Multiple Integrals in The Calculus of Variations Oct 2008

Classics in Mathematics Charles B. Morrey, Jr.
Multiple Integrals in the Calculus of Variations
Charles B. Morrey, Jr.

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. Hirîebruch 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
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
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 Vû denotes the tensor u ,oc^ where oc and /5 run independently from t o r . Likewise V ^ ^ = {^,a/Sy}, etc., and \\/û\^ 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 CîG), 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 qîasi-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'â2.,(^,^>^)â^^*
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 leêak 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
dxÔ
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''^ ôc^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)â^^[x(y)-]dld$, 'c{y) =^-c[x{y)].
lo , otherwise J r2 :== (^2)2 + . . . + (y')2. Then if we divide 99''(0) by the measure of the support of co and then let h and H -> 0 so that hlH -> 0, we conclude t h a t ^ ^^ ^-^ -^ ^ 'all (0) = â^ (^^) ^1 ^1 ^ a''^{xo) > a A^ > 0 ^ which is the stated result. This is called the Legendre condition.
1.5.4)
co(y^,...,y'')
=y
f ( A - | y i | ) ( 1 - r / i f ) , if | y i | < / ^ ,
'-^ '^^
' ^
0<r</f\
1.5. Other classical necessary conditions
11
If we repeat this derivation for the case of the general integral (1.1.1), a s s u m i n g / ^ C^ of course, we obtain
^"(0) =I{ZI<CUC$
+ 2 2-bf,C'C\ + CijC'en dx > 0

c^ J- = /î 2?.
< f (^) == hi 4 [^> ^ (^), V ^ [x)], 6?,. = /g^ pi^, co^(yi, . . ., y") = fâ)(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^,
/î/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{ô,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{ô,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 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) ÂijC^dxÂi
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) ^^î^("^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 ^ = î 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(ô) = 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 [Ô>Ô]) 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, ô], 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 <û^ 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[wâ'û,ocU,p
G
uâ)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
+ ÎKA
+ + 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 ^^^^ ê ,a(^) = {u ,OC)Q(X) for X^GQ SO that UQ ->u
and UQ ôc ->u ôc 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(^)âpS^ + 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 /[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 /[VêW]<êW> 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|î-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"^^ îo(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^ê 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'dlVWDîrlRy 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 / / =
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^piîx, 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
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) -^Â^ = 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\^Â^{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\^
^hOîj = 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^^PWô. + DP^W)]dx ifC^Lip,{G), C{x)ÔonG, ^ 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)zôc^ = 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^ CîG), then z^ Q ( S ) . / / we merely know that the A'(p)^ C^ and satisfy the condition ^p^{P)ôck>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
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) ^ 'ûôôi[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) Juôcdx^=^ 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 uôi{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 uôc{xo) + ^^^'^^ \R(x,xo)\ +-^{^"ô) (^^ - ^ 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^LîG), 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 \\/û(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
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'îx) 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 + fôi^
- ô)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
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-^^^Gôc = i ^ - i r ^ - i f ^ p l - x\-'{^^ - A;)

;|:^|/i^)2-r||_^'|-v(|a_^â)
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 ^î2, 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*(ô)]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 ^î? B{ô, Q) where we m a y choose g so t h a t | ^ * ( | ) w*(fo)| < e/2 for | ^ ^jBi? Pi B(Ô,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
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) - ô,.(^' - 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{xî)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
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/î),
\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 Kôi 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
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) jrô.[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).
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 |:î :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) ô(^) = | l o - : ^ | , n{v) From plane geometry, we conclude t h a t (2.6.14) = \î-x\>
ô(^) '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( ^) logvî(^) -~logy)i(
4*
52
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 lfjÂ{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 = | :î :^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 [î| > -?^ ^, 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.
53
To prove the second inequality, we extend / t o Bû 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 lî^ 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 - îh (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)=IÂix-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) Wîxi)] 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/Û) <Ch^(f), Proof. From Theorem 2.5.1, it follows that U ^ C^ with
u,4x)=fKo,4^-i)mdi.
BR
54
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 iô,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^[^(ô, 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=jÔ, 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/(î) / 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^ Cônlx] < 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\-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^ CfîRi) 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^ Cî^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 rjÎ! 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^lîi^^, ^-^^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) - /^(ô + [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 Koôc{x ~ d
G
i)uôc(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'ê(xi,X2, + Ci + R), v = 2
58
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].
gÔ 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\), î-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
< 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)hôi(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) uô:^{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) {î}> 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'^îx^' ~ 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 vôc{xs) = 0) (2.8.5) a^'îxs) vôc,3{xs) > 0 but vûpixs) 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)-Ê0, then u has neither a positive maximum nor a negative minimum.
Chapter 3
The spaces H"; and ilô

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 Tâ(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'û^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 Vû^ H^-^(G), 1 < ^ < m. Ifu^H^{G) and V'û^Hl{G), then u ^ H^+'{G). Ifu^HîG) 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 ^ uôc. (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 = uôc, 1<|^|<m. If m ~ \ and uôc(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.
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 vÔL given by the limit of the right side of (3.1.6). If, now the generalized derivatives u ,^(XQ) and xô,{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 (ô) ^"fa(yo)] ^y I

I e I
(3.1.7)
<\e\-^ f lÂîy)] - ^,p(ô)\'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,
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)-^/ lô(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)-Î dx'fd^'='f\u,cc(i'', 0 | ^ < < /3. 0<h<ho.
Then (setting 6 = x'^ x^)

I a;o+/i xfî+h x'xo+h [
( 2 A ) - / |wo(|) - uo{xQ)\di - {2h)-^J

I tco^ xîh
/ |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 Uâ{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 \Vû[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) = '^(ô + ^ 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\'7û(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^HîG), DcG, v^H^(D), v - U\D^ Hôi^), 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 ||iî|| = ||/|| and iî 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 -^îth 0 < | ^ | < w are uniformly bounded on G, then l^ Un7 iû 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 ô^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^,aÂ; + 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 -û 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'î^^) 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)~ô''~^.

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 Dûix) = 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 Kâ 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 Kôc{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
<Mfî 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 = (û, 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
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'â}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^îQ ^^^ Hô^{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
<âe(|^| <= ^ ^) 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 = î,ocy q^xQ = UQÔC 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 î 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^"î^"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-û in H^{G), then Un -û 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,
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 ==Ûns 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 îG 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.î^ ^^ 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 î 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^~^{î) (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-Ûs 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\-ÎPI^^\\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)!]-î^^~M fvû[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'û(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)\-Î
\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
(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 \ 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 {Wvôcfi)'^"
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 \uî[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 \ uî{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 \\Û\\l,j,^C-'\\u\\l^, where {'\\UIIJ,)P =J{\V
BR
u\^ +
R-û^)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 ylÔR, 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" Ô"^ 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.
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 ~ÛQ 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 BuoÔona. 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 Vô = 0 but ||ô||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-ûo 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
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) \\Vû\\l<e\\V"û\\l + C\\u\\l u^H^{G). and Proof. For otherwise there is a sequence {un} with Un^ H'îQ I] Un 11^ = 1 such t h a t
(3.6.4) I VHin I t > 4 ^Ûn \\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 || \7Ûn\\^ - ^ 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'û\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 VÛn-Ô 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 î^ - ^ ^ 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^ ÎdQ ^^^ ^^^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|2Â;<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 hîf.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 - î-'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(ô;^)).
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^ + 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 Â; < y IV2 Fi |2 Â; < 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) ô < 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)Â{\ + \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)â\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 >ô and jtn -^no- Since / is convex in p for each t, we have
(4.1.4) f{tn, Rnn) < (l -Îifn. ) + ^ ^^ " 0' 0<i^<i^^.
Holding R fixed and letting ;^ - ^ o o in (4.1.4) we conclude that /(ô, 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 ^ Vêl^:^ < 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).
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, ^ô = 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)
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)+ Îg{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' 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) ^ CîQ- 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)ôcPi^ ={\ - (o) Al^XAp s = rlR, Since for large r, \A{p)\ Al^ÂccA^^ (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 wâA^'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-^î^)

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\ + ||V9î)
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 ^î 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 î? 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 .ô 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)uôcp(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 ) , Mîy) = 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 awû>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-Ôin 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 ~ CnWlû
= 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) Âr (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)îA(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(^+^)Jî^ 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 î(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) > ?î-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, +î+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-ÎIJ,^ 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^ÛY
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) -^(î)! 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 + ô.^^)>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 (fgpîp) 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 ô(^) = ^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 î? 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
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|^Â;<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 î^ 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^î^)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) -
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-Ô. 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-^Êl 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 = înt where {%} are those integers for which ^li^îwtPaThis process is continued until oc = v. We define x'^^ = a"", x^^ = ^* a = 1, . . .,v. Next, we define ^J+i^2i = î,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 Îtp^- 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 îiÊ]^'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 Zvî,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]).
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 î, let iî, , . ., Rqbe the cells of Djc^ and for each k>k\, Rici[i= 1 , . . .,^-2*'(^-î))
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/[^,ôW, ^ 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(|ô(^) 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 + |î|^-l + |î|^-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'-') '(\ô-zn + \po-pt\)>

+
ô|.
K(\Po(x)\s
+ \zo{x)\^ +i).\x-x^(x)\
(4.4.10') K(\Pn(x)\s + |ô(^)|^ + \ ) ' \ x - x^(x)\ + + K{\ + \zo\^-^ + \Pn\'-^+\zS\^-^ + \Pt+7tn\^'^){\z-Z^\

(4.4.11') K(\ + lôl^-^ + \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 Â-^ for some s > 1 or satisfies those of Theorem 4.4-5 "iîth 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 î, . . ., 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^^+^+^ ocpî
^=1
+ C^)
bx^^^
a=l
d;i;* d(;irl.
Aâ-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
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^--+ Aî::::-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) aîL?i^m^Ô 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'î 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 = iîrjj 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)î = 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,...,Xrî), where F is continuous in the Xi and
Thenf is strongly quasi-convex inp if and only i/Fis
convex in
theXt.
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^Xjcviâ^^' p^XjcÔ (^ = 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=^ -MXiî), ^==\,...,v). XÂX^^DtX^î, 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) XiîÔ 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.î = 0 for all A, f for which (4.4.33) HmX^
i=|=fc
- DlXi) Lî = 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^Â% = KD^).
From (4.4.29) and (4-4.35) it follows that (4.4.36) AfX-î == KD^X^î + maXiî C^={KXjc-L-pl).
CÂXj,,
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 hî = XjcAXjc (Dtho: + hv+iXi)ii = 0 = DtL^K with XiîÔ] 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. CÂXu.
Chapter 5

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^ Lî^{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
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 vôcd^"^ u^pdx,
G
L(v) = j {e'^vôc + fv) dx

G
C(u,v) = f uv dx,
G
Bîu,
v) ^= J (vôcb' u -\- vc^ uôc + 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
Rî'{\\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'^CoZirî2;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'î.
(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 Îp < \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
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 llî.^ + llî.^. + ||/||,,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), / ^ LîD), and V b and d satisfy (5.2.11) fl\Vb\^
B (a!o.r)n2)
+ \d\^f^dx<(Corf'iY'\
î>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'uôc + 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'ûôi + du -\- f, 0<\h\ < H, 9*
132
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 Â^ (5.2.15) + ^^r) + hlix) u(x + hey) + b^(x) un[x) +
1
jv,^{a^ûn,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^LîD"), 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 LîD') to u^^{x), hi -^h'^y in Lv{D'), u(x + hey) û(x) in L2s(D'), Ufi{x) ->u^y in LîD'), el -ê'^^ 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^î^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 îf ^f/3 ^^>
C(u,v) = f u^ v^ dx
B2{U, V)= f p; bfÛ^ + Vi{clûi^ + 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\^âl^{x)7Ti7t^, SIITI, \a(x)\<,M, m>0
/(l^l^ + kl^ + IdlY^^dxîcorfîfK

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\^\^\^âf^{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 / îi Û ^!/s ^ ^ = ^ ^ / î^ 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
/^ 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) ipcoôc and ipôcco^ L2(G). Then there exists a sequence { C } - > ^ *^ ^\o{^)> *^ leÂ^'c/^ âcA 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, Â^ 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) <îpO),
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 ^nÊipcG with 0 < C w ( ^ ) < - ^ , such t h a t CU-ÔJL 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,
UÛs{D), PUÛiD), VPF, PW, PÛ^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)
C2Tâ-^ Jw^dx,
B{x^,R+a)
B(xo,R
a)GG,
^ =: 1 + 4fî^, 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
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'-'Û^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-Î
Ul'-^dx,
0 < a < oc,
B {xo,a)
where (5.3.16)
2 Z2 Z4 r^ ocî =
\. 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 ôc 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^â)
<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{\VWnî\^
[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-^,
iî = 4 s^.
From this recurrence relation, we conclude t h a t W}!^"" < i^î-^"'^"' 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'û^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~û,
if^LipdG),
y)(x)ÔonG
in the equation above. An easy approximation shows that this is legitimate. Since V V^x = u uôt so a'^V,ocV,^ = V-^ u^ a""^ u,ocU,^ <, a''^ u,ocU,p,
138
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 ' "û + / ) ] 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)dxÔ
for each
C^Lipc(G),
C(^) > 0.
(5.3.22) Remarks. This condition is formally equivalent to the condition -^râ^'^VR C'vô^ > 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 blîVv
a'S/v-\-C'S/v)
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
[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 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
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 = IIVî.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-^î^,
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
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< [LîSjRy-^^^'']' (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) 'îth' 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^|2Â;<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
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\^^ -î(^) ^'^^ (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(|| Viî.)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 _ ^) î^(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 Qfî 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 rî^' / I V % P Â;

G 'P G 'P G
<, 2C Corfî f \\/u\^ dx + 2C Corfî 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\\%'Rî
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 Rî], I V ^ i . , <.C[{RI r ) - i ||^||o,^ + ||.||,^ + Coi^î], 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
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 fô, 0 < /ô < /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'û + e^", (5-4.10) F = c'^'uôc + du + f, + vF]dx = 0, x^B(XQ, a). E and F being 0 elsewhere. Then u, E, and F satisfy l[v,4a^û,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 
fFvdx->
Bixo.B)
-Zi{L2
-
+ Z 2 L a n l ^ ) [Rja)^ ||V?;I^R
{Z^Ll+Zaân)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'ân){Rla)'î^
5.4. The case v = 2
147
where we have assumed that /j, < /î/2 and ju = (//i/2) (A/4). If we now choose juo so that /ô < /*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^î) from which we easily conclude (if 0 < ytt < //o), that cf>[R) < [^() + Z8(Z| + L| + Z2 Âî) (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 yjâOJ^ 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 -îp o),oc and PipCn-^Pyôj 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
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 ^^^ = wôc 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. JlKiîj
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, ., îv^), 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
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^û,pdx
G
= f vôcd'^'û^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 Ê^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,4alû,^
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. < CLRî'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 /î > 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
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) 'iîth 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 rî-%
\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) Iiâ-^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$p{^)\dx^l(Lsî, Bi$.s)nB{xo,r)nr [^ [Lrî, 0< s< r s ^r.
From (5.5.14) and (5-5.15), we find that (5.5-16) l2<Zi{v, /Lii, o)Lrî-^(''-^)('\\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(LRî) '|| UR ||J, (ii) II en ||S < [rj{R) + C{LR^î)^'^] '\\ 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{\ + LRî). 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\Â
(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
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,û, (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^%û,^ 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 CLRfî 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 || î? ||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
Proof. It remains only to prove t h a t the second term in (5.5-20) C1{Â) 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 Vîô,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(iÎ||.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'S7û-\-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-û 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^RFÛS -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'û,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
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 î? = 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'Vû-\-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 bÛiF) 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^' k, then V^* ci, V^'-i b, and V^'-^ c^Lp with y"(|V^*a| + |V^'-i&| + IV^'-^cD^dx^RRfî, 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'llJLÎII^^, and spt u C BR or GR\J GR, respectively.
if
0<R<.R2
160
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 - 9îW]^f
22
since the integral of V ' V^ iô (^ 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^ < Jî, 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''Vû-\-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, WunllaÔ 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^ Cî^), 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)).
162
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-û\\la, Areal, A > A o , u^Hl{G) 0 HIQ(G), Proof. For any given >î > 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'^Ûs,ccpUsdx
G
j{a"-^ a^^)us,xÛsdx
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, jalÛs,aÛsdx = Ja'^Ûs,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(î) (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-^ Xûf - 2K(Lu,u) > I L i. 1 + A2II u IP - 2rii(\\ V2 ^ IP + A2II ^ IP) _ I 2 | ^ - 2A C(rji) \\u\\^ > IJL ÎP (1 - 2rji C2) + + [A2(l - 2rji) - 21 C(f]i) - 2rji C2] || u |P (II Ill).
where C2 is related to the constant of Theorem 5-6.2. If we first choose fji so small that (1 2 iî) > 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 > ô- 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
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\\7û\^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 Vû, we obtain
|v*W| <ci?-4^<?^||v+î,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.,,
PÔ ^ >- 2
= [^!]-i sup (R~-r)^+P\\\7^+û\\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>+lî2 + ^-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
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 + \)ÎPHP 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, ^î?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? < iô- 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,
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^-^'û\

3=1
+ ^(?)i^-4Q'g'!(|V^-^+2^| + i ^ ^ - ^ + i ^ j ^ |V2^-^|) (5.7.16) =LAR\S7v-^'û\

2> + l
+L{\
+pA)\V^^'û\
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^+ÂR)^
(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+Â 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
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''wî.
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'îGR) 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'î'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
M6C~(GB),
P>0.
q>-2.
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'îG-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 Aî?,^,, respectively. As before we then obtain (5-7.21) N'j,^^(u)<M^Kf which gives estimates for all derivatives Dl S/û where ^ < 2. In order to estimate the remaining derivatives^, we note that we can solve 5.6.1 for Dû 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^-^ÂR)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
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 (^/iî)^].
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<PrAô,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 vâ^ = b^vôc + 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^'uôci- 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) = Vû{0) = 0, the norm \\u\\2 being h^(Vû,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) \Vû{x)\<LRf', \Vu{x)\^LR^+f', \u{x)\<LR^+f' (A<^) u{xi)I
1/(^2) - f(xi) I < IÎ 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 ||î?||2 - > 0 . Thus ||î?||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
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|
(yiô) -î^)\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
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.
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
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 VÛ{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: Definition 5.8.5. For x = xi + i X2 ^BJIR, we let d{x) be the distance of X from dBfiR and r(x) be the distance in the plane X2 = const, of X from the intersection of that plane with dBjiR. We define Si{x) to be the surface defined by h = l i i ( s , ^) = s , Si{x): h = h2{s,x) =
0,
S^BR B{xi,r),
r = r(x)l2
(R~
\xi\ - r - \s - xi\) X2l{R - \xi\ -2r) r <.\s xi\ <, R \xi\ ~ r

S^BR~B{XI,R-\XI\ - r ) .
Lemma 5.8.1. (a) If x xi + i X2 ^ BJIR, then r(x)=R-\x2\lh-\xi\, d{x) = hr(x)iyi + h^ (b) BfiR is convex (c) Any two points zi and Z2 of BUR can he joined hy a path x = x{s), 0 < * < /, in BfiR such that .?
I
f {^ [x (s)]Y-^
ds<C{v,fi,h)'\zi-Z2\f'.
177
Proof, (b) is evident. To prove the formula for r(x), we note t h a t (fi, X2)^BnR ^\X2\< h{R - llil) and \ h \ < R ^ \ h \ < R - \x2\lh. To prove t h a t for d(x) we note that (fi, ^2) dB^R ^ (5.8.30) and the minimum of for all (li, ^2) satisfying (5.8.3O) is just that of (r - \xi\)^ + [h(R -r) - \x2\]^ iovO <r <R which is just the square of the distance in R2 from the point (\xi\, \x2\) to the line ^^ = h{R C^), (C^, C^) being the coordinates in 7^2Using (a), it follows t h a t the distance of Bfi,R-r from dBjiR is > hrl]/i + h^. Also if {xi, X2) ^ B^R, then (txiJx2)^Bh^R^r for all ^ < 1 rjR (r <. R). Accordingly the desired path could be the polygon .^1 2:3 2:4 ^2 where zs and 24, are the nearest points of B^^R^r to z\ and Z2 respectively, r = \z\ ~ Z2\We need also the following two lemmas: Lemma 5.8.2. Suppose XQ^B^R, Si(xo) is the surface defined above and V is defined near XQ by the formula \h\<R, \i2\^h{R-\h\)
V{x)=lKo{x-^)d^.
Siixo)
Then for all x near XQ, V(x) = v~^ XQ X"^ + const. Proof. It is easy to see t h a t for each x near A;O, a surface S2{x) can be defined by ^ =^ ^2[^\ ^) with 5 ^ B{xio, r), so t h a t i2{s, x) = fi(s, XQ) for s^dB{xio,r), and the whole surface S{x) = Si{xo) + S2{x) is a admissible for x near XQ. Thus V{x) + V2(x) = Vo{x) = JKoix
Six)
- I) d^,
V2{x) =fKo{x
Szix)
~ f) d | .
But for X real, Vo{x) just reduces to the potential of 1. Thus Vo(x) = {2v)-^ 2 ' M ^ + const., and V2(x) = {2v)-^ 2 ' i^^" - 4 ) ^ + const.
a
for the same reason, using a translation of axes. The lemma follows. Lemma 5.8.3. Iff^ *Cl{BhR) and L = ||/||o, then \Vf{xo)\ <Ci{v,f^,h)L'[d{xo)r~K xo^BuR. I V V W I ^C2{v,fi,h)-L'[d{xo)r-^,
Proof. There is a constant x, which depends only on v and h, such that the poly cylinder \^^ x'\ <Cxd{x), oc = 1, . . ., 1^, lies in BfiR.
Morrey, Multiple Integrals \2
178
Since / ^ C^^{BfiR), we have (using Cauchy's integral formula with y : II :\;| = Kd{x))
f{x) = {271 i)-^j... j m di^... d^-m^ - ^1)... (1^ - x^)

V V V V
for any complex t,"-. The first result follows from the fact that |(|._^.)-lâ|<||q|.[;,^(^)J-l. The second result is proved similarly. We now prove the analog of Theorem 5.8.1 ior BUR. Theorem 5.8.4. PR(f) is a bounded operator from *C^(Bfiji) into *C^(^/iij) with bound independent of R. Proof. It is sufficient to show t h a t V^F ^C^(Bjiji), which we do by estimating \/^ F{x) in terms of d(x), F being the potential as defined in Theorem 5-8.2. So, let XQ = :vio + i ^20 ^hR and let Si (XQ) be the surface defined above. As was seen in the proof of Lemma 5.8.2, this surface may be completed to a surface S(x) by adjoining a surface S2(x) with the same rim, if x is near enough to XQ. But the derivatives may be computed by differentiating only in real directions. So, we may choose S2(x) as the set of f = s + i X20, where s^B [XIQ, r), r = r(xo)l2. Thus, we have
F(X) = Fi{X) + F2(X), Fi(X) ^fKo{xSi (xo)
|)/(|) i |
(5.8.31)
F2(x) = jKQ{XI
- s)f{s + ix2o) ds,
X = Xi + i X2Q .
Clearly F\ is analytic for x near XQ and we can differentiate under the integral sign as often as is desired. Thus Fi,o.^y{xQ) = JKo,ocpy(xo - I) [/(I) -f{xo)]
St (xo)
di +f(xo)'
V,ocPv{xo)
in which the second term vanishes, using Lemma 5.8.2. Now, from the form of our integrals, it follows t h a t \di\ < i^-dimensional element of area. From our construction it follows that
\ds\<{i + h^r^dh
where dî is the projection of the element on the real Rv. Thus
\V^Fi(xo)\^Zi'L'l\h
(5.8.32) Iiv-B(xio,f)
- xio\^-^-^dii <Z2'L'[d(xo)r-^
(L = ||/||i).
179
Now jp2 is just the potential of a complex-valued function. First we differentiate F2,cc(x) = j Ko(xi
B ixio, r)
s)f(s + ix2o) ds
= +
Ko(xi s)f(s + i:r2o) ds^ + Ko(xi s)fô,(s -Iix2o)ds.
B(xio,r)
Differentiating twice more and using Theorem 2.6.5, we obtain F2,ocpy(xo) = f Ko,^y{xio s)[f(s + t X20) ~ f(xo)] ds^ + + C^vf,o^(ô) + f Ko,fiy(xio s)[fôc(s + iX20) f,oc{xo)] ds. Thus
B(xio,r} 5J5(ccio.r)
I V3F2(^0) \<ZsLr^-i
+Z4L[d(xoV/-^
+Z^Lr[d{XQ)^-^
using Lemmas 5.8.1 and 5.8.3. The theorem follows from (5.8.32}, (5.8.33), and Lemma 5.8.1(c). Theorem 5.8.5. / / i J ^'5 harmonic and H $ *C^(B/?), then H ^ *Cf^(BhR) for each h < \ and
\\H\\2n^C{v,^.h)\\H\\2.
Proof. From formula (2.4.4), it follows that H{x) = -f [H{S) KoM^ -i)+Ko(x-
i) HR,4m
d^'^.
I t is clear that H is analytic in any Bhu with A < 1. Differentiating, we obtain [x Xi -{- i X2) H,^y6[x) == - l{Ko,.^y3{x V2 H(xi)' - f) [/f (f) - H(xi) - sjH(x^)' (f - xi) -
(I - î)2/2!] + Ko,py6(x - I) [H,4^)
- H,.{x{)
V/f,a(^l)-(|-^^l)]X since it follows from Theorem 2.6.5, differentiation, and analytic continuation that
j Ko,oc6yd(x dBR i) d^'^ = J Ko^ccfiyd(x
BBR
^) - (^ -
x) d^'^ =
. . . =
and also i -~ x = ^ xi -]- xi x, etc. Thus, using the properties of H for X real and the fact that | | xi\ < |f ^ | < (1 + h^)^\^ xi\, we obtain \^^H(x)\ ^ZiL j \ ^ - xi'f-^d^^Z2L{R - \xx\Y-^ ^ZzL[d{x)r-K from which the result follows. L = \\H\\2
12*
180
We can now prove our interior analyticity theorem: Theorem 5.8.6. Suppose u^ CKBRQ) with u{0) = \7u{0) = S7û(0) = 0 and satisfies (5.8.5) there, where ip is analytic in the {x, r)-space near (0, 0) and has the properties described there. Then u is analytic near x ^= 0. Proof. Let us choose ^ < 1. From Theorems 5-8.2 and 5.8.4 it follows t h a t I|i5/î2||2fe <.Mz uniformly for R < iô- If we restrict ourselves to UiR and U^R in *C^(i5/iij) with norms < M3, we see that (5.8.11) and (5.8.12) hold for complex x in B^R SO that if we take J?2, 0 < i?2 < ^ 1 , small enough TR satisfies (5.8.9) with e[R) = 1/2, considering it as an operator over '^C^[B}IR). Thus, there is a unique solution UR of (5.8.10) with ||wi?||2/^ < Mz and ||î2||2 < M^- These must coincide so t h a t UR also
''CI[BUR). The result follows.
We now wish to consider a solution 1^ of (5.8.1) where X[i is a point on an analytic part oi dG along which u vanishes. A simple analytic transformation of the X coordinates enables us to assume that AJQ = 0 and a part of G corresponds to GRQ with a part oi dG corresponding to ORQ. The substitutions made earlier lead to the equation before (5.8.5). Then an affine transformation, followed by a rotation of axes yields (5.8.5) on GRQ where ^ = 0 on ORQ and | V ^(0) | = | V^ w(0) | = 0 . We may assume that ^ has its previous properties. We write (5.8.6), where we make the following definitions; the extra difficulty in defining PR{f) arises from the fact that the hemisphere G^ is not a smooth domain. Definition 5.8.1'. The space *C(G/?) consists of a l l / ^ CI{GR) with /(O) = 0, the norm ||/||o being A^*(/, GR). The space '^C^{GR) consists of all u^ C^ (GR) which vanish along GR and for which also V ^(0) == V ^ ^ (0) = 0, the norm ||f^||2 being hfi{\J^ U,GR). For f^*C^^{GR) we define U = PR(f) b y the formula
U(x) = - Q(x) +lKo(x

(5.8.34) ^'"^
- f)/i(f) di - 2fKo(x
^'^^
- f)/i(f) d^
where Q is t h a t quadratic polynomial chosen so that V C/(0) = V^ ^(0) = 0, a n d / i is t h a t extension o f / t o B^R defined by the conditions , , , (5.8.35) /iW==/W, ^^GR, /l(^^<)=/l(-A^<), 0<r<2i^, x^B^R
/l(^^)=/i[(2^-^)fl,
fâ^(0,1).
Theorem 5.8.1'. Theorem 5-8.1 holds with BR replaced by GR. Proof. As in the proof of Theorem 5.8.1, we find that AQ =^ Q. Also since t / = 0 along ji'' = 0, it follows that Q is also. We need only prove the HOLDER continuity oi\/Û. Let F\[x) and F2[x) denote the first and second integrals, respectively, in (5.8.34). That h/^{V^ Fi, BR) '<Zih(i(f,BR) follows from Theorem 2.6.7, Corollary 1. Obviously,
181
hfji{f, BR) = h^{fi, B2R). Differentiating F2, we obtain ^2,a^yW - - 2 / X o , a ^ y ( ^ - f) [/l (f) - fl(x)] d^ + fl(x) V{x) =
(5.8.36) ^^^
V,a^y{x)
-2jKo{x-^)d^.
Obviously V is harmonic, and if a, say , < r , then (5.8.37) V,4x) = 2lKo(x-i)d$:, V,.^y(x) = 2lKo,^y(x-i)dS'^,
Since V is harmonic, V^vw can be found in terms of the other F,a/3y, and we find from (5.8.36) and (5-8.37) that \S/^F2{x)\ <ZiL J \S - x\f^-'-^ di + Z2LRÎ X^GR, \^ x\-^dS(.^)
<,ZzL[d{x)r-^,
d {x) being the distance of x from d GR. The theorem follows from Theorem 2.6.6(b). Again we define VR and TR by the formulas (5.8.8) and conclude easily as before that the following theorem holds. Theorem 5.8.2'. Theorem 5.8.2 holds with BR replaced by GR. Now it would be desirable to be able to introduce spaces of analytic functions as was done above and prove the analogs of the remaining theorems. However, it turns out to be more convenient to introduce first spaces of functions which are analytic only in (x^, . . ., x^-^), keeping x*" real. In this way, we shall be able to e x t e n d / t o / i and to generalize easily the formulas (5.8.21). We first define our domains GhR and our spaces of functions. Definition 5.8.2'. We define BohR to be t h a t part of B^R for which x" is real, GUR as the part of BO^R where x^ > 0, and G^ as the part of BQ^R where x'^ < 0. The space '^C^(GfiR) consists of a l l / ^ '^C^{GR) which can be extended t o ^ C^{GhR) and to be analytic in (x^, . . ., x^-'^) for x^ GJIR, the norm \\f\\oh being hfi(f, GhR). The space *C^{GfiR) consists of all
^ *C^(GR) which can be extended t o ^ CJ^{GJIR) and t o be analytic in
{x^, . . ., x'^-'^) for x^ GhR with u = 0 for all x^ GhR with x^ = 0, a n d V w(0) = \7û{0) = 0, the norm ||^||2ft being h^^^û, GhR). Theorem 5.8.3'. Suppose (i) that / i ^ C^(BQhR) and is analytic in [x^, . . ., x"-'^) for x^ BohR] (ii) XQ^ GhR] (iii) Si : ^ = | i ( s ) , s ^ BR and 5 * : f = ff(s), s^G^, are admissible with respect to 5i(f;:ro) = KQ {XQ f) / i (f); (iv) that Si and S* satisfy also the conditions (5.8.20) and also the conditions that (5.8.38) ir,(s) = illis) =0 on their respective domains. Then the integrals of ^1 over Si and 5 * exist. If S2 and S2 satisfy their respective conditions, the integrals of ^ i over Si
182
and S2 cire equal as are the integrals of ^ i over 5 * and 5 | . Finally if, for each X ^ GuRy S {%) and S * {x) satisfy all the conditions then the functions F and F * defined by (5.8.39) F(x)=lKo(x-i)fi(^)d^,
S(x)
F^x)=fKo{x-S)fi(^)d^
S*{x)
are analytic in {x^, . . ., x*'~^) for x on GhR and of class C^ in x and F,4X) (5.8 40) = JKO,4X--f)/i(l)^l, Six) ^^""^
S*{x)
F^îx) = / ^ "^ = 1 KoM^--f)/i(l)^f.
\<oc<v.
Proof. The statements concerning F(x) and the integrals over Si and 52 follows from the proof of Theorem 5-8.3, since the condition (5.8.38) guarantees that |J'(s, ^) E^ 0 so that the derivative of ^ i with respect to ^, which does not necessarily exist, is not involved anyway. An entirely similar analysis shows that the integral over 5 * equals that over S^: If we set | * ( s , t)=s
S*{s,t)
+ i[{\ - ^ ) | * (s) +tSUs)]
(1*^(5, t) = s"),
(p{t)=fKo{xo-i)fi(i)d^,
a repetition of the analysis in the proof of Theorem 5.8.3 yields
vl v1 V
OR
V = 1
ER
V = 1
<5=1
The first integral vanishes since |*' = s" and the second vanishes since I * (s, ^) = s on ^ ^ . The analyticity in the case F follows as before and t h a t for F* follows similarly; the formulas (5.8.40) are obtained in the course of t h e proof. Definition 5.8.4'. F o r / $ *C2(GAi?), we define U = Pnif) b y the formula U{x) (5.8.41)
BZR-BR
=-Q(x)+Fi{x)+F^{x) + f Ko{X - I ) / i ( f ) dS
S ix)
Fi(x) =fKo(x-i)/i{$)d$ F2{x) =f-2Ko(x

GZR-GR
- | ) / i ( a dS - 2JKo[x
S* {X)
- | ) / i ( | ) d^
where / i is the extension of / to BohR (J {B2R BR) which is uniquely defined by the formulas in (5.8.35) and Q is as usual.
Definition 5.8.5'. For x = Xi -{- i X2[x\ = 0) in BQJIR, we let ô(^)
be the distance from dBouR (in the 2v \ space) and r[x) be the distance in the i^-plane x^ = const, of x from the intersection of that plane with dBfiR. For x^GfiR, we define d(x) as the distance of x from dGjiR.
183
Lemma 5.8.1'. (a) BQ^R and GUR are convex. (b) r[x) = R \xi\ - |4i/A, do[x) =hr{x)j]l\ + h^ d{x) = min [do{x), x''], 4 = [x\, . . ., xl~^). (c) Any two points zi and Z2 of GUR can he joined by a path in GUR as in Lemma 5.8.1. The same is true of BQJIR. The proof is like t h a t of Lemma 5-8.1. Lemma 5.8.2'. Let S*{x) satisfy the conditions in Theorem 5-8.3' and, for X $ GfiR, define (5.8.42) V*{x) = -2fKo{x-i)d^-2lKo(x-S)d^.
S*{X) G2R-GR
Then F * is analytic in x^for X^GRU,
is harmonic there, and
(5.8.43) I V 3 F * ( ^ ) | < C(v,fA,h)R-K Proof. Choose XQ ^ GhR. For x sufficiently near XQ, V * (x) is given by (5.8.42) with 5 * (x) replaced by S * (XQ) SO that F * is analytic in all x'^ and its derivatives of any order can be found by differentiation under the integral sign. Thus F * is harmonic. For real Xi (5.8.44) V%(xi) == - 2/Xo,(î - i)d^ = 2fKo{xi - i)d^',, oc<v.
Since the right side of (5.8.44) is analytic, (5.8.44) must hold for x in GfiR. Hence V%py(xo) = 2fKojy(xo i)dii from which a bound (5.8.43) follows if {(x, p, y) ^ (v, v, v). But since F * is harmonic, this derivative has a similar bound. Lemma 5.8.3'. Iffi ^ C^ (^0hn) ^ ^ ^ / i ^^ analytic in x^for all X^BQnu, then ^ ^ ^ \fAx^)\<C\(v,ii,h)-[d(x,)-\>^\ Proof. The proof of Lemma 5.8.3 with v replaced by 1 1 demon^ strates (5.8.45) with (5(:vo) replaced by the distance d{xQ) in the plane x^ = xl from the intersection of t h a t plane with d G^R ; obviously d (XQ) > ^ ( ^ o ) , so the result follows. Theorem 5.8.4', PR is a hounded operator from *C2(Gftjj) to "^CKGUR) with hound independent of R, Proof. From the fact t h a t Fi, AFi, and / i are analytic in x^ for x on BofiR and AFi = fi for real x, it follows that this holds on BohR- It is also seen that A F2 = 0 on G^R. Choose xo^BohR. For x = xi + X<ZQ, | ^ I ÎO| < ^, we can write Fn(x)=fKo(xi-s)fi{s
B(xio.r) ^ZR-^R
(5.8.46) Fi{x) = Fii[x)+Fi2{x),

Si(Xo)
+ ix'^^)ds,
Fi2(%) =^fKo{x - S)fi (S)dS +JKO{X - S)fi(^)di.
184
As in the proof of Theorem 5-8.4, F12 is analytic in all x' and
(5.8.47)
I VSFiaWI < ^ i | | / | | o [ ^ W ? " ! .

+ ix2o)ds'^ +fKo{xi s)/i,(s + ix'ôjds.
In the case of Fn, we first choose oc < v and differentiate to obtain Fii,cc{x) = ~jKo{xi~s)fi{s
We can then estimate V^-Fii,a by \V^Fii,4ô)\ < Z2\\fh[d(xo)r-'as was done in the proof of Theorem 5-8.4, using Lemma 5-8.3' this time. Thus all second derivatives iî,a/3 with (oc, ^) ^ [v, v) satisfy the desired Holder condition. B u t then Fi^w does also since AFi / i . In the case of F2, we see t h a t F2[xi + ixm) =f-2Ko(x~
S*{Xo)
I ) / i ( I ) dS - 2fKo(x
G2R-GR
f)/i(f)di
for X = xi -{- i X20 near XQ. Differentiation yields (see Lemma 5-8.2') V3F2(ô) = / i ( ^ o ) V3F*(ô) - ifv^Koixo
8*{xo)
- f) [/i(f) -fi{xo)]di
(5.8.48)
-_2fv^Ko(xo
GZR-GR
- I) [/i(|)
-fi{xo)]d^.
The last two integrals can be estimated as in the proof of Theorem 5-8.1' and t h e first term on the right in (5.8.48) has the estimate 1/1(^0) V3F*(ô) I < ^3||/l|l R^-^ < ^ 3 | | / i l [d(Xo)r-^ using Lemmas 5-8.1' and 5-8.2'. The theorem follows. Theorem 5.8.5'. If H is harmonic on GR and H^"^C^ÎGR) , then ^ *C'^(<^/ii?) OL'y^d the inequality of Theorem 5.8.5 holds. Proof. This follows from Theorem 5.8.5 and the reflection principle.
Theorem 5.8.6'. Suppose U^CI{GR^) with VU(0) ='7Û(0) = 0,
u satisfies (5-8.5) on GRQ, and vanishes along OR, where y) is analytic near (0, 0) and has the properties stated near that equation. Then u is analytic in GR and can be continued analytically across GR. Proof. As in the proof of Theorem 5.8.6, one deduces t h a t there is an J?2 which is such that u^ '^C^{GhR) for every R, 0 <i R -< R^. From Theorem 5-8.6, we conclude that u is analytic in any such GR. NOW, let V be replaced by v + 1 and let x^+^ = y. Then u satisfies on GR an equation of the form
(5.8.49) Uyy = (p(y, X,U,U,a,Uy,Uy,u,Uôcp) (oC, ^ = i , . . ., v)
in which 99 is analytic near the origin; this is obtained b y solving (5.8.5) for Uyy. If we now set
w24a ^ ,, ^ = 1 , . . ., V
185
we see from (5.8.49) that the vector w^, . . .,IL^+^ satisfies a system of the Cauchy-Kovalevsky type
in which the/^' are analytic near the origin. Since the origin could correspond to any point on an analytic part of the boundary, we shall simply prove that the w3 are analytic at the origin. We already know that the w^ are analytic in (x, y) on GR and in x for {x, y) on GhR. Our method is to use the proof of the Cauchy-Kovalevsky theorem to show that a solution can be obtained with our given values q)^ (x, yo) which has a radius of convergence independent of yo, if yo is small enough. For each real yo > 0 but sufficiently small, we make the usual substitutions to prepare to prove the Cauchy-Kovalevsky theorem as it is proved in the cited book by GOURSAT, pp. 26: (5.8.50) *a)^x,r]; yo) == i2)^{x, yo + 7]) (pj{x; yo).
This leads to equations
(5.8.51)
(5.8.52)
-^ =p[^,yo
+ ri> *<^ + 9{^'>yo), *co^ + (px{^,yo)]-
Subtracting off the constant term and making the final change coHx,7]; yo) = *w^x, TJ; yo) a^yo) rj ^^(yo) - f^ [0, yo, (p(0, yo), (fx(0, yo)], we obtain the equations (5.8.53) j ^ = gH^>V>(^^<^x]yo) = -aHyo) +
+ f^ [^, y,co + (p [x, yo) + a (yo) r}> ^x + (fx [x, yo)]. Now the f^ [x, y, w, p) are analytic with
\p{x,y,w,p)\
|^^*(:^,yo)| < ^ ,
< M i for \x\''<^r,

\(pUi^>yo)\<^ if
\y\ <r,
\w^ < ^ \pi\
<r.
Moreover, we know t h a t the <p^ are anah^tic in x and 1^1 < ^ , 0<yo<(5. n)
Thus we see that there is a o" > 0 such that the g^ are analytic m(x,rj,a), for each small yo with \g^{x,y,a),n',yo)\<.Mi if \x'\â, 0 < yo < (T. |?7|<o', |a>|<or,
|7r| < c r ,
Consequently there are numbers M, r > 0 and ^ > 0 such that each g^ is dominated by the function C ^ [1 ^ y-i(7? + Afi + ^ 1- ;v' 4- col H . 1 tt,P)]. u _ M Q'Ênr]
186
The remainder of the proof as given in GOURSAT shows that there is a (3' > 0 such that the co^ {x, r); yo) are analytic with \co^x,rj;yo)\ <. [A^ for |:^| < ^', 1 y | < ^', 0 < yo < ^' 7 where d' and the [x^ are independent of JQ. Thus the w^ obtained as above by solving the differential equations have the same properties and must coincide with those obtained previously, for y > 0. Their analyticity at the origin follows. 5.9. Properties of the extremals; regular cases In this section we complete the proofs of the theorems and remarks made in Chapter 1 concerning the differentiability of the solutions of variational problems and of the weak solutions of the quasi-linear differential equations mentioned in the introduction. We begin by completing the proof of Theorem 1.11.1 which states in the case of the quasi-linear equations (1.10.13) and (1.10.14), where the A^ and Bi satisfy (1.10.7") or (1.10.7'") with ^ > 2, that the derivatives ^ j , and the function U = VÎ^^Hl{D) for each D C C G and the derivatives satisfy the differentiated equations (1.11.9) there. I n Chapter 1, we introduced carefully the difference-quotient procedure and proved that the difference quotients ^h(^) = h~Âz, (5.9.1) fAf,\Vzh\^dx^{Ci Az = z[x -\- h ey) ~ z[x) + C2a-^) j Ah{x)zl{x)dx satisfied the equations (1.11.3) and that for each D C DaC D' CG,
where Ci and C2 are independent of h. Now^, from the existence theory, it follows that each z^y^Ljc with ^ > 2. Consequently, since
1
Zh{x) = J z^y(x + t h ey) dt,

0
it follows from Theorem 3.6.8 that Zh ->z,y in Lfc{D') for each D' C(ZG. Moreover, b y approximating strongly to z in H\[D) hy Zn^^C^ and by using the techniques of Theorem 3.6.8 and Lemma 3-4.2 we conclude that Afi~->A in L]cj{jc-2)(D') if ^ > 2 and A^^ \ ii k = 2. Thus, from the Holder inequality we see t h a t the right side of (5.9-1) tends to its expected limit as A -> 0. Next, since Aji (x) > 1, it follows that zn u, say, 7 in H\{D) for a subsequence of h ->0. But from Theorem 3-4.4 (since D a a strongly Lipschitz domain C D', etc.) and the fact t h a t zji -^py in L]c(D), it follows that u = z^y so z^y py^H\(D). Also, we see t h a t An(x)-Â(x) = F^-2(jv) a.e. and the coefficients af^^îx), hl^^{x), eH ^hip ^hip ^^^ fhi ^^^^ ^'^' ^^^ bouudcdly t o their expected limits as h -^0, Pji {x) -> V {x) a.e. and in L^r {D'). From the uniform boundedness
5.9- Properties of the extremals; regular cases

