMosER theory (presented in 5.3 and 5-4) has not been extended to systems so we would still be unable to prove the Holder continuity of the py, or even their continuity. In case a part of ^G is smooth and the boundary values are smooth along t h a t part, there is a gap in the differentiability results. By making a correspondingly smooth (CJJ, n^ 1) change of variables and subtracting off a smooth function having the given boundary values, one reduces the equation (1.10.14) (or (1.10.13)) to one of the same type in which one works in a hemisphere and the solution is supposed to vanish (in some Hl^ sense) on the flat part GR {X^ = 0) of the hemisphere Gn (see 1.2). The transformed equation or system has the same properties as the original except possibly for the bounds. One then carries out the difference quotient procedure of Theorem 1.11.1 in the tangential directions and shows that the py ^ HI (Gr) foTr<.R and satisfy the equations (1.11.9), y = "i, . . .,v \. Since it is true that pv^y = py,v for y <.v, all the derivatives pv,y with y <.v^ L2{Gr). Since we already have interior differentiability we may solve the equations (1.10.13) for the z^^^ = pl^, in terms of the others. Thus all the py,y,y = 'i,.. .,v, satisfy (1.11.9) and (1.11.10) follows. However, Theorem 1.11.2 has not been shown, in all cases, to hold in such a boundary neighborhood. But the following theorems cover what is known to the author about the boundary behavior of solutions: Theorem 1.11.3. Suppose a portion y of dG, containing a point PQ, is of class C'^, n > \y suppose that the Af and Bf satisfy the common condition and any one of the sets (1.10.7'')> (1-10.7'")' ^^ (1.10.8") for some r > 2 and ^ > 1, and suppose that z^Hl {G) and is a solution o/(1.10.14). / / , in addition, N = \, z satisfies a Lipschitz condition in a neighborhood ru y{rG G) of PQ on G and vanishes along y, then z is of class C^ in any smaller neighborhood^ of PQ on G. If N ^ 1, the same conclusion holds if

2.1. Introduction

39

we know that z^ C^{r[J y). Ify,A, and B^ C^ or analytic, respectively, so is z on Syi. Theorem 1.11.4. Suppose G is bounded, (p ^ Hl{G), N = i, B = 0, and the A^" = A'{p) ^ C^ and satisfy the inequalities (1.10.7''') with k 2. Then there exists a unique solution z of Equation (1.10.14) such that z (p^ HIQ (G) . / / G is of class C^ and cp ^ C^(G) then z^H^ {G) for any r. If the A'^^Cl, G is of class C^ and (p^ C^iG), then z^ Q ( S ) . / / we merely know that the A'(p)^ C^ and satisfy the condition ^p^{P)^ock>0 if ;. #=0, p^Rv

and if G is of class C^ and cp^ C'^(^) then there exists a unique solution z^ ^(0) of Equation (1.10.13) such that z (p on dG. This is just a combination of Lemma 4.2.5 and Theorem 4.2.3. Additional boundary value results are found in Theorems 4-2.1 and 4.2.2. For N = \ and the py all bounded, the equations (1.11.9) take the form (5.3.21) with G replaced by GR, assuming t h a t the above-mentioned transformations have been made. Since each py with 7 < r vanishes along OR, it follows from Lemma 5.3.5 t h a t these py^ C^(Gr) for each r <. R and, in fact these py each satisfy the ''Dirichlet growth*' condition (5.3-27). Since the derivatives of pv are all determined as above from those of the other py, it follows t h a t pp also satisfies such a condition and hence ^ C^ {Gr) for each r <CR (Theorem 3-5.2). The C ^ results are evident and the analyticity is proved in 5.8. The results for A^ > 1 follow from those in 6.8, 6.4, and 6.7. Chapter 2

Semi-classical results
2.1. Introduction In this chapter, we begin by proving some of the elementary properties of harmonic functions. A proof of W E Y L ' S lemma (Weyl [2]) is inserted in 2.3 for later reference; the proof is included at that point since it is closely related to the mean value property. Then in 2.4 the classical notions of GREEN'S functions and elementary functions are introduced and these notions and POISSON'S integral formula for the circle and half plane are carried over to the r-dimensional case. In 2.5, the study of potential functions is begun and the formulas for and continuity properties of their first derivatives are derived. In 2.6, the formulas for and continuity properties of the first derivatives of certain "generalized potential functions" are studied and the HOLDER continuity of the second derivatives of ordinary potentials of HOLDER continuous density functions follow from the general results; an example of a con-

40

Semi-classical results

tinuous density function whose potential is not everywhere of class C^ is given. In 2.7 we present a proof of the now famous inequalities of CALDERON and ZYGMUND ([!]) and ([2]) for singular integrals; we confine ourselves t o reasonably smooth functions in order t o retain the essential simplicity of their proofs. The chapter concludes with the proof of the maximum principle for certain second order elliptic equations which was given by E. H O P F ([1]). 2.2. Elementary properties of harmonic functions
V

Definition 2.2.1. u is harmonic on G <> u ^ C^{G) and A ii{x) ^ '^u^o^oi[x) = 0,

x=l

x^G,

Theorem 2.2.1. If u is harmonic on G, D (ZC G, D is class C^, and B{xo, R) C G, then (2.2.1) (2.2.2) (2.2.3) Ju^ocdx^=^ 0
dD

u{xo)=r-^lu{x)d2: u{xo) = \B{xo,R)\-^fu(x)dx.

B{xo,R)

Proof. The first follows from GREEN'S theorem. If D = B{xo,r), 0 <r <R, (2.2.1) takes the form (2.2.4) furdS
dB(xo,R)

=0

where Ur denotes the radial derivative. Let (r, p) be polar coordinates with pole at XQ, r = \X XO\ and p on U = dB{0,\). Then (2.2.4) is equivalent to (2.2.5) j Vr[r,p)dZ = 0, v(r,p)=u(xo + rCp), Cp = qp,

Integrating (2.2.5) from 0 t o jR yields (2.2.2). Using (2.2.2) with R replaced by r, we obtain (2.2.3) by multiplying both sides of (2.2.2) by Fv r^-^ and integrating with respect to r from 0 to jR. Theorem 2.2.2. Suppose u is harmonic on the domain G, XQ^G, U takes on its maximum value at XQ. Then u[x) = U{XQ). The proof follows easily from the mean value Theorem 2.2.1 and the connectedness of G. The following important uniqueness theorem follows immediately from the maximum principle: Theorem 2.2.3. There is at most one function u which is continuous on G and harmonic in G which takes on given continuous boundary values on dG (G being a bounded domain).

2.3. W E Y L ' S lemma

41

Theorem 2.2.4. If u is continuous and {2.2.}) holds for every B{XQ,R) G G G, then u is harmonic. A harmonic function has derivatives of all orders which are harmonic. Proof. Since (2.2.3) holds, we see t h a t u ^ Ci(GO) with u^oi{x) = \B(xo,R)\-^fu(x)dx^
d B (ajo. R)

if B{xo,R)C.CG,

(x= \, . . .,v.

But then by GREEN'S theorem, we see that u^^ also satisfies (2.2.3)- By induction, we see t h a t all derivatives are continuous and satisfy (2.2.3). So, suppose XQ ^ G. Expanding in TAYLOR'S series, we obtain u{x) u(xo) = {x' xfj u^oc{xo) + ^^^'^^ \R(x,xo)\ +-^{^"^o) (^^ - ^ g ) ^,cc^M + R{XO, X) B{xo,R)CCG.

<M ' \ x - xo\^ if \x - xo\ <R,

It follows, by integrating (2.2.6) over B(xo, r), dividing b y r^ \B{xo, r)\, using (2.2.3), and letting r ->0, that Au{xo) = 0. From this one easily obtains the following corollary: Corollary. If u^ G^{G) and satisfies (2.2.4) (i.e. (2.2.1)) for each sphere B{xo, r) C Gy then u is harmonic on G. For, b y repeating part of the proof of Theorem 2.2.1, one concludes t h a t (2.2.3) holds for each such B{x{), r). Theorem 2.2.5. Suppose each Un 6 C^{G) and is harmonic in G, and suppose that the sequence {un} converges uniformly on G to a function u. Then u^ C^(G) and is harmonic on G. For each Un satisfies (2.2.3) so the limit function does also. The reader can prove the following inequalities: Theorem 2.2.6. Ifu is harmonic and in Lp{G), then \u(x)\^ ^\B{x,r)\-^ f\u(y)\Pdy
B(x,r)

if

B(x,r)c:G.

Theorem 2.2.7. If u is harmonic and u^L^iG), then \S7u[x)\ < C{v) \\u\\\ d~'^~^''^ where dx denotes the distance from x to dG. Theorem 2.2.8. / / u is harmonic onG, \u{x)\ < Af there, dx denotes the distance of x from dG for x^G, there exists a constant C, depending only on V, such that \\/^u(x)\^k\eJ^-^C^MS-^, k>\. 2.3. W e y l ' s l e m m a We begin with a generalization of the Lemma of du BOIS-RAYMOND : Lemma 2.3.1. Suppose f <^ Li on the interval [a, b] of Ri and suppose that (2.3.1)

lf(x)g(x)dx

=0

42

Semi-classical results

for all g^ C^ [a, b]. Then f(x) = 0 almost everywhere. If (2.3-1) holds only for all g ^ C^ [a, b] for which
b

(2.3.2)

jg(x) dx = 0,
a

then f{x) = a const, almost everywhere. Proof. In the first case, it follows b y approximations that (2.3.1) holds for all g which are bounded and measurable on [a, b] from which the first result follows immediately b y setting g{x) = sgnf{x). In the second case, let ^1 ^ C^ [a, b] with

f gi(x) dx = \.
a

Then if g is any function ^ C^[a, b], we have

b b

g{x) = g*{x) + gi{x)Jg{y)

a

dy,

f g'^ix) dx = 0
a that u ^ Li(D)

and, of course, g * ^ C^ [a, b]. The second result follows.

Theorem 2.3.1 ( W E Y L ' S lemma, W E Y L [2]). Suppose

for each D G G G and satisfies (2.3.3) f u{x) Av{x) dx = 0

G

for all v^ C'^(G). Then u is equivalent to a harmonic function. Proof. Suppose B{XQ, a) C G, 0 <C s <, a, (p(r) ^ C"^ for r > 0, (p(r) = 0 ioT r > a, (p{r) = (p{e) if 0 < r < , and (2.3.4) v{x) =(p{\xxo\), X[r) =
dBixo.r)

fu(x)d2:.

Then v^ C^(G) and X^Li[s,

a

a]. Then (2.3.3) becomes

a

fx(r)
e

[I ( . - 1 <p')] dr=jX(r)y,
e

{r) dr = 0,

where ip m a y be any function ^ C^[s, a] for which (2.3.2) holds (with [a, b] = [s, a]). I t follows t h a t X{r) = a constant a.e. on [e, a]. Since e and a are arbitrary this is true for 0 < r < d(xo, dG); let us call that constant Fv u(xo). Using the definition of X{r) in (2.3.4), multiplying by y*'-i and integrating, we obtain (2.3.5) u{xo) = \B(xQ,r)\-^ Ju{x) dx,
B{Xo,T)

B(xo,r)

CG.

By holding r fixed, we see that u is continuous on G and by letting r ^0 we see t h a t u(x) = u{x) a.e. so that (2.3.5) holds with u replaced by u on the right. Thus u is harmonic on G.

2.4. POISSON'S integral formula; elementary functions; GREEN'S functions 2.4. Poisson's integral formula; elementary functions; Green's functions

43

Suppose G is a bounded domain of class C^ and u and v are of class C^ onG = GU dG. Then, from GREEN'S theorem, we obtain the formula / {uAv vAu)dx=
G

/ f^T
dG

^T~) dS =
dG

{uv,oc vu,oi) dx'^

(2.4.1)
n being the exterior normal. Next, it is a well-known and easily verified fact that

is harmonic ii y ^ 0, Moreover - {v -2) fg^-'^-dS J( dn^ d S~~ ] dB(o,Q) ^^<o.e) [271, iip=2 So, let us define wi I I ., [ {27i)-^log\y\ if )^ = 2 Now, suppose G is bounded and of class C^, u^ C^(G), Au(x) = f{x) on G, and XQ^G. Suppose B{X{),Q)(ZG and we apply (2.4-1) to the domain G B{x{i, Q) with v{x) = Ko(x XQ). Then we obtain Ko{y)=\^^ (2.4.2) = - {v -2)rr, iiv > 2

dB{xo,Q)

dG

G-Bixo.Q)

(2.4-3) Letting ^ - > 0 in (2.4-3), we obtain (2.4-4) u{xo) =J[^J^

SG

- ^ ^ ) ^S + f^oi^

- ^o)f(pc) dx

[v(x) = Ko{x xo)). Definition 2.4.1. The function KQ, defined in (2.4-2) is called the elementary function for Laplace's equation Au = 0. If f[x) = 0, then (2.4.4) expresses u{xo) in terms of its boundary values and those of its normal derivative. However, from the maximum principle, a harmonic function is completely determined by its boundary values alone. If, in (2.4-4), we have/(A;) = 0 and could take
(2.4-5) ^(^) = Ko(x xo) + H[x, XQ) -= GO{XQ, X)

44

Semi-classical results

where H is harmonic in x for each XQ and is so chosen that ?; = 0 on 3G, then (2.4.4) would express U{XQ) in terms of its boundary values. Such a function v, if it exists, is called a GREEN'S function for G with pole at XQ. By the maximum principle, the GREEN'S function is unique if it exists at all. From the discussion so far given, it follows that {a) if a GREEN'S function v exists for a given domain G and point XQ and [h) if u is of class C^ on G and harmonic on G, then (2.4-4) expresses U{XQ) in terms of its boundary values. If a GREEN'S function could be found and if it could be shown to be harmonic in xo for each x m G, then the function u defined by (2.4-4) w i t h / = 0 and u in the boundary integral replaced by a function u * would be harmonic; it would then remain to show t h a t u{x) ->^**(xo) d.s X -^XQ for each x^^ on dG. And, of course, proving the existence of the GREEN'S function requires proving the existence of harmonic functions having given boundary values. This problem is called the Dirichlet problem. Because of all the problems mentioned in the discussion above, the DIRICHLET problem is not usually solved by proving the existence of the GREEN'S function. However, there are two cases where this is possible, namely the case when G is a sphere which, obviously, may be assumed to have center at the origin and the other is a half-space which we may assume to be that where x"^ > 0. We now derive' the GREEN'S function for such a sphere. Let x^ BR B(0, R) and let x' be the point inverse to x with respect to BR, that is, the point where

'x^^ = R^x^'llx]^.
Using the spherical symmetry, it is easy to verify that the ratio \^ x\l If x'\ is the same for all ^ on OGR SO that (2.4.6)

H-x'

Thus we note that if we define ( ^ [ l o g | f - x \ - log\i - x'\ - log(|;^|/i^)], (2.4-7) G(x, I) _ (^ _ 2)-i p-i [|| _ ;^|2-v ._ ||: _ V= 2

x'\^-''{\x\IR)^-']

v> 2

then G(x, f) is of the form (2.4-5)- Moreover, by using the formulas for x' and f', we see that (2.4-8) G{1 x) = G{x, I) so that G is harmonic in x for each f and vanishes for f interior to BR and x^dBR. Thus the function u defined by (2.4-4) with / = 0 and v{^)

2.4. POISSON'S integral formula; elementary functions; GREEN'S functions

45

= G{x, f) is harmonic on BR, Finally, since the function w = 1 and G satisfy the hypotheses of t h e argument in the paragraph containing equations (2.4-3) and (2.4-4), we conclude that (2.4.9) | ^ , S ( | )
dBR

= 1,

.^B,.

By computation from (2.4.7), we see t h a t

dn{i)

R-^^^G^oc = i ^ - i r ^ - i f ^ p l - x\-'{^^ - A;)

;|:^|/i^)2-r||_^'|-v(|a_^^a)

For I on dBn, we may use (2.4-6) to obtain

(2-4-10)
(2.4.11)

^ ^ = (/;7?)-i if - x\-r{m - \x\^) > 0

u{x) = {ryR)-'^ j\^-x\-''{R'^\x\^)u*(^)di

and thus obtain Poisson's integral formula

for the harmonic function u which takes on given values ^ * on SBR. To see t h a t u{x) ->w*(fo) as x->io if u* is continuous, X^BR, fo ^^i2, we note from (2.4.9), (2.4-10), and (2.4-11) that u(x) - M*(|o) = ( r . i ^ ) - i / | | - x\-^{R^ - |^|2)[^^*(f) u*(^o)]dS.

(2.4.12)

^^

To show t h a t this difference -> 0, we break t h e integral on the right in (2.4-12) into integrals / i over SBR H ^ ( I O , Q) and I2 over ^^i? B{^o, Q) where we m a y choose g so t h a t | ^ * ( | ) w*(fo)| < e/2 for | ^ ^jBi? Pi B(^O,Q), S being given. The reader m a y complete t h e proof. Thus we have the following theorem: Theorem 2.4.1. There is a unique function u which is continuous on B(0, R), coincides with a given continuous function u* on dB(0, R), and is harmonic in B(0, R). It is given in B(0, R) by Poisson's integral formula (2.4.11). We can now prove the following important reflection principle for harmonic functions: Theorem 2.4.2 (Reflection principle). Suppose u is continuous on G and harmonic on the {possibly unbounded) domain G which lies in a halfspace bounded by the hyper plane 77, suppose Q n U = S, and suppose u{x) = Oforx^ S. Letr = G\J G' \J S^^\ where G' is the domain obtained by reflecting G in U and S^^) is the non-empty set of interior [with respect to IT) points ofS. Define U(x) = u{x)forx^ G and define U{x) = u{x') for x^r G, where x' is the point obtained by reflecting x in II. Then U is continuous on P and harmonic in P.

46

Semi-classical results

Proof. That U is continuous on T and harmonic near each XQ in r S<0) is obvious. So suppose XQ^S^^^ and choose 2^ so small that B{xo, R) c r and let G(xo, R) = B(xo, R) n G- Let H be that harmonic function in B{xo, R) which coincides with U on dB(xo, R). From the symmetry in Poisson's integral formula, it follows that H(x) = 0 along 5(0) n B(xo, R) and hence H{x) = u(x) on dG(xo, R). Thus H(x) = u{x) on G(xo, R) and hence, using the symmetry in Poisson's integral formula, we conclude that H{x) = U(x) on B(xo, R). The following Liouville-type theorem is another immediate consequence of Poisson's integral formula: Theorem 2.4.3. A function which is harmonic and hounded on Rv is a constant. Proof. For, by differentiating (2.4.11) with u*=u and letting 7^ ^ o o , one finds easily that each u^x(x) = 0. Suppose we now consider the half-space i?+ : J^" > 0 and define G{x, I) = Ko{x i) Ko('x f) where G(f, x) = G(x, ^), V = x\ ' < = < , i.e. 'A; is the reflection of x in IT : x^ = 0. We notice that G(x, f) = 0 for f ^77. Thus G is harmonic in | for each x and in x for each | and thus would appear to be a candidate for a GREEN'S function for R^. Clearly ^G(A;, f)/ dn{i) is just dG(x, f)/9|''. Along I*' = 0, we have 2.4.13) - dG{x, f)/a|- - KoM^ - f) - ^o,.(^' - f) ,

A straightforward computation shows that L is harmonic in x for each il, is positive, and (2.4.14) lL(x,QdC= 1.
Rv~l

Thus if ^*(fi,.) is continuous and bounded and we define (2.4.15) u{x)=lL{x,Qu*{Qdil,

Rp-l

we see, as in the case of Poisson's integral formula for B{0, R), that u(x) is harmonic in R^ and hounded there hy the sup of ^*(|^), is continuous on R^, and coincides with /* along 77. We now prove: Theorem 2.4.4. There is a unique function u which is harmonic in R^ and continuous and hounded in R^ and which coincides on x^ = 0 with a given hounded continuous function u*; u is given hy (2.4.15) (^f^d (2.4.13). Proof. It remains only to prove the uniqueness. If ui and U2 both satisfy all the conditions of the theorem, then u = ui U2 is bounded

2.5. Potentials

47

and continuous on R^ and vanishes on 77. If we then extend u by reflection as in Theorem 2.4.2, then U is harmonic and bounded everywhere and is therefore a constant which must be zero. 2.5. Potentials In formula (2.4.4) with v(x) = Ko(x XQ) = Ko(xo x), we see that if u is of class C^ on G with (2.5.1) Au{x) =f{x) where G is of class C^, then the boundary integrals are harmonic so that the function U defined by (2.5.2) U{x)=lKo{x^i)mdi
G

would differ from ^ by a harmonic function and hence would also be a solution of (2.5.1). Definitions 2.5.1. The equation (2.5.I) is known as Poisson's equation and the function C7 in (2.5-2) is called the potential of f. Unfortunately, if / is merely continuous, it does not follow that U is necessarily of class C^; we shall give an example of this in 2.6. Since there is no difference in the proofs, we shall consider the more general integrals

(2.5.3)

U{x)^JK{x-^)md^.
G

The following definition is useful in this section and in the study 0/ higher order differential equations: Definition 2.5.2. A function (p{y) is essentially homogeneous of degree p in 3; if 99 is positively homogeneous in case p <C0 or, ii p is an integer > 0, 99 has the form

(2.5.4)

q^{y) = My) logbl = niy)

where cpo is a homogeneous polynomial of degree p and 991 is positively homogeneous of degree p. Theorem 2.5.1. Suppose f is bounded and measurable with \f{x) \ < M and has compact support G G, K^ C^{Rv {o}) and K is essentially homogeneous of degree 2 v, and U is defined 6y (2.5.3) Then U^ C^ everywhere with (2.5.5) (2.5.6) U,4x)==fK^4x-i)f(i)di, |VC/(^)|<M||Vi^||?.2:-^

rM,[3|ivi^||;,^ +
\'VU{x2)~VU{xi)\^l+\\^^K\\l,:\og{\+AlQ)], if ^ > ^ |M,||v2i^||;.2;log(1 +zJ/^), if Q<d, where q = \xi X2\, d is the distance of G from the segment x\ x^ and A is the diameter of G.

48

Semi-classical results

Remark. This theorem holds if U and / are vector functions and K is a matrix function. The proof is the same with indices inserted. Proof. We first assume t h a t / ^ ^ c ( ^ ) ' "the proof in the general case follows by approximations. We define (2.5.7) U,{x)=JK{x-^)f{^)d^.
G-B{x.Q)

It is easy to see that UQ converges uniformly to U on any bounded set and that (2.5.8) U,,4x) = -fK(x - |)/(f) di: + lK,4x - f)/(|) di;
dB{x,Q) G-B{X,Q)

this holds even if B (x, Q) C G, since we m a y first replace G by a large f region G' Z) GU B{x, Q). Since V K is homogeneous of degree 1 r, it is easy to see that the vector V UQ tends uniformly on any bounded set to the vector V defined b y

(2.5.9)

V4x) = UM^) ==IK,4X - i)m d^

G

from which (2.5.5) follows easily. Next, define (2.5.10) Then

(2.5.11)

V.Q[x)=JKAx-^)f[^)d^.
G-B{X,Q)

\VQ[X)-V{X)\<.\

"

"

.,

^,

Differentiating as in (2.5.8), we obtain

Fa,,^(^) = - fK,4x - f)/(f)^|^ + IK,.^{X - i)fi^)di,

dBix.Q) G-B(x,Q)

Thus, if we denote the distance from :r to G by ^1, we see that fM[|lVi^l|;.^+|lV2X||?.2:log(1+zJ/^), (2.5.12) | V F e ( : ^ ) i < | likr||v2i^l|S.2:log(1+/J/^i),
\V(X2) V{xi)\ < \V{X2) VQ(X2)\ + \V{Xi) - VQ(XI)\

if B{x, if
+

Q)nG4^0 B(x,Q)nG^0.
\VQ(X2)-VQ(XI)\

The result (2.5.6) follows b y setting |:V2 ^ i | = Q, observing that and using (2.5.12) t o estimate | VQ(X2) VQ{XI)\. 2.6. Generalized potential theory; singular integrals From Theorem 2.5.I, it follows that the first derivatives of potential functions are given b y formulas of the form (2.6.1) V(x)=fr{x~i)mdS

2.6. Generalized potential theory; singular integrals

49

where F satisfies the conditions (2.6.2) (a) / ' i s positively homogeneous of degree \ v, and (b)r$ci(i?.-{0}). We shall begin with a study of such integrals. All the results hold if V a n d / are vector functions and Z' is a matrix function. By repeating the second paragraph of the proof of Theorem 2.5.1 with K^oi replaced by F, we conclude the following theorem: Theorem 2.6.1. / / F satisfies (2.6.2), / satisfies the conditions of Theorem 2.5-1, and V is given by (2.6.1), then V is continuous. Theorem 2.6.2. Suppose F satisfies (2.6.2) and F^ C^[Rv {O}), / is hounded and measurable on the bounded domain G and f^ Cf^(D) for every D C G G, and V is defined by (2.6.1). Then V^ C^+f" on each D C C G and V,o.{x) = (2.6.3) Wo.{x) = Urn C4{x)+Wo.{x). fF,4x - f)/(|)^|, x^G
e-*0 G-B(x,Q)

C.= -.jr(-y)dy'^,
55(0,1)

the convergence being uniform on any DC. C G. Proof. We define VQ[X) by (2.5.10), with \/K replaced by F, and obtain the inequality corresponding to (2.5.11) and the formula

(2.6.4)

F,,(%) ^ -JFix - f)/(|)^f; + W.,{x)

SBix.Q)

where W^Q is defined by the integral over G B{x, Q) in (2.6.3). It is clear that the first term on the right in (2.6.4) tends to Cocf{x), uniformly on any D G G G, Next, choose a domain D' of class C^ such that D GGD' GGG and break up the integral for W^Q in (2.6.4) into an integral over G D' plus one over D' B{x, Q). In the latter integral and in the first term on the right in (2.6.4), write / ( | ) = f(x) + + [/(f) - / W J - Using the fact that F^cc(x - 1 ) = -dF{x - f)/^f and integrating over D' B{x, Q), we obtain

V,,4x) = C.[x,

U)'f(x)-^\F,.{x-l)f[i)dlG-B' B'-Blx.Q)

(2.6.5)

- jr{x ~ i) [/(I) -f{x)]d$', +frM^ - i) [/(I) -/(^)]^f

BB(x,g)

C4x. D') = -lr{x-i)dr^,

dD'

x^D.

Goi{x, D') and the integral over G D' ^ C^{D') and the last two terms in (2.6.5) tend uniformly to their limits. Thus the convergence of WXQ to Woe in (2.6.3) is uniform on D,
Money, Multiple Integrals 4

50

Semi-classical results

Differentiating WOCQ and using the ideas employed to obtain (2.6.5), we obtain W.,,p(x) =. C^,(^, D')f(x) +fr,.p{x - f)/(f) d^ G-D'

(2.6.6)

- / r , . (A; - 1 ) [/(f) -fix)]

8B{X,Q)

di^ +Jr,
B'-B{X,Q)

^ (X - i) Ua) - /{x)]

dl

The first two terms are bounded and are independent of Q ; the last two are 0(^'^""i). Thus, if ^ = |A;I :V2| and B[x, Q) C D' for every x on the segment xi X2, we see that \W,(X2) -W,(xi)\ Using (2.6.5) and the fact that ^CQ^.

v,A^) - F ,(^) = -jr{x - i) [/(f) -f{x)]di'^ + w,4x) - w^x)

I Wg{x) - W{x) I < C ^/^ if B{x, Q) C D'. The HOLDER continuity follows. Theorem 2.6.3. Suppose F satisfies the conditions (2.6.2). Then (2.6.7) jr^o.[y)dZ(y)
aj5(o. 1)

we obtain

dB{x,Q)

= 0,

oc==\,...,v.

Proof. Let Act denote the left side of (2.6.7). Then, from the homogeneity, we conclude that ^,log2=/[
1 SB {0,2)

j r,oc[y)dS{y)\dR=
[dB{Q,R) dB{0,l) \

r,o.{y)dy

( 0 . 2 ) - ^ (0,1)

= jr(y)dy'^-jr{y)dy'^

= 0.

From Theorems 2.6.2 and 2.6.3, we see that the first derivatives of functions V defined by (2.6.1), and hence the second derivatives of the potentials (2.5.2) (or (2.5.3)) are given by the formulas (2.6.3), where the Woe are given by the singular (or CAUCHY principal value) integrals of the form (2.6.8) W{x) = lim fA [x ~ ^)f[^)d^ = \imW^{x)

where A satisfies the conditions (a) A is positively homogeneous of degree v (2.6.9) (b) A^C^R,{0})

(c)
E

JA[y)dS{y)=0.

A proof like that of Theorem 2.6.2 yields the following theorem: Theorem 2.6.4. / / G and f satisfy the conditions of Theorem 2.6.2, A satisfies (2.6.9), and W is given by (2.6.8), then W ^ Cf*{D) for every D G G G. If f^ C^{Rv) and has compact support, then W^ Cf^{Rv).

2.6. Generalized potential theory; singular integrals

51

Proof. The last follows since G can be arbitrarily large. Using (2.6.9) (c), we see t h a t ii 0 <. a < Q, then Wa(x) - W,(x) =fA(x - I) [/(f) -fix)]di.

Clearly we may allow a -> 0 to obtain

I W(x) Wg(x) I < Ci Qf', B{x, Q)CD'CGG.

Proceeding as in the proof of Theorem 2.6.2, we obtain

W,,4x) =JAA^ - f)/(|)^| -fix) IA {X - ^)d^: ~

G-D'
6B{X,Q)

dB'

+ /zl,a(^-|)[/(f)-/W]if,
D'-B{X,Q)

X^D,

(DCCD'CCG). Thus if |:^i :V2| =Q and B{x, Q) C D' for every x on the segment xi X2, we find, as before, t h a t \W,{xi)-W,(X2)\<C2Q^. The result now follows. We now prove some theorems which will be useful later and which yield additional information in the case t h a t A also satisfies (2.6.10) (2.6.11) (2.6.12) jA(x-i)d^
B{xo,a)B{x^,Q)

A(-y)=A(y). = 0, = 0, x^B[xi, Q)CCB{XO, a)

Theorem 2.6.5. Suppose A satisfies (2.6.9) (^nd (2.6.10). Then

lA{x-i)di'^
d B ixo, a)

x^B[xo,a).

Proof. For each rj ^dB(0,\), let the ray through x in the direction rj intersect dB[xi, q) in the point fi and dB{xoy a) in the point fo and define (2.6.13) ^o(^) = | l o - : ^ | , n{v) From plane geometry, we conclude t h a t (2.6.14) = \^i-x\>

^o(^) 'Wo{~v) = a^ \x xo\^, Wi(v) 'fi(v) = Q^ \x xi\^. Taking polar coordinates with pole at x, we conclude t h a t f A{x-i)di = 2J^(v)\}^Sn{v) = 0 on account of (2.6.10), (2.6.9) (c) and (2.6.14). = jAirj)[logyjoi?]) - logtpi{ri)] dZin) rj)']d2{v)

+logi/;o( ^) logv^i(^) -~logy)i(

4*

52

Semi-classical results

From (2.6Ai),
(l ft AZ\

we conclude that
B{xo,a)B{xuQ)

dBixo.a)

dBixuQ)

Using (2.6.15) with xi = x, we conclude that lA(x-i)d^',

dBixo.a)

= fA(x
dBix.o)

- I; ^ 1 ; = ^-1 lfj^A{rj)dS
dBiO.l)

{rj) = 0

on account of (2.6.10). Theorem 2.6.6. (a) Any two points xi and x^ in BR can he joined by a path X = x(s), 0 ^ s ^ I, in BR such that
I

J {R - \x(s)\)f'-^ds<(2iu-^
0

+ \)'\xi

^2|^ 0 < ^ <

1.

(b) There is a constant C{fji) such that any two points of GR can be joined by a path x =^ x{s),0 < s < /, in GR such that j{d[x[s)]}f-^ds
0

< C ( / / ) - | : v i ~ x^^,

0 < / i < 1,

where d[x) denotes the distance of x from OGR = ZRU OR. Proof, (a) Is easily verified as follows: Let Q = | :^i :^21- li R <. Q <. 2 R, choose the polygonal path ;%:i 0 A;2. liO <.Q < R and | A;I | < i^ ^, \x2\ <,R Q, choose the segment %i :\:2. If | ^ i | > R Q, \x2\ <: R Q, choose the polygonal path xi xz X2 where xs is on 0 Xi with | ;^;3| = i^ ^. If [^i| > -?^ ^, 1^2! > ^ ~ ^, choose the polygonal path xi xs X4 X2 where xs is on 0 xi, x^ is on 0 X2, and \xz\ =:^ \x/^\ = R Q. (b) It is also sufficient to prove this f or 0 < | A;I ;V21 == Q = kR where 0 < y^ < 1/3, say. For such Q, the set SQ of x such t h a t d{x) > Q is the part of the ball \x\ ^R Q where x^{=y) > ^ . For such Q, we choose the paths xi X2 if xi and X2 ^ SQ, xi xs X2 if xi ^ SQ and X2 ^ SQ, and xi Xs X4 X2 if xi ^ SQ and X2i SQ; here xs is the nearest point of SQ to xi and X4 is t h a t nearest X2. A straightforward analysis verifies the result in all cases. Theorem 2.6.7. Suppose that A satisfies (2.6.9) cind (2.6.10). Suppose that f^ C^{BR), and suppose W is given by (2.6.8). Then W^ Cf'(BR) and \W{x)\<[2i^-l^-^\A\\'iS^K{f)R'' ^' K{W)<C{v,ix)WA\\l^+\\vA\\ls]K(f). Proof. From Theorem 2.6.5, it follows that

W{x) =
BR

\A{x-m{^)-f[x)]d^

from which the first inequality in (2.6.16) follows immediately.

2.6. Generalized potential theory; singular integrals

53

To prove the second inequality, we extend / t o B^u so that max/ and hii{f) are unaltered and write W =Wi- W2, W^{x) =JA{x(2.6.17) , ^'^ W2{x)=fA(x-i)mdi.

i)f{i) di

Differentiating l^i^ as in (2.6.6), we obtain Wu,Jx) = - f(x) lA{x-^)di:-lA(x-i)

dB2R dB{X,Q)

im - fix)] rfi; +
X^BR, 0<Q<R,

(2.6.18)

+ lA^4x-^)[m-f{x)]di,

in which the first term vanishes. From (2.6.18), we obtain \WlQ(X2) - WiQ(Xi)\ < 2h4f)[\\A\\l^ + (1 - /^)-l||VZl||?. J \X2 - ^ih (2.6.19) (1^:2-^li = 2^). From Theorem 2.6.5 and equation (2.6.17), it follows that I Wuix) - Wi{x)\ = 11A (X - i) [/(I) - fix)] di\ < h,if) \\A\\l^-fi-^ Q". (2.6.20) ^'^-^^ The result for Wi follows from (2.6.19) and (2.6.20). To prove the second inequality for W2, we first note that VW2ix)=I^Aix-i)fi^)di = JvAix-$)[fi^)-fix)]di, X^BR
BZR-BR

using Theorem 2.6.5 (or (2.6.15)). From (2.6.21), we conclude that

I vPF2(^)! < / 1 vzi (^ - f) I m - f(x) I di

RvB{x,d)

(^^^^)

< (1 - /.)-! Kif)' II Vzl 11?.^. (i? - iA;|)^-i

(^ = 7^_|:^|). From (2.6.22), Theorem 2.6.6 (a), and the fact that \W2{x2) W^ixi)] is dominated by the integral of j V TF21 over any path from Xi to X2, the result for W2 follows. Corollary 1. Suppose that f ^ C^^(BR) and U is its potential. Then U ^ C^'^^{BR) and there is a C(ILI, V) such that (2.6.23) AU(x) =f{x), X^BR, h^S/^U) <Ch^(f), Proof. From Theorem 2.5.1, it follows that U ^ C^ with

u,4x)=fKo,4^-i)mdi.
BR

54

Semi-classical results

Then, from Theorem 2.6.2 with r= for each r <. R, with (2.6.24)

KQ^CC, we see that VU^ ~ f)/(f) d^

Cl{Br)

C/,a^ W = C a ^ / W + Hm / KoM^ '"^^ ^.-^(-..)

5 5(0,1)

as one finds, using the formulas (2.4.2) for KQ. Since each i^o,a)8 satisfies (2.6.9) and (2.6.10), the second result in (2.6.23) follows. The first result follows from (2.6.24) and the fact that AKo(y) = 0 ii y z^ 0. By combining this result with POISSON'S integral formula, we obtain the following corollary: Corollary 2. Suppose/^ C^{BR) and u* is continuous on dBn, Then there is a unique function u which is continuous on BR, coincides with u * on dBn, ^ C^(Br) for each r <, R, and satisfies POISSON'S equation (2.5.1) on BR, We shall now give an example of a f u n c t i o n / ^ C^{Ba) the potential of which is not in C^(Ba). To do this, we first note that i f / i s bounded
and measurable on Ba and / ^ C^[B(XO, R)] {E{XO, R) C Ba), then its

potential V^ Cj^[B(xo, R)] since one can write V{x) = fKo{x~

B{xo,B)

i)f(S) di + fKo{x
Ba-B(xo,R)

- f)/(^) di

in which the first term $ C^[^(^o, R)] and the second is harmonic. Thus we have the formulas F, cc^{x) = Coa^fix) + I Ko^ ^^(^ _ I) [/(^) _ f(^x)] dS
Ba

{x^B{XQ,

R)) .

So now we define /(O) = 0 and f(x) = [log\x\]-^\x\-^x^x^, x=j^O, 0<a<\. Then its potential ^ C^{Ba {O}) and the formulas (2.6.22) hold if X ^ 0; but F,i2(:v) -> + ^^ as ;\; - > 0 . Remarks. However, if / satisfies a D I N I condition on a domain Q, then F ^ C^(G). Indeed if, in Theorem 2.6.2,/satisfies a D I N I condition on each D C G, then V F is continuous and given by (2.6.3), the convergence being uniform on any D G G G. Similarly if, in Theorem 2.6.4, /satisfies a D I N I condition on each D C G, then W^ C^{G). However W does not necessarily satisfy a D I N I condition. / is said to satisfy a D I N I condition on a compact set S if there is a positive, non-decreasing, continuous function (p{Q) such that l/(^i) / W l < 9^{\xi X2\), xi, X2^S,
a jQ-^<P(Q) dg <oa^ a> 0.

and

I.'], The CALDERON-ZYGMUND inequalities

55

2.7. The Calderon-Zygmund inequalities These very useful and now well-known inequalities are based on the following well-known (see ZYGMUND) inequality for the HILBERT transform which is actually the one dimensional case of the general inequalities. As was stated in 2.1, we shall restrict ourselves to HoLDER-continuous functions. The proofs are greatly simplified; moreover these results are sufficient for our later purposes. Theorem 2.7.1. Suppose f ^ Cl,{R^) (0 < ^ < 1). Then the HILBERT transform g, defined hy

(2.7.1)

^ = l-i/f^' W
\X^\>Q

^ C2 {R^) and, for each ^ > 1, ^ Lp{Ri) and (2.7.2) ||^||<C(;^).||/||. Proof. The integral (2.7.1) is the one dimensional case of the integral (2.6.8). Accordingly g^ C^^{Ri) by Theorem 2.6.4. It is sufficient to prove the result (2.7.2) for f{x) > 0 since we may always write / = / i /2 where each/^^ C^^(Ri) and is non-negative. So we define (27.3) ( ^ ' > ' ) = i f ( 7 y/ms

o o Since (2.7.3) is just POISSON'S integral formula for the upper half-plane (see 2.4) we see that u is harmonic for y > 0 and is continuous for y > 0 with u{x,0) =f(x). Since for some A > 0, f^ C^onlx] < A + \ snd f{x) = 0 for \x\ > A, we see, by writing/(f) =f(x) + [/(f) f(x)] in (2.7-4) that v{x, y) tends uniformly in A; as y ->0+ to g{x) for |A;| < ^ + 1/2, the convergence for larger x being obvious. Moreover ?; is a conjugate harmonic function to ^. If we let F{z) = u + i v, then o o o o Now, consider the function [F{z)]^ for y > 0, where w^ denotes that branch which is real and positive when w is (remember z* > 0 since / > 0). From (2.7-5), we see that \F(z)\P <M ' \z\-P for 1^1 > ^ + 1,

56

Semi-classical results CAUCHY'S

say. A simple limit argument using GR shows that (2.7.6)

theorem on hemispheres

j [F{x)]v dx = I [f{x) + ig{x)]Pdx = 0.

By integrating along the segment of length u = f{x) in the w plane from i V to {u + i v), we obtain for each x (2.77) (2.7.8) ( / + ^*^)^ (^'^)^ = p J wP-^ dw
segment

I p f WP-'^ dw\<p

\f{x} I [/2 + g2](i>-l)/2

<^-z(^)(!/|p+ 1/1-1^1^-1).

Integrating (2.7.7) from 00 to 00 and using (2.7.6) and (2.7.8), we obtain (2.7-9) l[ig{x)]Pdx <Ci(^)[(||/||)^+||/||-(|U||)^-i].

Since [ig{x)]P = \g{x) l ^ e x p [ i | ^ ] ^^^ I ^ W I ^ k | . we see that

p7T

k D ^ < C i ( ^ ) [ ( | | / | | o ) ^ + ||/||o-(|l^||)^-i]

from which the result follows for any p for which cos(^ 7r/2) ^ 0. We take care of those cases as follows: We note that cos(^ 7r/2) ^ 0 if 1 < ^ < 2. So suppose p > 2 and let p' be the conjugate exponent (p-i +p'-i = 1 ) ; then 1 < ^ ' < 2. Let h^ Cf^iRi) and let k be its HiLBERT transform. Then (2.7.10) \\kr^ < c{p') \\h\\i,. Moreover, it is easy to see that fh(x)g(x)dx = / - k{x)f{x)dx < C{p') ||/1|- \\h\\l, O O O O by applying the HOLDER inequality to the right side of (2.7.11) and using (2.7.10). Since the functions h are everywhere dense in Lp,, the theorem follows in this case also with C[p') C(p), Theorem 2.7.2. Suppose/^ C^^d^v), A satisfies (2.6.9) end (2.7.12) and suppose (2.7-13)
^^

(2.7.11)

A(-y) ^W = 1

~A{y),

/ ^ ( ^ - ^)/(f)^f
RVB{X,Q)

Then g^ C^i^v) Pi Lp{Fv) for each p > 1 and (2.7.14) \\gt<C{P)\\A\\l^-\\f\\l

2.7. The CALDERON-ZYGMUND inequalities

57

The hypothesis that A ^ C'^{Rv {O}) is not necessary; it is sufficient for /I^C0(i?,-{0}). Proof. That g ^ C^ (Rv) follows from Theorem 2.6.4 and that g ^ each Lp{Rv) follows since g(x) = 0 ( | ^ I"**) at c>o. Now, if we write ^ = x -\- rrj where r > 0 and rj^I! and use (2./. 12), we conclude that (27.15) lA{f])h(x, ri)d^[ri), i h{x, f]) = lim r r-^f(x J^rr])dr.
\T\>Q

g(x)=l

For each rj, let us write x = XQ + SYJ where XQ- rj = 0, Then h{x, + sri,n) (2.7.16) = \im r^l^ii^^, ^-^^s-*i>. ' (p{t;xo,rj) =f(xo + trj) in s for each

From Theorem 2.7.1, we conclude that h(xo -\- srj,r))^Lp {xo, rj) with (2.7.17) (2.7.18) ||v^|| < Cp Ml. yj(s; xo,rj) - /^(^o + [dx =

srj;r]).

By raising (2.7.17) to the p - t h power and integrating, we obtain / \h{x, fj) \Pdx < Cl{\\f\\l)P dx^ds).

The result (2.7.I4) follows from (2.7.15) and (2.7.18) by first applying the Holder inequality with measure -\A{y)\d^{y) to (2.7.15) to estimate |^(:v)|2? and then integrating with respect to x. Before we can prove the theorem corresponding to Theorem 2.7.2 when A satisfies (2.6.10) instead of (2.7.12), we need the following useful lemmas: Theorem 2.7.3. If u ^ Cl,(G), then (2.7.19) u(x) = f Ko^oc{x ~ d
G

i)u^oc(S)di

=r-^ I \x - i\-^ {x- - t)u,4i)ds.

Proof. We m a y assume G C B{x, A), u = 0 in B(x, A) G. Taking polar coordinates at x, the integral is just /
0

f
Li:

-Ur(r,rj)d2^(ri)]dr.

Lemma 2.7.1. If xi and X2 $ B^, (2.7.20) / | | - : ^ i | l - ' ' * | f - ^ 2 p - ' ' ^ l

BR

ClKo{x2 xi) + e{xi,X2,R), if V > 2 -{27z)^Ko{x2-xi) + 27tlogR'^e(xi,X2, + Ci + R), v = 2

58

Semi-classical results

where Co and Ci are constants and e{x\, X2, R) converges uniformly on BA X BAto 0 as R ^ o o . Proof. If r > 2, it follows t h a t as J? -> oo, the integral tends to a function oi \x2 xi\ only which is homogeneous of degree 2 r and so must be a negative multiple of KQ. Clearly also 8{Xi, X2, R) =
By-BR

-f\S-Xi\^-'''\i-~X2\^-''dS = {xi + X2)l2 and Q = \x2 xi\ and suppose - X2\-^dS < f \S - Xi\-^'\^
BR

converges as stated. In the case v 2,\etx R> A +Q. Then

B(x,R-\x\) Bix,R+\x\)

f \^' - Xi\-^'\S (2.7.21)

X2\-^d^

Evaluating the integral on the left, we obtain C2 + 2 7rlog[(i? - \X\)IQ] + cp[{R where C2 =
Bix,Q)

\x\)lQ]

l\i-Xi\-^'\i-X2\-^

-/|c-I
B(O.l)

ei

1 + ^.1
(i=(1.0))

<p{S) = J {\^ - Xl\-^-\^

B{x,R-\'x\)-Bix,Q) B(0,S)-B(0.1)

- X2\-^ -\^

x\-2)d^

= / (I f - eil2 |-l i I + l/2 r^ - II r^) '^l.

5 = (i? \x\)lQ.

In like manner, the right integral in (2.7.21) is given by

C2 + 2 7r log[(2^ + 1^1)/^] + (p[(R + 1^1)/^].

Clearly (p{S) tends to a limit 990 as S - > o o . If we take Ci = C2 + 9^0, then 27r log(1 - l^l/i?) + (p[(R - \X\)IQ] -(po<s < 27rlog(1 + \x\IR) + 99[(i? + 1^1)/^] - 990. Theorem 2.7.4. Theorem 2.7.2 holds, with \\A\\^^ replaced by MQ + Mi in (2.7.14), if A satisfies (2.6.9) and (2.6.10) instead of (2.6.9) and {2.7 A2). Here MQ and M, are the respective maxima of \A{x)\ and \\/A{x)\ for x ondB{0,\). Proof. We suppose first t h a t / ^ C'^(Rv), the general case follows by approximations. Let us define (2.7.22) h{x)==fR{xi)f(i) di, R(y) = Co-i I y | i -

2-7. The CALDERON-ZYGMUND inequalities

59

Co = 2 7Z if V = 2 and otherwise is the constant of Lemma 2.7.1. Then, by Theorem 2.6.4, A Q [Rv) with (2.7.23) A aW = H m fR,4^ - S)f(i)di ==fR{x-i)f,4i)di

from which it follows t h a t h^ ^^(Rp). follows t h a t (2.7.24) Vh(x) = 0(\x\-^), Next, we define

S i n c e / has compact support, it a t CXD. + 8R(x}

Ah{x) = 0{\x\-''-^)

(2.7.25) k(x)=^-fR{x-^)Ah{i)di=-fR{x-^)Ah(i)di
Ry BR

where SR converges uniformly t o 0 on any compact set on account of (2.7.24). Since t h e support of / is in some BA, we m a y write (since jAf(r))drj = 0)
BA k{x) = en{x) - I \ J R ( X BA [BR = IR{X) + fKo{x BA i)R(^ v)dS]Af{rj) drj

J
ri)Af{rj)dri

= 1R + f KG,cc{x~r])f,a{r])d7] BA

= SIR(X)

+f(x),

iit(x) = SR{X) jei{x,

BA

rjy

R)Af(rj)dr)

for all r > 2. Since k and / are independent of R, SIR ^ 0. Hence, substituting the right side of (2.7.25) for/(:^) in (2.7.13), we obtain g{x) = lim JHm (2.7.26) =:lim|Hm j A(x-^)\ jAh(7i)\ J -- R(i - r])Ah(r])dr] d^
[R^-Bix.o) Q-*0 \a^0 R^-B(x.Q)

J - A(x - S) R{i - rj)di]

[R^-B{X,Q) \

di].

g^O ya-*0 B^-B(x,a)

If we now set f = ;t: C on the right in (2.7.26), we obtain O lim jlim fS(x r]; Q)Ah[rj) drX a))da).

(2.7.27)
^{y'y Q) = fA(co)R{y
Rv-^Q

From the homogeneity, we conclude t h a t

(2.7.28)

S{y\Q) = \y\^-'S(rj;Ql\y\), ^i-yiQ)

rj = \y\-^y, = s{y>9)*

\rj\ = 1

Now, if \rj\ = 1 a n d \co\ > 3/2, then \rj co\ > \co\l}. If \7]\ = 1 and 1/2 < \(o\ < 3/2, then \A{co)\ < 2"Mo if h co] < 1/2 and R(rj co)

60

Semi-classical results

< Z"-!- Co if \r) -a>\> Finally, if 0 < ^ < 1 / 2 , 5(7-/; Q) = lA(co)R{f]

Bv-BiOA/2)

1/2, MQ being the max. of |Zl(co)| for |co| = 1. - co)dco + JA(a))[R{rj

B{0,li2)-B(0,Q)

- co) - R(7])]dco

(2.7.29) and as Q -^0, S{r}; Q) converges uniformly for ^ on i7 to S{rj). From the y analysis above, we see t h a t (2.7.30) (2.7.31) \S(y; Q)\ < C{v) MQ | y | i limS(y; Q) = S(y)

so t h a t we may let o* - > 0 and then ^ - > 0 in (2.7.27) to yield g{^) = I S(x- rj) Ah(rj) drj.

Now, in (2.7.29), we assume t h a t \r)\ is near 1 and 0 < ^ < 1/2. By defining
Sl{V'> Qy'^) ^ J ^ (^) ^{'^
Rv-B(Q,l/2)-B{rj;r)

~~ <^)^<^>

differentiating with respect to rj^" and letting r - > 0 , we find that 5 ( ^ ; Q) is of class C^ in r] and that
Ry-B{0,l/2)-B{r],l/S) B {r],l/Z)

(2.7.32)

+lA{a})[R,47]-co)

- R,4rj)]dco;

(0 < ^ < 1/2)

B{0,l/2)-B{0,Q)

in this development, we used (2.6.9) (c) for R^^^ In (2.7-32), we may let ^ - > 0 and conclude from the uniform convergence in rj for \r]\ near 1 that S^ C^{Rr {0}) and that (2.7.33) | V 5 ( y ) | < [Ci(v) Mo + C2{v) Mi] \y\-^ where Mi is a bound for |V^(co)| for |co| = 1. Thus, from (2.7.31), we deduce t h a t ~g(x) = lim JS,oc(x fj)h^oi(r])dr]. From (2.7.28), etc., we see that V 5 and Vi^ satisfy (2.7.12). From (2.7.23) and Theorem 2.7.2 we conclude that V A ^ Lp(Rp) with (2.7.?4) \\^m<C{p,v)lft. From (2.7-34), we see that V h can be approximated strongly in Lp by functions kn^ C'^(Rv) satisfying (2.7-34) uniformly so that, by Theorem 2.7.2 again, g is the strong limit (uniform on any bounded set) in Lp of functions gn so that \\gt<C*{p,v)\\VS\\i^-\\Vh\\l from which the result follows, since || VS||?_^ < C{v) {MQ + Mi) from (2.7.33)From Theorems 2.7.2 and 2.7.4, we obtain the corollary.

2.8. The maximum principle for a linear elliptic equation of the second order

61

Theorem 2.7.5. If A satisfies ( 2 . 6 . 9 ) , / C^^c{Rv)y and g is defined by (2.7.13), then g^ Cl(Rv) O Lp(Rr) for every p > 1 and

||g||o<C(r,^)(Mo + Mi)||/||o
where MQ and M\ are defined in the statement of Theorem 2.7A2.8. The maximum principle for a linear elliptic equation of the second order We conclude this chapter with the well known proof of this principle
d u e t o E . HOPF ([!]).

Theorem 2.8.1. Suppose that u ^ C^{G), that ^^^ 6, c and f ^ C^[G), c and f satisfy (2.8.1) (2.8.2) c[x) = 0, /(^)>0, x^G x^ G, and that u is a solution of the equation a'='^{x) u^o:^{x) + b^'ix) u,a(x) + c(x) u{x) =f{x), assumed to he elliptic on G. Suppose that x^^G and that u takes on its maximum value at XQ. Then u[x) = u{xo) for all x on the domain G. Proof. Suppose u[x) ^ U[XQ) = M. Then the set where u{x) < M is open. There is a ball B[x2, R)ClG such that u{x) <.M for x^B{x2,R) {^i}> where xi^dB{x2, R) and u{xi) = M. Finally, there is a ball B[xi,R[)CG with Ri < R. Let Si = B{x2,R) ndB{xi,Ri) and 5 , = dB{xi, Ri) - B(X2,R), so that StU Se = dB{xi, Ri). Then u{x) <, M s on Si and u{x) < M on Sg for some > 0. Now, let h(x) = e~yr^ e-y^^, r = \x X2\. Letting L 99 stand for a*^ 99^ a/5 + ^" (p,oc, we see that evr^Lh = 4y2 a'^^ix^' ~ xfi{x^ :vf) 2y[a^^6a^ + h''[x'' - %|)]. We can choose y so large that Lh[x) > 0 in B[xi, Ri). Finally (2.8.3) Let (2.8.4) h{x) < 0 on Se, h(xi) = 0. (3 > 0,

v(x) = u{x) + dh(x),

where 6 is small enough so that v(x) < Af on 5^. From (2.8.3), we see t h a t v(x) < M on dB(xi, Ri), v{xi) = M so t h a t v has a maximum at a point xs in B{xi, Ri) while L(v) > 0 there. But this would imply that (since all v^oc{xs) = 0) (2.8.5) a^'^ixs) v^oc,3{xs) > 0 but v^upixs) r)'rj^ ^0 yv = cl[x^ xl), C^ = clr)', for all rj. Now, we may define new variables y and w by the rotation w(y) = v{x)

62

The spaces H'^ and H^^

where the matrix c is chosen so that c"^ a(xs) c = d(0) is diagonal. Then (2.8.5) is equivalent t o (2.8.6)
y=i

2'^^^(0)'^\yy(0) > 0 b u t w^y6{0) C'^ C^ < 0 for all C

where all the d^y > 0. But the first inequality in (2.8.6) implies that some one w^yy(0) > 0 which contradicts the second. Corollary. / / u and the coefficients satisfy the conditions of Theorem 2.8.1 except that we require c[x) < 0, then u cannot have a positive maximum. If, also, f{x)-^E0, then u has neither a positive maximum nor a negative minimum.

Chapter 3

The spaces H"; and il^o

3.1. Definitions and first theorems In this chapter we collect statements and proofs of the theorems which we need concerning these functions. As we stated in Chapter 1, these a n d similar spaces have been discussed at great length by many authors and have certainly proved their worth in connection with the study of differential equations. Definition 3.1.1. A function u is said to be of class H^ on G iff u is of class Lp on G and there exist functions r^ of class Lp on G, a = oci, . . ., ^v, 0 < 1^1 < w, such t h a t (3.1.1) fg(x)r4x)dx
G

= {-iy-\fD-g{x)u{x)dx,
G

g^C-(G).

The functions ra (or rather the classes of equivalent functions determined by them) are called the distribution derivatives of u and r^ is hereafter denoted by D^" uovu,; we make the convention that 0" u = uii\(x\ = 0 . Remarks. I t is clear t h a t if u is of class H^ on G, its distribution derivatives are determined only u p t o additive null functions; moreover \i u* differs from w b y a null function, then ^ * is also of class H'^ on G and has the same distribution derivatives. The definitions above extend t o vector functions and t o complex-valued functions. The LEBESGUE
derivatives (SAKS p . 106) of the set functions T^a(A;) dx
e

are called t h e

generalized derivatives of u at points where they exist. Theorem 3.1.1. The space H'^{G) of classes of equivalent vector functions of class H^ on G with norm defined by

(3.1.2)

||i| = { /

pl2

] lip

dx\

(u = u^, . . ., u^)

3.1. Definitions and first theorems

63

is a BANACH space. 7 / ^ = 2, the space is a H I L B E R T space if we define

(3.1.3)

(^,^)?= f Z

Co^D'^u^Wvidx.

In (3-1-2) and (3.1-3), Co, denotes the multinomial coefficient \(x\\j(Xi\ . . .ocpl Proof. The only property requiring proof is the completeness. So suppose {un} is a CAUCHY sequence in H^. Then each of the sequences D^ Ufi, with 0 < 1^1 < w, is a CAUCHY sequence in Lp(G) and so converges in Lp{G) to some vector function roc(=u if | ^ | = 0 ) . But, for each oc with 0 < | ^ | < w, it follows that (3.1.1) holds in the limit so that the r^ are the corresponding distribution derivatives of u. Remark 1. I t will be convenient to call these elements of H^ functions and to say that u is continuous, harmonic, etc., iff some representative of the class forming the element has these properties. Naturally, also, there are many different topologically equivalent norms which could be used. Remark 2. The spaces H^{G) have been defined for all real m (see, for example LIONS [1]) and interesting results in the theory of differential equations have been obtained using these and other spaces such as those
introduced b y ARONSZAJN and SMITH [1], [2], CALDERON and others.

A discussion of these matters, though interesting, would lead us rather far afield into functional analysis and is not really relevant to our discussion. Accordingly, we shall not define them here. However, a special class is used in Chapter 8. Theorem 3.1.2. (a) Ifu^ H^(Q and D C G, then U\B ^ H^(D). (b) Suppose u is defined on the whole of G and each point of G is in a domain D such that U\D^ H^{D). Then there are functions r^ defined on G such that {0^ U\D) (X) = ro:{x) for almost all x on D, and any D C C G. (c) (d) (e) (f) / / , in (b), the roc^Lp{G), then u^H^(G). Ifu^ H^(G), then V^u^ H^-^(G), 1 < ^ < m. Ifu^H^{G) and V'^u^Hl{G), then u ^ H^+'{G). Ifu^H^iG) andg^ CZ'^G) (i. e. g^ CrHG)

and all its derivatives of order < m 1 are LIPSCHITZ, then (3.1-1) holds with roc = D'^ u ^ u^oc. (g) If u^Lp(G), u is absolutely continuous in each variable (on segments in G) for almost all values of the other variables, and if its first partial derivatives {which consequently exist a.e. and are measurable) ^Lp(G), then u^Hl{G) and its partial and generalized derivatives coincide almost everywhere. Proof. Parts (a) (e) are obvious and part (f) follows b y a straightforward approximation of the function g. To prove (g), we notice t h a t if g^ C^(G), then g[x) u{x) has the absolute continuity properties of u

64

The spaces H ^ and H^^^

and has compact support in G. From FUBINI'S theorem, it follows that we may take the ra as the partial derivatives dujdx^" in (3.1-1)We recall the definitions of a moUifier 99 and the 99-mollified functions UQ which were given in Chapter 1 (Definition 1.8.3). Theorem 3.1.3. Suppose u^H^(G), cp is a mollifier, and UQ is the (p-mollified function of u. Then UQ ->U in H'^{D) for each D C. d G and (3.1.4) IP(XQ(X) = D'Ue(x) if y)a = u^oc, 1<|^|<m. If m ~ \ and u^oc(x) = 0 a.e. on G, oc = \, . . .,v, and G is a domain, then u{x) = const, a.e. on G. Proof. If B {x, Q) C G, the function gx Q defined on G by gx,(^) = cpta - ^) ( y ) == Q-^cp{Q-^y))

^ C^(G). Hence, using Theorem B, 1.8 and the definitions, we see that D-u,(x) = f ( - 1)'^' ^^,,a(f) u{^) di = fgxQd)
G G

D-u(^) dS = ^/',(^).

The remaining statements follow from Theorem B, 1.8 and the connectedness of G. Remark. It is not known (and the writer believes it is not true) that a function u ^ H'^ on a domain G with sufficiently wild boundary can be approximated over the whole of G by functions of class C^ on a domain containing GU dG. We shall prove this, however, for a surprisingly wide class of domains in 3-4. The following theorem is an immediate consequence of Theorem 3.1.1, 3-1-2 b and c, and 3-1-3 and the details of its proof are left to the reader. Theorem 3.1.4. Ifu^ H^{G) ^^^ C ^ Cf~'^(G) and C and its derivatives are bounded on G, then C u^ H'^(G) and its derivatives are obtained from those of t, and u by their usual formulas. Proof. For if 99 is a mollifier and Ce and UQ are the 99-mollified functions we conclude that D^ ^Q converges almost everywhere and boundedly to D" C and D^" UQ converges strongly in Lp to 0" u on each D (Z G G, 0 < |(%| < m. Definition 3.1.2. A mapping T: x = x(y) of class C ^ or Cf~^ of a domain H onto a domain G is said to be regular iff it is 1 1 and all the derivatives of order < m of the mapping functions for T and T - i are uniformly bounded. Theorem 3.1.5. Suppose that x ^= x[y) is a regular mapping of class C^ of a domain H onto G, suppose u^ H^{G), and suppose v(y) = u[x{y)]. Then v^ H'^{H) <^f^d, if all of the generalized derivatives D'^ U{XQ) exist at XQ = x{;yo), then all of the generalized derivatives D* v(yo) exist and are connected with those of u at XQ by the usual rules of the calculus.

3.1. Definitions and first theorems

65

Proof. From Theorem '}A.'}, it follows t h a t we may approximate strongly to u in H'^'{A) for eachZl C C G. If Z) is the counter-image of such a A, the transformed functions converge strongly in H'^(D) to v. The last statement follows easily since a regular family of sets (see SAKS, p. 106) about yo corresponds to one about XQ. We now digress to give a simple proof of the famous theorem of
RADEMACHER concerning LIPSCHITZ functions.

Theorem 3.1.6. Suppose u satisfies a LIPSCHITZ condition on G. Then u possesses a total differential almost everywhere on G. Proof. Let ZQ be the set of points where one of the generahzed derivatives of u fails to exist. Let E be an everywhere dense denumerable set of points C on Z. With each such f, let x = xz[y) be a rotation which carries the vector ei in the y-space into f and let v^[y) = u[xc(y)]. The generalized (first) derivatives of v^ exist at all points y for which xc(y) ^ ZQ. Also, for each C, there is a set Z^ of measure 0 such t h a t the partial derivative dv^dy^ = Dyi v if xc(y) i Z^ (Theorem 3.1.2 g). If AJQ ^ ZQ U (U Z^), then all the directional derivatives in the directions C exist and are connected with the generalized derivatives D"^ u{xo) b y t h e usual formulas. At any such point, it is easy to see, using the LIPSCHITZ condition that u has a total differential. From this, we conclude the following generalization of Theorem 3- L 5 * Theorem 3.1.7. Suppose x = x{y) is a regular mapping of class C 7 - i ( m > 1 ) of H onto G, u^H^(G), and v{y) ^ u[x(y)]. Then v ^ H^(H) and if all the generalized derivatives 0^ U[XQ) exist and all those of the xy at yo exist, and if XQ = x(yo), then all the generalized derivatives D'^ v{yo) exist and are given by their usual formulas. Proof. Since x = x{y) is of class C^-^, v $ H^-^[D) on each D CCG. If m > 1, we conclude from t h a t theorem t h a t any derivative D^^ v of order \(x\ < m 1 is given b y a formula of the form (3.1.5) D^v{y) = j:A^^{y)u,p[x{y)]

where 1 < /5 < m 1 and the AQ are polynomials in the derivatives of the xy of order | ^ | + 1 | ^ | < m 1 and hence are LIPSCHITZ. Since the u^^ all ^ H^{G) at least, the theorem will follow from Theorem 3.1.2(g) and the special case m \. If UQ is a mollified function of u, defined on G^, if Z) C C the counter image of GQ, and v*{y) = Ue[x{y)], then v* is LIPSCHITZ and (3.1.6) vl4y) = u,,p[x(y)] ^x^Jy) (a.e.)

It is clear that we m a y let ^ - > 0 and conclude v^ Hl[D) with v^OL given by the limit of the right side of (3.1.6). If, now the generalized derivatives u ,^(XQ) and x^o,{yo) all exist and e runs through a regular family of sets
Morrey, Multiple Integrals c

66

The spaces H^ and H^Q

at yo and E is t h e image of e, these forming a regular family a t XQ, we conclude:

11 H " V [^ Ay) - ^ ,i8 (^o) ^"fa(yo)] ^y I

I e I

(3.1.7)

<\e\-^ f l^A^iy)] - ^,p(^o)\'Kiy)\dy +

+

IH~^ ^.)8 (ATO) 'f\x^^(y) - :^f(yo)

dy.

The second term on the right clearly tends to zero and the first does also since it is <,C\E\-^ f \u^p{x) u^^{XQ) I dx
E

where C depends on the LIPSCHITZ constants. We now prove two theorems concerned with the absolute continuity properties of functions ^ H\[G). Lemma 3.1.1. Each class of functions u in H\[G) contains a representative uo which for each oc is absolutely continuous in x"- on each segment in G with endpoints in GU dG for almost all values of x^ and UQ tends to limits as [x^, x^ tends to the end points of any such segment. Proof. If i^ = [a, b] is any rational cell in G, there is a sequence of values of ^ - ^ 0 such t h a t
lim { [|UQ{X^", X^) u{x', x^\'P + \UQ OC{X', X^) U X(X' X^)|^] dx = 0

(3.1.8) for almost sIL x^, oc = \, . , .,v. For any x^ for which (3.1.8) holds, the UQ are uniformly absolutely continuous and hence tend uniformly to an absolutely continuous function. The result follows by ordering the rational cells in G and choosing successive subsequences. Theorem 3.1.8. Each class of functions in H\{G) (C H.\{G)) contains a representative u which has the absolute continuity properties of UQ in the lemma and which retains these properties under a regular LIPSCHITZ mapping. In fact we may define U{XQ) as the LEBESGUE derivative of f u(x) dx for each XQ for which that derivative exists. ^ Proof. Since regular families of sets correspond under regular L I P SCHITZ maps, we see t h a t ^[^^(y)] = u[x{y)] and so it suffices to show t h a t u has these properties in some one coordinate system. Also, it is clear that u (XQ) is defined if there is an A such that
xo+h

(3.1.9)

\im{2h)-''

f\u{x)

A\dx = 0

in which case u (XQ) = A. Now, let Uo be a representative as in Lemma 3.1.1 and let [a, b] C
{A,B) C[A,B]C G. Using F U B I N I ' S theorem and L E B E S G U E ' S theorem,

3.1. Definitions and first theorems

67

it follows t h a t for each (x, there is a set Za of x^ of measure 0 such that if x'^^ is not in Za then U{s[x'^, x\^ is A.C. in [A'', B^^] and (3.1.10) lim ( 2 ^ ) 1 - ' ' / / l%,a(l) I ^f = / 1^0,(!, ^Lo)! ^ 1 "

for a given dense denumerable set of x"^ on [^'^, 5^^], including A^ and J5 and
xo+h

(3.1.11)

lim (2A)-^/ l^o(f) - uo{xo)\di

= 0 (;i:o = K , < o ) )

for almost every XQ on [^^, JS'^]. Since the functions of x"^ in (3-1.10) are all monotone and continuous, the convergence in (3.1.10) is uniform. Now, let XQ = (XQ, X^Q) be any point for which X'^Q is not iii Z^ and ^"^ < ^S < b^" and let > 0. Using the uniform convergence in (3.1.10) and the absolute continuity of uo{x'^, A;^Q), we see that we can choose an xf > x^ on {A"", 3") such that (3.1.11) holds at (xl, x^^) and such that

/|^0,a(|^<o)|^f"<l/3 (2h)-^I dx'fd^'='f\u,cc(i'', 0 | ^ < < /3. 0<h<ho.

Then (setting 6 = x'^ x^)

I a;o+/i xf^i+h x'xo+h [

( 2 A ) - / |wo(|) - uo{xQ)\di - {2h)-^J

I tco^ x^ih

/ |MO(I) - o ( 4 , < o ) | '^f I

x'xO~h I

< 1^0(^,40) - w o W ! + ( a ^ ^ ' / l M o d + d)~

Uo{i)\dS < 2e/3,

0<h<ho,
so t h a t (3.1.11) holds and u{x', X'^Q) is defined and = Uo(x', X'^Q) for all x"" on [^', 6]. Definition 3.1.3. A function u is absolutely continuous in the sense of ToNELLi (A.C.T.) on G if and only if u^H\{G) and u is continuous on G, Corollary. If u is A.C.T. on G, it has the absolute continuity properties described in Lemma }AA and if x = x{y) is a regular LIPSCHITZ map of a domain H onto G and v(y) = u[x{y)], then v is A.C.T. on G. Theorem 3.1.9. Suppose F ^ C^(Rp), suppose each U'P^H\ {G) for some A3? > 1, /> = 1, . . ., P , suppose U{x)=F[uHx),...,uP(x)],

(3.1.12)

68

The spaces H!^ and H^^

and suppose U and the Vu^ Lx(G) for some A < 1. Then U^ H\(G) and U^a{x) = Voc[x) a.e. / / the u^ are A.C.T., so is U. Proof. For each ^, let ^2 be a representative of the class U'P which ^> has the absolute continuity properties described in Theorem 3.1.8 and suppose7(:\;) =^F[u [x)]. Then tJ has these absolute continuity properties, is equivalent to U, and its partial derivatives dU/dx'^ = Fa almost everywhere. The result follows from the hypotheses and Theorem 3 1.2 (g). We conclude with two inequalities which will prove useful later (the inequality f v(x) d x\ <i f \v(x)\d x for vector functions v is useful); Theorem 3.1.10. Suppose z^H'^(G), D(ZG GQ, and (p is any nonnegative mollifier. Then (a) / 1 / l ^ i V^- ^(1) - V^z{x)\^f'^ <p*{i - X) dAdx < {Q \\z\\Z^)P

(b) \\^,\D-z\D\\ll,'<Q\\z\\Za iP>^). ZQ denoting the mollified functions. Proof. It is easy to see that (b) follows from (a) and (3.1.2). By regarding the derivatives D"^ z"^ with |ix| < m 1 as components of a vector, (a) is reduced to the special case where m = 1. There is a domain D' CC.G such that D CD'^, Using the mollified functions, we may approximate to z\i) by functions of class C^, so we may assume z^C^{D'). Then

\z{S)-z{xyf<^Qvj\\/z[x
0

+ t{^~x)]\vdt,

^^B(X,Q),

X^D.

Thus, we obtain

/I

j\z{^)~z{x)\vcp^^[^-x)dAdx

^ Qv] dt ^ (pliOll j \V z{x + K)\J>dx\d!: = gpfdtl^*{0\

0
BQ [D\^

f\Vz(y)\Pdy]di:<.(e\\z\\l^)P
J

if D denotes, for each fixed vector | , the set oi sXl y = x -\- i where x^D\ clearly Dt^ CD' CG for each [t, f) considered. 3.2. General boundary values; the spaces H^Q{G); weak convergence We begin by defining the spaces H'^^ (G): Definition 3.2.1. The space H^Q(G) is the closure in H^{G) of the set of functions C'^{G).

3.2. General boundary values; the spaces H'^Q{G);

weak convergence G(ZB(XQ,R)

69 and

Theorems 3.2.1 (POINCARE'S inequality). Suppose U^H^Q(G). Then (3.2.1)

G G

j \V^u[x)\i>dx <,p^-^R^rn-k)p j ^\jmu{x)'fdx,

0 < ^ <

m.

Proof. We shall prove this for w = 1, ^ = 0; the general result

follows easily by induction, since if U^H^Q(G), S/^ U^H^^^[G)

( = Lp(G) if ^ = m). From Definition 2.3.1, it suffices to prove the theorem in the case that u^ Cl[B{xo, R)]. Taking polar coordinates (;', C) (C on dB{0,\)) with pole at XQ and setting v(r, C) = '^(^o + ^ f), we obtain (3.2.2) R
T E

^[R-r)v-ij

j\v,r[sX)\^dsd^{^)

using the HOLDER inequality. The result (3-2.1) follows for m = 1 , ^ = 0 from (3.2.2) and the fact t h a t
B

j\u(x)\Pdx
B{Xo,B)

=
0

Jr''-^J\v(r,C)\^drd2JE

Corollary. The space H'^Q{G) is a closed linear subspace of H^(G) and, if G is hounded, the norm || u |^Q defined by (3.2.?) M7o = [j\'7^u(x)\vdxY''

is topologically equivalent to the norm \u\^ for u on H^f^{G). If p = 2, the inner product may be taken on H'^Q to be (3.2.4)
-

{^,'v)2o = j Z

2'^^,W^,='W^^

(u = u^,...,u^,

etc.).

For functions in H^ on arbitrary regions, we have the following theorems: Theorem 3.2.2. (a) Suppose u ^ H'!^Q(G) and V(x) = u[x) for x ^G and V(x) = 0 elsewhere. Then V^ H^[Rv) and V^ Hpo0) for any open set A DG. Moreover 0" V[x) = 0" u(x) on G and D"" V{x) = 0 for x^Rp G if 0 <.\(x\ < w [almost everywhere). [h) Suppose u^H^iG), DcG, v^H^(D), v - U\D^ H^oi^), U(x) = v{x) on D, and U{x) = u(x) onG~D. Then U^ H^{G), U - u^ H%{G), and Z)^ U{x) == Z) v[x) on D and Z) U[x) = D^ u(x) on G D ifO^\oc\ < w (a.e.). (c) If u^ Hl{G), E is measurable, E G G, and u(x) const, a.e. on
E, then Vu(x) = 0 a.e. on E (see MORREY [16], p . 254).

Proof, (a) follows immediately from Definition 3.2.1 and the theorems, of Section 3-1-I- If we define V{x) = v(x) u(x) on D and V(x) = 0

70

The spaces if^ and

H^^

elsewhere, then the conclusions of part (a) hold with G replaced by D. But then U(x) = u(x) + V{x) on G so the results in (b) follow. Part (c) follows by choosing an absolutely continuous representative u of u. Theorem 3.2.3. If p > 1, the most general linear functional in H'^{G) has the form (3.2.5) m = J S ZAt{x)u:,{x)dx

where the A^ ^ Lq(G) where p-^ + ^~^ = 1- ^/:^ = 1, ^^^^ linear functional has the form (3-2.5) i'^ which the A^ are hounded and measurable. Proof.It is clear t h a t any expression (3.2.5) defines a Hnear functional on H^\ Conversely, let Bp be the space of all tensors {99*} where each (pi^ Lp{G) with norm

lki= / 2" Sc^WUm

KQ Li=l 0^|a|:^m

dx\ .
J

Then, from Theorem 3.1.1, it follows t h a t the subspace M of all tensors (p where (pi = u^^ and u^ H^(G) is a closed linear submanifold of Bp. If we define Fi{(p) = f{u) for such tensors (p, we have ||i^i|| = ||/|| and i^i can be extended (Hahn-Banach Theorem) to a linear functional F over Bp with the same norm. But any linear functional F on Bp has the form

F{<p) = fS

Q i = l 0<|al^w.

ZAtix)<pUx)dx

where the ^ * have the stated properties. From Theorem 3-2.3, we immediately obtain: Theorem 3.2.4. (a) A necessary and sufficient condition that u^ converges weakly to u(un 7u) in H^{G) is that each component u^^^ of u^ converges weakly in Lp{G) to u\^. (b) / / Un-ru in H^'{G)^ then Un-7 u in H^(D) for DcG (i.e.
Un\ D 7 "l^l D)'

(c) If Un7 u in H'^{G), x = x{y) is a regular transformation ^ Cf ~^ of H onto G, Vn(y) = Un[x(y)], and v{y) = u[x(y)], then Vn7 v in H'^(H). (d) Ifun -V u in H^(G),C^ Cf-i(G), and all the D'^ C -^^ith 0 < | ^ | < w are uniformly bounded on G, then l^ Un7 i^u in H^{G). (e) If p'> \, bounded sets in H^{G) are conditionally compact with respect to weak convergence in H^{G). 3.3. The Dirichlet problem In this section, we interrupt our study of the spaces H^ and K^Q in order to illustrate the variational method for proving the existence of the solutions of certain differential equations. We also establish DIRICHLET'S principle.

3.3. The Dirichlet problem

71

Definition 3.3.1. If u and ^* ^ Hl(G) and u u* ^ HIQ(G), we say that u and u* coincide on dG or u and u* have the same boundary values on dG. li u^ HIQ{G), we say that u vanishes on dG. Remark. The content of the definition depends on p (but see Theorem 3.6.2). Theorem 3.3.1. (a) Suppose that u^H\{D) for each DCCG and satisfies (3.3-1) /C,a^,a^^ = 0,
G

^^Cr{G),

Then u is harmonic. (b) If u^ H\[^) ^^^ satisfies (3.3-1), then u minimizes the DIRICHLET integral (3.3.2) D{u,G) = f\Vu\^dx
G

among all functions with the same boundary values. (c) / / G is bounded and u* ^ Hl{G), there exists a unique function u = u'^ on dG which minimizes D{u, G) among all such functions. That function is harmonic. Proof, (a) The hypotheses imply that u is locally summable on G. Since the C,a <^r{^) i^ f ^o^s, we conclude from (3.3-1) and (3.1.1) that juAI:dx
G

= 0,

C^C'^'iG).

The result follows from WEYL'S lemma ( 2.3). (b) If u^HKG) it follows from Definition 3.2.I that (3.3.1) holds forallC^ ^lo {Q- Soif z^*^ H\{G),u'^ = uondG, andwesetC = u* ~ u so u* = u + C and C -^20 {Q> we conclude that D{u*, G) = D(u, G) + 2fC,ocU,ccdx + ^(f, G) > D{u, G)
G

unless C = 0. (c) Let {un} be a minimizing sequence and let f/j = ^* Un. Each Cn^HlQ(G) and D(Cn, G) is uniformly bounded. By POINCARE'S inequality IICwIII,^? and hence l^wfLc? is uniformly bounded. Thus a subsequence, still called Unyu. Since the norm of an element in a BANACH space {L2{G) in this case) is lower semicontinuous with respect to weak convergence, it follows that the DIRICHLET integral (=^(|i^,a||2.^)^) has this property so that D{u, G) < lim inf D{uny G). But u^H\ [G) and f%-7f = ^* ^ s o that C Hloi^) ^^^ ^ = '^** on dG. Thus u minimizes the DIRICHLET integral. So if f ^ HloiQ, u + ^C = u* on dG for all Z and D{u + ZC,G) =D(u,G) + 2A/C,a^,a^A; + A2 7)(C,G).
G

Since A = 0 gives the minimum, (3.3.1) holds and the results follow.

72

The spaces Hi' and H^^

3.4. Boundary values In this section we introduce the class of strongly LIPSCHITZ domains and prove that if G is strongly LIPSCHITZ and u ^ H^ (G) for some arbitrary w > 1, then u can be approximated in H^ (G) by functions each of class ^^{r) for some Fz^G. We then prove a variant CALDERON'S extension theorem for such domains and discuss bundary values on the boundaries of smoother regions. For the case m = \, some of the results are extended to ordinary LIPSCHITZ domains i.e. domains of class C\. An example is given of a LIPSCHITZ domain which is not strongly LIPSCHITZ. Definition 3.4.1. A domain G is said to be strongly LIPSCHITZ iff G is bounded and each point XQOI dG is in a neighborhood 3? which is the image under a rotation and translation of axes of a domain \y[,\ <.R, I>'*'!< 2L 2^ in which XQ corresponds to the origin, "St H dG corresponds to the locus of y^ = f (y^) where / satisfies a LIPSCHITZ condition with constant L, and 31 0 G corresponds to the set of y where \y!^\ <, R and

f(yl)<r<2LR.
Remark. A bounded domain is strongly LIPSCHITZ iff it has a regular boundary in the sense of CALDERON (p. 45). Lemma 3.4.1. Any bounded open convex domain is strongly hiFSCunz. Sketch of proof. Let B (XQ, R) be the largest sphere C G, the given domain. Let xi^dG. From the convexity, it follows t h a t any point interior to any segment xi x with x ^ B(xo, R) lies in G. If we choose axes so the y^ axis runs along the ray xi XQ, it is easy to see t h a t a part of dG near XQ can be represented in the desired form. We need the following well-known lemma: Lemma 3.4.2. Suppose f ^Lp(Rp), e is a unit vector, and fh is defined hfh{x) = f{x + he). Thenfh ->f in Lp(Rv). Proof. Let > 0. There is ag^C^c(Rv) such t h a t \\gh fh\\ = l k ~ / l l < ^ / 3 - Since g is also uniformly continuous, there is a ^ > 0 such that \\gh g\\< si} for \h\ <d. Theorem 3.4.1. Suppose G is strongly LIPSCHITZ and ti^H^{G). Then there is a sequence {un}y each^ ^^{Gn) where G (Z C. Gn, such that Un -^u in H^{G), Proof. GUdG can be covered by a finite number of open sets 9^^, each of which is either a cell with 3^^ C G or a boundary neighborhood of the type described in the definition. There is a finite partition of unity {Cj}, each of which has support in some one 3^^- and is of class C ^ everywhere; it can be arranged t h a t Ci{^) + * * * + CJ(^) = 1 on a domain FZ) GU dG. For each j , let Uj(x) = Cj{^) u{x). For each j for which Cj has support in G, Uj can be approximated by mollified functions which are 0 near dG and can be extended. So, let us consider a j where Cj has support in a boundary neighborhood 3t, and let Vj(y) = Uj(x) and for \yl\ < R, let us extend Vj{y) = 0

3.4. Boundary values

73

for y^>:2LR. Clearly, the function Wjn defined by Wjn{y) = Vj{y + n~^ Cp) H'^i^^) where 9^+ is the part of 9^ where >' >/(y;) - n-^ and also ^ '^jn ->^j- in H^{^+) where 3^+ is where y^ >/(yi). There is a sequence Qn ->0 such that Q^ <, (2L n)~^ and < distance of the support of Cj from d 31, and also so small that if we define Vjn = "Wongny then Vjn ->Vj on 9^+ and each Vjn^ C'^ on Slg^. If we define Ujn{x) = 0 for x not in 9^ and equal to the transform of Vjn(y) for x in 9^, we see that Ujn^C in a domain"Z)G\J dG. We define Un =2unj on their common domain. Lemma 3.4.3. Suppose 0 < s < r and S is a measurable set of finite measure. Then j \ x - y\-^dy^rv{v ~ s)-ic'-^ yvC'= | S | = \B{x, o)\, s Proof. Since \x y\~^ <: G~^ ii y i B(x, G) and \x 3^1"^ > a~^ if y^B(x, a) and since |5(A;, cr)| = \S\, it follows that f\x
S

y\~^dy <, f \x y\~^dy = Fviv s)~^o''~^.