1 1
187
of the left side of (5.9-1), it follows that Alaliaz{j,

1
weakly in L2 {D) to certain quantities and A^hh ' zji, A\Ph
'i
Alc^^^-zi^^
1*
tend
' el, A^^dn zji,
and A^fl' Pfi tend strongly in L2{D) to their expected limits a.s h-> 0 through a further subsequence. By multiplying the first two by bounded measurable functions and integrating, using the weak and strong convergence already established we see that the first two terms tend weakly
1
1
to their expected limits, li C^Lipc(D), A^S/C and A^C tend strongly to their expected limits so t h a t we may let A - > 0 to obtain (1.11.9) and (1.11.10). Before proceeding further with the general theory, we now discuss the extremals of / {z, G) in the case (1.10.8) with v = k = 2, N arbitrary. We have already seen in Theorem 4.3-1 that any minimizing function z satisfies a condition
(5.9.2)
f\Vz\^dx<L^(rlR)^f',
B(xo,r)
0<r<R,
B(xo,R)cG,
ju > 0
from which it follows that 2 ^ C^ (G). But now, if we consider a domain D (Z D'ad D' C d G, we know that the difference quotient Zh satisfies the equations (1.11.3) in which ^ ^ = 1, the alf^ satisfy (5.2.17), the hhy Cfi, and dfi satisfy (5.2.18), and Ch and//i satisfy (5.4-8) uniformly in h and, as we have already seen above, each Zh ^ HI [D) with norm uniformly bounded. It follows from Theorem 5-4.1, which has been seen in 5-4 to generalize to systems, that the zji satisfy a condition like (5-9-2) uniformly with G replaced by D, so that this is true of each z^y which therefore ^ C^{D). The whole analysis carries over for a solution ^ of (1.10.14) if the A^ and Bf satisfy (1.3.8") with k = v = 2, N arbitrary. Once the z^ and pi^ are seen to ^ C^{G), the equations (1.11.9) assume the form (5.1.1) (with B {u, v), C {u, v), and L {v) given in (5.2.16) with coefficients e, and / Holder continuous. Thus, in the case N = 1, we conclude from 5-5 that z^C^{G) and hence satisfies Euler's equation; then the theorems of 5-5 or 5.6 can be used to obtain the results stated in Chapter 1 that z^Cîi f^ 6% or A^ and Bi^ C^''^ in case # = 2 or if / and fp ^ C^-i or Af^ C^'^ and Bt ^ C^'^ in case n > 2. The cases where iV > 1 are treated in Chapter 6. Clearly ^ ^ C^ if/^ C ^ or A and B ^ C^. That z is analytic if / of ^ and B are analytic follows from the theorems in 5-8 as generalized in 6.7We now prove Theorem (1.11.1') for which proof we need the following lemma due to LADYZENSKAYA and URAL'TSEVA [2]:
Lemma 5.9.1. Suppose the A^ and Bf satisfy (1.10.8") with ^ > r and that z^Hl{G) D Cl{G), 0 < / ^ < 1 and is a solution of (I.IO.14)Then there are numbers Ri and Ci > 0 which depend only on v, m, M, mi,
188
Ml, k, K, fi, and A^ {z, G) such that JV^ (x) |2 (x) dx < C2 ie^/F*-2
B{xo,R) B{xft,R)
\vS\^dx, R)-].
0<R^Ri,
SeHl,[B{xo,
Proof. Suppose, first that | ^Lipc[B(xo,

C^X) = f 2 ( ^ ) [^^(^)
R)], and define

-z^Xo)]
this choice being allowed in (1.10.14) as one sees by an approximation. Using that C in (1.10.14), we obtain * /{|2- [Pi^t + (^' - 4)Bi] + 2i{zi - zi)S,.At}dx
Bixo,R)
= 0.
The hypotheses (1.3.8") (iV = 1) imply that

1
PiAUx, z,p)=piAUx,
z, 0) +pip}JAy(x,
0 1
z, tp)dt
> / ) i ^ J ( ^ . ^, 0) + W i | ^ | 2 | ( 1 + ^ 2 | ^ |2)-l+A:/2^^
0
> W2 F*^ i^2,
^ 2 , K2 =^ (p (v, Wi, M l , i^, ^)
M(;t:, ^,:?{))| < M i F * - i , It follows that

B{xo,R) B{xo,R)
|5(:\;,^,^)| < Mi F*.
/ F * P ^^k; < | { Z i |2 + Z2 Z i^'' F* ^ 2 + Z3 i: Z?'^ F^-21VI |2} dx ^ L = hf, [z, G), Zi = Zi {v, mi, Mi, K, k) from which the lemma follows easily, since li^dx<
Bixo.R)
2-1 R^f\ VI |2 dx < 2-1 Rf^JVf^-^ I VI12 dx

B{xo,R) B{xo,R)
by Poincare's inequality, k being > 2. The lemma follows easily for any 1$ HIQ[B (XQ, R)] by an approximation. The proof of Theorem 1.11.1' starts with the difference quotient procedure which leads to equations (1.11.3') obtained from (1.11.3) by replacing bh, cu, 4 , and/;, Pu by b^ Ph, cn Ph, 4 Qh, and 4 Qh, respectively, where Ph is defined in (1.11.4) along with Uh and Ch and Qh, bh, Ch, 4 , and fh are now defined by
1 1
AhPhbh=fAzdt,
0
AnPhCh
0
=fBpdt,
(5.9.3)
AhQh=j[\ +\z{x) + tAz\^ + \p(x) + tAp\^]J^'^dt,

0 1 1
AhQhdh =fBzdt,
AhQhfh
=JBxdt.
5.9- Properties of the extremals; regular cases
189
From our hypothesis (verified for any extremal if ^ > r and by any minimizing extremal if ^ = '^), it follows that z satisfies (1.10.9) and is therefore ^ C|J {D) for any D C C G^; so we may assume that G is already a subdomain of the domain of z and z^ C'|J(S). So, let us pick B {XQ, R -{- a) C GjiQ, 0 < a < i?, and define r] ^ L ipc[B {XQ, R + a)] with ?^ = 1 on B (XQ, R) and define
C = 7] Z, Z = rj Zh.
Substituting this f in the equations (1.11.3'), we obtain fAh{{VZ + Vrj'Zn) - [ah - (VZ -Vrj - z^) + b^PhZ + rjehPh] +
+ Z ' [cnPn- {V Z -V j Ah\VZ\^dx^j

B{xo.E+a) B(xo.R+a)
7} ' Zh) + dnQh Z -^ U Qh]} dx = 0 {ZiAhQnZ^+Z2,\Vri\Âhzl+ZÂhQh)dx
from which we conclude in the usual manner (since Pf < Qh) that
Zi =
Zi{v,N,mi,Mi,k,K).
Now, we observe that the function z {x -\- h ey) ( z -+ A z) satisfies the hypotheses of Lemma 5-9.1 for the functions * ^ and * 5 defined by *A(x,z,p) = A (x -\- hey,z,p), etc., which obviously satisfy (1.10.8") etc. Thus we may apply that lemma with V{x) replaced by V{x + hey) = 1 + \z + Az\^ + \p +Ap\^. By using the lemma in conjunction with the inequalities
(5.9.4)
< / [ ( 1 + \pi\ +tAp\^]rdt
{Ap=p2-pi, {r=-^
etc.) + kjl or kjl) + a)\), that
< C [ ( 1 + | ^ i P ) r + (i +\p2m we find, ii R-\- a <^Ri{z{x

B{xo.R+a)
+ h ey) also ^ CI[B{XQ,R

B{xo,R+a)
jZiAhQh\Z\^dx<^Z^[R
A-ajf'fAhlVZl^dx.
Thus, a R -{- a <, R2, depending on the same quantities as Ri, we obtain the inequality (1.11.8) with D = B (XQ, R), D' = B {XQ, R + a) and also JAhQhzldx <Z5a-2 R^jAhQhdx.
B(xo,R) B{xo,R+a)
Using this last inequality as well as the others wê see that the rest of th e proof can be carried through as before except that now AlbhPh'Zh,
1 1
A\ch'SIZh.
1 1
AQ\^^ tend strongly to their
converge only weakly in L2 \B (^0, ^ ) ] to their limits and, for each ^^Lipc[B[xQ,R)], limits. AlVC> AlPhC, ^nd AlQlC
190
The proof of Theorem 1.11.1" is deferred to the next section since a new idea is required. We proceed to the proof of Theorem 1.11.2. Here we must assume that N = \ (see Remark below). However we need assume only the hypotheses and conclusions of Theorems 1.11.1, 1.11.1', and 1.11.1". From Lemma 5.3.1 and these results we conclude that we may substitute C = coV~'py, co^LipcD,
1
a)(x)'>0
in the equations (1.11.9) or (1.11.9') and sum on y. Let us consider the case (1.10.7) in which V = {\ + z^ -\- \p\^)^, similarly. Using the relations
VV,oc==pypy,u + ZPc., pypy =
the other case is treated

\ Z^ ,
\ P \^ = V^ -
|VF|2<|Vi^|2 + |^|2, we find that V satisfies f V^-^-'{a),4a-^V,p

D
Coc =-V-'[co,o:py + co{Py,c. +'B^V) + co('C^V,oc
eV-^PyV,a)]
+'DV)}dx<0
where '5 = V-^l-a'^^zp^ 'D =^ -g + ^(F2 \ z^) + e'^PVpy] 1 - ^2) 4_ Vfypy] + gV^ > 0.
+ F - 2 [ - (& + c'^)zPo. + d{V^ -
g and > 0 being chosen so that a''^py,ocpy,p + Ve^ypyôc -sa^^V,o.V,^ It is clear that g may be taken bounded so all the coefficients are bounded and the result (5.3-4) follows with
W =y^-'=U\
?i = 2(k-8)lk,
1
P=1.
In the case (1.10.8), where F = (1 + l^p)^ we obtain (5.3-4) w i t h P = F and somewhat different coefficients B, C, and D. Remark. The reason t h a t we must assume A^ = 1 in this proof is t h a t T if we set C* = co V-^ p\, above we obtain terms like which give co,a F"^ a'^^{VV,p ~ zp^) in the case N = \ but yield nothing tractable ior N '> 1. The results above apply only to interior domains and, except for the cases k = V = 2, to cases where A^ = 1. However, if we have a weak solution of (1.10.14) for the general iV which is known to ^ ^jiiD), at least for Z) C C G, then the equations (1.11.9) have bounded coefficients ^ C^ and the theory of 5.5, as generalized in 6.4 allows us to conclude t h a t the p^ ^ CJ so the z^ $ Q . Then repeated apphcations of the theorems of
5.10. The extremals in the case 1<CA<<2
191
5.6 and 6.3 to Euler's equations allows us to conclude the higher differentiability, analyticity, etc. If A^ = 1, all the higher differentiability results follow if we have a weak solution known to be Lipschitz. Thus, in the case of an integrand of the form/(^), we conclude the differentiability of the Lipschitz solutions found in 4- 2 and announced in 1.7 and 1.10. And, as was stated in 1.10, Ladyzenska and Ural'tseva have showed t h a t any bounded weak solution of (1.10.14) in the case (1.10.8") is differentiable, in the case N = \. But, for the general system (1.10.14) for N arbitrary, it is possible to conclude all the higher differentiability of a weak solution if it is known t o $ C^ (D), D C G. The writer (MORREY [10]) showed this using the space L2,^ of 5.5 (see also NIRENBERG [1] for the two dimensional case). But it is more interesting and the techniques are results are more important using the Lp theory which has been highly
developed for higher order equations by AGMON, AGMON, DOUGLIS, and
NIRENBERG ([1], [2]), BROWDER ([1], [2]) and which has been presented in 5-5 and 5-6 for a single second order equation. We shall present the theory as it applies to higher order equations in Chapter 6. 5.10. The extremals in the case 1 < A: < 2 In this section, we prove Theorem 1.11.1''. The only difficulty in this is due to the fact t h a t it is no longer true that Ah(x) > 1 and the places where A his, near zero do not necessarily occur at the places where z^ is large. So to avoid this difficulty, we introduce a new idea, namely to replace our minimum problem by that of minimizing / {z, G) among those functions for which a certain other integral / (z, G) < some number K. This same idea will be used again in 9-5 in connection with the parametric problem. The writer feels t h a t a further exploitation of this idea could lead to fruitful results. Definition 5.10.1. In this section we define (5-10.1) J(z, G) = (1/2) / I
G
Vz\^dx.
For each z*^Hl(G), we denote by ZQ the harmonic function in G such t h a t ZQ ~ z*^ H\Q[G) and we define KQ = J(ZQ, G). Theorem 5.10.1. Suppose z* ^Hl(G). Then for each K > KQ, there exists a function ZK which minimizes I {z, G) in the family FK of all z such that ^ ^ * ^ H\Q (G) and J [z, G) < K. The vector ZK satisfies the equation
(5.10.2)
/K:a(/*i!>i+/p^) + c/.]rf^ = o, cîo(G).
for some /u = /LC [K) > 0. There exists a sequence {Kfi) _> _[_ 00 such that Kn ' [x (Kn) - ^ 0 . If ZK 1^ ^0, fJ' is unique. Proof. The existence follows immediately from Theorem 1.9.1. That fi is unique if ZK 9^ ZQ is evident. If / {ZK, G) < K, it is obvious that
192
(5.10.2) holds with ju = 0. If K > KQ and / {ZK, G) = K, then ZR ^ ZQ and is not harmonic so that there exists a function Ci^ L ipc (G) such that (5.10.3)
G
/ C l a 4 . a ^ ^ = 1.
It follows t h a t J {ZK + Afi, G) is increasing with L Suppose that C satisfies (5.10.4) and let us define /C:a4.^^ = 0
G
Z{x, A) = ZK(X) + A C(:v) - C A2 Ci W where c is a constant. We easily conclude that J[Z {', A), G] < iC for all A with I A| < Ao if c is sufficiently large and positive. Since / (ZK, G) = K and ZK minimizes / {z, G) as above, we deduce that (5.10.5) l{C:Jpi
G
C'fzi)dxÔ
for each C^ L iPc {G) which satisfies (5-10.4). Now, let C be any vector ^ L i pc (G). We can write (5.10.6) C = C* + ACi where ?^ = f
G
C^z^^^dx.
Then C* satisfies (5.10.4). Accordingly (5.10.5) holds ioi C* = C ^ Ci so that (5.10.2) holds with (5.10.7) f^= - l('^Ci'fp
G
Ci'fz)dx
as one sees by substituting the value of 2. from (5.10.6). Since I (z, G) > I {ZK, G) whenever / (z, G) <. K and since J [ZK + ^ Ci, G) is increasing, we conclude that A-i [/ {ZK - A fi, G) - / (ZK, G ) ] > 0 for 0 < > l < h, (^^^^) limA-i[/ {ZK -?iCi,G)-I (ZK, G)] = fx
using (5.10.7). Finally, let cp [K) = I {ZK, G). Evidently 99 is non-increasing and (p{K+AK) <I(zK + ?iCi>G), AK = J(zK + )iCi,G)-J{zK,G). Since A KjX - > 1 as A - > 0, we find t h a t
(p'{K) = [JL[K) for almost all K.
Then, it follows that fx is summable on [KQ, 00) and ff^{K)dK<I(zo,G) z\ being a minimizing function for / {z, G). -I[zi,G)
5-10. The extremals in the case 1 < A < 2
193
We now prove Theorem 1.11.1". Proof of Theorem 1.11.1". Suppose first that ^*$Ci(r) for some T D G . Then, if Z~ z"^^H\Q[G), there exists a sequence {zn], each Zn being in Ci{F) with [zn Z*)\G^C\ [G) such that Zn->Z in HI(G). Consequently ^{^Ky G) -înf I(Z, G) ioT Z z* ^ HIQ(G) SO that ZK r some z 2.^ K runs through a subsequence of the sequence {Kn\ of Theorem 5'10.1 and z minimizes I{Z, G) among sll Z yZ z*^ HIQ(G). Now for K fixed and equal to one of the Kn above, we replace C ^^ the equation 5-10.2 by C/^ W = ^"-"^ [C(^ ^ ^y) C W] and conclude that the difference quotient ZRU satisfies D' (5.10.9) + C*^ft(6ri^4,.a + Ci,-.4ft ^ni.Pn)]dx = 0. By putting t,^ = rf z\^, we find in the usual way that
Since / (ZR, G) < K, we may let A -> 0 to obtain the limiting equations corresponding to 5-10.9) and the inequality l(f, + F|-2) I VPK P dx^C
D
a-^f(f, VI + V\) dx
D'
(5-10.10)
= 2C3r^K'iLi(K) +Ca-^JY%:dx
{DcD'^).
The right side of (5.10.10) remains bounded as i^ -> 00 in our subsequence, since K fJi{K) -> 0 and / {ZK, G) -> min. Also /I VpK 1^dx = / F | V^Î VpK 1 ^:v ^ D D ^lfV^^dxY-''^ÎV^^\\/pK\^dxY^
(5-10.11)
(h =
k(2~k)l2).
From this it follows that pK-y p in Hl{D), that V^^+^^^pK -7 F-1+^/2^^ in H\[D), that f/ = V^^^^Hl(D), and that we can let X ->00 in the limiting equations [h = 0) (5.10.9) to obtain the equation (1.11.9). Finally, suppose merely that z* ^ HI [G). Let ZQ be a function minimizing / (Z, G) among all Z ^Z z*^ Hl{G). Let {Gn} be a sequence of domains such that Gn C Gn+i for each n and G = \J Gn- For each ^, there exists a sequence of functions {z*^}, each 2:*^,^ ^^(GTI+I), such that ^tp -^ ^0 in HI (G n) as ^ ->oo. For each n and ^, choose a function 2:^^ = lim-s:^^^ as in the first part of this proof. For each n, we can choose a subsequence of the Znp which converges weakly in H\[Gn) to some ZnThis function Zn is minimizing for / on Gn and Zn ZQ(^ H\Q{Gn). For each n, the function Zn = Zn in G^ and ZQ in G Gn is minimizing
194
for / (z, G) and Zn z"^^ HIQ{G). The bounds for the integrals of \Vpn\^, V^~^ \Vpn\^, etc., on each D C G G are uniform, so that there exists a subsequence of the Zn which converges weakly toward a function z of the type sought. 5.11. The theory of Ladyzenskaya and Ural'tseva^
In this section, we assume that ^ is a bounded weak solution of (1.10.14) with A^ = 1 in which the A^" and B satisfy (1.10.8") with k> \. Since we may write
1
poc A^ [x, z,p) =pocA^ {x, z, 0) +pccp^ I Al^ {x, z, t p) dt
it follows easily t h a t the hypotheses (1.10.8") imply the existence of numbers mi and M\ such that pocA^{x,z,p) > mi \p\^. Ml, 1^1 < M i ( l + 1^1*^-1) (5.11.1) \B\<Mi(\ + \p\^). ^^^^^ The most difficult part of this analysis is that of showing that z^ C^ (G) for some // > 0. For t h a t purpose, it is convenient to introduce the following D E GIORGI type space: Definition 5.11.1. z ^"^djciG; L;y; 6) iff (i) z^Hl(G), (ii) \z{x)\ <L on G, (iii) for each sphere B {XQ, r) C G, we have (5.11.2)
B{xo,r/2)
j\\Jz\^dx^yr''-T^, in B (XQ, r), we have + {a r)-^^ msix [z (x) A j H ,
(iv) for each hi for which z(x) hi ^d (5.11.3) / I Vz\^dx<y\A(xQ, hi, r)\l\
0<a<
f\Vz\^dx<y\B(xo,
\,
A2, ^) | | l + (o-r)-^ max [;^2 '2^(^)]^},
(v) for each ^2 such that A2 z {x) < (5 on B (XQ, r), we have
0 < a < 1. Here A {XQ, h,r) = B (XQ, r) D A (h) and B (XQ, h,r) = B [XQ, r) Ci B [h) where A (h) and B [h) are respectively the sets where z [x) ^ h and z [x) < h. Theorem 5.11.1. Suppose z ^')8]c[G', L',y',d), XQ^G, and a is the distance of XQ from dG, Then osc. [z,B(xo, r)] < Co(rlaYo, 0 < r < a, /ô > 0 where CQ and jbio depend only on k, L, y, d, and G.
1 For a complete account of their work, see their new book (LADYZENSKAYA
and URAL'TSEVA [3]).
5.11. The theory of LADYZENSKAYA and URAL'TSEVA
195
It is convenient first to prove the following lemmas: Lemma 5.11.1. Suppose w ^ H\\B[xQy r)]. Then I j\w{x)-w |'/('-i) dxy-^^'" < C{v)j\ Vw {%) I dx
where w denotes the average value of w. Proof. From the homogeneity of the situation, we may assume that y = 1 and w = 0. From Theorem 3-6.5, it follows that [\w[x)\dx
5(0.1)
^ Ci{v)(\ S/w{x)
5(0.1)
\dx.
From Theorem 3-5-4, we conclude t h a t there is a bounded operator @ from H\[B (0,1)] to H\Q \_B (0,2)] such that iiW=^w, then j\VW{x)\dx<^C2J\\Vw{x)\
-B(0.2) 5(0.1)
\w{x)\^dx W{x) ^w{x), x^B(0, 1).
<.Czj\Vw{x)\dx,
5(0.1)
Then, from SOBOLEV'S lemma (Theorem 3.5.3), we conclude that l\w{x)

5(0.1)
\^^(^-^)dx<^f\W{x)
5(0,2)
\^^(^-^)dx
<(C4f\'7W{x)\dx^^^''-^^
\ 5(0,2) /
from which the lemma follows. Lemma 5.11.2 ( D E GIORGI [1]). There is a constant ^ == P(v) such that if u^H\ [G) and B {XQ, r) G G then (A - h)r^-î^{K A; r) < ^ / |
(5.11.4)
\/u(x)\dx K,r)\,
A:^ JB (XO, (XQ,
^(eo.7i.r)-^(a!o.A.r)
h<X,
r(h,X\r)
= mm[\A(xo, r 0, if
yvr" \A(xo, r)
A [XQ,
h,r)\].
Proof. Let us define h, r) A, r) w (x) = I u (x) h, a xÂ h,r)

A {XQ,
[k h, if ^ ^ ^ (XQ, Xy r) Then Vw (x) = 0 unless xÂ (XQ, h,r) A {XQ, A, r) when V w (x) = Vu {x) (almost everywhere). Applying Lemma 5.11.1 and using the definition of r, we obtain T-[(A A wy + wt] <iCJ\Vw(x)\dx}\ t = vl{v 1)
from which the lemma follows easily. Lemma 5.11.3. For each (T$ (0,1), there exists a 6 {a, y, k,v) "> 0 such that if u^S8]c(G, L, y, d), B {XQ, Q) C G, and H = max [u(x) h] ^d (xÂ (XQ, h, Q)), then I ^ (^0, h + H + a^ Q, Q a Q)\ = 0 whenever \ A [x{), h, Q)\ <,d Q^
13*
The corresponding result holds for the sets B {XQ, h, g).
196
Proof. Define Qj = Q (T Q + 2-3 a Q, hj = h + G H + a^Q 2-^ {a H + a^ Q), j = 0, 1, 2, . . . We shall show t h a t | ^ ( ^ o , ^hQi)\ - ^ 0 - We apply the estimate (5.11.3) with Q = Qj and Q ~ a Q = QJÎ to obtain (5.11.5) / I Vu\^dx<y' \A(xo, hj,Qj)\ (1 + {QJ Qui)-^H^y
We now estimate the left side of (5.11-5) using (5.11.4) and the Holder inequality and assuming that d ^ Q^ <: yp (Q a QYJI to obtain / I Vu\^dx^\A \^-^( f\ Vu I dxY" > \A [1-^ / f\Vu\ A \A I \A-A' (5.11.6) > j 5 - ^ | ^ " | i - ^ ( V i - Ay)î .4'I*<!-!/').
[A = A(xo, hj, Qj+i), A' = A{xo, hjî, QJ_Î),
dxY I
A" = A (XQ, hj, QJ);
for \ - k < 0 , | ^ ' ' | > | ^ | , and \A'\ <\B {XQ, QJ+I)\ ~ \A\ since \A (xo, ho, Qo) I = \A {xo, h, Q)\<ê'^'' < \B {XQ, Q - a Q) | / 2 . From (5.11.6) and (5.11-5) and the fact t h a t \A'\ < \A''\, we conclude that l ^ ' l < i 5 y i / ^ i ^ " | i + i / ^ ( V i - h,)-HQJ - QHI)-'[H' (5.11.7) We now prove by induction that (5.11.8) \A(xo,h3',Qj)\ <d'2-^*'^Q\ y = 0, 1, 2, . . . provided t h a t we take
e = min | i Q-^' I B (XQ, Q-GQ)], a^^l{2^''+^ ^ y i / * ) 4 .
4-
{^fj'-
For j = 0, (5.11.8) is just t h e condition \A {XQ, h, Q)\ < 0 ^^ If we assume that (5.11.8) holds for some y, it follows from (5.11.7) and the definition of 6 that it holds for y + 1 Proof of the theorem. We consider the balls Bj ~ B {xo, Qj), Qj = 4~^' ^, y = 0, 1, 2, . . . We denote by coj the oscillation of u on Bj and set
Ci = m a x [2MIQO, 2^+1], (ô = a),
where the constant s will be chosen below. Then we have coo < Ci QoSuppose for some j that we have the opposite inequality
a)j > Ci Qj,
Then (5.11.9)
a);>2^+V;.
197
We shall show that 5 can be chosen so that (5.11.9) implies
(5.11.10)
wj^rja)jî
+ Qj,
^ = i - - 1- < i i 25+2
To t h a t end, we set
(O = CDj_l, Jl = (jLli + JL12)I2 jUi == m a x w (x), fji<z = mmu {x), x^ Bj_i
Dt = A (xo, jLCi ~ 2-* (JO, 2 Qj) A (XQ, fii 2 " * - ! co, 2QJ),
t = 1, . . ., s.
We assume t h a t \A [XQ, ]l, 2QJ) \ < \B(xo, 2 Qj)\l2; if not, we must have
\B{XO,]2,2QJ)\ <\B(xo,2Qj)\l2.
We now apply Lemma 5.11.2 and the Holder inequality to the integral of I V ^ I over Dt to obtain \A(Xo, fii - 2-^-1 CO, 2Qj) |Ml-l/') ( - ^ ) ^ <|SA;|i)^|A:-lJ| û\^dx,
(5.11.11)
t=\,...,s.
^'
We estimate the integral on the right in (5.11.11) using (5.11.3), (5.11.2) in case 2~*co > 6, and t h e facts t h a t \A(xo,]l, 2QJ)\ < \B(xo, 2^;)|/2,
Jt = jui 2-1 0), and u{x) ^ jui on B (XQ, QJ-I). We obtain
l\Vu\^dx<:V
;^ l r M ( ^ o , 1^1 - 2-^CO, 2^;) K ^ + (2^;-)"^(2-co)^], 2-*co<^.
(5.11.12)
Using the fact t h a t | A {XQ, JLII 2"* o), 2^^) | < | ^ {XQ, jui ~ 2-i co, 2^;) ^\B{xo, 2Qj)\j2, (5.11.12) becomes JJ ' ~[Ky ^ p ^ [QI + (2-^-1 co)*], 2-* C < ^, O K = 2"-! y,.
(5.11.13)
But now K[Q^ + (2-^-1 co)^] > 2"-! 7v (dl2)^ so t h a t we m a y write (5.11.13) in the form (5.11.14) f\Vu\f^dx<KiQ^.-^[Q^
Dt
if 2-* C > a O Ki^Ki(y,d, k,v).
+ (2-^-1 co)*],
Taking into account our assumption (5.11.9) and noting that co = coy_i > (Oj > 2^+1 Qj > 2^+1 Qj, we obtain, from (5.11.11) and (5.11.14), I A (XQ, //I - 2-^-1 CO, 2Qj) |A:(i-i/*') (2-^-1 co)^ < / 5 ^ | / ) i i ^ - i 2 i ^ i - ^ p ^ ( 2 - ^ - 1 co)^. In addition, | A [x^^, jui 2-^-i co, 2QJ) \ >\A (XQ, /LII 2--i co, 2^^) so that
(5.11.15) \A{Xo, JUi 2--l C , 2^;) \Mv-l)/vik-l)^(^2Ki^^ O gv-k^l/ik-l) . | / ) ^ | .
198
We sum these inequalities over t from 1 to s to obtain S'\A [xo, iix - 2-s-i CO, 2 q^) \ < (2iî ^^ ^p^)i/(*^-i) | B (x^, 2 QJ) \ (5.11.16) (r = k(v \)lv{k1)). If s is so large that [s-l (2iî |5^) )/. 2". ^J+ (v-fc)/(&-i)]i/r ^ Q g j (2 ^j) (the Qj occur to the same power on both sides) then IA (^0, /î - 2-^-1 a>, 2^,-) I < 0 ( i ) (2^J)
\2/
If, in addition, 2 > Mjd, then 2--i co < 6 and it follows from Lemma 5-11.3 (with h = fj,i~ 2-s-i (o, a= 1/2, Q = 2QJ, H = 2--i co) t h a t IA {xo, ^1 2-s-i CO + 2--2 C + 2-1 ^;, ^;) I -= 0. O This means (since z (x) > jU2 on ^Byî) t h a t coj < co;_i 2-^-2 co;_i + ^j which is the desired inequality (5.11.10). Thus we have shown that, for any j , (5.11.17) coj < Ci Qj or (Oj < rj coj_i + QJ, 0 < ? < 1. y coj < rj coy_i + (Ci + 1) 4~^' ^0. By induction, we find t h a t C , < rj^{wo + (Ci + 1) [(4^)-i + (4^)-2 + + (4 77) W ] } < C2 T^^', O C2 = coo + ( C i + \)l(4r)1) since we must have rj > 3/4. The theorem then follows from the argument in the last part of the proof of Theorem 5-3-3Theorem 5.11.2. Suppose z is a solution o/(1.10.14) on G with N = i, where the A ^ and B satisfy (5.11.1),^'^ which z{x) <,L onG and z^ H^ {G). Then z^ Bjc {G; L;y; d) for some numbers y and d ^ 0 which depend only on V, m, M, K, k, and L. Proof. First we set ^{x) = e^'^'^^rjîx) where rj is defined as usual to equal 1 on 5 (XQ, rfl) and to vanish near dB {XQ, r) and in G 5 {XQ, r) and A is a positive constant. We obtain ê^^[XrjÂ^po, -\- krj^-Â'rj,oc -^ Brj^]dx Using (5.11.1), we conclude t h a t Je^^[Xmirj^\
Bixo,r)
Hence we certainly obtain
= 0.
S/z\^ 2, Mirj^ k Mi\S7rj\ Mirj^(\ + I Vz\^)]dx
- rj^-^ {\ + \ \7z\^-^) < 0.
(5.11.18)
199
Using Young's inequality, (5.11.19) fe^'{[Xmi

B(xo,r)
a^-^b == k-^[k \) sa^ + kMi](r]\'7z\)^ - {X + k)Miri^
k-^s^-^b^, ~ 2Mx\\7rj\^]dx ^0.
we conclude from (5.11.18) and (5.11.19) with e = 1 t h a t
By choosing A = 2 ^ Afi/wi, for instance, we find that

B{xo,rm
since e^'^ is bounded above and below. Next choose h, a with 0 < a < 1, and B [XQ, r) C G, and define ^{x) = rj^(x) co{x), x) (z{x) h, if z(x) > h, a}(x) = i if z(x) < h.
where ^(:\^) 1 on ^{XQ, r ar) and vanishes on and outside dB(AÔ> >') Since a),a(^) = zô:{x) if z{x) > h, we obtain
B{oco,r)
>:f[rj^(mi\
A{xo,h,r}
Va>i^ Mi)
kMico\
V?y |^*~i(| V-^ |^-i + 1) + \)]dx.
(5.11.20)
~rj^coMi(\\/z\^
Now, we choose h so t h a t co (x) <,d = mil4 Mi and use (5.11.19) with = mil4 Ml {k \) and we conclude from (5.11.20) that flVzl^dx^yJli+ia
A{xo,h,rra) A{xo,h,r)
r)-^' co^ {x)] dx
from which property (iv) follows. Property (v) is proved similarly. For what follows it is convenient to note that the hypotheses (1.10.8") imply (5.11.1) which, in turn, imply t h a t (5.11.21) pocA'{x,z,p) ^m2V^ M2, I ^ I < M2 F^-i, \B\<M2V^ where m2 and M2 depend only on m, M, K, and k. Lemma 5.11.4. Suppose that the hypotheses of Theorem 5-11.2 hold with k '>2. Then there are numbers x, 0 <. K <C i, and Ci which depend only on m2, M2, Co, /no, and v, such that I V^ f 2 (x) dx < Ci {rlaô I F^-2
Bixo^r) B{Xi,,r)
\\/^\2dx, a < 1,
0 <r
<Ka,
B (xo, a) C G,
for any bounded ^^Hl^lB {XQ, r)]. Proof. By approximations, it is easy to see that if r < ^ we may set C{x) = [z{x)^z{xo)]mx)
200
in (1.10.14). From this substitution and Theorems 5.11.1 and 5-11.2, it follows t h a t l{^^[A-p^
B(xo,T)
+ {z-
zo)B] + 2f | , a ( ^ - zo)A^}dx
= 0
> / { f 2 [W2 V^ -M2-2C0

JB(a;o.r)
(rjayoM2 V^] - M2 Co (rja)f'oF*-21 y f |2^x
from which the result follows easily, since
fPdx^(r^l2)f\Vi\^dx
B{xo,r) B(Xo,r)
by Poincare's Inequality and V^~^ > 1. As a consequence of this lemma and Theorems 5.11.1 and 5.11.2, we obtain the following useful theorem: Theorem 5.11.3. Suppose the hypotheses of Theorem 5-11.2 hold with k >2. Then if D (Z C G, there are constants RQ > 0, C2 and C3, depending only on v, m, M, K, k, L, and D such that z^ C^{D) with hfîz, D) < C2, fV^^^dx
B{xo,r)
< CsRf'oJv^-^
B{xo,R)
\V^\^dx
R< Ro
for each bounded ^^H\Q[B[XQ,
R)], B[XQ, R) (Z D.
We can now prove the analog of Theorem 1.11.1. Theorem 5.11.4. Suppose that the hypotheses of Theorem 5.11.2 hold with N = 1, k ^ 2, and suppose D (Z G G. Then the py and the function U = V^'^^HliD) and they satisfy (1.11.90 IV^'-^{C,4^''^py,P + h^Vpy + e-yV) + =0
V^dx.
+ ^V{c<-py,a + dVPy+fyV)]dx
(1.11,10) / F ^ - 2 | ^p\''-dx < C{v, m, M, K, k, L, D) f
In addition V^+^^Li{D). Proof. We let Z)' be a domain such that D C C D' C C G and apply the difference quotient procedure. This leads to the/^-equations (1.11.3') ((1.11.3) modified as described in 5.9 in the proof of Theorem 1.11.1'); for 0 <i \h\ <C ho, Zfi is Holder continuous. Now, let us suppose t h a t R + a < 2ô ( - -yD'), suppose B^â ^ B (xo, R + a) C D', let 7 = 1 on BR, etc., and set ^ C =^r)Zh, Zh = 7] Zh
in the equations (1.11.3'). The result is (see the notation of 5-9). JAn{{VZn
Bn+a
+ V?? zn) [an (VZ/, - V^ zn) + hnPhZh + r} enPn] +
+ Z[cnP [VZn - Vf]z) + dnQuZu + / A Qh'\]dx = 0
201
from which we easily conclude as usual (since P | < Qn) (5.11.22) lAu\VZh\^dx<lAh{ZiQnZl + Z27]^Qh + Z3\Vri\^zl)dx.
Now, by using the fact t h a t z (x -\- h ey) is a solution of (1.10.14) for the functions A'{x + h ey, z, p) and B [x + h ey, z, p) and using the inequalities (5.9.4), we see as in 5.9 t h a t JAnQnZldx<^Z^j[V^x)
BR-i-a BR+Q.
+ V^x
hey)-\Zl{x)dx \^dx
< Z4 Cs{R + aYoJlV^-îx)

BR+Q
+ V^-^{x + hey)] \ VZ^ fdx.
<Z5{R
+ ayojAnl
BR+U
"^Zn
So, if we assume Zi Zs(R + )''o < 1/2, we obtain (5.11.23)

BR+U
fAn\vZh\^dx<,Z6jAh(ri^Qk+\Vr]\2z^)dx
BR+U
(5.11.24)
lAnQkZldx<Z^lA{rj^Qh
BR+a BR+a
\Vri\^zl)dx.
We may now let A -> 0 in these inequalities and in (1.11.3') to obtain the results. Next we prove a preparatory lemma like Lemma 5.3-3Lemma 5.1L2. Suppose the hypotheses and conclusion of Theorem 5.11.4 hold and suppose w := V = F ^ 7 ^ ^ L2{D) for some r > 1. Then w^Hl{A) for each A CC D and if B {XQ, R + a) G D, J\ Vw\^dx
B{xo,R)
< C^râ-^ fw^dx,

B{xo,B+a)
0 < a < R,
Q= i + 2/^î,
C4 = C4 (v, m, M, K, k, D);
C4 does not depend on r. Proof. It follows from Theorem 5-11-4 and Lemma 5-3-1 with ip = 17-1+^/2 and co = rj^ ^LPV ^^lat we may set in the equations (1.11.9')> VL being the truncated function for V (see (5.3.1)). Since VL,OC = 0 when V > VL, we see that C,.=-Vi(r]^Py,o.
yV,oc=pypy,cc
+ r]^V-^PyVL,.+
| V F | <
2rjrj,.Py)
| Vj^ |
Using these results we obtain

| F ^ - 2 ] 7 ^ ^ 2 ( | vF|2
BR+U
;.| V F L | 2 ) ^ : ^
< / ( Z i F^+2 F i ?^2 _j_ 721 v ^ |2 F * Vi) dx.

BR+U
202
Now let
COL = > F F i / 2 , ^ WL^^n F*^/2 17A/2
V OJL = Vi^ (>y V F + F Vi^ + A ^ V F L / 2 )

Vz^^L = F-i+^'2 VJ!^[(kr]VVl2] + (A^ V F L / 2 ) + F Vry].
Then, since ^ r = (A + ^)/^, we conclude that 1^2 vfc+2 v]^dx = jF* coi dx
BR+U BR+U
< C3(ie + )''o J { Z 3 F * - 2 F i r ? H ( | V F j 2 + Aj V F i | 2 ) +

BR+O.
< C3ZiZ3T(i^ + ay^ jjf + C3{R + a)fol(Z2Z3r

BR+a
V^^^Vidx +
Z4)\Vri\^V'cVidx.
Let us now define a b y 2C3 Zi Z3 T ^''o = 1, a < 2?o, replace a b y a/2, set i^ = a/2, assume Bad G and suppose a < ^ . Then f / \VwL\^dx<,Z^x ^\ Vrjl^V^V^dx.
We may now let L - ^ 00 and use the definition of rj to conclude f I \/w |2dx <Z5Ta-^
B{xo,a/2)
J w^ dx
B{xo,a)
which is exactly (5.3-16). The remainder of the proof is the same as that of the proof of Lemma 5.3-3; the different result for Q here is due to the different definition of oc. Theorem 5.11.5. Suppose the hypotheses and conclusions of Theorem 5.11.4 hold. Then U is bounded on each A G G D. Proof. The proof is exactly the same as that of Theorem 5-3-L Remarks. The higher differentiability results now follow as in the other cases. But we have assumed above that ^ > 2. LADYZENSKAYA and URAL'TSEVA have extended Lemma 5-11.4 to the cases 1 < A < 2; it follows that Theorem 5-11-3 extends to those cases also. So the proof of Theorem 5-11-4 goes through as far as the equations (5.11.2 3) and (5.11.24)However, the writer has not found a way in these cases to show that
J\ VrjlÂhzldx
BR+a
is uniformly bounded. Since the authors above {L and U) gave no hint as to why this is so, they may have overlooked this point. This is just the difficulty which caused the writer to resort to the argument in 5-10 which does not apply in this case since we have assumed z bounded and
5.12. A class of non-linear equations
203
would have to show that the ZK were uniformly bounded. However, once the results of Theorem 5.11-4 are shown t o hold, then Theorem 5.11.5 will follow also. 5.12. A class of non-linear equations In this section we present essentially the treatment of LERAY and
L I O N S of a theory of non-linear equations initiated b y M I N T Y , BROWDER
[3], and VisiK. Each of these latter three authors has written several papers on this subject; we have quoted a recent one of each author. The equations which can be solved by this method are in some respects more general than those discussed in the preceding three sections, since higher order equations and systems can be handled but not all of those previously discussed, namely those satisfying t h e conditions (1.10.8"), can be handled. is a mapping from Rm into itself Lemma 5.12.1. Suppose u=f{x) such that (5.12.1) lim|;^|-i^-/(^) = + oo. Then the range of f is the whole of RmProof. Let UQ ^ Rm and define /*(:v) = f(x) -~ UQ. Then / * satisfies (5.12.1). Consequently it is sufficient to prove that the range of any map satisfying (5.12.1) contains the origin. Using (5.12.1) we see that we may choose R large enough so t h a t (5.12.2) \x\-^X'f{x)^\ for \x\=R. But from (5.12.2), it follows t h a t the mapping
w=f{RmfiR^)i
ifi = i
from dB {0,\) into itself is homotopic to the identity. I t follows that there is a solution x in B (0, R) of / {x) = 0. We begin with the statement and proof of the abstract theorem. Theorem 5.12.1. Suppose that V is a separable, reflexive, Banach space, V is its dual (the space of linear functionals), and A is an {possibly non-linear) operator from the whole of V into V which is continuous in the weak topology of V on finite dimensional subspaces of V (i.e. if Xn ->xo in a finite dimensional subspace of V, then A Xn r {weak convergence) A XQ in V). We suppose further that (see notation below) (5.12.3)
(A (w) v)
I ,[ - > + 00
as II7; II ^ + C 3 X
{coerciveness).
Finally we assume the existence of an operator % from F X V to V which has the following properties: (i) 51 (w, ^') is defined everywhere and carries bounded sets into bounded sets and (5.12.4) %{u,u) = A{u).
204
(ii) 51 {u, v) has the continuity properties of A as a function of v and (5.12.5)

(5.12.6)
{^{u,u)
%{u,v),u
v) > 0 for all {u, v) (monotonicity).