B(x,a)

We next prove the following important theorem: Theorem 3.4.2. Suppose K is essentially homogeneous (see Def. 2.5.2) of degree m v {m'> 0) and ^ C^+i \Rv {O}] and suppose f is defined on Rp and u is defined by

(3.4.1)
Then (a.) (3.4-2) (3.4.3)

u[x)=JK{x~y)f{y)dy.
IffeQ{R.),ueC^{Rv),and D^uix) = j K,^{x y)f{y)dy, D''u[x) = Co^f{x) + Yim 0<||<w1 \oc\=m,

I K,o:(x-y)f(y)dy,

^ \y\ == ^ (b) / / / $ Lp(Rv) and has compact support C G C C Rv, then u ^ H^{D) for any bounded D, the formulas (3-4.2) hold almost everywhere, and \\u\\l^<C{v,N,m,p,D,K)-\\f\\l^. Remark. Formula (34-3) holds in (b) also, almost everywhere, but we shall not prove this. Further properties of the functions in (34-1) in case (b) will follow most easily from certain SOBOLEV type theorems to be given in 3.5 and 3.7. Proof. Part (a) follows from the methods and results of 2.5 and 2.6. To prove (b), let us assume that f^Lp. Clearly K^cc is essentially homogeneous of degree m \(x\ v.li this is > 0 , then K^a is continuous

74

The spaces H^ and H^^

and it follows easily by approximating to / t h a t D^" u is continuous and given by (3.4.2). It m \oc\ v = 0, then K^oc{x, y) = Clog|;\; y\ + -\- Kioc(x y) where C is a constant and Kix is positively homogeneous of degree 0. In that case, if _/> > 1, the right side of (3.4.2) is bounded by a constant times || / | | J ; Up = 1, it follows t h a t the right side ^ Lr for every r with r norm bounded by a constant times ]| / | | ^ . So, let us assume t h a t m \oc\ v = s, 0 < s < r . Let Ux denote the right side of (3.4.2). Then, for almost all x, \Uo:{x)\<M^,J\x-y\-^\f(y)\dx (34.4) \U4x)\p<Mf,^

f\x-~y
IG

{"^dyY ^ / I ^ y i"^ \f{y) \^dy

G

<Mf^i CP-^{V, S) 1 G 1(2^-1) (i-^/')Jj x y\-' \f{y) \Pdy by Lemma 3.4.3- From (3.44) and the preceding discussion, it follows t h a t II Uoc ||p,2) < C'll/JlJ for 0 < |(:x| < w 1. So, suppose D is a fixed bounded domain and we approximate / strongly in Lp b y functions fn^ C^(G). Then all the Unoc->Ux in Lp(D). Moreover, from the CALDERON-ZYGMUND inequalities, it follows t h a t the D^" Un converge strongly in Lp{R) to some limits f/a when | ^ | = m. The results follow. Theorem 3.4.3. (Variant of CALDERON's Extension Theorem) Suppose G is strongly LIPSCHITZ and G (ZG D. There is a linear hounded extension operator @ which carries H^{G) into H^Q{D) and which has the property that if V = (^u, then v(x) = u(x) for x^G. Proof. We begin as in the proof of Theorem 3.4.1 assuming, as we may, t h a t each 3^^ C Z). For each j for which Cj has support C G, we define (Bj u as the extension of Uj{x) t o D obtained b y setting Uj(x) = Oior x^D G. For a j for which ^j has support in a boundary neighborhood ?ft, we define ^ u first for u^C^ [G) and show t h a t it is bounded. Then @ u = 2JJ '^' To define j u for such j and u, we define Vj{y) Uj\_x[y)'] for y $ 9 t + (notation of the proof of Theorem 3.4-1), extend the function / to the whole space t o have LIPSCHITZ constant L and extend Vj{y) = 0 for y* >- fiy'p) where it is not already defined; clearly Vj{y) ^ C^ in t h a t domain U. Now, let y ^ C7 and let f be a unit vector with | f | <: L-^ C" ^ and let w{r, f) = Vj{y + r C)- Then
0

(3.4.5)

t^dO) =f'0^D'-w^{r.
4: LB

C}dr = V](y).

Let us choose a function a)j{z) which is homogeneous of degree 0, of class C^lRv {0}], which has its support in the cone 2:*' > L | 2:^ j and is such that its integral over ^ is 1. Multiplying (3.4.5) by co(C) and integrating

3.4. Boundary values

75

over 2! we obtain v^{y) (3.4.6) vf{z)

la|=m

)dz

ziU

K4z)
m V.

^ J z \-^
V" (m

2'C<xZ'^a}j\ (2) w ' \x\=m

(Coc = | a | ! / ^ i ! . . ,avl)

where X ^ C^ [Rv {O}] and is positively homogeneous of degree To define &j u, we begin by extending Vj to the whole space by (3.4-6) and letting uf{x) = ^^{y). Then uf{x) = Uj{x) for x^G and uf ^ H^(A) for any bounded domain A, b y Theorem 3-4.2 with \\ur\\:_,<C{v. Finally we define uf^ = (S; u by uf'^{x) = C{x) uf(x) where C C'^(Z)) and C{x) = 'I on G. The result follows. A special case of the following Theorem was proved by RELLICH. Theorem 3.4.4. Suppose G is strongly LIPSCHITZ and w > 1. Then bounded subsets of H^'{G) are conditionally (sequentially) compact as subsets of H^^-^G). If Un-7U in H^{G), then Un->u in H^'^{G). The theorem is true for any bounded domain G if we replace the spaces H'^(G) and H^-HG) by H^^iQ ^^^ H^o^{G), respectively. Proof .J Suppose GdC D CdRvy Gc DQQ, ( is the extension operator of Theorem 3-4.3 and Vn &Un. Then, from Theorem ')AAO we conclude t h a t (3-47) II Vn, - Vn WZa' <Q\\vn \\ID ^MQ. First, let us suppose t h a t ^ > 1. Then, according to Theorem 3.2.4(e), g a subsequence, still called {vn], such t h a t Vnzvm H^{D). It follows from the formula in Definition 1.8.3 for VnQ and VQ and the weak convergence that D'VnQ converges uniformly to 0" VQ on G, for 0 < | ^ | < < w 1. Then, taking norms on H^~'^{G), we obtain
^ Un '^ II ^ II '^n '^TiQ II + II ^n Q ^ e II + || ^ e '^ || = V\G, Un = Vn\G)y

N. m. p, A)-

\\u\\Z^.

<2MQ (UnQ = VnQ\G>

-\-\\UnQ Ue\\(u '^Q = VQ\G)

using (3.4.7). The result follows easily. If ^ = 1, all the D^^ Vnq are equicontinuous and uniformly bounded so t h a t a subsequence, still called {un} exists so t h a t the Z)* Unq converge uniformly to some functions cpocQ for each of a sequence of ^ - > 0 . Since (3.4-7) still holds the sequences

76

The spaces H^ and H^^

<^ae(|^| <= ^ ^) form a CAUCHY sequence in Lp{G) and so tend to limit functions (fa in Lp(G). If we set u = cpx when |<%| = 0 then we see t h a t u^H^~^(G), (foe = ^i,ocy q^xQ = UQ^OC aud (3.4.7) holds for UQ and u. The result again follows, but now, we do not know that u ^ H^. However, if Un -7 u in H^, then this u must be the same as the preceding one. The last statement is now evident since we may assume G C. D where D is
strongly LIPSCHITZ.

Remarks. It is clear from Theorem 3.1-5 and 3-1-7 (change of variable theorems) how to define the elements of the spaces H^{Tl) for a compact manifold Tl, with or without boundary, of class Cf ~^. For a given finite covering U of 9Jl by coordinate patches ri with domains Gi in Rv and ranges 3^^ in 9D^ one can define a norm by

(3.4.8)

(II u \\;\)p =2'(ll ^i Wla,)^- W =

i

<Mx)]

for example and any two such norms are topologically equivalent. In the case p = 2, the corresponding inner product would be

(3.4.9)

{u, v)^==2:f

i Q

0:^1 a l^m

Z(^-J^"^i^"ndx,

The definitions can be carried over for tensors on Wl, but then, usually, Wl must be of a higher class than C^~^ since the relations between the components in different coordinate systems usually involve at least the first derivatives of the coordinate transformations (see Chapters 7 and 8). Also, an extension can be made to non-compact manifolds which possess coverings in which no point of W is in more than K of the 3t^, K being independent of the point; but then the topological equivalence of two norms (3-4.8) is not usually true. If G is a bounded domain of class Cf ~^, then we may consider ^G as a manifold of class Cf "^, the structure being defined by the mappings allowed in 1.2, Notations. Theorem 3.4.5. / / G is bounded and of class C^~'^ the functions u ^ Cf ~^(G) are dense in any space H^{G) with ^ > 1 and there is a bounded operator B from H^{G) into H^'^dG) such that B u = U\QQ whenever u^ Cf-i(G). If Un-^u in H^{G), then Un -^u in H';-^{dG). If p > 1, the mapping B is compact. Proof. To prove the first statement, we select a finite covering of G by neighborhoods 9^^, each of which is a cell with closure interior to G or is a boundary neighborhood in the sense of the definition in 1.2, Notations; in the latter case we suppose that 9^^ is mapped onto / \ U 0*1 by a regular map x = Xi(y) of class Cf"^, A being the set 0 < y" < 1, |y^| < 1. We select a partition of unity {fs}, s = 1, . . ., 5, of class Q - ^ ( G ) , each Cs having support in some 9^^. Now, suppose u^H^(G), let Cs ^ = i^s, and let Vs{y) == Us[Xi{y)] for y^Fim case 3^^ is a boundary neighborhood. For those s for which 3^^ C G, we can approximate to Ug,

3.4. Boundary values

77

by {uns}, in which each Uns^ ^TiQ- ^ ^ r the other s, we can approximate to Vs by similar Vns on / i U cfi, using Theorem 3-4.1; clearly each v^s niay be chosen to vanish near G D dFi since v does. The first statement follows b y setting Un ==^Uns where, of course Uns[xi{y)] = Vs(y) when y ^ A and Uns{^) = 0 elsewhere. Next, suppose u^ Cf~^(G) and let % and Vs be defined as above. We define BsU = Us\eQ. Then, clearly Bu = u\Qg. For each s for which ^iG G, BsU = 0. So let s be one of the remaining indices; evidently Vs = Us u, Us being bounded. Since Vg^ Cf~^(A), we have l\l\'7^vs(y\ (3.4.10) < {yn^-^fflz\'^'^s(v'> lim(^) = : 0 ,
Q-*'0+

yl) - V^'Vs(0, yl) \^\ ' dyl yl) \^Y'^dyldri-<e(y-) = \\vs\\:^,r.

0<^8(r)<8(\)

SO that ||^s(0, y^) llr.^i^ ^^ evidently bounded, since y*' is arbitrary in (3.4.10). For u^H^(G), it is easy to see t h a t BgU = Ws where Wg = 0 on dG ^i and Ws is the transform of Vs{0, yj,) on dG O^t where Vs is an A.C. representative of Vs (Theorem 3.1.8). Now, suppose Un-yu in H'^(G). Then each Vns^'^s and (3.4.10) holds with a function 8[Q) which is independent of n, since in casep = \, the set functions f \V^Vns\'^dy are uniformly absolutely continuous.
e

From the fact (Theorem 3-4.4) t h a t Vns-^'^^^s in H^~^{^i) (since A is strongly LIPSCHITZ) it follows easily from (3.4.10) t h a t the limiting values ^ns(0, yl) -^Vs(0, yl) in H^~'^(ai), The last statement now follows from Theorem 3.2.4(e). The case m = \ in Theorem 3.4.4 can be generalized to LIPSCHITZ domains, i.e. those of class CJ. Theorem 3.4.6. If m= \, Theorem 3.4.4 holds for LIPSCHITZ domains, i.e. those of class CJ. Proof. Using the notation of the proof of the preceding theorem, we conclude, since the interior 3i% are strongly LIPSCHITZ as is / \ , that if {un} is any bounded sequence in H^(G), then a subsequence {q} of {n} exists such t h a t all the Uqs ->Us in H^~^{G) if % C G and Vqs ->^s in H^~Hri) and hence Uqs-^Us in H^-HG). li Un-ru in /f^(G), then Uns->Us and Vns ->Vs- The theorem follows. An example. That not every LIPSCHITZ domain is strongly LIPSCHITZ is seen by the following example of a bi-LiPSCHiTZ map in the case v = 2. Let {r, 6) and {R, 0) be polar coordinates in R^. We define the mapping T:R = r, 0 == 0 logr, O < 0 < 7 7 : , 0<y<1, T (origin) = origin

78

The spaces H^ and H^Q

By computing the derivatives Xx, etc., where X = R COS0, Y = R sin, x = r cosO, y = r sinO one easily verifies t h a t the map is bi-LiPSCHiTZ. Remarks. We have not proved an extension theorem hke Theorem 3.4.3 ioT m = i with LIPSCHITZ domains. 3.5. Examples; continuity; some Sobolev lemmas

As was pointed out in the introduction ( 1.8), the functions in H^{G) have been studied from many different points of view by many writers. We shall not go deeply into real variable properties of these functions here but will merely give a few examples to indicate the generality of the functions allowed and then shall prove some SoBOLEV-type lemmas which will indicate when such functions are continuous or have several continuous derivatives. I t is easy to verify t h a t the function / defined by (3.5.1) f(x) = \x\-f^, x^B(0,i), (h + m)'P<v, ^ H^. From this, it can be easily shown t h a t any function / defined by f[x) = J \x S\^^djLi{e),
G

{h + m)'p

<v

where G is bounded and /^ is a finite measure on G, ^ H'^ on any bounded domain D. (3.5-1) shows also that, for a given m and p, the wildness of functions of class H'^ increases with the dimension v. We now prove a SoBOLEV-type lemma guaranteeing continuity. Theorem 3.5.1. Suppose u ^ H^{G) where G is hounded and strongly LIPSCHITZ and m > v\p. Then u is continuous on G and there is a constant C{v, m, p, G) such that \u(x)\^ (3.5.2) C-\G\-^IPI^^\\V}U\\O ^-^ (A = diam G). + (m-vlp'n^^,\\ V^!"}

/ / G is also (see Lemma 3-4.1) convex, we may take C = 1; u may he a vector function. Proof. If G is strongly LIPSCHITZ, U may be extended to ^ H'^Q{D) where Z) is a given open set D G with || u|j^j) < Ci(v, m, p, G, D) - \u\^(} and then may be extended to be 0 outside D. So, by allowing a factor C, we may replace G by its convex cover. Since u may be approximated in H'^[GQ) by functions ^ C^, for each ^ > 0, we may assume G = GQ and u^ C^(G) and then let ^ - > 0 in (3.5.2). By expanding in a TAYLOR'S series with remainder about each y in G and then integrating with respect to y over G, we obtain (for notation.

3.5- Examples; continuity; some SOBOLEV lemmas see 5.7 below)

79

(3.5.3) \G\'u{x) =^zu^rH-^y I '^^<y)'{y - ^ydy +

1

m.-i

_ | - y ( _ > l ) m [ ( ^ 1)!]-^i^^~M fv^u[x+

t(y x)]'(y

x)^dy

dt.

The inequality follows by applying the HOLDER inequality to each term, first setting z = x + t(y x) to handle the last integral; when this is done the last integral is dominated b y
1

[A^l(m

\)\]'Jt^-^-'^l 0

[Oixj)

J\V'^u(z)\dz\dt J

where G(x, t) consists of all z = x + t(y x) for y^G and \G(x, t)\ = f\G\. Corollary. If u ^ H\[G) with p '> v, then u is continuous. We now present a "DIRICHLET growth'' theorem guaranteeing continuity (MoRREY [4] and [7]): Theorem 3.5.2. Suppose u ^ Hl[B(xo, R)], 1 < ^ < r, and suppose (?-5-4) f \\/ u\p dx < LPirldy-p+Pf",
B{x,r)

0 <.r <d = R ~ \x ~- xo\,

0 < ju < \ for every x^B(xo, (3.5.5) for R). Then u^ C^[B{xo, r)] for each r <, R and \u{^)-u{x)\<CLd^-^'P{\i-x\ld)^ \^-x\<dl2, C= C(v,p,fi).

Proof. By approximations, we may assume u ^ C^{BR). Let x and f be given, let ^ = |f x\l2, and x = {^ + x)l2. For each 7]^ B{x, Q), we observe that
1

u(r)) - u(^) = (rj- - |) / ^ , a K + Hv - f)]^^

0 1

\u{7]) u(^)\

< 2 ^ | | Vu[i
0

+ t(fi ^)]\dt.

Averaging over B{x, Q) we obtain (3.5.6) \B{x,Q)\-if\u{rj)-u{^)\d7i

B (a.
Q)

:.2Q\B{X,Q)\-^I

\j\s7u[i
B(X,Q) LO

t{7i-^)]\dt drj.

Interchanging the order of integration, setting y I + ^(^ I) and noting t h a t y ranges over B{xt, tg) where xt = {i t) i -\- t x which is at a distance dt >: S 2Q -\- tQ '>dl2 from dB{xQ, R) and then using

80

The spaces H^ and H^^

(3.5-4) and the HOLDER inequality we obtain (3.57) l\Vu(y)\dy<

Bixutg)

CiL{dl2)^-^-^lP{t

Qy-^+^

Ci = C[v, p).

Thus we see t h a t the right side of (3.5.6)

1

(3.5.8)

< 2C2 Q^-''L{dl2)^-'f'-''lPQ''-^+^ftf'-^dju,

0

C2 = C2(v, p).

Using the same result for the average of \U[YJ) U[X)\, we obtain the result.
We now prove another SOBOLEV Lemma (SOBOLEV[1], N I R E N B E R G [ 3 ] ) :

Theorem 3.5.3. Suppose u ^ Hl{Rv) with \ <= p < v, and has compact siippOYt. Then u^ Lr{Rv), where r = v pl{v p) and

(3.5.9)

I r < tri{\\.4i?''^v-"^t\\^M\i
a=l

i <p <v, t=p{v \)l{v ~p);

u may he a vector function. Proof. I t is sufficient t o prove this for u^ Cl{Rv). The case where p '> \ can be proved b y applying the inequality (3-5.9) for ^ = 1 to the function v defined b y v[x) = \u(x)\*. For, if this is done, we obtain (3.5.10)
->

f\u{x)\rdx=

''

f \v{x)\^dx
a=l

< fj {Wv^ocfi)'^"
a=l

<i^(\\u\^yHP-i)lp]J{l\u

40)^1",

(s = vl{v - 1))

since r = p[t \)j{p \), The first result in (3.5.9) follows easily since r [r s(p \)lp] = s and the second result follows from an elementary inequality. We shall prove the inequality for ^ = 1 and v = } , t h e proof for r = 2 or r > 3 is similar. Clearly
00 00

\u(x,y,

z)\ < / | ^ , i ( f , y, z)\dS,

/ I ^,2(^,^7, z)\d7],

0 0

J\u^s(x,
Thus u{x,y,z)\si^^[
~
00

y, OI^Cf\u,i{i,y,z)\di\
M/2 \l/2 / ~ \1^2

j\u,2(x.ri,z)\drj]

I f \u,3{x, y, 0\dU

3.6. Miscellaneous additional results

81

Integrating first with respect to x then y, and using the SCHWARZ inequahty, we obtain J \u{x, y, z)\^l^dx
OO

<l

I \u^i[S, y, z)\d\
\CO
/

I j j\u^2{x>
\OO

'^J, !^)\dxdr]

J J \u,2{x, y, C) IdxdCj
OO / OO

\ i ; 2

ff \u(x, y, z) l^l^dxdy
OO

< j ff \ u^i{S, y, z)
\ CO

\didy\
/

j J \^l',2(^,'r], z)\dxdrj\
OO /

[j J J \^,z{x,yX)\dx
\ OO

dy d^\
/

from with the result follows by integrating with respect to z. Theorem 3.5.4. There is a constant C[v) such that if u ^ Hl[B{xo, R)], there exists a function U^HIQ[B{XO, 2 R)] such that U{x) = u(x) for x^ B(xo, R) and \\^U\\l,j,^C-'\\u\\l^, where {'\\UIIJ,)P =J{\V
BR

u\^ +

R-^u^)Pl2dx.

Proof. From considerations of homogeneity, it follows t h a t it is sufficient to prove this for the unit ball B(0,i). But then the result follows form Theorem 3.4-3 (the extension theorem). Theorem 3.5.5. IfG is LIPSCHITZ and u {a vector) ^Hl{G), \<p <v, then u^ Lr{G), where r = v pl{v p) and (3.5.11) \\u\\?<C(v.p,G)-\\u\\l.

If u^ Hl{G), then (3.5-11) holds for each p <Cv. If G = B{xo, R), we may replace (3.5.11) by (3.5.12) \\u\\?<C{v,p)-'\\u\\l. Proof. The last statement follows from Theorem 3-4-6 (the Extension Theorem for m= \), If u^Hl(G) then, of course u^Hl{G) for each ^ < 1^. In order to prove the first statement, we cover G by neighborhoods 3li and choose a partition of unity {Cs} as in the proof of Theorem 3-4-5. For each 5 for which the support of ^sClG, (3.5-11) holds for Ug, For the remaining s, we may extend Vs(y) by positive reflection {vs{y^, yl) = Vs{ y'^, y'v)) and then Vs has compact support in \y^\ < 1, \y'^\ < 1. So t h a t (3.5.11) hold for such t;^ and hence also for the corresponding Ug. 3.6. Miscellaneous additional results In this section, we include a selection of theorems which are useful in discussing boundary value problems in the calculus of variations and a few results useful in the case of differential equations of higher order.
Morrey, Multiple Integrals 6

82

The spaces H'^ and H^^

Theorem 3.6.1. Suppose u ^Hl(G) and u is the absolutely continuous representative of Theorem 3.1.8. Then (a) if R = [a, b] is any cell in G, U\QJ}, is a representative of BR U, the boundary value operator of Theorem 3-4.5 for R\ (b) if x = x[y) is a &^'-LIPSCHITZ map of a part '31 of G onto FR U OR in which OR corresponds to 31(1 dG, v{y) = u[x(y)], and (p{yl,) = -^[^(0, y^)] for yl^OR, where yj = B u, then ^(y^ y^) ^ ^2?tei?) fo^ ^^^^ y^> varies continuously with y^ as an element of LJ,{GR) and, for almost every y^, ^(y^ y'v) - ^ ^ ( 0 . y'v) ^^ y" ^O"^ andv(0, y^) is a representative of(p(yl). (c) / / B{xo, R) CZ G, then U\QS{X^^R) is a representative of B u, B being the boundary operator for u; u\QB{xo,B)^Hl[dB(xo, R)]- for almost all R for which B(XQ, R) C G. Proof. This follows from Theorem 3.1-8 combined with the proof of Theorem 3-4.5. Theorem 3.6.2. Suppose G is LIPSCHTTZ. Then U^HIQ(G) ifu^Hl(G) and Bu = 0 a.e. on dG. Consequently, if U^HIQ{G) 0 Hl{G), then Proof. The second statement follows from the first. To prove the first, cover G by neighborhoods 9^^, choose a partition of unity I^s, and define Us and?:;^ as in the proof of Theorem 3.4-5- By using the mollified functions USQ, we see t h a t those Ug for which Cs has support interior to G ca,n be approximated strongly in H\{G) by functions ^ C^(G). So let s be one of the other indices. From Theorem 3.6.1, it follows t h a t the function Vs{y) extended to be 0 for y" < 0 $ 77^(71 U J'f). From Lemma 3.4.2, it follows t h a t Wn -^Vs in Hl{ri) if ^;^(y^ y^) = Vsiy" n-^ h, yl) and the support of each Wn ^Fiii h is small enough. By mollifying the Wn, one obtains a sequence % , each ^C^(ri) such that Vns->^s in HKFi), and hence one obtains a sequence Uns^Lip^{G), etc. Theorem 3.6.3. (a) Suppose G is LIPSCHITZ, u^Hl(G), and u is continuous on G. Then U\QQ is a representative of B u. (b) Suppose G is LIPSCHITZ, Unru in H\[G), B Un has a continuous representative (pn for each n, and xpn -^cp uniformly on dG. Then (p is a representative of B u. Proof, (a) follows from Theorem 3.6.1. (b) follows from Theorem 3.4-5Theorem 3.6.4. Suppose G is LIPSCHITZ. Then there are constants Ci(v, p, <T, G) and C2(v, p, r, G) such that f \u\p dx < Ci\ f \V u\v dx + i f \ B
G

u\dSY]
) \

[G

\a

J\u\Pdx G

<C2\f\Vu\Pdx
[G

+
IT

if\u\dxY] JJ

for every u^ Hl{G); here a and r are subsets of positive measure of dG and G, respectively.

3.6. Miscellaneous additional results

83

Proof. We prove the first, the proof of the second is similar. If the statement were not true, there would exist a sequence {un}, each Un ^Hl{G) such that (3.6.1)

j\un\Pdx>

n\j\Vun\^dx
[G

+ \j
\O

\Bun\dSYy

J J

On account of the homogeneity, we may assume that ||^w||^^(? = 1 for each n. Thus there is a subsequence, still called {^^71}, such t h a t Un -> UQ in Lp{G) and Bun->0 in Lp{a) (Theorem 3-4.6). Thus (see (3.6.1)) Vun-^0 in Lp{G), so t h a t Un ~^UQ in Hl(G) so t h a t B Un-^B UQ in Li(or), on account of Theorem 3.4.5. But, from Theorem 3.1-3, it follows that uo=^ const, which must be 0 since Buo^Oona. But this contradicts the fact t h a t || Un ||J = 1 for each n. Theorem 3.6.5. Suppose G is LIPSCHITZ. Then there are constants Ci{v, p, G) and C2{v, p, G, c) such that

Q G

{ \u\'P dx <, Ci ( \\J u\P dx

if if

u = \G\~'^ f udx = 0

j\u\'Pdx<,C2J\Vu\'Pdx

|5|>c|G|

(c > 0)

for every u^ Hl{G); here S is the set where u(x) = 0. In case G = B(xo, R), we may write Ci(r, p, G) = Ci(v, p) RP, C2(r, p, G, c) = C2{v, p, c) RP.

Proof. Again we prove only the second, the proof of the first being similar. The last statement follows by homogeneity. If the statement were not true, there would exist a sequence {un} such that ||^^||J=: 1 and Un->uo in Lp{G) and

f \un\Pdx

> n J \\/Un\Pdx,

\Sn\>i

c\G\>

for each n. Thus Vun - > 0 in Lp{G) so Un ->uo in H\{G) and wo = const, on account of Theorem 3.1.3 ^^^ ^0 ^ 0 since V^o = 0 but ||^o||J = 1But since Un ^= 0 on Sn, we conclude from the convergence t h a t 0 = lim \ \un UQ^P dx =^ lim j \un % \P dx = lim | ^/o |^ * | 5,^ | which contradicts the fact that \Sn\ >: c\G\ > 0 for all n. Theorem 3.6.6. Suppose G is a LIPSCHITZ domain and where p '> v. Then u ^ C]^(G) and (3.6.2) h,{u,G)^C[v,p,G)'\\ull, fz=\~vlp.

u^Hl(G)

The corresponding result holds if G is replaced by a compact LIPSCHITZ manifold {i.e. one of class CJ) with or without boundaryy the norm being any one of the norms discussed in the remarks after Theorem 3.4.46*

84

The spaces H^

and

H^^

Proof. We prove the first, t h a t of the second being similar. The theorem follows for any interior domain D from Theorems 3.5.1 and 3.5.2 and the HOLDER inequality. The theorem follows for G by covering G with neighborhoods 9^^-, choosing a partition of unity {Cs}> and considering the functions Us and Vs of Theorem 3-4.5- Since Fi is convex and each interior 3^^ is also, the theorem follows for each Us on G. Theorem 3.6.7. Suppose, for each n, that Un ^ Hl(G) with p '> \ and is continuous on G. Suppose also that 1|V ^7^1^,6? ^'^ uniformly bounded and Un converges uniformly to u on G. Then u^ru in Hl{G). If G is LIPSCHITZ, the same result follows if each Un is continuous on G and Un converges uniformly to u on each D G G G. Proof. We prove the first statement; the proof of the second is similar if one makes use of Theorem 3-6.4. The hypotheses imply that \\un\\l g are uniformly bounded. Let {q} be any subsequence of {n}. There is a further subsequence {r} of {q} such t h a t Ur 7 some ^0 in ii\(G) (Theorem 3.2.4(e)). But then we conclude from Theorem 3.4.6 that Ur-^uo in Lp{D) for each LIPSCHITZ D G GG, SO t h a t u = UQ. Hence the whole sequence Unr u in H\[G). Theorem 3.6.8. (a) Suppose (p ^ Lp{G), D G Ga, e is a unit vector, and ipji is defined on D hy
1

ipfi{x) = \ (p[x -\- t h e) dt,

0

I h\ < a.

Then fh^ ^P{^) fo^ ^^^^ ^^^^ ^ ^'^^ '^h-^(p i'^ I^pi^)(b) Suppose u^H\{G), D GGa, Cy is the unit vector in the %y direction and vji is defined on D hy i^hi^) = h~'^[u(x -\- h Cy) u{x)]. in Lp{D). Then vu Proof, (a) Let F(x, u) = (p{x -{- u e), \u\ <i a. Then F ^ Lp{D) each u and is measurable in {x, u) over D X {a, a). Also
a a

for

f \ f \F{x, u) \Pdx]du -a [D I

= f \ f \ (p(y) \Pdy] du<2a -a iDiu) J

f \ q)(y) \Pdy < ?

since D(u) G G for each u on {a, a), D{u) being the set oi y == x + u for u^(a, a). Hence F^Lp{F), T = (a, a) X D. Now
'fPh [x)

= J F{x, th)dt

= h-^ f F(x, u) du

so y)h(x) is defined a.e. on D a n d $ Li(D) by FUBINI'S theorem. Also \Wh(x)\^<h-'^

f \F[x,u)\'P du

lb

3.6. Miscellaneous additional results

85

Consequently ipn^Lp for each h] if h === 0, iph{x) = (p{x). Now, from Lemma 3-4.2, it follows that lim f \F(x, u) (p(x) \Pdx = 0.
u-t-O J)

Accordingly, we see that

0

Wh{^) (p(x) = h~^ J [F(x, u) (p(x)]du

h

f \ fh{x) (p{x) l^dx <h~^ J i f \F(^> ^) ^{^) \^dx]du D 0 ID I from which the result follows. (b) Let u be the representative of Theorem 3.1.8. Then vji is given almost everywhere by
1

'^h(^) = f u^y(x +
0

they)dt.

The result follows from part (a). We conclude this section with three theorems which are useful in the study of differential equations of higher order. There are a great variety of additional theorems like those above and those below which can be proved by the methods of this section. Theorem 3.6.9. Suppose G is bounded and strongly LIPSCHITZ, m > 1 and p '> \. Then, for each j , 1 <,j<,m 1, and each > 0, there is a constant C {v, m, p, G, j , s) such that (?.6.3) \\V^u\\l<e\\V"^u\\l + C\\u\\l u^H^{G). and Proof. For otherwise there is a sequence {un} with Un^ H'^iQ I] Un 11^ = 1 such t h a t
(3.6.4) I VHin I t > 4 ^^Un \\l + n\\ Un i|^

From Theorem 3.4-4, it follows that we may assume t h a t Un > u in H^-^(G). Since f V^'^^[J < i|^^!|^-i < ||^^||^^ = 1, it follows t h a t Un->u in Lp(G), so u = 0. But then || \7^Un\\^ - ^ 0 so that jj V ^ Un\\l and hence || Un \\^ -^ 0. Theorem 3.6.10. / / u^H^(G), there is a unique polynomial P of degree < w 1 (or 0) such that the average over G of each D"-{u P) is 0 if 0 < 1^1 < M 1. This is easily proved by induction on m. Theorem 3,6.11. Suppose G is a strongly LIPSCHITZ domain and G C B[XQ, R). Then there is a constant C{v, m, p, G) such that \\V^ u\l <. C R'^-qSl'^u\l, for every u^H^^{G) 0 < y < w - 1 such that the average over G of each D^ u is 0 for

0 < k l < w 1.

86

The spaces IT^ and H:^^

Proof. We first assume t h a t G CB[0,\). Suppose the theorem is not true. Then there is a ; , 0 < y < m 1, and a sequence [un} of functions of the type described such that (3.6.5) l|V^-^^K>^||V^^^||, \\unt=^\. A subsequence, still called {uri) converges in i7^~^ {G) to some u. From (3.6.5) it follows that V^Un-^O in Lj,{G) so t h a t Un->u in H'^{G) and V*^ ^ = 0. By induction on m and t h e use of moUifiers, it can be shown t h a t ^ is a polynomial of degree < w 1. Clearly the averages over G of each i)'^ is 0 if 0 < | ^ | < m 1. From Theorem 3.6.10, it follows t h a t u == 0. Since ^^^ - > ^ in H'^{G), this contradicts (3.6.5). 3.7. Potentials and quasi-potentials; generalizations In this section, we introduce t h e notion of ''quasipotentials** and deduce some of their differentiability properties as well as some additional such properties of potentials. Definition 3.7.1. Suppose ^ is a vector ^Lp(G), p > 1. We define its quasi-potential u b y
G G

(37.1)

Theorem 3.7.1. Suppose e^Lp(G), ^ > 1, u is its and V^ is the potential of e'. Then (a) Each F ^ HI [D) and u^H^ (D) with

quasi-potential,

I V2F||o.p < C{v,p) Wet^Cr. I \\VU\IJ, ^ C(v,p) \\e\\la \\VV\\% < C{v,p)-A'\\e\\la, \\u\\% < C ( . , ^ ) - z l |kK.(. for any bounded domain D, A being the diameter of D;
(b) u(x)=
G

F^(A;),

X^G

(a.e.),

(c) fv^cc(u,<x-i-e')dx

= 0,

v^Lipc(G)
then u^ CJ^{ER) and

(d) if G BR and e^ C^(BR),

h^(Vu, BR) < C{v, fj) h^{e, BR)

(O < ^ < 1).

Proof, (a) Follows immediately from Theorem 3-4.2 and its proof, (b) follows from (3-7.1), Theorem 3.4-2, and t h e definitions. To prove (c), let {en}^ ^ r ( ^ ) ^^^ suppose ^i^ - ^ ^ in Lp(G). For each n, we see t h a t u,(x) = lim so t h a t fei{i)^Ko(x - | ) ^ | = - / i ? o ( * - I) < . . ( ! ) d^
G ^^^G-B{x,Q)

Aun==~el^ and (c) holds for each n. The result follows b y passing to the limit, (d) follows from Theorems 2.6.2 and 2.6.7.

3.7- Potentials and quasi-potentials; generalizations

87

Theorem 3.7.2. If u^
G

H\Q(G),

then u(x) is given a. e. on G by

G

(2.7.19) Proof. We have seen in Theorem 2.7.3 that this holds if ^ ^ C\^{G). So, let U{x) denote the right side of (2.7.19) and let Un^ ^IdQ ^^^ ^^^h n and Un ->u in HIQ{G). We note that

\u{x)\ ^ r-^ I \S - x\^-' '\v u{i)\d^

G

11 U(x)\dx < f \Vu{i)\- \r-^f\i

G G [ G

- x\^--dx] di
J

<g'\\Vu\\la

(3.7.2) y,g- = \G\. Using (3.7-2) for the difference Un U, we see that Un ->U in Li(G), Theorem 3.7.3. (a) Suppose thatf^L\[G) and satisfies (3.7.3) /1/(1)\d^<.L^ f-2+;., o < A < 1 V^H\{D)

for each hall B{x{),r)y and suppose V is its potential. Then for any hounded domain D and satisfies (3.7.4) f\vV\^dx<C^(v,iu)-L^r''-^+^^,
B{Xo,r)

XQ^G, 0 < r < i^

where R is the diameter of G. Moreover (3.7.5) fv(x)f(x)dx==-fvAx)VAx)dx,

G G

v^Hl^{G).

(b) If v>2 and f^ Us'{G), s' = vl{v + 2), then V^Hl(D) for any bounded D and (3.7.6) /lVF|2^^<C2(r)-(||/||,.)2. D //, also, f satisfies f \f{x) \dx<Li i^^-i (rlRy-^+f", 0<r<^R, B(xo,R)cG 0 < ^ < 1 , i:i>||/||, r = vl2 for all xo in G, then (3.7.8) /|vF|2^A;<C(i^,/^)Lf(r/i?)''-2+2A^.
-B(a;o.r)

(c) If f^

L2(BR)

and satisfies a condition f\f{x)\^dx'<L^r''-^+^f'

BiXo,r)(\BR

for every B{xo, r), then V^ CJ,(BR) and h^if.Bn) <C{v,iu)L.

88

The spaces H^ and iJ^^

Proof, (a) Let Wa(x) be the right side of the equation

(37.9)

v,4x) = - rr^/ II - ^i-'d"' - x^)M) di.

o By proceeding as in the proof of the preceding theorem, we easily conclude that V and Wct^Li{D) for any bounded D. By approximating to / in L\{G) hy fn in C^ci^), we see as in the preceding theorem that F $ Hl(D) for any bounded D and that (37.9) holds a.e. Now, we select xo^G and write fii) =M^) +m) where /i(f) = m in B{xo,2r)

and/i(f) = 0 elsewhere and let F be the potential oi fjc. Let & (37.10) <P2{Q\x) = j\m\d^.
R -B(a;,e)n[(?-B(!Bo,2r)]

Then, from (37-9) for V^ we have

\vV2(x)\ < r-^ f \i - x\^'^\f{^)\dC' = r-' f Q^~^ <P'^(Q; x) dQ

r ^ 1 < Lr,-! R^-^ + {V-\)I Q^-^ dQ\
since obviously <(^L r 1)r7i(1 -X)-^Lr^-\ 0<.Q <r Accordingly / I V F2 (A;) |2 i:^ < C {V, ;i) L2 ;^^-2 + 2 ^ A
B{xQ,r)

\ x^B{xo;r)

(:v$5(^o;^)).

From the SCHWARZ inequality, we obtain |VFi(A;)|2<r;-i/i/2, where 0 < a < A (37.11) h=^r-^j\^-x\^-^\m\d^,
B(XO,2T)

and

B{Xo,2r)

since/i (I) = 0 outside B{xo, 2r). In order to evaluate I2 define Then we see that
B{x,Q)nBiXo,2r)nG

0 < ^ <r,

3.7- Potentials and quasi-potentials; generalizations

89

Proceeding as with 992, we see t h a t (37.12) l2<C{v,A){A-o)-^Lr^-\ x^B(xo,r). Integrating (3.7.11) over B(XQ, r) and using (3.7.12), we see that I \VVi{x)\^ dx < o-^(2r)'' L{2ry-^-^^ C{VJ,G) Lr^-\

Since this holds for any a with 0 < cr < A, we see that f\VVi(x)\^dx<C{v, Using (3.7.9), we see t h a t A) Z2 r^-^+2^.

using Theorem 3-7-2 (with x and | interchanged). To prove (b) we note first t h a t it follows from Theorem 3.4.2 t h a t V^HIAD) for any bounded D with ||V^ F | , , . ^ ^ < C(r) ||/||L.. If we take a n y A, we find, as in the proof of Theorem 3-4.2 that

\\/v{x)\^^' <g^ '-i. r-^ j \^ - x\^-^ 1/(1) |2 s' d^

So let Wx rj V ^x where rj = i on BA on rj = 0 on and near dB2A- Then W^ Hlg,(Rv) and has compact support, and Wo:,^(x) =rjV,oc^ + r]jV,oc

I VTF||L,.2^ < I V^FIIL,.., + C'{glA)i-y2s' ll/IIL,. I I

Hence, for each A, we find that V F ^ L2{BA), letting yl - > + oo, we find t h a t using Theorem 3-5-3. B y

J\vv\^dx<CHp){\\f\\o)2.
To prove the second statement, choose a point XQ^G and let i^ be the distance of XQ from ^G. Write

f{x)=Mx)+Mx)
where /2(A;) = 0 in B{xo, Rjl) and/i(A;) = 0 in G 5(A;O, i^/2), and let

F^fc be the potential of fjc. Clearly/g ^ L2., (G) with ||/2||L^(? < 11/11.'.^, so F2 ^ Hl(D) on any bounded D. But since F2 is harmonic on B{xo, i^/2), it satisfies a condition (3.7-8) with // = 1 and Li = Ij/IL'- Finally, i t is easy to see t h a t / i satisfies the condition of part (a) with L replaced b y Li i^-^-^+i. The result follows.

90

Existence theorems

(c) Choose xo ^ BR and r > 0. Write f = fi + f^ where fi{x) = f[x) on B[XQ, 2 r) and/i(A;) = 0 elsewhere; let Vjc be the potential oifjc. Then, from Theorem 3.7-1 or Theorem 3-4.2, we conclude that 11V2 Fi |2 ^A; < y IV2 Fi |2 ^A; < Zi L^ y'-2+2/^. Moreover, we conclude that V2 is harmonic on B(xo, 2 r) and |V2F2W|<Z2/||-^|-^|/(|)|^|,
BiiB{Xo,2r)

x^B(xo,r).

If we define
B{x,Q)n[BR-B{Xo,2r)]

we find that
00 00

I v2F2wI < ^ 2 / r'(P2{e> ^) ^Q < ^ ^ 2 / r""^ 9^2(^;^v) ^o < Z S L r ' ^ - i

since from the s inequality^ SCHWARZ inequality,
(P2(Q>".-^, .J,

it follows that g>R

'LR'-^+i',

Then V F satisfies the growth condition for V Vi so the result follows from Theorem 3.5-2. Chapter 4

Existence theorems
4.1. The lower-semicontinuity theorems of Serrin In Chapter 1. we proved two theorems concerning the lowersemicontinuity of integrals I(z, G) (Theorems 1.8.1 and 1.8.2). Although each of the functions considered was required to be in the space H\{D) for each D <ZC. G, it was necessary in the first theorem only to require that the functions Zn converged weakly in Li(D) to z for each D (Z <Z G, nothing being required concerning their derivatives. In the second, however, we required 2:^ ^ in H\ (D) for each D G G G which implies 7 strong convergence in Li and the uniform absolute continuity of the set functions f \VZn\ dx on such domains D. The writer, in his work referred
e

to in Chapter 1, required essentially this latter type of convergence in his lower-semicontinuity theorems. In this section we present some of SERRIN'S recent generalizations ([1], [2]) of these theorems and some of those of ToNELLi in which SERRIN merely requires that Zn ^-z (strongly) in Li{D) for each D C G G. Our general hypotheses in this

4.1. The lower-semicontinuity theorems of SERRIN

91

section concerning/ are as follows: (4.1.1) (i) f{x, z, p) is continuous for all [x, z, p) (N arbitrary), (ii) f{x, z,p) ->0 for all [x, z, p), (iii) f{x, z, p) is convex in p for each (x, z).

For much of this section, we shall abbreviate (x, z) to t. We need the following well known lemma which is an immediate consequence of the definition given in 1. Lemma 4.1.1. Suppose {/} is a family of convex functions defined on a convex set withf(p) < / o ( ^ ) < + oa for each p in S. Then sup f(p) is a convex function <.fo{p) on S. Suppose t h a t f(typ) satisfies (4.1.1) (t (x, z)) and that its derivatives fp are also continuous. For each L > 0, we define (4.1.2) where fL{t, p) = max fO, sup [F{t, p, q)]\ F{t, p, q) =f{t, q) + {p - q) 'fp{t, q).

Lemma 4.1.2. The function fL{t, p) defined above has the following properties: (i) (ii) (iii) function / L satisfies the conditions (4.1.1); fL(t, p) ^f{t, p) for all (t, p), the equality holding if \p\ < ^ ; for each compact subset ^ of the t-space, there is a constant A and a A((T), > 0 for a > 0 with X{a) ^^0 as a -^0, such that fi^[tyP)^A{\ + \p\) \fL[s,P) -fL{tyP)\<.X{\s -t\)'[\ +\P\), Sj^Z \fL{t,P2)-fL[t,Pl)\<.A\P2-Pl\. Proof. Obviously / L > 0. Using Lemma 1.8.1, we see t h a t 0 and ^{^yp>^) ^fi^'P) ^^^ each t and q, so t h a t / / , is convex and < / . Since F{t,p, q) is uniformly continuous on any compact set, it follows t h a t / / , is continuous. Obviously fhif,p) = / ( ^ , p) for \P\ ^L. Now, clearly uniform inequalities of the form (iii) hold for each function F(t,p,q) when \q\ <:L and s and t are in ^ ; also the sup in (4.1-2) is taken on for some q, depending on t and p. Clearly the first inequality in (iii) follows. To prove the second define foL{t,p) =m^xF(t,p, q).
Then fL{t, p) = max {0,/OL(^, P) ],

Also

\fL{s,P) -fL(t,p)\<\foL(s.P)

-foL{t,p)\.

foL(s,P)=F[s,p,q{SyP)l foL{t, P) = F[ty p, q{t, p)] > F[t, p, q(s, p)] foLis.p) -foL{t,p)<F[s,p,q{SyP)]-F[t,p,q{s,p)] < max {|/(s, q) - fit, q)\ + \p - q\ ^ |/^(s, q) - fp{t, q)\}.

92

Existence theorems

One obtains the same result with s and t interchanged. The third result is proved similarly. Definition. An integrand function / is normal if and only if (4.1.}) f{Xy z,p) -> -\- oo as \p\ ->+ c>o for each {x, z). Lemma 4.1.3. If f is normal and ^ is any compact subset of the {x, z) space, there is a constant a such that f{x,z,p)^a\p\i, a>0, {x,z)^2^. Proof. Suppose this is not true. Denote (x, z) b y t. Then / sequences {tn} and {pn} such that f{in,pn)<'-\pn\-^^ Since f{t, p) > 0, we see t h a t R^ = \pn\ ->c>o. Let jtn = Rn^pn', we may, b y taking a subsequence, suppose that tn >^o and jtn -^no- Since / is convex in p for each t, we have
(4.1.4) f{tn, Rnn) < (l -^Iifn. ) + ^ ^^ " 0' 0<i^<i^^.

Holding R fixed and letting ;^ - ^ o o in (4.1.4) we conclude that /(^o, R n^) < f{k, 0), 0 < 2^ < CO which contradicts (4.1.3), Lemma 4.1.4. Suppose g(x,z,p) satisfies (4.1.1), vanishes outside a compact set 2^= SQ X TQ in the {x, z) space, and has continuous partial derivatives gp which are Lipschitz continuous in x, z, and p and satisfy \gp\ < M for some M. Suppose Zn and z^ Hi(D) ^'^d Zn -^z in Li{D) for each D CC.G. Then Ig {z, G) < lim inf Ig {zn, G). Proof. Let e > 0 and choose Q SO small that 5o C GQ and (4.1.5) / I V ^ V^el^:^ < e
-So

ZQ being the 99 mollified function for some moUifier cp. Now the W E I E R STRASS ^-function associated with g is defined b y E{x, z, p, q) = g[x, z, q) - g[x, z, p) - {q - p) ' gp{x, z, p). Since g is convex and \gp\ < M , we conclude that 0 < E{x, z,p,q)
g{x, Zn, pn) g{x, Z, p) (4.1.6) > g{x, Zn, PQ) + [pn PQ) ' gv{^> Zn, PQ) g{x, Z, pg) ip-po) 'gp{^> ^>PG) - 2 M \ P P,\.

<^2M\p

q\.

Making the abbreviations \J Zn = pn, V ^^ = PQ> \l Z = p, we obtain

where

2M \p - PQ\ ^ g{x, z, p) - g{x, z, pg) ~ (p - pg) gp(x, z, Pg).

4.1. The lower-semicontinuity theorems of SERRIN

93

Now, from the dominated convergence theorem and FATOU'S lemma it follows t h a t lim inf f [g {x, Zn> PQ) g{x, z, pg)] dx>0,

(4.17)

hm

/ PQ [gj,{x, Zn, PQ) gp{x, z, pe)] dx = 0.

Q(X,Z) =gp[x,Z,Pg{x)].

Now, let us write From our hypotheses, it follows t h a t Q is continuous and has support in 5o X To and
I Q{Xi, Z) Q{X2, Z) \<M'\Xi-X2\.

We may approximate uniformly to Q by similar functions of class C^, all with the same M ' ; so we may assume Q^ C^. Define
Znix) zlx)

Then, from Theorem 3.1.9, it follows t h a t P^Hl{G) support in G, and so J VP{x)dx = f {Q[x,Zn(x)']Vzn
G i Zn

and has compact +

Q{x,z)'7z}dx

+ j{lQcc(x,
G [z

C)dC\dx = 0,

From this it follows that (4.1.8) I f[gp{x,Zn,pe)pn

So

gp{x, z,pg) p] dx\ <M'

f\zn
So

z\ dx.

Integrating (4-1.6) over So and using (4.1-5), (4-1-7) and (4-1-8), we obtain
I(Zn, G) I(Z, G) > En 2M S M' J\Zn So z\ dx,

Ihnsn = l i m M ' f \zn z\ dx = 0. The lemma follows easily from this and the arbitrariness of e. Lemma 4.1.5. If q? is convex on an open convex set S, it satisfies a uniform LIPSCHITZ condition on aity compact subset of S. This is well known. Lemma 4.1.6. Suppose f(t, p) is strictly convex in p for each i (and satisfies (4.1.1)), suppose ^ is a compact set in the t-space, and suppose K, L, and R are positive numbers. Then there exists a positive constant K such that (4.1.9) Rt, P) -f(t. q)-A-[p-q)>K whenever t^^, \p\ <,K, \q\ < L , \p q\ '>:R, and A is a vector such thatfif, r) f{t, q) A ' [r q) ^0 for all r (Lemma 1.2.2).

94

Existence theorems

Proof. Let {tn}, {pn}, {^n}, {An} he chosen to satisfy the conditions and so t h a t the left side of (4.1.9) tends to its inf. Since ^ is compact and / satisfies a uniform Lipschitz condition in p for p on any bounded part of space, it follows that the An are uniformly bounded. We now prove our first lower-semicontinuity theorem (SERRIN [2], Theorem 12): Theorem 4.1.1. Suppose that f{x, z, p) satisfies one of the following conditions (in addition to (4.1.1)): A. / is normal. B. f is strictly convex. C. The derivatives fx, fp, andfpx

are continuous.

Then if Zn and z^ Hl{D) and Zn -^z in Li{D) for each D G CG, it follows that I{z,G) < l i m i n f J ( ^ ^ , G). Proof. We shall prove the theorem in case A first. Let 5 a n d T be compact subsets of t h e x and z spaces, respectively, let L > 0, let SQ and TQ be bounded open sets containing 5 and T, respectively, and let > 0. From Lemma 4.1-3. it follows that there is a constant ^ > 0 such t h a t
(4.1.10) f{x,z,p) >a\p\1, X^SQ,Z^TO,\P\ < . L + \ .

Let 99 be a moUifier in [x, z, p) space and let /^ be t h e 99-mollified function of / where Q is chosen so small t h a t (4.1.11) \f{x,z,p)~f,[x,z,p)\<.e for {x, Z)^SQ x TQ^P\ <.L + \. Next, we let oc{x, z) be of class C^ everywhere, have support in So X To and equal 1 on S X T ; then define (4.1.12) g'{x,z,p) =oc(x,z)feL{x,z,p). It is clear t h a t g' satisfies the conclusions (i) and (iii) of Lemma 4.1.2 and vanishes for (x, z) outside So x TQ. Now, letting {x, z) = t, ^^o = So X To, Z = ^ X T (4.1.13) g'it.p) ^f,L{t,p) <.fS>P) ^f{t>P) 1. + e if ^ 2 ^ 0 , \P\^L+ So now, suppose \p\ > L + 1. Then

fQL(t,p) =fQ{t, qp) + {p qp) 'fep{t, qp) for some \qp\ <. L = fS> qv)+S'(rp (4.1.14) - qp) 'fMi> qv) [S = r f ^ >^

= S[Mt. qp) + [rp - qp) f,p{t, qp)] + (1 - S)fS. qp) <Sf,{t,rp) + {\-S)fS.qp) <S[f{t, rp) + e] + (1 - S) [f(t. qp) - s] <{2S-1)s+f{t,p)

4.1. The lower-semicontinuity theorems of SERRIN

95

using the convexity of the segment qpp tion of the ray Op line 0 p such t h a t shows that Sp^ B(0, (4.1.15)

and the fact that S > 1. Here, rp is the intersection with dB(0, L + 1). But now let tp be the intersecwith dB{0, L -\- \) and let Sp be the point on the qp Sp \\ fp tp. Then a simple geometric argument L) so t h a t S = i f ^ < | ^ | - I .

If, finally, we let g{t,p) =g^[t,p) (mollified) with a sufficiently small we see, using the bound \g'(t,p)\ < ^ (1 + \p\), t h a t (4.1.16) g{i>P) <f9L{t.P) + e(i + 1:^1) <f{t,p) + s{3 + 1^1) ^mp) + C8[\+f(t,p)] k[t,p)-f{t,p)\^Ce for t^Z>\P\<^L

using (4.1.10) for t^^Jo and the fact that g(t,p)=0 otherwise. Using the conclusions (iii) of Lemma 4-1.2 for g', we see t h a t g satisfies the hypotheses of Lemma 4.1.4. Now, suppose 7(2:, G) < + oo and 7 > 0. There are compact sets ^ 5 C <^ in the x space, T in the z space, and a number L so large that if 5 * is the subset of S where z^T and \p\ < 7 , then (4.1.17) I(z,S*)>I(z,G)^r) (if I(z, G) = + 0 0 , let M be arbitrarily large and S, T, and L can be found so that I{z, S*) > M). Then, using (4.1.16) I{z, G) < I{z, S*) + rj ^ Ig(z, S*) +7)+ ^Ig{z,G) Ce\S*\ +r} + C e | S * | < lim inf/^(^^, G) + ly + C e | S * |

< (1 + C e) lim inf/(^^, G)-\-r] + C e{\ + \G\). Since r] and e are arbitrary, the result follows in this case. In case B, we proceed in the same way. But, in this case we find, using Lemma 4.1-6 for/^ t h a t (4.1.14) can be replaced by
(4.1.14') f,L{t, p) <f{t, P) + ( 2 S - \ ) 8 - K S

where x is for fg where ^ $ ^ o > \q\ < 7 , | r | < Z, + 1, and |^ ^| > 1 since, (4.1.14)
fQ{i> ^p) + i^P - qp) 'fQp{i> (1P) ^fQ[t> ^p) - ^

Since a convex function is Lipschitz, it follows that the supporting vectors A in Lemma 4.1.6 are just/27(^, q) for almost all q; consequently one can mollify both sides of (4-1.9) to see that the K above is that f o r / with L replaced by L + ^ and R = i replaced by 1 2Q. Thus, if we choose e < x/l, we see that (4.1.16) can be replaced by g{t.p)<f{t.p) \g{t,P)-nt.P)\<Ce The rest of the proof is the same. + Ce for t^Z-\P\<I'.

96

Existence theorems

In case C, we proceed b y defining

g'[x, Z, p) = 0C{X, Z) fL{x, Z, p)

where (x[Xy z) was discussed above. B y inspecting the proof of Lemma 4.1.2, we see t h a t g' satisfies the conclusions of that lemma with A((T) < ^ cr if ^ is not varied and g'[XyZyp) ^f(x, z, p), equality holding if {x,z)^^ and \p\ < L.

If we now mollify with respect to p only we obtain (since 1 + |_/>| satisfies a LiPSCHiTZ condition for all p) that g{x, z, p) = ga{x, z,p) satisfies all the conditions of Lemma 4.1-4 and \g{x, z,p) g'[Xy z,p)\ '<AG for all {x, z,p). The rest of the proof is the same. We now remove the restriction in Theorem 1.8.2 that fp be continuous and generalize t h a t result somewhat (see SERRIN [2], Theorem I3 and MoRREY [7], Chapter I I I , Theorem 4.I): Theorem 4.1.2. If f satisfies (4-1.1) only and Zn and z^Hl(D) with \\zn\\\^D ^'^^ ll-^lli.D 'i^'yiiformly hounded and Zn->z in L\[D), all for each D CGG, then I{z,G) < l i m i n f / ( ^ ^ , G). Proof. Choose D CC G. If lim inf I{Zn, G) = + co, there is nothing to prove. Otherwise, let > 0 and M be an upper bound for ||>?W||I.D and ||-2:||i jr>. Define g(x,z,p) =f(x,z,p) + e\p\. Then g is normal. Using Theorem 4.1.1, we obtain I(z, D) < Ig{z, D) < lim inf7^(^^, D) < M + lim inf I{zn, D) <sM + lim mil{zn, G).

The theorem follows. If, in Theorem 4-1.2, we require only t h a t Zn -^z in Li{D) (without the uniform boundedness of ||'2:?^||i^z)) then the conclusion of the theorem doesn't necessarily hold. An example, due to ARONSZAJN, is to
be found in t h e book b y PAUC.

A third and perhaps less interesting theorem, stated below, requires the following definitions: Definitions. An integrand / is of type I if and only if there are two positive functions X{a) and ^(c) defined for cr > 0 with limA(<7) = lim//(o') = 0, //(a) <, B a for a large, such that \f(Xy z,p) ~f[y, w,p)\ ^X[\x - y\) [1 + / ( ^ , z^p)] + ii[\z - w\)

4.1. The lower-semicontinuity theorems of SERRIJSL

97

for all X, y, z, w, p. An integrand is of type II if and only if there is a function X with the properties above such that \f{x, Z,P) -f[y, W,P)\ < A(|^ -y\+\u~ V\) [1 + / ( ^ , Z,P)],

Remarks. Important examples of integrands of type I and I I are / = / ( ^ , P)=A{X) F(p). f = f(x. z,p)=A (X, z) F{p),

respectively, where A [%) and A {%, z) are continuous functions with some positive lower bound. Theorem 4.1.3. (a) Suppose f is of type I. Then the conclusion of Theorem 4.1.1 holds. (b) Suppose f is of type II, z is continuous on G, Zn cind z^ HUD) for each D G G G, and Zn converges in measure to z on G. Then I(z, G) < liminf/(^^, G). We refer to reader to SERRIN'S paper (SERRIN [2]) for the proof. In t h a t paper SERRIN defines the functional I{z, G) in much the same way as LEBESGUE defined the area of a surface: ^{z, G) = inf flim inf/(^^, Gn)\ for all sequences {zn} in which each Zn^ C^{Gn), GnG G, U Gn = G, and Zn -^z in Li{D) for each D CCG. The scope of the discussion is enlarged by allowing ''weakly differentiable'' functions z in the discussion. A vector function z is weakly differentiahle on D if and only if it ^ Li(D) and there exist measures fii such that I zicoi^^dx= f wi djui for all oj ^ C ^ {D);

in case z is continuous, it is of bounded variation in the sense of TONELLI (bounded variation in the ordinary sense iiv 1). SERRIN then investigates in some detail the relation between the integrals I{z, G) and I(z, G). In many cases it is shown that I{z, G) < I{z, G), the equality holding when z is ''strongly differentiable*', i.e. z^H\ (D) for each D CC.G. As an example, if we allow functions of bounded variation and define I[z, G)= f yr+V~^ dx
a

{G = {a, b)),

then I(z, G) is just the length of the arc^: = z{x), if z is continuous. In this case, one can say t h a t I{z, G) < ^{z, G) and the equality holds if and only if z is absolutely continuous. For details of the results, the reader is referred to the paper of SERRIN already referred to.
Morrey, Multiple Integrals y

98

Existence theorems

4.2. Variational problems w i t h / = / ( p ) ; the equations (1.10.13) with N = I, Bi = 0, A'^ = A<^(p) In this section, we prove the theorems stated below. The first represents a complete extension of t h e results of A, HAAR discussed in Chapter 1. STAMPACCHIA [ 3 ] I required some additional smoothness of the domain G and of the boundary values as did GILBARG. Definition 4.2.1. Let G be a bounded strictly convex domain. A function (p defined on dG is said t o satisfy the bounded slope condition with bound M {BSM condition) iff with each xo^dG, there exist two linear functions h{x) = q^{xo) + afc^x"^ %g), | % | < M, k= 1,2 such t h a t li{x) <,(p[x) < l2{x) for all :\; on ^G. Theorem 4.2.1. Suppose G is bounded and strictly convex, cp satisfies the BSM condition on dG, f = f{p) is convex and I{z, G) is given by (1.1.1). Then there is a function ZQ which coincides with cp on dG, which satisfies a Lipschitz condition with constant M on G, and which minimizes I(z, G) among all Lipschitz functions z = q? on dG. If f is of class C^ with (4.2.1) fv,vSP)^^^^>^ ^^ ^^^' then ZQ is unique and minimizes I{z,G) among all z^Hl{G) such that z ZQ^H\Q{G). If (p^HliG) with p'> 2, and G is of class CJ, then ZQ^ Hl{G). If f^ Q , C^, or analytic, respectively, then so is ZQ on G. If, also, G and 99 ^ C^, C"^, or analytic, then so is zo on G. Theorem 4.2.2. Suppose G and cp satisfy the hypotheses of Theorem 4.2.1 and suppose that the A'^(p) ^ Cj^, 0 <C /J' <i 1, and satisfy (4.2.2) (4.2.3) ^i^iP)^-^^>0 if A 9^0. Then there is a unique Lipschitz solution ZQ of the equation lc,.A'^(Vz)dx = 0, C^Lip.G
G

which coincides with 99 on dG; ZQ satisfies a Lipschitz with coefficient M on G. If the A'^ C^ with w > 1, then z^ CJ5+^(G). If G and the boundary data also ^ CJJ'^\ z^ CJ5+^(G). The C^ and analytic cases also hold. Theorem 4.2.3. Suppose G is of class C^ and cp ^ C%{G) and suppose the A^^ CJ and satisfy mi{p)\X\^<.Al^{p)Xo.h<^Mi{p)\X\^, mi(p)>0, Mi(p) <ymi(p). (4.2.4) Then there exists a unique solution ZQ 0/(4.2.3) ^ C^iQ- T^^ higher differentiability results of Theorem 4.2.2 hold. Our method is first t o solve the equations (4.2.3) in the cases where G, cp, and the A^ are reasonably smooth (these include the Euler equations for t h e variational problem) and satisfy (4.2.4) with m{p) a n d
1 He has recently improved his results

4.2. Variational problems with f = f (p)

99

M{p) constants. We then obtain bounds for the first derivatives which enable us to handle the general cases by approximations. The proofs of the first statements in Theorem 4-2.1 depend on the following two lemmas: Lemma 4.2.1. Suppose f(p) is convex for all p. Then there exists a sequence {fn} of convex functions such that each ^ ^^{Rv), fn[p) ^f{p) for all n andp, andfn{p) converges uniformly tof{p) on any hounded set in Rp. Proof. S i n c e / h a s a supporting plane at 0 (Lemma 1.8.1 (a)) we may subtract that off and assume that f(p) > 0 and /(O) = 0. Define gn{p) as / L O was defined in the proof of Lemma 4-1.2 with L = n. It follows (since /(O) = 0, f(p) > 0) t h a t / L O = / L and so satisfies the conclusions of t h a t lemma. From the Lipschitz condition in (iii), it follows that, if gna denotes the cp cr-mollified function for some mollifier 99, then

gna^C"" and gna{p)<gn{p)+Aa<f(p)

We may clearly take
fn{P) = gnan(P) ^Gn,

+ Aa,
On = ^jnA

\p\<.n.
.

Lemma 4.2.2. Suppose f and fn are convexy fn(p) ^fiP) for each n and p, and fn - > / uniformly on any hounded set. Suppose that Zn and z all satisfy a Lipschitz condition with coefficient M on the hounded region G and that Zn ->z uniformly on G. Suppose Zn = z on dG and suppose, for each n, that Zn minimizes In{^y G) among all Lipschitz Z =^ z on dG. Then z minimizes I{Z, G) among all such Z. Proof. Let ^ == inf / (Z, G) among all Z = 2: on dG. Let > 0 and choose di Z z on dG such t h a t I{Z,G) <C d + e. Then In[Zn, G) < In[Zy G) < / ( Z , G)< Then lim sup/^(;2f^, G) ^d. d + 8.

On the other hand (using the uniform Lipschitz condition and the uniform convergence of fn t o / ) , I{z, G) < lim inf7(^^, G) < lim mi[In(zn, G) + En] ^ d (lim en = 0). Lemma 4.2.3. Suppose G is hounded and strictly convex and suppose (p satisfies the BSM condition on dG. Let r= {(x, z)\x^dG, z = (p{x)} and let E he the convex cover of F. Then there are convex functions q)~ and (p^ on G such that (4.2.5) E= {{x,z)\x^G,cp-{x) ^z<,(p+[x)}. (p~ and (p+ satisfy Lipschitz conditions with coefficient M on G and (4-2.6) (p-{y) = (p{y) = (p+(y) for y^dG. If (p~{xo) (p+{xo) < 0 for some XQ^ G, then (p~(x) 99+(A;) < 0 for all x^G.

100

Existence theorems

Proof. The hypotheses on G and cp imply t h a t Z has t h e form (4.2.5) where9?" and 99+ are convex and (4.2.6) holds. Suppose Xi and X2^G and suppose xs and X4, are the intersections of the line xi X2 with dG. At xs we have linear functions /g" and l^ of slopes < M such that I-(x) < 99-{x) < 99+ {x) <l}(x), x^G, /g-(^3) = (p(X3) = It (^3) with the same at x^^. I t follows easily from the elementary properties of convex functions of one variable (along Xi, x^) t h a t \(p-{xi) -(p-[x2)\ <.M '\xi-X2\, \(p+{xi) -(p+{x2)\ <M\xi -X2\. The last statement is evident. Lemma 4.2.4. Suppose the A^" ^Cl{or C ) , n > 1, and satisfy (4-2.2). Then, for each R, there exists a vector A%^ CJJ {or C^) which satisfies (4.2.4) with Ml and Mi constants and also satisfies
\ARJ,{P)\^MI{R) for all p,A%{p) = A^[p) if \p\<.R^

Proof. From (4.2.2) and the continuity we have (4.2.7) Al^{p)K.Xp^mi{r), \A^[p)\<.Mi{r) for \p\<^r. Letco^ C ( i ? i ) withcw(s) = Ofors < l,co(s) = 1 fors > 2, andco'(s) > 0 for all s. Clearly mi{r) is non-increasing and Mi(r) is non-decreasing. Let us define (4.2.8) A%{p) = i \ - coirjR)] A-[p) + c[R) w{rlR) ^ . , \P\ = r, + c m |A|2 + 5 (o'{c H^ - f-1 HA-Aoc) \p\-^{p'X). 2Mi|^||A|), where c == c{R) is to be chosen. B y computation we obtain ^Rp^(P)^ocPi^ ={\ - (o) Al^XAp s = rlR, Since for large r, \A{p)\ Al^^AccA^^ (4.2.9) H=

< 2 Mi(r) r, we find using (4-2.7) that 2k ^ mi{2R).

[2k{\ - M) + ca)]|;.|2 + SO)'{cm

If we take c >: 2 k, we see t h a t the right side of (4.2.9) > ^ |AP for all r on \R, 2 R] if 4kc{sa)') >4M2(sco')2 (Mi = Mi{2R)) which holds if ^ c > Mf max(s a>'(5)). Lemma 4.2.5. Suppose G is bounded, cp ^ H\{G), the A^{p) ^ C^ and satisfy (4.2.4) with mi and Mi constant and the condition (4.2.10) \A^(p)\<M^. Then there exists a unique solution of Equation (4-2.3) such that z (p^ HIQ{G). If G is of class C^ and 9?^ C^{G), then z^ H^{G) for every r. If G is of class CI, the A"^ are of class C\, and cp^ Gl^{G), then z^ Cl{G). Additional theorems concerning differentiability on the interior and at the boundary hold as stated in Theorem 1.10.4 and 1.11.3.

4.2. Variational problems with f = f [p]

101

Proof. We use Theorem 5.12.2 to prove the existence. We set (4.2.11) (4.2.12) z = (p + u, A'{x,p) = A'^[\/(p{x) + p], '^l{u,v) = A(v), {A(v),w) = J w^aA^'iVq) + Vv) dx.
G

and define A{u) and 51 (w, v) b y

We conclude, from (4.2.4) and (4-2.10) t h a t A-{q)=A-(0) + q,plA-^{tq)dt,

| ^ (^)| < M i ( 1 + | ^ | ) , -hpocp^f^P^l-^^i^)

1

pocA^lVcpix) +p] =Po,A<'[V(p(x)]

> m i | ^ | 2 - M i | ^ | - ( 1 +|V(^(:v)|). (4.2.13) {A^V^(p[x)-^p2-\-A-[\J<p{x)+p^-\}{P2.-pia)

1

+ ^Pl^^

+ {1 ~t)Pi +
Thus the A'ix.p) {A(v),v)
G

0 tp2]dt>mi\p2-pi\^.
- Mi\\v\\l{\G\ + ||V9^i)

satisfy the conditions (5.I2.13) and (5.12.14). Also

= Iv,,A-(x,Vv)dx>mi(\\vr2,)^

SO that A satisfies the coerciveness condition (5.12.3). Thus A satisfies all the conditions of Theorem 5-12.2 so there exists a solution u^ HIQ(G) of Equation (4.2-3). The solution is easily seen, using (4.2.13), to be unique. The interior differentiability results follow from the results of 1.11 and Chapter 5. We must now prove t h e differentiability a t the boundary assuming G of class C^ and 99^ C^{G), To t h a t end, let xo^dG and m a p a neighborhood 31 of XQ on G onto GR U OR (see 1.2, 5.I, etc.) using a diffeomorphism of class C^ so that XQ corresponds to 0 and '^ 0 dG corresponds to OR. Then Equation (4.2.3) assumes t h e form
(4.2.14) lv,ocA'{x,Vu)dx = 0, V^G\{GR)

where the new A'^ C^ and still satisfy the stated condition, possibly with different bounds. We may rewrite (4.2.14) in the form
1

fv,cc[af{x)u,^
GR

+ A^(x,0)]dx

= 0,

af{x) = JA^
0

[xJVu{x)]dt

and the aj^ are bounded and measurable and satisfy (4.2.15) ^ ? ^ ( ^ ) ^ A ^ > wi|A|2, \a{x)\ < M i and t h e A''{x,0)^ C^. From Lemma 5.3.5, it follows t h a t ^ ^ Cl{Gr) for each r < R and satisfies a Dirichlet growth condition as in t h a t lemma

102

Existence theorems

Since this is true for each XQ, we conclude t h a t (4.2.16) flVul^dx^L^r-^+^f", 0</^<l

for all B(xo, r). We now revert to Equations (4.2.14). Applying the difference quotient procedure of 1.11 in a tangential direction, we see t h a t Ufi = h-^[u{x + hey) u(x)] satisfies fv,oc{K^'^h,^ + et)dx = 0,

where a^^ and e^ are given by formulas like (1.11.4), the a^^ satisfy (4.2.15) uniformly for small h, \\uji\\l^G{r) i^ uniformly bounded for each r <C R, and e^ satisfies a condition (4.2.17) J\eh\^ dx < JJr^-^+'^f^, 0 < r < i?.
-S(i/o,r)n(?j5

It follows, again from Lemma 5-3-5 that the u^ are uniform^ Holder continuous on each Gr and \/ uji satisfies (4.2.17) with R replaced by r with an L and [A, independent of h. Thus we may let A -> 0 and find that the SIpy with 7 < v satisfy these conditions. Since pv,y = py^v and we can solve for pv^p in terms of the others, we see t h a t Vpv also satisfies (4.2.17) as do the py, and therefore py^ C^^(G-r) for each r < R. The higher differentiability results follow from Theorem 1.10.3. Lemma 4.2.6. Suppose G is of class C^, z ^ ^ J ( ^ ) ' ^^^ ^ satisfies an equation of the form fv,oca^^(x)z,^dx
G

= 0,

v^H\^{G),

a^^^Cl{G).

Then z takes on its maximum and minimum values on dG. Proof. Approximate to the a""^ by a^^ ^ Q ( ^ ) and let Zn be the solution coinciding on ^ G with z of the equations

Then Zn^ ^li(^)> ^n^ ^^(-^) for each D (Z G G, each Zn takes on its maximum and minimum values on dG and Zn -^z in C^{G) (see 5-2, 5.5, and 5.6). Proof of Theorems 4.2.1 and 4.2.2. In case 99 coincides on dG with a linear function l(x), then l{x) is the desired solution in either case. So we assume that this is not the case. We first prove Theorem 4.2.2. We first assume t h a t the A^^C^ and satisfy the conditions in Lemma 4.2.5. Let ^ be a moUifier. For each n, we define '(p^{x)^ (p{x)^ Mn-^Tn~^\^-M^^ ^^G, xo^G,

4.2. Variational problems with f = f {p)

103

Then (p~ and 99+ ^ C(GeJ and are convex. Let Then Gn is convex and of class C^, Gn C Gii+i, G^T^ -> <^, and 99^ converge nicely to cp- on any D C C G. For each w, let Zn be the solution of Lemma 4.2.5 with (p replaced by cp~, say. Each Zn<^ ^^(Gn). By a difference quotient device like that in 1.11, starting from (4-2.3) we find t h a t each ^n,y == :^wy satisfies lv,.[a-^{VZn)pny,(i]dx = 0, V^Hl,{Gn) a-^{q) =^ A^ iq)

and so each linear combination of the pny takes on its maximum on dGn so t h a t \'7Zn\ takes on its maximum on dGn. Now, since (p~ is convex, we conclude that

as is seen b y rotating axes. Accordingly

'al'{x) = JAl^[{\ -t)V<p^ + t^z]dt,

0
so t h a t Zn q)~ ^0 by the maximum principle. In like manner Zn < <p^ on G,j. Consequently \\7Zn\ < M^ on dGn and hence on Gn and M^ - > M . Then a subsequence, still called {zn}, is such t h a t ^^i converges uniformly on each D C. G G to 3. Lipschitz function z = (p on dG and z satisfies a Lipschitz condition with coefficient M on G. Moreover, on any D G G G, it follows from the theorems of 5-2, S-S, and 5-6 t h a t the derivatives of all orders converge uniformly on D to those of z. Thus ^ is a solution. If the A'^ C^ b u t do not satisfy the conditions (4-2.4) with mi and Mi constants, we may find functions A%{p) which = A'(p) for \p\ <R,R> M, which do satisfy (4-2.4). For R> M, the solution z obtained above for ^i? is just that for A. Finally we can approximate any A'^^ C^~^ by those ^ C^. This proves Theorem 4.2.2. The higher differentiability results along dG follow from Theorem 1.11.3. In c a s e / i s regular and of class C^, we conclude from Theorem 4-2.2 just proved that there is a solution ZQ of Euler's equation on G where z satisfies a Lipschitz condition with coefficient M on G and ZQ ^ H^ {D) for each D GG G and each r > 2. It is immediately evident that .^o is the unique function minimizing I[Z, G) among all Lipschitz functions Z = ZQ = q) on dG. The higher differentiability results follow as above. I f / i s not smooth, we may approximate it from below as in Lemma 4-2.1 and complete the proof of Theorem 4.2.1 by using Lemma 4-2.2 and the uniform Lipschitz bound M.

104

Existence theorems

Proof of Theorem 4.2.3. Let us first assume t h a t the A^" ^ CJ and satisfy the conditions of Lemma 4-2.5. Then there is a unique solution z^ Cl(G) with z = acp on G oi Equation 4.2.3 for each a, 0 < cr < 1. As above we see that V z takes on its maximum on dG (for each a). We note t h a t z{x, 0) ^ 0 and, b y difference quotienting with respect to c, we see t h a t z and V z vary continuously with a. In fact Za^ C"J(G) and satisfies (4.2.18) fv,aAl{aV(p+Vu)Za,^dx = 0, v^Cl{G), z = a(p + u.

Now, let us suppose t h a t the A^" satisfy the conditions (4.2.4) with mi(p) and Mi(p) variable. Let us choose R > max|V9!?(^)| and define the A% as in (4.2.8) so Lemma 4.2.4 holds. Then, as long as | c V 9? + V ^ | = \Vz{x, a)\ < R, u is 3L solution of a'^(x,G)u^ocp(x,G) = aa'^^ (p^x^, Z{X,G) = a(p{x) + a-^(x,G) = {A-^^[VZ{X,G)] (4.2.19) Then we notice t h a t (4.2.20) W2(7)|A|2<t2^(^,a)/laA)8< 7 ^ 2 ( 7 ) , M^iy) = 2j[\ +y) for such {x, a). From (4-2.19) and (4.2.20), it follows t h a t where l\ip\\\o = m a x |^(JV;) |. Now, let xo^dG. Since G is of class C^, there is a ^ > 0, independent of xoy such t h a t there is a unique ball B{xi, Q) such that B(xi, Q) D G = {XQ]. If we define w(x) = K-P(Q-t we see t h a t Lw ==KPtr-i'^ < K Ptm2{y)
{^/^(x)]^

U{X,G)

+ A^^^[Vz{x,G)]}{mi{\7z)

+ Mi(V^)}.

-r-i),

r=\x~-xi\,

P = |||9^|!|g,

IZ^""" - {^ + 2) r-^ ^"'^(^"^ ^1) (^^ ^f)] r-i-^[vy (t + 2)]

W{XQ)=0, W{X)>:0 for x^G.

KPtQ-^-^,

Clearly we m a y choose K{v, y, D) and t{v, y, D) so t h a t Lw{x) Then we see t h a t L{a w ^ u) ^0 inG, a w :^ u >:0 on dG. in G and so, since
<GCi(v,y,D)*P,

< C{v,y)P,

x^G.

The maximum principle implies that aw^u>0 (T te^ ^^ = 0 at ^0, it follows that du , :.G\\7W{XO)\ (4.2.21) \Vu(xo)\ =: '-(xo)

4.3- The borderline cases k =v

105

provided that max (cr | V 951 + | V ^|) < i^. But if we take

i?>liv<pK + Ciii|v2^|||S
the bound (4.2.21) will hold for all ;^ on G and all c, 0 < o- < 1. This proves the theorem. 4.3. The borderline cases k v In this section, we complete the proof of the continuity properties of the solutions of minimum problems which were stated in Theorem 1.10.2. We assume t h a t (i) f{x, Zyp) satisfies (4-1.1); and (ii) there exist numbers w > 0 and K such t h a t (4.3.1) / ( ^ , z,p)>mV'' K for all {x, z,p), F = (1 + \p\^)^l^. Lemma 4.3.1. Suppose ^^H\{^), k'>\, and ^ is its average over ^ . Then there is a constant C = C{y, k) such that
/{if(a)-?|2+|VaC((T)|2p/2^2^(^)<C/'|VaC(or)|^^2'((7).

Proof [k and v fixed). If this is not so, 3{^'n] ^ f c ( ^ ) such that u z On account of the homogeneity, we may assume the left side = 1 for every n. Since the H\ norms of the functions f^ C are all 1, a subw sequence y some fo in Hl{^). But then I^n Cn tends strongly in HK^) to Co since VaZn-^Oin Ljc{y^).Thus f 0 is a constant, which must be 0 since its integral over ^ is 0. But this contradicts the fact that \\Cn ~ CnWl^u
= 1.

Theorem 4.3.1. If f satisfies (i) and (ii) above and also (4.3.2) f[x,z,p)^MVand if z minimizes I{z, G) among all vectors in Hl[G) having the same boundary values, then z is Holder continuous on interior domains and satisfies (1.10.9) { SV^dx\^l'^C^[\Vn'{rlRY if ^ < ^ ,

where C and [JL depend only on m, M , K, and v. Proof. The hypotheses (4-3.1) and (4.3-2) imply that (4.3.3) (4.3.4) ^ Ipl" K <f(x,z,p) <Mi\p\^ + Ml, Ml = 2"-!^. (p[r) = flpl'dx, Br == B{xo, r). Let us define

106

Existence theorems

Using (4.3.3) and the fact t h a t z is minimizing we find t h a t (4.3.5) m(p{r) < I(z, Br) + K\Br\^ I{Z, Br) + \\7Z\'dx and is therefore essentially
Br

K\Br\

< ( K + M'i) \Br\ + MiJ where Z is any function = ^ on dBr. Now, for almost every r, z^Hl{dBr) continuous. For such r, let Z{s,a)

= z + (s/r) [z{r, a) z] and z being the

(s, a) being "spherical coordinates", a being on dB(0,i) average of z{r, a) over dB{0,\), Since 1 V 2 |2 = Zf + 5-21 V a Z |2 = r-^{iz(r, we find t h a t J\\/Z\''dx==v-^f{[z{r,
Br dBir,a)

o) - z]'^ + \ Vaz{r, o) |2},

G) - ^]2 + I "Vazir, o) l^Yl^d^^ Crcp'[r)

(4.3.6)

< C{v) f I Vaz{r, o) \'d2; <

dB (r, a)

using Lemma 4.3-1 From (4.3-5) we conclude t h a t (4.37) (p(r) ^Ar (p'{r) + B r\ A = M i / w , B = yv[K + Mi)lm.

From (4.3.7) and the absolute continuity of (p it follows t h a t

q^(r) < (rlR)^iA(p{R) + j ^ - {T^'ARV-HA ^v)

(4.3.8)

fV'dx^

2"-! [yvT'' + ^(r)]

from which the result (1.10.9) easily follows. The HOLDER continuity follows from this and Theorem 3-5.2. Theorem 4.3.2. Suppose k = v = 2 and f satisfies (1.10.8*) m\p\^ <f{x,z,p) <M\p\^, 0 < w<M", instead of (4-3-1) ^'^^ (4.3-2). Then if G is bounded hy k> \ disjoint JORDAN curves andz*^H\(G) and is continuous on G, then any minimizing function z with z 2:* ^ H\Q[G) is continuous on G. Proof. Since (1.10.8*) implies (4-3.1) and (4.3.2), the continuity on the interior follows from Theorem 4.3-1. If, in the argument in (4.3.5) we replace Br by an arbitrary domain gC.G, we conclude, using (1.10.8*) that (4.3-9) D{z,g)<.{Mlm)D{Hg,g) Hg denoting the harmonic function coinciding with z on dg. Now, let Po^dG; suppose it is on the bounding curve F, liFis the outer boundary m a p the interior of F conformally (using the RIEMANN mapping theorem, see for example Chapter 9) onto a domain G' in the

4.3- The borderline cases k = v

107

Upper half-plane so t h a t PQ corresponds to the origin, a part of dG corresponds to aa and a part of G corresponds to Ga. If JT is an inner boundary, t h e same result may be obtained by first performing an inversion with respect to a point interior toF. The transformed function 'z* is continuous on 'G a n d ^ / f | ( ' G ) and the transformed function 'z is continuous in 'G, ^H\{'G), satisfies (4.3-9) for subdomains '^ C 'G, and 'z 'Z*^HIQ('G); all these things follow from Theorem 3-1-5, the invariance of the DIRICHLET integral under conformal mappings, and an approximation. Let us extend '^*, 'z and 'w = 'z 'z* by the formulas 'z*(x^,x^) (4.3.10) = 'z*{x'^, x^)y 'w{x^,x^) = 'w(x^, x^), 'z{x^, x^) = '^*(A;1, X^) Tf- 'w(x^, x^)

where 'G~ is the reflected region. From Theorems 3-l-2g, 3-1.8 and its corollary, it follows t h a t '2:*, 'z, and 'w^H\i^G{i) and '2:* is continuous on 'Go = 'G U 'G- ('Go = 'G\J 'G~ U part of x^ axis on d'G). Let (p be a mollifier and let 'ZQ, etc., be the mollified functions. Now, for each R, Q <,R <C a, there is a set of r of positive measure between Rje and R such t h a t i^z is already A.C. in a for almost all r, see the corollary to Theorem 3.1.8) 'z[r, a) is A.C. in c with 'za^, a) ^ L2 and
2jr

R 2 71

(4.3.11)

f\'zo(r,G)\^dG<2 0^' "

f fr(\'zr\^ + r-^'za\^)drdo BH ^ ^ ' ' <.2D{'z, BR)<e^(R)l27z.

Using the method of proof of Theorem 3-1-8, we conclude that we m a y choose an r satisfying (4.3-11) and also the condition t h a t % ( r , a) converges uniformly to 'z{r, a) and %,ty(^, o) -^ '2,a{r, a) in L2 for a sequence
of Q ->0. And, on a^, 'ZQ(X^, 0) = 'ZQ{X^, 0) and so converges uniformly
in x^ to '2*(A;1, 0).