->0
(iii) If Un 7 u in V and
(^(^71, Un) ^ ( w , u), Un u)
then 51 (un, v) 7 ^ {u, v) for each v^V. (iv) If Un-7 u and %{un, v) ^v"^, then l^{un, v), Un) -> (t^*, u). Then the range of A is the whole of V\ Notation, li v ^V and v' ^ V, then {v\ v) means v'{v). Proof. Let {wj\ be a basis for V and let F ^ be the space spanned by 2^1, ^w- The equation > (5.12.7) {A [u), v) = (/, v), fi = Ai[xi, . . ., Xm), u,v^Vm t = i, . . .,m is reduced to a mapping and the condition (5.I2.3) is reduced to (5.12.1) by the substitution m m m u=2^xj Wj, v=2yj ^'j> (/> ^) = 2P yjy etc. Thus, for each m, there is a solution Um oi (SA2.7). Since
(A [Um), Um) = (/, Um) < | | / | 1 * ' || Um \\
it follows from (5-12.3) that \\uni\\<M and hence that 1^ (um) ||* < M * (using (i)) for some M and M * and all m. We may therefore extract a subsequence {^7^} such that (5.12.8) (5-12.9) ^^-^-^^ in V, A[un) 7 %vciV\ x=f ^(un, u) -rip in V. Since (5-12.7) holds with u = Un for all v in Vn, it follows that since (/, ?;) = {A (un), v) -^(x, ^) for ^ ^ F i which is dense in V. We next note that
lim(5l (Un, Un) 51 {Un, u),Un u) = lim{A (Un), Un)
(5.12,10)
\im.[A{un), u) lim(5l(^w, u), Un) + Vnn{%[un, u), u) = lim(/, Un) (;^, u) {ip, u) + {ip, u) = 0
on account of (5-12.8), (5.12.9), and hypothesis (iv). Now, from this result and (iii) it follows t h a t 51 {un, t") 51 {u, v) for each 7 Then as in (5.12.10), we conclude t h a t
\im{^ {Un, Un) "^{Un, v), Un v) = lim(A (Un), Un) ~
v^V.
\im{A (un), v) lim(^ (un, v), Un) + lim{'tK(un, v), v)
= {x> ^) ix> ^) (^(^' ^)> ^) + ( ^ K ^)> ^)-
20*5
Thus (5.12.11) {x %[u,v),u v)>0 for all t;. w^V, If, in (5.12.11), vêsetv = u ^w, w^ V, i > 0, we obtain
(X ^(u, u S '^), w) > 0 for all
By letting | - ^ 0+ and using the continuity along lines we obtain (X 91 (^^ ^)> ^ ) > 0 for all z^ ^ F e from which it follows at once t h a t
A(u) =
S^{u,u)=x-f'
We now present the existence theorem of VisiK essentially as extended b y L E R A Y and LIONS. We have allowed ^ to be a vector. The
theorem concerns equations of the form (5.12.12) (Au, w) ^ [ 2 Zî{^> <5^> D'û)D^w^dx = (/, w)
where u ^= [u^, . . ., u^) and w = {w'^, . . .y w^) range over a reflexive Banach space F F i x X F^v in which
and, as in Theorem 5-12.1, (/, w) means / (?x^),/being a linear functional on F ; here d u denotes the tensor {D^ u^} where 0 < | {x | <.nii 1 and D^ u denotes the tensor {D^" u^] with | ^ | = mi. I t can be shown, as in 1.3 and 1.10 t h a t if the vector u minimizes the integral J F (x,d u, D^ u) dx
G
among all vectors ^ in a space such as F , then u satisfies an equation of the form (5.12.12) where Af(x, d u, D^ u) == dFldpi (Pi = Z) u^).
Thus the existence theorem is an alternative method to that of the calculus of variations. We shall first prove the theorem under somewhat simplified assumptions concerning the A^; we shall indicate later how our simplified assumptions may be relaxed somewhat. We shall assume that G is strongly Lipschitz (see 3.4) unless all the F^ = Hîîfi) in which case we shall allow G to be merely bounded. Concerning the A^ ê shall assume. (a) Each Al is measurable in [x, | , rj) and continuous in (f, rj) for almost all x on G; here f = {f^} where | ^ | <,mi 1 and YJ = {rjl} where | ^ | = mf. (b) We assume t h a t for all {x, f, rj), we have
(5.12.13) \At(x,^,rj)\<Cfi
+2'f 2'|f^l^-'+2'I^M^"'ll-
206
(c) We assume a strong monotonicity condition:

i |a|=mi
(5.12.14)
>Q[\ri2-riilR^S,x)
if
\^\^R,\rii\^S,x^G
where Q{r,R,S, x) > 0 whenever r > 0 and is non-decreasing in r. Theorem 5.12.2. We assume that G and the Af satisfy the conditions above and we assume that the operator A defined in (5.12.12) satisfies the coerciveness condition (5.I2.3). Then the equation (5.12.12) has a solution u^V for each f in V\ Proof. Let us define the functional '^{u, v) by (^(u, v), w) =[21 (5.2.15) +ZAi[x,
\oc\<mi
2Af[x, du{x),
du(x), D^v(x)]D^wi(x) Dû{x)]D'w^{x)\dx.

j
I t is clear that the conditions on the A^ and G ensure that % {u, v) is defined for all (u, v), 31 carries bounded sets in F X F into bounded sets in V, and iiun-û and Vn ^v ina, manifold of finite dimensionality in V X V, then ^ {un, %) -7 91 (u, v). Clearly 51 {u, u) =^ A {u). The condition (5.12.5) follows immediately from the condition (5.12.14)- We need only to verify conditions (iii) and (iv) of Theorem 5.12.1; our theorem will then follow from t h a t theorem. To prove (iii), let us suppose that Un7 u and that (5.12.6) holds. Let V and w^V and let {r} be any subsequence of {n}. Since D"^ u\ -> Z)** W^ in Lp (G) if 1^1 < mi (Rellich's theorem, Theorem 3.4-4), it follows t h a t we may choose a subsequence {s} of [r] such that dus[x) ->du (x) for almost every x. Thus, for such x, d Ug {x) is bounded with respect to s. Thus, the condition (5-12.6) together with (5.12.14) implies t h a t
lim jQ[\D'^{US u)\, R(x), \Dû{x)\, x] dx = 0
R (x) being a common bound for | d Us (x) \ and | (5 ^ (:v) |. It follows that we may extract a further subsequence {t} such t h a t we also have D^^ ui(x) -^D'u^{x) almost everywhere if \(x\ = mi. Thus, if we define {At[x,dun{x),Dmv[x)l if|^|=w^ (5- . ) J^^^^^ -\Al[x,dun{x),DÛn{x)-], ii\oc\<mi / f being defined similarly, then we see that f^;^ {x) / J (x) -^ 0 almost everywhere. From (5.I2.I3) we conclude that (5.12.17) \Af{xî,rj)\2>' < C{\ + | | | P + 1^1^}, f=pliP - 1). From (5.12.17) and the uniform boundedness of \\un\\, it follows t h a t ll/wilU' ^^^ uniformly bounded. Also, since t; is a fixed vector and d ut{x) ->du{x) in Lp{G) we see t h a t the set functions
207
are uniformly absolutely continuous. Thus/^'^ -^ ft ^^ ^v (^) if 1^1 'îSince |Z)* w'^\ ^ Lq(G) for some q '> p ii \oc\ < mi, it follows that {SE(ut, v) 51K v), w) ->0, w^V.
Since {r} was any subsequence the result follows. To prove (iv), we assume t h a t v^V, Un -i^ u in V, and ^ {un, v) 7 v^ in V. Again, let {r} be any subsequence of \n]. Since #^ ^ it follows 7 as above that the norms f/^^Hj?/ are uniformly bounded. Thus we may extract a further subsequence {s} of {f} such that b u^-^b umL^(G), 6 Us{x) -^d u {x) almost everywhere, and //^ some functions /* in 7 Lp, {G). As in the preceding paragraph it follows that/g\ - ^ / f in Lp, (G) if 1^1 = m^ where / f has its previous significance if | ^ | = mf. Thus v^ is given by
From the strong convergence of /*^ to / f and the weak convergence of D"^ u\ to 0" u^ ii $x\ = mi and from the weak convergence of//'^ t o / f and the strong convergence of D^ u^ to D"^ u^ when | ^ ] < w^-, the desired result in (iv) holds for the subsequence [s]. But since {r) was arbitrary, the desired result follows for the whole sequence {%]. Remark 1. The reader can easily verify t h a t the theorem still holds if the condition (5.12.13) on the ^^ is replaced by
M?(^,f,^)|<cfi^(%)+21
2'|f^^h^''^^+2'h^M> 1^1 = mi,
(5.12.18)
r-HJ, p)>{pt{h P) <[P -
l ) - i [ 1 - v-Hmj '^)+ v-^p(mi ~\oc$
\^\)pi
^"(^'^ ^) > . ( / - " i f ? ; 7 J - V i ) - ^(^'^)' '^^^^^' '^^'^^ ^ ' '

K^Lp,, L^Lqîôc), q{i, oc) > 1, q~'^{h oi) > \ p-'^ + v-'^[mi I ^ I). The proof is essentially the same but makes stronger use of Sobolev's lemma. It is instructive to specialize to the case where all the nti == \ and the ^ f are differentiable with respect to all arguments. Then the condition (5.12.13) is implied by the first inequality in (1.10.7")- The last inequality
208
in (1.10.7") implies t h a t
1
(5.12.19)
= 2 " (VL - rjU) ivy - VU) I Ay^ [x, i, rn + t(rj2 - f]i)] dt

1
0
If ^ > 2, the condition (5.12.14) holds w i t h ^ = mi r^. In case 1 r,
\i\<.R,
\rji\<S,
0
If Q{r, R, S) = 0 for some R, S, and r > 0, H {f^}, {rjin}, and {s^} such that \in\ < ^ , lîTil < 5 " , and \sn\ > r and / (s^, 1^^, "^yiî) - > 0 . If, for some subseequence, \sn\ remains bounded, a compacness argument shows that / -TÔ. Thus we must have | s ^ | - ^ o o . But
[1 + \Sn\^ + \rjln + tSn\^] < 1 + i^2 + 2 S 2 + 2 | 5 ^ | 2 .
Consequently we see t h a t / (%, ^n, îw) -> + oo, a contradiction. Thus (5.12.14) holds in this case also. Clearly this monotonicity condition, then, is implied by the very strong elhpticity of the system as assumed in (1.10.7"). Using (5.12.19) with rj2 = rj and 7^1 = 0 and using the facts that
tP-2 (1 + | | | 2 _!_ | ^ | 2 ) - l + p / 2 < (1 + | | | 2 + ^2 |^|2)-l+2>/2
we conclude that
i lal=l
(5.12.20)
^m2\r]\^VP-^
-Mi\rj\
(1 + \^\2)ip-i)i2
_ Mi\i\
-VP-^
> m 3 F ^ - M 3 ( l + |f|2)2>/2,
V =(\
+ | | | 2 + |7^|2)l/2.
In case all the Vt = H'^'oiG) (Dirichlet problem) then (5.12.20) guarantees the coerciveness condition (5.I2.3) provided the diameter ofG is sufficiently small (Poincare's inequality). The conditions (1.10.7'") and (1.10.8") are sufficient for the theorem provided that p '> v and the coerciveness condition holds. But, in these latter cases the coerciveness condition is more of a restriction than it is in the case of assumptions (1.10.7")- The coerciveness condition corresponds roughly to the condition t h a t f{x, z,p) >mVP K for all {x, z, p).
6.1. Introduction
209
Chapter 6
Regularity theorems for the solutions of general elliptic systems and boundary value problems
6.1. Introduction In the preceding chapter, we developed rather completely the regularity properties of weak and strong solutions of certain first order variational problems and second order differential equations involving only one unknown function. Complete results concerning the solutions of the corresponding problems involving several unknown functions were obtained only in t h e case where v = 2; these results were obtained b y the writer before the war and were described in Chapter 1. In the preceding chapter some first differentiability results were obtained for certain systems of equations b u t these results did not imply the continuity of the first derivatives. In 1952, the writer presented a paper ( M O R R E Y [ 1 0 ] ) at t h e Arden House Conference on Partial Differential Equations, in which it was proved that a n y vector solution of class C^ of a regular variational problem of class C^, 7i > 2, was also of class CJJ. These results still leave a gap in the theory for systems, which can only be filled b y an extension of the D E GIORGI-NASH results, developed in 5-3, to systems or some entirely new device. B u t , in proving the CJJ differentiability results for systems of equations, t h e writer was forced to use the very important formulas of F . J O H N ([1], [3]) for fundamental solutions and to use methods which are appropriate for the discussion of elliptic systems of higher order. Hence the inclusion of this chapter in this book. Our presentation will be a modification of that to be found in the two
i m p o r t a n t papers of AGMON, DOUGLIS, and N I R E N B E R G ([1] and [2])
although some ideas essentially due to BROWDER ([1], [2]) have been used. A brief b u t rather comprehensive treatment of the topics in this chapter is to be found in the last chapter of the recent book b y HORMANDER [1]; he includes some topics such as existence theorems on t h e ''index'' of a problem, etc., not included here. Many people, particularly among the Russians, have contributed t o the development of this theory and have proved supplementary results not stated here. Many additional papers are referred t o in the papers to which we refer in this chapter. The theorems proved in this chapter will enable us to conclude the differentiability of the solutions of the higher dimensional PLATEAU problem discussed in Chapter 10. They will suggest differentiability hypotheses on the functions Af(x,^,rj) in 5.12 which, together with an a priori assumption that each u^ ^C^j, will imply further differentiability properties of the uK
210
Regularity theorems for general boundary value problems
We shall begin b y studying linear systems of the form (6.1.1) Liu{x,D)u^{x)=fj[x), j=\,...,N, x^G, and shall be concerned also with the solutions of (6.1.1) subject to boundary conditions of the form (6.1.2) Brjc(x, D) u^{x) = gr{x), r=\,...,m, xondG, where m is determined below. I t is assumed that there are integers Si, . . ., sjsf and h, . . .,t]s[ such that each operator Lj^ is of order Sj + t]c\ since we may add an integer r to all the tj if we subtract it from all the Sj, we may as well assume that maxsy = 0. We allow complex-valued functions throughout and allow the coefficients in the operators to be complex. We require the system (6.1.1) to be elliptic in the sense of the following definition: Definition 6.1.1. The system (6.1.1) is said to be elliptic if and only if the determinant L [x, X) of the characteristic polynomials L'^^ {x, A) is not zero for any real non-zero A; here L'^jî^' ^) ^^ ^^^ principal part of the operator Ljj^, i.e. the part of order exactly Sj + tjc. It is to be noted that L is a homogeneous polynomial of degree
(6.1.3)
P-Z[s^
+ h)-
It is assumed that if 5; + Â; < 0 for some (;, k), then Ljic ^ 0. From the ellipticity condition, it follows that for each k one of the operators Zj.jj. ^ 0, so that, for each k, max(sy + t]c) = fe > 0. Systems of this sort were introduced b y DOUGLIS and NIRENBERG. They showed that if P = 0 in (6.1.3) then the equations (6.1.1) m a y be solved for the u^ in terms of t h e / ; and their derivatives. In fact, with our normalization, vndCKSj = 0, it turns out that if several fe = 0, then the corresponding u^ may be found in terms of the remaining u^, fj, and their derivatives. This is easy to see as follows: (1) By relabelling the u^, we may assume that ifi < ^2 < * * ' < ^JV- (2) B y reordering the equations, we may assume si < S2 < * < Sisr = 0. If ti ^ t2 ^ ' ' ' =' tr = 0 <C < tr+i < < îv, we must have Siv_e+i = - - - = s^ = 0, where Q ^ r, and the ellipticity guarantees that the matrix of functions Ljjc(x) with
N Q + 1 < y < i V and 1 < ^ < r has rank r. DOUGLIS and N I R E N -
BERG then showed that if these u^ are eliminated from the remaining equations, the reduced system is still elliptic. So it is sufficient to consider reduced systems and we restrict ourselves to such systems; we present such results in 6.2. To sum up, we assume (6.1.4) maxsy==0, tjc'>i, k = i, . . .,N, P > 0. In order to study boundary value problems, we impose the following conditions of proper ellipticity (sometimes called the root condition) and uniform ellipticity:
6.1. Introduction
211
Definition 6.1.2. The system (6.1.1) is said to he properly elliptic or to satisfy the root condition or to satisfy the supplementary condition if and only if P is even, say P = 2 m, and, for each pair | and | ' of Hnearly independent I'-vectors the equation (6.1.5) L{x,S + zn = 0
has m roots with positive imaginary part and m with negative imaginary part. We say that the system is uniformly elliptic on a set 5 if and only if there is a number M such that \L(x,K)\->.M-^>0, |;i|-1, x^S (A real)
and the absolute values of all the coefficients in the operators are < M for X on 5. Lemma 6.1.1. (a) The conditions of ellipticity and proper ellipticity are invariant under transformations of the independent variables. (b) / / r > 2, any elliptic system (6.1.1) is properly elliptic. Proof, (a) This can be easily verified by the reader. (b) Suppose, for some ô, lo, and f^, that 2:0 is a root of (6.1.5) corresponding to the pair (fo, fi), i-e. L {XQ, fo + -^Ô fo) = ^- Then obviously, 2:0 is a root corresponding to (fo, fo)- Since, if r > 2, the pair (|o, IQ) can be deformed continuously into (fo, fo)> always keeping the vectors linearly independent, and since no root of (6.1.5) is real, and since the roots of (6.1.5) [XQ fixed) vary continuously with (f, | ' ) , it follows that, for any fixed x^, f, f, the number of roots of (6.1.5) with positive imaginary part equals the number with negative imaginary part. The result (b) is not necessarily true if r = 2 as is seen by replacing the system (6.1.1) by a single operator Z> where ox by \hx by)
The reader can easily verify that L2U = 0 for any u of the form u = {\ x'^ y2) f{^z), z = X -\- y i
where / is any holomorphic function. Thus the DIRICULET problem for the operatorL2 fails very seriously to have a unique solution. Another example of this sort was given by BICADZE. The operators L\ and L2 are not properly elliptic. An example is given in 6.5 of a properly elliptic operator L for which the Dirichlet boundary condition is complementing (see Definition 6.1.3 below) but which is such that every complex 1 is an eigenvalue forL. The following system is seen to be properly elliptic and is even strongly elliptic as defined in 6.5: (6.1.6)
,^ , ^ v '^xx %2/ + '^xyyy =/ 14*
212
The determinant L(A, //) (A^ = A, A^ = ^) is
I -XfJi^
(^2-/^2)3 I
-(A2-/^2)4_|_^2^6.
We now consider the boundary conditions (6.1.2); we shall assume t h a t the portion of the boundary dG on which we consider these conditions is sufficiently smooth, the degree of smoothness being specified in each case being considered. First of all, the m in (6.1.2) will be precisely P/2. We require also that there are integers hi, . . ., hm, some of which may be negative, such that the order of Brjc is <:tjc - hr', we let ho be the largest of 0 and the hr. If tjc hr<0, we assume t h a t Brk^ 0; otherwise we let B'^j^ denote the principal part of Brk- At any point XQ of ^G, we let ^ denote the unit normal at XQ and | any (real) vector tangent to (9G at XQ. Let z^ (XQ, | ) , S = \,.. .,m,he the roots of L(XQ, ^ + zn) =0 with positive imaginary p a r t ; these exist since we assume that the system (6.1.1) is properly elliptic. Define
(6.1.7)
L^ {xo. i;A = n [ z - z^ {xo, m
vt
and let \\D^(xo, ^ + z n)\\ be the matrix adjoint to \\Ljjc(xo, f + zn)\\. Definition 6.1.3. For any XQ^G and any real vector | tangent to ^ G at XQ, let us regard L^ (XQ, f; z) and the elements of the matrix I Qrk {^0, l i -^r) I as polynomials in z, where I I Qrk(xo, I ; ^) = 2 ; Bl^{xo, ^ + zn) D^{xo. ^ + zn).
The system (6.1.2) of boundary operators is said to satisfy the complementing condition {with respect to the system (6.1.1)) if and only if the rows of the Q matrix are linearly independent modulo L^ [XQ, S; Z), that is, the polynomial
r=l
2* ^r Qrkixo, i;z)^0
(mod L+)
only if the Cr are all 0. In the system (6.1.1) the maximum order of any derivative of a particular u^ which occurs is tjc. It is to be noticed t h a t some (or all) of the hr are allowed to be negative, in which case derivatives of u^ of order higher than fe may occur among the boundary conditions. As an example, we might have Lu~Au= Bu = aDlû [Dlr + Dl. + D|s) u, I = (fi, h, 0), + ^ = (0, 0, 1), + {hiDaî + h^DaP' + hz)Dl.u
+ (Cll ^ l i + 2Ci2 Daî D^2 + C22 Dl. + ICizDa^l + 2C23Da^2 + C33) X
6,1. Introduction
213
P3 being a cubic polynomial. Then B{^ + zn)=az^-^{hh^ b^ h + 63) z^ + (cii If H )z +
and the complementing condition is merely that B{^ -\- z n) not be divisible by L^ (considered as polynomials in z) for any f 9^ 0. Notations. In this chapter, we shall obtain estimates for the solutions of various problems which involve constants which do not depend on the particular solution. We shall often denote such constants simply by C b u t C s occurring in different places may denote different constants. On the other hand, we shall sometimes distinguish different constants by using subscripts; however, we do not guarantee that C2, for example, always denotes the same constant. We shall frequently use Zi, Z2, . . . to denote similar constants. Our notation here is not consistent, but the situation is usually clear from the context. Estimates for solutions of boundary value problems in a half-space where the operators Ljjc and Brk have constant coefficients and coincide with their principal parts will involve constants which depend only on a common bound for these coefficients, for M {Dei. 6.1.2), and for J - i where / = inflmaxig^'l}, lf|==1, i ' n = 0,
the Q'^ being the various m x m determinants in the matrix (J of (6.3.19), below, which is related to the ^-matrix above. Such constants will be said to depend only on E; interior estimates will not depend on / . Constants which depend also on bounds and/or moduli of continuity of all the coefficients and possibly on the domain G will be said to depend only on E and E'. If a constant depends on other quantities such as h, q, /LC, etc., we may write C = C{h, q, /bt), etc. In 6.3, we shall replace v by V -\- \, let X = (%i, . . ., x^), and y = x^+'^; we then often let X = {x, y). We say that a vector function has support in GR U CTR if and only if it vanishes outside GR and on and near ^R; it need not vanish on GR. If a = {oci, . . ., { v is a multi-index, we define X) oc\ = (ail) (^2?) . . io(v\), x^" = (x^)""! . . . (x^'^v, etc. Definition 6.1.4. A function or vector function 99 is said to be essentially homogeneous of degree s if and only if (1) 99 is positively homogeneous of degree s or (2) s is an integer > 0 and (p{x) = (pi{x) log!A;] + (p2(x) where (pi is a homogeneous polynomial of degree s and 992 is positively homogeneous of degree s. Remark. If 99 ^ C^ (Rp (O}) and is essentially homogeneous of degree s, then any first partial derivative is essentially homogeneous of degree s 1.
214
In 6.2, we introduce certain fundamental solutions of equations (6.1.1) with constant coefficients and present interior estimates for the solutions of systems (6.1.1) with variable coefficients; the principal results are stated in Theorems 6.2.5 and 6.2.6. In 6.3, we introduce the
"Poisson kernels" of AGMON, DOUGLIS, and NIRENBERG ([1], [2]) and
VoLEViCH and introduce certain solutions of certain boundary value problems for a half space in which the Lj]c and Bric have constant coefficients and coincide with their principal parts. These results are used to obtain the principal results on higher differentiability, given in Theorem 6.3.7, and on ''coerciveness" as given in Theorem 6.3.9. In 6.4, we first obtain results on local coerciveness and higher differentiability on the interior for "weak solutions" of equations (6.1.1) and then obtain similar results at the boundary for weak solutions of (6.1.1) which satisfy the boundary conditions in a certain weak sense. In 6.5 we treat the Dirichlet problem for the "strongly-elliptic" systems introduced by NIRENBERG [2]. In 6.6, we extend the analyticity results for linear systems given in the recent paper by MORREY and NIRENBERG to apply to solutions satisfying analytic general boundary conditions along an analytic part of the boundary; the developments are like those of 5.7. I n 6.7, we extend the results of the author (MORREY [12]) and A. FRIEDMAN ([1]) concerning the analyticity of the solutions of non-linear systems to the case of general non-linear analytic boundary conditions; the developments are like those of 5.8. Finally, in 6,S, we obtain results on higher differentiability of the solutions and weak solutions of certain non-linear problems and then prove a perturbation theorem generalizing that (Theorem 12.6) in
the large paper of AGMON, DOUGLIS, and NIRENBERG [1]. Our estimates
are of two types, the Lp type involving Lp estimates of the solutions and their derivatives and the Schauder type involving estimates of Holder norms. We have not attempted to present the latest developments in the theory of elliptic differential equations nor have we attempted to trace the development of the subject. But a great many references are to be found in the recent books and expository papers by HORMANDER [1],
LIONS [2], MIRANDA [2], ROSENBLOOM, B R O W D E R [1], [2], GARDING [2],
STAMPACCHIA [1], NIRENBERG [3], and the author (MORREY [15], [18]).
We have not presented an account of what might be called ''natural" or 'Variational" boundary conditions. We present only special cases of such problems in Chapters 7 and 8. We have in mind here the regularity at the boundary and the boundary conditions satisfied by the solutions u of equations of the form (6.4.1) where we require that these hold for all V in some linear manifold V and perhaps require u to satisfy (possibly in a weak sense) fewer than m boundary conditions. STAMPACCHIA [2] has shown t h a t if u{N = 1 ) ^ H\[G) and satisfies an equation of the form
6.2. Interior estimates for general elliptic systems
215
(5-2.1) for aU v in some V such that HIQ{G) C V C Hl(G) and if dG satisfies a mild degree of smoothness, then u^ C^(G). For higher order equations this approach to boundary values was formulated by VisiK and SoBOLEV [1]. Quite a few results along these lines have been obtained recently b y SCHECHTER. N O doubt other workers in the field, such as
AGMON, BROWDER, N I R E N B E R G , etc., have contributed results b u t the
author is not familiar with them. 6.2. Interior estimates for general elliptic systems In this section, we consider interior estimates for the solutions of the general elliptic systems of the form (6.1.1). We shall be able to include a simple proof of the Lp estimates. Our method is essentially that of 5-5 and 5-6. We begin b y writing the equations in the form (6.2.1) Lojjcu^- =fj - {Ljk - Lojk) u^,
where LQJ]C denotes t h a t operator with constant coefficients obtained from the principal part of Ljjc b y evaluating the coefficients at the point XQ. Since these equations with constant coefficients cannot, in general, be reduced to a simpler form b y affine transformations of the x variables, we can only assume XQ to be the origin and have to consider the general case. Moreover, for such systems, there is generally no simple law of reflection, so the boundary neighborhoods require the more complicated treatment given in the next section. We shall assume t h a t the coefficients in Ljjc, and the functions u^ a n d / j satisfy the following minimum conditions: Minimum conditions. In case Sj = 0, we require the coefficients in the principal part of each Ljjc to be continuous and the others to be bounded and measurable. Ifsj < 0, we require merely that all the coefficients^ C^'^~î[G). We shall assume that m > 0, m being the total order of the system. We assume that fj ^ H~^^ (D) and i^Jfc {D) for some ^ > 1 and each D G G G. We do not assume that the system is properly elliptic. Remarks. The reader will see t h a t some of the results can be generalized b y relaxing somewhat the boundedness requirements as was done in 5-5 and 5-6. However, an attempt to state the most general theorem possible would add greatly to the complications without adding much interest. After reading those sections and this one, the reader will be able to formulate and prove whatever results he m a y need in this line. As in 5-5, 5-6, we shall first assume t h a t the vector u has support in BR and satisfies (6.1.1) there. Next, for each sufficiently small R, we alter the coefficients of the Lj]c outside BR thus obtaining new operators LRJJC, the coefficients of which reduce to those of LQÂ; on, near, and outside of dB2R, the requisite continuity properties being preserved. Then
216
we define P2i2(/) to be a particular solution UR on B2R of the equations (6.2.2) LojicU%=fj. Next, we extend u to be zero on B2R BR and write u =^ UR-\- HR where UR = P2R[Lo{ti)] so that HR is a solution of the homogeneous equations (6.2.2). It is then shown that HR is a polynomial of bounded degree with \\H\\^ < C \\u\\^ in a convenient space. If we regard HR as known, then UR satisfies an equation of the form
(0.2.3)
UR TR UR = VR,
TR UR
VR = P2R(f) P2R [{LR Lo) HR]^

[LQ U
= P2R [ (LR LQ) UR]^
= f (LR LQ)
U)
If R is small enough, ||TR\\ <, i/2 and it follows that Mt<C,\\f\\r+C2{R)\\u\\0, where the norms are specified below. In case / and the coefficients have additional smoothness properties, it is shown that u then possesses additional smoothness properties. These results are sufficient for interior boundedness theorems and, together with the results in the next section, are sufficient for the estimates in the large including the boundary. We now set about defining the operator P2R. Suppose that we let Lo(A) be the determinant of the matrix LO;A;(^) of characteristic polynomials, let Ll^(?i) be the cofactor of Loijc(k), and let Lo(D) and Ll^(D) be the corresponding operators. Then a solution of (6.2.2) can be obtained by setting (6.2.4) U^ = Ll\D)Fi where LoFi=fi and our problem is reduced to solving the second equation in (6.2.4). We now do that heuristically by Fourier transforms. Introducing the transforms by their usual formulas F(y) = (In)-'' f e-^^-yF(x) we find formally i' Lo(y) P(y) = f(y), F{y) = (-i)^ L^^ (y) fiy^dx, (r ^ vjl),
Taking the inverse Fourier transforms, we find formally
F(x)=fK(x-i)mdi
(6.2.5) ^' K(x) = (In)-''( i)"^ j e^^' y L^^(y) dy.
1 Of course {LR LQ) U denotes the vector F defined by

Fj = LEJJCU^ LQ j jc U^
2 Recall t h a t m is the total order of the system and may be odd if v = 2.
217
If we evaluate the integral for K by introducing polar coordinates in the y space, we obtain (6.2.6) K{x) = (27)-^(-i)^fL-^(G)lfr^-^-êirix-o)Â^2;.
If Im (x ' a) were > 0, the integral in the brace would converge to (6.2.7) Jv-i-m{^'(y)> JhM = hl{iw)-^-^, A>0, m<vi. In case m '> v 1, the integral would diverge. But we notice that (formally)
(6.2.8)
If we define (6.2.9)
A^'M-^'^Jh^vMj ^ ^ ) = ^ : ^ ^ - ^ ^ ^ ^ Co - : 0, Ct= \ + 2-^ + " ' + t-^ ^ ^ ' ^ > 1,
if
we see that (6.2.8) principal log and Then, instead of Jv-i-m, we choose (6.2.10) and define Kl(x)
holds for all integers h siadp; for the log, we take the cut the w plane along the negative imaginary axis. defining K by (6.2.6) with the brace replaced by P as a positive integer such that V \ 2P = (27i)-^{~-t)n^+^P <0
fL-'(a)Jr-i-m.2p{a'x)dZ i: K (x) = AP K% [x], Mp {x) = Lo K* (x). This function K will be seen in Theorem 6.2.1 below to be the desired function. We first prove the following lemma: Lemma 6.2.1. If yo is a real vector 9^0, there are functions cl[y), analytic for complex y near yo, ^^^^ l^^l t^^ transformation (6.2.11) (6.2.12) Cr = cl{y)X^ is a rotation of axes for each real y near yo ^^'^^ such that Proof. Let 7 = y ' / | y | . Making a, fixed rotation of axes, we may ^^ assume 770 = (1, 0, . . ., 0). If, in (6.2.12), ^ denotes the row, we define
cl(y) =7]", (X = \, . . .yV
and then notice that there is a unique way to complete the matrix c^ (y) in such a way that the matrix is orthogonal, cJJ (y) = 0 for ^ > )/ > 2, and each determinant in the upper left hand corner is positive. The cJJ are analytic. Theorem 6.2.1. (a) The functions K%, K, and Mp are analytic for all complex X = xi -j- i X2 1^ 0 with \x2\ <. hi\xi\, where h\ depends only on the ellipticity hound E.
218
(b) Kp and K are essentially homogeneous of degree m -\- 2P v and m V, respectively, and satisfy (6.2.13) K*(-x) = (-\)^K*{x), if v>2; K(-x) = {-\)^K{x). (c) Zl^-i Mp{x) =Ko{x)^) + const.
(d) I f / ^ CÎ(^R)
a v = 2, z1^-i Mp{x) = Ko{x)
^^^ ^ ^5 defined by
(6.2.14)
then F^ C^(BR)
F(x)
BR
=lK(x-^)f(i)d^,
< C hf,(f, BR) and if f^ CÂBR),
with hf,{V^F,BR)
F^ Q+fe(5^) uîth A^(V^+^F, BR) < C h^{V^f, BR), where C depends only on E and on fi. Moreover LQ F = f. (e) If f^Lq(BR) for some q > i and F is defined by (6.2.14), then F^H^BR) with | | V ^ i ^ | | , % < C - | | / | l o ^ and if / ^ / f g , ( 5 ^ ) then F^H'^+îBR) ^^^'^^'||V^î^||J,jj < Ci||V^/li^B where C depends only on q and E and C^ depends on q, E, and k. Proof. We begin by using carefully the definition of Jh{w) and breaking u p the integral for Kp into integrals over ^ + and over ^ - where or X > 0 and < 0, respectively. This yields K(^) = CplL^â)'{a'x)h[-log$y'x\
+ Cp j L^^{G)
E-
+ Cn + inl2\d^
' x\ +
' {G ' x)h[-\o^\a
Cn~i7il2\dZ,
Cp=
\lh\(27iiY
{h = 2P + m v).
Now, since LQ is of degree m, L^'^{a) - {a - x)^ is even or odd according as V is even or odd. Thus 2Cp f L^^{a)' (6.2.15) K(x) = {a' x)f^ [-log\a' x\ + Ch]dZ> v even , V odd.
\i Ji Cp J LQ^ (o)' [G ' x)^ d2^

2;+
The properties of X J in (b) follow from this representation and the analyticity of Kp follows by making the rotation
(6.2.16) G''==c^{x)Ty
of Lemma 6.2.1, which is analytic in :\; near an :vo ^ 0. When this is done, ^+ is just the subset of ^ where r^ > 0 and
G ~ G{X,T), G' X ^ \x\r^.
To prove (c) we note t h a t Mp(x) = {27z)'-^{-$Pfj,-i-2p(o'x)d2;-
^ KQ is the kernel for POISSON'S equation, as defined in (2.4.2).
219
We shall prove (c) in the case v 2k; the proof for v odd is similar. Then we notice that
and C is a constant. If ^ = 1, so i^ = 2, this is the result, since Pr 2n. If ^ > 1, we find by taking zl^~i of this, t h a t zd^-l Mp[x) = - (27l)-Tr 2 4^-^{k - 2)\ {k - 1)! ^2-"
which is seen to reduce to KQ (X) ; one uses the formula Fv = vyv, I f / $ CI[BR), (6.2.17) it follows that LoFi(x) = fMp(x
BR
JTV = yv-i 2 J cos" 9 dd.
F is given by (6.2.14), and F% is given by F%[x) =JK%[x - I ) / ( f ) di

BR
- f ) / ( | ) d^, ^JKoix
BR
APF'l.(x) = F{x) - ! ) / ( ! ) d^
LoAP-^F*(x)
so that it follows from Theorem 2.6.7 that F^ C'^^{Bji), satisfies Lo F = f, and has the stated bound. I f / $ C^^(Bji), we see b y integrating by parts successively that V^' F is given by (6.2.14) in terms of V^/, j < k. Thus the remaining results in (d) follow. The results in (e) follow from these, the usual approximations, and Theorem 3-4.2. We now define our spaces and norms. Definition 6.2.1. For each integer A > 0, we define two pairs of
spaces, *H^Q{BR) and '^'^H^{BR) of vectors having components in
various J^*-spaces (^ > 1), and *CjJo(5ij) and **C^(jBi?) (0 < ^ < 1) of vectors having components in various Cj^{Bji) spaces with respective norms defined as follows:
(6.2.18)
0
where we define the norms
**\MtB =
I'\MCR''
'll9'llU=i^-^l|V*-'9'||%
(6.2.19) 'MU = S[R-'h,{Vt-^<p) + R-'-^WW^-'CPWIR]
s-0
\\\(p\\\R=îm,x\(p(x)\.
XSBR
220
We now define our operator PRhoif) depending on the integer ho defined in 6.1 in connection with the boundary value problem. Definition6.2.2. I f / $ * H%{BR) for some A > Ao > 0 and ^ > 1 or / ^ some *Cô (BR) for some h ^ ho and /^, 0 < // < 1, we define PRhoif) = U where
m = Ll^[Fi - F*], Fi = fK{x(6.2.20) F* (x) = y
$)MS) d^
siv
^ V Fi(0) x\ Q = m-{-ho
(the sum being 0 if ^ < 0). The following lemma simplifies the handling of the derivatives of low order.
Lemma 6.2.2. (a) If u ^ H\>{BR) p ^ q ^ 1, > 0, and s is an in-
teger with 0 < 5 < ^, then Ci = Ci{s,p,q,v, Thus sj), Q = v{p q)lpq.
1kllU<C2(||V^^llS^+i?-^-^||ÎW,
(b) Ifu^ CJ,{BR),
C, = C,(p,q,vJ).
0<s<t, 0<s<t; a = vjq.
0 < / / < 1, e > 0, and q > \ , then
hf,{\Ji-û,
BR) <eR'h^{Vû,
BR) + CsR^-t-^-f^M^,
11 V ^'' u \\\R < e R^+^ h^Vû, BR) + C4 i^^"^"^ \\u\\%, 1 Cs = Cs{s, qy fJiy Vy s, t), C4 = C4(, q, JLC, V, s,t), Thus 'M\R^Ci{q,li,v.t)[h,(ytu.BR) +
R-*-'-^'\\u%j,].
(c) The same results hold with BR replaced hy GR. Proof. In each part, the last statement follows from the first. From considerations of homogeneity, it follows t h a t it is sufficient to prove the first statement in each part for the case R = 1. If (a) were not true on Bi, there would exist a sequence {un} with '||^7i||^ = 1 such that Un 7 u mH%{Bi)hut Then u = Q and (since s > 0) V^~^ ^î ^ 0 in Lp (Bi). But then \7* Un-^0 (strongly) so that w^ - ^ 0 in H\{Bi), contradicting the fact that '||^w||î = 1. Suppose one of the first inequalities in (b) fails to hold on Bi for some s > 0. Then there exists a sequence {un} such that '|||^?i|||i == 1, V^~^ Un = converges uniformly to V*~^ u for some ^, 0 < s < ^, and Ki'^^-'Un) > eh^VÛnyBi) +n\\un\\^,. It then follows that uÔ and it is easy to see that hf^{\7i-s Un, Bi) ^ 0 if s > 0. B u t then h^{\/tun,Bi) - > 0 so '|||^^|||*,i->0, a contradiction.
221
The other inequahties in (b) are proved similarly. Part (c) is proved in the same way. Theorem 6.2.2. PRUQ is abounded operator from '^H^Q{BR)to '^'^H^(BR) and from *C^Q(BR) to **C^(5i2) with hound independent of R if h > ho > 0. The bounds depend only on h, ho, p, or /u, and E. Moreover if U = PRhoif), then U satisfies the equations Loj,UHx)=f^{x)-f*(x) where ff is the Maclaurin expansion of fj out to [and including) the terms of degree ho Sj v. Proof. By integrating by parts we see that Fi{x) - Ft(x) =f
BR
^ ^
C.K4x,i)D''Mi) lvX.(-f);c^
\oc\=ho~si
di,
C = ^ ^ ^ T ^ Q= m + ho-si~v
\ix\=hosi
K4x,^)=K{x-^)(6.2.21) K^x)
= D-K%^{x),
C,D-K^x)
K(x).
Since (see Theorem 6.2.1) Ko, is essentially homogeneous of degree Q, we have Kcc(y) =Ki4y)log\y\ + Kôîy), V^K^y) = V^ iîa(y) log|y| + + K2scc(y) where iCia is a homogeneous polynomial of degree q and K^cx. and K^so: are positively homogeneous of degrees q and Q s respectively. Thus, if we SQi X =^ R y, ^ =^ R 7}, y, Tj ^ Bi, then K4Ry, (6.2.22) + RlogR-[Kicc{y-fj)and the second term vanishes-. Thus (6.2.23) f\Koc{x,S)\di^CRQ+^,
BR
Rfj) = [K,{y - f,) ~ ^
V'
K4-f,)-y^]-
R + '
*ê ^^y'Ki4-'>j)-y^]
j\Ko.{x,^)\dx<.CRQ^\
BR
X.^^BR.
The analysis above assumes that ^ > 0 . I f ^ < 0 , the results (6.2.23) follow from Lemma 3-4-3- Using the Holder inequality, we find that \Fi(x)-F*(x)\^ (6.2.24) <(Ci?e^-)-i / 2: C.\K4x,^)\\v'^-^'fi(^)\9dS
so that we can say that \\Fi - Fff, < Zi R^^^ \\ V'^--^/z||? < Z2 R^^-^^-ô I v'^-Vî?. I The fact that Fi ^ H'^+h-'' with || v^^'^"^^ Fi||o < Z3 || V'^"'^ fi||^ follows from Theorem 6.2.1. Using Lemma 6.2.2 we obtain the first
222
result. The Holder results are obtained similarly. The last statement follows from Theorem 6.2.1. I t is convenient to introduce the following terminology: Definition 6.2.3. The operator L (i. e. the matrix [Lj]^) is said to satisfy the h-conditions [h > 0) on a set G if and only if (i) ii Sj = h = 0, the coefficients of the highest order derivatives in the operators Ljjc are continuous on G, the others being bounded and measurable there; and (ii) if ^ Sj > 0, the coefficients in Ljjc^ C\~^^~'^{G). The operator is said to satisfy the h fi conditions on G if and only if the coefficients in the operators Ljjc^ C^~^^{G). Given the operator L defined on BA, we define LR by (6.2.25) LRJJC - Lojk + (p(R-^ \x\) {Ljjc - Lojic), 0 < i^ < Ajl, where LQJJC is defined near (6.2.1), and 99 is a fixed function ^ C^{Ri) with (p(s) = \ for s < 1, (p{s) = 0 for s > 7/4, and 99' (s) < 0 for 1 < s < 2. Theorem 6.2.3. (a) Suppose that L satisfies the h-conditions on B^ and LR is defined as in (6.2.25). Then LR satisfies the h-conditions on B^R and {LR LQ) is a bounded operator from **i7^(52i?) ^fîo '^H^Q{B2R) with bound e (R), which depends on E and E', and which - > 0 as R ->0. (b) 1 / L also satisfies the h-ju-conditions (0 < â < 1) on BA, then so does LR and LR LQ is a bounded operator from **C^(B2R) into *C^Q(52i?) with bound of the form C R^^ where C depends only on E and E'. (c) / / in (a) or (b) above h> \, then the bound of the operator LR LQ is <,CR where C depends only on E and E'. Proof. The proof of (a) is similar to but simpler than that of (b) and the proof of (c) is similar. We sketch the proof of (b). Suppose u^
* * CI{B2R) and/ = {LR LQ) U. Then
l,'fcW-ô,-fc = 9^(^"MÎ)[^?fcW-<fc(0)]
if
\oc\=sj
+ tjc
^hk(^) = (P{R~^ l^D^hi^)
if IÎ < 5; + h
the difj^ being the coefficients in Ljjc. If \^\ = h sy(> 0), D^fj is a sum of terms of the form Dy (pR' D^ {a^j^ - a^^j,) D^ u^ (pR{x) = where y, d, etc., are multi-indices and Irl + 1^1 + 1^1 =h + h, \y\ + \d\ <h-Sj, 1^1 + \M + \Q\<h + h, \>c\ + |Jl| <h-Sj. The derivatives of the afj^ and their Holder constants are bounded independently of R while we easily see from the definition of ' | w|||Jij and and D'' cpR D^ afj^ - D^ u^, (p{R-^\x\),
223
the form of (pR that ill^^^^|||2i?< (2i^)*'^+'^-|^l+î^, \lâ a^W^R ^ C'R^, hf,[Dû^,B2R] < [2Ry^^^-\^\'K
a = JLC ii h Sj = 0; cr = 1 if A s^- > 0,
(6.2.26)
( ^ = **||||||^J.
It follows easily that hfi(\/^~^^fj B2R) <, C K and the rest follows since
the fj vanish on B2R B^^RJ^.
Definition 6.2.4. If L satisfies the A-conditions on Bj^ and R < A/2, we define the operator TR on the spaces **-fif^(52i?) and **C^(52i?) by the condition t h a t TRU = P2R [{LR LQ) U] . From Theorems 6.2.2 and 6.2.3, we immediately conclude the following theorem: Theorem 6.2.4. / / L satisfies the h-conditions on B^, then TR is a hounded operator on '^*H^{B2R) for each p '> \ and R ^ Aj2 with hound 8(R,p) where s depends also on E and E' hut - ^ 0 as R -^0. If L also satisfies the hju conditions, then TR is hounded on **C^(^2i2) ^^^^ hound e{R, fi) of the form C(ILI, E, E') R"^, a = /u or \ as in (6.2.26). Remark. TR depends on the integer ho which we may, of course, regard as fixed. We can now prove the following interior differentiability theorem: Theorem 6.2.5. Suppose that L satisfies the h-conditions on the bounded domain G, u^^Hf+^'(G) for some q > i, and the fj = Lj]cU^^H\~^j{D) for each DCCG and some p^q and h > AQ- Then u^ ^ H^^'^'^ (D) for each
DGGG.
If, also, L satisfies the h JJL conditions on G and the fj^C^~^j(D) for each DCCG, then u^^C*^+^{D) for each such D. If L and f^C^{G), thenu^C^{G). Proof. The last statement follows from the preceding ones. To prove the first statement suppose D C C G and suppose D C C A C G G. Then the fj^H^"^) {A). We first assume A > 1 + Ao and shall show that w^ ^ H*^^^ (D). With each point XQ of D is associated an i? > 0 such that B{xo, 2 ^ ) c z l and the bounds of TR on the spaces '^*H^[B{xo, 2R)] and * *i7J [B {XQ, 2i?)] are both < 1/2; we take ho = 0 for this proof. We can find a C^ partition of unity fi, . . ., Cs defined on a domain F Z^ D and such t h a t each Cs has support in some one of the B [xo, R) Let ti^ ~ ^g u^; we shall show that each ^lsÎI*q^^(G). To do this, we suppose that the support of Us is in BR .Then, clearly
US^*'^H^,(B2R) (6.2.27) ^nd Ljjc < = Csfj + Msjjc U^ = fu if3 = Ljk U^)
and fs^*Hl{B2R)
since the operator Mgjjc is of order ^Sj
+ tjc 1.
224
Regularity theorems for general boundary value problems (B2R),
Now, write Ug = USR + HSR where USR = P^R [ÔS ^ S ] . H US ^ C^
then
mlK(x
- ^)LosimK(i)d^
B2R
= LllFis(x) = u^(x).
B2R
= mLosimjK{x-i)uf{^)di=LosjK(x-i)u^($)d^
Thus, from the definition of P2R, we see that H^j^ is a polynomial of degree <,tk V. Thus UsR satisfies
UsR = P2R{fs) P2R[{LRS LQS)USR] P2R[(LRS
[{LRS
LQS) HSR], ^Q s) HSR] .
Or
UsR TsR UsR = VsR = P2R ifs) ~ P2R
Since ITîj |] < 1/2, it follows that USR and hence US^^'^H\{B2R). Accordingly ^ ^ * * ^ J [A] for somezl'Z^D.The argument may be repeated choosing R smaller if necessary. I n a finite number of steps we find that To raise the exponent from q to p, we proceed similarly b u t choosing R so that | i r i j l | < 1 / 2 in **HI^{B2R) where qi = vql{v q), the Sobolev exponent corresponding to q. Then, from (6.2.27), we see that fs^ *H\[B2n). A finite number of repetitions accomplishes our purpose. In case L satisfies the h fji conditions and / ; $ C^~^j, we may take p so large that u^^Cf^^^^^'^ and then fg^ *C^ and we m a y conclude that u^ ^ Cfl''^^ b y taking R small enough. For later use, we need the following lemma and theorem: Lemma 6.2.3. There exists constants Ci(v, s, t, p, q) and 2(7], s, t, ju, q) such that 1 /f IIU < Cx R-t-^ IIH ||o^, 'IH11^ < C2 R-^-^-^ IIH ||o^, Q=v{p-q)lpq, G = vlq, p>q for all polynomials H of degree < s . Proof. Let n be the larger of s + 1 and t + \. Then, from Lemma 6.2.2 with = 1, ^ replaced by n, and s replaced hy r = n t, WQ see that (since V H = 0) I V^HWl^ < C[R-t-q H\\l h,[VtH) < C^R-t-o-^\\H\l^. The lemma follows from Lemma 6.2.2. Theorem 6.2.6. Suppose L satisfies the h-conditions on B^. Then there exist constants i?2 > 0 and C2 depending only on v, h, q, E, and E' such that
** i u |i,\ < C2 [* IIL IIJ^ + *'\\u 11?^], *'\\u 11;^ =^R-h-'^-^ 1 ^ 1;^ 1
0 < i? ^ i?2, M* H*g''{BR), Q = v\q- \)lq. Also there exist constants Rz> 0 and Cz, depending on v, h, fi {0 < ju < \) E, and E' such that if L satisfies the (x h conditions on BA, then * * | | | f c < C 3 [ * | L ! l | ^ + *'|||||?^], 0< R<R3, M^ CJf+''(5j;) . * *'\M\i^^R^'-"*'\\u\\u
6.3- Estimates near the boundary; coerciveness
225
Proof. We prove the second, the proof of the first is similar. We omit the subscripts /^ i?. We confine ourselves to R so small that || TR || < 1/2 in the space **CjJ;(52i2). Suppose U^Ô^;^^{BR), define LR, P2R, and TR as above (with some fixed ho < A); we assume u = 0 on B2R BR. Let UR = P2R(LO U). Then, as was seen in the proof of Theorem 6.2.5, u^ =u\ + H\, where H\ is a polynomial of degree <,tjc v, and (6.2.28) UR-TRUR = PR(f)-PR[{LR-Lo)HR], f=LRU^Lu. Using the definition of UR, we obtain (Z^ independent of R)
* * II î? Ill ^ < z f * III ^ III ^
* * II /iT/j II ^ < z f * IK ^ i p .
Using (6.2.28), theorems 6.2.2, 6.2.3, and 6.2.4 and the fact that || TR\\ < 1/2, we find further t h a t (6.2.29) *1l|î^fc<^3ll|/|||Ê + î(^)**|||^fc '\\\uRr^n<2,Z*\\\f\\\lJ,+Z,el{R)**\\H\l^ ..*'\\\HRrÎ,<*'\\\ur^B + *'\\\uE\r^n. **\MtR (^4 = 77^) From the definition of the norm, we conclude t h a t (6.2.30)
From (6.2.30), (6.2.29), and Lemma 6.2.3, we conclude that **|||/^flill^ <Z*'|||!||?B + Z*\\\f\\\lj, + e2(R) from which the result follows. 6.3. Estimates near the boundary; coerciveness
**\\H\''uE<zrtH\'in+^nf\&R + ^s{R)*nMiu
We now carry out the boundary estimates for the boundary problem described in 6.1, namely (6.3.1) LjjcU^=fj in G, BricU^ =^ gr on dG, We shall begin by considering functions defined on hemispheres GA instead of full spheres BA but otherwise shall try to imitate the developments of the preceding section. For functions u having support in GR U OR we obtain estimates for u in terms of those f o r / a n d ^ as defined in (6.3-1). The fact that the coefficients in the boundary conditions are also variable causes an extra complication which we take care of in the following way: We define the operator PR (/, g) for the pair (/, g) to be a certain solution of the equations (6.3.2) LojkU^ = / y on R+\ %"+! > 0, BQrTcU^ = gr on a where the fj and gr have support on G2R U (y2R and a2R, respectively. Then if u is given with support in GR U CTR, we alter the coefficients of L and also B outside GR to obtain operators LR and BRrk as in the preceding
226
section and then write u =^ UR + HR where (6.3.3) UR = PR[f[LR - Lo) u, g-{BR~ Bo) u].
I t turns out again that HR is a polynomial of bounded degree. We define