Now, let us choose an arbitrary point P i in G^ and map Gy conf ormally onto 5(0,1) so that P i corresponds to the origin. Then the transformed function "ZQ is continuous on 5(0,1), ^ H\{Bi), "ZQ -^"Z in Hl{Bi) and '^(1,99) converges uniformly on dB{0,\) to a function "C{^) which is the transform of';2:(r, a) on the part of^J5(0,1) corresponding t o ^ V ^ ^ ^ 'z*{x^, 0) on the part of dB{0,\) corresponding to Cf. I t follows from Theorem 3,4.5 t h a t "^ has the continuous boundary values "C. Moreover
2n

(4.3.12)

Osc "C < oj{B) + / I 'za{r, a)\da<.

co{R) + s(R)

using the SCHWARZ inequality; here co(P) is the oscilation of -2:* along OR. Finally, "z still satisfies (4.3-9) for subregions " g C 5(0,1) which we

108

Existence theorems

shall now take as circles Br = B{0, r). Replacing Hg by Z as in (4.3-5) we obtain ^(r) = D{"z, Br) <.C' (MIm) r (p'{r) (C = C(2)) from which we conclude t h a t cp{r) < ^(1) yi/^, Setting
0 r

^(1) <

e^(R)l27z.

ip[r) = J J s i / 2 | "zs{s,
0 2jr

r 2n

(p)\dsd(p, (L == Mjm)

^{r) = J J\ "zs (s, G) I ds do

00

we find, successively y)^{r)<27ir(p(r)

r

<^ e^(R)r^+^l^,

\p{r) < e{R)r(^+^)J^i^ ds

y(r) = js-^l^ y)'{s)ds = r-'^j^xp{r) + ^ / s-^l^y){s) 0 6 <e(R)'{L + 1)yi/2L. Thus we see t h a t (since "z is continuous inside)

<X{^)<iL+^)e(R). Using (4.3.12) and(4.3-13) we see that \"C{(p) "z(0)\ < e if i^ is sinall enough. Since Pi was arbitrary, the continuity at PQ follows. Before developing the notion of a continuous function monotone in the sense of LEBESGUE, which was defined in the introduction, we introduce the following notations and definitions: Notations. For scalar functions u continuous on G, we define M(u, S) = m2ixu{x),
xes

m{u, S) = min^(^),
xss

S compact C G (o{u, S) = M(u, S) m{u, S) ju+{u, D) = M[u, D) - M(u, ,dD) 1 dD) ] ^ a domain C G. fji~{u, D) == m{u, dD) m{u, D) Definition. For a scalar function u continuous on G and a real number w, we define T+{u, w) as that function U such that U{x) =^ w whenever %^ a domain D for which u[y) '> w for all y in D and u{y) = w on dD and U(x) = u{x) otherwise. We define T-(u, w) as t h a t function U such t h a t U{x) ~ w whenever x^ D on which u{y) < w, etc. The reader can easily verify the facts stated in the following lemma:

4.3. The borderline cases k=v

109

Lemma 4.3.2. Suppose u is continuous on G, w is a real number, and U = T+ {u, w). Then (i) U{x) < u[x) for x^ G and U{x) = u(x) for xf^ dG; (ii) oj(U, S) < (jo{u, S) for every continuum S G G; (iii) fi+iU, D) < iA+{u, D) and fj.-{U, D) < ^~{u, D) for every D G G; (iv) ifM{U, dD) < w, then M(U, D) < w. Corresponding results hold if U == T~ {u, w). For (i) is obvious and the others follow easily from the fact that the set of X where U(x) < u(x) is an open set Q G G which is the union of domains D on each of which ^i(x) > w with u{x) = w on dD] U{x) w on Q. Theorem 4.3.3. Suppose f satisfies (4.1.1), (4.3-1), ^^^ the supplementary conditions fix, z, p) > 0, f{x, z, p^, . . ., pr-^, 0, ^^+1, . . ., p^^) = 0, (1.3.10) r = 1 , . . . . A', ^* :^ {^j }, 1 fixed. ^ * Suppose z is continuoiis on G and ^Hl(G). Then there is a function ZQ^ Hl(G) which coincides with z on dG, for which I(zo, G) < I{z, G), and of which each component is monotone in the sense of LEBESGUE. Proof. We shall show t h a t we may replace each component z'^ in turn by a function 4 which is monotone in the sense of LEBESGUE, reducing the integral at each step. We begin with z^. For each n and each i, 1 < ^ < 2^-i, we define z'^^, +z\^y and +2:^ b}^ induction as follows: wl^ = M 2-'^'i-{M
Zi = -L (Z , ^ i i ) ,

m),
Z^^ =

M == M(z^, G),
Z^^2n-i*

=-m{z^,G)

Clearly each +.'r^ is continuous on G. We note that if D is a domain in which M ( + 4 , a Z ) ) > < + i ^ then +4+1.1W ^ *"4W < ^ for x^D. Also, if D is a domain in which 5^^+1,2 < - ^ ( ^ 4 + i . i ' ^ ^ ) ^ ' ^ i - i . i ' then M ( + 4 + i . i ' ^ ) ^ ^w+ia (otherwise H an ^0 in D where ^^+1,1(^0) > ?^i-i,i and hence a sub-domain A in which '^4+1.1 (A:) > ^^+1,1 with the equality on the boundary). Thus, if Z) is a domain in which ^ ( + 4 + 1 . 1 . dD) > <H,i.2, then /.+ (+4+i.i' ^ ) ^ 2-n-i{M - m). Now, suppose we have proved t h a t /<+(+4+i.v ^ ) ^ 2-^-1 ' [M m) (4.3.14) whenever M ( + 4 + i . az)) > < + , . , + ! ( ^ < 2^+1 - 1).

Then suppose M(+zl+^ i+^, dD) > w++^ ^^^^ Th_en, from (4-3.14), we conclude t h a t M(+zC,^,^,, D) <M{-^zl^,j, D) < M(+zl^,^,, dD) +

110

Existence theorems

+ 2-n-i(M - m); _an(i also if w^^^^,^^ < M{+zl+,^,+,, dD) < < + i . , + i , then M(+4+i^^+p D) < ^^++1,^+1. Thus we see that (4.3-14) holds with i replaced by ^' + 1. Thus it follows that, for each n, +^i+i(^) < '^z\[x) for x^Gy the equality holding ii x^dG. I t is also true t h a t the -^z\ are equicontinuous on Q (property (ii) of the lemma) and, from (4.3.14) for i = 2^+^ 1, it follows t h a t /^+(+'ii, D) < 2-^(M m) for every D C G. Hence the +zl converge uniformly to a function +z^ for which (4.3.15) f^+{+z^, D) = 0 for every D C G. Obviously the integral is reduced at each step and from the lower-semicontinuity we conclude t h a t it is reduced if z^ is replaced b y +z'^. Then, starting with +z^, we m a y form the functions ~zl^ and ~^^ in the analogous way using T" and setting w^j^ m -\- 2~^ ' i [M m). From the lemma it follows t h a t iJi'^{~z\, D) = 0 for each D and n so t h a t the ~zl converge uniformly to the desired function z^ which is monotone
in the sense of L E B E S G U E , since

^ - ( ^ 1 , D) = iu+(zl ,D) = 0 for every D C G, Theorem 4.3.4. Any family of functions z ^ Hl(G), each of which is continuous on G and monotone in the sense of LEBESGUE, for which IIV-^IIJ,^ is uniformly hounded ^ is equicontinuous on each D CC. G. If G is Lipschitz, each z is continuous on G, and the functions are equicontinuous along dG, then the functions are equicontinuous on G. Proof. Since w[z,B{x{i,r)'] = aj[z, dB{xo,r)], we see t h a t the latter function is non-decreasing in r. From Theorem 3.5.2 for ^5(0,1) with v replaced b y v 1, it follows t h a t {co[z, B{xo, S)] Y < {co[z, dB{xo, r)] f (4-3.16) ^Clj\Vaw(r,a)\'d2;, 6 <.r -< a, B(xo, a) C G, (w{r, a) = z{x)) for almost r. Multiplying (4.3-16) by r-^ and integrating from ^ to a yields
a

{o)[z, B{xo, d)]yiog(ald)

a

< Cl J ar--^ f \Vaw{r, G)\''d2J dr G)\^]''I^ dr d^^

<Clf
0

/ / ' " - i p Wr{r, o) |2 + r-^\Vaw{r,

dB{Xo,r)

{4.3.17)

^(G||V2||.)"

which proves the first result.

4.3- The borderline cases k = v

111

Suppose, now, that G is LIPSCHITZ, Z is continuous on G, and z is monotone. Let xo^dG and map a neighborhood 9^ of XQ onto Ga b y a bi-LiPSCHiTZ map X = x{y) so that XQ is carried int the origin. Let yi be any point on aal2, let R ajl, let e > 0, let Q{<R) be so small that the oscillation of C(C(y) = ^[^{y)]) along the part of aa with |y y i | < ^ is < e/2, let (3 < ^, and suppose that a)[C, G-(yi, (3)] > e. Then
COLC, dG{yi,r)] > g, d <r ^ R ,

so that

(4.?.i8) co[C, 2'(n> ^)] > /2, a < / < ^, (2'(yi, ^) = G n ^^(yi, r). ^
Then, proceeding as in (4.3-16) and (4.3.17), we find t h a t (4.3.19) {el2Ylog(QlS) < ( G l v C r r ) ' < {K4^^UY

where Kv depends only on G and v. Thus if the z are equicontinuous along dG, we may first choose Q small enough, independently of z, and then, using (4.3-19), choose d small enough so t h a t co[^, G{yi, 6)] < e. Remark. I t is easy to see that Theorem 4.3-4. holds for vector functions z if we interpret aj{z, S) = sup|2:(:\:2) -^(^i)! and define a continuous vector function to be monotone in t h e sense of
L E B E S G U E if and only if

a)(z, D) = co(z, dD) for every D C G. From Theorems 4.3-3 and 4-3-4, we easily deduce the following Theorem: Theorem 4.3.5. Suppose G is Lipschitz, z* ^ Hl(G), z* is continuous on G, and f satisfies (4.1.1), (4.3-1), ^'^d (1.10.10). Then there exists a function z-^ Hl[G) which is continuous on G and coincides with <^* on dG and minimizes I(Z, G) among all such Z. If z"^^ Hl{G) and is continuous on G, there is a similar function z which minimizes I{Z, G) among all similar
Z such that Z z"^ ^ HIQ(G).

Proof. I n each case, let {zn} be a minimizing sequence. In the first case, each Zn m a y be replaced by a zon, in which each component is monotone, and the {zon} form an equicontinuous minimizing sequence. In t h e second case, we choose an expanding sequence {Gn} (Gn C Gn+i) of LIPSCHITZ domains such t h a t U Gn =' G; and replace Zn b y a function zon such t h a t zon{x) = Zn{x) ioT X ^ G ~ Gn and zo{x) is monotone on GnThe resulting minimizing sequence is equicontinuous on each D Cd G by Theorem 4-3-4Remarks. If is clear t h a t theorems like the main existence Theorem 1.9-1, involving variable boundary values can be proved. We leave their formulation and proof to the reader.

112

Existence theorems

4.4. T h e general quasi-regular integral In this section, we study integrals in w h i c h / i s of class C^ and satisfies
t h e general LEGENDRE-HADAMARD condition

(^5.5)

fviv^^ {^> z> P) ^- h ^' ^' > 0 (all X, z, p, A, I ) .

We shall be interested in the way such functions vary with p, so we shall suppress the arguments x and z for the present. We begin b y generalizing such functions slightly as follows: Definition 4.4.1. We say that a function (p[p) is quasi-convex'^ if and onty if q){pi, + Aa I*) is convex in X for each p and | and convex in | for each p and X. Lemma 4.4.1. Suppose (p is convex on an open convex set S. Then cp is continuous on S.If\(p{^)\ <, M on S, then cp satisfies a uniform LIPSCHITZ condition with constant 2 Mjd on S^. If (pn ^s convex on S for each n and <Pn{S) ^9^(1) for each f, the convergence is uniform on each compact subset of S. If cp is convex on the interval [a, h] in R\ and a <C c <C d <C b, then cp is bounded on [a, b] by a number which depends only on a, b, c, d and its values at these points. The proof is left to t h e reader. Theorem 4.4.1. / / / is quasi-convex everywhere, it satisfies a uniform LIPSCHITZ condition on any bounded part of space. If p is given, there exis constants A- such that (4.4.1) f{K + ^o.^^)>fm + A^X.^^ foralU,f. / / / is also of class C^, then A'^ = fpo (p). Iffis of class C^, then f is quasiconvex if and only ^/ (1.5.5) holds. Proof. If is clear that if / is quasi-convex, it is convex in each single Pi for fixed values of the others. Thus it is sufficient to prove t h a t a function (p{i^, . . ., i^) which is convex in each i^ satisfies a uniform LIPSCHITZ condition on each hypercube RQ : \^P\ < M. Let R denote the hypercube 112? | < M + 1. In the case P = 2,(p, being one dimensionally convex, is continuous on dR and dRo and is bounded there b y a number Q. Hence cp satisfies a uniform LIPSCHITZ condition in f^ and i^ with constant 2 Q on RQ. In the case P = 3, we similarly get (p continuous on each face of dRo and dR. and bounded by a number Q on the part where the varying f 2 satisfy | f ^ | < M. Then again we obtain the result > t h a t (p satisfies a LIPSCHITZ condition in each ^^ of bound 2 Q on RQ. The result may be proved b y induction. The last statement is obvious (using Lemma (1.8.1)). I t is clear (from the definition of convexity) that any mollified function (pg is also quasi1 The terminology here is slightly different here than in the writer's paper (MoRREY [9]) where functions (p satisfying this condition are called weakly quasiconvex; the functions (p which we call strongly quasi-convex below were called quasi-convex in the papers cited.

4.4. The general quasi-regular integral

113

convex and ^ C^ and so satisfies (1.5.5)- From Lemma 1.8.1, it follows that Since 99 satisfies a uniform LIPSCHITZ condition near p, we may choose a sequence of ^ - > 0 so that the numbers (fgp^ip) tend to limits A'j, thus verifying (4-4.1). We now obtain a necessary condition for lower-semicontinuity: Definition 4.4.2. We say that z^ couverges Lipschitz to z ofi G if and only if Zn converges uniformly to z and the Zn and z satisfy a uniform LIPSCHITZ condition (which may depend on the sequence). Remarks. It is clear that the LIPSCHITZ convergence of Zn to z on G implies t h a t Zn 7 zin H] (G) for any s > 1 b u t does not imply the strong convergence of Zn to z in any Hj {G). Theorem 4.4.2. Suppose I{z, G) is lower-semicontinuous with respect to this type of convergence at any z on any G and f is continuous. Then (4.4.2) / / [ ^ o , -2:0, po + V C{x)] dx >/(%o, -^0, pa) ' 'yniG)
G

for any constant (xo, zo,po), any bounded domain G, and any LIPSCHITZ vector C which vanishes on dG. Proof. Let XQ be any point, R be the cell XQ<,X'^XQ-{h, ZQ be an}^ vector of class C^ on RU dR, Q be the cell 0 < .v" < 1, and f be any vector which satisfies a uniform LIPSCHITZ condition over the whole space and is periodic of period 1 in each x'. For each n, define f7, [x) on R by
CJ,(A;) = 7^-1 h CJ[n h-Hx ~ xo)].

Then the Ci tend to zero in our sense. Then, for each n, I(zo + Cn> R) can be written as a sum of integrals over the sub-hypercubes of R of side n-'^ h. If r is one of these the integral over it is n-'' h" Jf[xi
Q

+ n-^ h | , Zn{xi + n-^ h | ) ,

where r ixl^x^" ^ x1 + n-^ h, x^ = x^ + k^ n-^ h, 0 <k' <n 0 < ^ < 1. ~\ Zn{x) = Zo(x) + Cn{x), x^" = x'^ + n-^h^"", Thus we see t h a t lim7(;?o + Cv, R) = / 1 / / [ ^ > M^)> Po{x) + V C(l)] d^\ dx > I{zo, R). ^-^^ R [Q J By letting ZQ and po be arbitrary constant vectors, setting ^o(^) = ^0 + + poa' (x<x XQ), dividing by m(R) = h^ and letting h -^0, we obtain (4.4.2) for G = Q and C periodic of period 1 in each x'^. But if G is any
Morrey, Multiple Integrals g

114

Existence theorems

bounded domain and f vanishes on dG, we may choose a hypercube Q' of side t containing G and extend C{^) to be zero in Q' G and of period t in each x'. Then a simple change of variable obtains the result in general. Theorem 4.4.3. / / / satisfies the conclusion of Theorem 4.4-2, then f is quasi-convex (in p). Proof. Clearly we may suppress [XQ, ZQ). We next note t h a t any mollified function fq again satisfies the same conditions and ^ C^. Then it follows from the derivation in the introduction that (1.5.5) holds for/^. It then follows from Lemma 4-4.1 t h a t / i s quasi-convex. Definition 4.4.3. If / satisfies the conclusion of Theorem 4.4-2, we say ihdit f is strongly quasi-convex in p\ ii f depends only on />, we say t h a t / i s strongly quasi-convex. The following theorem is interesting. However, we shall not prove it here. The proof is much like but simpler than t h a t of the more general lower-semicontinuity Theorem 4-4.5 below. A proof is to be found in the writer's paper referred to above (MORREY [9])Theorem 4.4.4. If f is continuous and strongly quasi convex in p, then I(z, G) is lower-semicontinuous with respect to LIPSCHITZ convergence. We wish now to prove a lower semi-continuity theorem which involves weak convergence. The theorem is proved after four lemmas. Definition 4.4.4. Suppose C ^Hj (G) and suppose i^ is a cell with RG G. Then C is said to be strongly in Hl{dR) if and only if the boundary values B C, as defined in 3.4, ^ Hl(dR) and also there is a sequence {Cn} of class Ci on R such t h a t Cn-^C^ri Hl(R) and f^ - ^ J B f in Hl(dR). Remark. In this instance a representative of the boundary values BC is obtained by just taking the restriction to ^i? of an absolutely continuous representative , of f. Lemma 4.4.2. Suppose f ^ Hl(G). For each oc, i < oc <v, let {a', h') he the open projection of G on the x"- axis. Then there exists sets Z^ of measure zero stich that if R : c'^ <^x'^ -< d^ (oc = \, . . .,v) is any closed cell in G with
C^f {a', h"") Z and ^ ^ [a^", h^") Z, ex = 1, . . ., r

then f is strongly in Hl{dR). Proof. This is proved using the method of proof of Theorem 3-1-8. Lemma 4.4.3. Suppose R is a cell with edges 2h^, . . ., 2 A" and center XQ. Let h = minh', L = h-'^{h' /i)i/2 Suppose also that 0 <C k < h, that f * *^ Hji^) strongly in H][dR) with U*tsR<k, for some D 3 R cind f * is (s>1).

||VC*||?.<,B<M

4.4. The general quasi-regular integral

115

Then there is a function C Hl{R) which coincides with C* on dR, is zero except on a set of measure and satisfies
R

m[R) [1 - (1 - h--^ ky],

/|VC|^^A;<C-^-(1 +MS)

where C depends only on s and L. Proof. For each x ^ R, x 4^ XQ, let x * (x) be the intersection of the ray XQ X with dR, and for each x^ R define
''^^^ ~ |

- I

\X*(X)

(x = -Xo\~K\x-:Xo\

XQ)

{x =?^ Xo)

Let U^ be the pyramid in R with vertex XQ and base the face F^ where ^ = x^ A. On the pyramid J J ^ , introduce coordinates |^, . . ., 1""^, r by x"" = xl -{- r h"", xy = xy + r ^y (0 <r <\, y = i, . , ,, V i).
h^ + x-].

Then, if r and ^y are considered as functions of x, we have

r(x) - r, x*{x) = [|1(A;) + X], . , ., ^^-^x) + x^-^

Similar coordinate systems may be set up on each of the other jfj . Define r 0 ( 0 < r < 1 ^ A-i), ^^^^ ^ I hk-^{r1 + ^ A - i ) (1 ~kh-^<r< 1). Choose a sequence f* satisfying the conditions of Definition 4.4-4; and for each n, define ^n{x)^<p[r(x)]-C:[x*{x)l The each Cn(x) is LIPSCHITZ on R. B3' computing the derivatives of f in JJi^, using the equality and the fact that each ||''| ^ hv, we find that
SCHWARZ

in-

I Vf W |2 < (1 + 2L2)^-2 ^^r) I VC:(f)|2 + 2(A-)-2 9,'2| C*(|)|2 Using the facts that <p(r) = (p'(r) = 0 for r < 1 /j~i ^ and that we obtain y-i ^(r) < 1, (p'(r) = k-^h, 1 A-i ^ < r < 1, /j'' > h. l\VU\'dx<Ci{s,L)kf[\vCS\' + k-s\C^\s]dS.
8*

116

Existence theorems

Also

f\Cn\'dx:=fh-r^-^(p^(r)drf\CS\sdS<LkJ\Ct\sdS.
n; i-fc/i-^ F; F;

Adding these results for all the J J ^ . we obtain the result for each n; and also ||C||s,E is uniformly bounded. Thus, we may extract a subsequence which tends weakly in H]{R) to some function I^^H][R). Since each Cn = C* on dR, C* tends strongly in to ^* on ^i^ and l^ t,^ ondR. From the lower semicontinuity, the result follows. Lemma 4.4.4. Suppose f is strongly quasi-convex and suppose for all pi and p2, that \fip2) -f{pl)\ ^K(\Pi\s-l + \P2\^-1 + 1) \P2-Pl\. Then ifpo is any constant vector^ D is any hounded domain, andC^ HIQ(D), it follows that f[pQ + n[x)'] is summable over D and fflPo
D

+ 7t{x)] dx > m{D) -fiPo).

Proof. There exists a sequence of functions Cn, each of class C^ on D and vanishing on and near dD, such that C^ ->f in ^ J (D). For each n and almost all x on D, we have \f[pO + M^)]-f[Po + 7l(x)]\ (4.4.3) <K[7Zn(x) - n{x)] [1 + \Po + nn{x)|^-i +\Po + n{x)Y-^ . Using the HOLDER inequality, and so on, and the strong convergence in / / i p ) , we see that lim (f[p{s + nn{x)] dx = / / [ ^ o + n[x)] dx. Since / is quasi-convex, the result follows. Lemma 4.4.5. Suppose that f satisfies the hypotheses of Lemma AAASuppose also that each Cn Hl[D) and is strongly in Hl(dR) where R is a cell with R G D, and that

C^0 in LSR).

||C[J.a^M, ||CIU^M.

Then for each po, f[po + 7Zn(x)] is summable for all siifficiently large n, and liminf jf[pQ -I- Tinix)] dx > m{R) -fipo), < a W = Cix^^i^)Proof. Let K be the number in Lemma 4.4.4. For each n, let kn = Un\\leE and let h be the quantity of Lemma 4-4.3 for R- Since kn->0, we have kn <.h for all n > some n\. For each such n, let r]n be the function of Lemma 4-4.3 which coincides on dR with f^, and let
Xn^ ^n rin> 4c.:=>^L^' ^4a = zl^^' {^n =^ ^n + Mn) -

4.4. The general quasi-regular integral

117

Then, since y,^ 0 on dR, we have f f[7to + con{x)] dx >

R

m(R)f(po)-

We also have (using Lemma 4-4.3) (4.4.4) IxnlU-^O. WnnWU^M, || o) i|B < M + || ;<||o^. As in (4.4.3), we see that, for each n, and almost all x on D,
\f[pO + COn(x) + >Cn{x)] f[pO + COn(x)] \ < K ' \xn{x) I {\Po + a)n(x) + >Cn(x)\^-^ + \Po + OJn{x)\'-^ + 1).

Using the HOLDER inequality (4.4.4) and so on, we see that lim f \f[po + 7tn(x)] f[po + o)n{x)] \dx = 0, from which the result follows. Theorem 4.4.5. Suppose f is continuous in (x, z, p) and strongly quasi-convex in p. Suppose also that there are numbers m and K, K y> 0, such that (i) f(x, z, p) > w (ii) \f{x,Z,p2) -f{x,Z,Pi)\
(iii) 1/(^2,^2,:^) -f{Xi,Zi,p)\

<K[\

+ [Pi\^-^ + |/>2|^-l] \P2 - Pl\

+ \Z2-Zi\]

<i^(l + |^|.)[|^2-^l|

for all (x, z, p) with various subscripts. Suppose also that Zn 7 ZQ {weakly) in Hl(G) and that either (a) each Zn and Zo are continuous on G and Zn converges uniformly to ZQ on each closed set interior to G, or (b) the set functions Ds(zn, e) are uniformly absolutely continuous on each closed set interior to G, where
Ds(Zn> ^) = f \V Zn\^ dx. e

Then

I{zo, G) < liminf/(^^, G).

Remark. If s = 1, weak convergence in G implies the hypothesis (b). The conditions (ii) and (iii) are closely related to the corresponding ones in (1.10.8). Proof. We note first that hypothesis (ii) implies (4.4.5) (4.4.6) \f{^,z,p) -f{x,z,0)<.K\p\{\ + \p\^-i). Also, hypothesis (iii) similarly implies \f(x, z, 0) - / ( O , 0, 0)1 < K{\x\ + |.'|). Thus, for all [x, z, p), we have

(4.4.7)

\f{x, z,p)\ < 1/(0, 0, 0)1 + K{\x\ + 11 + 1^1^ + 1^1). ^

Therefore I{ZQ, G) and the I{zn, G) are uniformly bounded.

118

Existence theorems

We shall select subsequences {zrt} and form gratings &r as defined below^. We begin by selecting a subsequence, still called {zn} such that I{zn, G) - ^ i t s former liminf. For each oc, 1 < ^ < r, let (a', h') be the open interval projection of G on the %" axis, we let Z^ be the union of all the sets (for the Zn) in Lemma 4-4.2, and for each a, P, and {r, t) (or n), we let ^^^p be the set of all x'(i[a'y 6*) Z^" such t h a t ll^r||l,c?(/) < - P , G{x'') being the set of all x'^ such t h a t {x',x'^)^G, Suppose t h a t M is a uniform bound for ||^%||s,(? there being one on account of the weak convergence. Evidently |^?P I > (^^^ - a) MP-i. Now we choose P i so t h a t ilf P f ^ < (6i ai)/3 and define

Evidently \E\\^ 2{b^ a^)l} so t h a t there is a point xl-^^El which is in the middle third of the interval {a^, b^). By definition, x\^ is in E^p^ for infinitely many ;^; so we let zn = z^ where {fit} are just these n arranged in order. Next, we choose P2 so t h a t MP^^ < {b^ ^^)/3 and define ^ ^
S=l n^S

As before, there is an xl^ (^ 7t| and in the middle third of (a^, b^) and we define Z2t = ^int where {%} are those integers for which ^li^^iwtPaThis process is continued until oc = v. We define x'^^ = a"", x^^ = ^* a = 1, . . .,v. Next, we define ^J+i^2i = ^i,i ^^^ ^' = 0, 1, 2, and choose Pv+i so large that MP~^^ < (1/5) of the length of the shorter interval {XIQ, X\^) and {xl^, x\^). We define EIJ^^ = n U ^Itp^- Then there is a point :v;J+^ 1$ ^l+i and in the middle fifth of the interval [X\Q, X \ ^ . We choose 4 + i j = Zvn\ where n[ are those integers for which ^ii^E]^'p^, define '^J+i = 0 U J ' p^, choose x]^^ 3 in the middle fifth of the inter-

val {x\^, x\^ and in T J + j and define Zv^i,t = 4+i.r, where rt are those i such t h a t ^J+1.3^ ^J,^' ,p,. This process is repeated for ^ = 2, . . ., i^ f using the middle fifth of each interval (x'^^, x^^) and {x^^, x^^. The we define xl^^^^^^ = xl^^j, i = 0, . . ., 4, chooseP2V+1 so large that MP^}^^ < ( 1 / 9 ) the length of any of the intervals (:^J+i^^_i, ^J+i,^). This process is continued indefinitely taking the middle 9th of each interval for ^ = 2, . . ., r, then I/I7 for oc \, , . .,v, and in general the middle 1/(2^ + 1) of each interval. We let @jt denote the 2*'^ cells whose faces lie along hyperplanes x"" = x"^^, ^' = 0, . . ., 2^, r == (^ 1) r + a. We now let {zri] denote the diagonal sequence. Then if R is any cell of any
1 This procedure is essentially due to TONELLI (see TONELLI [8]). I t is also similar to the e d grating process used by L. C. YOUNG ([1]).

4.4. The general quasi-regular integral

119

(SA; for which R(Z G, we have each Zn strongly in Hl(dR) with ||^w||s.aiJ uniformly bounded and, of course (since that was true of the original sequence) Zn->z in Ls(dR). Moreover, the quantity L of Lemma 4-4.? is uniformly bounded for all cells in any (^jcNow, we first consider the alternative (a). Let s be any positive number. For each k, let Djc be the union of all the cells of (^jc which are interior to G. Since / is bounded below and I{zo, G) is finite, we first choose ki so large t h a t (4.4.8) I{zn, G - Dkd > - els (w 1, 2, . . .) let I{zo,DJc^)>I,{zoG)-e|5^ For this ^i, let i^i, , . ., Rqbe the cells of Djc^ and for each k>k\, Rici[i= 1 , . . .,^-2*'(^-^i))

be the cells of ^jc in Djc^. For each k, define x^{x), z^[x), p%{x) on Djc^ by defining them on each cell R of ^ to be equal, respectively, to the average over R of x, ZQ[X), and PQ{X). Then, from hypotheses (ii) and (iii), we conclude t h a t \f[(x, z^{x), Po{x)] - f[x*(x), (4.4.9) < K(\Po{x)\^ + 1) (\x - xt{x)\ z* {x), p*{x)] I + \zo{x) - zf{x)\) +

+ K[]po(x) 1-1 + \pl{x) 1-1 + 1] \Po(x) - ptix) I; the method of proof is similar to t h a t of (4.4.7)- If we let
Cn = 2n ZQ, jCn = pn pO,

we see similarly t h a t
\f[x, Zo{x),po(x) + 7tn(x)] f[x^ {x), Z^ (x), p* {x) + nn(x)]\

(4.4.10)
(4.4.11)

<K(\Pn{x)\^ + K(\Pn{x)
\f[x, <.K{\Pn(x)\^+

+ 1) (1^ - xi(x)\
-f[x, Zo{x),

+ \zo{x) - zl{x)\) + 1)
pn{x)]\

\s-i + \p^{x) + nn{x)\s'^

i)'\Zn(x) -Zo(x)\.

\Po(x)~pn^)\>

Zn{x),pn(x)]

Now, by the HOLDER inequality on each Rjd, we see t h a t (4.4.12) f \p^ {x)\s dx < I \po(x) \^ dx.

By applying the MINKOWSKI inequality, we see t h a t the integrals

(4.4.13) f\7ln(x)\'dx, f\pt{x) +7ln(x)\'dx

are uniformly bounded. Finally, (4.4.14) 1

JJkl

[\pQ{x)-pl{x)\^dx=:^0

120

Existence theorems

Hence, using (4.4.9) (4.4.14), we may choose a ^ so large that

(4.4.15) /l/[^,^oW, ^ o W ] - / K W , ^?M, A*W]K^<^/5,

(4.4.16) f\f[x, Zo(x), Pn{x)] -f[x^{x), (n = ZS(X), Pt(x) 1, 2, . . .), + 7tn(x)] \dx<8lS

and then choose ni so large that

(4.4.17) / \f[x, Zn(x), Pn{x)] ~ f[x, ZQ[X) , Pn{x)]\ d X < e/5,

vSince x^{x), z^[x), pf {x) are constant on each Rjdy it follows from Lemma 4.4.5 that liminf r/[^J(:^), zt(x),pl[x) + nn{x)']dx (4.4.18) >:jf[4{^),zt{x),pt{x)]dx.

Using (4.4.8) and (4.4.15)(4-4.18), we see that liminf/(%, G) > I(zo, G) s. The result follows in this case. We now consider the alternative (b). For each natural number q, we define
/^(A;, Z, P) = [\ ~ aq{x, z)] f(x, z, p) + m aq(x, z),

(0 ^Q{X> ^) =\^{R~ q)^ ~2(R-q)^

(0 < i? < q), (q<:R:Cq+ 1),

[l

(i^> ^ + 1 ) , i ? = (|^|2 + |^|2)i/2.

It is easy to see that e a c h / ^ satisfies hypotheses (i)(iii) with the same m and some Kq. Moreover fq is independent of {x, z) for i? > ^ + 1, and also fq {x, z, p) < /g^.1 {x, z, p), lirnfq {x, z, p) = f(x, z, p). Thus it is sufficient to prove the lower semicontinuity for each q. For a fixed q, we note that we may replace \zo{x) z^(x) \ by (pjc(x) in (4.4.9) and (4.4.10) and \zn(x) zo(x)\ by^^(%) in (4.4-11), where (pjc{x) == min(|^o(^) z^{x)\, 2q + 2), ipn{x) = mm{\zn(x) zo{x)\, 2q + 2). From the uniform boundedness of the q)]c and ipn (q fixed), the uniform absolute continuity of the set function Ds(zn, e), and the facts t h a t \ivci(p]c{x) = 0, lim^^(%) = 0

4-4. The general quasi-regular integral

121

almost everywhere, it follows that the argument can be carried through as before for each fixed q. Using substantially the same proof, we can prove the following theorem, the hypotheses of which are closely related to the corresponding conditions in (1.10.7). Theorem 4.4.6. Suppose that f satisfies the hypotheses of Theorem 4.4-5 wth conditions (ii) and (iii) replaced by (ii') \f{x,z^,Pi)-f{x,Z2.p2)\

< K ( 1 + |^i|^-l + |^i|^-l + \Z2Y-^ + 1^2^-1)

{\Z2-Zl\ (iii') + \P2-P^\) z,p)\^ K{\ + \z\s + \p\s) . \X2 xi\. 1/(^2, z,p) -f(xi,

Then I(z, G) is lower-semicontinuous with respect to weak convergence in Hl(G). A corresponding theorem holds if f = f(x,p) and zi, Z2, and z are omitted in (ii') and (iii'). Proof. The right sides of (4.4-9), (4-4-10), and (4.4.11) can be replaced respectively by (4.4-9')

+ K(\ + \zo\^-^ + \po\'-^ + iz^''^ + \pn'-') '(\^o-zn + \po-pt\)>

+
^o|.

K(\Po(x)\s

+ \zo{x)\^ +i).\x-x^(x)\

(4.4.10') K(\Pn(x)\s + |^o(^)|^ + \ ) ' \ x - x^(x)\ + + K{\ + \zo\^-^ + \Pn\'-^+\zS\^-^ + \Pt+7tn\^'^){\z-Z^\

(4.4.11') K(\ + l^ol^-^ + \Zn\'-^ + 2\Pn.\'-^) \Zn -

\Po-Pt\),

Inequalities (4-4.12), (4.4.13), and (4.4.14) hold with similar inequalities involving ZQ, Zny and z^. Also Zn -^ZQ in Ls(G). Consequently the remainder of the proof may be carried over. We can now prove an existence theorem similar to Theorem 1.9.1 for quasi-convex/. Theorem 4.4.7. Suppose that f satisfies the hypotheses of Theorem ^A-^ for some s > 1 or satisfies those of Theorem 4.4-5 "i^ith s > v, condition (i) being replaced in both cases by (4.4.19) f[x,z,p)'>m\p\^--miy m > 0.

Suppose also that F* is a family of vector functions which is compact with respect to weak convergence in H] [G) and that F is a non-empty family, closed under weak convergence in Hj {G), such that each z in F coincides on dG with a z* in F*. Then I(z, G) takes on its minimum in F. Proof. Let {zn} be a minimizing sequence, suppose Zn z^ = Wn ^ H\Q{G), Z^ being in F * . From (4.4.19) and our hypothesis on F * , it follows t h a t ||V^?z||?, llV^*||s, and HV^^wi? are uniformly bounded. From POINCARE'S inequality, it follows that \wn^s i^ uniformly bounded.

122

Existence theorems

Consequently, for a subsequence ot n, z^ 7 z* in F* and Wn7W in z* + w in Hl{G) and z = z* on dG. If / HIQ(G), SO t h a t Zn satisfies the hypotheses of Theorem 4-4.6, / is lower-semicontinuous. If / satisfies those of Theorem 4-4.5, then the Zn are equicontinuous on interior domains and we may assume Zn converges uniformly to z on such domains. Again, / is lower-semicontinuous and z is minimizing. Remarks. We have seen in Theorems 4-4-2 through 4-4-7 t h a t it is the strong quasi-convexity which plays the important role in the lowersemicontinuity and existence theory. It is an unsolved problem to prove or disprove the theorem that every quasi-convex function of p is strongly quasi-convex. We now prove a general sufficient condition for strong quasi-convexity and then prove two theorems giving examples of special forms of / for which quasi-convexity implies strong quasi-convexit}^ Lemma 4.4.6. Suppose v '>2 and ^i, . . ., f""! ^ C'^{G), Then

Proof. This is proved by induction on r. If 1 = 2, (4-4.20) is just ^

which is true. So, suppose the theorem to be true for 2 < r < X and then let C^, . . ., C^ be functions of {x^, . . ., XK+^)^C^(G). Then, if we call 5 the resulting sum on the left in (4.4.20), we obtain

5=^(-1)
a=l

dx''

2 ' ( - i ) ^ + ^ aC^ c)(Ci,

^=1

^-1^

^(S+l^

rK\

bxK+1 b{x^

+ (_1)K+1

bg\.
bxK+'^ b[x^. bx^bxK+1

, XK) b{x'^

C^)

b 6i(Ci, K+1 =(-') bxK+lb{x^,

., ;va-i, ;va+i, . . ., ;r^)

...,C^)
XK)

-f2;(-i^^+^+^ ocp^i
^=1

+ C^)

bx^^^

a=l

d;i;* d(;irl.

A^a-l,

Ar+1,

the last term of which vanishes by our induction hypothesis. But the first two terms also cancel as one sees by using the well-known method of differentiating a determinant. Lemma 4.4.7. Suppose that the functions C^, . . ., f^, // < r, satisfy a uniform Lipschitz condition on G and that one of them vanishes on dG. Then

(4.4.21)

CbJ^ J b{x^, .

.r/')

dx = 0.

Proof. Suppose C^ = 0 on dG. Choose a large cell R containing G in its interior, extend 4 to satisfy the same LIPSCHITZ condition over the

4.4. The general quasi-regular integral

123

whole space with C^ = 0 outside G, let 9 be a moUifier, and CQ he the 9 mollified functions. Clearly the integral in (4.4-21) formed with CQ converges to that for f as ^ ->0, C^ = 0 on and near dR, and Ce C' So we may as well assume f ^ C'^(R) and G = R. Then

a=l

dR

a=l

a==l

where ^^
d {x'^, . . . , ;i^-l, ;^+l, . . . , ;ir/^) '

the last equality holding by GREEN'S theorem. But the boundary integral vanishes since C^ = 0 on dR, and the integrand in the second integral vanishes on R by Lemma 4-4.6. Definition 4.4.5. A form is said to be alternating if and only if the coefficient is zero unless all the a 's and all the i's are distinct and the interchange of two as or two i's change its sign. Theorem 4.4.8. A sufficient condition that f he strongly quasi-convex is that for each p there exist alternating forms with constant coefficients such that for all n we have f{p + n)>f{p) + Atni^--+ A^i::::-y 7^:;... TTI; . Proof. This is an immediate consequence of the preceding lemma. Theorem 4.4.9. If the ajk are constants and (4.4.22) f{p) = '^npipi a necessary and sufficient condition that f be strongly quasi-convex is that (4.4.23) a^iL?i^m^^O for all X and | . Proof. If f 0 on Z) *, we see from Lemma 4-4.7 with / = 1 that > (4.4.24) jf[p
G

+ n{xy\ dx =f(p)

m{G)+f
G

afinU^) 7ij(x) dx.

The condition (4.4.23) is just (1.5-5) and so is a necessary condition. But if we introduce the FOURIER transforms (see VAN HOVE) CHy) =^ {27i)-'"^fe-i^-yCHx)dx,

124

Existence theorems

the integral on the right in (4.4.24) becomes Re f a'^i y^ y^ CH^ dy ^ 0

00

if the condition (4.4-23) holds. Thus (4-4.23) is sufficient in this case. Lemma 4.4.8. Suppose
m n

for all X and y for which

i=l?=1

Then there is a constant K such that aij = Kbij (i = i,..

X = c ^,

.,m; j = i, . . .,n).
3/ = drj,

Proof. We may introduce new variables f and rj by c and d being nonsingular matrices. Let a and b be the matrices of the original forms and A and B those of the transformed forms. Then A = c' a d, B = c' b d (cl^ = Cji). We may We shall show that there is a scalar K such t h a t A KB. assume t h a t Bii = \ {i =^ \, . . .,r)\ Bij = 0 otherwise, r

^m,n,

unless ^ = 0 in which case . 4 = 0 also and the theorem holds. B}^ taking tjs = i^irjj z=z 0 (j z^ s, s = 1, . . ., n) in turn we see that Ais = 0{i== i, . . .,m, s > r);Ais = 0(i^s,s=i,...,r,i= 1, ..., m).

Then, by choosing 1 < 5 < ^ < y and setting rj^ =.7]^ = \, nJ = 0, j :^ s, j ^ t, we have (Ai, + Ait)^i = 0 for all f with ! + f = 0. such t h a t Ats + Att = K{sJ). Thus there exists a constant K(s,t) Ass + Ast = K(sJ), Hence Aii = A22 = ' " = Arr = K, so that A = KB. Theorem 4.4.10. Suppose that N v + \ and (4.4.25) f{p)=F{Xi,...,Xr^i), where F is continuous in the Xi and

Thenf is strongly quasi-convex inp if and only i/Fis

convex in

theXt.

4.4. The general quasi-regular integral

125

Proof. If F is convex in the Xi, it follows from Theorem 4-4.8 that / is strongly quasi-convex in p. Hence suppose / is given by (4.4-25) and is stronglv quasi-convex in p. If AXjc = X^{pi + X.S^)-Xjc{pi), then it is easily seen that (4.4-26) Also, since we have (4.4.27) AXjc^Xjcvi^a^^' p^Xjc^O (^ = i, . . . , r ) ,

p^^X^j,i^=-d^fXi.

Now, choose a set of Xi not all zero and choose any p such that Xt(p) = Xi. Since / is strongly quasi-convex and hence quasi-convex, there are constants Ai such that

f{Pi + Li')^f{P)
(4-4-28) AfXocS^ < 0

+ At?i.i'.

Since / depends only on the Xi, we must have for all A, S with Xjcp^LS^ = 0 (k=z \, ., ., V + 1). Obviously, then, the equality must hold in (4.4.28). Using (4.4.26) and (4-4-27), we see that (4-4-29) P^pAXjc=^ -MXi^i), ^==\,...,v). X^AX^^DtX^^i, Dt^XjcXjcviHence since the determinant of the coefficients of the A Xjc is not zero, we must have (4.4-30) for all X, I for which ^-;.,f = 0

(4-4-31) Xi^i^O and Z)?Aaf^-0. Now, since not all the Xi are zero, assume Xjc ^ 0. Then (solving for 1 ^ in terms of the |^ with i ^ k) we obtain * (4.4.32) Z{AtX,c-AtXt)l.^i = 0 for all A, f for which (4.4.33) HmX^
i=|=fc

- DlXi) L^i = 0.

From the preceding lemma, it follows that there is a constant K such that (4.4.?4) At X^ - At Xi = K(Dt X^ - Df^ Xi).

126

Differentiability of weak solutions

Hence (4.4.35) A-,^KDt + L-Xi, L-== X^^A% = KD^).

From (4.4.29) and (4-4.35) it follows that (4.4.36) AfX-^i == KD^X^^i + maXi^i C^={KXjc-L-pl).