(6.3.4) TR((P) = PR[~{LR Lo)cp, (BR - Bo)cp]
and the equation (6.3.1) becomes

(6.3.5) UR -TRUR^VR = PR [f (LR Lo) HR, g (BR Bo) HRI
We define PR(Lg) (6.3.6) = UR+VR-WR, UR = PR(0,g), y, = BorkV% WR=PR(0,y),
and VR is defined in a way very similar to the Pho,2R (/) of the preceding section, / being a smooth extension of / to B2R', PR(0, g) and PR (0, y) are defined b y means of the 'ToissoN kernels" introduced by AGMON, DouGLis, and NIRENBERG ([1], [2]). We write PR instead of P2R in this section, b u t still work on G2R. In order to obtain the Lp results, it is shown how to express the derivatives of highest order b y means of singular integrals of GALDERON-ZYGMUND type; this is an idea similar to one used b y BROWDER [1], [2]. We shall replace r by 1 + 1 and x^+^ by ^ y and shall write the operators LOJTC(L>) and Bork(L>) in the forms Lojjc (Dx, Dy) and Bork (Dx, P>y) where x = (x^, . . ., x"). We now develop formulas for the Poisson kernels by attempting to solve the equations (6.3.7) LojTc(Dx, Dy) u^ = 0 on R+, Bork(P>x, Dy) u^(x, 0) = gr(x) on a, formally b y introducing the Fourier transforms u^ and gr with respect to X defined b y (6.3.8) uk(X,y) = (27t)-'' f e-'^''û^(x, y)dx, (r = vl2)
with a similar formula for gr. This yields the equations (6.3.9) Lojk(iK Bork(il (6.3.10) Dy) u^(X, y) = 0, y > 0 Dy) u^(l 0) = gr(X), y = 0. |A| y)
It is convenient to normalize these equations b y setting ^^(A, y) = (i \X\)-^k um gr(}i) = (6.3.11) (6.3.12) Lojk((r, -iDn)U^(lu) Bork((r, ~iDu)Um (im)-^rGr(X). =0, 0) - Gr(?i) u>0, (a = |>l|-U, \a\ = 1).
Then the equations (6.3.9) in w = |>l| y become
227
We begin t h e solution of these equations by stating the following well known lemma: Lemma 6.3.1. (a) Every solution U of a single ordinary differential equation {with constant coefficients) of the form L[-iD)U is of the form = 0, L[z) = aoftiz ^ -rj)^j
U(u) =2'^;M^*'''^"^
where each Pj is a polynomial of degree <,mj ~ 1. Moreover, every function U of this form is a solution. (b) / / s ^ any rj, the equation L{iD)U=P(u)e^^'^ ( P a polynomial) has a unique solution U of the form Q [u] e^ ^ ^, where Q (u) is a polynomial of the same degree as P. If s = rj, the equation has a unique solution U of the form u'^jQ {u) ^*'^?* where Q is a polynomial of the same degree as P. Definition 6.3.1. A function of t h e form P{u)e^^'^, [P a polynomial) where / w s > 0 is said to be exponentially decaying.
We now prove the following fundamental theorem of AGMON, DOUGLis, a n d NiRENBERG ([1], [2]):
Theorem 6.3.1. We suppose that the matrix of operators LQJJC is properly elliptic. Then there are fixed rectangles R+ in the upper half-plane and R~ in the lower half-plane such that R^ contains all the zeros with positive imaginary part of LQ {a, z) for each a with | or | = 1 and R~ contains all those with negative imaginary parts. The operators Bork satisfy the complementing condition with respect to the LQJJC if and only if there are exponentially decaying solutions Us of the homogeneous equations in (6.3.11) which satisfy (6.3.13) BorJc{(^>-^Du)U^(a,0) =:drs, r,s=\,...,m. If they exist, these functions are unique and are analytic near \a\ = 1. / / we introdtice the functions U^{o, z) by the formulas (6.3.14) U^,{o, z) = {27z)-^fe~i^Û^{a,
0
u)du
we see that the U^ are analytic in G near \G\ = 1, are rational functions of z having denominator L^ [a, z), defined as in (6.1.6), and we have (6.3.15) U'l{a, u) = jei^Û\{G,
dR+
z)dz.
Finally there are polynomials Pj {G, Z) in z such that

(6.3.16) Lojjcia, z) U^{G, Z) = PJS{G, Z).
Moreover, the U^ satisfy the conditions. (6.3.17) fBork{(r,z)U^{G,z)dz

OR
= drs> r,s=
\, . . .,m..
15*
228
Regularity theorems for general boundary value problems-
Proof. Since the zeros of Lo(a, z) vary continuously with a, the set \a\ = 1 is compact, and no zero is real, the existence of the rectangles R+ and R- follows. Until the last paragraph of this proof, we assume that o* is a constant and suppress it. Let us consider the matrix LQJ]C[Z) of characteristic polynomials. By applying the elementary operations of (a) interchanging two rows and (b) adding to one row another row multiplied by a polynomial, one can replace || Lo;A; (-s^) || by a triangular matrix || ZyA;('2^) |i in which Ljjc(z) ^ 0 if y > ^. Clearly
L,{z)=-nLjj{z).
If we factor Ljj{z) into Lf^ (z) L~' (z), then, clearly
N^
^ (^) = i t n^ti
(6.3.18) Ljjc{-iD)U^
(^)'
= 0
^ = ^î'
nij being the degree of L^^. Moreover, the system is equivalent to the homogeneous system (6.3-11). Next we see that every exponentially decaying solution of (6.3.18) (and hence (6.3.11)) can be obtained as follows: First let U^ be any exponentially decaying solution of L^]:^{iD)U^ = 0. There are Wiv linearly independent solutions U^' (Lemma 6.3.1). With each given such U^, we associate the particular solutions of Lemma 6.3.1(b)
for the equations LN_I,]^T_IU^-^ = ~L]S^_I,N U^, LN-2,N-2U^~^
= L N - 2 , N - 1 U^~^ J^N-2, N U^, etc., in turn. Next, take U^ = 0, let U^-^ be any exponentially decaying solution of L^-i, N-1 U^~^ = 0, and determine the other U^ as the particular solutions obtained by solving the other equations in turn. There are m]s^_i such solutions. This process may be continued yielding m == Wi + * * * + îv linearly independent exponentially decaying vector solutions of (6.3.11). Thus the totality of these vector solutions is an w-dimensional complex vector space V. If U^V, each expression Borjc(iD)U^(0) is a linear functional Fr{U). Thus the transformation Xr = Fr{U) is SL \ 1 transformation from V to Cm the range of the transformation is Cm. To show this, we set Lo+ (z) = z^ -i- ai z^-^ + ' + am and define Lt^{z) =z^ + ai z^-^ + '" + a^, ^ = 0, . . ., w - 1. We notice that if R+ is a rectangle enclosing the roots of L J , then
6.3. Estimates near the boundary; coerciveness
229
For, ii y <: p, the degree of the numerator is < w 2 and the result follows by replacing î^+ by a large circle; if y = ^, the result follows similarly; i f ^ < } / < w 1, the numerator differs from zy-^-^ L^ {z) by a polynomial of degree < y 1 < m 2. Next, we write
m1
(6.3.19)
BorJc{z)Ll'{z)=Sq'/^^
{modL^);
we consider the matrix Q = (ql^) as having m rows (indexed with r) and m N columns. The complementing condition is equivalent to the statement t h a t Q is of rank m. Thus, if Xr are any numbers, there are constants cip such that (6.3.20) With these ci^, we define (6.3.21) Then
Borici^*i))^^(0) = - 1 .
ci^ql^ = Xr.
UHu)^-leR+
fci^^L^.,^,(z)ei-^dz.
rCz^Ôr;fc(^)i:S^(^)^%^=^^^
It is clear from (6.3.2I) that the vector U^ is an exponentially decaying solution of the homogeneous equations. It is unique in spite of the fact t h a t the cip satisfying (6.3.20) are not unique. Thus, for each a, there is a unique solution U^{a, u) of (6.3.11) and (6.3.13)- Since Lo(or, z)'^ and the LOJA: and L* and B^rk are analytic in or, it is easy to see that we may choose, for each s, functions Csi^ [a) which are analytic in a near any given co which satisfy (6.3.20) with Xr = drs' Thus U^{a, z) is analytic in a also. If we introduce U^ by (6.3.I4), it is clear from (6.3.21) t h a t f7j(a, z) = (-i)csip(a) [ ^ ( a , z)]-^Ll\a, z)Li_,_^{a, z)
1 To see t h a t LQ {a,z) is analytic in a, we note that cnicf) = fz^lLîc

dR+
Z)]''^LO'{G,
z)dz =-j z^ L^'^

dR+
L^dz
so that each Cn is analytic in a and

(L^)-^ L^' = m z-^ 4 - f ^n (O) z-n-l
n=l
where the series converges uniformly for | ^ | > s o m e A, independent of a; of course i?+ is a fixed rectangle in the upper half-plane which contains all roots of LQ {a, z) for all a with | cr | = l .
230
where, near any co, the Csip{a) satisfy (6.3.20) with Xr = drs- The statements about the U^ follow. Thus, from this theorem it follows that the equations (6.3.11) and (6.3.12) have a unique exponentially decaying solution which is given by Um u) =ZGsWU^s{^> ^)> U^s = ieinzxjic{a, z)dz
where R+ is a fixed rectangle in the upper half-plane which contains all the roots of Lo+ (cr, z) for each o with \a\ = 1. If we now use the substitutions (6.3.10) and formally apply the inverse Fourier transform, we obtain (6.3.22) (6.3.23) u^{x, y) = J L^^x L^r(^x,y) f, y)gr{S)d^, where z)dz
= C^r j \X\-h+^rdXJe^^^-^-^\^'^y'W^,[a,
RV
dR+
C^r ^
(2n)-''{-i)h-K.
The integral in (6.3.23) is divergent, but if we introduce polar coordinates (s, a) with s = |A[ and \G\ = 1, we obtain L^rf^x^ y) _ C^^ jdj;
i:
jU^,{a, z) \js^-^-h-^K
dR+
Lo
^^^(^'+^^)^s
J
dz.
By replacing the bracket by Jv-i-tk+hr {a x + y z) where the Ju {w) are defined in (6.2.7) and (6.2.9), we obtain (6.3.24) L^^r^^^ y) _ Ci^rjdZJJvî-t,^hr[o'X
s eR+
+ yz)U!^{a,
z)dz.
It is convenient also to have the following integrated kernels. Definition 6.3.2. We define, for large P , *L^/(^, y) = {-\ra^r *L%{x, y) = Dl^L%^,,{x, jd^jJ._i_t,^nr-2p[o'X
E
dR+
+ yzm{a, y).
z)dz,
y),
Ll^{x^ y) =A^^Lf,{x,
(6.3.25) The following lemma is very useful: Lemma 6.3.2. Suppose (p is essentially homogeneous of degree n > 0 and (p ^ C^ (Rv+i {0, 0 }) and suppose (6.3.26) Then 1 Z ) | ^ ( Z , S)\ < C ^ | . Y | ^ + i - l ^ l | S | - i if 0<|X| <^|^1, <p(X, S) = f{X-S) -2lv^<p{-S)-X^.
\P\<n,
\D^^(p{X, S)\ = \DP(p{X ~ S)\ < C^\XS\n-\^\
A <\
if |i8| > ^ ,
where CA, and Q depend also on bounds for the derivatives of 99 on ^B(0,1).
231
Proof. The last statement is evident. Moreover, if \^\<,n, then D^(p{X, S) is obtained from D^cp{X E) by the formula corresponding to (6.3.26). So it is sufficient to prove the first statement only for ||^| = 0 . We write 3 = S^, | l i = 1, <p{X) = cp^{X)\og\X\+<p^{X) where (px is a polynomial of degree n and 992 is positively homogeneous of degree n. Then we have (6.3.27) D-<p(X) = D-<px{X) \0g\X\ + Cp2a:{X) where 992a is positively homogeneous of degree n | | . Thus
9)(Z, SI) =S'9)i(S-iX- I) [logs + log|S-iZ - f |] + S9P2(S-iZ-|) - S 2 < ^ ^ " [ i 5 - 9 , i ( - f ) l o g S + 9)2(-l)] = SUS-1Z-|)-^<^^Z)X-|)
oc\ = {oci\)...{ocv\),
since, from (6.3.27) it follows t h a t D'(p{^) =(p2{^). The result follows since | S-i X | < ^ < 1. Theorem 6.3.2. (a) The functions *L^^ "^L^, L^\ and L\' are analytic for all real {%, y) 9^ (0, 0) with y > 0 and are essentially homogeneous of degrees 2P + tjc hr ~ v, \a\ + 2P + tjc hr v, tjc hr v, and \oc\ + tjc hr V, respectively. (b) For each r, P, and oc, the functions above satisfy
k
Jc
(6.3.28)
|al=t
2C,cD^^*Lf, = *L''/.
|a|=i
ZC-DlLl'
= Lkr,
y>o
ZB,,,*m.,0)^6lM,{x),
C. =
^ ^ - J ^
where Mp is defined in {6,2.\\); it is assumed that P is large. Proof. The analyticity for y > 0 is evident from (6.3.25) and the definitions of the Jn given in (6.2.7) and (6.2.9). The essential homogeneity follows from an argument simpler than that given in the first part of the proof of Theorem 6.2.1. Since | < 1 = 1, it follows immediately from (6.3.25) 7 (and (6.2.8)) t h a t 4"'*^p+iai(^,y) = *-^'/(^,y). y>o. Thus the first formula in the second line of (6,3.28) follows. From (6.3.25) if follows t h a t ZLojjc''L%'{^,
fc
y) = C]pjdZ
S
jJQ{O'X
dR+
+ yz)ZU^jc{a,
^
z)U^,(a, z)dz (27t)-''(--\)Pi'j+K
= 0, Q = v -
\ +Sj + hr-2P,
y > 0, C]p =
232
on account of (6.3.16). Likewise
^Botic'^Lfix,
^
y)-\-'Clpfd2;fMa'X
2 dR+
+ yz)ZBotk{a,
k
z)W,{a, z)dz ht-2P.
'dp = (-\)P(27z)~*'iK-h, ZBotit^L^'ix, k 0) =
l==v - i -h hr-
If P is large, we m a y let y - ^ 0+ and obtain
(-i)P(27t)-^dlfj,_i.2p{a'x)dZ
from which the last line of (6.3.28) follows, using (6.2.11). The remaining formulas in (6.3.28) follows easily. To prove the analyticity of ^V^^ at a point (x^, 0), A:O =^ 0, we make the 5 rotation a = a{x,T) '. G"^ dî\x\~^ ' x) ly, G - x = \x\' T^ of Lemma 6.2.1 for x near X{^. Then we set r i = COS99, T1+' = of^ sin99, |co| = 1, *iV5r(A;, cp, z) = sm''-^(pJU^[G{x,
|a>|=l
T),z]dco.
Then *N^/ is analytic in its arguments for 99 near the segment [0, jt] in the 99-plane, X complex b u t near XQ, and z outside and on dR+. Also, for a suitable integer n and constant C,
n
"^L^/ix, y) = C n 0
UJ2
fjn (\x\cos(p +
+ yz) *N^/{x, cp, z) dz\ dcp, J
y > 0.
We conclude the analyticity at {XQ, 0) by noticing t h a t the segment [0, 7r] m a y be replaced by a nearby arc from 0 to JT along which / m cos99 > 0 except at the end points. Theorem 6.3.3. Suppose h > /ô > 0 and ho + hr >: i for each r and suppose that the gr satisfy the following conditions for some L: (i) (ii) gr ^ C^-^^^-(GA) with hf,(V^+Kgr,GA) <L, for every A.
(LRf^+K+f'-K 0<.t^h-{ I V^gr(X) I < \LRf^^K+f'-t(XJ2R)e-i[\ [LRf^+K+f'-*{Xl2R)6-t^
hr, \X\^2R, y > 0, + l o g | Z / 2 i ? | ] , 0 < ^ < ^, Q<t^h + hr
where Q = ho -\- hr v 1 and \X\ ^ 2 R, y > 0 in the last two lines of (ii). Suppose also that u is defined by u^x, y)=fZ
a r lai=ar
^C.Ll^(x,
y; S)D-gr{S) d^,
qr =~-ho + h -
(6.3.29)
Ll^{x^ y; I) == Llr{x - | , y) -^D^^Li^(~l

|/5|<T
o)|-f,
T ^ l ^ l + tjc hr V = tjc + ho V 1.
233
Then u^ ^ O^^^ (GA) for every A and (6.3.30) 2'\\\^'\rt2M<cL.

k
If the gr satisfy (ii) only for 0 <t <^ h + hr \, then U^Ô^+^-'^{GA) for any A and (6.3.30) holds with h replaced by h 1. In either case u is a solution of the boundary value problem (6.3.7) with gr replaced by gr gf where gf is the expansion of gr about 0 out to and including the terms of degree Q. Of course if Q <,0, the conditions for 0 < ^ < ^ are to be omitted. Proof. We m a y extend the L^^{x,y) to be of class C^ on the whole of Rvî preserving the essential homogeneity and the analyticity along y = 0, ^k; 9^ 0. From Lemma 6.3.2 we conclude that
rc-|xK*+;io-'-iî.|l|-i if \B\<:T, | x | < ^ | f | , A<\

Thus we see that u^ ^ C"^ (analytic in fact) for y > 0 and we may differentiate under the integral sign as often as desired to conclude that ^ is a solution of the equations (6.3.1) with L;^; replaced b y LQÂ; and t h e / j being possible polynomials of degrees <, ho Sj v 1; the convergence of the integral (6.3.29) at C D is guaranteed b y the fact t h a t \oc\ = r + v. X Next, we note that any derivative of the form D^^""^^ D'^^^ u^ {h ~ ho + \ oi the :\;-derivatives can be shifted onto gr) is of the form (6.3.31)
jCD^r{x~^,y)y[^)d^,
y = D'gr,
\d\=h
+ hr,
where F is positively homogeneous of degree v. Since DxFipc | , y) is absolutely integrable over a with value 0 (if y > 0), we m a y replace y (^) by y (f) r [x] in (6.3.31). Thus we find t h a t (6.3.32) < CLJ[\X - f 12 + y2]-('+l)/2 . I f _ ;t:|^^| < CL y^^"!.
By differentiating the j th equation in (6.3.1) (as modified above) ho Sj times with respect t o y and using the ellipticity, we see that we can solve for the D*^*+'* ^^ in terms of the other derivatives D^^'^û^. By performing suitable differentiations we obtain the derivatives j)h-h^+iDh+ho^k in terms of the i)^-'^''+2 2)|+/io-iÂ:^ and so on. Thus we obtain a bound (6.3.32) for \/h+^+û^. From Theorem 2.6.6 it then follows t h a t (6.3.33) h^{\/^k+û^, GA) < CL for all A, We now wish to show that w is a solution of the boundary problem (6.3.7) except for the polynomials ^f. Let (p^C'^{Ri), (p{t) = 1 for ^ < 1, (p[t) = 0 for ^ > 2, 99' (^) < 0 for 1 < ^ < 2. If, for large n, we define
234
we see that each {gnr} and its derivatives of order <^h -\- hr are equicontinuous, converge uniformly on any ball to those of gr, and the gnr satisfy conditions (ii), at least with the log term added. Thus the corresponding u^ and all their derivatives of order <.t]c + h converge uniformly on any GA to u^ and its derivatives. Then, for each fl, Hfi Id'On ~ Un where uon is defined by (6.3.29) with L^'^{x,y',^) replaced by L\^ [x I, y) and Un is the Maclaurin expansion of UQn out to terms of degree r. I t follows from Theorem 6.3.2 t h a t U{)n satisfies the homogeneous equations (6.3.7). Since LQJJCU^ is just the corresponding Maclaurin expansion of LQJJCU^^, it follows that u^ also satisfies (6.3.7). Hence u satisfies (6.3.7). Let us define
u^^x,
y) ==12 2C.L%^/{x,
r ial =Sr
y; i)D-gr{^)d^
. V^.
For each n, we find by integrating by parts, etc., that ^*fc=C7*^-F*^

a r
<^^\-y\.
Alu-^^ = uK etc..
(^^^"^^ Vt^{x^ y) = fZL%''(^ - I y)gnr{S)d^

where V^ and F * ^ are the Maclaurin expansions of U^ and U^ ^ out to terms of degree tjc + ho v \ and tjc + ho v \ + 2P, respectively, and U^ has the same formula with L^^^ replaced by L^^. By using Theorem 6.3.2(b) with P large, we find that BorlcUt^X, 0) =fMp{x
a
i)gnr(^)d^
Bor1cUl[x,0)
=gnr{x).
Since F^ is a polynomial of degree ^ t + ho v \, BorkVl^ is a polynomial of degree <,ho + hr v ~ 1 = q.^ v. Hence (6.3.35) BorJcUi(x, 0)=gnr{x) - ^ ~DPgnr{0).
Since V*^^(0, 0) = 0 for 0 <,t ^tjc-{-ho--v \, the bound (6.3.30) will follow from (6.3.33) and bounds which we now establish for \7û^(x, y) with t = tjc + ho v. From (6.3.29) it follows t h a t
V ^ ^ ( ^ , y ) = / 2 ' ZCccV'Ll^{x-ly)D-gr(^)d^,
t = h + ho ~ v,
235
Using the bounds (ii) for D'^gr {\<x\ = ho + hr 1, V*L\^ degree 1) we obtain I V * ^ ^ ( ^ ) | <.JC[\X~ f |2 + y2]-l/2.i:.iÂ-V/^+l^f +
pos. horn
Rp-âR
The second result on differentiabihty is proved similarly. Corollary. Suppose A > Ao > 0, ho -{- hr^ i for each r, and suppose that gr^Ct't^~^{(y2R)' Suppose also that u is given by (6.3.29). Theî (6.3.30) holds with h replaced hy h \. If gr^Cl-l^'{a2R), then (6.3.30) holds. In either case, u is a solution of the boundary value problem {6.} J) with gr replaced by gr g* where g* is the polynomial in (6.3.35)Proof. oT u = U V where U = AÛ*, U* is given by (6.3.34) and V^ is the Taylor expansion of U^ out to terms of degree tjc -\- ho v 1. We now develop a representation of u^ in which the highest order derivatives D^k^^ u^ are expressed as integrals of Calderon-Zygmund type. Theorem 6.3.4. Suppose /^ > AQ > 0, q > i, and ho + hy ^ \ for each r and suppose the gr satisfy the following conditions: (i) gr $ H^^h {GA) with I V ^"^^r gr ilg ^ < -^ for every A ; I
(ii)1krr3'i<^;
(iii) if Q = ho -\- hr V \ > 0, then V ^ gr is continuous on any GA with 0 < ^ < ^; and (iv) for \X\ > '^R we have 'Rh+h,-{v+i)/q~n\^X\l2R)Q-t{\+log\Xl2R^^, if 0<t<ZQ |V*gr(X)|<{^^ ^Rh+h,-iv+l)fq--t^^X\l2R)6-t, if Q<t<^h + hr. Suppose, further that u is given by (6.3.29). Then u^^ H^^'^^{GA) for any A and Z'W^^falR^^^' HW'^^^'^^^^tA^L for every A. Proof. We begin by defining r^^^6^ ^ " ( ^ ' ^ ) ^ ^ ^ ^ + ^ ( ^ ' ^ ) (^, >') 4= (0, 0), y>0 K^-^'^^) Nl^[x,y) = ~DêLlUx,y) \e\ = \ {e=\,.,,,v) and where we extend L^^ to belong to C^ on Rvjî {O, 0} and to be essentially homogeneous of degree fe + ^0 v. We then define u^{x,y)=U^[x,y)V^[x,y), vi^{x,y) = Tu^{x,y), [qr =-ho + hr \)
UHx, y) =[2
ICJ
2^M|-(X - S)Dr^gr{l
rj)\ di drj, t = tjc +
f]) +
R^+i * ' ' ' " ^ ' '
liî=i
+ mj{X (6.3.37)
- S)DID^gr{l
T^ ho-\.
D^^Dl
\^\==h-ho,
236
t/** being the Taylor expansion of U^ out to terms of degree fe + ô

-V 1.
By using moUifiers and the methods of the proof of Theorem 6.3.3. we may approximate to gr by functions ^Cc{R^+-^. For such functions we may integrate by parts in (6.3.37) if fe + ^0 > 'I and conclude that uk(x^ y) = u^(x, y), since Ni'(x, y) = -Ll'{x, y), y > 0, (x, y) 4= (0, O) MZ'ix, y) + Nl%{x, y)=0, {x, y) =# (0, 0). In case Â; + ^0 = 1, we may obtain the same result by first removing a small sphere B {x,y\ Q) from GR, integrating by parts, letting ^ ^ 0, and using the result j [x'M)J'[x, y) + yNl'{x, y)]^5 = 0;
6^(0.1)
since this integral depends only on the values of the L^^^^ for {x, y) near dB[0, 1), its vanishing is seen to follow from (6.3.36) by altering the U s near (0, 0) so that they ^ C^ [B{0, 1)]. Then v^ has the form v^x,y)=f2
jl+
ZC.S
r |a|=7io+7ir-l l l e | = l
Z D'j, M ^ ^ {X - S) Z)f ^+^ gr{S) + dS
4 D^^N^r^x - S)Dri>D^gr{S)j
which is of the form 2.6.1. The results concerning the derivatives D^-hoj)t^+ihuJ^ then follow from the Theorems of 2.6 and 2.7. The same results for the other derivatives of order h + h are found by differentiating the equations (6.3.7) as in the proof of Theorem 6.3.3. From (6.3.37) we see that u^ ( = u^) is a sum of terms of the form
(D{X) =IM(X, E)D^gr{S)dE, 1/5 I = >o + Ar ^
M {X, E) = M{X - E) -^^DyM{-S),

y
Q = tjc +
ho-v-\
and M is essentially homogeneous of degree h + h^ v 1. By the method of proof of Theorem 6.2.2 we conclude that (recall that v has been replaced by r + 1) (6.3.38) j\M{X, E)\dX<CRh-^fô, j\M{X, E)\dE <CRh+ô.
Letting. C = 0)1 + 0)2 where coi is the integral over GZR, we obtain O (6.3.39) I o)i[X) |ff < [CRh-^HY-^ f\^(X,
G3R
E) I \DPgr{E) l^dE
using the HOLDER inequality. From (ii), we conclude that (6.3.40) IID^gr IS < CR^-h'Wgr t%^ < CLR^-fô. I
6.3. Estimates near the boundary; coerciveness By integrating (6.3.39) and using (6.3-40), we obtain Using Lemma 6.3.2 and (iv), we see that if | ^ | < 2 i?,
I^v+iGzn
237
Thus, also ||C02|.2B^CL7?.+*. Corollary. Suppose A > ô ^ 0, Ao + ^r ^ 1 /<5^ ^ 3 c ^, ^'^d suppose < Â that gr^H^'^^'ifi^R) and has support in G^RÔ^R. Suppose also that u is given hy (6.3.29). Then U^^H^^^^[GA) for any A and u satisfies the conclusions of the theorem with L replaced hy ^'Igr^'^^RMoreover u is a r solution of the boundary value problem (6.3.7) with gr replaced by gr g* where gf is the polynomial in (6.3.35). Theorem 6.3.5: Suppose A > ô > 0, ô + ^r > '1 for each r, u^ ^//<*+7ii^Q^"^ jgy ^^Q}I ^ ^^^ ^(IQ}I J^^ ^(iQJi ^k^C"^ jgy | ^ | > i^Q and y > 0 for some i?o > 0, and (6.3.41) (6.3.42) LojJcU^{x,y)=ft> y>0, BorkU^=gr 0^ \Dlu^(X) I < C ^ ( 1 + I Z D^r ô-^-i-^^+i/^ ^ = 0,1,2,..., \X\^Ro, y>0, y = 0,
where ff and gf are polynomials of degree <, ho Sj v 1 and ho -\- hr V 1, respectively. Then u^ is a polynomial of degree < fe + + ho V \. Proof. Let 99 be a mollifier in the A;-variables and let w^ {x, y) be the mollified functions:
^0(^ y) = J fPei^ ~ l^)^^(f, y)d^ = J (p^{x S)u^{L y)d^ y = I f < ( ^ - f)<(f^ 7])didr].
R B{X,Q)
I t is easy to see that all derivatives of UQ of order <,t^ ^ h 1 are continuous in [x, y) and C^ in ji: for 3/ > 0 and the u^ clearly satisfy equations like (6.3.41) ani(6.3.12)if^ > 0 (the polynomials are altered but are of the same degrees). By repeatedly using the equation (6.3.41) and its derivatives, we conclude that each u^ ^ C (R^'+i) if ^ > 0. Since u^ converges uniformly with all derivatives to w^ if \X\ <. Ro and y < 0, it is sufficient to prove the theorem for u ^ C"^ {Ri'+-^). Let v^ Dl u^ where t > max(^;fc + ho, ho Sj, ho + hr). Then v^ satisfies (6.3.41) with/^^ and ^* ^ 0 and, from (6.3.42) we conclude that
\vHX)\<C{\
+ \X\)-'"^'^
X^RUi-
238
Now, if we let v^ (A, y) be the Fourier transform of % as defined in (6.3.8) and introduce F^(A, ^) by (6.3.10), i.e. V^{X,u) = {imykV^{X,ul\u\), we see that | F*^ (A, w) | < C | A | *fc for all (A, u) and satisfies Lojjc(a,-iDu)V^{ku) = 0, B^rkio,-iDu)V^[X,0) =0, r=\,.,.,m. Since V^ is bounded for w > 0 and A fixed, it must be exponentially decaying and hence ^ 0 by Theorem 6.3.1. Thus all the Di u^^ 0 and we conclude from this and (6.3-42) that uJc{x, y) =Z4{y)^^> pic<.ric=^h + hQ v \.
By applying the operators D ^ L Q for T > TA; 2 W we see t h a t each u^ satisfies Z)^ LQ U^ {%, y) = 0. Of course, if pjc < 0, ^^ ^ 0. Otherwise, since DILQU^ == 0, it follows that Dl+^^ c^^(y) = 0 iov \^\ = pjc, this being the coefficient of x^ in D^ LQ U^ {X, y). Thus, using (6.3.42), we conclude that the c^ {y) are polynomials of degree < TA; ~- pjc- Next, suppose t h a t 0 ^ p < pjc and that all the c^ (y) with \P\ '> p are polynomials of degree zjc \P\. Then, since DlLoDl^-û^{x, y) = 0 and Dl^~û^(x, y) contains no terms with \p\ > _/) we see as above t h a t Dl'^'^^'^^'^~''^c^ (y) = 0 ii \p\ = p. Thus such c^ sere polynomials which, b y (6.3-42) must be of degree ^r^ p- The fact that each u^ is a polynomial of degree < Tjc now follows b y induction. Definition 6.3.3. For ^ > 0 we define the spaces '^H^[GR), *C^{Gji), **H^{GR), and **C^(Gi?) just as they were defined in Definition 6.2.1 with BR replaced by GE\ i f / *^g or *C^ we require V^fj to vanish on ^R b u t not necessarily along OR, t 0,. . .,h SJ ~ 1. Definition 6.3.4. li hoÔ, h "> ho 1, and ho + hr '> 1 for every r, we define the space *Q(o'i?) to consist of vectors g with grKiC^^%^{GR) with norm
* ' 1 1 1 . o r IIITi V I I I a \\\l^+'hr Ills |!Ui2 2 J lllferlllÊ r
If /^ > ho, we define the space * 'H^ [OR) to consist of all vectors g such that gr[x) = gr{x, 0) where gr^H^'^^'{GR) and V^gr = 0 along 2!R for 0 < ^ < ^ + A,- 1 with norm
* 'II ^ IIJK = inf 2 " 'II gr U \ gr {X, 0) = gr (x).
Lemma 6.3.3. (LIONS) There are bounded extension operators T from

*^J(Gij) into ^H^{BR) and from *C^^{GR) into ""CUBR).
Proof. Let t max h Sj, We m a y define xf=F, where = f{x, y) if [x, y) ^ GR, F(x,y)=Oiiy^ 0, (x, y) i GR,
t+i
F{x,y)
F{x, y) =2!Qp(^>
t+i
^y)> y < 0
where the C* are uniquely defined by the conditions
2 ; ( - s ) ^ Q - 1, u = ,0 ..., t.
239
Definition 6.3.5. Suppose h^ ho>0 and ho + K ^ \ for every r. F o r / i n *CJ(G^) and ^ in *'C^(ai,) or f o r / i n *H^,{GR) a n d ^ ^ * ' ^ J M we define
PR{Lg)=Un+VR-WR V%=Ll'(Fi~Ff),
(6-3.43) ^ x
= PiRig) + P2R(f) +P3R{f) Fi{X)=JK{X-S)fi{E)dS,

^"
where UR is defined by (6.3.29) in terms of g,
f = xf ho-si-v-i,
F^{X) =2ji\^^^'^^^'
g = 2m +
T being the extension operator of Lemma 6.3.3, and WR is defined by (6.3.29) with g(x) replaced by y(x, 0) where yr{X) = BorJcV^^(X)^yro{X) -y:^{X), yf {X) =^ B^ricLl'Ff {X), Theorem 6.3.6. PR is a hounded operator from *H^(GR) X'^'HÎOR) to **H^{GR) and from '^C^{GR) x^'C^{aR) to **C^(Gi?) with hound independent of R. If u = PR (/, g), then (6.3.44) Lojjc u^ = fj / f in i?++i and B^rk u^ = gr gf on a, ff and gf being Maclaurin expansions of fj and gr out to the terms of degree ho Sj V \ and ho -\- hr v \ if these are > 0, heing 0 otherwise. The second result holds if h = ho 1 if ho 1 > 0. Proof. From the two corollaries it follows that PIR is bounded from *'H^{aR) to **H^(GR) and from *'C|^(o-i?) to ''*CI{GR) and t h a t UR = Pmig) is a solution of (6.3.44) w i t h / ; = / f = 0. Now V% coincides with the function U^ in (6.2.20). If f^*C^{GR), the norm of / in '^C^'ÎGR) < C ' R^~K times its norm in *C^(Gi2). We see by integrating by parts ho si times t h a t
Fi{X) = f
2C.K4X
S)D^fi[S)d3
K4X)=DfK,^,(X)
(see (6.2.11)) so that Fi ^CI'^+^'''{BA) for any A. Thus, we see that yr^C^'^^^(BA) for any A and that yô is a sum of terms of the form
lr{X~S)f{S)dE,
/Q"o'^(^2i?),
f = D'Ji,
\a\^ho-si.
r ess. hom. degree ho + hr v 1.
We see also from (6.3.43) that y* is the Maclaurin expansion of yro out to and including the terms in X*, t ho + hr v 1. Accordingly the Maclaurin expansion of yr out to terms of degree t vanishes. From this representation it is easily seen t h a t the yr satisfy the conditions on the gr in Theorem 6.3.3- Thus in this case P2R and PSR are seen to be bounded operators from *C^(G2R) to **C^(G2R) and from Theorem 6.3.3, it
240
follows that VR = VR WR satisfies (6.3.44) with^^. = gf = 0. The case A Ao 1 ^ 0 follows by inspection in the *CjJ; case. In c a s e / ^ *iJj(G2ij), it follows from Theorem 6.2.2 and its proof that the yr satisfy condition (i) for the gr in Theorem 6.3.4. Condition (iii) follows from condition (ii) and the relevant SOBOLEV lemmas ( 3.5). For almost all X in R^+^, we see that \7^y{X) is a sum of terms of the form co{X)=JM{X,E)ME)dE, MS)=D-MS), \oc\=ko-si, M(X, E) ^ M(X - E) - V ^D^M(~E)
where M is essentially homogeneous of degree Q t and Q = ho -{- hr V \. From this it follows from Lemma 6.3.2 t h a t j\M[X, E) I dX < CRH+^-i, {ii ho + hr ~ t> j\M[X,E)\dE<^ 0). CRH^K'K
Using the method of proof of Theorem 6.2.2 and the facts that (6.3.46) ll/o i ^ CR'^-'ôZ'Wfi
I
nW ^ CR>^-H
*\\f\l^n
we obtain the result (ii) for the yr by setting t = 0, L =C*\\f\\^^2E above and using Lemma 6.2.2. Now, if ^ ^ < 0, | Z | '> '}R,\E \ < 2R, we see t h a t (6.3.47) \M{X, E) I <C\X\^-t, f \fo{E)\dE
B2R
< CRi^^^ni-ifQ) ||/ol|o ,^
so that the bound (iv) holds in this case. If ^ / > 0, it follows from the essential homogeneity that \M(X,E)\ <C\X\^-t[i+log\Xl2R\] {\X\^}R,\E\ <2R)
from which the bound (ii) again follows as in (6.3.47). The remaining results follow easily from (6.3.45), (6.3.46), and (6.3.47)Definition 6.3.6. Suppose ho is the smallest integer such that Ao > 0 and Ao + ^r > 1 for every r. Ii h ^ ho, we say that the operators Brjc satisfy the A-conditions in B(xo,A) if and only if the coefficients all ^ C^+^'-'^ [B {xo, A)]; they satisfy the h /u conditions there if and only if the coefficients all ^C^'^'^ [B (xo, ^ ) ] ; we allow the h /u conditions if we merely have A > AQ 1 provided that A > 0. We can now imitate the developments at the end of the preceding section. We first prove the following theorem like Theorem 6.2.5. Theorem 6.3.7. Suppose A > /ô > 0 and Ao + ^r > 1 for every r. Suppose G is of class C^'^'^"^, ^0 = max^^, and suppose the operators L = {Ljjc} and B = {Brjc} satisfy the h-conditions on a domain F Z) G.
6.3. Estimates near the boundary; coerciveness
241
Suppose that u^ ^ H*^^^^ (G), fj^Hl~^^{G), gr^Hl+^(G) and

(6.3.48) p>:q,
and gr = gr on dG, where
fj = Ljjc U^, gr = BrJc U^.
Then u^^W^'^^{G). If G is of class O^'^^, L and B satisfy the h ix conditions on a domain J ' D G , and u^ ^H*^^^ (G), fj ^ C^"^^ (G), and gr^C^+^^{dG), then ^ ^ ^ Q + ^ / / G, fj, gr, L, and B^C"^, so does u^. If h^ho, fj^Cl+'^{G), gr^Cl+^r^sG), / ^ o > 1 , wêQ+'^-i(G), then Proof. The proof is similar to that of Theorem 6.2.5 from which the interior results follow. But now, also, with each A;O on ^ G is associated an i? > 0 and a mapping of class C*{'^^~'^ (in the first case) of a neighborhood of XQ onto G2R', R can be taken so small that the transformation TR defined in (6.3.4) (in terms of the coordinates on G2R) has bound < 1/2 in a given desired space. Suppose C $ C^ (G) and has support in the neighborhood of Xo corresponding to GR U (TR and let U be the transform of C u. Then, as in (6.2.27), we see that LjjcU^ = Cfj +
BrJcUf^ = Cgr +
Mjjcil^=ff
NrJcil^-=g?
where the tildas denote the transformed functions and the operators Mjk and Njjc involve only derivatives of lower order. Since these are smoother than those of highest order, the proof proceeds as before. The basic fact is t h a t HR is again a polynomial on account of Theorem 6.3-5. Next we prove a local theorem like Theorem 6.2.6. Theorem 6.3.8. Suppose B and L satisfy the h-conditions {h > ho) on BA. Then there are constants R2'> 0 and C2 depending only on v, h, q, E, and E' such that **\\ulln<C2[*\\Lutj, ulc^Hf+îGM), + *'\\Buf^j,
k
+ **'\\u\U,
0<R<R2, e = (v+l){q~ \)lq
**'iMi?B =2'^-*-''-el|*!|?B,
whenever u also has support in GR U OR. Also there exist constants i?3 > 0 and C3, depending only on v, h, /u, (0 <C jix <. '^) and the quantities above such that if L and B satisfy the h fi conditions on BA, then
0 <
i? <
i^3, SUpt UCGRU
GR.
This last result holds if h == ho \ if ho 1 > 0. Proof. The proof is exactly analogous to that of Theorem 6.2.6. The differences involve only PR: we let u = UR -\- HR,
UR = PR [LO U, BO U), Morrey, Multiple Integrals TR(P= PR [{LR Lo) (p, (BR Bo) 16 99]
242
and then we have

UR -TRUR = PR{f,g) PR [{LR Lo) HR, {BR B U. Bo) HR,]
f = LRU^
Lu,
g = BR U^
We conclude this section with the following well-known ''coerciveness" inequality. Theorem 6.3.9. Suppose h '> ho ^ 0, ho -{- hr ^ \ for every r. Suppose G is of class O^'^^"'^, to = max^;^, and suppose L and B satisfy the h-conditions on a domain F Z) G. Then (6.3.49) 111 u^ fto' <C\Z mia'^ +1 \\gr tâ'^ +2\\^' fc L ? ^ k ufc^Hl^+^(G), gr^m.+^r^G), gr = gr on WIG] > \ dG,
where C depends only on v, h, q, G, E and E', and fj and gr are given by (6.3.48). In case G is of class C^^'^^ and L and B satisfy the h fi conditions on r, then a similar inequality holds for the various ||| \j^-norms. This last result holds if h = ho 1 if ho 1 > 0. Proof. We prove the first statement, the proof of the second is similar. G can be covered by a finite number of neighborhoods 9^p each of which is either a sphere B (P, R) with B(P,2R) C.G or is the image of GR U OR under a mapping r of class Cî^^"^ [to == max^^:) which maps BA onto a neighborhood D9^p with A ~>2R. Since the transformed operators under such mappings still satisfy the /^-conditions, it is clear that we may choose each R < the number R^ in Theorems 6.2.6 or 6.3.8. And there is a partition of unity fi, . . ., Cs of class C\^^~'^ (G) each Cs having support in some one such neighborhood. Now suppose there is no such constant. Then H a sequence {un} such that the left side of (6.3-49) is unity for each n, u\ 7 u^ in H*^'^^{G) and the bracket on the right in (6.3.49) - > 0 . Then we must have u = 0 and hence < - > 0 in H*^+^~^{G). Now let Uns = CsUn. Then < s - > 0 in Hf'^^~^{G) and Uns satisfies
(6.3.50) Ljjc Uns = ^sfnj + ^jks '^n, ^rk ^ns = Cs gnr + ^rks ^w
where the operators Mjjcs and Nrjc s involve only derivatives of lower order. Clearly, then, the right sides of (6.3.50) -> 0 strongly in the respective spaces. Thus, from Theorems 6.2.6 and 6.3.8, it follows t h a t each u^g -> 0 strongly in H^^'^^{G). But since u^ =2'^ts> ^î^ contradicts our assumption that the left side of (6.3-49) is 1 for each n. 6.4. Weak solutions In this section, we consider solutions of equations of the form dx = 0,
(6.4.1) G ? = 1 lal^-î
fc=l \^\<^QiTc '
v^^ CîG),
Qjjc = tjc + Sj mj
6.4. Weak solutions where we shall assume t h a t the Sj, tjc, and MJ satisfy (6.4.2) min(sj + fe) > 1, inm(tjc + Sj MJ) > 0, minw,- > 0,
243
maxs^- = 0, and where we also assume t h a t there exists an integer ho which may he negative such that (6.4.3) It follows that (6.4.4) 2m = 2^ (sj + tj) >N,
9=1
min(/A; + ô) > 0,

N
min(Ao ^j + %) > 0. mj <sj + tj < 2m.
We note t h a t if the u^, a^^, and the f^ are sufficiently differentiable, then the u^ satisfy the equations (6.1,1) where LjkU^'= 2 2 (-iy^^^D-aîDûK fj= 2 {-\T^D-fi.
(6.4.5) For each XQ we define the operators LO;;A: hy (6.4.6) Loj'jcu^ =- I

\oc\=mj
2' (\^\=Q}jc
iy^âfi(xo)D-^û^
and then define the operators LQ and L^^ as in 6.1 6.3. We assume that the operator Lo is elliptic and shall denote the collection of numbers Sj, tjcy mj, ho, V, and a bound for the coefficients in (6.4.6) and for Lo(l)~^ for A real with |A| = 1 b y . In the case of variable coefficients, E' shall stand for the moduli of continuity of the coefficients in (6.4.6) and bounds for their derivatives and for the other af^ and their derivatives as required. We notice that the equations f[v,oc(a'='û,^ + hû + / ) + v{c^ ^,/S + du + / ) ] dx = 0
G
f [v,ocp(a'='û f^^) + v,oc(b'u ~f)

G
+ v{cu f)qdx
= 0
are special cases of (6.4.1) (here oc and ^ are single indices). As in 6.2, we begin by considering solutions u having support in a small ball BR and define the coefficients a|/^.^ by
'^lU^) - <ii. = <p{R-^\A) [rt^W - ag{x^)-\

' 0 , otherwise
where 99 is defined in (6.2.25). Definition 6.4.1. If ^ > ho, we say that the a's satisfy the h-conditions on a set F if and only if
16*
244
(i) ii h = ho and sj = 0, the aj'^ with \a\ = nij and \^\ = Qjk are continuous; and (ii) ii h Sj + \oc\ > 0 then the a^^^ c^-sj+icci-i f^py^ otherwise the af^ are bounded and measurable. The as satisfy the h ju-conditions on F if and only if the coefficients af^ with \a\ = mj and | ^ | = QJJC^CKF) a t least, those for which h Sj + \oc\>0 belonging to C^-'^+^'^îr), the remaining a/ s belonging to C(r) a t least. Definition 6.4.2. For h > ho, we define the spaces **H^(B2R) and **C^{B2R) of vectors u just as they were defined in 6.2 and 6.3 exept that now h may be < 0. We define the space '^H^{B2R) to consist of those vectors / = {ff} such that
ff \^Lq(B2R)
(^m^-sj+\oc\^B2R), ,
if if h~
h-Sj-^\oc\>0, Sj+\oc\<0,
We define the space *C*(52ij) to consist of those vectors/such that
j^C%^>+\''i(E2R),
\C%{E2R)
7i-s;+|al>0 '
if
, if
h-sj+\a\>0
h~sj+\cc\<0
h-Si + \a.\<Q
Definition 6.4.3. If /^ > AQ a n d / C *H^{B2R) define PhoR (/) = UR where (6.4.8) .