C^AXj,,

Finally, if we are given any values of the A Xjc, the quantities hi = p\AXic {i = \,, . .,v) and Vi = X^cAXjc are determined and the AXjc are also uniquely determined b y the hi. Using (4.4.29), we may determine the Aa in terms of the hi = {i = 1, ...,v), and substitute them into h^i = XjcAXjc (Dtho: + hv+iXi)ii = 0 = DtL^K with Xi^i^O] and we merely have to choose the f* to satisfy the equation this is always possible unless all the Df ha. = 0. Thus, unless these linear relations in A Xi hold, we have F{X + AX) = / ( ^ ^ + A.IO >_f{p) + AtXo.^^ = F{X) + The result follows in general by continuity. C^AXu.

Chapter 5

Differentiability of weak solutions

5.1. Introduction In this chapter we supply the details which were omitted in the sketch of differentiability theory which was presented in Chapter 1. I t was seen there t h a t this theory involved a study of the solutions of generalized linear equations of the form /[C,a('^^,)8 + h'u + e"") + C(c''^,a + du+f)]dx
G

= 0,

C^LipciG),

(5.1.1) in which the a*^ are bounded and measurable and the coefficients b and c^ Lv{G) and d^ L^i^{G)y and satisfy (5.1.2) (5.1.3) m | A p <a^(:^);,aA^, \a[x)\ < M f o r a.e. A;^ G and a l U ; < {Corf^^yi^. / ( l & P + |c|2 + \d\Yi^dx

We do not assume t h a t a^'^ a"-^ or t h a t h"- 0". The proofs of the theorems in this generality are new.

5.1. Introduction

127

Although it is not strictly necessary for the applications, we present in 5.2 a general existence theory for such equations and prove an interior boundedness theorem and an approximation theorem; using these theorems a theorem on further L2-type differentiability is proved. If the coefficients ^ C^, the solutions ^ C^ on interior domains. If a portion oi dG^ C ^ and the boundary values ^ C ^ along t h a t portion, then t h e solution^ C along t h a t portion. These results follow by repeated application of the general theorems. This entire analysis carries over to systems, even those where the a'lf merely satisfy the general Legendre-Hadamard type condition (5.1.4) </WAaA^|^f^'>mlA|2|||2, \a{x)\^M, m>0,

corresponding to (1.5.5), provided in this case t h a t the a'^^ are continuous at least. In 5.3 and 5-4, we present a much simplified generalization of the DE GIORGI-NASH-MOSER results from which the interior boundedness and Holder continuity of the solutions of (5-1.1) follow. I n 5-5 we present Lp and Schauder estimates of the solutions of (5-1-1) under various hypotheses concerning the coefficients. In 5.6, we treat t h e equation (5.1.5) Lu = a""^ u,o:^ + b^" u,oc + c u = f;

in this case the coefficients a"^ must be a t least continuous. Higher differentiability, both of L^-type and Holder continuous (Schauder) type are obtained in terms of the differentiability o f / a n d the coefficients. In all these sections, regularity a t t h e boundary is proved whenever possible. In 5.7, we prove that the solutions of linear analytic elliptic equations are analytic on the interior and along an analytic portion of the boundary in case the boundary values are analytic. The method consists merely in finding bounds for the derivatives. This method has been used b y
A . F R I E D M A N [1] for non-linear equations, using t h e device of G E V R E Y

(see also for systems 6.7). But the writer presents his proof for non-linear equations in 5.8 (see also 6.7 for systems) since it uses a different method and the author considers it more interesting. In 5.9, we supply the details of the proofs of Theorems 1.11.1, 1.11.1', and 1.11.2 which were omitted in 1.11. The special argument needed to prove Theorem 1.11.1" is presented in 5-10. The results of
LADYZENSKAYA and URAL'TSEVA ([1], [2], [3]) mentioned under Theorem

1.10.4(v), with k restricted to be > 2 , are proved in 5-11. In 5-12, we present t h e version of LERAY-LIONS of an abstract existence theorem, involving ''monotonicity", for the solutions of non-linear functional equations and apply it to show the existence of solutions of certain non-

128

Differentiability of weak solutions

linear elliptic equations of higher order. BROWDER, MINTY, and VisiK have each written a great number of papers on this subject; we refer in t h a t section to a recent paper of each. 5.2. General theory; v > 2 We begin b y definingi B(u, v) == Bi{u, v) + B2{ii, v), (5.2.1) Bi(UyV) == f v^ocd^"^ u^pdx,
G

L(v) = j {e'^v^oc + fv) dx

G

C(u,v) = f uv dx,
G

B^iu,

v) ^= J (v^ocb' u -\- vc^ u^oc + du v) dx

G

and we consider the equation (5.2.2) B(u, v) + I C{u, v) = L(v) which is the same as (5-1-^) except for the additional term XC{u,v). The equation (5-2.2) corresponds formally to the differential equation (5.2.3) -A(''^^,i8 + b'''^) ^''^.a dulu =f e'^^.

Lemma 5.2.1. Suppose v > 2, q ^ Lv/2(Q ^^^ satisfies (5.2.4) J\q\''^^dx<^(Corf'iY^^

GnB(x,r)

for all

B{xo,r).

Then, for each e, there is a Ci depending only an e, v, jui, CQ, and R, such
that
G

j\q\\u\^dx<ej\S/ufdx
G

+
G

Cij\u\^dx,

U^HIQ{G),

if

GcB(xi,R).

Moreover, there is a constant C2, depending only on v, such that f \q\\u\^dx

G

^260

R^i'{\\Vu\\l)^,

u^Hl(G),

GcB{x^,R).

Proof. For each r, there exists a finite sequence {C3?}, ^ = 1, . . ., P , each C^(G) and having support in some B{xj,, r) such that

Then, using the SOBOLEV lemma, Theorem 3-5.3> we obtain j \q\'\u\^

G

dx = j \q\'\u\^
G

^l:ldx
33 = 1

=^ ^
P

j \q\'\uj,\^
B{Xjr)G

dx

<(Corf'i)2:lf\up\^'dxY'^CoZir^i2;f\'7up\^dx <2ZiCorf'if
G

2:(Cl\Vu\^
V

+ \VCp\^\u\^)dx

{up=Cvu)

^2ZiCornJ

l\vu\^

+ [2'1VC2>P)|^PWA;

1 If complex functions are allowed, replace ?; by z;.

5.2. General theory; v > 2

129

from which the first result follows. The second result follows from the Holder inequality, Theorem 3.5-3, and (5-2.4). Theorem 5.2.1. There exist numbers M\ and Ao, which depend only on i^(> 2), m, M, Co, /ui, and R such that ^^^^^^ \B{u, u) I > (w/2) \\u}^ - Ao C{u, u), G C B(xi, R).

Moreover, there are constants C3 and C4, depending only on v, such that \B{u, u)\^[m^ ei(R)] (\\Vu\\l)^ >[mei{R)] (1 + i?2/2)-i (||^|1)2,

e^[R) ^ CsiCo R^^Y'^ + C4 Co R'^i.

(ll^ll = ll^lll)

Proof. This follows immediately from (5.1.2), (5.1.3), Lemma 5-2.1, and the Schwarz and Poincare (Theorem 3.2.1) inequalities. Theorem 5.2.2 (Lemma of LAX and MILGRAM). Suppose in a real {or complex) Hilbert space , BQ (U, V) is linear in u for each v and (conjugate) linear in v for each u, and suppose (5.2.6) <^ ^ \B^{u,v)\<M^\\u\\\\vl mi>0. (Tou,v). (ii) \BQ{u,u)\'>mi\u\^, Suppose To is defined by the condition (5-2.7) Bo(u,v) = Then To and T^^ are operators with bounds Mi and m^^, respectively. Proof. I t is clear that To is a linear operator with bound Mi. From (5-2.6) (ii) and (5-2.7), we see t h a t wi I ^Ip < \Bo{u, u)\ = I (To ^> ^^)| < II'^11 * I To u\\ I so t h a t

IITo ^/II > mi \\u\\.

It follows easily that the range of To is closed. If the range were not the whole space, there would be a t* such that BQ(U, v) = {To u,v) = 0 for every u. But, by setting u = v, if follows from (ii) t h a t v = 0. Thus TQ^ is a bounded operator with norm < w^^. Theorem 5.2.3. Suppose the transformation U is defined on HIQ{G) by the condition that (5-2.8) C{u,v) = {Uu,v)l veHl,{G). Then U is a completely continuous operator. Proof. That U is an operator follows from Poincare's inequality (Theorem 3.2.1), since ||C7^|| = s>uip{U u,v) = snp fuvdx
G Morrey, Multiple Integrals g

^2-^

R^\\u\\

if

||?;|| = 1.

130

Differentiability of weak solutions

Next, suppose i% 7 u in HIQ(G). Then Un -> u in L2 (Theorem 3-4.4) and \\U(un u)\\ = sup J {Un u)v dx <, 2'^^^ R \\v\\ I ^^4 -^llo - ^ 0 I
V Q

so that U is compact. Theorem 5.2.4. If X is not in a set b, which has no limit points [in the plane), the equation (5-2.2) has a unique solution u in H\Q{G) for each given e in L2{G) and f in L2sf{G), where s' = vj{v + 2). If l^h, there are solutions of (5.2.2) in which ^ 9^ 0 and e = f = 0, but the manifold of these is finite dimensional. If Ao is defined as in Theorem 5-2.1, then no real number Ai > Ao is in b. Proof. Let us define AQ as in Theorem 5-2.1 and BQ by Bo(u, v) = B(u, v) + Ao C{u, v) and define To by (5.2.7). Then, equation (5-2.2) is equivalent to (5.2.9) To ^ + (A - Ao) Uu = w, where (w, v) = L {v), L being a hnear functional since v^L2s{G) and (2s)~i + (2s')~^= 1. Moreover, from Theorem 5.2.1, it follows t h a t BQ satisfies the conditions of the Lemma of Lax and Milgram with mi = mj2. Accordingly To has a bounded inverse so (5-2.9) is equivalent to u + [X Ao) T^^ U u = TQ^ W . Since T^^ U is compact, the theorem follows from the Riesz theory of linear operators. Theorem 5.2.5. Suppose the coefficients, e, and f satisfy the conditions of Theorem 5-2.4. There is a constant C, depending only on v, m, M, Co, ^ i , and R, such that

||V^li2% < C[a-i ll^i.^ + ll^i.^. + ||/||,,d,

DcGa, GcB(xi,R), HKD) for any solution u 0/(5.2.2) with X = 0 in which u^ L2{G) and u^ for each D (Z C G. V = rj U, B{U,U) U == rj u. +

Proof. We define r] as in the proof of Theorem 1.11.1 and define Making these substitutions in (5.2.2), we obtain

(5.2.10)

+ ( [a^'^ri ocuU p a^^^r] ocuU^^ a'^^rj ocf] p\u\^

_ _

'

+ V ^"{^,0. + ^^,) + ^,aw(&' ~ c) C + rjfU] dx. Using (5.2.10), (5.2.5) for U, the CAUCHY inequality. Lemma 5.2.1 for U, and the inequalities / \fU\ dx < \fl%, II Ull, < C 11/111,, II U\l, <e{\ C7io)2 + C(||/||L,)2
G

we obtain the result.

5.2. General theory; v > 2

131

Theorem 5.2.6. Suppose G = GR, R ^ \, suppose the coefficients and e and f satisfy the conditions of Theorem 5-2.5 on G, suppose u^L2(G), u ^ HI (Gr) for each r <, R, suppose u = 0 along an, and s^lppose u is a solution 0/(5.2.2) with A == 0. There is a constant C, depending only on v, m, M, Co, and jui, such that

||Vi., < C[(R - r)-i |K||.^ + ||g||^ + ||/||,,.B] .

Proof. The proof is identical with the preceding proof except that we define Y\(%) = \ if |A;| < r , r\[x) = \ 2 [R r)~^ {\%\) r) if / < 1^1 <.{r + R)I2, r]{x) =0 ii \x\->(R + r)j2. Then u, v, and U all vanish on SGR. Theorem 5.2.7. Suppose that the coefficients a'^/ b^, cJJ, and dn all satisfy the conditions of Theorem 5.2.5 uniformly on G and converge almost everywhere to a"-^, b', c"-, and d respectively, and suppose that e'^ 7 e^ in L%(P) and fn^f in Lzs'iG). Suppose that Un-v u in H\{G) and that Un is a solution of [S.2.2)n for each n. Then u is a solution 0/ (5.2.2) imth the limiting functions. Proof. For each fixed v ^ HIQ(G), we see that <^t;,a->a^^,, in L2(G), so t h a t
Bn(Un, v) -^B{u, v), C[Uny v) -^ C(u, v), Ln(v) -^L[v).

blv,oc->h'v,oc,

etc.

Theorem 5.2.8. Suppose the coefficients and e and f satisfy the conditions of Theorem 5-2.4 dnd suppose that u^H\{G) and satisfies (5.2.2) on G. Suppose also that a^"^^ C\{D), b^"^ Hl{D), ^^ Hl{D), c^'is bounded, d^ Lv{D), / ^ L^iD), and V b and d satisfy (5.2.11) fl\Vb\^
B (a!o.r)n2)

+ \d\^f^dx<(Corf'iY'\

^i>0

for each D (Z G G. Then u^H\{D) (almost everywhere) the differential (5.2.12)

for each such D where it satisfies equation

- ^ ( a ' ^ ^ ^ , ^ + b^'u + e^") + C'u^oc + du + / = 0 .

then u^ Hl(G).

/ / G is of class CJ, the coefficients and e and f satisfy the conditions above on
G, and u^ HIQ(G),

Proof. This is proved b y the difference quotient procedure of Theorem 1.11.1. Let Z) C C G, choose D' and D" so D C G D' C C D" G C G and Z)' C D'^, let vf^ C'^(G) with support in D\ let Cy be the unit vector in t h e x^ direction and define
1

Vji(x) = h-^[v{x hey) v(x)] = f v y[x th Cy) dt,

(5-2.13) 0

uji{x) = h-^[u(x -^ hcy) u{x)], (p(x) = c'^u^oi + du -\- f, 0<\h\ < H, 9*

132

Differentiability of weak solutions

Replacing v by vn in (5.2.2) and making the indicated changes of variable, we see t h a t ^lfl satisfies the equation (5.2.14) where ^y/iW = < ' ^ W ^A^ (5.2.15) + ^^r) + hlix) u(x + hey) + b^(x) un[x) +
1

jv,^{a^^un,p

+ E^j,) dx = 0,

v^LipcD'

+ el{x) - A^yfcp{x + they)

dt.

From our hypotheses and Theorem 3.5.2, it follows that b is continuous on D". From our hypotheses and SOBOLEV'S lemma (Theorems 3.5.3 and 3.5.5), it follows that (p^L^iD"), Thus (see 3.6) for a subsequence of A - > 0 , al^{x) converges a.e. and boundedly to a^y(x), u,p{x + hey) converges in L^iD') to u^^{x), hi -^h'^y in Lv{D'), u(x + hey) ^u(x) in L2s(D'), Ufi{x) ->u^y in L^iD'), el -^e'^^ in L2(D'), and f (p{x -{-they) -^99 in L2(Z)'), so that
1

dt

It follows from Theorem 5-2.5 t h a t || V^;^||2,D is uniformly bounded, so that UfiyU^y in H\(P) for a further subsequence of A -> 0, since un -> u^y in L%{D). From Theorem 5.2.7 it follows t h a t u^y satisfies the limiting equations on D. But then one may integrate by parts in (5.2.2) to arrive at (5.2.12). In order to prove the last statement, we pick any point XQondG and map a boundary neighborhood N of XQ in the proper way onto GR by a regular map of class C\. Then (5.2.2) goes over into an equation of the same form on GR. We may repeat the argument of the preceding paragraphs for each y <,v 1, since each uji vanishes along Cr. Using Theorem 5.2.6 this time, we conclude that each u^y with y <,v 1 ^ Hl(Gr) for each r < R. This implies t h a t all u^yd with 7 < v 1 and 1 < ^ < r $ 1,2(Gr). But now a x'' > e > 0, u^v is also in HI and u satisfies (5.2.12). Since a^^ > 0, we can solve that equation for u^vr and thus conclude that it $ L2(Gr) also so that ti^v also H\{Gr) so t h a t u^ Hl(Gr). Thus, since each point XQ oi dG is in a neighborhood on which u(^ HI, we conclude that^^i^KG). Corollary 1. / / , for some boundary neighborhood N of XQ on G \J dG, the part N DdG is of class { and if the coefficients and e and f satisfy the conditions of the Theorem on N, and if u is the solution 0/ (5.2.2) in H\Q{G) then u ^ H\{N') for each boundary neighborhood N' of XQ with N' C N.

5.2. General theory; v>2

133

In case the coefficients and e a n d / $ C^ a repetition of the argument leads to the following result: Corollary 2. If the coefficients, e, and f ^ C^ on the interior of G and u is a solution o/ (5-2.2), then u^ C interior to G. If G is of class C^, the coefficients, e, and f^C'^{G), and U^H\Q(G), then u^C'^(G). Local results corresponding to that in Corollary 1 also hold. Further results are proved in 5.5. The entire discussion of this section carries over to systems of equations like (5-2.2) where (if complex functions are allowed) B(u,v) =Bi(u,v) + B2(u,v), L(v) = - J {etv% + fv^) dx

(5.2-16)

B\ {u, ^) = / ^Ja ^if ^f/3 ^^>

C(u,v) = f u^ v^ dx

B2{U, V)= f p; bf^U^ + Vi{cl^ui^ + dij U%dx

u = {u^, . . ., u^), V = {v^,. . ,, v^).

The assumptions on the coefficients analogous t o (5.1-2) and (5-1.3) are t h a t they are all measurable or in some Lp and satisfy (5-2.17) (5-2.18) ni\7z\^^al^{x)7Ti7t^, SIITI, \a(x)\<,M, m>0

/(l^l^ + kl^ + IdlY^^dx^icorf^ifK

BiXo,r)nG

In case the a'^f are continuous on G, {S.2A7) may be replaced b y the more general Hadamard type condition (see (1.5.5))* (5.2.19) m\A\^\^\^^af^{x)AocA^'^H^, \a{x)\<M, m> 0.

When this is done, we confine ourselves t o ^ ^ H\Q[G). The reason why t h a t can be done is because we see b y taking t h e Fourier transform i^ of ^ that / ^ii ^U ^!/s ^ ^ = ^ ^ / ^i^ y"^ y^ ^* (y) ^^ M ^y G ' R
^m^J\y\'^'\U{y)\^dy
Bv

= m J \\7u\^dx,
G

U^HIQ{G)

in case the a'^f are constants satisfying (5-2.19). I n case the aff are continuous on some JT D G we may, for each sufficiently small r > 0, cover G with a finite number of spheres B{xs,r), and can then choose Cs, s = 1, . . ., 5 , ^ Cl[B{xs, r)] so that ff + + f| = 1 on some domain

134

Differentiability of weak solutions

/^ D G. Then, if E(r) is a modulus of continuity of a, we find that

Bi (U, U) = j ^
G s

<f {Ul^ - Cs. a Ui) ^ - C,, ^ U^) dx

> I ^ / < f <a </J ^^ - Ci MJlul^ dx

8 G G
> I 2 / < f

sG s G - CiM (\u\^ dx >'^ y] (\V Us\^ dx -- CiM j\u\^ dx G s G G G

\V u\^ dx Cs(r,m,M)

(^^) ^ l a < ^ dx-C2

8(r) ^ f \ V

Us [2 dx

> ~

\u\^dx,

Us = CsU.

5.3. Extensions of the de Giorgi-Nash-Moser results; v > 2 Lemma 5.3.1. Suppose that (i) co ^ HIQ(G), (ii) tp and xp m ^H\(G), (iii) \p{x) > 1 on G, and (iv) ipco^oc and ip^occo^ L2(G). Then there exists a sequence { C } - > ^ *^ ^\o{^)> *^ le^A^'c/^ ^acA Cn^Lipc[G), such that y> Cw,a - > ^ co,a m Z2(<j^) ^^ ^ Cn -^CO ip in L2s(G) where s = vj{v 2). If a)(x) > 0, ^A^ Cn '^^y be chosen > 0 . Proof. We prove the lemma in the case where co{x) > 0, the other case is easily proved by writing 0}{x) = a)+{x) a)~(x) where a>^(x) > 0. By using the fact t h a t co^ ^2o(^) ^^^ ^^^^ ^Y choosing absolutely continuous representatives of ip a n d ^ co that the latter function ^ H\Q{r) on any JT I) G if we define it = 0 outside G, Thus ip oy^L^siG). For each L, we define the truncated function OOL by

{
Evidently
0 < ^

coix), ;
L,

\i x^G EL, .r ^p
II

Ej, =

{x\<o{x)>L}.

x^hL
0 < |^,a{cO COi:)| < X{EL) | V ^ | 'CO,

(co OOL) <^ipO),

ip ' (C0,a 0)L,oc) = ip0),o: '

%[EL) being the characteristic function of EL. Consequently

(5 3 2) W'{^-(^L)->0 in UsiG)

ip,oc'(o) (OL)-^

0 and

^(co,a co/,,a)-> 0 in

L2(G)

as L ^ o o . Moreover, there exists a sequence {^n}, in which each ^n^EipcG with 0 < C w ( ^ ) < - ^ , such t h a t CU-^OJL in H\Q[G) and Cw(^) -^COL(^) and VC?i(^) -> VCOL(^) a.e. in G. Then it is easy to see t h a t (i) ip' {COL Cn) - ^ 0 in L2s(G), (ii) (tp ipK) VCOL(%) ^ 0 in L2(G) as X -> cxD, Z being fixed and yjK being the truncated function of ip, and (iii) that ipK ' {V OJL V Cn) ^ 0 as n -^ oo, K and L being fixed. Con-

5.3- Extensions of the de GIORGI-NASH-MOSER results; v > 2

135

sequently (5.3.3) y){a)L Q - > 0 in L2s{G) and! } asw->cx3.

The lemma follows easily from (5.3-2) and (5.3-3) Lemma 5.3.2. Suppose that U $ H\{D) for each D C C G, that U{x) > 1 and W = U^ for some A, 1 < 2 < 2, and that P^ Lv(G). Then for each

DCGG,
U^Us{D), PU^UiD), VPF, PW, P^U^L2sr{D), PVW and s = vl(v-2), P^W^Li{D). s'== i;/(^ + 2),

Proof. Suppose t h a t D G Ga and rj is defined as usual (i. e. as in the proof of Theorem 1.11.1, etc.). Then rj U^ H\Q {G) and it follows from the SoBOLEV lemma (Theorem 3.5.3) t h a t U^L^siD)The remaining inequalities follows from the HOLDER inequality. Lemma 5.3.3. Suppose (i) that U ^ H\{D) for each D (ZG G, (ii) that U(x) ^ i in G, (iii) that w = V^^ L2(G) for some r > 1, and (iv) that W = U^ satisfies {see Lemma 5-3-2)

(5-3-4)

fCoc(a^^W^ + b'^W) + C{c'W oc +

dW)]dx<0

for each C ^ ipc(G) with C(^) > 0,

for some A with 1 < A < 2; ze^^ suppose also that a, b, c, and d are measurable, a satisfies (5.1.2), and b, c, and d satisfy (5.1.3) for B(xo, r) C G, Then w^H\ [D) for each D C. G G and j\\7w\^dx<,
(5.3.5) B(Xo,R)

C2T^a-^ Jw^dx,
B{x^,R+a)

B(xo,R

a)GG,

^ =: 1 + 4f^i^, 0 <a <R, where 2 depends only on m, M, v, A, jui, Ci, and upper bounds for R and the coefficients: C2 does not depend on r. Proof. It follows from the hypotheses. Lemma 5-3-2, and Lemma 5.3.1 with y) = t/^-i and w = rj^ U^'^ Uf-^ t h a t we may set (5-3.6) C = ^2 f72-A ^2r-2 {ri^LipcG) in (5-3-4), UL being the truncated function (see (5-3-1)) of U. Since UL,OC{X) =^ 0 a.e. on EL, we conclude t h a t Co. = 7]^ U^-' Ul^-^[{2 - A) ?7,a + ( 2 T - 2) UL,a] + 2r]r],ocU^-' (5-37) The inequality (5-3-4) becomes (again using \/ UL = 0 on EL) j {rf lPjf'^[X[2 - X)\/ U' a-V U ^ (2 - A) U b-V U + AU cS/U (5.3.8) + dU^ + {2r - 2){A^ UL' a-^^ UL + ULb'VUL)] + 2rjUUl''-^S7r]-(Aa'\7U + bU)}dx = 0. + + Ul'-\

136

Differentiability of weak solutions

Using (5.1.2), the bounds for the coefficients, and the inequaHties of ScHWARZ and CAUCHY, we conclude that (5.3.9) fv^ ^ r - ' [ | V t7|2 + (T o < Z i / ( r ; 2 T P 2 + \\/ri'^)Ul'-'^U^dx.
G

1) IV UL\^] dx (P2 =,|5|2 j ^ |c|2 -|_ | ^ | ) .

Now, let us define (5.3.10) (5.3.11) (5.3.12) VWL= UUl'^Vi]

WL =

rjUUl'K U + (r ~ 1 ) V C / L ] . + ZstI
a

Then, as was done in (5.3-7), we find t h a t + riUl-^[V It follows from (5.3-9), (5.3-10), and (5.3.11) t h a t f \\/ WL\^ dx < Z2r^ f P^ wldx
G G

\Vr]\^ U^ ' Ul""-^ dx.

Suppose that rj is defined as usual with a replaced by a/2, G by B{XQ, a) C G, and Dhy B {XQ, aj2). Then the inequalities I wldx (5-3-13) J ^ (a^l2) f \V WL\^ dx (wL^Hl^[B(xo,a)])

l\wL\^^dxY'<C3f[\VwLf-i-a-^wl]dx <.{Cz+il2)f\VwL\^dx,
BUo.a)

s=vl{v-2)

of PoiNCARE and SOBOLEV (Theorem 3.5.5) hold. Consequently / P2 ^jdx < I / P^dx^'f f \wL\^^dxY' < Z4 a ^ i / jVze;z,|2 dx,
B(Xo,a) [Ba j \Ba J B{x,a)

(5.3.14) From this and (5.3.12) we deduce that

j \VWL\^
(5.3.15) JS(a;o.a)

dx <Z^r

a-^I

Ul'-^dx,

0 < a < oc,

B {xo,a)

where (5.3.16)

2 Z2 Z4 r^ oc^i =

\. 0 < a < oc.

We can now let L -> 00 to conclude t h a t J \S7 w\^ dx ^ Z^r a-^ J w^ dx,
B{Xo,a/2) B{Xo,a)

Next, if 0 < a ^oc and B{xo, R + a) C G, we can cover B(xo, R) by a finite number of balls B{xi, a/2) such t h a t each ball B{xi, a) C B(xo, R -\- a) and there are no points in B{XQ, i^ + a) which belong to more than K{v) of the B{xi, a). Then from (5.3-16), we obtain f\\7w\^dx<2Jif\Vw\^dx<Z5ra-^2!fw^dx
B{Xo,R) B(Xi,a/2) Bix^^a)

<KZ^ra-^

fw^dx.

B{Xo,R+a)

(5.3.17)

5-3. Extensions of the de GIORGI-NASH-MOSER results; v > 2

137

For larger values oi a, (5-3-17) holds with a replaced b y oc. Thus we conclude t h a t f\Vw\^dx< KZ^ra^ oc-^ a-^ f w^ dx,
B(x,R) B(Xo,R+a)

a being an upper bound for a(<,R). The lemma follows b y using the definition of/x used in (5.3.15) Theorem 5.3.1. Suppose that the coefficients P, and U satisfy the hypotheses of Lemma 5.3-3 ^^^^ T = 1. Then U is bounded on each domain D C <Z G and \U[x)\^ ^C a-* j \U{y)\^ dy, X^B{XQ,R)
(5.3.18) B{x9,R+a)

0 < a^R,

B{xo, R +

a)cG

where C depends only on v, m, M, Co, and fx\. Proof. Let us define 5 == vl{v - 2), Wo = C7, wn = U^"", Bn = B(xo, R + 2-^ a), Wn =
Bn

jwldx.

Using Lemma 5.3.3 we conclude in turn t h a t wo^L2s{Bi), wi^ H\{Bi), W2, = \wi_\^^ L2[B2), W2^Hl(Bs), etc. Then, using the inequalities (5.3.5) and (5.3.13), we obtain

W}!^=if\wn.i\^^dxY'<Csl{\VWn^i\^
[Bn J Bn

R-^wl_,)dx

(5.3.19)

< 2C3 C2 s^^-e 4 a - 2 1 wl_^ dx = Ko Kl Wn-1,

Bn-1

where Ko = 2C3 C2 s"? a-^,

i^i = 4 s^.

From this recurrence relation, we conclude t h a t W}!^"" < i^^i-^"'^"' i^Ji-^"'^"' 'Wo = C a-' WQ . The theorem follows b y letting ^ - > ^ ^ . Corollary. Suppose the coefficients satisfy the conditions of Lemma 5.3.3, that u^L2(G), and that u^ H\{D) for each D C G G and satisfies f [v,cc{a'^u^p + b'u) + v(c'u,o: + du)] dx == 0,
G

v^Cl(G).

Then there is a constant C, depending only on v, m, M, Co (ind jui such that 1 +u^(x) <Ca-'l[\
B{Xf^,R+a)

+u^(y)]dy, B{xo,R + a)cG.

x^B{xo,R),

0<a^R,

Proof. Define V^{x) = 1 + u^{x) and set

v = ipV~^u,

if^LipdG),

y)(x)^OonG

in the equation above. An easy approximation shows that this is legitimate. Since V V^x = u u^ot so a'^V,ocV,^ = V-^ u^ a""^ u,ocU,^ <, a''^ u,ocU,p,

138

Differentiability of weak solutions

one sees t h a t V satisfies fy^,cc{a^^ F,a + 'b^ V) + ipCC^ F a + 'dV)

G

dx<0 F-2) ^.

'h^ = (1 -

F-2) 6,

'c^ = c''+

F-2 6 ^

/^ = (1 -

For the purpose of proving the differentiabiHty of the solutions of the variational problems and the weak solutions of the other differential equations (1.10.13). it is sufficient to study equations of the forms (5.3.20) (5.3.21) / ( v C - ^ - V w + C-c-V^)i;^ = 0
G

/ [VC( 'Vu
G

+ e) + C{c ' "^u + / ) ] dx = 0

CLiPeG,

in which the a^ satisfy (5-1.2), c satisfies (5.1-3) {b = d = 0), e^ L2(G), and f^L2sf(G), s' = vl{v + 2), and e and / satisfy certain additional conditions; we suppose, of course, t h a t u^H\[D) for each DGGG, The more general equations (5.1.1) are interesting in themselves and are treated in a paper by the author (MORREY [14]). Definition 5.3.1. A function v ^ Hl{D) for each D C C G is a subsolution of (5.3.20) if and only if f{V^
G

-a-Vv

+ Cc'Vv)dx^O

for each

C^Lipc(G),

C(^) > 0.

(5.3.22) Remarks. This condition is formally equivalent to the condition -^r^a^'^VR C'v^o^ > 0. Lemma 5.3.4. Suppose that (i) F is non-negative and convex on the interval [0, 00), (ii) H = e~^ is convex on that interval^ (iii) u is a nonnegative solution 0/(^.3.20) on G (iv) V{x) =F(u{x)], and (v) v^L2{G). Then v is a sub-solution of (5.3.20) on G and J\Vv\^dx<C
D

a-^\G\

ii DcGa,

GcB{xuR)

where C depends only on v, m, M, fxi, Co, and R. Proof. First, we assume t h a t i / e C2[0, CXD), -.\
^H(U)

<-e

{s>0)

and t h a t H" is bounded on [0, 00). Then F^ C2[0, 00), and F, F', and F" are bounded there with F"[u) > [F'(u)]'^. Let us set C = ^^ F'[u) in equation (5.3-20), where rj is defined as usual. I t follows t h a t 0 =^ j [ZrjSJrj' a-Vv
G

-{- rj^F''{u)

^ u - a - S7 u -^ rj^ c ^ V v] dx + 2riyrj'a'SJv'] dx,

^ f bl^iVv

a'S/v-\-C'S/v)

5.3- Extensions of the de GIORGI-NASH-MOSER results; v > 2

139

since F" > {F')^, Finally (5.3.23) jri^\S7v\^dx^

G

Cf(rj^c^
G

\Vrj\^)dx

from which the inequality follows easily, using Lemma 5-2.1. In the general case, H is convex with 1 ^ H(u) < 0 on [0, oco). It is easy to see t h a t H can be approximated from below by functions H having the properties in the preceding paragraph. I t follows t h a t the functions Vn(x) ->v(x) from below and hence strongly in L2{G). Clearly, also, Vn 7 V in H\[D) for each D (Z G G, on account of the inequality (5.3.23) which holds for each n. The inequality holds in the limit by lowersemicontinuity. Theorem 5.3.2 (HARNACK type). Suppose that (i) u is a non-negative solution 0/(5.3.20) on B^R^ B(xo, 2R) and (ii) the set S where u[x) ^ \ has measure > ci |52i?|, ci > 0. Then u{x) > C2 > 0 for x^
BR

where c^ depends only on v, m, M, Co, /ui, and ci. Proof. There is a k, 1 < ^ < 2, such that \B2R BJCR\ = (1/2) ci 1^2121. Then | 5 Pi BJCR\ > (1/2) ci \BJCR\. Let us define F(u) = max[log(w + s), 0], where 0 < e < 1. It is easy to see t h a t F satisfies the hypotheses of Lemma 53-4. Consequently f\Vv\^dx<Ci R"-^ where v{x) = F[u(x)'\.

Since v{x) = 0 on S and \S C\BJCR\ Theorem 3.6.5 t h a t

> (ci/2) |5A;i2|> if follows from

fv^dx^C2Rr
The theorem follows from this and Theorem 5-3-L Theorem 5.3.3. Suppose u is a solution of (5.3-20) on G. Then u^ C^^(G) where 0 < //o < 1 ^'^d juo depends only on v, m, M, Co, and /Lii, More precisely \u(x) u(xo) I < C L a-^ (\x xo\IRY\
L = \\U\\IE-,S, B{xo,R + d)GG,

x^ B{xo, R),
r = vl2,

where
d<R,

and C depends only on v, m, M, Co, and jui. Proof. It is sufficient to prove the inequality. It follows from the corollary to Theorem 5.3-1 that \u{x)\ < CiLd-\ X^BR-^B(XO,R). Let us define w * and M * as the essential inf. and sup. of u{x) on BR and let us choose m (unique) so t h a t 15| ^\BR\I2, S+ and S " being the sets of points x^ BR for which u(x) > m and u{x) < m, respectively. If m* <m <M*, the functions [M* u{x)]l{M'^ fn) and

140

Differentiability of weak solutions

[u(x) m'^]l{fn m*) satisfy the hypotheses of Theorem 5.3-2 on BR with ci = 1/2. I t follows t h a t wi < u{x) < Mi for x^ BR/2, where fn^ =: m h(m m*), Mi = m + h{M* m), A = 1 C2 < 1, C2 being the constant of Theorem 5.3-2 with ci = 1/2. The same results hold if w = m * or w = M * or both. Now, let us define
q)(r) = [ess sup ^(A:)] [essmiu{x)] for x^Br, r <R.

We conclude from the preceding paragraph t h a t (p{2-^R) <^h^S, Thus log(p(r) < logS log A -\- {n + i) log A <log(5/A) + [(log/^)/(log2)]log{2?/r), if n log2 < log{Rlr) < (^ + 1) log2. From this it follows that <p (r) < h-^ S {rjRy', //o = log A/log 2. Theorem 5.3.4. There are constants Ri > 0 and C which depend only on V, m, M, Co, and jui, such thst
\\Vu\\l,<CL(rlRy-''+"\ 0 < r < R , L = \\VU\\IK, r = vl2,

S = 2CiLd-\

n= \,2,,.

for each R,0 <iR <,Ri, and each solution 0/(5-3-20) with || V ^HH.B < + ^><^Proof. Evidently we may suppose t h a t the average value of z* = 0. From Theorem 3-6.5, we conclude t h a t
\\U\\IJ,<CLR.

From Theorem 5-3-3 we then obtain I u(x) - u(xo) I < Z . h||o^^ . ( i ^ / 2 ) - (1^ <.Z2L'R^---f'^'\x-xo\''\

xolIRr

\X-XQ\<.RI2.

We define Y} as usual with a, G, D replaced b y ;', B{xo, 2r), and B{xo, r) respectively, and p u t (5-3-25) C{x) =r]^'[u(x) -u{xo)], x^B(xo,2r), 0 < r < i?/4 Vu}dx, in (5-3-20). We obtain 0 = J^rj^[\/u ' a ' \7u + c{u UQ) ' Vu] -\- 2rj{u UQ) S77]' a -

The theorem follows easily by using (5-3-24) and the inequalities of

CAUCHY and SCHWARZ.

Theorem 5.3.5 (Existence and uniqueness in the small)- There is a constant JR2 > 0 which depends only on v, m, M, Co, and jbti such that there exists a unique solution C/ o/(5-1 -1) which^ H\Q \B (XO, R)] for each R < R^, e^Lz B(xo, R), / ^ L9,sf [B(xo, R)], such that B(xo, R) G G; this solution

5.3- Extensions of the de GIORGI-NASH-MOSER results; v > 2

141

satisfies where C depends only on the quantities above. If u*^ Hl[B(xo, R)], there is a unique solution H of the homogeneous equations (5.1-1) such that H u*^ HIQ[B{X, R)]. This solution satisfies lVH\\l^<C-'\\u*lln; if b = d = 0, '11^*111.^ f^^y be replaced by ||VW*||2.RProof. Using (5.1-3) ^^nd the notations of 5-2, we see that Bi(u,u) ^mU', 1^2(^,^)1 ^L'\ using also the L = IIV^i.R f\d\'U^dx / ( I ^ P + \c\^)u^dxY^ +

inequahty (5.3.14). Thus B{u,u)^(mj2)L^ if 7^ < i^2- Then B satisfies the hypotheses of the Lemma of LAX and MiLGRAM. We note also that \L(v)\<{\\e\\U+C\\m_^) where C is the CJ^^ in (5.3.14). The last statement follows from POINCARE'S inequality (Theorem 3-2.1), Theorem 3.5-4 and the fact that U = H u"^ satisfies (5-1-1) with e^= ^^^*^ + 6w*, f=c'u% + du''.
SOBOLEV

Theorem 5.3.6. Suppose that u is a solution of (5-3-21), e^ L2(G),f^ L2st{G) {s' = vj[v + 2)), and e and f satisfy j\e\^ (5-3-26) dx < Lf (;'/i^)^-2+2A^, 0<^</io,

u^H\{G),

0<:r<i^<i?2, B(xo, R)cG.

/ I / I dx < L2 R'-^ {rlRy-^^i^,

Then u^ C^ (G) and, in fact, u satisfies a condition of the form (5.3.27) I|Vi|r^||VMi,B<,..rt<if(r/2?r-'+", 0<r<R<R2. Proof. Let V be the potential of / over G. Then it follows from Theorem 3.7.3 that V satisfies a condition like (5.3.27) and also that
G

Thus equation (5.3.21) reduces to j [\7 C(a ' V u + e V V) + C c ' V u] dx = 0

G

in which e ~ \7 V satisfies (5-3-26) with Li replaced by another L. Thus we may assume that / ^ 0.

142

Differentiability of weak solutions

By virtue of Theorem 3-5-2, it is sufficient to prove that (5.3-27) holds for some K. Since R < R2, we conclude, using Theorem 5.3.5 that u ^ U + H on BR where U is the solution of (5.3-21) on BR which
^ HIQ(BR) and H is the solution of (5.3-20) such t h a t H u^ H\^{BR),

and t h a t (5.3.28) l l V f / | l i ? < Z i l ^ l | ^ < Z i | ^ | | ^ , ||Vi/||/?<Z2||V^||i?<C2||V^||(?. Then it follows from Theorem 5-3-4 that Now, let us define (p{s) = L " i sup ||V t/f^s for all e which satisfy (5.3.26) with Li replaced b y L, i^ replaced b y 5 < i?, U being the solution of (5.3.21) ^HIQ{BS). Next, choose an arbirtary e which satisfies (5.3.26) (Li replaced b y L). We may write U = Us + Hs on Bs where Us is the solution of (5.3.21) ^ HIQ(BS). Obviously e satisfies l\e\^dx< [L^iSjRy-^^^'']' (rlSY-^+^^, 0<r<S.

Thus, using the ideas of (5.3-28) and the definition of 99, we conclude t h a t IIV UsWs ^ZiL^ (S/i^)-l+^ IIV HsWs < Z2 IIV t7||^ < Z2 L (p{SIR). Now, suppose t h a t 0 <r < S <: R. Then | | V ^ | | r < | | V t 7 ^ | | . + ||Viy5||r ^L{SIRy-^+^(p(rjS) +ZsL <p(SIR) {rlSy-^+f^-. Since e is arbitrary, we conclude (setting s = rjR, t = SjR) t h a t (5.3.29) (p{s)^t^-^+f'q)(slt)+Zsq^{t)'{slty-''+f'\ Obviously cp is monotone and 99(1) < Zi. So let us choose a, 0 < cr < 1, Then, obviously ^(s) < So s^-i+^, (5.3-30) (p(s) < Si 5^-1+'^, cT < s < 1, where So < Z i a-^+i"". co = a^^'''. Using (5.3.29) with cr^ < s < a and t = (r~i s, we obtain Si = So(1 + Zsco), Since Si > So, (5-3-30) holds for o ' 2 < s < 1 . cr^ < s < 0*2 and t = a~^ s, we conclude that Using (5-3-29) with

^(5) < S2 5^-1+'*, 0r4 < S < 1 ,

S2 = So(1 + Zs (O) (1 + Zs 0)2) .

By repeating the argument, we obtain

(p{s) <Ss^-i+^, 0 < s < 1, S = So(1 +Z3C0)(1 +Z3a>2)(l +Z3ft>4)...

from which the theorem follows immediately. The following theorem concerning the boundary behavior of solutions of equations of the form (5-3-21) is very easily obtained. More general theorems concerning equations of the form (5.1.1) are presented in the author's paper referred to above (MORREY [14]).