I |a|<:mi
or *C^(^2ij), we
and F\'^ is the Maclaurin expansion of -Ff out t o terms of degree af where af = 2m v \ if Ao s^ + | ^ | < 0 a n d erf = 2m + ho ~ si + $x\ ~ V \ \i ho si + \(x\ > 0. Using the methods and results of 6.2 and approximations to t h e / ? b y smooth functions we conclude the following theorem: Theorem 6.4,1. PJIQR is a bounded operator from '^H^(B2R) into **/fJ(jB2ij) and from *C^(E2R) into '^'^C^(B2R) with bound independent of R if q :> \, 0 < ju < 1, and h > ho. If UR = PhoR{f), then UR satisfies (6.4-1) with G = B2R and a'^^ replaced by ag;\ and / ? replaced by ft "" fj^" ^^^^^ fj^ ^^ i^^ expansion of f^ out to terms of degree ho s^ + -\- \(x\ V \ {if this is > 0). Moreover if u has support in B2R and satisfies the same equations with the f*'^ ^ 0, then u^ U\ ^ 0 or is a polynomial of degree < max(Â; + ho - v \ytjc + si v $. It is understood that the ^g^^ are defined in (6.4.7)-
6.4. Weak solutions
245
The last statement follows since the mollified functions u^ satisfy the same equations as does u with t h e / ? replaced by their mollified functions fi^; of course Q is assumed so small t h a t UQ has support in B2R (we may work in B2R+d with ^ < ^). Then
^0
Tkl
IK(x
- i)LoimU^{i)di]=
Ll'Loimf
K{x -
^)i,^(^)d^
J J>2R
!a
\^mj
{6.4.9) since UQ has compact support. Using (6.4-9), integrating by parts and letting Q -^0, we obtain (6.4.10) u^(x) = Lf 2' (-i)'''i^î^r(^)
where Ff is defined in (6.4.8). The result follows by comparing (6.4.10) and (6.4.8) and calculating the degrees of the polynomials Lg^( 1)l'^'Z)'Fj\ Remark. We could have defined PhoR(f) by defining Fl^" as the expansion of Ff out to terms of degree 2m v 1. This would give a definition of PR which is independent of ho and, moreover U% would satisfy the equations without polynomials / j * . However, the presence of the ho and the polynomials J^ is not a serious drawback and their presence is necessary for the treatment of boundary values. Definition 6.4.4, We define the coefficients UR as in (6.4-7) and define TR on the spaces **i7j(J52i?) and **C^(B2R) by (6.4.11) TÛR = UR = PnoR(fR), fU = . - <.) D-<-
Remark. TR obviously depends also on ho. Then if u has support in BR and satisfies (6.4.1) and if we define UR = PhoRifit) where/R is defined in (6.4-11) with UR replaced by u, we see from Theorem 6.4.1 that
(6.4.12) U=:UR + HR
where H\ is a polynomial of degree < max(fe -}- ho v 1,fe r 1) and the equations (6.4.1) are equivalent to (6.4.13) UR-TRUR = VR = PUORif) PkoR[ - 0 D^HR]. The remaining developments of 6.2 can be repeated to yield the following results: Theorem 6.4.2, If h'> ho and the a's satisfy the h-conditions on B^, then TR is a hounded operator on '^*H^(B2R) for each p '> \ and each R < Ajl with bound e(R, p) where e also depends on h, ho, E, and E' hut 8(R, p) -^0 as R -> 0. / / the a' s also satisfy the h ju-conditions, then TR is hounded on **C^(B2R) with 6{R, ju) of the form C{ILI, h, ho, E, E') x X i^^, a = ju or \ according as h = ho or h ^ ho.
246
Theorem 6.4.3. Suppose the a' s satisfy the h-conditions on G, u^ ^ H\^+h^[G) for some q> \, the f^^Hl-'^+^^^'iD) if h - sj + \oc\ > 0 and fj $ Lj)(D) if h Sj + \(x\ < 0 for each D G O G and some h > ho and p ^ q, and u satisfies (6.4.1) on G. Then u^^ H^^^^{D) for each DC C G. If, also, the a' s satisfy the h ju conditio7ts on G and theff^ C'*~^^+I'Î(Z)) ifh Sj + \oc\>0 andff^ C^^(D) ifh-sj + $x\^ 0, then u^^ Cy^{D) for each D C G G. The proof of this is just Hke that of Theorem 6.2.5. The equations replacing (6.2.27) for Us are of the form
k l/S|<ei/t (6.4.14)
dx = 0,
v^CriB^R)
g!s=Csf^ + k%,
where if | ^ | = mj the kf^ involve only derivatives of u^ of order <^;A:The k'jg are not uniquely determined b u t a definite set can be obtained by replacing v^ in (6.4.1) by Cs '^^ and writing
D- {Cs v^) = Cs D- v^ + [D- {Cs v^) - Cs D v^

and then shifting the Cs factor to the bracket in (6.4.1) to obtain
Lfe1 1 / 5 k 11 / 8
Thus the equations (6.4.14) must be equivalent to
X [DO (Cs u^) - & DO u'''\ + 2 2 C^" (fs "^) - fs D- vi] X X \22atuDûi'-mdx
= 0,
v^Cr(B^R).
Theorem 6.4.4. Suppose the a' s satisfy the h-conditions on B^. Then there are constants R2'> 0 and C2, depending only on ho, h, q, E, and E' such that ifuis a solution 0/(6.4.1) withu^^H*l+^(Bii), a n d / ^ "^HîBji), then **\\ u lU ^ C2 [ * | / | | J B + **'|| u WU. * * ' i nn = 2 R-''-"-' I * liJj., I
k
0<R<R2, Q=v(q$lq. Also there exist constants C3 and R^, depending only onho, h,jbt {0 <C JLI < 1), E, and E' such that if u is a solution 0/(6.4.1), u^ '^*C^{BR) and has support in BR and if f^ *Cg(5ij), then
**NU^ 1 1^ <r Cc r*lll f lll'^- u **'!IU^ l'i|oi **'l!Uy No^ J?Q-V-/^ . **'i| ^1 0 1^ ' /1 \\\U\\\^R S=, ^21 \\\j\\\txR\^ |,||^|||li?J' .!i!^|||lR ^ l^lll-KRemarks. It is clear that we may allow thefj with h Sj -\- \oc\ < 0 to be distributions in some properly defined space H^s^+l^l-^ such
6-4. Weak solutions
247
spaces have been defined and used extensive^ in the study of differential
equations (see HORMANDER [1], A. FRIEDMAN [2], LIONS [2], LIONS a n d
MAGENES, and many others; see Chapter 1). We could say that 99^ H~* if and only if cp =^ A* (p where (p^H\\ no doubt spaces C~* could be defined similarly. Moreover the coefficients a'jl may be allowed to be somewhat more general as in 5-5 and 5-6. We now wish to consider weak solutions satisfying general boundary conditions in some weak sense and would like to obtain global results like those in 6.3. We shall consider boundary conditions of the form
(6.4.15) (Z I DyCnBrJcyU^-gry]dS==0, C ^ C^ {G)
where the operators Bricy are of order ^tjc hr pr and where, if {x^ y) are boundary coordinates the equations (6.4-15) reduce to j Z
(6.4.16)
IDlCîx)
[Bricy{x, Da,, Dy) u^x, 0) - gry{x)] dx =: 0

C'C-{G2R).
We assume that the integers pr and ho satisfy (6.4.17) as well as (6.4.3)We define Borkv operator of order the principal part
(6.4.18)
min^r > 0,
minô + hr -\- pr > 1
^ ^ assume that Sj, tjc and mj continue to satisfy (6.4.2). = 0 if \y\ <pr) if I7I = pr, we define B^rkv as that tjc hr pr obtained by replacing the coefficients of oi Brjcy b y their values at the origin. Then we define
BorJc =
\y\ = 'Pr
2(-'^)^'DyBorJcy
We suppose that for each point on the part of dG in which we are interested, that the operator B^rk above and the operator LQ satisfy the root and complementing conditions 0/ 6.1. Remark. If Brk{x, y\ B>xy Dy) is any operator of order fe hr with sufficiently differentiable coefficients of degree <tjc hr pr in Dy, it can be written in the form Z{-^)^''^DlBrjcy{x,y;D^,Dy)
\y\^Vr
in which the operators Brky are of order <.t]c hr pr- This justifies the restriction of our attention t o boundary conditions of the form (6.4.16). Definition 6.4.5. We define the space *'^J(c72i?) [h > Ao, 5 > 1) t o ^ consist of those vectors gry for which there exists a gry^H^Q^^^'^'^{B2R) if /^ + A r + l r | > 1 and HIQ{B2R) if h + hr + \y\ < \ such t h a t grvi^, 0) = gry(^) on a2R, and define the norm by
*'l|g||J,2=inf/ Z 'll^rrir.V"" + i: R'^''^^''-"\\grr\\l,A. [h-i-hr+lyl'^l h+hr+\y\<l J
248
We define the space *'C^(a2ij) to consist of those gry^ Cg+'''-+'^l{a2i2) in case h + hr + \y\^ \ and gry^ Cl^[a2R) otherwise, with norm
We define the space *'C^(a27?) in the same way as *'C^(o=2i?) but replacing 1 by 0. Definition 6.4.6. F o r ^ in one of the spaces above, we define PhoR{0, g) = UR, where
u^l,{X) = fL^r(^x~
ly)gry(f)
di
{r not summed),
Uy'^ is the Maclaurin expansion of ^gj out to terms of degree h hr v, and the L^^ are defined in (6.3.24). Theorem 6.4.5. PhoR is independent of ho and is a hounded operator from *'H^(a2R) into *'^H^{G2R) and from either *'C^(a2i?) or *'C^(a2R) into **C^(ff2/2) îth bound independent of R. In any case UR is a solution of the homogeneous equations (6.3.7) for y > 0 and satisfies the boundary conditions (6.4.16) with Brky replaced by B^rkyProof. Using the methods of proof and the results of Theorem 6.3-3 we conclude that Pn^R is a bounded operator as stated. For y > 0, we see from Theorem 6.3.2 t h a t
(6.4.19)
k
i:Ujicuii{X) = o, y = i,...,iv
fc
for each fixed r and 7. Now LoJA; is an operator of order tjc + s; and so the degree of ^ LQJJC U*^'^ is ^ Sj hr v. If this is negative,
(6.4.20) ILojjcu;^'(X) = 0.
k
If it is positive, then tjc hr v'> tjc + Sj so that

k
vanishes like X^ where g = \ Sj hr v. In this case (6.4.20) follows on account of (6.4.19). Now, suppose the gry^ C^{O2R)' Then from Theorem 6.3.3 and its proof we conclude that
(6.4.21)
k
?^-<(-') = {Lw; I T=r.

lBosku;^'(X)=^0.
Also ^ Bosjc ^J?*" (X) is of degree hs hr v. If this is negative, (6.4.22)
6.4. Weak solutions
249
If it is positive we conclude (6.4.22) using (6.4.21) and the fact t h a t [ul'!^(X) 14*!''^ (X)] vanishes at the origin to the power 1 + hg hr v. We wish now to define PR{f, 0). By virtue of Lemma 6.3.3, we may as well assume that / is already extended to B2R and we then use the norms defined in Definition 6.4.2. Definition 6.4.7. F o r / i n the space *C^(G2i2) or *HI{G2R), h > ho, we define (we omit the subscript R)
F]'{X)=dl'(~\y'^D'^l{X), %t{X)=fK(X-S)fl(S)dS,
(/, e not summed) f = rf
W^^^[X)=Z
g%{X) =
li-^r^Dlf
BorlcyV^^^{X),
ZCpLY{x,y;i)Dlgry{lO)dS
qri = m3ix{h + hr + pr "^, hr + pr + Si \e\ "i) where i^J** is the Maclaurin expansion of-F|^ out to the terms of degree 2m l^l r 1 (if this is > 0, F p * ^ 0 otherwise). Theorem 6.4.6. PR(f, 0) is a bounded operator from *C^(G2R) into **C^(G2i?) and from *H^{G2R) into **^J(<^2i?) ''îth bound depending on E and not R. If Y = PR(f, 0), then Y satisfies the equations (6.4.1) "i^îth a replaced by a^ and satisfies the boundary conditions (6,4.16) with Br icy replaced by Borky ^^^ gry ^y 0. Proof. By virtue of Lemma 6.3.3, we may as well assume that / *C"J^(^2i?) or '^Hq{B2R)' It is easy to see, using the carefully constructed definitions of the norms that the mapping via PR from / t o F is bounded as stated. Let us define Fo, F * , go, and ^ * by F = F o - F * , ylcle^ _ Lg*J^J^*, g^go~g* gly' = Borky V^^^''.
Then we see easily by checking the degrees of the polynomials that the vector F^^* satisfies the homogeneous equations (6.3.7). If, for some (l> )>fi^ C7{B2R) it follows that V^' satisfies the equations (6.3.1) with a replaced by ao and //^ = ( 1 )iI D'fl' d\ (1,8 not summed).
By virtue of our hypotheses (6.4.17), we see t h a t h + hr-\- pr hr + pr + Si-\e\-\=(h \>ho + hr-\- pr \>0 \8\) + hr + Pr-i)-(h-Si+
250
and qri > 0 in all cases. Also, in all cases, gj^ behaves at oo like a function which is essentially homogeneous of degree hr + pr + si \e\ v ~ \ and gj.y* is a polynomial of degree < this. Since this < ^r; r in all cases, the polynomial g^^ may be neglected in the expression for W if desired. Since Lf[x, y\ | ) is, for a fixed | , of the form G\(X) G^*(X), where Go satisfies the homogeneous equations (6.3.7) and GJ is a Maclaurin expansion of Go, it follows that W satisfies the homogeneous equations (6.3.7) for 3/ > 0 where it is analytic; the convergence at 00 of the integrals follows since L^ has degree 1 in 111 and D^ ^J^ has degree < 1 in ^ | | | at CO. A proof like t h a t in the first part of Theorem 6.3.3 shows that the mapping f r o m / t o W via Pnif, 0) is bounded as stated; one sees that ^ r r ^ î'^'^îÂ) for every A with the expected bounds. Now for some (Z, e), let us assume that / j ^ C^{B2R). Then the gly C'^{BA) for every A and W^' can be seen to ^ C'îW+i)- Let us approximate to ^J.^ by gly^^ ^Ti^v+i) using the device in the proof of Theorem 6.3.3. Then the derivatives D^W}^ converge as desired in Rv+i to \/^W^^ and we have (omitting [I, e)) Wl{X) = Wi,(X) Wl*{X), grn
lyl==2)r
^'L{-\r^m'gryn
Pfô(Z) =
/L*'-(^-|,y)^,(f.O)^f,
and TF^^^* is the Maclaurin expansion of W^^^ out to the terms of degree tiz V \ -{- max (A, sz | |) =^ tjc + qri hr pr v- By the method of proof of Theorem 6.3.3, we conclude t h a t BorJcWUlO)=grn(lO). Thus, by passing to the limit, we see t h a t Boric W^'^(^, 0) = ^[^(1, 0) - ^[^* (I, 0) where gl^ * (X) is the Maclaurin expansion of gl"" {X) out to the terms of degree qri pr v. Since the polynomials gly^ could be neglected in the formula for W, we must have ^J.^* ^ 0. The results follow. From here on the developments parallel those of 6.3. The theorem corresponding to Theorem 6.3.5 is: Theorem 6.4.7. Suppose u^ ^ HI^"'^{GA) for each A and each k, u^ ^ C for \X\ <, some RQ and y > 0, and suppose u satisfies (6.4.1) with a replaced hy a^ and f^ replaced hy ff^ and satisfies (6.4.16) with Brky replaced by Borkv ^^^ Sry replaced hy gf^ where ff"- andgf^ are polynomials of degrees < Ao S; + 1^1 V 1 and ho -\- hr + pr v 1, respectively. Suppose also that u satisfies \Dlu^(X)\ < Cp(1 + iXlY'^+^^-^-^-^+^f^ ^ = 0, 1, . . ., \X\ > Ro. Then each u^ is a polynomial of degree ^tjc + ho v 1.
6.5- The Dirichlet problem for strongly elliptic systems
251
Remark. Since if UR = PR(f,g), then UR satisfies (6.4-1) with a replaced by UQ and satisfies (6.4-16) with Brky replaced hyBorjcy with no residual polynomials / * and g*, we need Theorem 6.4-7 only for the cases / * = g* = 0. By paralleling the previous developments we may conclude the following results: Theorem 6.4.8. Suppose ho is the smallest integer satisfying all the conditions of this section, suppose h > AQ, suppose the a' s and the coefficients in the Br icy satisfy the h-conditions on a domain Fz^G, suppose that for each point XQ of G the operator LQ is properly elliptic and that LQ and the operators Bork satisfy the root and complementing condition for each XQon dG, suppose G is of class Cl'^^^~^{to = maxfe), and suppose u satisfies (6.4-1) and (6.4-15) "i^^here f^^H^{G), Q = max(0, h Sj + \oc\), gry^Hl(G), T =: w a x ( l , h + hr+ \y\), and u^^ Hl^+^{G). Then
k [j.oc r,v k J
where C depends only on h, q, G, E, and E'. If the a' s and the h' s satisfy the h ju-conditions on rz:)G, G is of class C*^^^, u^^ C*^'^^{G), ff^ Cf,{G), andgry^ Cl(G), r = max(0, h + hr + \y]. Then I
fc
I I I ^ ' \ m ' ^(^\I
L?,a
U\\\U
+ I
r.y
\kry\ila
I\\u'II?.
where C depends only on h, ju, G, E, and E'. If G, ff, gry, and the a' s and h' s satisfy the conditions in the first statement for some h' >h and q' >q and u^ ^ H^^^'^[G), then u^ ^ H*"^^^' (G). U ^y ffy gry> <^^^ i^^ ^' s and b' s satisfy the (Holder) conditions in the second statement with some h' ^h and ifu^^ H^^^'^[G), then u^^ Cf^^'^^' (G). If Gy fjy gryy ^^d the a's and b's^CîG) and u^^Hl'+îG), then u^^ C^(C) 6.5. The existence theory for the Dirichlet problem for strongly elliptic system It is clear that the results of 6.3 imply that for every A, the set of solutions of the homogeneous boundary value problem (6.5-1) LjkU^ + {\)'^û^ = 0 onG, BrkU^=0 ondG,
form a manifold of finite dimensionality, provided the Ljk and Brk satisfy the supplementary and complementing conditions of 6.1. It is evident (since for one equation we regard L^ ^ as the identity) that the Dirichlet boundary conditions are always complementary for a single equation of order 2 m which satisfies the supplementary condition. However the following example due to SEELEY and communicated orall}' to the author (essentially) by F. BROWDER shows t h a t every complex
252
number A may be an eigenvalue: Letting {r, 6) be polar coordinates in the plane, we consider the operator
on the annulus TT < r < 27r. For each complex A, the function U(r, d; X) = Jo{w) sinr, w = X^''^ e'^ is an eigen-function for the given eigenvalue A, Jo being the Bessel function. That is, U satisfies LU -XU = 0. In this section we present the existence theory for the Dirichlet problem for a strongly elliptic system, as defined by NIRENBERG [2]; in this problem, the eigenvalues are isolated. The developments parallel those in 5-2 and 5-6. In discussing strongly elliptic systems, it is convenient to change the notation somewhat from that used in 6.16.4. We assume t h a t the operator Ljk is of order Sj + sjc where we also assume (6.5.2) sy>1, y = 1 , . . . , TV; thus m = si + h SN ^ N. In the previous notation tjc was the maximum order of any of the Ljic for fixed k. Thus if we define s = maxs;, then (6.5.3) 5 = maxs;, tjc = s -\- sjc, old Sj = s + new Sj, where tjc refers to the previous notation and the Sj are new. Definition 6.5.1. If we let
LM^, D)= z <. W D",
L;^
= z ?. ^".
L'y^ denoting the principal part of Lj^, then the system is said to be strongly elliptic if and only if
(6.5.4) Ia|=sj+Sfc ?=1
x^G, (6.5.5) V^-iw*^ = 0,
A real, ^ complex. r = \, .. ,,S]c on dG.
A vector u will be said to satisfy 0 Dirichlet data on 5 G if and only if In order to fit these special boundary value problems into our general framework, we replace the index r by the pair {j, r); then the operators Brk, in [x, y) coordinates, are
fo
, if y ^ ^
Since, in the old notation, Brt was of order tjc hr, it follows that (6.5.7) hjr = s + Sj+\-r.
253
In order to prove our existence theorems, we begin as in 5.2 by assuming, in addition to (6.5.4), that
(6.5.8)
afjc^C[-^-''^-^{G), | a | > s j , + 1,
a'^j^ bounded and measurable, all oc,j, k.
In that case, it is possible to integrate by parts to eliminate derivatives of u^ of order higher than S]c and thus obtain (6.5.9) ^ ^ *
The integrations by parts are not unique but we shall assume that the)^ are done in some fixed way, obtaining the definite formula (6.5-9) for B (u, v). However these integrations are performed, it follows easily that
(6.5.10)
(6.5.11)
SA?i{x)?if>i^''
= z
Bjcrjui
2'r*Î^J*= 0 ondG
With these assumptions, the equations LjjcU^ + ('^)'iU3' =fjy B{u,v) ^ '^*^^^ C(u,v)= may be solved in the weak sense by solving the equations + A C {u, v) = L (v) f 2Ju^v^dx, L{v)= j2;{~\Y^f^vJdx
G ^
G ^
in which B (u, v) is defined in (6.5-9) and G need only be bounded. As in 5.2 we first prove the following theorem: Theorem 6.5.1 (Garding's inequality). We suppose that G is bounded and the coefficients Af^ are bounded and measurable and that those whith \(x\ s; and \^\ = sjc are continuous and satisfy
(6.5.13)
ReZ
^'^ |a|=sj i/5|=Sfc
IAî{x)Lk&'^^>M-^Z\M^'A&\^>
i
M > 0,
A real, f complex.
Then there exist constants Mi and Ao, depending only on v, M, the Sj, G, bounds for the coefficients, and the moduli of continuity of those Af^ for which \oc\ = Sj and |/51 = sjc, such that |5(^,^)|<Mi.||^li.||t;||, ^ ^ ReB(u,u)^ u,v^ô mi = Af-i (mill) || u [p - XQ C {u, u),
where the space ^ 0 ^5 the space of vectors u in which u^^ H^^[G) and (6.5.15) ( ^ , ^ ) = / 2 ' ^CaD^'u^D'^vUx, Ca = \(x\\l(x\.
Q J \<x\ = Sj
Proof. It is sufficient to prove this for u, v^ ^TiQ- The existence of Ml follows immediately from Poincare's inequality (Theorem 3-2.1). To
254
prove the second inequality, let e > 0. G can be covered by a finite number of spheres B {xi, ri) such that \A'^^{x) A^^(xi) \ < e whenever x^ B(xi, Yi) and \(x\ = sj and |j^| sjc. Clearly there is a sequence {Ct}, t = \, . . ., S, in which each f^^ C^{Bv) and has support in some one B (xi, Vi) and Cf + * * * + f| ^ 1 on G. Then, as in the proof of the corresponding inequality at the end of 5.2, we see that ReB(u,u) =ReB'{u,u) + Re f Z I
Q t | a | = s j |<8|=Sfc
^^Â^^DûWû^dx
= ReB''(u,v) + RejZZ
2 ^?i D^^D^ ^"l dx,
ul^^Ctu^
(6.5.16) ^ * '*'^*^ '^'^^'^ where B"{u,v) is a form like B(u,v) in which, however, "Af^Ô whenever \(x\ = Sj and |/9| = sjc. But now the last term in (6.5.16)
-R^flS + Re Zi:
G t \oc\=sj l/Sl=Sfc
ZAt{^t)D-u\DPu\dx I[Ag{x)-Afi(xt)-]DuiDûidx.
Clearly the absolute value of the second term in (6.5.17) (6.5.18) < Zi(r, si, . . ., SN,N)'8' Z WM^t
If we introduce the Fourier transforms ul of uf by their usual formulas (see between (6.2.4) and (6.2,5)), the PLANCHEREL theorem shows that the first term in (6.5.17)
(6.5.19)
=flReZ
ZAîy^y^ûfdy^mi:S\\M^-
Now, using the fact that Vî w| = Ct VîU3 + Mlu^, where the Mf are of order Sj \, and using (6.5.16) (6.5.19), we see that Re B {u, u) > (mi Zi e) |1 u f Re B'" [u, u) where B'" is a form like B". Almost exactly as in 5-2, we conclude the following two theorems: Theorem 6.5.2. Suppose the transformation U is defined on 0 (see Theorem 6.5.I) by the condition that (6.5.20) C{u,v) ={Uu,v)^Q, v^^Q. Then U is completely continuous. Theorem 6.5.3. If X is not in a set g which has no limit points in the plane, the equation (6.5.12), with L {v) defined by the more general expression (6.5.21) L[v) = [2: Zf^D-vJ'dx,
has a unique solution u in ^^ for each set of ff in L2{G). If X^î, the manifold of solutions u of the equation with L = 0 has a finite non-zero dimension. It is assumed that the coefficients satisfy the conditions of Theorem 6.5.1.
255
Now we notice t h a t the equation (6.5.12) with L{v) given by (6.5-21) is of the form discussed in 6.4 with
(6.5.22) h = S, Mj = Sj, hjr = S -}- Sj -\- 1 r, pjr 0 .
Consequently we may conclude that if the afj^ satisfy (6.5.8) the / , ? $ i7r^+li(G) and G is of class Cf-i, then the u^^ Hl^'^(G) H - ^ | J ( Q and the integrations by parts yielding (6.5-9) may really be performed (in reverse) to yield the following theorem: Theorem 6.5.4. Suppose the a"-^ satisfy (6.5-4) and (6.5.8) and G is of class Cf *~^. Then if 1 is not in a set g which has no limit points in the plane, the equations (6.5.23) LjTcU^ + {^Yiui =-fj onG have a unique solution u with uJ ^ H^^^^ [G) f\ H^^ (G) for each f for which each p^ H2~^^(G). If X^^, the manifold of solutions of the homogeneous equations has a finite nonzero dimension. Of course, higher differentiability follows from more assumptions about the coefficients and the fj. We shall not state these since it is possible to eliminate many of the assumptions made in (6.5.8) about the coefficients, as was done in 5-6. We prove below a theorem (Theorem 6.5.6) corresponding to Theorem 5-6.4, but only for systems in which 5i = 52 = sjv = w > 0; the more general systems introduce difficulties. We must first prove the following theorem which is strongly suggested by the results above and has been proved in AGMONDOUGLIS-NIRENBERG [2].
Theorem 6.5.5. / / the system (6.1.1) is strongly elliptic, it satisfies the supplementary condition and the Dirichlet boundary conditions are complementing. Proof. That the matrix Ljjc satisfies the supplementary conditions follows from the fact that the set 5 of coefficients a"-^ satisfying (6.5-4) is convex and contains the set
(6.5.24
a% = 0 if j ^ k or if j = k and oc i^ 26 with
?A -^ -^ t-
\t^\
li^l = Sj
J
obviously the coefficients above correspond to the system

A^JU3 =fj, y = 1,.. .,N.
Clearly the determinant L{X] x) of the matrix || a^^. Xoc|| cannot vanish for any real X or there would exist a complex f so that contradicting (6.5.4). The rest of the supplementary condition requires proof only if 1 = 2. Then if G and T are orthogonal unit vectors then the ^ roots z oi L{a + zr\ x) = 0 are never real and vary continuously with the coefficients as they vary over S.
256
To show that the Dirichlet conditions are complementing, it is sufficient, according to Theorem 6.3.1 to show the existence of exponentially decaying vector solutions U^^ of Lojjc{a,~iDu)Ul{a,u) (6.5.25) where, here, we are using the {x, y) notation of 6.3. These equations made non-homogeneous are ordinary differential equations of the form (6.5.26)
fc=l
= 0,
Dr-i Uf,{a,0) = df dl,
\a\ == \,
2 " Ibjjcp(a)i-iDu)i'UHa,u)
29=0
Sj+sjc
=fj{a.u)
where (6.5.4) implies that (6.5.27)
Re 2]
i.k
Sj + Sjc
2 =0 >
|^*c?^> W i 2 ' ( l +iW2)s, |î|2, i [\o\ = 1), mi = M - i .
Let us now define H^ as the set of all vectors U (henceforth we suppress a) such that U^^ Cîc-i with Z)fc-i U^ absolutely continuous and satisfying (6.5.28) Dl m{0) =0, p = {),. . .,sic\ and with norm defined by
o
(6.5.29)
S j
\\Uf=j'^'î^'\\DvUyfdu<o<..
For U and V having finite norm, we define B{U.V) = ( ^ \ v i j:hicv{-iD)vm (6.5.30) 0 '-^^ """
we also define Bl{U, V) similarly with [0, 00) replaced by [a, h]. Now, let us introduce the Fourier transforms U and F of ^ and V in HQ. Then, for such U and F , we see that (6.5.31) B{U,V)= Z lhicvZvU^[z)V^{z)dz.
From (6.5.27) and (6.5.31), we see that B satisfies the hypotheses of BQ in the Lemma of Lax and Milgram (Theorem 5.2.2). Consequently if fj $ C^~^5 [0, 00) and vanishes for i/ > 1, say, there exists a unique U in ^ 0 such t h a t
(6.5.32)
B{U, V)=I
2 ' / ; V^'du,
V^Ho.
257
The usual difference quotient procedure shows that U^^ C^+^fc and satisfies (6.5.28) and (6.5.26). So, to find U\^, let a)\^ be any function ^ C^^^k [0, 00) and vanishing for w > 1 and satisfying Then, let U be the solution above with
fl = 2
Then we may define
Sj + Sk i:hicv{-iD)P(4^.
39 = 0
Uft = o>fe - UK The following theorem and its proof are essentially due to N I R E N BERG (see AGMON-DOUGLIS-NIRENBERG [1], p. 693):
Theorem 6.5.6. Suppose that G ^ Cf^~^, that Sj = s for every j , that the a-^ with \(x\ =^ 2s ^C^[G), the other coefficients being hounded and measurable. Then the conclusions of Theorem 6,SA hold, In fact, there exist real numbers Ao and C, which depend only on v, G, E, and E' such that
h J \ k J
u^^Hf{G)r\Hl^{G), A real. Proof. We first prove the last statement. The proof of the first is like that of Theorem 5.6.5. For any given r]i > 0, each point X{) is in a neighborhood or boundary neighborhood in which \a[x) a (:Vo) | < iqi where a {x) = {ajic {x)} with 1^1 = 2s. We choose a sequence f^, ^ = 1, . . ., T, such t h a t each Ci ^"^ [G) with spt f i in some one such neighborhood and Cf + + C| = 1. We let u^ = ^t u^ and note that where ZJj^ is the principal part of Ljjc and Mjict is of order < 2s. Then, using (/, g) to denote the L2 inner product, we obtain (6.5.33) ^ ^''* = 2 ' [ ( - ^Y{L]k < uf) + {M;,, UK ui)]
where M'^-^^^ is of order < 2s. But {6.5.34) '

i
'
where Z^-^^ is the operator with the constant coefficients afj^{xt) where 1^1 = 2s and Xt is in the small support of Ct- Since for each t {ci. (6.5-9), (6.5.17), and (6.5.19)) ReZi(6.5.35) ^
^Y{L-kt<
4) =Rej
Afi(xt)
i
D-
u{D^<dx
G9,k.ocJ \n
Morrey, Multiple Integrals
258
we see, using (6.5.33)(6.5.35) that

\\2s.
(6.5.36)
C(?i)2"lh'ÎP-Ml*^ll''> - 2 2 ? i i F M l | i o
j,k
Z3(ir)2.
But now, for real 1, (6.5.37)

j \ k I j \ k J
From Theorem 6.5.5, it follows that the boundary value problem satisfies the conditions of 6.3, so t h a t (6.5.38) (1*^112^)2 < Ci f2'(ll^;fc*IIT + (IIÎITl-
Using (6.5.36) and (6.5.38), we see that (6.5.37) becomes
>(i-z2îCi)2'(iii;fc*'n2
+ [A2(l - Z2V1) - 2AZ3 - Z2V1 C I K I M H ^ 2
SiWLjicW'Wr + ^^HMr
if A > some Ai.
The result follows easily from this and (6.5.38). Higher differentiability results follow from 6.3. 6.6. The anlyticity of the solutions of analytic systems of linear elliptic equations In this section we carry over the methods and results of 5-7 to apply to systems of the type (6.1.1) on the interior and to those of the type studied in 6.3 at the boundary. The developments follow those of MoRREY-NiRENBERG except t h a t we treat general boundary conditions. For our results on the interior, we consider solutions on a sphere BR. Since we already know that the solutions ^ C^, we shall assume that
U^C^(BR). We define
eAf,Br)
(6.6.1) dj) (u, Br) =
/'2'|V-^.-+^/,-(:^)|2^^
Br ?
1/2
, p'>So
= minSj ( < 0 ) ,
1/2 Br
^ > t, t = max fjc
^ '
where, in general, \/Q(p = 0 if q < 0. With [p\] given b y (5-7.6), we define MR, P if) = [P l]-i sup (7^ - ry-^P e^ (/, 5 , ) , p ^ s^ ( > - ^) (6.6.2 NR,^{U) =.-. [p\]-^ sup (R-rY+P
R/2^rt.
6.6. Analyticity for linear systems
259
As in 5.7, we shall show that (6.6.3) iVij,p(w) < M L ^ p ^ - t for some constants M and L. We begin b y studying equations of the form (6.6.4) L%u^=fj, y = 1,...,iV, where the L^j^ are operators with constant coefficients and involving derivatives only of order Sj + tjc. We use the notation of 6.2 until we start considering boundary value problems. Since, for eachy, there must be a t least one non-zero operator Z^^, we must have
(6.6.5) SO + J 5 > 0 .
Lemma 6.6.1. Suppose u ^C^{BR) there is a constant K\ {v, E) such that do{u,BR)
and satisfies (6.6.4) on BR. Then <.Kieo{f,BR),
Proof. Multiplying both sides of (6.6.4) b y K{x f) (see (6.2.11)) and integrating with respect to | we find t h a t (6.6.6) L%m[x) = Fj(x), m[x)^ j K{x- | ) u^^(f) ^f,
BR
Fj{x)^lK{x-^)fj(^)dl
BR
Operating on both sides of (6.6.6) with L^' and summing, we obtain u^{x) = V^'Fj{x) from which the lemma follows, using Theorem 6.2.1. Lemma 6.6.2. Suppose u^C^ (BR) and satisfies (6.6.4) on BR. Then there is a constant K2 [v, E) such that, for 0<Cr<Cr-^d<.R, r>d,
do (U, Br) < i^2 U (/, Br^s) + Z^~^ ^-Q (/> ^r+s) + Z ^~^ ^-Q (^ > ^rÂ '
(6.6.7) Proof. Let 9 be befined as in (6.2.25) et seq. and let 9
{6.6.S)
C(^) = ^[^"Mkl - ^ + ^)].
U^{^) = C(^) ^^(^).
Then U has support on Br+6 C BR and satisfies

Sj + fc
(6.6.9)
L%m:^Fj = Cfj+SIV9C-L^jcau''
for appropriate operators Ljjcq. Moreover

(6.6.10) V-^/F;-=CV-^.7^-+^V*:->/ZM/;+ 2 ;
q=l
k q=l
^V^C'MjTcqui^
where Mjq and Mjjcq are appropriate operators with constant coefficients of orders Sj q andfe ^, respectively, being zero if these integers
17*
260
are negative. Since (6.6.11) I V ^ C W I <.Z^[v,(p) ^-^, the lemma follows by applying Lemma 6.6.1 to the vector U and using (6.6.10) and (6.6.11). Lemma 6.6.3. Suppose u ^C^ {BR) and satisfies (6.6.4) on B^. Then there is a constant Kz [v, F) such that
NR,P{U) <KS\\MR,j>{f) +IMn,j,_,{f)
[
(6.6.12)
Q=i
+iNR,^_^{u)\,
a=i
p>0.
Proof. Suppose r is any number with Rjl ^ r < R and let (6.6.13) d^(R-r)l{p+ \) so R - r - d = [i -^'l'{R - r).
Then, if _> > 0, V ^ ^ also satisfies (6.6.4) w i t h / r e p l a c e d by V ^ / , so that / Lemma 6.6.2 holds with all the indices increased by p. Multiplying both sides of this result by (pl)~^ {R r^+P and using (6. 6. 13), we obtain (p\)-^R-ry+Pdp(u,Br) < (p\)-^K, [e^f Br^s) {R~rdY^p[\ - j ^ ^ " +
+ f^{P + ^)'e^_,(l Br^,) '(R-r-
^)^+^-<?(l - ^ ) " ' " ' ' ' ' ' +
from which the result follows by taking the sup. We turn now to the general analytic system (6.6.14) which we write in the form (6.6.15) L%^''= fi + {Lîic - Lik) *= = Fi where L^^ is the operator whose coefficients are just those of the highest order derivatives evaluated at the origin. We set (6.6.16) V-'i-^^'Fj =- V-'j-^^fj + 2 ' 4 l ( ^ ) Vh-^^-ûf^, 5o < A < 0, Ljjcuf^==fj
where the aj^, to be abbreviated by a^^, are appropriate analytic tensors. (We recall t h a t V^9? = 0 i f ^ < 0 . ) Accordingly, there are numbers ^ > 2, L, and RQ, such that in B^Ro-
{^'^"^y)
\\7â^^<^p\LAv, aOO(o)=.0, \a^^(x)\ <^LA '\x\,
pô
261
Applying the vector inequality to (6.6.16) and using Lemma 5-7-4, we obtain
\2:1 v-'i-^pF^|2ii/2 ^ i^; I v-'^+^fj|2ii/2 + 1 0 J L ? J \Z I v-^y+^7î 1211/2 < [2; I v-^.-^^/; :/><o. +2^1 ^^^W I [2'V^fc+2?-^^^|2]i/2^ |2]i/2 +
a=i L fc J
Using (6.6.17) and separating out the term where q = 0 and }c = p, in case ^ > 0, we obtain for 0 < ^ < i^ < -Ro (setting r = vj2), ep{F,Br)<plLAPyl^^r^ + (6.6.18) t V ,^\ + Z in'(p->c)\LAP--d,.,(u,Br), ep(F,Br)^[p\]LAPyl^^r' +
2)-l
pÔ, 2:L[(p->c)\]dp_,(u,Br),p<0.
Thus, from the definitions (6.6.2), we find that Mn,p{F) ^yl'^L{AR)pRt^^

t
+ ^L(AR)P--NR,4U)
p
(6.6.19)
+LAR'NR,P(U)+^^^^^^^L(AR)P--R'NR,,,-,{U)
MR,^[F)^yl'^LAvRP^t-^'
+ZLR'NR,P_,[U),
t+p
so<p
<0,
We now apply Lemma 6.6.3. If we take R so small that (6.6.20) KsLAR< 1/2, we see that we can solve for NR^ p {u) in terms of the non-homogeneous term and the NR^q{u) with q <p. The result is, for p > 0,
NR,^{U) < 2KJyli^LRt^^Z{AR)v- SQ (which is the general case), the double sum forp+ \ <.q < So is missing and q ranges from 1 to so in the triple sum. We note
262
< 2i-2> yl'^ LAô iR+%+^
if
^ i^ < 1 / 2 .
Thus, if we assume that (6.6.22) we see t h a t NR,q{u)<MRi+So-^'''PQ, t<q<p, P>~SQ
iVi?,pM < iK^R^+ôÂl^-Vyl'^LA^ +^ZML2^-vp^ + +Z ZÂ'^' 2^-^-^MP^-^ +2! Z "z"^^'^ 2Q+^-p-^MP'<-^ + {AR< 1/2)
+ VMP2>- + 2-iL2'P^"4 <^^-^^''ô+^P^
(6.6.23) if M and P are chosen so t h a t (6.6.22) holds for ^ < ^ < S Q 1; it is easy to see that P may be chosen > 1 and large enough so that the coefficient of M in the brace is < P ^ / 4 i ^ 3 and then M may be taken so large that the remaining terms in the brace are ^M P^j^K'^. We therefore conclude the following theorem: Theorem 6.6.1. If u is a solution o/(6.6.14) and fj and the coefficients in the opefators Ljk are analytic at XQ, then u is analytic at XQ. We now consider the analyticity at the boundary. Using an analytic change of independent variables, we may transform a boundary neighborhood along an analytic portion of the boundary into GRQ, the part of the boundary corresponding to ORQ. We begin by considering solutions u of (6.6.4) which satisfy the boundary conditions (6.6.24) (6.6.25) Borjc u^ = gr on an, R <Ro, where Bork has its significance in 6.3. We assume t h a t ^0 > 0, Ao + ^r > 1 for each r, ho and hr being the numbers in 6.3. We assume t h a t the boundary conditions and equations satisfy the conditions in 6.3 and we use the (x, y) and other notations of t h a t section. Instead of (6.6.1) and (6.6.2), we define (SlVô-îVl/jhr, P > 0. ep (f. ^') = j 2 ' I V *o-,+/j hr, I f2'l|V%+*oV?M*|k, (6.6.26) ^J.(^.Gr) = [|.||^,^+^^+^^,.||^^^ 0>p>so-ho, p>0, 0 > p > - t - h o ,
|2'||vAo+^v^|s|k,i'>o,
cp(g. Gr) = j^-ii v ^ o + V i ' ^ s ||G 0>p>-ho-h,h^m3.xK,
263
the norms being the L2 norms. Since all functions considered ^C^(Cr), we may take ^s ^"^ (^r) We then define NR^P(U) (6.6.27) = sup[p!]-!{R - r)+Ao+2>dp(u,Gr)^
R/2<r<B R/2^r<R
p^
- t ~ ho
MR, P (/) - sup [p !]-i {R - ry+ô-^P e^(f, Gr), QR,p{g) = sup [pirHR - r^'-'ô-^Pcpd Gr),
p ^ so - ho p^ -ho - h
We shall first establish the analog of (6.6.22). Lemma 6.6.1'. Suppose u, / , and g ^C"^ (GR) and have support on GRUOR, suppose g^C^ {OR) , suppose gs (x, 0) = gs (x), suppose u satisfies (6.6.4) on GR and (6.6.24) on OR. Then there is a constant K[{v, E) such that (6.6.28) do (u, BR) < K[ [eo (/, BR) + co (g, BR)] . Proof. We note that u, f, g, and g satisfy the hypotheses on any GA with A ^R and vanish outside GR in R^+i. Let U = PR (/, g). Then, from Theorems 6.3.6 and 6.3-5, it follows that u^ U^ is a. polynomial of degree <,tjc + ho v 1 and hence that (6.6.28) holds, since only the highest order derivatives of the u^ are involved. Lemma 6.6.2'. Suppose u, f, and g ^ C^ (GR) , suppose g ^ C^ (CR) and g (x, 0) = g (x) on OR, suppose u satisfies (6.6.4) on GR, and suppose u satisfies (6.6.24) on CR. Then there is a constant K'^{v, E) such that
{
(6 6 2Q^ ^^^ a=i
hoSo
eo{f> Gr+d) +2d-Qe_q{f *^^^ a=i
Gr+6) + Co(g,Gr+d) + ^ J
0 <r <r + d <R,
r > d.
Proof. Define C and C7^ b y (6.6.8). Then U satisfies (6.6.9) and