5.4. The case v = 2

143

Lemma 5.3.5. Suppose u is a solution 0/(5.3.21) which ^ H\[Ga) and which vanishes along Oa- We suppose that e and f satisfy the hypotheses of Theorem 5.3.6 with B{xo, r) replaced by B{xo, r) 0 Ga for all XQ such that B(xo, r) C B{0, a). Then, if we extend u to B{0, a) by (5.3.31) u(x^,xl) = -u{-x\xl) it follows that u^ C^{Ba) and satisfies (5.3.27) '^ith' G = Ba. Proof. For, suppose C Cc(Ba), and we extend u, a^^ a"^, e"", c'^, and / , for oc <v, b y (5-3.31) and extend a^"^, a"", e", and c^, for a a n d /? < 1^, b y the formula (5.3.32) Then / [ V f (a-S/u (5.3.33) = J[V C*'(ci-S/u
Ga

(p(x\xl)=(p(-x-,x:). + e) + C{c' Vu + / ) ] dx + e) + C*{c'Vu+f)]dx

where

C* (^^ K) = C (^', < ) - C ( - x\ < ) , ^* = 0 on dGa.

Accordingly the extended function u satisfies (5.3.21) as extended to Ba and e a n d / s a t i s f y conditions like (5.3.26) there. The lemma follows. Theorem 5.3.7. Suppose G is Lipschitz, u is the solution 0/ (5-3.21) in HIQ(G), and e^ Z,2(G),/^ L2s,(G), and e and f satisfy

(5.3.34) (5.3.35)
(5.3.36)

J\e\^dx<Llr'-^+^f' f\f\dx<L2r^-^^^
B(x.f,,r)nG

Then u^ C^{G) and satisfies a condition /|V^|2^A;<i^2;..-2+2^^

B{Xo,r)f\G

K'^<^C\L\

+ CILI

where Ci and C2 depend only on v, m, M, jui, Co, JLC, and G. Proof. The condition for interior spheres is just Theorem 5.3.6- Each point of 5G is in a set A^ C ^ which can be mapped on Ga b y a bi-LiPSCHITZ map as in 1.2, Notations. The form (5.3.2I) is preserved b y the mapping and the transformed e and / satisfy the conditions of Lemma 5.3.5. The result follows. 5.4. The case v = 2 The theory of 5-2 and 5-3 carries over t o case v = 2 with only rather minor modifications. I n fact the theory can be greatly simplified in this case. However, there is a slight complication caused b y the fact that the exponent s' v j{v + 2) = 1/2, so the exponent 2s' = \

144

Differentiability of weak solutions

and we cannot conclude from the CALDERON-ZYGMUND inequalities that the second gradient ] V^ F] of the potential F of a function/in L\ is again in L\ and so cannot immediately conclude that V F ^ L2 as in Theorems 3.7.1 and 37.3. Lemma 5.4.1. Suppose u ^ H\Q{G) and q ^ Li{G) and satisfies (5-2.4) j\q{x)\dx<.CQrf', for all B{xQ,r).

Then q u and \q\ \u\^^ -^i(^) ^'^^ (5.4.1) (5.4.2) j\q{x)'u[x)\dx^C^'C^'\\Vu\la'g"^'r-^-"\ / I q{x) I I u{x) \^dx < C2 Co(|| Vi^i.)2g'/V'^--^/2^

0 < / < /^, q and u may he tensors; Ci and C2 depend only on X and fzi. Proof. It is sufficient to prove this for vectors u^ Q(^)- Then (5.4.3) Ui(x) = - (27r)-l / I I - :^ |-2 (|a _ ^) ^i^(l) ^1
G

using Theorem 2.7.3. Hence

B(Xo,r)nG

Jl q{x) u{x) I ^:\; < (27i)-^f

B{xo,r)nG G

j\q[x)

\'\S x\-^'\\/u{^)\

d^dx.

(5.4.4) Applying the

-B{o.r)n^

SCHWARZ

inequality to (5-4.4), we obtain

1/2 lGGf]B{Xo,r)

/ I ^(^) ^^(^) I ^^ ^ (2^)-^ f /^^ / I ^ - ^ P~^ * I ^(^) I ^^X X [ / ^ f / I q(x) I I ^ - ^ !"^^ I Vw(|) |2^:vl^^^
[G GnB{Xo,r) J

(5.4.5)

0 <;.<//].

Using Lemma 3.4-3, we see that (5.4.6) / ^ f / I f - ^ M k(^) N ^ < 27r A-ig^ Co ^'^^
G GnB(xo,r)

Next, we define, for each | ,

5(l.e)n-B(a;o.r)nG^

From our assumption on ^, we see that Accordingly

.r)r\G B{x^,r)nG
00

(P^{Q) < Co Qf^i and

Co ^^1.

/ 11 ^ r^ i $'(^) \dx<,

j Q-^ (P'^(Q) dQ
0

(5.4.7)

5.4. The case v = 2

145

The first inequality follows easily from (5.4.5), (5-4.6), and (5.4-7). The second follows from two applications of the first with numbers A' and A" with A' + A" = 1 Lemma 5.4.2. Suppose q satisfies the condition of Lemma 5.4-1. Then, for each e > 0, there are numbers C3 and C4 depending only on e, v, Co, //i, and R such that f\q\'\u\^dx<C^CoRf'^J\Vu\^dx
G G

u^ + CsJ\u\^dx
G

^1^(0), R).

J\q\'\u\^dx<ef\Vu\^dx
G G

G<ZB[XQ,

Proof. The first follows by setting g = r = R in (5-4.2). To prove the second, let / > 0 be given. There exists a set of ^p as in the proof of Lemma 5-2.1. Then, as in the proof of t h a t lemma,

Jl ^ I I ^ |2 ^:v < 2 ^ / I ^ I I ^3? |2 ^^ < C Co r^i^' / I V % P ^A;

G 'P G 'P G

<, 2C Corf^i f \\/u\^ dx + 2C Corf^i f 2^\ S/Cp M ^ P dx

G G

using the first result and the fact t h a t each ^p has support in some B (xp,r). Using these lemmas instead of Lemma 5.2.1, we easily prove Theorem 5.2.1. In Theorems 5-2.4 and 5-2.5 it is necessary to require t h a t / s a t i s f y a condition of the form (5.2.4) in which case we find (using(5.4.1)) that J\fv\dx^CiCo-\\Vv\\%'R^i
G

f\fU\dx<CiCo-\\VU
G

||.e i?"! < (IIU l|.o)2 + C -1 R^"! Q

in the proofs of those theorems. In Theorems 5-2.5 and 5-2.6, the results are, respectively, I V u \\% < C [a-i 11^ ||o^^ + ||. Ijo^^ + Co R^i], I V ^ i . , <.C[{RI r ) - i ||^||o,^ + ||.||,^ + Coi^^i], provided that / always satisfies (5-2.4). In Theorem 5-2.7, it must be assumed in addition t h a t t h e / ^ a n d / s a t i s f y (5-2.4) uniformly. The statement and proof of Theorem 5-2.8 carry over, using Lemmas 5-4-1 and 5-4-2 to show that d u and hence 99^ L^. The theorems in 5-3 can be proved in much the same way as they were but replacing vjiv 2) by an arbitrarily large but fixed finite number 5 > 2; the SOBOLEV inequality (5.3.14) holds for any s in the case r = 2. However, the theory can be greatly simplified in the case y = 2 by using the original proof of the writer as modified to make use of recent simplifications and to take care of the lack of self-adjointness in the equations (we have not assumed a^' = a'^ nor b"^ = c*). We now
Morrey, Multiple Integrals |Q

146

Differentiability of weak solutions

present this simplified theory. All the theorems and their proofs generalize immediately to vector functions except, of course, the last theorem which concerns sub-solutions and corresponds to Theorem 5.3-'ITheorem 5.4.1. Suppose u is a solution ^H\{G) o/(5.1.1) in which the coefficients satisfy (5.1.2) and (5-1.3) wth v = 2. There are numbers C and f^o, 0 < /^o < /Mi> which depend only on v, m, M, Co, and /ii such that if e and f satisfy (5.4.8) / I e\^dx<
B{xii,r)

L\{rla)^f',

j\f\
Bixo.r)

dx < L2(r/a)^

0 < jbt < jLto, 0 < y < a, B (xo, a) C G, then u^ CI (G) and satisfies (5.4.9) J\Vu\^dx< C{L\ + Ll + L^ an) {rlafi",

L^ - (11 ^^ Wia)^ ^ / (I V ^ |2 + a-^\u |2) dx,

B{xo,a)

B(xo,r)cB{xo,a)cG. Proof. Let us consider a ball B {XQ, a) C.G and let us redefine u in B{XQ, 2 a) B(xo, a) so that the new u^ HlQ{B2a) and \\'^\\U,Bix.,2ay<.Zl{y)L. (Theorem 3-5.4)- Let us define E^" = b'^u + e^", (5-4.10) F = c'^'u^oc + du + f, + vF]dx = 0, x^B(XQ, a). E and F being 0 elsewhere. Then u, E, and F satisfy l[v,4a^^u,p-\-E^)
B(Xo,a)

v^Hl^{Ba),

(5-4.11) ||i7||o^^<i:2(y/^)^4.i:[ci/V"^/2 _|_ CiCo(2a)^/v^^-^/2]^

as one sees from (5.1.3). (5-4.8), and Lemma 5-4.1 with G = B{xo, 2a). Now, let us choose (i) polar coordinates with pole at XQ, (ii) a representative u which is A.C. in d for almost all r and (iii) an i?, 0 <iR <i a, such t h a t u {R, d) is K.Z. in d with UQ^L^, Write v = u - H, H(r,d) =H + {rlR) [u [R, d) - H], on BR

where H is the average oi u(R, 6) with respect to 6. We note that (5.4.12)

>

fFvdx->
Bixo.B)

-Zi{L2
-

+ Z 2 L a n l ^ ) [Rja)^ ||V?;I^R
{Z^Ll+Za^an)e-HRIa)^^.

-e{\\Vu\\lR)^

8(\\VH\\IR)^

Using (5.4.11) and (5-4-12), we conclude from (5.4.10) that (5.4.13) j\Vu\^dx<.Z^j\\/H\^dx+Z^{Ll
B{xo,R) B{xo,R)

+ Ll +

L'^an){Rla)'^i^

5.4. The case v = 2

147

where we have assumed that /j, < /^i/2 and ju = (//i/2) (A/4). If we now choose juo so that /^o < /*i/2 and juo < 1/2 Z5 and if we set (p(R) = J \Vu\^dx, we see that f\VH\^dx
o,B) B{Xo,R)
2JT.

B{Xo,R)

= I fr{Hf +
0 0

r-^Hl)drdO

= lf{ul{R.
0
2JT

6) + [M(i?, e) - ] 2 } i 0 0)d6<Rcp'(R).

<ful{R,
0

Then from (5.4.13)> we conclude that cp{R) < (2/.o)-i i^ 9^'(i^) + Z,{Rla)^^, Z, = Z^{L\ + Ll + L^ a^^i) from which we easily conclude (if 0 < ytt < //o), that cf>[R) < [^() + Z8(Z| + L| + Z2 ^A^i) (2^/^)2^ On order to handle variational problems of degree other than 2 (i.e. y :5^ 2 in (1.10.7) or (1.10.8)), it is necessary now to prove the lemmas ^ corresponding to Lemmas 5-3-1 5.3-3 and then prove the theorem corresponding to Theorem 5.3-1Lemma 5.4.3. Suppose (i) that co ^ HIQ(G), (ii) that yj and xp co ^H\[G), (iii) \p{x) '> \ on G, (iv) y)co^x and yj^aOJ^ L2(G), (v) P^L^ {G) and satisfies (5.4.14) J P^{x)dx<Corf'i, P{x)^0.
B(Xf,,r)G

Then there exists a sequence {^n} -> o) in H\Q [G) , in which each f^ ^ Lipc{G), such that ip Cn,oc -^ip o),oc and PipCn-^Py^oj in L2{G). If oj{x) ^ 0, the Cn ^(^y be chosen > 0. Proof. As we noted in the proof of Lemma 5.3-1, ipo)^H\^^{r) on any r ' D G if we merely define ip a> =^ 0 on F G. Thus it follows from Lemmas 5.4-1 and 5-4-2 that Pipoj^L^ (G). The remainder of the proof is essentially the same as that of Lemma 5-3-1Lemma 5.4.4. Suppose that U ^ Hl{D) for each D CG G, that U{x) >: i on G, that W{x) = [U{x)]^ for some A, 1 < A < 2, and that P satisfies hypothesis (v) of Lemma 5.4-3 on each D (ZC.G with Co = Co [D). Then, for each such D, VW, PW, PVW, and P^w^Li(D). Proof. For PU, P U^-\ and V C/ ^ L2{D). 10*

148

Differentiability of weak solutions

Theorem 5.4.2 (Existence and uniqueness in t h e small). There is a constant i?2 > 0 which depends only on v, m, M, Co, and jui such that there is a unique solution U of (SAA) which ^ HIQ[B [XQ, R)] for each R < R2, each e^L2{Bji), and each f^Li{BR) satisfying (5.2.4); this solution satisfies llV?7i.c<Ci||e|lB+C2Coi?''i, where Ci and C2 depend only on the quantities above. If u"^^ Hl[B {xoy R)], and R <. R2, there exists a unique solution H of the homogeneous equation
such that H u*^ HIQ[B (XQ, R)].

The proof is like that of Theorem 53-5Theorem 5.4.3. Suppose (i) that U ^ Hl{D) for each DCCG, (ii) that U(x) > 1 in G, (iii) that U^ L2 (G), (iv) that W = U^ satisfies (5.3.4) for some A with 1 < A < 2, (v) that a satisfies (5.1.2) and that b, c, and d satisfy (5.1-3) on each D CCG with Co = Co {D). Then W^H\{D) and is bounded on each D G G G. Proof. Since U ^ Hl{D) for each JD C C C, it follows that i7^ ^ L2(D) for every T > 1 and every D G G G. We m a y then repeat the first part of the proof of Lemma 5-3-3 setting where rj has support in D' and D G G D'a. The result is equation (5.3-12) with G replaced b y D' and T b y A: /1 \7WL Pdx < Z2 A /P2 wldx + Z^XJ\ Vfj p U^ Uf'^dx. 2
D' Df D'

Using Lemma 5-4-2 with Z2^^ < 1/2, we find t h a t (5.4.15) j\VwL\^dx<^ j [Z^rf +Z^\\/rj\^)mUf-^dx,

It is now possible to let L -> 00 and this obtain W ^H\ {D). But now, pick a hal\B{xo, R) G D' where R <. R2 (which now depends on D') and let H be the solution of the homogeneous equation (5.1.1) which coincides with W on dB {XQ, R). Let w = W H. Then z^ =: 0 on SBR and satisfies (5-3-4) there. If we then set C = ^"^, we see that ^^^ = w^oc and C = ?^ if z' > 0 ! C,a = f = 0 if ze; = 0 Accordingly (5-3-4) with w and this f becomes (a.e.)

/ [vC ^ vC + (^ + c) c vC + dC^] dx<o ^

B{xo,R)

which implies that f = 0 since R ^ R2. Thus w <,0 or W <, H on B{XQ, R). But from Theorem 5-4-1, it follows that H is Holder continuous interior t o B{xo, R).

149 5.5. Lp and Schauder estimates We obtain further estimates concerning the solutions of equations (5.1.1) when the a'^ are continuous and the coefficients and e a n d / satisfy various supplementary conditions. In case the a^"^, }f-, and e"^ ^ C^^ and the c, d, and / satisfy certain supplementary conditions, we shall show t h a t \l ufC^^. We shall obtain corresponding results for weak solutions of systems in 6.4. These results were obtained in the two dimensional case in MoRREY [7], Chapter VI, 6, where a different method was developed to treat the case where the coefficients a'^f were constants satisfying (5.2.17) and b = c = d = 0. Let us set an = aij, alf ^ afl = bij, a^f = Cij. It turns out t h a t if the u^^ H\{G) and satisfy (5.1.1), i.e. JlKi^ij
G

K + hj K + di) + oji(bji ul + Cij u\ + ei)] dxdy

= o,

then there exist conjugate functions vi such t h a t -Vix = [hi K. + Cij u% + ei), Viy = atj 4 + btj u{ + di. If we let w be the 2Ai^-vector [u^, . . ., u^, v^, . . ., v^) and g be the 2 A^vector [di, . . ., d^, ^1, ., ^iv^), these equations become Aw.x + Bwy + g=^0 where A and B are 2N X 2N square matrices having a particular form. By using the ellipticity and certain theorems on A-matrices (see BOCHER, 9196, it is seen t h a t we may find affine transformations w = D w' and g' =^ C g so t h a t the transformed equations have the same form as the conjugate equations above but with

where h is arbitrarily small. Moreover, the maximum and minimum magnifications of the transformations and their inverses are bounded by a constant depending only on m, M, and h. An interesting generalization of this method was employed b y C . MIRANDA ([1]), who introduced exterior differential forms (see Chapter 7) instead of the conjugate functions. We begin by considering equations of the form (5.1-1) on spheres BR or hemispheres GR where the symmetric part of the a^^ matrix is just d""^ and prove local differentiability. A neighborhood of each interior point can be mapped on such a sphere by an affine map in which the maximum and minimum magnification and those of its inverse are uniformly bounded by a constant depending only on m and M. In the case of a

150

Differentiability of weak solutions

boundary point, one can first perform the affine transformation as above, follow t h a t b y a rotation to make x"^ ~Q t h e tangent plane, and follow t h a t with a m a p of the form
(5.5.1) V = ^ ^ ^ = 1, . . ., 1; 1, V = X'' f{x^, . . ., A;''-!)

which carries a part of dG into the plane a : x^ = 0. The antisymmetric part of the constant matrix UQ^ m a y as well be omitted since (5.5.2) f v,o:a^^u,pdx
G

= f v^ocd'^'^u^pdx
G

if

V^HIQ{G)

as is easily seen b y approximations, using Lemma 4.5.7Having mapped a neighborhood of XQ onto B2R or G2R, we then alter the coefficients b y defining ^E^W = ^f + (p(B~^x ^^'^'^^ h%[x)=^(p'h-, I) [a-^{x) af(x)], c%=(pc-, dn=(pd,

where 9?^ C(Ri), q){s) = \ for s < 5/4, (p{s) = 0 for s > 7/4 and cp is non-increasing. If ^ is a solution of (5.1.1) with support on BR or GR U (TR, it is also a solution of (5.5.4) fv,4al^u,^
G

+ b%u + e"") + v(c%u,y, + dRU + f)dx = 0,

V^HIQ(G), G = B2R or G2R.

We then write (5.5.5) u = UR + HR, UR = Q2R[(aR ^0) Vu + bnu + e]

P2R[CRVu + dRU +f]

where, in the case52i?, Q2R andP2i2 are the respective quasipotentials, as defined in 3.7, and potentials, except t h a t in the case 1; = 2, we define pR(f) as the ordinary potential o f / o v e r BR minus its average value. The reason for this definition of PR in the case r = 2 is t h a t if V = PR{f), then, in addition to the results in Theorem 3-7.1, we also have

\\Vr,<C{v,q)RHf\\o
in all cases. This follows from the theorems of 3.6. In all cases, we see that Q2R-^0 as :V - > 00. This is true of P2R also if r > 2; b u t if r = 2, P2R might become logarithmically infinite. So if u has support in BR, it follows t h a t HR = 0 if 1 > 2, or HR = const, if ^ r == 2. I n t h a t case
(5.5-6) UR TRUR=^VR= Q2R {e + bR HR) P2R {dR HR + / ) BR UR] P2R (CR ' V UR + dR UR) . TR UR = Q2R [{aR ~ ao) '\7UR+

It is then shown t h a t if R is small enough, \TR\ < 1/2 in HI(B2R), where u and UR are known to be, and also in a space H\{B2R) for some
^ > 2 or a space C\{B2R). We then conclude t h a t u e HI{B2R) or C]^{B2R)

151 as the case may be. A similar program is carried out for hemispheres G2R. The general bounded domain G is handled b y a partition of unity. Definition 5.5.1. In the space CIQ{B2R), we define |||^||||J gij = hAe. B2R)_+ R-qe\\\lj,, J!|^|||o^ being the norm in C^ {B2R). In the space C]^[B2R), we define In the space H\{B2R), we define the norm such that

We define the space L2>f^(B2R) to consist of Ql\f^Lp{B2R) J \f{x) \P dx <L2> r^'-'P+'P^ < LP (2 Rf-v+vi^
B{xo,r)r\B2R

for every B{xo, r) and we define ||/||p,^,2i? =" ^^^- -^ Then, from SOBOLEV'S lemma (Theorem 3.5.3) and Theorems 37.1 3.7.3, we obtain the following results: Theorem 5.5.1. QR is a hounded operator from Lq{BR) to H\[BR) for each q'> \, and from C^^^Q{ER) to C^^BR) with hounds depending on q and /Lt (0 <C ju < \), respectively, hut not on R. PR is a bounded operator from
LP(BR) to HI(BR) and to HI{BR) with hound independent of R if \ <

p<^v and q = p vl{v p); PR is also a bounded operator from LP^BR), 1 ^p <.v, or LP(BR) with p = r/(1 /^) into C^BR), with hound independent of R. In fact if u QR{e) and C^LQIBR) then U^HI(BA) for any A and \\Vu\\ls^ ^ C ||^||J^; if e^ C%{BR), then u^ CI(BA) with hf,{S7u, BA) < C hf,{e, BR) for every A] if v = PR(f) and f^ Lp (BR), \<P<v, then V^HKBA) for any A and || Vt; ||,%^ < C ||/||o^ if q =pvl{v p)\ iff^Lp,f^(BR), \ <p <v, or to Lp withp = r/(1 ju), then v^ C'^{Rv) with hfi(^v, Rv) < C ||/||, the norm being in the proper space. Finally, if v = 2, and f satisfies the condition (5.57) / \f{x) \dx^L
-B(a!o.r)nBB

rn,

all

B{XQ, r), pix > 0,

then Pnif) C HI (BA) for any A and ^ PR(f) ^ L2,II(BA) for any A and \\PR{f) M,2P. < CLR^i'i, I V P ( / ) ||..., < CL. Definition 5.5.2. We say that the coefficients a, h, c, and d satisfy the i^^-conditions on a domain/"if and only if the a^^ are continuous on r and the b"-, c', and d are measurable there and (i) b^ and c$ Lv(r) and d^Lvl2(r) (ii)
rf]B(xo,r)

if vl(v \) < q <v and v > 2 /^i > 0,

/ ( | & | * ' + | c | * ' + \d\''^^)dx<L''r''n,

(m)b^^Lq(r),

for every B(xo, r), if ^ = r > 2, c^^Lr(r), and d^ Lp(r), p = v ql(v + q), ii q > v.

We say t h a t the coefficients satisfy the C^-conditions on F if and only if

152

Differentiability of weak solutions

the a""^ and b^^ Cl{r) (5.5.8)

-B(a'o.r)nr

and c^ and d^"^ Li{r)

and satisfy xo^Rv, r > 0.

J (\c{x)\ + \d(x)\)dx^Lr'-^+^,

It is desirable to prove the following lemmas: Lemma 5.5.1. (a) Suppose F is a hounded strongly Lipschitz domain and (p^ Li{r) and satisfies (5-5.7) 'i^ith BR replaced by F. Then, for each > 0, there is a C [s, jbti, v, F) such that (5.5.9) / I ^ W I i ^ W I ^ ^ < CR'L u^HliF), (b) Suppose u^ HI{B2R) (5.5.10) \u[x)\ with q > r. Then x^B^R. '\\ u \\l r^i-%

\r\=y,R^

<^C[v,q)R^-->^'\\u\\,

Proof, (a) On account of the extension theorem (Theorem 3.4.3)> we may assume that u^ HIQ(F). Then, from Theorem 3.7-2, we conclude that / I (p{x)\\u{x)\dx<F-^fdxf\ (p(x)\'\i x\^-''\Vu(i)\di, B{xo.r)r\r B(rco,r)nr r (5.5.11) Applying the Holder inequality to the double integral on the right in (5.5.11), we see that our integral / on the left satisfies

/<r-i4^-1)/*' 7l/^ ii =

B(xQ,r)rr) r

Jdxf\^-x\-^+^\ip{x)\d^

(5.5.12) l2 = fdxfli -X l-^C-i) I (p{x) I I V ^ ( l ) j " ^ ! . B(xo,r)nr r From Lemma 3-4.3, it follows that (5.5.13) Ii^a-^F,{\F\lyy)^'''Lrf'i =
00

G-'^FrR''Lrf'i.

In order to estimate I2, we first note that / 11 ^ [-^("-D \q)(x)\dx<^ B{xo,r)nr (5.5.14) where (5.5.15) (p{s;
0

f s-^('-i) (p'{s; ^)ds 0

==a(v l ) ! ^ - ! - ^ ^ - ! ) ^ ( 5 ; ^)ds i)=l\(p{^)\dx^l(Ls^i, Bi$.s)nB{xo,r)nr [^ [Lr^i, 0< s< r s ^r.

From (5.5.14) and (5-5.15), we find that (5.5-16) l2<Zi{v, /Lii, o)Lr^i-^(''-^)('\\u\\lY, 0 < a <fiil{v - \). The result (5-5-9) follows by taking s = a{v ^)lv < //i/r. From this, it follows that (5-5-9) holds for values of s such that juijv < < //i. (b) follows from Theorem 3.5.I with p = q ^ v and m = \ and from the definition of the '|| u IJ^^gR-

5-5. Lp and SCHAUDER estimates

153

Theorem 5.5.2. (a) Suppose the coefficients a, b, c, and d satisfy the H\-conditions on BA where q'>vl{v ~ 1) and the equality holds only if V = 2. Then for each R <, Aj2, the operator TR is hounded on H\[B2R} and has abound of the form e{R) where e{R) ->0 as R -^0\ e{R) depends on the moduli of continuity of the a'^^ and on v, m, and M and on sup J ( | 6 | ^ + |c|' + l^l^'S) dx L and fii in case (ii), and sup J {\b\Q + |c|^ + \d\P) dx
B{xo,R)nBA

in case (i) (of def. 5.5-2),

in case (iii).

(b) / / the coefficients satisfy the C^-conditions on BA, then, for each R <, Ajl, TR is bounded on C\[B2R) with a bound of the form CR^ where C depends on v, m, M, L, /bt, and the C^-norms of the coefficients a and b. Proof. We set eR = {aR ao) '
VUR

-\- bR

UR,

fR =

CR- \/UR

+ dR

UR,

rj(R) = ma.x\aR{x) ao] . In the H^ case we see, using Lemma 5-5-'I (with g){x) = l&i?!" + \CR\^ + + IdR]""^) and Theorem 3-5.5, that (i) \\eRt^[v{R) + C\\bR\\^]^\\uR\\l4fR\\l<Cq UR t% < C(LR^i) '|| UR ||J, (ii) II en ||S < [rj{R) + C{LR^^i)^'^] '\\ UR ||J, (iii) 1 CR ||o < [rj (R) + C\\bR ||o Ri-^'a-i '\\ UR \\l, 1 \\fRt^[\\cRr. + C\\dRtRi-^'ay\\uR\\l in the respective cases; here we set ^ = r ql(v + q) and omitted the subscript 2i? on all the norms. If )^ = 2 in (ii), the e inequality still holds and fR satisfies (5.5-7) with L replaced by CL'\\UR\\1{\ + LR^i). In the C}, case, we note first that I bR{x2) - bR{xi) I < I (PR{X2) (PR{XI) I I bR{x2) I + (PR{XI) | bR{x2) - bR{xi) | (5.5.17) <{CBo'R-f^ + Bi) \x2-x1\f', X^B^R,
BA).

X^^X^^B^R, BQ = m2ix\b{x)\,
\x\^A

(pR{x)^(p{R-^\x\),

\bR{x)\^BQ, Bi - h^{b,

Using this idea again, we obtain \aR{x2) aR[xi)\ < 2Ai\x2 / \fR{x) \dx<LR^{\+R)''\\\u
Bixo,r)riB2R

^ l | ^ Ai = h^{a, BA), IIIJ.2E ^^-1+^

The theorem follows from Theorem 5.5.1-

154

Differentiability of weak solutions

We can now prove the following interior boundedness theorem: Theorem 5.5.3. (a) Suppose that (i) the coefficients satisfy the H\ and H\, conditions on the domain G, where <.q <. q', (ii) suppose e^Lq, (D) and/^ Lp, {D) for each D G G G, wherep' =v q'l(v + q') > '^ and (iii) suppose u^ Hl{D) and satisfies (5-2.2) (with ^ = 0) on such D. Then u ^ Hi, [D) for such D. (b) Suppose that (i) the coefficients satisfy the H\ and CJ conditions on G, q~>vj(y 1) and 0 < /^ < 1; (ii) suppose e^ C^^{D) and / ^ Li^(Z)) for each DCC.G, and (iii) suppose u^H\{D) and satisfies (5-2.2) (A = 0) on such D. Then u^ Cl(D) for each such D. Proof, (a) Each point XQ in G is the center of an ellipse which can be mapped onto B^R by an affine m a p as described above, where R is so
small t h a t the bound TR is < 1/2 on H\[B2R) and on H\,{B2R) where

we assume first that q' <v ql(v q), q <v. Let C$ Cl{G) and have support on the image of BR and let U be the transform of C ^- Then, b y replacing z; b y C ^ in (5.2.2), we see that U has support on BR and satisfies the transformed equations (5-2.2) (A 0) with e a n d / r e p l a c e d respectively b y J^l = f e- - al^C,^u, (5-5.18) Since U has compact support, we see that
U=UR + HR, UR -TRUR=VR= Q2R (ER + bR HR) -

FR = Cf-c%^,ccu

+ C,4^%^u,^ J^blu

+ e-).

P2R {FR + dR HR) (HR = COUSt.). Since q' <,v ql(v q), it follows that p' <,q so t h a t E^ Lq,(B2R) and F^ Lp, (B2R) SO that U^ HI,(B2R). Since it is clear that the coefficients satisfy the ^^-conditions if they satisfy the //"J,-conditions and q" <iq' or if they satisfy the CJ conditions we m a y use the result to prove the result for any ^'. If Z) C C (^, we m a y find a partition of unity on D each member of which has small support as above. In the CJ-case, we first show that u ^ HI, (D) for some q' > r, for each Z) C C G. Then we see that E^ Cl(D) so that U^ CI(B2R). Thus, as above, u^ Cj^(D) for any D CC.G. We need the following lemma for later reference: Lemma 5.5.2. Suppose the coefficients satisfy the H\-conditions on BAThen there is an RQ, Q <C RQ <. Ajl, such that if u has support on BR and satisfies (5 -1 -1) there with e^Lq andf^ Lp where v>2 and p = qvl(q + v) > 1, then 'MU<Cx{le\ln + \\f\lE), 0<R<Ro. If V = 2 = q, so p = i, the same result holds with \\f\\l^B replaced by CLRf^i if f satisfies (5-5.7). If the coefficients satisfy the Cji-conditions

5.5- Lp and SCHAUDER estimates

155

on BA, there is an Ri, 0 < i ? i < ^ / 2 , such that if u^ ClQ(Bji) and satisfies (5.1.'I) there with e^ C^(Bji) and f^Li^(^[Bii), then
WuR

^^'-<C^{'\\e\\U + \\ni,,.n).
specified in

The constants RQ, RI, Ci, and C2 depend on the quantities Theorem 5.5.2. Proof. Since u has support in BR, we have seen t h a t U = UR+ HR
SO WR UR TR UR = VR + WR, VR =

Q2R [c] P^R (/),

=- Q2R {bR HR) P2R (dR HR)

where HR = 0 if i' > 2 and is a constant if > == 2 and TR has bound < 1/2 if i? < i?o or Ri. li V = 2, we see, since \/ u = V UR and u has
compact support, that '\\U\\IR < C'WURWIH SO '\\HR\\IJ^ < C'\\UR\\IR.

Setting p = qvl(q -\- v) we find as in the proof of Theorem 5.5.2 that if r/(r 1) < q <iv, then II bR Hn tn < C 'II UR || j] hn fk, \\ du HR \l^ < C 'f UR ||I\\ dR ||o,,.^.

Since || &i? IJjj and || ^i? ||J/2,E ^ - 0 as R->0, t h e result follows in this case. The other cases are proved similarly. Definition 5.5.1'. We define t h e norms and spaces of functions on G2R as they were defined on B2R b u t with B2R replaced by G2R and the ftmctions u (in C1(G2R) or H\[G2R)) being required to vanish along a2R. The functions e'^ are supposed t o vanish along ^ 2 i ? b u t not necessarily along a2R. We define P2i2(/) as the restriction to G2R of the former P 2 R ( / ) where / i s the extension to B2R b y ''negative reflection"

We define Q2R [e) = U hy the formula (5.5.20) U[x) =^ -fKo,4^

^2R

- I) eHi)di + 2S

fKoA^

- ^ye^[^)d^

""^^ G2R

where we assume t h a t e has been extended b y ''positive reflection" (5.5.21) f l[x\ < ) = ] elx"", x') ^ , if x'' > 0, -

KQ is the elementary function for Laplace's equation. Using t h e theorems above, we obtain the following result: Theorem 5.5.1'. Theorems 5.5.1 and 5-5.2. and Lemma 5.5.2 hold with BR replaced by GR; the H\- and C\-conditions on the coefficients being defined as above.

156

Differentiability of weak solutions

Proof. It remains only to prove t h a t the second term in (5.5-20) C1{^A) for any A. Let this be denoted by V. For %" > 0, we have

= 2j,

f^2Ko,4x

- i)'[e-{S)

-e-(x)]di

since e vanishes along ^ ^ and the integral of V^i^o,a overJf?" converges absolutely to 0. From this we conclude t h a t I V2F{:^)| < C{v,iLt) K{e, G2R) M ' * - ! from which the result follows, using Theorem 2.6.6. By the method of proof of Theorem 5.5-3. using the result of Theorem 5.5.2 for G2R, we can prove the following theorem: Theorem 5.5.4'. Suppose that G ^ C^, the coefficients satisfy the H\and H\,-condition on a domain F Z) G, where vj{v 1) ^q <iq' and p' = vq'l(v + q') ( > 1 ) , suppose e^Lq,(G) and f^Lp,{G), and suppose u^H\{G) and is a solution of (5-1.1) on G. Then u^H\,[G). If, also, G^ CJ^, the coefficients satisfy the Cj^-conditions on F, e^ C^(G), and feLi,^(G),thenu^Cl{G). We now prove the following a priori bound: Theorem 5.5.5'. (a) Suppose that G ^ C^ and that the coefficients satisfy the H\-conditions on F Z) G with q ^vl(v 1), the equality holding only if V = 2. Suppose also that e^Lq{G) and that f^ Lp{G), where p == qvl(v + q) {>\) and where f satisfies (5-5-7) if p = 1- Suppose that u satisfies (5-1-1) on G. Then

(5.5.22)

l k K . < c ( | | . | | + i / K + ||||?)

where C depends only on v, m, M, q, G, the moduli of continuity of a, and hounds for the coefficients) if v =^ 2 =^ q, so p =^ \, the term \\f\\^ must be replaced by L. (b) If G^C^^, the coefficients satisfy the Cj^-conditions on Fz>G, e^ C^{G) and f^Li^[x{G), then (5.5.23) l|||||i.e<C(i^I||.e + l/i?,. + ||M|!S). Proof. We prove (a) with ^ > 1, the proof of (b) and the last case of (a) are similar. Each point X{i of G is in a neighborhood ^ C a neighborhood It which can be mapped as described at the beginning of the section onto B2ROT G2RUa2R so that 9^ is mapped onto BR or GRU(TR, where R is so small t h a t Lemma 5.5.2 holds, and we may choose a partition of unity {Cs} of the proper class, eachf^. having support in some one neighborhood 9^. Then, if Us and Cs are the transforms of Cs ^ and fs, respectively (on BR or GR), we see t h a t Us satisfies (5-1-1) with e a n d / r e p l a c e d by Eg and Fs as given in (5.5.18) with the obvious changes of notation.

5-6. The equation a'S7^u-\-b-Vi^-\~ou~?.u=f

157

Now, suppose there is no such constant. Then there are sequences {^n}, {en], and {fn} such t h a t
(5.5.24) II Un Wio = 1, II ^^ \\U + Wfn WIG + II Un ||S.c.-> 0 ,

and Un7 ti in H\{G). Since Un-^u in L^(G) and hence in Ll{G) it follows that u = 0, Thus we conclude that Ens ->0 in Lq[G), [G = B2R or G2R) Fins-^0 in Lp{G) and i^2ws-7 0 in Lq{G), F2ns = Cs,oca'^jf Un,p, Fins = FnsF2ns' Thus V^ ^2/2 ^ i w s - > 0 in Lp{G) and S/'^P^RF^US -V 0 in Lq(G), so that P2R{Fns) - ^ 0 in Hl(G). But then Uns->0 in H\{G) for each s (Lemma 5-5.2) so that ^^ - > 0 in H\{G). But this contradicts (5.5.24). Remarks. From Theorems 5.5.4 and Theorem 5.5.5 and the proof of Theorem 5-2.4 we conclude the following: Suppose G, the coefficients, and e andf satisfy the H\-conditions and either the conditions in Theorem 5-5-5 (^) or (b) and suppose u^ Hl(G) and is a solution 0/(5.2.2) withX = Ao, where Ao is defined in Theorem 5.2.1. Then u^ Hl{G) if q > 2 or C]^{G), respectively, and the hounds (5.5-22) or (5-5-23) hold without the term \U^^Q on the right. For if not, sequences [un], {%}> and [fn] would exist satisfying (5.5.24) without the term || ^ ||i,6? on the right. But, from the proof of Theorem 5-2.4, it follows t h a t w^ ~>0 in H\Q{G) and so in L\ (G) and the remainder of the proof proceeds as before. 5.6. The equation a* V^ u -}- 6 Vw -{- cu Xu =f

Using the general theory of 5-2 together with Theorem 5-2.8, we easily deduce the following preliminary existence theorem. Theorem 5.6.1. Suppose G is of class C\, a ^ C^ on G, b is bounded on G, c^ Lv{G) and satisfies (5.2.11) there, and a satisfies (5-1-2) there. Then, if X is not in a set b without {finite) limit points, there is a unique solution of (5.6.1) L u X u ^ {a ' \7'^ u -\- b ' \/ u -\- c u) X u =^ f which ^ Hl(G) n H\Q[G) for each f^ L2{G). If X^h, there is a non-empty hut finite-dimensional manifold of solutions in HI (G) 0 HIQ{G) with / = 0. Proof. Let us define C(u, v) as in 5.2, and define B{u, v) = J{v,cca'^u,p
G

+ v[(a^^ b'^)u,cc
G

cu]}dx,

L{v) = ffv dx.

Then, from Theorem 5.2.8 it follows t h a t (a) if X is not in a set b, there is a unique solution ^ of (5-2.2) in H\Q[G), and (b) t h a t solution (if it exists) ^ Hl{G) and satisfies the differential equation (5.2.12) (a.e.). But (5-2.12) reduces to (5-6.1) in our case. In the theorem above, we were forced to assume that a^ G\(G) in order to be able to use the HILBERT space proof given above and in 5.2.

158

Differentiability of weak solutions

It has long been known in the case v = 2 (see the article by L. LICHTENSTEIN in the Enzyklopadie der mathematischen Wissenschaften [3]) and was proved in 1934 by J. SCHAUDER for any v t h a t if the a, h, c, a n d / were merely HOLDER continuous, then the solutions ^ (^%,(G) and to C^ (G) if ^ = 0 on ^ G and G is of class C^. We shall first obtain an existence theorem for the case where G is of class CJ, a is merely continuous and h and c satisfy the conditions above. We shall then obtain SCHAUDER'S results. There are some a priori bounds which have been obtained by CoRDES ([1], [2], [3]) for the case that the a"^^ are merely bounded and measurable with & = c = 0, Although these are important for the study of quasi-linear equations, the results are not completely general and require certain special techniques for their proofs. Accordingly we shall not consider these results. Our method of proof is to reduce the problem to one on spheres or hemispheres BR^ or GR^ where a"^^ = Zl"'^ as in the preceding section; here we may take a^^ =^ a^^. As in t h a t section, we shall consider first functions with support in BR or GR. Without altering the fact t h a t u satisfies (5-6.1) on BR (or GR), we may alter the coefficients as in (5.53)- We then write, assuming t h a t A = 0 in (5-6.1), (5-6.2) u == UR + HR, UR = PR [f (UR ao) ' V^ U bR ' \/U CRU] where in the BR case, PR is the potential as defined before, and in the GR case, PR{f) denotes the potential of/* w h e r e / * is the extension of / t o the whole of BR by negative reflection: (5.6.3) f*{x', K) = - / ( - ^ ^ < ) , ^' < 0; H^Q(BR) HR{X)=\ and spt. u C BR, ^ / ^ / ^ ^ (const, if 1 = 2. ^ / * may not be continuous across x^ = O.liu^ then we easily see that (5-6.4) u{x)==fKo{x~^)Au(^)d^ B^ n HIQ{GR)
BR

so

If U^HI[GR) (5.6.5)

and has spt. in GR U CR, then this time - f ) / * ( f ) di, f{x) =Au{x),

u{x) - UR{X) =fKo(x

a n d / * is defined by (5.6.3); this follows from our definition since the average of ^i? = 0 even in the two dimensional case. Thus if r > 2
(5.6.6)
TR UR

UR TRUR

= VR = PR (/),

= PR [(aR ao) ' V^ UR + bR ' \7 UR + CR UR].