(6.6.30) BoskU^ = Cgs+'i:'7^C'BsJcqU^^ys
Q=l
on GR.
If we define ys b y (6.6.30) with gs replaced b y gs, we see t h a t ys{x, 0) = Ys (x) and, in fact U, F, y, and y satisfy the hypotheses of Lemma 6.6.V. Just as in the proof of Lemma 6.6.2 we find t h a t eo (F, Br+6) < eo (/, Gr+s) + Zi 2'V^^_g(/, oo(y, Br+d) <co{g, Gr+d) + Z2 2:d-^c__,{g,
g=l
a=i ho+h
Gr^6) +*i;Vî_^(^, Gr^s)

Q=i t+ho
Gr^s) +Id-^d_,(u,
g=l
Gr^s)
The lemma then follows from Lemma 6.6.1'. By differentiating in the tangential directions only and proceeding as in the proof of Lemma 6.6.3, we prove the following lemma:
264
Lemma 6.6.3'. Suppose u, / , g, and g satisfy the hypotheses of Lemma 6.6.2' on GR. Then there is a constant K'^[v, E) such that NR,^{U) ^KAMn,^{f)
I ho+h
+ZMR,p.a{f)
Q=l t+ho ^ )
+ QR,p{g) + p>o.
+SQR,P~Ai)+INR,p_,{u)\,
q=l q=l
Remark. In this proof, we use the facts that dp{u,Gr) = do(Vlu,Gr) if pÔ,e_q{V%fGr)êp_q(fGr),p,q>0,
Gr), p,qÔ.
C_q(V%g, Gr)<Cp_q(g,
To handle the general analytic system, we write it in the form (6.6.31) L%u^=Fj, BoskuJ' = ys = gs+{Bosk-Bsic)u^ on OR, where Fj is defined as before, in (6.6.15) and gs is analytic on GR with gg [x, 0) = gs {x). As before, we have
t+ha
\/hQ-Sj+^F.^^hQ-Sj+^f.
+â]%{x, y)Vh+h-^^-û^(x,
q=0 t+ho
y), 0 > A > so /ô
Vô-V^ Ys == Vô+V^gs + ZHk (^> y) V*A-+ô+^-û^ {x,y), 0 > A > ô -^ aOO(o) = bOO(0) =0, \a^^(x,y)\ <LA]/\x\^ +y^,
I b^^{x, 0) I <LA ]/\x\^ + y^,

with bounds like those in (6.6.17) for V^fj, V^ ^^^ V%b^^, V^gsUsing the previous analysis and the idea of the remark above, we obtain the inequalities (6.6.18) and (6.6.19) with t replaced by ^ + AQ and we obtain the corresponding inequalities for QR,p(g). As before, if we take R(<Ro) so small that IK^LAR <,\j2 this time, we can obtain an inequality like (6.6.21) with t replaced by ^ + ô and terms on the right coming from the analysis of y. An analysis like the previous one demonstrates the existence of numbers R> 0, M, and P such that (6.6.32) NR^P{U) <MRi+ô+'^ PP, p^t ho, 0<R<.Ri. Now, in order to prove our analyticity theorem, we must obtain bounds like (6.6.32) for all the derivatives of u. For this purpose we define dp,q{u,Gr) ^^ NR,P,^{U) =2;i|Z)t''"'""''VS^*^k, = snp[(p + q)\rHR p>:0, q ^ - t - h o Gr).
- ry^ô^P^^dp,,(u,
Using the idea of the remark above, we conclude t h a t (6.6.34) NR,p^q(u) < NR,p^q{u) if g' < 0 and j ^ > 0. \, We shall show that there exist numbers R2, 0 < R2 < Ri, M, P ^
265
and 0 < 1/2, 7?2, P, and d depending only on the bounds for the coefficients, / , and g and their derivatives, such that
If we differentiate the j t h equation Sj-\-a in terms of the other derivatives, obtaining (6.6.36) D^û^ =f-^ +2 ;iblfjx,
times with respect
to y, 0 < ( ; < A o , the eUipticity imphes that we can solve for D!^''^'^ u^ y)D'i^^-^\/tu\ 0 < ( 7 < Ao
where/<^^ and 5^^^ are suitable functions analytic near (0,0). We may assume that t h e / ' s and h' s satisfy (6.6.37) ^
for each set {a,Q,co). Differentiating (6.6.36) with a = ho and using (6.6.37) and Lemma 5-7-4, we obtain, f o r ^ > 0 and q >0,
k
e = l co=0 A=0 n=0
(6.6.38)
X y ^ l D*^^+'^+^-^ v?+^^^ |.
From (6.6.38) and (6.6.33), we obtain, f o r ^ > 0, ^ > 0,
(6.6.39)
Q,0},2.,f.l.
+IBli,^LAP^^-'-^NR,a>^,,^_,(u)RP^^-'-f'+^-^,
D2>Q ^ P^-q^' [(A -f // + C O g) !]
Now, from (6.6.34), it follows that (6.6.35) holds for all - < 0 and all ^ > 0, provided merely that (6.6.40) 0<R<^Ri and MPP+Q < M(P(9)^+^(9-ff 0 >:q'> t ho. for all p>0,
Then, from (6.6.39), we see that if (6.6.35) holds for all {p\ q') w i t h ^ ' > 0 and t ho<.q'<q{>^0), then it will hold also for all {p', q') with q''^q, if, for example, we choose AR2=--, Pd^P, P l 9 > 1 , 6><1, M > M , 22>+^M>2L-(2/l)-ô+ô, P > 1, ^ P 0 > 1, and 0 < 1/161. Since A was supposed > 2 , these conditions may be satisfied.
266
6.7. The analyticity of the solutions of analytic nonlinear ellipticsystems In this section we extend the results of 5.8 to general elliptic systems; the analyticity at the boundary is proved for the systems and boundary conditions discussed in 6.3. As usual, we consider analyticity on the interior first. We consider solutions u {= u^, . . ., u^) of systems of the form (6.7-1) (pj{x,Du) = 0, t=\,...,N in which the cpj are analytic for all values of their arguments near the values of those arguments along the solution at XQ. The system is elliptic along the solution in the sense that the linear equations of variation (6.7.2) Ljjc (x, D)vi^^~ %' [x,Du-^XD v]^^Q - 0
form a system of the type discussed in 6.2. We shall use the notations of that section. In (6.7.1), we may suppose t h a t XQ is the origin. In addition, we make the substitution (6.7.3) ui' =,vk j^pk^ k=^ \, . . .,N where p^ is t h a t polynomial of degree < fe such that (6.7.4) V'"P*(0) = V^^^(O), 0<r<tk. We therefore assume t h a t XQ = 0 and 0^ u^ (0) = 0 for | ^ | < fe in (6.7.1). If we expand the q)j about the origin, we may write (6.7.1) in the form (6.7.5) L%u^ = Mfjû^ + WJ(^> D^) where the operators L^^ have constant coefficients and are zero or of order Sj + h* the M^^j. have constant coefficients and are of lower order, and the ipj are the remainders. Since the derivatives D"^ u^ (0) = 0 if \oc\ ^tjc, we see by differentiating the j t h equation in (6.7-5) up to Sj times, t h a t the Taylor expansion of ipj begins with a homogeneous polynomial in x of degree 1 Sj, linear terms in the D'^ u^ with coefficients homogeneous and linear in x, and quadratic terms in the D"^ u^ with constant coefficients. Definition 6.7.1. We define the spaces *C^(5i?) of v e c t o r s / f o r which fj^^J^'î^R) with \7''fj{0) = 0 for 0 < f < - - s ; and **C^(5ij) of vectors u for which U^Ô^{BR) with V^ u^ {0) = 0 for 0 <.r <.tjc', we define the norms by
(6.7.6)
(6.7.7)
*||/|Ui, =ZK[V-îf^,
i
En), **lklki? -2'^4V^fc^^ Bn),

k
Definition 6.7.2. F o r / ^ *C^ {BR), we define P/j(/) = U, where U^ = M^ - PK u^ = Ll^Fu Fi (x)=fK

BR
(x - i) fi (f) i f
6.7. Analyticity for non-linear systems
267
where P * is the polynomial in (6.7.4). We are here supposing t h a t the order m of the operator LQ is > 0 ; otherwise LQ is just a constant and we set Fi = L^^ fh In order to show that PR is a bounded operator from *Cfi{BR) we need some more refined results than those in Theorem 6.2.1(d). Theorem 6.7.1. PR is a hounded operator from '^Cfjt{ER) to **C^{BR) with bound independent of R {dependent only on v,ibi,E). Moreover, if U = PR (/) and w > 0, then U satisfies (6.7.8) L%U>'=fj. Proof. We extend/y to ^C'l^ [B^R) (properly; cf. Lemma 6.3.3) and write F^[x) = Fn[x) - Fn{x), Fjiix) =jK{x|)/y(f)^| (6.7.9) Fj2{x)=fK{xB2R-BR
f)/,.(I) dl
From Theorem 6.2.1, we conclude that (6.7.10) FIJ^C^~^^(BR), Dmî-s,Fj2{x) (6.7.11) (6.7.12) h^^^-îFi^, ==lDA(x-i)
B2RBR
BR)<Zih4V-'ifj, [D-s,fj{i) - D-s,fj(x)] X^BR.
BR). di,
By differentiating Fj2 and using Theorem 2.6.5, we obtain
A(y) = D^K(y),
A(-y)==A{y),
From (6.7.11), we see easily that I v^-î-/i^y2(^)| < ^ 2 ^ ^ ( v - ^ . 7 ; > BR)'{R - 1^1)^-1. The result (6.7.10) for Fj2 follows as usual from Theorem 2.6.6. Thus, if we define u by (6.7.7), we see that U^Ô^{BR) and u satisfies (6.7.8). Thus U^ satisfies where Pj is a polynomial of degree < Sj. But by differentiating and setting X = 0, we find t h a t Pj ^ 0. The theorem follows. Now, as in 5.8, we assume that w is a solution of (6.7.5) and that u^ **C^(î?o)> ^ 0 > 0. For 0 < i^ < Ro, we write (6.7.13) If we define VR = PR[MHR
(6.7.14) TR{UR,
u = UR + HR, UR = PR[M U + y)(x, Du)'\, + yj {x, D HR)]

DHR)1
HR) = PR[M UR + ip{x, DuR + DHR) - yj(x,
then the equation for UR becomes {6.7.15) UR TR [UR, HR) = VR and HR is seen to satisfy [6.7-^) with fj = 0 on BR,
268
Now, if ^^ is a solution on BRQ, then it is clear that **||^||i? < Mi independently of R. The proof of the following theorem is identical with t h a t of Theorem 5.8.2: Theorem 6.7.2. Suppose u is a solution ^*'^C[x{Bii^ of (6.6.5) and suppose UR, VR, HR, and TR are defined as above. Then **|| î? {{R, **[| VR\\R, and * *|| HR \\R are uniformly bounded by some number M^ for 0 <C R <. Ro, HR satisfies (6.7.8) with f = 0 on BR, and
TR (0; HR) = 0, **|| Tn(UiR; lim **\\VR \\R = lim * *|| ^ ^ ||^ = 0,
2?->0 R-*Q
HR) - TR{U2R;
HR) \\R < e(R) **|| UIR - U^R \\R,
lime{R) = 0 ,
if * *|i HR \\R, * *\\ UIR \\IR, * *1| U2R \\R < M2.
Thus, if 0 <C R ^ Ri, UR is the only solution 0/ (6.7.15) for which *ÛR \\ < M2, provided * *|| HR \\R < M2, HR being supposed given. Now, as in 5.8, we introduce the following spaces of analytic functions ; we use the notations of that section. Definition 6.7.3. The space *C^h{^hR) consists of a l l / ^ *C^(-S/j) which can be extended to E^R SO that each/y$C~^^(5î?) and is holomorphic in BJIR. The space '^*C[jtfi{BjiR) consists of all z/^**C^(5i?) which can be extended to BUR so t h a t each ^^^CJf (5/^/?) and is holomorphic on B^R. We define the norms by
(6.7.16) mUR^ZKiy-'ifhSuR),
j
**||iU=2'A^(v%w* 5;,^).
k
The following theorem is proved by imitating the proof of Theorem 5.8.3: Theorem 6.7.3. Suppose 0 <, h <, hi, hi being the number in Theorem 6.2.1_(a), suppose / $ '^C^h{BhR), suppose XQ^B^R, suppose 5 : f = f (s), 5 C BR, is admissible with respect to the function ^ (|; XQ) =K(xo~i)f {i), and f satisfies (5.8.20). Then the integral (5.8.18) exists. If ^ and f* are both admissible and satisfy (5-8.20), the integrals have the same value. Finally, if for each x^BfiR, S {x) : f ^{s; x) is admissible with respect to g ( f ; x) and satisfies (5.8.20) for each x, then the function F defined by
(6.7.17)
is analytic on B^R and (6.7.18) LoF{x)=f{x),
F[x)=JK[x~^)f[^)d^
8{x)
D<'F{x)==lD^K{x-S)f{^)dS, \(x\<m.
Six)
Definition 6.7.4. For f^*Ch/z{BhR), we define PR{f) =U, where U is defined by (6.7.7) with Fi replaced by its extension (6.7.19) Fi{x)=lK{x-^)fi{^)dl
Six)
6.7- Analyticity for non-linear systems
269
Theorem 6.7.4. PR is a hounded operator from "^Cjifx (ShR) to ^'^Chfx [BUR) with hound depending only on ii, h, and E and not on R. Proof. Let f^*Chf^ {BJIR) and suppose / is extended to BhR as in the definition and is also extended so that es^ch fj^C~^^{B2R) and is zero near 2^2R, the norm of / being increased at most by a factor C (E); the possibiHty of this is proved as in Lemma 6.3.3 using r instead of x^. We then write Fi{x) = Fii{x) -Fi2{x), Clearly F12 is analytic on BfiR and ym^l-s,Fi2{x) - j \7^-HA{X-^)
B2R-BR
Fi2[x)=JK{x
- i)fi{^)
dl
/ Z ( | ) -^^D-fi{x)
L |aj=o
(^ - x)-\
J
d^,
A{y) = \7^K(y),
X^BUR
for, since A satisfies (2.6.9) and (2.6.10), we see by induction and differentiation using Theorem 2.6.5 t h a t (6.7.20) Thus I V^+i-îFz2 (^0) I < ^ ( ^ , h E) h^ {V-Hfi, BUR) [^(^0)]^-^, ^0 B^R. j\J'^-hA(x
BZRBR
S)'{^ xYd^^
0,
0<|^| <
-sz.
(6.7.21)
Since the derivatives up to the order s; of the/y join up continuously across SBR, we find that Drri-^-HFu{x) (6.7.22) = ID^-^K(X
Six)
~ i)D-îfi(S)d^ - ^)D-Hfi[^)d^ s)D-Hfi(s +
+ = (pi(x) + ix2o)ds, r=r (ô)/2. (p2(x),
+ JDm-iK{x
B2R-~BR
(p2(x) = J D^-^K(xi
BiXio,r)
x=xi+i
X2Q,
xo=XiQ+i X2Q s BjiR,
where S {x) Si [XQ) U ^2 (XQ), SI {XQ) is the surface in Definition 5-8.5, and 82= {x = s -{- i X20 \ s ^ B (xio, r)}. In case m = \ and s? > 0, this has to be proved by removing a sphere B (xi, g) as is done for 992 below. For x near XQ, we may find D^ cpi [x) by differentiating under the integral sign to get Z)2^^(x)=fDA{xSiiXo)
I) [D-Hfi(S) - D-hfi{x)-]d^ [D-Hfid) D-Hfi{x)-]di
+ lDA{x-i)
BZR-BR
270
since, by replacing Si (XQ) by a surface near it but containing a thin annulus of points ^ =z s + i X20, s^B {xio,r + e) B [xo, r), we can prove that , , JDÂ (X -S)di + JDo:A{x - i)d^ = jA{x
dB-2R
- D ^ f ; - jA{xi
dBiXio.r)
- s)ds', = 0 ,
xi^B(xio,
r).
Using Theorem 2.6.5 and analytic continuation. Thus (6.7.23) (6.7.24) I V2 991 (^0) I < ^ (/^, h, E) ' [r (ô)]-i. 952e(^) = / ^ ( ^ i ^)/(^ + ^'^20)^5,
B{Xio,r)-B{Xi,Q)
Now, for points x = xi + i:V2 0, X\^B {XIQ, r) we write
r{y) = D ^ - 1 K (y), / ( I ) = D - i / z ( | ) . Differentiating (6.7.24) with respect to x', integrating by parts, and letting ^ - > 0, we obtain (6.7.25) Dccp'zix) = jr{xi
+ f r{Xi
B(%o,r)
s)f{s + ix2o)ds(x'
s)/,a(5 + iX20)ds.
Since these are real integrals of a complex-valued function / , we may differentiate (6.7.25) using Theorems 2.6.2 and 2.6.5 to obtain D^Da(p2[x) = Jr,p(xi
dB(Xio,r) + Cpfâ(Xl + iX20)+J rj(xi S) [ / , a ( s 4 - ^'^20) f,o:(Xl + iX2o)]ds
s) [f{s + 1x20) f(xi
tX2o)]ds'^ +
Cp =
jr{s)ds^
55(0.1)
since F^p can be taken as a zl in Theorem 2.6.5. From this and Lemma 5.8.3 and the fact that d {XQ) < r [XQ), we conclude that (6.7.26) I V2 (p2 {x)\^Z ill, h, E) . [d {xo)Y-^. 0. The result follows from (6.7.21), (6.7.22), (6.7.32), and (6.7.25). Theorem 6.7.5. Suppose H ^* '^Cfi {Bji) and satisfies (6.7.8) withf= Then H^ **Cjifi(BhR) and **\\H\\nR<C{f^,h,E)-*^HU. Proof. We m a y define V to be an extension of H to B2R to have
support in B2R and so t h a t * * | F||i2 < Z i ( ^ , r) **\\H\\R. A S in the
proof of Lemma 6.6.1, we conclude t h a t Hi(x) = Ll'Fi(x),

(6.7.27)
Fi{x) = JK{x B^R-SB
f)/;(f)if,
Mx) = L%VHx),
xfB2R.
271
From (6.7.27), we immediately conclude that H is holomorphic on BJIR. Moreover

S2RBa
SO that I S/rn+l-SjF(x) I ^ Zîfl. h, E) h^(S7-îf], < Z3(fi, h, E)ZK(VhH,

k
S^R) [^(%)]"-! BR) [d{x)r-\ x^BuR.
The result follows easily from this. The proof of the following interior analyticity theorem is essentially like that of Theorem 5.8.6; one verifies the conclusions of Theorem 6.7-2 for the spaces "^^ChfiiBhR). Theorem 6.7.6. Suppose u ^ **C/,(-Bij) for some JR > 0 and satisfies (6.7.5) there where L^j^, Mjjc, and the ipj have the properties as described above. Then u is analytic near x = 0. We now wish to prove analyticity along an analytic portion of the boundary of the solutions of a system of the form (6.7.1) which satisfy non-linear analytic boundary conditions of the form (6.7.28) ; ^ r ( ^ , ^ ^ ) = 0, r=\,...,m where we now use the notation of 6.3; the order of LQ is now 2 m. We assume t h a t the )^r contain derivatives of u^ of order ^tjc hr, where some of the hr may be negative, and we assume that the linearized boundary conditions (6.7.29) ^^Xr(X>D^+^D^ IA-0 = 0
satisfy the conditions of 6.1 and 6.3. We assume that ho is the smallest integer such that (6.7.30) ho > 0, Ao + Ar > 1, ^ = 1, . . ., w, and we assume that our solution u is such that u^ ^ O^'^^^ (9^) in a neighborhood on G of a point Xo on dG, assumed analytic near Xo. We begin by making an analytic change of independent variables which carries xo into the origin, 9^ onto G/jo U Cijo, and dG C)^ onto (^RQ] we then use the {x, y) notation of 6.3. In the resulting equations (6.7.1) and (6.7.28), we make the substitutions (6.7.3) where P^ is that polynomial of degree tfc + ho such that (6.7.4) holds for all r, 0 < r '<tjc + ho. If we then expand the new 99^ and Xr about the origin in the (x, y, p) space, we obtain the system (6.7.5) and the boundary conditions (6.7.31) Bork U^ = Cork U^ + COr {x, 0, D u) where Bork has constant coefficients and contains all and only derivatives D^û^ of order t]c hr, Cork has constant coefficients but is of lower order, and the Taylor expansion of cor begins with a homogeneous
272
polynomial in (x, y) of degree ho -\- hr + 1, linear terms in the 0" u^ with coefficients homogeneous in {x, y) of degree > 1, and second powers of the derivatives; the expansion of ipj has its previous form except that the homogeneous polynomial in {x, y) has degree /ô + 1 ^jThe following lemma can be proved by the method of proof of Lemma 6.3.3: Lemma 6.7.1. There exist bounded extension operators ti JE from
Ci(Gjt) to 0^{G2R), r^jR from CI[BR) to 0^[B2R), T^JR from 0^(GR) to
0^{BR), and T^JR from CjîaR) to 0^(G2R), with hounds independent of R if the norms are defined using A^, and which are such that ti JR (/) vanishes near ^^^ and r2jR(f) vanishes near dB2R. Thus The construction in Lemma 6.3.3 is to be replaced by (6.7.32)
t o y i e l d T3;i?.
F(x,y)=^CiF[x,-j^),
S= l
y<0
Definition 6.7.1'. We define the space *C^ {GR) as the space of vectors / s u c h that fj^C^^-'i{GR) with V^/y (0, 0) = 0 ior 0 < Q < ho Sj, the norm *||/||i? being ^h^{\7ô~^jfj, GR). We define the space *'C^(î?) as the space of vectors g such t h a t gr ^ C^^^^^" {GR) with V^^r(O) = 0 for 0 < ^ < Ao + ^r and norm *'\\g\\ = = X " ^ ^ ( W ' ^ r ^ r , OR). We define the
_ r _
J
space **C^{GR) as the space of vectors u such that U^^C*/^'^^^{GR), VQu^(0,0) = OioYO<^Q<tjc + ho, and norm **|| i^l| =2'î"(V*"+''^^^ Gij). P'or f^'^Cfi{GR) and g^*'C{GR), we define PR(f,g) as follows: Let
fj = T2j,2R[rsjRfj], gr = TÔrRgr, Or = ho + hy.
We then define PR (/, g) = u, where u^ = U^ ^ V^ ~W^ P^ where U^ is given in terms g by (6.3.29), V^ is given b y (6.3-43) ^ ^ is defined b y (6.3.29) with gr replaced by
yr{x) =BorkV^(x,0), [sll x) ]
we assume, of course, that gr{x) = 0 outside a2R and, as usual, P^ is chosen so that \/^ u^ (0,0) = 0 for 0 < ^ < Â; + ôSince PR (/, g) as we have defined it differs from the PR (/, g) of Definition 6.3.5 by a polynomial of degree ^tjc + ho, the following theorem follows from our normalization at the origin and Theorem 6.3.6: Theorem 6.7.1'. PR is a hounded operator from *Cft(GR) X *'C^(ai?) into **C^(Sij) with bound depending only on [fJiyE). Moreover, if U = PRify g) then U satisfies [6.7.S) with f replaced by f everywhere on i?+^^
6.7- Analytic!ty for non-linear systems
273
and also satisfies (6.7.33) BQricU^{x,0)=gr[x), all X. Now we assume t h a t w is a solution of (6.7.5) on GR and (6.7.31) on OR and that u^ **C^{GR^, RQ > 0. For 0 < i^ < jô, we write u = UR + HR, UR = PR [MQ U + ip (x,y, D U), CQU + CO {X, D U)]. If we define VR = PR [MO HR + ^P {X, y, D HR),
CO HR
+ W (X, D HR)] + D HR) HR)]
(6.7.34) TR {UR, HR) = PR [MO UR + ip [x, y,DuR

~ y) {x, y, D HR),
COUR + CO (X, D UR + D HR) (JO{X, D
then the equation for UR reduces to (6.7.15)- An analysis like that in the proof of Theorem 5.8.2 demonstrates the following theorem: Theorem 6.7.2'. The conclusions of Theorem 6.7.2 hold with BR replaced by GR. In addition HR satisfies (6.7-33) ^^^^ ^r 0 on GR. As in 5.8, we now introduce certain spaces of functions which are analytic in the ;^:-variables only. Definition 6.7.3'. We define BohR (as in 5.8) to be the part oiBhR where y is real, G^R as the part oi BohR where y > 0, and G^^ the part of BohR where y < 0. The space "^Chfz (GUR) consists of all v e c t o r s / ^ *C^ {GR) which can be extended to GhR so that fj^C^^~^^(GjiR) and fj is analytic in X for all (x, y) in GJIR. The space *'C/i^(G/îj) consists of all ^ ^ *Ci^{aR) such t h a t gr can be extended to OJIR SO that gr^C^^'^'^'[GhR) and gr is analytic in GUR] here GUR is the part of BQ^R where y ^ O . The space '^*Chfi{GhR) consists of all u^ **Cf,{GR) which can be extended to GhR so that u^ ^ C*^^^^ {GhR) and u^ is analytic in x for {x, y) in GhR. We define the norms as usual:
*\\fUR =ZK[WVy.
ÛR].
etc.
We wish now to extend the definition of the operator PR (/, g) to the space *Chfi{GhR) X '^'Chtx{GhR) and wish to show then that PR is a bounded operator from that space to '^'^Chf^{GhR), We consider the separate terms in the definition 6.7.1'Theorem 6.7.3' (a). Suppose g ^'^'Chfji{GhR), gr = r4srRgr where Sr = ho + ^r, (^'y^d we define uTc = U^pJc^ Ui^{x^ y) = f^J ZC.L^^{x - I , y)L-gr{S, 0)dS
where Cx = \(x\\l(x\ and P^ is the usual polynomial. Then u^**Chfz{GhR), satisfies the homogeneous equations {6.7.S) on Rj^+i, satisfies the boundary conditions (6.7-33) for all real x and all complex x on GhR, and finally (6.7.35) '^*\\u\\h<C{iu,h,E)'
Money, Multiple Integrals
*'\\g\\ provided
0 <h <h2{E)
^g
<hi{E).
274
Proof. From Theorem 6.3.2, we conclude that the L^'^(x,y) are analytic for all {x, y) " 0 such t h a t y is real and > 0 and x == Xi + i X2 is complex with | ^ 2 | < /^]/|yp + \xi\^ if 0 <C h <h2(E). For such h, we extend gr to ajiR and then extend U^ to GhR by the formula
UHx, y) =f^Hx, y; f)^f + J ^Hx, y; f)i|,

^Hx, y; f) =Z
where
r
ZC.Ll^[x - i, y)D''gr{i),
\\=Sr
D. = ||!/!
S(x,y) :i == s + i^2 (s),
SÔR,
is admissible. We have seen in 5.8 that such integrals are independent of the path 5 {x, y) as long as it is admissible so that U^ is analytic. Near any point [XQ, y^), we may choose a fixed surface containing a disc f == s + ^' X2Qy I s XIQ\ < Q, independent of [x, y). We may calculate the derivatives of U^ by differentiating in the real directions, so that we would find t h a t a derivative Dx Dh'^ô U^ {x, y) is a sum of terms of the form
jD^rix
- I y)y{i)d^ + JD^rix
O-GR
- | , y)y[^)d^
S{X,y)
where y (f) D^rgr, r{x, y) Dh'^h Ll^(x, y) so that P i s homogeneous of degree ~v. Now, if y > 0, we find, as in the proof of Theorem 6.3.3 and by using analytic continuation, that
fD^rix
S {X, y)
- I y)d^ + JD^rix
a-aR
- 1 y) d^ = 0,
[x, y)^GnR
SO that we may replace y (|) above by 7 (|) y {x) and thus obtain a bound like (6.3.23) for all such derivatives. Using the differential equations we get a similar bound for i^^fc+'ô+i and hence conclude the inequality (6.7.35) and thus the theorem. Theorem 6.7.3' (b). Suppose f ^""Cnii {GR) and we define V^ by (6.3.43) where fj = T2;.,-2i2 T3A,-2i?//, ^j = ho Sj. Then V^ ^J^*^'^^^ {BA) for any A, V^ is analytic in x for (x,y)^BohR, V^^Cy^^{BohR), hf,(Vh+ôV^, BohR) < G {/u, h, E) *||/||/i, and yr{X) ^ Bork V^(X) satisfies the conditions at 00 in Theorem 6.3.3 and the analyticy conditions in Theorem 6.7.3'(a). Thus W^^C*^'^^''{GA) for any A, W^ it is analytic in x for (x,y)^GfiR, Wfc^Cyô^GA), andhf,{Vh+hoW^, GQUR) <.G{fA, h, E) *||/||/,. Ifv^ = V^ TF^ P^, P^ being the usual normalizing polynomial, then v satisfies {6.7.K) with fj replaced by fj on R^'+i U G^R and satisfies (6.7.}}) with gr =" 0 on a U OhRProof. The results for real [Xyy) follow from Theorem 6.3.6. We may extend/y toi^v+i U BQUR by defining/,- as above on B2R, defining/; {x,y) = 0 on Rv+iB2R, extending/- to GhR as in the definition, and then extend-
275
ingfj to BohR using the formulas (6.7.32). I t is clear t h a t the extendedîs analytic in x for {x, y) on BQUR. We m a y then extend Fj to Rv+i U BQJIR
by defining Fj{x,y) ^ l^j{x,y;

S(X,y)
| , fj)dS drj + J ^j(x, y;

B2R-BR
^,rj)d^drj,
S i K y]Lr]) where (6.7.36) S{x,y)\^^s
=K(x-^,y
~- rj)fj(^, rj),
[x, y) ^ EQUR,
+ ih (s, t) X, y), r) = t (real), {s, t) ^ BR
is admissible for g ; . An analysis like that in the proof of Theorem 5.8.3 shows that Fj is analytic in x for [x, y)^BojiR and that we may differentiate Fj 2m + ho Sj i times, shifting ho Sj of the differentiations onto the ^. The result is (p{x, y) =.lr{x-ly
(6.7.}7) s^^'^y)
- fj)f(l
rj)didrj
+ J r(X ~
BZR-BR
E)f[E)dE
where / $ C^^[B2R U BohR) and F is homogeneous of degree r. Near a point {xo, yo) in BohR, we may take S {x, y) as the union of a fixed surface 5 i , given b y (6.7.36) with (5, t) ^ BR B (jîo, yo', y) and the disc ^ = s -\- i X2,o, ri = t with (s, t)^B (^10, yo; ^). Then, for points [x, y) with X = xi-^ i xô and [xi, y)^B (A;IO, yo',r), {6.7.'^7) becomes (p {x, y) -= (pi (x, y) + (p2 {x, y), (6.7.38) ^ i ( Z ) = fr(XE)f{E)dE + f r(X ixô, E)f{E)dE, t)dsdt.
q)2{x, y) ~ f F(xi s, y t)f{s-\B{Xio,yo'^r)
An analysis exactly like that applied to the cpi in the proof of Theorem 6.7.4 shows t h a t (6.7.39) I V2 (pi{X)! < Z(/^, K E) h,{f, BOUR) [d[X)r-^
d{X) being the distance from X to dBohR where BOUR^^ considered as a set in i^2+i. We may differentiate 992 with respect to an x' obtaining the formula analogous to (6.7.28). This m a y then be differentiated with respect t o y or an x^ so that we obtain a bound (6.7-39) for \\J x \/(pi[X)\. It then follows that any derivative Dh'^ô~^r^ ^C^(BohR) with the desired bound. The results for V then follow from the differential equations. The proof of the results for W differs from the proof of Theorem 6.7.3(a) only in the consideration of the convergence at 00, since y does not have compact support. Thus we obtain our first desired result. Theorem 6.7 A'. PR is a hounded operator from *ChM{G-hR) X '^' [CuixOhR) into **Chpi(GjiR) with bound C(ibt, h, E).
18*
276
Theorem 6.7.5'. Suppose H^**CJÎ{GR), satisfies (6.7.8) with the fj = 0 on GR, and satisfies (6.7.33) withgr = 0 on OR. ThenH^ * *QA* (^hR) and (6.7.40) * *|| H WnR < C (//, h, E) * *|] H \\R if 0 < /Z < A2 < Ai, = 0 elsewhere on
^2 heing the number in Theorem 6.7.3'(a). Proof. Let H^ = ri,t^+ho,RH^ and define H^(x,y) F^+i and define (6.7.41)
fj(x, y) = Lo3k HK gr(x) = Bork H^ {%, 0).
Then | r = 0 on OR and on o* G2R and/,- = o on GR and on R^^^ G2R. We now define f = r' f where (p is defined as usual ( = 1 for s < 1, == 0 for s > 7/4, etc.) f,[X)=<p{R-^\X\)r,{X) f.[x,y) =ZCJ^[x, -(T,- + 1)-My]
where the CJ are the unique constants such that

Tj + l
Then T' is a bounded extension operator and we see that fj (X) = 0 for X ^ BR and in i?vî B^R. Then we define F by (6.3.43) and then define W as indicated immediately below and define U by (6.3.29) in terms of g. Then U, V, and W have the requisite differentiability bounds everywhere, V is analytic in BUR, U is analytic in GRH U ORh, if h is small enough, and the fr are analytic in BfiR. From a minor extension of Theorem 6.7.3'(a), it follows t h a t W is analytic in x on GRhUaRh. Finally, from Theorem 6.3.6, it follows t h a t u = U ^ V W satisfies (6.3.44) w i t h / * = ^ * = 0 s i n c e / and g are 0 near 0 {gf could be a polynomial vanishing on a). From Theorem 6.3.5, it follows that H ~ u is a polynomial of low degree. The results follow. Theorem 6.7.6'. Suppose w^**C^(Gi?Q) and satisfies (6.7-5) on GR and (6.7.31) along GR, where the ipj and cor satisfy the analyticity conditions specified near the respective equations. Then u is analytic at the origin. Proof. Suppose 0 <C h <, h2 (E) < hi {E). From Theorem 6.7.2' and 6.7.5', it follows that **\\HR\\h,R < Ms for 0<R < R2. The theorems above show t h a t the conclusions of Theorem 6.7.2' hold for the spaces **Ch,.(GhR). Thus if 0<R<R2<Ri, UR is the only solution of (6.7.15) with **|lwi?||;^ < M3. Thus the previous UR must ^**C}if^{GhR). Thus, this is true of u also. The ellipticity shows t h a t we may, after differentiating the j ~ t h equation in (6.7-5) Sj times, solve for D*^ u^ in terms of the other derivatives. Introducing the various derivatives of
6.8. Differentiability. A perturbation theorem
277
the u^ as new variables w\ we see that the w^ [x, y) are analytic in x for [x, y) on GfiR and satisfy a system of the form (5.8.49). The analyticity at (0, 0) follows from our previous discussion of that system in the proof of Theorem 5.8.6'. 6.8. The differentiability of the solutions of non-linear systems; weak solutions; a perturbation theorem elliptic
In this section we first obtain results concerning the differentiability of the solutions or weak solutions of non-linear elliptic systems which satisfy general non-linear boundary conditions. We assume that t h e equations of variation, as defined in (6.7.2), are properly elliptic along the solution and the linearized boundary conditions, as defined in (6.7.29) satisfy t h e complementing condition. We make t h e corresponding assumptions regarding t h e non-linear systems analogous t o those in 6.4. We conclude t h e section with a perturbation theorem for such non-linear boundary value problems. We shall obtain our results on the interior first and shall make use of the spaces ^H^Q(BR), etc., defined in Definition 6.2.1, on balls BR interior to our domain. Theorem 6.8.1. Suppose that u^ ^ Ch(G) and satisfies the equations (6.8.1) (pj(x,Du) =^0, y=i-...,iV,
on G, where (pj{Xyp)^ C^-^;(9^), h'> i, and ^ is an open set in [x,p)space containing all the points [x, D u); we assume that the equations of variation are properly elliptic. Then u^^ H^''^^{D) for each p'> i and each D C. C. G and the derivatives D^ u^, | ^ | < h, satisfy the corresponding differentiated equations [almost everywhere if \d\ = h). If, also, the ^3^ C^^'î^) then the u^ ^ Cl^+^{D) for each D C C G. If the cpj are of class C^ (analytic) on Jl, then the u^ are of class C^ (analytic) on G. Proof. We first prove the first statement ior h = 1. Let XQ be any point of G and let B(xo, A) C G; we hereafter assume that XQ = 0. Let 0*0 be a positive number < ^ / 2 , let y be a positive integer <i^, let Cy be the unit vector along the x^ axis, and define ul(x) = a-^[u^(x
N
+ acy) u^(x)],
\a\ < (TQ.
Then the u^ satisfy the equations U (6.8.2) 2 <fc (^) D^ â (^) =fja (x),
1
(or not summed)
ci'ji(x)=f(pjî{x+atey,p(x) + t[p{x+aey)-p(x)]}dt
fjo(x) = - f (fja^y {same} dt,
{p = {p^}).
278
Clearly a^-> Ujla"^ = {^"ji}' ^^^) ^^^foj-^fj (6.8.3) (^^jci^) = (pjp^ [x, Du{x)],
in C~^j, where fj(x) = - (pj:^v.
So the equations (6.8.2) are properly elliptic and the constants C2 and R2 in Theorem 6.2.6 are independent of a. So, choose R < min(i^2, ^ / 2 ) and let CR ^ T (^R) with C^j == 1 on BRf2 and define 17^ = ^ â- Multiplying (6.8.2) by C, we see t h a t Ua satisfies
(6.8.4)
IL^, VI = Fi. = C/y. + 2" ^"n <

k k
where the operators Af^^ (which involve the derivatives of CR) a-re of order < S j + fe 1 and the Z^^ have the a^f as coefficients. It follows that the Fja converge in C~î to certain Fj which are expressible in terms of x and the D^ u^, \^\ < tjc. From Theorem 6.2.6 it follows t h a t the norms of Ua are uniformly bounded in **H^{BR) for any given ^ , if JR is small enough. Thus a subsequence of values of or -> 0 exists such that Ua 7 in '^*H^{BE) to some U. By writing (6.8.4) in the form
( 6 . 8 . 5) Z Ulc Ul = Fja k Gja, Gja = y; ( L ? , k Ljj,) Ul
and noting that Gja->0 in H-^^(Bji), we see that Ua-Û in '^*H^{BR) and t h a t t/ satisfies the limiting equations. Thus U^^^^HI^{BRI2) and satisfies the differentiated equations on BRJ^. Since XQ was arbitrary the first result follows for h = \. The remaining results follow from repeated apphcations of Theorem 6.2.5 since the # y satisfy. {6.%.6) Ljjc u\ (x) = Z a^j^ {x) D^ u^^ = fjy {x) = (pja^y (x, D u).
k,P
Since u^^ H^^-^^{D) for each ^ > 1, it follows that u^^ C^^,{D) for each ju\ 0 < / ^ ' < 1. Thus, if the cpj^ CJ-^^(9^) for some //, 0 < / / < 1, it follows first the af^ and/y in (6.8.3) C'^J, so that u^^ Cl^+^. Then the a s a n d / ' s $ C''^ so that ^ ^ ^ Cj*+i. If,"now the (p^^ Q " '"''it will follow t h a t the as and f s^C^-'^ so that u^^H*'^+^ for each p> \. If (fj^ C^"*, it follows as before that u^^ C*^^^. The argument may be repeated. The following theorem is proved in a similar way using Theorem 6.3.8 for boundary neighborhoods. Theorem 6.8.2. Suppose that ho is the smallest integer satisfying the conditions . , , r , ho>0, ho -}- hr > \ for each r. Suppose u^^ C*^+ô(G) and the u^ satisfy (6.8.1) on G and (6.8.7) ;fr(^, ^ ^ ) = 0 ondG. Suppose, for some h > ho, that G is hounded and of class C*"^'^, ^0 = max^^:, (fj^ C^~^î^), and ir^ C^'^^''{W), where 9^ and W are appropriate
279
neighborhoods in the {x, p)-space containing all the points [x, D u{x)] for x^G. We assume that the linearized equations are properly elliptic and the linearized boundary conditions satisfy the corresponding complementing conditions on dG. Then u^^ H^^'^^{G) for each _> > 1 and the D^ u^ / satisfy the differentiated equations on G. If, also, the g)j ^ C^"^' (%) and Xr^ C^'^î^), i^^'^ ^ ^ C^^'^^{G). Corresponding results hold in the C"^ and analytic cases. / / ô > 1, G is of class Cjo+ô^ u^^ C^*^+'ô~i(C), (fj^ Cl^-'^{^), and the Xr^ C^^^^^^Wj, then u^^ Cyôf^Q). Proof. We prove the first conclusion for h = \. The interior results have been proved in Theorem 6.8.1. So, let XQ be a boundary point and map a neighborhood of XQ onto GA in the usual way. Let Cy be the unit vector in the (new) xv direction, assumed tangential. If we define the u^ as before, we see t h a t they satisfy (6.8.2) and the boundary conditions
where 6^^ and gj" are defined in terms of the Xrp'^ and ~Xrxy as in (6.8.2). From Theorem 6.3.8 and the convergence in C^^'^^'' of b and g'^ to &r and gr, respectively, we conclude that we may find an i?2 > 0 such that the bound in (6.3.49) holds on GR, independently of cr if 0 < i? < R2. Then we define Ci? and U^ as before, we may repeat the argument of the preceding proof to show t h a t U^ ~>u^y in H*^'^^^ {GR) . Using the differential equations, we find t h a t the normal derivative u^^ also H*'''^^^(GR). The further differentiability results are obtained by repeated application of Theorem 6.3.7- The last statement is proved in the same way. The equations analogous to (6.4.1) are of the form {6.S.S) [^
G
j=\
||^
2 ' ^ ' ^ ^ ^ * - ^ ? ( ^ ^ ^ ) ^ ^ = 0,
v^C^[G)
where we shall assume that S;, t)c, mj, ho, and QJJC satisfy the conditions in (6.4.1), (6.4.2), (6.4.3), and (6.4.4) and 99^ involves derivatives D^ w^ where \P\ ^Qjk' We assume that each cp'^ with \(x\ = wy$ C^ at least and that the system of operators in (6.4.6) is properly elliptic for each set of numbers (6.8.9) <f0 = 9^^ (^0, po). \(X I = W;, j /5 I = QJJC for (xo, po) in a neighborhood 9^ of all points [x, D u(x)']. The theorem analogous to Theorem 6.8.1 may be stated as follows: Theorem 6.8.3. Suppose that Sj, tjc, mj, QJ^, ho, and the 99^ satisfy the conditions stated above and suppose u^^ C*'^+'*o(G) and satisfies equations (6.S.8) on G. Suppose, for some A > Ao + 1, that the cpf^ C''('j!fl) where T = max(0, A S; + 1^1) and ^t has its usual significance. Then u^^H^^+^{D) for each D C C G. If the cp^^ Cl(^), 0 < / ^ < 1, then uic^ Cl^+^^{D) for each D C C G.
280
Sketch of proof. We prove the first statement ioi h = ho + i. The second statement follows from the theory of 6.4. The results for larger h follow by repeating the argument. We begin by replacing v^ in (6.8.8) by v^ = a~^ [v3(x ~ a Cy) v^x)], Cy having its usual significance; we assume that v^ C'^(G) and that 10*1 is sufficiently small. By proceeding as in 1.11, we see that the u^ satisfy equations of the form (6.4.1) where the oF^^^ and/^*^ are defined by formulas like (6.8.2) for those {j, oc) for which 99^$ 0{^) for some T > 1, namely those for which HQ Sj -\- \(x\ > 0 (for each j , there is at least one such a since HQ Sj + nij > 0). For the remaining (;', a), for which the 99" are not differentiable, we write
JD^'vKx)-
(p'^[x,u(x)]dx
= J (p'^[x,u(x)]'i
( ?
1 0 .
jD'^viîx
taey)dt\dx
= jD^vi^[x)[-f^^{x)]dx,
G
f^^[x)
= f (p^[x + tGey,u{x
0
taey)]dt.
(6.8.10) Since any 99" with | ^ | = m^; is differentiable, we see that the u^ satisfy equations of the form (6.4.1) where the a^^^ = 0 and the/^"" are given in (6.8.10) a ho Sj + \(x\ < 0. The usual argument can be carried through using Theorem 6.4.4. The boundary conditions analogous to (6.4.15) are: (6.8.11) j Z
eo
ZDyC'xry[^'^î^)']^^
= ^>
C^C'-iG)
r \v\<Vr
where the ^ry involve derivatives D^ u^ with |/5| <.tjc hr pr and are such t h a t if C has support on a boundary patch mapped on GR, then (6.8.11) reduces to (6.8.12) [2^ ZDl^'''Xry[^,0\Du[x,0)]dx = 0
if we use the (x, y) notation of 6.3. We shall require that the ^ry with \y\ = pr shall be differentiable and that the operators
(6 8 13)
BorJcU^^2:
\v\=Pr
\d\ = Qr
Z(-^)'^'&rfcr0^^î^^ (^0, po) ^W,
, qr = h ~ hr - Pr (X= {x, y)),
Kicvo = Xryp^, [^0, po],
W being a neighborhood of all the [X, D u{X)], X^G, satisfy the complementing condition with respect to the Lojk- We assume that Sj, tjc, mj, Qjjc and ho satisfy their previous conditions and that ho, hr, and pr satisfy the conditions pr>0, ho + hr+pjc'>\, tk hrpr'>0 and we assume t h a t ho (probably negative) is the smallest integer satisfying all these conditions.
281
Using the methods of proof outHned above and in 6.4, we can prove the following theorem: Theorem 6.8.4. Suppose that the Sj, tjc, QJJC, ho, hr, pr, (p'j, ^^^ X^v satisfy all the conditions above, suppose that the u^ ^ Qic+hQ ^^^ ^^^ satisfy the equations {6.S.%) on G and (6.8.11) on dG, suppose that, for some h^ho+ i, G is of class C'+^ (p^^^ C^'St) and Xw^ C'îW), where T = max {0, h ~ Sj -^ \a\), co = m a x ( 1 , A + ^r + 171), and 9^ and W are neighborhoods of the sets {[x, D u (x)], x^ G} in the appropriate [x,p)spaces. Then u^^ H^^^'îfi) for each p^X.lf, also, G is of class C^^"^^ ^y C ^ ^ ) , and xry^ C^{^'). thenu^^ Cy^{G). Uho > 1, u^^ C^^+'ô'i, G is of class C*^^^^, 99^^ C^^^), and Xry^ Cîi{^')y where K =^ max (0, ho Sj + $x$ andX = m a x ( l , ho + h + \y\), then u^^ C^*+'ô(G).
We now wish to prove the perturbation theorem of AGMON, DOUGLIS,
and NiRENBERG ([1], Theorem 12.6). We first prove the following standard lemma: Lemma 6.8.1. Suppose the Ljjc and Brjc satisfy the ho JJL conditions^ on a domain G of class C^^'^^^ and suppose the system L is properly elliptic and the boundary operators Brjc satisfy the complementing condition on dG. Suppose also that the problem (6.8.14) Ljjc u^ = 0 on G, Bric u^ = 0 on dG
has a unique solution. Then there is a constant C, independent of u, such that if Ljjc u^ = fj on G and BricU^ = gr on dG, then
(6.8.15)
2'lll^*iy'^^'< ^fZ'lll/il!!^^^- + 2'lll^r|!|^+N
where gr is any function ^ C^^'^^'{G) such that gr = gr on dG and (6.8.16) |||a^|!|^ = ^^(V^9)) + i ; m a x | v X ^ ) ! .
Proof. Suppose this is not true. Then there is a sequence u^ such that the left side of (6.8.15) equals 1 for each n and the right side tends to zero. From the equi-continuity, we may assume then that Ljk u^~>0 and Brjc u^->0 uniformly. But, from Theorem 6.3.9 and the negation of (6.8.15), we conclude also that Un-û uniformly and u ^ 0 and also Cjl''^^^{G) and u satisfies (6.8.14). This contradicts the hypothesis. Let us now suppose t h a t the numbers Sj, tjc, mj, QJJC, ho, and hr satisfy the condition in Theorems 6.8.1 and 6.8.2, that G is of class Cô+'^+^, that u^^CI^'+ô+^{G), and that (pf^ ^ 6^^+'-'^ and ;^f ^ C^+'''-+^ in their arguments, and that the u^ satisfy the equations (6.8.1) on G and {6.S.7) on dG. These respective equations of variation, as defined in (6.7.2)
1 See Definitions 6.2.3 and 6.3.6.
282
and (6.7.29), reduce to LjTc v^ -=Q on G, Ljjc v ^ ^ 2 Brk z^^ = 0 on dG r = sj + h I < W D^^K
We wish to find a solution u = UQ -\- v oi the equations (6.8.18)

define C681Q^ ^0.0, lyj
(pf^ {x, Du) = fj(x) inG,
Xr i^' Du) =gr(x)
ondG,
we
for all / a n d g w i t h * | | | / | | 1 ^ a n d *'|||^||||^ sufficiently small, w h e r e *lll/"lll^ Z/ Illy; IIU Illy ilU T III/'Jll'io-sj- ' j *'III i? IIU "^^ Z( Ill&^llliW y Ills Iljo inf J. III ^^111^0+^'-
the norms on the right having been defined in (6.8.16). Suppose we define
m (^> P) = <P?' [^> ^ ^0 {X) + P] , Xr (^, P) = xT i^' D uo (X) + P] .
(6.8.20) Since we have assumed that UQ^ Cj^+ô+^(G), it follows that cpj^ cô+2-sj and Xr^ cô+hr+2 ^^ j^g arguments and (6.8.21)
(Pj(^>P) Z
(pj(x,0)~0,
;fr(^, 0 ) ^ 0 .
T = Sj + tjc X = h K.
Now, let us define yjj and cor by

I! 9JvHx> 0)p} = ipj(x,p), ZXrp^{x>0)P} =" COr{x,p),
(6.8.22)
Xr{x>P) Z
^"^'^^
Then the yjj and y)jp^ Qo+i-s^ and the Xr and Xrp^ d^+hr+i [^ their arguments and (6.8.23) WA^' 0) ^ Wjpi^> 0) = ^ ' COr(::t:, 0) = (JOrp{x, 0). The equations (6.8.18) then become, in turn,
(6.8.24) , ^ 3 25) (pj{x, D v) =fj{x) in G, Xr{x> -D v) = gr{x) on dG,
^^^ '^^ ^ ^^^^^ "" '^^ ^^' ^ ^ (^^^ ^ ^ ^ BrJc V^ = gr (^) C'^r [^, -C^ ^ (^)] on ^ G
where L^j^ and Brk are exactly the operators in (6.8.17). We now state the theorem. Theorem 6.8.5. Suppose (1) that all the numbers Sj, tjc, to = max^^t, '^jyQjk, ho, and hr satisfy the conditions in Theorems 6.8.1 and 6.8.2; (2) that G is of class Cô+ô+2^ (3) ff^^t u^^ Cy^^[G), (4) that cpf^ Cô+2-s, and Xr^^ ^ cô+hr+2 ^^ ^j^^^y arguments in % and W, (5) that the u^ satisfy (6.8.1) and (6.8.7) "îth (pj replaced by cpf^ and Xr by xf\ (6) that the Lsystem in (6.8.17) is properly elliptic, (7) that the Brk satisfy the complementing condition on dG, and (8) that the homogeneous system (6.8.17) has only the zero solution.
283
Then v^^ C^^'^^^^'^ [G) and there exist numbers A and J5 > 0 such that if the norms of f and g, as defined in (6.8.19), are <; A, then the equations (6.8.18) have a unique solution v of norm <CB. If ho> \, all the differentiability requirements may be reduced by 1 in terms of ho {G of class Cô+'io+i^ etc.) and we conclude that v^^ C^^^^^^^[G). Proof. The additional differentiability of v follows from Theorem 6.8.2. Using this differentiability, we introduce the functions cpj, Xry Wh and cor as in (6.8.20) and (6.8.22) and we conclude (6.8.23). From Lemma 6.8.1, it follows that the problem (6.8.26) Ljic w^ =fj on G, Brjc w^ = gr on dG has a unique solution which we denote by P (/, ^) == ze' in which P is a bounded operator from *CjJ X *'C2 (norms in (6.8.19)) into **C]J which consists of all vectors uû^^ C^^'^^ (G) with
**|| 1
III h+ho
k
^\\\n
Then the equations (6.8.18) and (6.8.25) are equivalent to (6.8.27) **\\\Tvi (6.8.28)
III
V -{- T V = w, w =-PiAg)^
Tv == Pbp,oj]. if * * lll^^lilS<^ i=\,2
We wish to show that ^

" |l!/
Tv2K<-** \\\vi
2
-M\\'
where L :is sufficiently small Let =y)j[x,Dvi:W], (6.8.29) f*{x) and let us suppose t h a t (6.8.30) (6.8.31) **lvill<L,
^* W ==
(Or [x, Dvi{x)],
1, 2,
**lvi~vtll = R. x^G.
Then, from (6.8.23), we see that \frM-fU^)\<CL-R, Now, letting the p^ be arranged in a single sequence p^, . . ., p^, we see that ^
/f w -/AW -y;*2 w = s^Hi^) [î w - î wi,

1
(6.8.32)
Aj'i(x)=fipjj,i[x,p{x,f)]df, + f[Pl{x)~P2{x)].
p(xj')=p2(x)
In case ho Sj = 0, so t h a t ho = Sj = 0, all we need to do is show that hf,{ff, G) ^CLf'R, From (6.8.32), (6.8.23), and the fact that y)j and yjjp^ CJ, we easily find that \Aji{x)\<CL, \Aji(x2) - Aj'i{xi)\<C 1/^1x2-xilf" if Z < 1 . The desired result in this case follows.
284
More generally, a derivative of order r = ho Sj (> 0) of /f^ /fg ^ff is a sum of products of a derivative of order q of an ^^-^ and one of order r q oi SL (p\ pl), 0 <, q <,r. Any derivative of order q of an Aji is the sum of terms of the form
1
Q =.ly,^^^j,[X,p{X,
0
f)] [Dr^P^l) W>0,
. . . {DympJCn) dt' + \-\yra\=q,
(6.8.33)
K=^[ki,...,km),
\oc\ + \yi\
7co > 0,
CO \,...,m\im>Oy
ipjiaK = D^Dpici. . . Dpkmipjpi
where oc and the yo) are multi-indices. If m > 2 or ^ < r, it is easy to see that every such term Q^ C^ and that
h^(Q) < m a x | V ^ W | <
XeG
C^>
\Q(^)\
^^>
the last holding even if ^ = T. If 5' = r, the only non-differentiable terms (assuming r > 0) ^ are of the form
1
Q = J tpj pi j)ic[x, p {x, t)] D^ p^ {x, t) dt,

0
\p\
=r.
Let Q be such a term and abbreviate the integrand ip Q. Then 1 ^(^2) - ^(^1) - /v^(^2, t') [Q{X2, n - Q(xi, f)] +
+ Q{xi,r)\jp{x,^j')-^[x^,t')-]dt'.
Using the definitions of p [x, t') and Q [x, t') = D^ p, and the further differentiability of y)j, we obtain
\Q{^2) ^ ( ^ I ) | <CL \X2 xi\f.
The other ff and the correspondingly defined g^ = g*^ gf^, can be handled in the same way. Thus if **||| z;^ |||^ < L, we conclude from these results and the boundedness of P that * * | | | r ? ; i - Tv2\\\^^< CL''**|||?;i-?;2|||^, if L is small enough. The result follows. The following special case of a variant of Theorem 6.8.5 can be extended to systems: Theorem 6.8.6. Suppose G is of class Q , the A''{x, z,p) ^C%, B{x, z,p)^ C^, zo^C^(G), zo satisfies the equations (1.10.13) (N = 1), and the equation of variation ^ a^P = Al^ [x, z{x),^z{x)-\, h^ = Al, c^ = Bp^, d =:^ B^ has only the zero solution if ^ = 0 on dG. Then there are constants C, ^0 !> 0, and C Q > 0 such that the boundary problem T (6.8.35) L{z]l:) = 0, C = 1 o n ^ G (6.8.34)
/ . . ..X
^(^^^) ^ r i (^'^^'^ + ^'^) - ('"'^'^ + ^ 0 = 0 with Z = ZO.
285
has a unique solution C ^^ [G) provided || 2: ô || < <7o. If we denote this solution by % [z), then we have \\%{z2)-%[z^)\\<C\\z^-zr\\ (6.8.36) Thus there exists a unique function ^C^[ Qo, Qo\ and satisfies (6.8.37) % = ^^^)' Z from [^0,^0] i'^to CJ^[G) which ^(0)=ô. if | | ^ i - ^ 2 | i < a o ( | ^ | | = l||9'|l!I).
Proof. From Lemma 6.8.1 it follows t h a t there is a constant C\ such t h a t if / ^ C^^{G) a n d ^ ^ C|(^)> there is a unique solution of the problem (6.8.38) and that (6.8.39) l|Ci<Ci(|||/|||o+|g||) (|k||=||k|p. So we set f = fo + o) in (6.8.35) where Co is the solution of the problem (6.8.35) with z = ZQ. Then we write (6.8.35) in the form
(6.8.40) L{ZO;CD) = [L[Z] W)-L(^0; CO)] [L{z; Co) L{zo\ Co)].
L(2:o;C) = / o n G,
C^ôn^G
From this, (6.8.38), and (6.3.39), we find that (6.8.41) l l c a | | < C 2 | | . - . o | | ( | | c . | | + ||Co||). llCi < llôl (1 + 2 C 2 I . - ôll). Thus, if C2 11-^ 2:0 I < 1/2, we see that I (6.8.42) \\w\\ < 2C2 ||. - ôll IKoll,
Now, if we have CA; = ^^{^k), \^k .ô|| <o'o, ^ = 1, 2, then we have L(^2, C2) - L{zi, Ci) = 0, C2 - Ci = 0 on ^G, and L{z^, C2 - Ci) - \.L{Z2) Ci) -L[zv, Ci)].
An analysis like that in (6.3.40)(6.8.42) shows that IIC2 - fill ^ C3 1^2 - ^lll llCli < C4 1^2 - ^ll|. Thus g (Z) satisfies a LIPSCHITZ condition so the theorem follows. Theorem 6.8.7. Suppose that G, the A^, B, and ZQ satisfy all the hypotheses of Theorem 6.S.6 except the last. If X{)^ G, there exists an R'> 0 such that the problem (6.8.35) "^Hh z ^= ZQ and G = B{XQ, R) has a unique solution C (ind C(x) > 0 in l^{xo, R). Proof. We notice from (6.8.34) t h a t i:(ô;C) = a^{x)C,ap + 'b-{x)C,. + 'c{x)C where the a^^ and 'c^ Cl(G) and 'b'^^ C^(G) and the coefficients depend on zo, of course. Now for each R such that B {XQ, R) C G, we define
VRiy) =CR(xo + Ry) = 1 +ojR(y),
y^B(0,\).
286
A variational method in the theory of harmonic integrals
Then the desired COR satisfies the conditions Al^WR^.p (6.8.43) + {Al^ - Al^) (DRô^p + RBla>^^, + R^ CRCOR + R^CR = 0 , COR = 0 on dB(0,\), Alf(y) = a'^^[XQ + Ry), etc.,
Af = A%^(0) = a-nô).
From (6.8.43), we conclude that IICORW <{ZiR The result follows. Chapter 7 + Z2R^+^ + ZsR^ + Z4R^) \\COR\\ + R^Z,.

7.1. Introduction In this chapter, we show how one can apply the variational method to the study of the theory of harmonic integrals. In h i s first paper on the subject, W. V. D. HODGE [1] used a variational method in this theory t o study certain boundary value problems for forms defined on domains in Euclidean space (using Cartesian coordinates). But, in order to carry his theory over to compact Riemannian manifolds, he and subsequent writers found it expedient to employ methods involving integral equations (see H O D G E [2], KODAIRA, D E RHAM-KODAIRA, and references
therein). More recently, MILGRAM and ROSENBLOOM ([1], [2]) and G A F F -
NEY ([1], [2]), have treated certain problems b y their ''heat equation" method involving parabolic equations. The variational method was applied t o general compact Riemannian manifolds by MORREY and EELLS and to such manifolds with boundary by MORREY [11]. A closely related method was employed concurrently b y FRIEDRICHS [3] in both cases. Certain boundary value problems had been discussed previously
b y D U F F and SPENCER and b y CONNOR ([1], [2]).
One of the principal results which was obtained early is now known as the ''Kodaira decomposition" and states that
(7.1.1) S2-ee:e or &l=^r@^r@c^r
where S j denotes the space of forms of degree Tg which are either all even or all odd, which have components in S2, the inner product being defined in 7.2; ^ denotes the space of harmonic fields in S2, and (S^ and ^ are the respective closures of the spaces {da} and {d^} where a and ^ are smooth (see 7-2 for definitions). I n case all the forms involved are smooth (which is the case if M and the given form in C^ are smoothsee below) this decomposition leads to the principal theorem of HODGE which is that there is a unique harmonic field in each cohomology class of closed forms (i.e. those co with dco 0) on M. This statement is
7.1. Introduction
287
equivalent to the statement: Suppose M is of class C, co^ C^ (M), and dco = 0; then there is a unique harmonic field H (^ ^^{M)) such that (7.1.2) 0) = H + d(p for some ^^C(M).
The writers mentioned above, except for E E L L S , FRIEDRICHS, and
MoRREY, were not especially concerned either with the differentiability of M and of the forms or with the nature of the elements in the spaces K and '^ above. Not only does the variational method yield a very simple proof of the decomposition in (7.1.1), the finite dimensionality of (in the case where M is without boundary), and the fact that in the general case in (7.1-1) where c ^ and d^^, there exist forms a and ^ in the SoBOLEV spaces HI (M) such t h a t (7.1.3) doc = c, dp = d, doc = 6^ = 0, but also shows that the general results hold for manifolds M of class CJ on which the metric tensor is merely Lipschitz. The success of the method
depends on (a) the use of the SOBOLEV spaces HI (M), ^ > 2, (b) the
recent results concerning the differentiability of ''weak solutions" of elliptic differential equations (see Chapter 5), and (c) an inequality due, in the case of forms, to GAFFNEY [1] and in the case of a single differential equation of higher order to GARBING [1]. The results in (b) above lead to very exhaustive results concerning the differentiability of various solution forms. However, some additional differentiability is obtained by using the special character of the equations (see Theorem 7.4-1). In the case of a compact manifold with boundary, it turns out t h a t the SOBOLEV space H\(M) splits into the two closed linear subspaces H\+ [M) and ^ | ~ {M) of forms a for which noy = 0 and to) = 0, respectively, on the boundary h M, n co and t co being the normal and tangential parts of CO (see 7.5)- The GAFFNEY inequality holds, on each of these spaces separately and this, together with the boundary differentiability theory of Chapter 5, makes it possible to carry over the preceding theory to each of these partial spaces. By using these two spaces and certain potentials defined therein, it is possible to prove first a decomposition like (7-1.1) in which oc^Hl+ and ^^H\-. In this decomposition however, |) m a y have infinite dimensionality (see ^7J)- After this is shown a great variety of boundary value problems treated by D U F F and SPENCER
and CONNOR ([1], [2]) are treated.
It was pointed out that FRIEDRICHS was working on these problems at about the same time as were EELLS and the author. After a number of conferences, it seemed as though there was enough difference between our results as well as our methods to warrant publishing both versions and this was done. The version presented here is mainly due to MORREY
and E E L L S and MORREY [11]. I n 7.27.4, we consider manifolds without
boundary and in 7-57.8, we consider manifolds with boundary.
288
We do not assume that the manifold is orientable; there are completely parallel theories involving only "even forms'' or only ''odd forms'* (see DE RHAM-KODAIRA). We shall assume t h a t all the forms considered are even or else that they are all odd, 7.2. Fundamentals; the Gaffney-Girding inequality Definition 7.2.1. We adopt the usual definition of a Riemannian manifold of class C^ (with or without boundary) and dimension v and define those of class C^, 0 < /* < 1, in the obvious way: any two admissible coordinate systems are related by a transformation of class C^, and if 6 is one such with domain G CRv, the induced components giAx)^Cl-^ on G, Assumption. We assume that any manifold M is compact and of class at least C[.* We shall be concerned with exterior differential forms of degree r {0 <,r '<v) onsi manifold M; we call these simply r-forms. In the domain of any coordinate system, such a form may be represented as follows: (7.2.1) 0) =2a)u...irdxi' A . . . A dx^^
where a)u...ir ^^^ the components of co in that coordinate system and A denotes the exterior product. If two coordinate systems with coordinates (x) and {x) overlap, the relation 0Ji,...ir(^) = lcoj,j,[x(x)\ (7.2.2) f"t~ ^ //I/ I ^ -^- ,
^^^ even forms for odd forms, / = ^{!/ ^g|,
holds between the components. The notation in (7-2.1) is often abbreviated to (7.2.3) oj='^co^dx^ where / [or other capital letter, correspondingly] denotes a sequence in which 1 < H < ^2 < * < V < V. The relation (7-2.2) allows us to define (2 as follows: Definition 7.2.2. An r-form o) onM will be said to be in 2p {resp. HI) ^ its components are in Lp (resp. HI). If oj and rj are ^'-forms of the same kind in S2, we define their inner product by (7.2.4) (CO, rj) = JF{P;
M
co,
ri)dS[P)
* Actually the reader will observe t h a t many of the results have proper generalizations to manifolds of class H^ where p > v. Such manifolds are of class Cj^ with ^ = 1 vjp.
7.2. Fundamentals; the GAFFNEYGARBING inequality
289
where in any admissible coordinate system F{P) oy,ri) î;gi^{x)coi(x)riK{x),

I,K
dS{P) = r{x)\dxK
.. dx'<,
(^^5)
...
I P'H ^ 1 . . . ^ * l ^r aij-fCi
r(x) = [g(x)],112
and g is the determinant of the gij (and the g'^^' matrix is the inverse of the gij matrix). It is easy to see that the value at P of i^ is independent of the coordinate system provided that co and rj are of the same kind. The following theorem is well-known and evident: Theorem 7.2.1. For each r, 0 < r < i^, the totality of even (resp. odd) r-forms in Q^ (^^^^ equivalent forms identified) forms a real Hilbert space with inner product given hy (7.2.4). Forms w are often regarded as alternating tensors and written in the form (7.2.6) co = Jy ZoH....ir dxh A . . . A ^xh.
When this is done the components are defined for all values of the indices from 1 to v and are antisymmetric in the indices, a component being zero if two indices have the same value. In this case, co transforms like an ordinary covariant tensor, except for the factor e in (7-2.2) and F(P;co, rj) has the form (7.2.7) F{P; w,ri)=y^
H
2'^*^^^ -^^'^^^ Wi....*.^^....fc.

ir
where the i's and k's all run independently from 1 to v. We shall not often use this form b u t it is very useful in computations (see the proof of Theorem 7-2.4 below and 7-5, 8.3, etc.). We now wish to introduce an inner product into the space H\. Definition 7.2.3. Let U = ( ^ 1 , - - -, UQ) be a finite open covering of M by coordinate patches Oq with domains Gq (LIPSCHITZ and in Rv) and ranges Uq. If co and rj^Hl (see Def. 7-2.2), we define (7.2.8) ({co, rj))vi = {co, rj) +Z
Q=l fY^ I a = l
SIco^'^Vit-^^'
where the co^^^ and tj^f^ are the components of co and r] with respect to d q. Then we define the corresponding norm (7.2.9) ||coiu = ( K c o ) ) r . Theorem 7.2.2. For each coordinate cover U and each r, the space of r-forms in HI forms a real Hilbert space HI''' with inner product given by (7.2.8). Any two such inner products are topologically equivalent. This is evident.
Morrey, Multiple Integrals \g
290
Definition 7.2.4. If o) is an r-form ^ H\ with r <^v, we define its exterior derivative dm SiS the r + 1form defined by (7.2.10) dw =2^coja^dx'Âdxl.
If r = V, we define dco = 0; we have assumed co given by (7-2.1) or (7.2.3), of course. Remark. We note that if we wish to write
do) = ^ {d(ji>)jdx^
then we must have (7.2.11) (dco) J

y=-l
=2;\-iy-^dcoyldxJy
^
(Jy =Jl>
"> h-1 > Jv+ly , jr+l)
This definition is set up so that the following version of Stokes' theorem holds. Theorem 2.3.7 (Stokes' theorem). If b is an oriented manifold of class C^ in M having dimension r + 1 <^^^ boundary b h of class C^ and if co is even and of class C^ on b, then
(7.2.12)
b
Jdco=Ja)
66
where b b has its usual orientation induced by that of h. We do not need this theorem and therefore omit its proof. Definition 7.2.5. If co is an r-form ^H\ given by (7.2.1) and r > 1, we define its co-differential ^ co by the condition t h a t (7.2.13) [oj,d(p) = (da),(p) for every 99^ H\'^~^. If co is a 0-form, we define ^ co = 0. Theorem 7.2.4. If co ^ HI and is given by (7-2.5) with r > 1, then {dco) = r - ^
y^ ~ ^)r-1 ki...kr-i
2:(dco)jc,...jcr.^dxKA...Adx^r-., + (gl^ +P-^ g'^r^-)copfc....jcr-. +
-(day)k,.jcr_,=g-^co^jc,...Jcr-^x(7.2.14) If CO is even (resp. Proof. We may patch with domain (7.2.4), (7.2.7), and (co, dcp) == ^ f^ghh
G (*).0') d=X I
+2IgVg%g'^C0^k^...Jc,_^tk^^^,.,Jc^_^.
odd), then d co and 6 co are even {resp. odd). choose cp in (7-2.13) to have support in a coordinate G; we assume cp given in the form (7-2.6). Then, from (7.2.11), we obtain . . . gVr coa) \^{~^)'-^
y=l
9^(,.;)
rdx
v=i
(),(>)
'^
(7.2.15)
{i) = i^...i,.
{j'y)=h...jy-ijy+i.--jr.
etc.
7-2. Fundamentals; the GAFFNEYGARDING inequality
291
If for each y we now choose (h, . . ., Zr-i) = (u) and set jy = oc, the last integral in (7.2.15) becomes
y=l
(i)Al)
(7.2.16)
If we now set iy = p, ki- - - kr-i = {iy) and use the anti-symmetry in the indices, the integral (7.2.16) becomes
(fc) (l)
which must, for all (p(i), be equal to

; ^ / 2 ' ^ ^ l ' l g^r-lh~l{dw)ilc) (piDTdx.
By equating coefficients of (p(i), multiplying both sides by gi^si gir-iSr.i and summing over (/), we obtain
from which the result (7-2.14) follows. The reader may verify the last statement. Remark. We notice t h a t if co^l...^,. is anti-symmetric in the indices {i), (dco)ki..,jcr-i^ 3^s defined by (7.2.14) is also anti-symmetric in ^ i . . . ^r-i. and {dco)j^,,j^^^ as defined in (7.2.11) (with J[, replaced by (7^) is antisymmetric in ji . . . jrî. Thus we may write dco or d co in either form (7.2.3) or (7.2.6). Definition 7.2.6. We define the Dirichlet integral by (7.2.17) D{co) == {da),dco) + {dco, dco). Theorem 7.2.5. d is a hounded operator from the whole of HY i'yito 82"^^, and 6 is a bounded operator from the whole of H\^ into S2~^. D[cS) is a lower semi-continuous function with respect to weak convergence in H\'^. If cojc tends weakly to coo ^^^ Hl^ on M, then cok tends strongly to coo in S2 on M. Proof. This is clear from (7.2.11) in the case of d and from (7.2.14) in the case of 6, since the gij are at least Lipschitz and have bounded first derivatives. Now if cojc tends weakly in HI to co, dojjc andScojc tend weakly in S2 to do) and dco, whence the last statement about D{co) follows from the lower-semicontinuity of the norm with respect to weak convergence. The last statement is an application of Theorem 3.4-4.
19*
292
Essentially the following lemma has been proved by GAFFNEY [1]: Lemma 7.2.1. Take any > 0 and r such that 0 < / < r. Given any point %{s^M there is a coordinate system 6 mapping the open hall B (0, q) of radius q in Ev onto a neighborhood U of XQ and a constant I such that (7.2.18) D(co) > ( ! - ) / Z(îxo^)^dx -l(oj, co)
for any r-form co ^ Hl'^ whose support is in U. Proof. Let us set Bg = B{0, q). Then for any co whose support is in U we can write (7.2.19) D(co) = f Z W'^'"^ îxo^ ^JxP + 2 ^^^^ ojixo^ CO J + c^^ coi coj] dx
B, iJ
and aiJ^(x) = aJi^^{x), ciJ[x) =^ cJi(x), where the as are combinations of the gij only and the h's and c's are similar combinations of the gij and their first partial derivatives. Consequently it follows from our assumption that the as are Lipschitz and the Vs and c's bounded and measurable at least. We begin by choosing a fixed coordinate system mapping 0 into XQ and BQ onto a neighborhood of XQ^ with ^^^(O) = dtj. Since the a's are LIPSCHITZ, we may choose q so small that (7.2.20)
J
faiJ^'^x)
o)ix<^cojx^dx > DQ[(O) - I