We shall assume r > 2; we have seen how to take care of HR. Definition 5.6.1. On the space H\Q{B2R) or HI{G2R) we define

1kEo.H=J^-^*llv^-^-^fc;

5.6. The equation a'V^u-\-b'Vu-\-cu~?,u=f

159

on the space C^{B2R)

or C^{G2R),

we define

A i j = ^2i? or G2R, III II being the norm in C^ {r.2R). I Definition 5.6.2. We say t h a t the coefficients a"-^, &", and c satisfy the h-p-conditions on a region F (h an integer ^0) ^ a is continuous and satisfies (5.1.2), b and c are measurable there, and (i) if A = 0, and p < v/2, then b^ U{r) and c^Lp,2{r) and (5.6.7) / ( l ^ | ' + \c\''l^)dx<KRf'i,
B{x,E)nr

jLCi > 0;

(ii) if A-= 0, and vl2<p<v, then b^UiF) and c^Lp{r) and 16|^ and I c|^ satisfy a condition like (5.6.7); (iii) if /^ > 0 and ^ is chosen so that (k - \) p <,v <Ckp, then V^' a, V^*-i S, V^-^c^L(yfj) for y = 1, . . ., y 1 with ^ / (I V^' < I + I V^'-i & I -1- I V^*-2 c |)W:/ dx<K ^
B{xo,R)nr

R^i

where meaningless terms are omitted (V~^ b, etc.); if y > k, then V^* ci, V^'-i b, and V^'-^ c^Lp with y"(|V^*a| + |V^'-i&| + IV^'-^cD^dx^RRf^i, j = k, ..., h + 2, Bixo,R)r\r where V'^^^ a, V^+2 ^, V^^^ b are to be omitted. The coefficients satisfy the /^-/^-conditions onF ^ they also ^ C^ (F). Definition 5.6.3. We let E denote the set of bounds and moduli of continuity of the coefficients and their relevant derivatives; it will usually be clear from the context what quantities are included. Lemma 5.6.1. (a) Suppose G = BRQ or GRQ, 1 < _> < ^ < CXD^ h'> 0, / the coefficients satisfy the 0-p- and the h-q-conditions on G, and a'^{0) = d"-^. Then there exists an Ri > 0, depending only on v, h, p, q, and E and a constant Ci, depending only on Vyh,p, and q such that ifu^H^^ {BR)
then U^H%^[BR) and if U^HI{GR) H HU[GR), then U^H^+^GR)

nHl^{GR),and '\\u\\i^^C^'\\Lu\\n, if 0<ie<i?i and spt u C BR or GR U OR, respectively. (b) / / , also, the coefficients satisfy the h-jbt-conditions on G, there exists a constant C2, depending only on v, h, p, and ju, and an R2'> 0, depending only on v, h, p, ^ , and E such that if u^ H^Q(BR), then u^ C^'^^(BR) and
ifu^HKGR) n HI^{GR), then u^ CI-'^GR), and

lll^lll^^^ < Cs'llJL^III^^, and spt u C BR or GR\J GR, respectively.

if

0<R<.R2

160

Differentiability of weak solutions

Proof. We first show t h a t PR is a bounded transformation from

HU{B2R) to HI^^B^R) and from CI,{B2R) to CI^^{B^R). If/ C^B^R)

and u = PR{f), then (5.6.8) V^" u{x) = I Ko{x - f) \/^m d^, 0 < ; < A.

The result then follows from Theorem 3-7.1 for the H\'^^ case and from Theorem 2.6.7 for the Holder norm case. I f / ^ C ^ ( G 2 i ? ) and s p t / C G2UcT2i?, and if u = PR(f), then u is given by (5-6.5) so that 'V^' u{x) = I Ko(x -^) ' V V * ( f ) ^l> 0 < y < >^,

where 'V denotes the gradient with respect to the {x^,.. ., x^-^) variables. This m a y be written in the form 'Vi u=Uj~ (5.6.9) Fy, Uj(x) = I Ko{x - I) (pj{S)dl Vj{x) = 2 f Ko(x ^)<pj{^)d^, (pj = ' V V ^

a-.

where / + is the extension of / by positive reflection: (5.6.10) /+(*', < _ i ) = / ( - * ' , x',_-,),
PI CI{E^R) with

x'KO.

For e a c h ; , cpj^ LJ,{B2R)

(5.6.11)

I <f>i\U < (2i?)*-^ 11/11*^, 'III mtn <

(2R)n-i'\\\f\\l^.

The desired inequalities for Uj follow as before (with h = 0). The H^ results for Vj follow from Theorem 3-7-1. For x"^ > 0, we have V'V2F;(:v) == 2 / V'V^K^{x ^ |)^,-(f)^f

= 2 / V ' V 2 i C o ( ^ - ^) V<Pm - 9^iW]^f

22

since the integral of V ' V^ i^o (^ I) over R~ converges absolutely to zero. Thus, using the fact t h a t Vj is harmonic, we find that I V3 V^[x) I < C A^((^y, B2R) (^^)'^-i. It follows from Theorem 2.6.6 that Vj^ CI{G2R) with the desired bound. It follows easily that PR is bounded in the G2R cases also. It is now easy to see, using (5-5-3) and (5-6.6), that 1 TR^.i^ < 8i(R;v, h, q, E), '\\TRI%\^ < e2{R',v, h, fji, E) \imej{R\ . . .) = 0, y = 1, 2.

Thus, if i^ < J^i, the norm of r^^ in HliP^R) and in H^+^(r2R) is < 1/2 (A/2 = B2ROV G2R), so there is a unique solution 7^^? of (5.6.6) in both of

5.6. The equation a''V^u-\-b'Vu-{-cu}.u=^f

161

these spaces. Since the second space C the first, these solutions must coincide. Since u = UR, the result (a) follows. The proof of (b) is similar. We can now prove the following a priori estimate: Theorem 5.6.2. Suppose that G is of class Cf+^ and that the coefficients satisfy the h-p-conditions on G for some p ^ \. Then ^ a constant Ci, depending only onv, h, p, E, and G such that (5.6.12) \\u\ll^<C^{\Lulla + \\ul\a), u^Hl*^{G) r\H\,{G). If G is of class C^"^^ and the coefficients satisfy the h-ju-conditions on G, then there is a constant C2, depending only on v, ^ , p, E, and G, such that (5.6.13) MU'<G4\\\Lu\\\lo) + \\u\\U, Cr'(G)nffio(G)Proof. We shall prove the second statement; the proof of the first is similar. Each point P of G is in a neighborhood on G which can be mapped as in 5.5, by a mapping of class C^'^^, onto either BR or GR in such a way t h a t the coefficients of the transformed operator, still called L, satisfy the /^-/^-conditions with a'^{0) = 6"-^, and R is small enough for the conclusions of Lemma 5.6.1(b) t o hold. Let {C^}, s 1, . . ., 5 , be a partition of unity in which each Cs has support in some such neighborhood. We let Us be the transform of C^ u and Ls the transformed operator. Now, suppose there were no constant C2 satisfying (5.6.13). Then, there would be a sequence [un]^ C^^^^[G) O H\Q{G) such that (5-6.14) \lu4\l%^=\, ||JL|||^^^0, Wunlla^O and we m a y assume t h a t Un, SI Un, and V^^^^ converge uniformly to w, V ^ , and V^w, respectively, where u must be zero on account of (5.6.14). Let Uns be the transform of f Un- Then Uns satisfies
(5.6.15) U Uns = Cf Ls < + Ms < ,

where u^ is the transform of Un and Ms involves only u^ and its first derivatives. From (5.6.14) and (5.6.15), it follows easily that '||| i^f LgU^ + + MsU^\\\^R->0, But then 'If ^ w s | | m ^ - > 0 for each s. B u t then |||^w|||^G^->0 contradicting (5.6.I4). We now prove the following differentiability theorem: Theorem 5.6.3. (a) Suppose that G is of class CJ"^^ and that the coefficients satisfy theO - p and h q conditions on G, where A > 0, 1 < ^ < $', and either A > 0, q^-py or both. Then if u^ H^[G) r\ H^Q(G) and Lu^H^G), it follows that u^H\-^^{G). If we know only that u^Hl(D) for each D G G G, it follows that u^ H^'^^(D) for any such D. (b) If G is of class C^'^^, the coefficients satisfy the 0 p and h /n conditions onG,u^ Hl(G) 0 HIQ(G), andLu^ C^i^), then u^ C^^^(G). / / we know only that u^H^(D) for each D G G G, we conclude that u^C^'^^{D) for such D. If G, the coefficients, and Lu^ C(G) then u^C-(G) if u^Hl(G)nHl,(G); if u^Hl(D) for D G G G, u^C^{G(o)).
Morrey, Multiple Integrals 11

162

Differentiability of weak solutions

Proof. The C^ results in (b) follow from the CJ"^^ results. We prove (a) first. We first assume q = p, h = 1. We may assume the results and notations of the first paragraph of the proof of Theorem 5-6.2 where, however, we assume that the neighborhoods are small enough for the conclusions of Lemma 5-6.1 (a) to hold for h = 0 and for h = \. Then Us satisfies (5-6.16) Ls Us = Cf Lsu* + Ms u* where Ms is of lower order as in (5-6.15). Thus LgUs^ H\{rR) so that each Us^ H^^[FR) so that u^H\[G). A repetition of the argument proves the result for any h and the same p. The argument for raising the exponent from p to p' = v pl(v ~ p) (the Sobolev exponent) if ^ < r is similar; since Ms i^* is of lower order, Ms u*^H%(rR) if u^ H^+^G). Up^v, then Ms u*^H^(rR) for any q if u*^ H^^^{rR). A finite number of repetitions raises the exponent from p to any desired q. In part (b), we first raise the exponent to some q^v, at which time we can conclude that Ms u*^C^(FR) iiu^ H^'^^(G). The following special theorem and its method of proof are essentially
due to NiRENBERG (see AGMON, DOUGLIS, and NIRENBERG [1] p. 693 and

also .6.5 for a more general theorem). Theorem 5.6.4. Suppose that G ^ C\ and the coefficients satisfy the 0 2-conditions G. Then there exist numbers Ao > 0 and C, which depend only on v, G, and E such that \\u\\lc^<C\\Lu-^u\\la, Areal, A > A o , u^Hl{G) 0 HIQ(G), Proof. For any given >^i > 0, each point XQ is in a neighborhood or boundary neighborhood in which \a{x) a{xo)\ <Crji. We choose a sequence Cs, s \, . . ., S, such that each Cs^ ^ i ( ^ ) with spt Cs in some one such neighborhood and f f _|_ . . . _|. ^2 ^ ^ ^ ^ jg^ Ug = Cs^ and note t h a t ^s L u = L Us -\- Ms u where each Ms is an operator of the first order. Then ~{Lu, u)l= -Zi^sLu,
s=l

Csu)l=

-Z{Lus,
s

% ) ? + {M'u,

u)l

(5-6.17) where M' is of the first order. But [Lus, Us)l = Ja'^^Us,ccpUsdx
G

j{a"-^ a^^)us,x^Usdx
G

(5-6.18)

f(b'Us,cc + cus)usdx,
G

af^ = a'^^{xos)

where XQS is in the small support of f. Since, for each s, jal^Us,a^Usdx = Ja'^^Us,ocUs,^dx > 0,

5.6. The equation a-^^ u -\- b ' Vw -\- c it 2.u = f

163

we see, by summing (5.6.18) with respect to s and using the facts that Us Cs u. etc., that
G

I- ^

where M is an operator of the first order. Thus -{Lu, u)\ > - ^ l i l V2 w|| |||| - Ci(^i) (II V|l + i|M||) |||| S : - 2ni II V 2 II 11II - C ( ^ i ) II u f . But, now for real A, \Lu-XuY^ \L uf-^ X^uf - 2K(Lu,u) > I L i. 1 + A2II u IP - 2rii(\\ V2 ^ IP + A2II ^ IP) _ I 2 | ^ - 2A C(rji) \\u\\^ > IJL ^IP (1 - 2rji C2) + + [A2(l - 2rji) - 21 C(f]i) - 2rji C2] || u |P (II <p II = II ?> Ill).

where C2 is related to the constant of Theorem 5-6.2. If we first choose fji so small that (1 2 i^i) > 1/2 and (1 2 ^1 C2) > 1/2, we may then choose Ao so large that X^l2 2X C (rji) 2 7 1 C2 > 1/2 for A > ^o- Then ^ ||i:_^||2>l[||L||2+||^l|2] from which the theorem follows. Theorem 5.6.5. The conclusions of Theorem 5.6.1 hold under the hypotheses of Theorem 5.6.4. In fact no real A > Ao is an eigenvalue. IfG is of class C| and the coefficients ^ C^{G) with c(x) < 0, we may take Ao == 0. Proof. First, suppose Ai > XQ. We may approximate to L by operators Ln whose coefficients satisfy our conditions uniformly with a^^ converging uniformly to a^"^ on C, each a^^^ C\{G). Define Lin = Ln h I as an operator on L2{G) with domain all u^HKG) which vanish on ^G. If, for some n, Ai were an eigenvalue forl,^, then Lin would carry some non-zero element into 0 which would contradict Theorem 5.6.4. Thus Xi is not an eigenvalue for any n. Hence, ii f^L2{G), L^^{f) = Un is defined for each n and || Un 1|| < C||/||| for all n. Hence a subsequence, still called {un}, 7um H\{G) and 1 = 0 on ^G. Then Lin Un -^ Li u ^ in L2 (cf. the proof of Theorem 5-2.7), so that Li u = f. Now, the equation Lu~Xu=f is equivalent to the equation
LQU

+ {X Xo) u =f

(LQ

= L

XQI)

which is, in turn, equivalent to u {X Ao) LQ'^ U = L^^f. As an operator on H^0^20> ^0^^ i^ compact, since weak convergence in H\ implies strong convergence in L2. The first part of the theorem follows from the RiESZ theory of compact linear operators.
11*

164

Differentiability of weak solutions

To prove the last statement, suppose A > 0. If A were an eigenvalue, t h e n L u Xu = 0 for some u^ C^(G) which vanishes on ^G b u t is not ^ 0. B u t since the coefficient of u in L u Xu is c{x) A < 0, this would contradict the maximum principle (Theorem 2.8.1). 5.7. Analyticity of the solutions of analytic linear equations In this section, we present a simple proof of the analyticity of the solutions of (5.6.1) on the interior when the coefficients a n d / a r e analytic and show that solutions of such equations which vanish along an analytic part of the boundary can be continued analytically across the boundary. By using the mappings described in the proof of Theorem 5-6.2, which are analytic in these cases, the problem of local differentiability is reduced to that for solutions on spheres or hemispheres with a'^^ Zl*^^. The local differentiability theorems imply a global theorem in case the entire boundary d G is analytic. The technique of proof is that given in a recent paper by MORREY and NiRENBERG; since the solutions are known already to be of class C^, it is sufficient to obtain bounds for the derivatives. In view of Lemma 5.7.2 below, it is sufficient t o obtain bounds for their L2-norms. Lemma 5.7.1. Suppose u ^ Hl(Bji), u ^ H\{Br) for r <i R, and (5.7.1) Au=^f on BR. + ^-4^2]^^! J Then there is a constant Ci depending only on v, such that (5.7.2) f\\7^u\^dx<Cii ff^dx + f[d-^\Vu\^ ^' \^r+6 ^r+d 0<r <r + d <R, 0<d<r.

The lemma holds with BR, Br, and Br+d replaced by GR, Gr, and Gr+s, respectively, and the spaces HI [BR) and HI {Br) replaced, respectively, by Hf[GR) and Hf{GR) (u = 0 along OR). Proof. The second statement follows from the first b y extending u and / to BR b y negative reflection. To prove the first, define r] as usual with a replaced b y d, G b y Br.^d, and D b y Br, and let U =^ ri u. Then U has support in BR, U ^ H\[BR), and U satisfies AU =^ rif + 2S/ 7] ' "7u + AT] ' u = F. The result follows from Theorem }.7.1.
Lemma 5.7.2. Suppose U^C^{BR) [or C{GR[J GR)] and suppose

that there are constants M and K such that (5.7.3) (5.7.4) \\\7Vull^<M'P\KvR-v^ \S/Pu(x)\ p = 0,\,2,... {r==vl2). Then u is analytic on BR [or GR] and, in fact < Ci{m,v) 'Mpl(i.\K)PR-^-^,

5.7- Analyticity of the solutions of analytic linear equations

165

Proof. Applying the SOBOLEV lemma, Theorem 3.5.1, with m = \ -\-f [T] (r = vl2) and p = 2, to V^u, we obtain

|v*W| <ci?-4^<?^||v+^i,K + (-T)-i^?^!|v'+*.|!l|

(5.7.5) ^'^' ' where C is y~^^^ or 2^^^yy'^^^> depending on whether we are considering BR or GR. Inserting the bounds (5.7-3) into (5.7-5), we obtain

{^

rnl

^ - ^ (P + ; ) J +
\ ' /

+ (m- T ) - I - ^ ^ ^ , (p + m)n<:C(m,v)'M'
^
^ m \\^^ '

R-^-^K^+-^-pi-^^-^'
r

p\m\

from which (5-7-4) follows easily. It is convenient to introduce the following notations: Notations. F o r / a n d u ^ ^^(BR), we define MR,p{f) (5-7-6) NR,J,{U) = ipl)-^
-R/2<: r<iJ R/2<r<R

sup (R - r)2+J'||V^/||i.,,

P^O ^ >- 2

= [^!]-i sup (R~-r)^+P\\\7^+^u\\l if ^ > 0 ,

[p\]=pl

[:^!] = 1 if : ^ < 0 .

Remarks. From Lemma 5-7-2, it follows that the analyticity of u on BR will follow from an inequality of the form
(5-7-7) NR,P(U)^M'KP, P ^ - 2

for some constants M and K. Our method of proof will consist in demonstrating (5-7.7). Lemma 5.7.3. Suppose u and f^ ^^{BR) and it satisfies (5.7-1) there. We suppose also that MR,p{f) and NR^P{U) < oo for each p. Then there is a constant C^, depending only on v, siich that (5-7.8) NR, P {U) < C2 [MR, P(/) + NR, ^ _ I {U) + NR, P_2 {U)-\ , ^ > 0. Proof. We apply Lemma 5-7-1 (for BR) to each component of V ^ u, square, and add to obtain [^R, V M l ' < ^1 {P 0 - ' sup (7^ - r)4+2 V ^^''''^^

j\\/Pf\^dx +

+/((5-2|V2>+l^i2 + ^-4|V^^^|2)^%]

We now obtain the following results using the notations (5-7-10) / I V ^ / | 2 doc < (/)!)2 M\^[f) '(R~r^)-(4+2p)

| | V 3 ? + I ^ | 2 ^ A ; < { [ ( ^ - 1 ) ! ] } 2 A r | ^ ^ _ 3 ^ ( ^ ) . ( i e - r - ^ ) - ( 2 + 2 2^ ^ > 1

/ 1 V ^ ^ | 2 i;k; <{[(j;^ - 2)!]}2lS\^_^(u) - (R - r - d)-^P, ^ > 2 .

166

Differentiability of weak solutions

Setting (5.7.11)

d^{R-r)l{p^

1),

p > \ ,

in (5.7.9) and (5-7.10), we obtain + (1 -\-p-^)^n[p + \)^IPHP l ) 2 } A l . , - 2 ( ^ ) ] , ^ > 2,

with analogous results in the cases^ = 1 and j ^ = 0 (taking^ = {R r)l2 in the latter case). The result follows. Now, in order to prove the local differentiability, we write (5-6.1) (we absorb the term A u into c u) in the form (5.7.12) Au = F ^f (a ao) V^w b V ti cu.

In order to use Lemma 5.7.3. we shall obtain bounds for MR^P (F) in terms of MR^p[f) and the ^R^q{u). As an aid in this work we state the following lemma, the proof of which is left to the reader: Lemma 5.7.4. Suppose V and W are tensors in C'^(D). Then (5.7.13) \'7^{V-W){x)\<.J:{^^)\\Iv-QV{x)\'\\/aW{x)\, x^D.

Suggestion for the proof. By arranging the sets of common indices and the sets of remaining indices into single sequences, one may assume that
fc=i

i ^ \, . . ., n,

j = \, . . ., p.

Then, if oc denotes the multi-index (oci, . . ., ocv),

fc=l iS+y=a

The result follows without much difficulty if one notes that

Now, since a, h, and c are analytic, there are numbers ^ > 2, L, and RQ < 1 such that /. Q..^ | V ^ / ( ^ ) | , | V ^ a ( ^ ) | , \S7Vh[x)\, (5.8.14) \^ T A\ \ \yvc{x)\^Lp\Av, ^^i?o-

Theorem 5.7.1. Suppose a, h, c, and f are analytic on BR^ and satisfy (5.7.14) there, and suppose that U^H\[BR^ and is a solution of (5-6.1) there where A 0. Then there is an R^ <; RQ, which depends only on v, A, L, and RQ, such that u is analytic in BR for each R <, R7 Proof. Suppose 0 < i? < i^o- From the results of 5.5 and 5.6, it follows that u^C^ {BR) so that NR, p {u) is defined for every p. Obviously,

5.7- Analyticity of the solutions of analytic linear equations

167

MR, P (/) is also defined for every p and, in fact, ^R,vU)-(5.7.15) sup ( ^ ! ) - i ( 7 ? - ; ' ) 2 + p f ^L2(^!)2^2p^;^l|m^T<R \i J <.Zi[v)'LR^+'='{AR)v. + LIV^^+i^l + L | V ^ ^ | +

Using (5.7.12), (5.7.14), and Lemma 5-7.4, we obtain | V ^ ( i ^ - / ) | <.LAR'\\7^-^'^u\

3=1

+ ^(?)i^-4Q'g'!(|V^-^+2^| + i ^ ^ - ^ + i ^ j ^ |V2^-^|) (5.7.16) =LAR\S7v-^'^u\

2> + l

+L{\

+pA)\V^^'^u\

X[\ + [p+

\-q)A

+ {p+

\ --^)(^_-^)^2].|vi^+i-^|.

Using the definition of NR^P(U), (5-7.15), (5-7-16), the fact that A > 2, and the fact t h a t MR^ p acts hke a norm, we conclude t h a t MR,p(F)^MR,p(f)-]-LARNR,p(u) +
39+2

2L2;(AR)^NR^p^q{u).

(5.7.17)
Now if we use Lemma 5-7-3 with/replaced hyF and choose 0 < R < JR?, where i^7 is chosen so that C2LAR7< we conclude that NR,p(u)<2C2ZiLR^+^AR)^
(5.7.I8) + 2C2NR,P_I{U) +

1/2,
V+2

+ 4C2L2;{AR)^NR,p,g(u)
2C2NR,P-2{U).

We shall choose M and K so that (5.7.7) holds. Evidently, we can choose them so (5.7.7) holds ior p = 2 and p = \. Suppose (SJJ) holds u p to _/) 1, ^ > 0. Then, from (5.7.18), we conclude that
NR,P(U) <MKP'2C2

M - i Z i L RI+^A R7lK)P + X - i + K-^ +

v+2 ^MKP 2L2{AR7lK)Q 2 C 2 M - i Z i L i ^ f + ^ < 1/2

for any j ^ ^ 0 if we first choose i^7 so small that (to take care of the case p = 0) and then choose K so large that 2C2

<l/2.

In order to prove analyticity of u along OR when u vanishes there and /, a, b, and c are analytic near 0, we proceed in much the same way. But now, it is convenient to consider first only derivatives with respect to

168

Differentiability of weak solutions

the x"^ with (x <iv since these all vanish along OR SO that Lemma 5-7.1 applies on GR. We therefore denote x"^ by y, derivatives with respect to x'^ by Dy, and denote by Va; the gradient involving only derivatives with respect to the x'^ with oc <,v. As an aid in later estimates we state the following generalization of Lemma 5-7.4: Lemma 5.7.4'. Suppose V and W are tensors in ^^{D). Then

\ynv-w)\<i{^)\viv^-\vr'w\,
9=0

l^?vj(F-Ty)i<i'i'(^)()|z)vFHz)r'"vr''w^i.
and corresponding inequalities hold for \DIVHV'W)\ and | V^ V ^ ( F M^) |, etc. We now introduce the following notations: Notations. For u a n d / ^ C'^iGR) with u = 0 along OR, we define

M'n.Af) = (^0-1 sup {R R/2^r<R

r)^-Pmm.r
^ > 0

f (^!)-i sup {R -r)^+P\\\7^V?.u\\l^,

^k.(

sup (R rj2+2>||v2+^w||g,, p = - 1 , - 2 .

Lemma 5.7.3'. Suppose u and f ^ C'^i'GR), u satisfies (5-7.1) on GR, and u = 0 along OR. There is a constant Cg, depending only on v, such that

Proof. We apply Lemma 5-7.1 for GR to each component of V^ ^ and add, to obtain

j \V'^\/lu\^

dx <^CiU

\Vlf\^ dx + j [d-^\S7\/%u\^ + p->0.

-{~d-^\Vluf]dA,

We note that V, V^c, and Dy all commute with one another and (57-19) iV^^|2<|v2vr^^|^ :^>2; |V^^|<|V^^|,

p=OA.

After taking (57.19) into account, the lemma is proved in exactly the same way as was Lemma 5.7.3In estimating derivatives of the form D^^^ D^. u, it is convenient to introduce the following notation: Notation. N'^,,Ju) = sup [(p + g) !]-i(i? - /)2+3'+ \\Df^
iS/2S:r<B

VMir>

M6C~(GB),

P>0.

q>-2.

5.7- Analyticity of the solutions of analytic linear equations

169

Theorem 5.7.1'. Suppose a, b, c, and f are analytic on GRQ and satisfy (5.7.14) there, and suppose u^ Hf*{GEQ) and is a solution of (5.6.1) there, where X = 0. Then there is an Rg < Ro> which depends only on v, A, L, RQ, and corresponding quantities for certain ratios of the coefficients andf, such that u can be extended to be analytic in BR for each R <, R^. Proof. As before, we know that u ^ C'^iG-R) for each R < RQ. The proof proceeds as in t h a t of Theorem 5-7.1, b u t estimating only the ||V^ V^^||2.r- The estimate (5.7-15) carries over without change. Instead of (5.7.16) we obtain \Vl(F-f)\^LAR\V^V%u\+L\VV%u\+L\V%u\ +

By using (5-7.19) and proceeding as before, we obtain first the estimate (5.7.17) with MR,P and NR^P replaced by M'^,^ and A^i?,^,, respectively. As before we then obtain (5-7.21) N'j,^^(u)<M^Kf which gives estimates for all derivatives Dl S/^u where ^ < 2. In order to estimate the remaining derivatives^, we note that we can solve 5.6.1 for D^u in terms of the other derivatives. (5.7-22) Dlu=g+ ''^B-DyDo.u + '^C^^^D.Dpu

in which g, the 3", and 0"^ are analytic on GRQ, We shall assume that g and the 3" and 6"^ satisfy estimates of the form (5.7:23)
(5.7-24)

|^?VS^|,

\DIVIB\,

\DlV%C\<piq\LAP^^
^ > 0 , ^ > - 2 ,

in GR,.
0<R<RS,

We shall show that if 7^g is small enough, then

N'ji,:p,a{u) <MKP+QOP,

where M, K, 6 are fixed constants with K > 1 and d < 1/2. Differentiating (5.7.22) and using the bounds (5.7.23) and Lemma 5-7-4', we find that (5.7.25)
m-Qn=0

\Dl+ \/lu\<p\q\LAP+Q

Noticing that A^^,J,,Q(^) acts like a norm, we next conclude that (5.7.26) ^R,^,M <LZ2(v) R^-^^AR)P+Q + + q) !]-i X + L 2: Z im) il) ^^' n\{p + q - m - n)\ {A R)^+^[(p

Ml1 The remainder of this proof could be replaced by a dominating function argument like that used to prove the CAUCHY-KOVALEVSKY Theorem. Or a theorem of HOLMGREN [1] may be employed.

170

Differentiability of weak solutions

Now, (5.7.24) will hold for 5^ = 2, 1, and 0 and all ^ > 0 if i^ is small enough for (5.7.21) to hold and (5.7.27) M KP+^QP'>M^Kl+^, p>0, ^ = 0,-1,2, since it follows from the definitions that (5.7-28) A^k^..M^^kp+aM if J ' > 0 , ^ ^ 0 . Let us suppose, then t h a t (5-7.24) holds for all q less than some positive integer, which we again denote b y q, and all ^ > 0. Then, using (5.7.26) and (5.7.24) and the fact that, for _^ > 0 and ^ > 0, the factorial coefficient of each term in the double sum is < 1, we conclude that

m=0 n=0

The right side of (5.7.29) will be < 1 always {R < Rs) and (5-7.27) will hold if we choose Rs, 6, K, and M in that order so that LRl+^^\l2, KO = max(iCi, 2ARs), d = mm{\l2, I/I6L), M = m a x [ 1 , M i (^/i^i)^].

5.8. Analyticity of the solutions of analytic, non-linear, elliptic equations In order to conclude the analyticity of the solutions of variational problems and of the quasi-linear equations mentioned in the introduction and treated in 5.1 5-6, it is necessary to give a separate proof for the non-linear case since the methods of the preceeding section do not generalize immediately. The analyticity of the solutions of a single linear analytic elliptic equation in two variables was proved by HADAMARD in 1890. The analyticity of any solution u [x, y) of class C^ of a single nonlinear analytic elliptic equation with v = 2 was first proved in a famous memoir of S. BERNSTEIN [1] in 1904. Other proofs of Bernstein's theorem
were given by M. GEVREY, T . R A D O [1], H. L E W Y [1], and b y B E R N -

STEIN himself [2]. In 1932, E. H O P E [3] proved the corresponding theorem for a single equation b u t with v arbitrary. In 1939, PETROWSKY proved the theorem for systems which are almost as general as the most general elliptic systems yet considered (see 6.2, 6.3). In 1958, MORREY [12] and A. FRIEDMAN [1] proved the theorem almost simultaneously but b y different methods for the most general elliptic systems; analyticity at the boundary was proved for the solutions of strongly elliptic (see 6.5) systems.
The methods of B E R N S T E I N , G E V R E Y , R A D O , and FRIEDMAN all

involve estimating the magnitudes of the successive derivatives; by

5.8. Analyticity for non-linear elliptic equations

171

t h a t method, FRIEDMAN was able to obtain corresponding results for quasi-analytic systems involving the quasianalytic classes of MANDELBROjT. Those of L E W Y , H O P E , PETROWSKY, and MORREY all involved

extension to the complex domain. L E W Y used his theory of hyperbolic equations to carry out the extension and H O P E extended certain generalized potentials to the complex domain, a procedure which had been used by E . E . LEVI. The proof of the writer combines this idea with a simple type of implicit function theorem in a Banach space and the idea of 5-5. We present this proof in this section. The proofs for more general systems will be sketched in ^6.7. We consider a solution u of an equation of the form (5.8.1) (p{x,D'u) = 0, 0<\(x\<2, oc = (oci, . . .,ocv) in which 99 is analytic in its arguments (x, Vx) for x near XQ and r^ near D' u (xo) and is real for real values, and the equation is elliptic at this point; that is, the linear equation (5.8.2) 2<PrA^o,ro)'D-v
a

=0

is elliptic. B y a translation of axes, we m a y assume t h a t XQ = 0. Next, let Q{x) be that quadratic polynomial such that (5.8.3) V*<3(0) = V^w(O), u=Q ^ = 0,1,2. Making the substitution (5.8.4)

+v

in (5.8.1), we obtain a new equation of the same type in which V^^(O) == 0, ^ = 0, 1, 2. If we expand the new <p about the origin, we obtain a new equation which we can write in the form a'^fi v^a^ = b^v^oc + c^v + ipix, D v) {0"^, h\, q const.). Finally, b y making an affine transformation of the x space, we obtain (writing u again) (5.8.5) Au=^ M u + ip{x,Du), M u = b^'u^oci- cu. The Taylor expansion of ip starts with a homogeneous linear function of X which is followed b y terms in the x'^ and TX of second and higher orders. We first suppose u defined and ^ C^ (^RQ) and introduce the idea of 5.5 in which we write (5.8.6) u = UR + HR, UR = PR[M U + y){x,D U)], HR = U UR where we make the following definitions: Definition 5.8.1. The space *C^(BR) consists of a l l / ^ C(5ij) for which /(O) = 0, the norm ||/||o being hfx{f, BR). The space *C^(BJR) consists of all U^CI(BR) for which u{0) = Vu{0) = V^u{0) = 0, the norm \\u\\2 being h^(V^u,BR). For / $ *C0(5i?), we define PR{f) as the potential o f / m i n u s the quadratic polynomial satisfying (5-8.3).

172

Differentiability of weak solutions *C^^{BB)

From the results in 2.6, we obtain the following theorem: Theorem 5.8.1. PR is a hounded operator from *C2(5/j) to with hound independent of R. If F = PR(f), then (5.87) AF=f.

Thus, in (5.8.6), HR is harmonic. Proof. The only thing not yet proved is (5.8.7). But, from its definition, AF [x) = f{x) -\- C, C Si constant. Substituting x = 0 gives C = 0. Now, as in 5.5, we regard HR as known and shall try to find an equation for UR. To that end, we define .5 g g. VR = PR[M HR + yj(x,D HR)] TR(uR;HR)=^PR[MuR + rp{x,DuR + DHR)-y){x,DHR)].

We easily obtain the following theorem: Theorem 5.8.2. Suppose u is a solution ^ *CI{BRQ) of (S.S.S) and suppose UR, VR, HR, and TR are defined as ahove. Then \\UR\\2, \\VR\\2, OLnd I HR |]2 are uniformly hounded hy some numher M2 for 0 <C R ^ RQ, I
TR(0; HR) = 0, lim \\VR\\2 = lim \\UR\\2 = 0

(5.8.9)

IITR(UIR, HR) R-*0

TR(U2R, if

HR)II2 < e{R) \\ UIR -

U^RU,

lims(R)

= 0

||//||,||C/iij||,||C/2is||<M2.

Thus, if 0 <i R ^ Ri, UR is the only solution with \\UR\\2 < M2 of the equation (5.8.10) UR-TR{UR;HR)=VR, ||//ij||2<M2, HR given. Proof. Obviously ||^||2 < L for 0 < 7^ < RQ. L e t / = M u + ^p(x, D u). Then, for x, xi, and X2 on BR, we have (5.8.11) \V^u{x)\<LRf', \Vu{x)\^LR^+f', \u{x)\<LR^+f' (A<^) u{xi)I

1/(^2) - f(xi) I < I^I IVu(x2) (5.8.12)

Vu{xi)I + IcI \u(x2) -

+ \\7xf\ \X2 - xi\ + IVr^l <,ZiRf' '\x2~ .ri|^

fc=0

2'|V^^(:v2) -

V*^(A;I)

where |Va;V^| and jVrV^j are bounds for the indicated gradients for x^ BR and the r^ in the range defined by (5.8.11); we know that j Va; V^j is bounded and \Vrip\ ^C R^ in that range. Thus ||/||i ^ 0 like Rf^ and it follows that ||^i?||2 - > 0 . Thus ||^i?||2 < i^ + \\UR\\2 SO \\VR\\2 - ^ 0 . Obviously TR (0; HR) = 0. To get the other bound on TR, we note that
TR{UIR,HR)-TR{U2R,HR) = PR[M(UIR U2R) + y^{x, D UiR + DHR) yj{x, D U2R + DHR)]

and estimate the bracket as in (5.8.11) and (5.8.12) above. If we assume that Ri is chosen so small that s{R) < 1/2 and \\VR\\2 < M2/2, then it

5.8. Analyticity for non-linear elliptic equations

173

follows that the equation (5.8.10) has only one solution UR with \UR\2 <M2. Now, after the manner of 5-5, we wish to introduce certain spaces of analytic functions, which we shall denote by *C|^(5/jij) and *Cf^{BhR)
and which are subspaces of *C^(BR) and '^C^(BR), respectively, such

t h a t every harmonic function in *Cf^(BR) is also in '^Cf^{BhR) and PR is a bounded operator from *C^^(BfiR) to ^CKB^R) as above. The functions in these new spaces can be extended to be analytic on the domains BfiR in the complex %-space which are defined below. Given / defined in some domain, we have to show how to extend its potential as given b y (2.5.2) to that domain and verify the other properties. Now KQ {y) is a function of | y | only and, for real y,

If, now, we consider complex vectors y = yi + i y^ (yi and yo real vectors in R^), we see that (5.8.13)
a=l

1^(^)2 = |yi|2 - |y2|2 +

2iyi'y2.

If we choose h, 0 <. h <c 1, and assume t h a t (5.8.14) then (5.8.15)

a=l

|y2|<A|yi|

(yi^o) -^i^)\yi 2

> Re2^(y<^)^ = |yi|2 - [ysp > (1

Consequently, it is clear that we can extend KQ (y) uniquely to be analytic in any region of the type (5-8.14); one encounters branch point trouble for h->\. In extending the formula (2.5.2) to complex values of x, it evidently will not do simply to let xbe complex, keepingf real, because iix = xi-\-\- i X2 and A;2 9^ 0, we note that X i = (Xi i) + iX2 and \x2\ is not <. h - \xi i\ for f near xi. Thus we must allow f to range over some r-dimensional ''surface" which evidently must pass through X, therefore depending on x, and must have OBR as its boundary. Consequently, in order for (x f) to be in a domain (5.8.14) for | on such a surface, x must be restricted to a set of xi + i ^2 such that (5.8.16) \X2\ <h\xif I for all | on OBR. Since the minimum of the right side of (5.8.16) is just R \xi\, we see t h a t X = xi -{- i X2 must satisfy (5.8.17) \x2\<h{R-\xi\).

174

Differentiability of weak solutions

Next, any surface S{x) over which we integrate must have the property that if S = h + ih^S{x), then |^2 f2| < ^ |A;I f i | . We m a y describe certain such surfaces b y
S(x) : | = | ( s ; A;), \h{s, X) X2\ <h\li{s,x) xi\, S^^BR,

Finally, the integral over such a surface is to be interpreted property. We begin b y defining the domains BfiR and the function spaces we shall use: Definition 5.8.2. The domain BhR in complex r-space consists of all complex X = xi + i X2 which satisfy (5.8.17). The space *C^(BJIR) consists of all / ^ *C^(BR) which can be extended t o be (single-valued) ^ C^{BhR) and analytic in BJIR, the norm being hfi(f, BJIR) (i.e. the Holder
constant on BhR). The space *C^(BJIR) consists of all u^ *CI{BR) which

have similar extensions to BhR, the norm being hf^(\/^ u, BJIR). We now define our surface integrals. Definition 5.8.3. Suppose ^ is defined, and continuous, and complexvalued on a set D in complex i^-space, suppose S : f = f (s), where f(s) is a Lipschitz vector defined for s^G where G is a Lipschitz domain in r-space, and suppose that f (s) ^ Z) for s^G except possibly a t s = SQ^G and ^ [ f (s)] is continuous except possibly a t SQ. Then we say that S is admissible with respect to g and define the integral (5.8.18) f^(S)d^=fm(s)]J(s)ds, J=

d(sl ,SV)

in case it is absolutely convergent. Theorem 5.8.3. Suppose / ^ *C^^{BfiR), XQ ^ B^R, and S : f = | ( s ) , s ^ BR is admissible with respect to the function (5.8.19) and I satisfies the additional (5.8.20) fi(s)-s, %[^)Xo)=Ko{xo-^)f{^) conditions 5^5,^, 12(5)= 0, s^dBR.

Then the integral (5.8.18), with g^ defined by (5.8.19), exists. If i and | * are both admissible and satisfy (5.8.20), the integrals have the same value. Finally, if for each x^BfiR, S{x) : ^ = ^(s; x) is admissible with respect to ^ ( f ; x) as defined in (5.8.19) and | ( s ; x) satisfies (5.8.20) for each x, then the function F defined by (5.8.21) F{x)=JK^{x--^)f{^)d^
Six)

is analytic on BhR and (5.8.22) F,4x)=fKo,4^-i)f(S)dS.

5.8. Analyticity for non-linear elliptic equations

175

Remarks. We notice that in order for 5 : | = | (s) t o be admissible with respect to ^ ( f ; XQ), i must satisfy also . g . 12(^10) = ^ 2 0 , \h{s) -X2Q\<h\s -xio\

Proof. From t h e definition of KQ and from (5.8.15), (5.8.20), and (5.8.23), and the fact that f is Lipschitz, it follows that | / [ f (s)]| and I / (s) I are bounded and KQ satisfies \Ko[xo - | ( s ) ] | ^ 2 i (1 - A2)i-/2 Is - ^ l o P ^ ^ ^^ (5.0.24) , , , , ( i ' > 2) ^ | V X o [ ^ 0 - | ( 5 ) ] | < Z 2 - ( 1 -^2)(l-^)/2.|5_;,^0]l-r ^ ^ so that the integrals (5.8.21) and (5.8.22) converge absolutely. There is an obvious modificiation iiv = 2. If f and f * are two admissible vectors ^ ^^(Bji {^10}), let f (5, t) = (1 Q f (s) + ^ {* (s) and define (5.8.25) (p(t, Q) =fKo[xo
BR-B(XIO,Q)

- i(s, t)]f[^(s,

t)] / ( s , t) ds,

Differentiation with respect t o t yields

(Pt

(5.8.26)

_ r ~J ^

'

5(ji_..., |v-i_ cj ^y ^.+i_..., J, d(s\....s'') '^^'

as is seen by expanding the determinant with respect to the y-th column and using the relations

d(% ^)lds' = i r ^f^l, + ^ 32 fv/5s dt.

One sees in turn b y using Lemma 4.5-6 and the proof of Lemma 4.5-7 (for complex-valued functions) t h a t

(5.8.27)

Mt'Q) = -

y.(5=l

2^-^^'""h'^*^'^'''6B{xto,Q) ^

From (5.8.24), (5.8.25) and (5-8.27), we see that (p{t, g) -^(p{t, 0) and (ft {t, Q) ->0 = (ft [t, 0) uniformly, so the integrals have the same value for f and f *. If | and f * are merely Lipschitz it is easy to see that one can approximate to each b y C*^ admissible functions. To see the analyticity, pick XQ = ;tio + i :^2o BJIR. I t is easy to construct a family S{x) of admissible surfaces i = i {s, x) which ^ C^ as above in both s and x for x near XQ. Let us define F{x,Q)=^J%[i{s,x);x]J{s,x)ds,
BR-B{XI,Q)

176

Differentiability of weak solutions

where g is defined in (5.8.26). Clearly for each sufficiently small o > 0, F{x, Q) ^C^ in X near :vo and we obtain
PX^Y{X, Q) = fKo^y[x| ( s , x)]f[S{s, x)] ds -

(5.8.28)

''"^'^' PX^Y{X,Q)

' ^ " " ''^' =ifKo,y[x i(s,x)]f[^{s,x)]ds -BK--B(a!i.e)

by making use of the devices used in (5.8.26) and (5.8.27). Using the bounds (5.8.24), we see that F(x, Q), Fx-^y{x, Q), and Fx2v(^> Q) converge uniformly to F, Fx^y, and Fx^y and t h a t Fx^y = iFx^y so that F is analytic with derivative given by (5.8.22). Definition 5.8.4. For / ^ *C0(J5/^^), we define U = PR{f) b y the formula (5.8.29) U{x) = -Q(x)+lKo(x-i)md^,
Six)

X^B^R

where S{x) is admissible for each x and Q is the quadratic polynomial satisfying (5-8.3). We must now estimate h*{\/^ U, BIIR) when U = PRI/) a n d / *Cl(BfiR); clearly we m a y assume, for this purpose, t h a t U is given by the integral in (5.8.29). As in 2.6, we estimate V^U{x) and then use a lemma like Theorem 2.6.6 which is included in the lemma we state and prove after introducing the following convenient notations: Defin