BQ
BQ
2 J J ^
Ziîxo^)^^^,
where DQ (co) is the Dirichlet integral in the euclidean case with cartesian coordinates. Since the h's and c's are bounded and since for any ?7 > 0 we have \2(X^\ <rjoc^ + rr^^^, it follows that D(co) > Do(oj) e f ^{o)ixo)^dx
BQ
l{o), OJ).
Now suppose that the MJ are of class C^ on BQ and vanish on and near SQ. Then by using the formula (7.2.13), we may write B>o{o)) = [do dco + ddo co, co)o. Using the formulas (7.2.11), (7.2.14), (7.2.4), and (7.2.5) with ^^ = ^^ we find t h a t (do dco + ddoco)o = 2jfîôoJidx= f ^{coix<^)^dx = Do{co).
(7.2.21)
^'
"'
Using approximations, the last equality in (7.2.21) holds for all forms co in HI with the stated properties.
7.3- The variational method
293
Theorem 7.2.6. For each r = 0, . . ., n and coordinate covering U of M^ there exist constants K^ > 0 and L]x such that (7.2.22) D{o}) > i ^ u ( K ^))i: - ^ u K oj) for every o)^H\^, Proof. From Theorem 7.2.2 it is sufficient to prove this for some particular It. Let IX = [U\, . . ., UQ) be an open covering of M by coordinate patches such that each point ;^^ M is in some Uu satisfying (7.2.18) with 6 = - , say. Let Gi, . . ., GQ be the domains in E'^ such that Ujc = Ok{Gk) for all k. There exists a finite sequence (pi, . ., cps of Lipschitz functions on M, each of which has support interior to some Uq, and such that
(7.2.23)
Z(ps[^) = 1
for all x^M. Now if (7.2.22) were false for the U just described, there would exist a sequence {co^} of r-forms in H\^ such that (D cop) and {cop, cop) were uniformly bounded but Ijco^) llu ^^=^- Then, for some s, q, and some subsequence, still called mp, we would have ( ^{(psoyf)l-dx-^ where cps has support in Uq, since s \\(Âvi<^Z\\^sO}p\\n and I 9^5 (0 11^ = (995 (.Op, (fsojp) +2^ I 2^ ((fs a>^^0|a dx. a==i4 I, a But it is easy to see that D{(psO)p) and {(psOOp, (psOJp) are uniformly bounded. From our choice of neighborhoods we have reached a contradiction with the fact that D {(fs (Op)^^
^ J I,oc
Q .
oo,
Z ((Ps cof)l dx.
7.3. The variational method We begin with the following lemma. Lemma 7.3.1. Let 9)1 be any closed linear manifold in the space 2 ^/ r-forms on M. Then either there is no form (joofW which is in H\'^ or there is a form COQ ^^ Wr\HY with (coo, coo) = 1 which minimizes D{(o) among all such forms. Proof. If Wl contains no form in Hl'^, there is nothing to prove. Otherwise let {CJOJC} be a minimizing sequence, i.e. one such that {cojc, (ojc) = 1 and (Ojc^Wl C) Hl'^ ioT k = 1 , 2 , . . ., and such that D((ojc) approaches its
294
infimum for all co^ 501 H Hl'^. From Theorem 7.2.6 it follows that the {{(jojc, a)jc))u are uniformly bounded. Accordingly, a subsequence, still called {cojc}, exists which converges weakly in Hl^ to some form COQ. But from Theorem 7.2.$. cojc tends strongly in Sj to coo and D{co) is lowersemicontinuous with respect to weak convergence in Hl'^. The proof of the lemma is now complete. Definition 7.3.1. A harmonic field co on M is a form in HI on M for which dcD = d CO = 0 almost everywhere. We will let ^ denote the linear manifold of harmonic fields on M of degree r. Theorem 7.3.1. For each r = 0, . . ., r ( = dim M) the linear manifold ^^ is finite dimensional. Proof. The ^|-forms are dense in Sg, since the Lipschitz forms are. Let Wli = 82- There is a form coi in Tli H Hl'^ which minimizes D (co) among all such forms with {co, co) = 1. Let WI2 be the closed linear manifold in S^ orthogonal to coi, and let C02 be the corresponding minimizing form in 50^2 By continuing this process, we may determine successive minimizing forms coi, C02, C03, . . ., each satisfying {cok, ojjc) = 1 and being orthogonal to all the preceding ones. Now if D (coi) > 0, there are no harmonic fields ^ 0 since D (coi) <Z)(co2) < . On the other hand, suppose D{co]c) = 0 for all values of k. Then bj^ Theorem 7-2.6, ((COA;, COA;))U is uniformly bounded in k, whence a subsequence {cop} converges weakly in Hl^ and hence strongly in | to some form coo in Hl'^. This is impossible since the co^r form an orthonormal system in S^. Theorem 7.3.2. For each coordinate covering U of M there is a constant Ao such that (7.3.1) D(co) >Ao((co,co))n for any oj in H^ which is orthogonal to ^'. Proof. For, let coo be t h a t form in H^^ (there is one since each harmonic field is in ^ | ) which minimizes D{a)) among all 00 in HY with (w, co) = 1 and CO orthogonal to ^. Then clearly D (coo) > 0 and by homogeneity (7.3.2) D{a)) >i)(coo)(co,co)
for all C in i^l*" and orthogonal to ^. By Theorem 7.2.6 we see that O (7.3.3) /û((ct>, co))u < {1 + i:u/^(coo)} /)(co), from which (7.3-1) follows. Theorem 7.3.3. Suppose coo is any form in S2 <^^<^ orthogonal to ^. Then there is a unique form QQ in H^ and orthogonal to ^ such that
(7.3.4)
(î3o, iC) + (îô, aC) = (coo, C )
for every C in H^. Moreover, the transformation from coo to QQ is a bounded linear transformation from S^ into H\'^.
7-4. Final results for compact manifolds without boundary
295
Proof. We see from Theorem 7-3.2 t h a t I(aj) = D{w) 2 {co, coo) (7.3.5) > Ao {{CO, oj))n - {hl2) (CO, 0)) - (2/Ao) {coo, coo) > ( V 2 ) {{co, co))n - (2/Ao) (coo, coo) for all CO in Hl^ and orthogonal to ^. Also, since {co, COQ) is a (continuous) linear functional on Hl'^, it follows that I{co) is lower-semicontinuous with respect to weak convergence in Hl'^. From (7.3.5) we see that {{cok, cojc))u is uniformly bounded in any minimizing sequence; the existence of the minimizing form QQ is established as usual. Since co = 0 is allowed we have (7.3.6) I {Do) < 0, (Ao/2) ((iô, Qo))n < (2/Ao) {coo, coo), from which the last statement will follow when the others are established. Now let f be any form in Hl'^ and orthogonal to ^. Then (7.3.7) /(iô + A C) - / (i3o) + 2A [{dQo, dC) + {SQo, dC) -(coo,C)]+A2i)(C). Since / {Do + ^ C) > -^ {^0) for all A, we see that (7-3.4) must hold all f orthogonal to ^ and hence for all C in Hl'^ since any such can writtem uniquely in the form ^ = H -{- Co, where Co is orthogonal ^^ and dH = dH = 0. If Qi also satisfied (7-3.4) for all C in Hl^, would have (7.3.8) {dDo ~ dQi, dC) + {SQo - SQi, dC) = 0 for be to we
for all C in Hl'^, in particular for C = Do Qi, from which it would follow that Qo Oi was harmonic. But then iô iî = 0, since both are orthogonal to ^^. Definition 7.3.2. The form Do of Theorem 7.3-3 is called the potential of cooWe observe t h a t if all the forms in (7-3.4) were sufficiently differentiable, then (7.3-4) would imply that (7-3.9) ADo^ddDo + ddDo=^coo.
7.4. The decomposition theorem. Final results for compact manifolds without boundary The defining equation (7.3-4) for potentials is a special case of the equations (7.4.1) {dco - <p, dC) + {doj - ip, dC) - {rj, 0 = 0 for all C in H\. We will have occasion below to use these more general equations. Referring to the defining formulas for d and d, for any f in H\ whose support is in a single coordinate patch, we see t h a t (7.4.1) is
296
of the form
G
IJ
(7.4.2)
+ t/[&^^ W cojaôc + ci'îx) CO J +f{x)]}dx
= 0
for all C in ^20^ ^^^ symmetry properties of the coefficients and their dependence on the gij and their derivatives have been noted in (7.2.19). The e' s a n d / ' s are linear combinations of the rp' s, ip' s, and rj' s with the coefficients of the 99' s and ^ ' 5 in the e' s and those of the rj' s in t h e / ' s not involving the derivatives of the gij; those of the 99' s and ip' s in t h e / ' s possibly do involve such derivatives. Thus if the manifold M is of class C^ (at least CJ), then the gij and hence the a s and the coefficients in the e' s and of the rj' s in t h e / ' s are of class C^~^; the b' s and c' s and the coefficients of the 99' s and y)' s in t h e / ' s are of class C^"^ (bounded and measurable if k = /bt = \). Moreover, if G = ^ (0, R) and the gij (0) = dij, then the a' s reduce to what they are in ^. By repeating the argument at the end of the proof of Lemma 7.2.1 and assuming that C vanishes on and near dG, we see that we may add certain constants to the a' s which do not affect the value of (7.4.2) for any such C, whence (7.4.3) a^^^(0) =d}l...did''^
Thus the equations (7-4.2) are of the type discussed in 5.2 and 5-5. From the results of 5.2 and 5-5 we may draw the following conclusions about the harmonic fields and potentials discussed in 7.3. Corresponding results hold for the more general equations (7-4.1): (i) / / M is of class C\ and co is an r-form in 2 on M, then the potential Qofo) is in HI on M if r > \;Q and its derivatives are in HI in any coordinate system if r = 0. If oj^2p with p'>2, then Q^H\ and in C|^ if p '> V, fx = \ vjp. (ii) / / M is of class C^, where 0 < /^ < 1 and co^C^^, then Q^ Cj^ if r>\]Q^Clifr = 0. (iii) / / M is of class C\ and 0 ) ^ 8 2 , then the components of Q and their derivatives in any coordinate system are in H\. If co^ S>p, then Q and its derivatives are in H],', Q^ C]^ if p ^ v and ju = \ - vjp. (iv) / / M is of class C^, 0 < /^ < 1 and ^ > 3 <^^<^ co $ C^""^, then ^ C^-^;ifco^Cf^^ then Q^Cîfr = 0. (v) / / M and co are analytic or C^, then Q is analytic or C^, respectively. (vi) / / 0) is a harmonic field and M is of class Q , then co is of class C^~i (CJ for 0-forms); if M is analytic or C, then co is also. We begin with a lemma which follows at once from the relations d d = d 6 0 whenever M and the forms are of class C^.
7-4. Final results for compact manifolds without boundary
297
Lemma 7.4.1. Suppose that (x and ^ are in H\ on M. Then (doc, d^) = 0 and [da, ^) - [oc, d^). Proof. Suppose U = (C7i, . . ., UQ) is an open covering of M by coordinate patches. There is a sequence (pi, . . ., cps of Lipschitz functions such t h a t 995 > 0, 991 + + 995 = 1 on M and such that the support of cps + (ft is contained in some Uq whenever the supports of 995 and (pt intersect. If we write as = cpsoc and ^t = ^t P, then [doc, d^) ^Z^ôcs.
s,t
d^t),
[doc, P) =Z(ôcs,
s,t
^t),
the sums being over all ordered pairs (5, t) such that the supports of tps and (pt intersect (and hence both lie in some Uq). Thus the proofs of both parts are reduced to the case where both have support interior to some Uq. But then in the domain Gq, we may approximate to the components of oc and ^ strongly in ^ 2 ^ y C^ functions; we may approximate to the gij by functions gptj of class C^ in such a way that the gpij satisfy the same Lipschitz condition as gij, and the respective first derivatives converge almost everywhere to those of gij. The result follows easily. Theorem 7.4.1. Suppose M is of class CJ, co is in S2 on M, and Q is the potential of co. Then (i) a = dQ and ^ = dQ are in HI and satisfy (7.4.4) CO = doc + dp = 6 (dQ) + d(dQ), da = dp = 0. (ii) If co^ HI, then a and ^ are the potentials of dco and doj, respectively, so that doc and d^ are also in HI. (iii) If a)^Sp(p '> 2), then Q, dQ, and SQ are in HI and in C^ if p '> V and /Lt = i vjp. (iv) / / M is of class C^ with 0 < /^ < 1 and co^ C^^~^, Ihen Q, dQ, and SQ are in C^~^ with k '> 2. Proof. We shall prove that (7-4-4) holds in some neighborhood of each point of M and will also show that each point J\: of M is in some neighborhood U such t h a t (7.4.5) (doc, dC) + (6 oc, 60 - (dco, C) = 0 = o =0 = 0 (dp,dC) + (dp,SC)-(dco,0 for all C in HI with support in U if co is in HI, or (7.4.6) (doc, dC) +(doc-co,dO (dp - CO, dt) + (dp,dC)
for such C if C is merely in S2. Since M can be covered by a finite number O of such neighborhoods and any l^^H\ may be written as the sum of a finite number of Cs, each of which has support in some one such neighborhood, (7.4.5) and (7.4.6) hold for all C in H\ on M. The regularity
298
results follow from equations (7.4-5) and (7.4.6), the results for systems of 5.2, 5-5 and 6.4, and the remarks at the beginning of this section. Choose PQ^M and an admissible coordinate system 6 with domain G = B{0, R) such that gij(0) dfj. Then the equations (7.3-4), when expressed in terms of the components Qj (and certain integrals of jacobians which have constant coefficients and which vanish if the components ^-^20 ^^^ added (cf. 7.2)) take the form (5.2.2), where B{u, v), Bi(u, v), and B2(u, v) are defined in (5.2.16) and the coefficients satisfy (5.2.17) and (5.2.18) (with//I = 2 since our coefficients are bounded). Let us approximate the components coj, assumed in H\[G), and gij by functions (Dpi and gpij^ C'^ (G) so that copi-~>a)i in HI (G) and the gpi j -^gtj uniformly so that all the gpij satisfy the same Lipschitz condition as the gij and their derivatives converge almost everywhere to those of the gij. It is clear from Theorem 5-2.1 t h a t we may assume that R is so small that Bp (u, u) > mi (II u 111)2, u $ H\Q (G), mi = w/2, for every p (and degree r) and t h a t (5-2.6) is satisfied for a l l ^ . Then, by writing Qpi = Oi + rjpj, rjpi^HlQ(G), we see using Theorems 5.2.1 and 5.2.2 that, for each p, there is a unique solution vector Dpi of the equations corresponding to (7-3.4) with coo replaced by cop and which coincides on dG with Qi. From the interior regularity theorem ( 5.2, C ^ case) it follows that each Qpi^C'^(G). Also, from Theorems 5.2.1 and 5.2.7, we see that Qpi -^Qi in H\[G). It follows that the vectors ^j>j = {^^p)j and ^pK = (dQp)K converge strongly in 22 {Q to ocj and ^K' For each p it is easy to see that the vectors ocpj and ^pK satisfy the equations corresponding to (7.4.5) and (7-4.6) with the coi replaced by the copi. From Theorems 5.2.1, 5.2.2, 5.2.5, and 5.2.7, it follows that the a^j ->(xj and ^PK-^PK strongly in H\[r) for each 7^ with F cG. Thus a J and ^K H\ [F) for such F and satisfy the equations corresponding to (7.4.5) and (7-4.6). If the coi are merely in 82 {G), we can still approximate to coi strongly in S2(^) by the copi. In this case we still have the strong convergence of apj to ocj and ^pK to ^K in HI (F) ior F ^G so that oc and P^H\ and satisfy (7.4.6). In case co^ H\y the equations (7-4-5) are the defining equations which show t h a t oc and ^ are the respective potentials of doy and d a>, so that d a, doc, d p, and d^ all ^ HI. We next prove a decomposition theorem for r-forms in S2 on M which strengthens the result of K. KODAIRA mentioned in our Introduction. Theorem 7.4.2. If oo ^21 on M, then there are H\ forms H, oc, and p, where H is a harmonic field, such that (7-4-7) o} = H + d<x + dp doc = dp = 0, oc = dQ, p = dQ,
7.4. Final results for compact manifolds without boundary
299
where Q is the potential of co H, The three forms in (7-4-7) (ire mutually orthogonal in fi^, and the totality of the elements in each is a closed linear manifold. If Hi, oci, î are any forms in HI, where Hi is a harmonic field and (7.4.8) a)=Hi-^d<xi-{-dPi, then Hi = H, d oci =^ d a, and dî = d^. Proof. Since *" is finite dimensional, there is a unique H ^^^ which is nearest co. Then co ^ is orthogonal to ^, whence its potential Q exists and has the properties mentioned in Theorem 7.4.1. Thus (7-4.7) defines a decomposition which is easily seen to have the desired type, using Lemma 7-4.1. Moreover, suppose {ocp} is a sequence of HI forms such t h a t dap = 0 for each p and docp->some TJ strongly in 2, each ap = dQp being orthogonal to '*; then we see that D (ocp ocq) = (d(Xp docq, 6ocp Sag) - > 0 ; it follows that the ap form a Cauchy sequence in HI, using Theorem 7-3-2. Hence the ocp - ^ some oc strongly in HI with da = 0, d oc = 7]. A similar proof holds for the manifold oi d^' s. Next, suppose t h a t Hi, oci and î are HI forms satisfying (7-4-8), where Hi is a harmonic field. From Lemma 7-4-1 we see that the three components are mutually orthogonal, from which we immediately conclude t h a t Hi = H and d (oc ai) = d (î P). Since these last two are orthogonal by Lemma 7-4-1 they must both be zero. Theorem 7.4.3. If w^ H\ on M, then in the decomposition of Theorem 7-A-2 we have d a and d^ in H\ also. Proof. This follows at once from Theorem 7-4-1, since H is always in H\. Theorems 7-4-2 and 7-4-3 lead immediately to the following theorem: Theorem 7.4.4. Each cohomology class of forms in HI contains a unique harmonic field. Proof. For if dw = 0, then doc must be closed and hence zero in (7.4-7)Remarks. The further differentiability properties of the forms doc and dp in (7-4-7) follow from Theorem 7-4-1 and the other regularity properties stated at the beginning of the section. In his book. Geometric Integration Theory, H. Whitney [2] considered a class of differential forms (the "flat" forms) analogous to those described in the following definition. Definition 7.4.1. An y-form co in 2 on M is said to be S2-flat <^ there exists an S2 form rj and a sequence {cop} in HI such that cop-^co and dcop -^r] strongly in S2 on M. Theorem 7.4.5. An r-form co in S2 on M is S2--fl^l tf ^'^d only if 6 a^H\ in the decomposition (7-4-7)-
300
Proof. Let co be S2flat, and let rj and {co^,} be as in Definition 7-4-1. If we write (7-4.4) for each p, we have Hp, d ocjt, and d^p in H\, where ocp is the potential of dcop. It follows that a is the potential of rj, whence ^ / is in H\. % Conversely, suppose d aisin H\. We may approximate to co strongly in 2 by {a>^} with each co^ in H\. Then if we write (7.4-4) for e a c h ^ , the forms /^^ and d^'^ are in H\ and d^'^ tend strongly in S2 to d^. Now, if we define ooj, by 0}p = H + doc + d^'^ we see t h a t {co^,} is a sequence as in Definition 7-4.1; it follows that C is S2-flat. O 7.5. Manifolds with boundary Definitions 7.5.1. For a ^^-dimensional manifold M with boundary 6 M of class C^(0 < /^ < 1) (C^, analytic) we adopt the standard definition : each point of M U & M is contained in some set 3^, open onMyjhM, which is either the homeomorphic image of the unit ball in Rv or of the part of it for which jt^ < 0 in which latter case, the points where x^ = 0 correspond to ^t O b M; any two coordinate systems are related by a transformation of class C^(C'^, analytic). Any coordinate system with a Lipschitz domain G is admissible if it is related in this way to one of the preferred ones and 'St H b M corresponds to a part of d G. We still allow even and odd forms and define the spaces 2p and HI and the inner products (co, rj) and ((co, rj))]x and their corresponding norms as before. The differential operators d and 6 are defined as before. Assumption. We assume M compact and connected and of class at least C\. Theorem 7.5.1. The spaces S2 (^nd H\ of forms of a given kind and degree are Hilbert spaces. The operators d and S are bounded operators from HI to S2, preserving even-ness or odd-ness, and D{m) is lower-semicontinuous with respect to weak convergence in H\. Finally, if con -^co in H\, then con -^co in Qp. In order to define boundary values and the normal and tangential parts of a form on b M, it is necessary to define admissible boundary coordinate systems. Definition 7.5.2. An admissible boundary coordinate system on a manifold M with boundary bM oi class C^ is a coordinate system of class C^ which maps its (Lipschitz) domain G\J G onto a boundary neighborhood SI in such a way that a, the part oidG on x'^ =^ 0, is not empty and open and is mapped onto Slr\bM, and such that the metric is of the form
v-l
(7.5-1)
ds'^ =:^ '^gy6{x^,0)dxydx^ ^[dx^'Y
on o*.
7.5- Manifolds with boundary
301
Lemma 7.5.1. (a) / / M (U hM) is of class C^, each point PQ of bM is in the range of an admissible boundary coordinate system. (b) / / ^ > 2, each point PQ of b M is in the image of a coordinate system of class Cj^"^ which satisfies all the other conditions of Definition 7.5.1 (^'yid is such that the metric takes the form (7-5.1) iîth [x^, 0) replaced by any x in G. If M is of class C^ {analytic) this coordinate system may be taken to be of class C^ [analytic) and hence admissible. (c) / / [x) and [y) are overlapping admissible boundary coordinate sytems [or two systems as in (b)), then,
(7.5.2)
y,(-;,o)=y:(v;,o)-o,
oc<v,
y:,(.<,o) = i .
Proof. Let [x) be any coordinate system of class C^ which maps G\J a onto a boundary neighborhood ^ of PQ in the usual manner. Suppose (I) is another similar coordinate system and let 0^^ and g^^ be the metric tensors of the (f) and [x) systems, respectively. Then we want C'va = dva, at least if k^ 0, so t h a t we want
at least on ^^ = 0. If we multiply (7.5-3) by d^'jdx^ and sum we obtain (7-5.4) g^^-^ = ^=.r6l, r = ^^,
since, of course, |^ = 0 when x^ = 0. Multiplying (7.5-4) by g^^ and summing, we obtain
(7-5.5)
fj; = ^ ^ '
where we find, by multiplying by d^^jdx^, using (7.5-4) and summing on X, t h a t (7.5.6) r'^g^^= 1. A coordinate system as in (b) may be obtained for | in a 1-sided neighborhood of a by solving the differential equations (7.5-5) with initial conditions
(7-5-7)
(7.5.8)
^M^;.o)=|\ x<v,
| | = ^., g
^-(i;,o)-o.
If the [x) system already satisfies (7-5-1) it is clear t h a t T=l, = ^J on 4' = 0
which proves (b). In order to obtain an admissible coordinate system (f), it is necessary only to define functions x^ of class C^ on a 1-sided neighborhood of a which satisfy (7-5.5), (7-5-6), and (7-5-7) along a. This may be done by
302
defining
(7.5.9)
x^(K, ^') = ^"f^^tWv - K)f{%) <^Vv>
for instance, where 5(f^, S*') denotes the (v 1)-dimensional ball with center at |^ and radius I*' > 0, cp is a mollifier in Rv_i and (p* (y) Definition 7.5.2. Suppose o) ^ HI on M. We say that the tangential part tco or the normal part n co of co vanishes on h M \i and only if (7.5.10) (7.5.11) ft>n...r'0) =-0 if all iy<v, or f^n...ir(^v> 0) = 0 if some iy = v,
respectively, x being an admissible boundary coordinate system. Remark. These are invariantly defined by virtue of (7.5-2). Lemma 7.5.2. (a) / / ojn 7 oj in H\ and t ajn(n cun) vanishes on bM for each n, then ta)(na)) vanishes on b M. (b) / / M is of class C* with ^ > 2 and co^ C^ and tco 0 (nco = 0) on b M, then t dw = 0{n d o) = 0) on b M. Proof, (a) follows from the theorems in Chapter 3 on HI functions. Using a partition of unity, one sees that it is sufficient to prove (b) for forms CO having support in the range of some admissible boundary coordinate system. For such an co, one sees immediately from (7.5-10) and (7.2.11) that if (7.5.10) holds, then all the (dco)j^,, ,j-^+^ with all;V<'^ satisfy (7.5.10). Also, if (7-5.11) holds we see from (7-2.14) and the facts that gpoc = g*'' = dvocin a,n admissible boundary coordinate system, that ( 5 co)^l... ^y_i = 0 if any iy = v. This proves the results. < Definition 7.5.3. If 99 and yj are given by (7-2.1), we define (7.5-12) {cp,^){P)=F{P;cp,xp) respectively, where F{P;q), \p) is defined in (7-2.5). Lemma 7.5.3. Suppose oc and ^ are r and (r \)-forms in HI which are of the same kind. Then (7-5.13) [oc,d) = [doc,P)+
bM
j{{-\y-^noc,t^)dS{P).
If either naort^ = OonbM, then {d oc, d^) = 0. Proof. We prove (7-5.13) first. There is a partition of unity Cs, s = 1, . . . , 5 , such that if the supports of ^s and ^f. intersect, their union is in the range of some one coordinate patch. Thus, by setting as = Cs oc, Pt = Ct P> we see that it is sufficient to prove the theorem for the case that oc and ^ both have their supports in the range 9^ of some one coordi-
7.5. Manifolds w i t h b o u n d a r y
303
nate patch. If ^ C M^^\ (7.5.13) follows from 7.2 with no boundary integral. So suppose 5y?: is a boundary neighborhood and [x) is an admissible boundary coordinate system with domain GR U OR (in this chapter ^^^ < 0 on GR). We may approximate the gyd and and the components of oc and ^ by smooth functions (Theorems (3.1-3) and (3-4.1) on GR. Then by looking at (7.2.15) we obtain
^ ' ' ^ "^ ^ r\a^ yd) (?)
Now if for each y we set jy = v and hence iy = v, since ^''^ = 0 if a < r, move the index iy ( = r) to the first subscript of oc, replace ^^ by ^ 1 . . . y^r-i and j!^ by /i . . . /^-i, we obtain (oc,d6) = {doc,6) + - r - ^ I Ig^^^^"'g^'-''^'-'^^vki...Jcr~i X
from which (7.5-13) follows since all the h . . . /r-i are < v, since jy = v. Clearly dS(P) = r(x) dS(x) since gvx(x) = dv<x for x on (TR. To prove the second result we may approximate as above and, since d^ (p = d^ (p = 0, we see that {da,d^) = f(nSocJ^)dS{P)
bM
= 0
if (X,fiand the metric are smooth, since it follows from Lemma 7-5-2 t h a t ^ ^ < x = = 0 i f ^ ^ = 0 o n & M . The result follows by passing to the limit. The following corollary will be useful later: Corollary. If M is of class C^ with ^ > 3 and co ^ C^(M), then {dco,dtp) + {dco,d7p) = (Aoj,if) (7.5.14) + J (-\y{(nda)JC)
bM
(idoj,nC}}dS,
A = dd + dd.
I t is now important to develop a formula for DQ (CO) , where (7.5.15) Do(o)) = (do), da))o + (SQCO, doco)o, do and the inner products being formed using the Euclidean metric, o) having its support in GR (perhaps 7^ 0 on OR) and being smooth. From (7.2.11) and (7.2.14), we obtain
(dco)mi r+l . . . mr+i = ^ ( 1 ) ^ " ^ CO{my) x^y = I
(7.5.16)
{do a))si . . . sr-l
^=1
COasi . . Sr-1 X^ -
We substitute the expressions (7.5.16) into (7-5.15) and form the inner products using the form (7.2.7) and the Euclidean metric in which g^^ = d^K etc. We obtain (7.5.17) (dco,da))o = , , \ , j I I(-'^y'-^co{m'y)xmyCO{m'y)xm6dx.
304
We break the sum on y and d into the sum where y = d, that where y <C d and that where y > ^. In the first sum, we replace m^ by ii . . . ir and call My = oc. In the sum where y <i d, we move the index ms to the first place in m^ and move nty into the first place of w^, using the antisymmetry of the indices; we then replace m^Q (in which both niy and ms have been omitted) by si . . . Sr_i and call ms = ^ and my = oc. We do the corresponding thing for the sum where 7 > (5. Then we combine these two sums with the integral for (^0 cjo,doO))o. The result is
^0(a>) = ^ J ^ K..Ar'^^^^ + Jy^Z^ ^ ^ {(ôcsi.,,Sr-ix-X
X CO^si. . . Sf-l xP OJxsi...
Sr-1 xP OJ^si. . . Sr--[ x^') ^^
for smooth forms, from which we obtain (7.5-18) Do(a)) = (^(D%<^dx + f ^ {OJVTCD^TXP COVTXÔJ^T) dS
T : \ < 51 < < Sr_i < V
for any form co $ C^ on Gji which vanishes near the spherical surface of GR. Lemma 7.5.4. (a) / / co (considered as a set of functions) ^ H\ on GR, CO is zero on and near the spherical part of the surface of GR, and if either t(jo = 0 or nco = 0, there exists a sequence cop of similar forms of class C^ on GR which converge strongly in H\ on GR to co and such that t a)p = 0 or n cop = 0 (respectively) for each p. (b) If o) satisfies the hypotheses o/(a), then (7.5.19) D^(M) = fS{îi)x^)^dx.
Proof. Clearly (b) follows from (a) and (7.5.18). Also, since the condition t CO = 0 is just the same as saying that the co' s with ir < v vanish on OR and w co = 0 is the same as saying that those with ir = v vanish on OR, part (a) is just reduced to proving the theorem for functions. If the function co is not required to be zero on CR, we extend co to the whole of BR by co(x^,xl,) =co(x^,xl) and then note t h a t the h mollified functions with respect to a spherically symmetric moUifier of co are of class C^ and have support C BR if h is small enough; these tend strongly in HI to CO on BR. If co is zero on OR, we begin by extending co to BR by co(x'',xl,) = co(x^, xl) and then proceeding as above; we note that the mollified functions vanish on OR. Lemma 7.5.5 (GAFFNEY [1]). With each point P of M and each ^ 0 is associated an admissible coordinate system with domain G and range 11 and a constant I such that D(a)) > (1 - e) / 2 ; (oJi:,o.)^ dx - l(ay, oy)
'].(i. Differentiability at the boundary
305
jor any jorm co^ HI with support on U and either t(o = 0ornco
0on
Proof. This has been proved for interior points in Lemma 7.2.1. If P is a boundary point, it is clear that we may choose an admissible boundary coordinate system with domain GR U GR which carries the origin into P and GR U OR into a neighborhood of P in which gij{0) = dij. Then, exactly as in the proof for interior points, we conclude that we may choose R so small t h a t
D[cd) >DQ{O)) e j 2^ {(Oixo^)^ dx l(co, w)
for some / and all oy ^ H\ with support in 9^, DQ {CO) having its significance in Lemma 7-5A- The result follows from that lemma. The following theorem follows from the lemma above in exactly the same way as in the case of Theorem 7-2.6. Theorem 7.5.1. For each finite system 91 of admissible coordinate systems whose ranges cover M, there are constants k and I such that D{o)) '>k\\ojf-l{oy,oy) (k > 0) for any form in HI with either tco = 0 or nco = 0 on b M, the norm being that corresponding to 9?:. 7.6. Differentiability at the boundary In this section, we discuss the regularity at the boundary of the solutions of the equations (7.4.1). The resulting systems (7.4.2) for boundary neighborhoods are not quite the same as those in Chapters 5 and 6 because we are requiring only that some (normal part or tangential part) of the components vanish on b M. However, the general ideas are the same as those in 5.2, 5.5 and 5.6. On account of equation (7-5.18) and Lemma 7.5.5 and its proof, it follows that if P o bM, there is an admissible boundary coordinate system with domain GR in which the origin corresponds to P Q and goi^{0) == docp, and if either nco or t(p = 0 on bM, then the corresponding boundary integral to t h a t in (7.5-18) (which then vanishes) may be subtracted off and the equations (7-4.1) take the form
/ [<a f ^j + K ^' + 0 + ^' K < + îi ^' + h)\ ^^=^^

GR
(7.6.1)
v^H\t,
at^{0)=dijd-^
i,j=\,...,N,
where H'^ denotes the set of all vectors v ^ H\ such that (7.6.2) all v^ = 0 on 2Ê, v^=^0 on (Ti? for i=\,...,k (0 <k <^n).
1 Here as elsewhere if (x> has support in a boundary neighborhood 9^, it is not required to vanish on 6 M Pi 5^. Morrey, Multiple Integrals 20
306
k being given. We have already noted that the as are combinations of the gap whereas the h's, c*s and d's involve also the derivatives of the g's] the e^ are combinations of the g's and the non-homogeneous terms in equations (7.4.1) and the ft involve these and the derivatives of the g's. We shall obtain results concerning the differentiability of u by obtaining such results for U == C ^ where f is a 0-form ^C'^(G/j) which has support on GR\J OR; results for interior neighborhoods have already been obtained in 7.4- Assuming that v has support in GR U CR and t h a t v^ Hl^, we replace v^ by f v^ in (7-6.1) and find that U satisfies (7.6.1) with e and / replaced by E and F, respectively, where
^ ^
Fi = Cfi + Â<f<p + K^' - ot,u^ + <).
So we begin by considering solutions u of (7.6.1) which have support on GR U OR and which $ ^ | J . We begin by altering the coefficients outside GR as in (5.5.3) to obtain new coefficients a^fn, etc. Then, if u^W^, is a solution of (7.6.1), and has support in GR U OR, it is also a solution of / [vU i<^R <P + KR^'' (7.6.4) We then write, as in (5.5-5),
U=ÛR + HR, + dR'U +f]
+ <) + ^' KR < v(^Hl*.
+ îJR^^ + /^)] dx = 0,
(7.6.5)
UR = Q2R [{aR ao) - Vu + BR-u + e] +

+ P2R [CR'Û
where we now wish to define the vectors U = Q2R{^) and V = P^nif) to satisfy the boundary conditions (7.6.6) U^ , 0) = F*' , 0) = 0, i= \, . . ., k. So we define Q2R and P2R as follows: Definition 7.6.1. We define the vectors U = Q2R(e) and V = P2R{f) as follows: li i <, k we define U^ [x) and F* (x) as in Definition 5.5.1' and Equations (5.5-19), (5.5-20), and (5-5.21). If i>k, we define P2R (/) as the restriction to G2R of the potential o f / o v e r B2R, w h e r e / i s defined, by e x t e n d i n g / t o B2R by positive reflection: (7.6.7) fK-x'')=f.^n> with the usual modification ii v = 2. li i "> k, we define e by positive reflection and then define Ui(x) JZ (7,6.S) W^x) = ^1
G2R
f-KoA^
- i)eni)d^
W%{x), me'(S)dS>
[Ko{x - i) - Ko{x' x^) if X = [x^, x"^).
x' = (xl,
7-6. Differentiability at the boundary
307
Lemma 7.6.1. (a) / / U = Q2R(e) and V^P^RIJ), e^LziG^R), / Lp{G2R), p > 2vl{v + 2), p > \, then U^ and F* vanish along OZR if i <.k and. U^ and F* are solutions of
^2R GzR
(7.6.9) If i > k, U^ and F* are restrictions to G^R of functions which are defined on Rv and satisfy {7.6.7); they are solutions of (7-6.9) for all V^H\{G2B) which vanish on ^^R(b) If H^ HI{G2R), vanishes along G^R and satisfies (7.6.10) fv,,d-^H,fidx
GzR
= 0,
V^HI^{G2R),
then H is harmonic on B2R if extended by negative reflection. IfH^H\ [G^R) and satisfies (7.6.10) for all v ^ HKG^R) which vanish along^^^, then H is harmonic if extended to B^R by positive reflection. Proof, (a) The results concerning U^ and V^ for i <,k were proved in Theorem 5-5-'i'- If the ^^^ C^ and have support in G^R U O^R, then the ^? G^^[B2R) and have support in B2R. Since K[s(x I) KQ[X' |) is the Green's function for the lower half space, it follows that W^, as defined in (7.6.S) ^ C^ on each closed half-space (but not necessarily on the whole space), this can be proved by writing Wi(x) = - jKoix
B2R
- I) eîS) di + 2jKo{x
Gtn
- |) e^{i) di
and using the method of proof in Theorem 5.5.1'. It follows that U^^ CJ^(G2R). If we approximate uniformly (with the same Holder condition) to the ^" by smooth functions e^^ and (7-6.9) holds for each n, it will hold in the limit. Then /^,a((5^ Ul, + el,) dx = fv{Ul,
GzR 02R
+ el,) dS~-Jv{AUi
G2.R
+ <,.) dx.
(7.6.11) By integrating by parts, we obtain
Ui{x) =f-Ko(x-^)
From Chapter 2, we conclude that AWi=el, (x),
s'elU^)
di +
Wl^x).
AUi(x) = - Z el,,, (x)
Ui,Ax) = Wl,,{x)ÂWi as ^^'-^0 since W^ = 0 along cr2/2- Thus the integral on the left in (7.6.11) vanishes for each n. If ft^ CI{G2R), then/^^ CI{B2R) iii>k and so Zl F^ = fi and F*y = 0 along G2R since F*' is even with respect to x^. 20*
308
In part (b), it follows from (7^6.10) and the interior differentiability theorems of Chapter 5 that H is harmonic on G^R- î H vanishes along (72ij and H is the extension of HtoB2R by negative reflection (see (5.5.19)), then H is easily seen to ^H\{B2R) by using the absolute continuity properties of Â. If, in (7.6.10), we replace G^R by ^272 and let
V ^ C ^ (B2R), we see t h a t
(7.6.12)
B2R v[x^,x^)
G2R^ =v[x,,x^) v[x^, X^)^H\^{G2R).
If we do not assume that H vanishes along a^R b u t do assume that (7.6.10) holds for all v which vanish along^2i?, we obtain a result like (7.6.12) if we let H be the extension by positive reflection, let v ^ C^ [B^R) and define v by
V {x[., A;*') ?; [x^, x'^) + V [x^, % * ' ) .
The developments in proving regularity at the boundary now proceed as in 5.5- We define the norms and spaces as in Definition S-S-"! except that the spaces C^^Q{G2R) will consist of those vectors in C|^(S2i?) which vanish along^^jj (the spherical part oidG^R', recall that, in this chapter G2R is the lower hemisphere); we do not require them to vanish along (72ij. However, the range spaces H\^ [G2R) and CJ* {G2R) of the operators Q2R and P2R are spaces of vectors which satisfy the boundary conditions in Lemma 7-6.1. With t h a t understanding we obtain the result. Theorem 7.6T. The results of Theorem 5.5-1 hold. In our case, the coefficients UR are Lipschitz and the hRy CR, and da (in (7.6.4)) are bounded and measurable if the manifold M is of class C\ and, more generally the UR^ C^~^ and the bn, CR, and dR^ C^~^ if M is of class C^, ^ > 2, and so in these cases satisfy the H^ conditions in Definition 5.5-2 for any q. We note again that if M is of class H^ with p "> V, then the coefficients satisfy the H\-conditions of Definition 5-5.2 with q <.p- So if we write the equation (7-6.5) in the form
UR TRUR=-VR TR UR + WR, VR - : Q2R{e) + P2,R[f) +
= Q2R [{aR ao) ' V UR + hR UR] + P2R [CR '\/UR
dR- UR]
dR-UR],
WR = Q2R [(aR ao) ' \7 HR + hR HR] + P2R [CR ' V HR +
(7-6.13) we obtain the theorem Theorem 7.6.2. The results of Theorem 5-5.2 carry over to our case. Now, we notice, if our given data e a n d / vanish outside G2R and i^~, we may carry over the analysis with R replaced by any A ; we then notice
t h a t Q2A{e) = Q2R{^) and p2R{e) = P2A{^) for any A and we m a y let
A -> 00. If ^ vanishes outside GR, we then see that the harmonic vector
'].']. Potentials; the decomposition theorem
309
H2A = H^R can be extended to the whole space where it is bounded so t h a t we have. Theorem 7.6.3. / / U^H\{GR), vanishes on^^, and satisfies (7.6.4), and if HR is defined in (7-6.5), then HR = const, if v = 2 and HR= 0 if r > 2. If V = 2 and i < k, H^^ = 0. Ifv = 2 and i > k, then 11% is the average value over B2R of the ordinary logarithmic potential of ff, where ff is the extension to B^R hy positive reflection of c%^^ u[^ + dRij u^ + ft. Remark. One must recall the definition given in 5-5 of P2i?(/) when r = 2. Thus, if e a n d / ^ Lq{G2R) for some q > 2, for example, R is small enough, depending on q, ^ vanishes outside G2R, U^H\{G2R), U^ = 0 along (y2Rii i <= k, and u satisfies (7-6.1), then u^ H\[G2R) and
Corresponding results hold if e a n d / ^ ^^0(^272)- But now, suppose that the components u are components of a form a> which satisfies equations (7.4.1) in the large on M and suppose we assume that the forms q),ip, and Tj^^qior some ^ > 2. We assume that n co ^=^ n ^ = 0 ior d^ forms o) and C considered or else that ^ co = ^ C = 0. Then we know that Q)^H\ at any rate and from our interior results o)^H\ on any D with D d G M. In any boundary neighborhood, then U = C '^ satisfies (7.6.1) with e a n d / replaced by E and F as given by (7.6.3). Now the terms ^ ^f, Cfi, and C,a ^f ^ &q but the second term in E and the terms V C (^ ^) ' uinF^ 2s, where s = 2r/(r 2) and the term V C * ^ * V ^ merely ^ S2. If ^ < s, we conclude from the theorems above that each U^H\', if q'>s, we conclude that U^H\. Since this is true in each sufficiently small boundary patch, we conclude that (jo^H\ \i q ^s ox to H] otherwise. In this latter case, then we conclude that u^ , where s' = V sl(v 5). We may repeat the argument (with smaller patches if necessary) to conclude that 00 ^H\ if ' < s' or in H], otherwise. Clearly any q may be reached in a finite number of steps. If M is of class CJ, 0 < // < 1, and if 99, ip, and ty ^ C^, we first show that co^Hl for some q '> V which implies that u^C^, so that -E and F^C^ so that co^Cl. The higher differentiability results can be obtained as in 5-5, 5-6. 7.7. Potentials; the decomposition theorem In this and the next section, we shall assume that all of our forms are of the same kind, completely parallel theories being obtained for each kind. Definition 7.7.1. We define the closed linear manifolds ^ J and ^ ^ (see Theorem 7.5-1) of i / | as the totality of forms in HI for which nw = 0 and t (JO = 0 on b M respectively.
310
Just as in Section 7-3, we obtain the following result: Lemma 7.7.1. Let 9K he any closed linear manifold of S2 such that 9)1 n ^ J ($^) is not empty. Then there exists a form co in ^ r\ ^t (^2^) which minimizes D [w) among all such forms with (co, co) = 1. Theorem 7.7.1. The manifold ^+(~) of harmonic fields in $ J ( $ ^ ) is finite dimensional. Theorem 7.7.2. If a> ^ $2^($2^) ^^^^ ^^ S>2-'Orthogonal to +("), then there are positive constants A+ and %- for each 3^ such that D{a)) >A+||a)l!2(A-||a)P). Theorem 7.7.3. / / ^ C S2 ^'^d is 22-orthogonal to "^(^~), there is a unique form Q+(Q-) in ^ J _L "^(2" J- ~) ^^^^ ^^^^
(77.1)
(î3+ dc) + {dD+, SO = (^, c), ce îm)>
Definition 7.7.2. The functions i^+ and Q~ are called the plus-potential and minus-potential of YJ, respectively. The defining equations [JJ-^) for the potentials are a special case of the more general equations (7.7.2) {dcD ~ cp, dC) + {dm - yj, dC) - {rj, C) = 0, C^ î or $"
which were discussed in Section 7.4. The differentiability results for such equations on the interior of M follow from the discussion there given. But now, suppose we select a point P on b M and choose an admissible boundary coordinate system with domain GR and range a boundary neighborhood H of P such that gij(0) = dij. In such a system the conditions no) =^ n ^ = 0 for ^ ^ and toj = tC = 0 for ^ ^ correspond under a proper ordering of the sets / and / t o the equations (7-6.1). Accordingly we see that if the support of f is confined to IX, the system (7.7.2) reduces to the system (7.6.1) discussed in Section 7-6- From the theorems of Section 7-^ we may conclude that the differentiability results for the plus and minus potentials stated near the beginning of Section 7A for potentials, hold right up to the boundary. We now extend these results as in Section 7A and summarize as follows: Theorem 7.7.4. Suppose co^ S2 0 "'"(22 0 ~) (^f^d Q is its plus {minus) -potential (i) / / M is of class C\, then Q, dQ, and dQ are in ^ ^ (^2^)(ii) / / M is of class C\ and oj is in Sp with p '>v, then Q, dQ, and dQ are in ^ ^ (^~) and C^ if ju = \ vjp. (iii) / / M is of class C^ and w ^ C^'^ (^ > 2, 0 ândco^ C^-^ then Q^ C^-^. (iv) / / M and OJ are of class C^ or analytic, then so is Q (v) If Q and co are 0-forms, then Q has an additional degree of differentiability in all cases above except the second half of [m).
T."]. Potentials; the decomposition theorem
311
In all cases, if we set a = dQ and ^ = SQ, (7.7.}) da + dp = d[dQ) + d{dQ) =00, da = d^ = 0; C^î {^2) (da, dC) + (doc - CO, dC) = {dp - (D, dC) + {dp, SC) = 0, (7.7.4)
Proof. The results for D follows directly from the discussion above and Section 7-6. The proof of the results for dQ and SQ is like that of Theorem 7.4.1 where it is already done for the interior of M. We choose a boundary point P and an admissible boundary coordinate system of the type described in the preceding paragraph and approximate (if necessary) to co and the gij by smooth functions. For each of the approximating functions Q, we see from formula (7.5-14) that AQ = co and (7.7.5) ndQ = 0 if i^^^2^ and tdQ -=^Q ii Q^^^ since, in the integral over h M, t^ is arbitrary if f ^ ^ ^ and ^ f is arbitrary if C $2^. From Lemma 7-5.2 we see that t d(p = 0 whenever t(p = 0 and 99 and M are differentiable. From Lemma 7.5-2 it follows also t h a t (7.7.6) n(p = 0 '->nd(p = 0. Hence, from this and (7-7.5), we see that both a and /^$2^($2^) if co and D ^ ^ J (^^) at each stage of the approximation, t h a t a and p satisfy (7.7.3) and hence (7.7-4) using Lemma 7.5.3. The approximation may then be carried through as before on each Gr with r <C R. Since a finite number of the smaller boundary neighborhoods cover b M, the results (7.7.4) for all C in ^ J (^^) follow and the differentiability of a and/?now follow from Section 7'6Remark. Except in the case of zero forms i2, the individual derivatives of the individual components of Q do not, in general have the same differentiability properties as do dQ and SQ (the coordinate transformations will not allow it). The following two theorems are useful and important: Theorem 7.7.5. Suppose rj ^^2 ^ HI, iJ+ and H~ are its projections in + and ~ and Q+ and Q- are the plus and minus potentials ofrj H^ and rj H~~, respectively, and a^ = dQ'^, /5 == dQ^. Then (i) A:+ is the plus potential of drj and p- is the minus potential of drj and dr)^22Q "^ ^^^ ^^ S2 0 ~. (ii) Ifr}^%^, then p+ is the plus potential dr]^ &2Q $'^. (iii) Ifrj^ '^2> ^^^^ ^~ ^^ ^^^ minus potential of drj^ S2Q ~. Proof. These results follow from Theorem 7.7-4, equation (7-7.4) and Lemma 7.5-3. Theorem 7.7.6. (i) If rj'^ ^^t ^'^^ V~ ^ '^2> t^^^^ ^^^ unique forms a and P where A : $ ^ J and is 22-orthogonal to + and P^'^2 ^'^^ ^^ S2orthogonal to ~ such that da dr]+ da =^ 0, dp = drj", dp = 0.
312
(ii) If 7]^ HI, there are unique forms y ^ ^ ^ O (S2 0 '^) ^^^ s in $2 n (S2 0 $~) S'i^ch that dy = drj, dy = 0, ds = drj, de = 0. Proof. The uniqueness is evident. To prove (i), leti?+ andi2~ be the respective plus and minus potentials of 7^+ H+ and r]~ H~ and let r = SD'^ and E = dQ~. From Theorem 7.7-5, we see that F is the plus potential of dr]+ and E is the minus potential of dr]~. Then, from Theorem 7.7.4 we conclude that a dP and ^ = SE have the desired properties. To prove (ii), let Q+ and Q~ be the respective plus and minus potentials of rj H+ and TJ R- and let A=dQ+, B = dQ-, y = dA, e = dB and (ii) follows from Theorem 7.7.5(1)Definition 7.7.3. We define the linear sets and 1) as the sets of all forms of the form doc and d^, where oc^'^^ ^^^ P^^2> respectively. We now can prove an analog for the Kodaira decomposition theorem for manifolds with boundary. Theorem 7.7.7. The sets and and the set ^ of all harmonic fields in S2 on M are closed linear manifolds in S2 <^^^ {7JJ) 2 = ee^. 2 Moreover, if 00^ H\, its 2 projections y, e, and H on ^,%, and belong to ^2", ^ ^ , and H\, respectively, and dy = ds = 0. Proof. That K and are closed linear manifolds follows immediately from Theorem JJ-^ and 7.7.2 and t h a t ^ is also follows from the interior regularity theorems of 5.5- Using Lemma 7-5-3, we see that and ^ are orthogonal and that (^ and % are both orthogonal to ^ 0 HI and, in fact, ii H^Hln{22Q^G'^), then / / $ $ ( ^ J and ^^ are both everywhere dense in 2)Now, suppose rj^H\ and let y and s be its projections on g and %, respectively. Using Theorem 7-7-(> we conclude the existence of unique forms a and /5 in $ J O (S2 0 +) and $ ^ 0 (S2 0 ") respectively, such t h a t {7.7.^) da = y, doc = 0, 3^ = 0, d^ = e. Since y and e are the projections of 7 on g and %, we see from {7'7-^) 7 that a and /9 satisfy (da, dC+) + (da, dC+) - (v, dC+) = (da, dC+) + (da, dC+) - (dri, C+) = 0
^^ ^^^
(d^, di-) + (d^, dc-) - (n. dc-) = (dp, dc-) +

+ (dp,dC-)-(drj,C-)=0
for all C* in ^ J and f- in ^ ^ . Thus a is the plus potential of drj and ^ is "" the minus potential of drj. The results follow from Theorems 7-7-3 and 7-7.5-
"J.']. Potentials; the decomposition theorem
313
The following theorem contains further information concerning the decomposition [7J.7), Theorem 7.7.8. Suppose co ^ S2 <^^^ y, , cind H are its projections on (, ^, and ^ , respectively, and suppose Q^ and Q~ are the plus and minus potentials of o) H, respectively. Then (8.7.10) y^doc, e=^dp, oc = dD+, ^ = dQ-, doc = 6^ = 0. If co^ HI, then <x and ^ are the plus and minus potentials of dco and dco, respectively. We have the following differentiability results on the closure
ofM:
(i) / / M is of class { and co^ &p{p ^ 2), then y, e, and H^ Qp and a, ^, Q+, and Q-^H],] if also w^ HI, then y, e, and H^H\ with y, d and H in C^ in case fx = \ vjp and p ^ v. (ii) / / M is of class C^ with ^ > 2 and 0 < /^ < 1 and if co^ C"^"^, then Hy y, and e ^ C^~^ and oc, ^, Q+, and Q- ^ C^~^; if also, co ^ C^~^, then H, y, and s^ C^~^. (iii) / / M and w are C^ or analytic, so are oc, ^, y, e, Q^, and Q~. (iv) In the case of zero forms, H is a constant, and = O.Ifw^Hl with dco = 0 or if 0)^ 22 cif^d w = dt] where 7]^ HI, then y = 0', if o)^H\ and dw = 0 or if 0)^ 22 ^nd m = drj where r}^ H\ then e = 0. Proof. Suppose, first, that o) ^ H\. If we then define oc, ^, y, s, i2+, and Q- by (7.7-10), the results follow from (JJ.S) and (7.7.9). In case co is merely in 2, we use the left sides of (7.7-9) and approximate, using the Theorems of 7.6 and Theorem 7-7-2. The regularity results and the last statement follow from the facts that oc and ^ are the respective potentials of dco and dco, since dH = SH = 0, in case co^ HI. The last results for co merely in S2 follow from Lemma 7.5.?. We may now prove a slightly strengthened form of an inequality due to Friedrichs [3]Theorem 7.7.9. There ^'s a A > 0 such that if co ^H\ j ^ ^, then D{co)-> Proof. For if a) ^ i / | J _ , then (7.7.11) w=y + 8, dy = de = 0, y^^J, ^"^2 > ^X[2\\y\\^ {y,s)=0. + 2\\s\\^] Hence from (7.7-9) and Theorem 7-7-2, we see that D{w) = D{y) + D{8) ^2,-^\\y\\^ + ?i-\\8r > A I C IP, I O ;i = min[A+/2, A-/2]. Kloof.
The following theorem completes the analogy with the case for a compact manifold without boundary. Theorem 7.7.10. If f] ^22 cc^d is 22-orthogonal to , there is a unique form Q in H\ and 22-orthogonal to such that (7-7.12) [dQ, dC) + (SQ, dC) - {v> C), C^Hi
314
Moreover, {7.7 A 3)
dQ = dQ+
and dQ = dQ-.
Q+ and Q~ being the respective plus and minus potentials of rj. The differentiability properties of Q are the same as those in Theorem 7-7AProof. The proof that Q exists in E\ and is unique is just Hke that of Theorem 7.7.3- Obviously ^ 2 0 $ ^ and 82 0 ~ so that its plus and minus potentials exist. Accordingly we have, for example, (7.7.14) (7.7.15) [dQ - dQ+, dC) + [dQ - dQ+, dC) = 0 for all C ^2" dC = dQ-dQ+, (5C-0. But, from Theorem 7-7-^, we m a y find a C^ ^ J such that Using (7.7.14) and (7.7.15) and a similar argument for Q-, we derive (7.7.13). Now, let us consider the decomposition {7.7.7) for Q, D+, and Q~. Using (7.7.13) and the last statement in Theorem 7-7.^ also, we obtain n-1- , -r , dK'r =.dK- ^0, K+ = Ht + E+, K- = H^-^r^ where the F' s^d, the E' s^^, and the H* s^^. From (7-7.16) and the uniqueness of the decomposition, we obtain (7.7.16)
Q = r+E,
D =H +r + E,
Q=Q+K,
H + Hf = 0, r + E = r+ + {E+ + E+) = (r- + r-) + {7.7.i7) r = r+, E = E-.
E-
The differentiability properties of dQ and SQ follow from (7-7.13) and Theorem 7.7-4 and those for Q follow from (7-7-16) and [7.7.^7) and Theorems 7.7.4 and 7.7-9. Remark. We cannot conclude the differentiability of Q directly from (7.7.12) and the theorems of 7.6 since the equations (7-7.12) are not the same as (7-6.1) since the boundary integral corresponding to the first term in (7-5.18) is not necessarily zero in this case. Definition 7.7.4. The form Q in Theorem 7.7-10 is called the potential ofrj. Important Remark. All the differentiability properties a t either interior or boundary points are local; all differentiability results extend immediately to cases where the given hypotheses hold in some coordinate patch, the conclusions then holding in t h a t patch. 7.8. Boundary value problems In this section, we derive briefly the results concerning boundary value problems for harmonic forms and fields which have been obtained b y
other methods by D U F F and SPENCER and CONNOR [1]. The differentiability
results on the interior have been obtained in 7.4; the corresponding
7.8. Boundary value problems
315
results on the boundary depend on the given boundary values as well as on the differentiability of M and are stated below. The following theorem is seen (from their proofs) to be equivalent
to Theorems 3 and 4 of the paper b y D U F F and SPENCER (pp. 150,15I):
Theorem 7.8.1. (a) / / co is any closed form in H\, there is a unique harmonic field H such that (r = degree of co). co = H + d^, tH = tco,^,dp^'^^{tp = td^ = 0), 0 < r < w - l . (b) / / CO is any co-closed form in HI, there is a unique harmonic field H such that cjo = H + d(x, nH = ncjo,oc,d^^} {noc = ndoc = 0), i < r ^ n . In either case, the differentiability results are as follows: (i) / / M is of class C} and co^Hl, p >2, then H is also and co and H^ C^ on M if p ^ V and // = 1 v/p. Ifr = 0,co and H are constants. (ii) / / M is of class CKC^, analytic) and co is of class C^-i(C"", C"), ^ > 2 , 0 < ^ < 1 , then H is of class C^~^ (C^, analytic). Proof. If ^co = 0, then from Theorems 7-7-7 and 7.7-8, we see t h a t the term 6 ^ = 0 in the decomposition which proves (a); (b) is proved similarly. The differentiability results follow from Theorem 7-7-S. The next theorem is a refinement of Theorem 2 of the paper of D U F F
and SPENCER.
Theorem 7.8.2. Ifrj is any form in H\, there is a form co in HI such that tco =^ trj and dco is a harmonic field. If, also, YJ = d%for some y^ in H\, then there is a unique co of the form co = d^ with | in HI which satisfies the conditions above. Proof. Let H- be the projection of drj on - , let oc be the minus potential of drj H~, and let (7.8.1) CO = rj y, ty = 0 = t e, y = da, s = doc. Then y and s are in ^ ^ and from (7-7.3) (^ = ^), we have dy -\- ds = drj H~, dco = dr) dy = H~ -{-de. From (7.8.2) and the last statement in Theorem 7-7-^, we see that
^cw(e)n(ee)==Now, suppose Tj = dx for some x in H^. Then co =^ d(x oc) from (7.8.1). Suppose coi also satisfies all these conditions. Then t(a) coi) = 0, CO coi == dv{v = X oc | i ) , d{co coi) ^ . But then, from the definitions of g and ^ and Theorem 7-7-S, we obtain
CO COi^ '^2> ' ' ^(^ ~ ^1) ^ n , - . d{cO COl) = 0, . ' . CO ~ COl
^ T ) , C C l = (5r^ g , . . O coi$ -, C coi^S2 ~ O O C O

(Theorem 7-7-5 (i)),
. . C C l = 0. O O
316
The d-NEUMANN problem on strongly pseudo-convex manifolds
We now consider boundary value problems for harmonic forms as distinct from harmonic fields. We begin by defining
//io = ^2"-n $2-, 0 = | ) + n r
= n i/|o-
Then HIQ consists of all HI forms co with tco = n(o = 0. Theorem 7.8.3. ^0 consists of the element 0. Proof. Obviously if H ^ ^0, then H ^ HIQ and minimizes D(H) among all forms in HIQ. We may clearly find a Riemannian manifold M' of the same class as M such that M C M'. If we extend H to vanish on M' M, then H satisfies the same conditions on M' and hence satisfies (7.8.3) {dH,dC) +(dH,dO =0 for all f in H^ and hence in ^2^ or ^ ^ and so has the interior differentiability properties on M' stated in 5-5 and 5-6. The theorem follows from the unique continuation theorem of ARONSZAJN et al. Definition 7.8.1. A form K is harmonic on M if and only if Ky dK, and dK^ HI on any domain interior to M with d dK + d SK = 0 there. Theorem 7.8.4. If oj^ HI, there exists a unique harmonic form K in H\ such that t K = too, n K = no). The differentiability results for K are the same as those in Theorem 7.8.1 except that in the case of zero-forms,
K^Clifm^Cl.
Proof. Write K = oj ~\- rj and minimize D{oj + ^) = D{rj) + 2{drj, dw) + 2(drj, dco) + D (co) among all rj in ^20 H (S2 $0). Then D{rj) > A || 7 p so that the minimiz7 ing function exists as usual. Then K is easily seen to satisfy (7.9.3) for all f in HIQ SO that K is harmonic on the interior of M (using the differentiability results of 7.4). Since YJ^HIQ f) (S2 0 0), the equivalent equation (7.8.4) [df] + doj, dC) + (drj + dco, SC) = 0 for all C$ HIQ is of the form (7.6.1) on boundary coordinate patches. The differentiability results follow from 7-^- The uniqueness follows from Theorem 7.8.3. More general theorems were proved in the authors paper [11]. But the statements of the results are too long to be reproduced here. Chapter 8
The 5-Neumann problem on strongly pseudo-convex manifolds

8.1. Introduction In this chapter we present a simplification of the recent solution due to J . J. KoHN ([1], [2]) of the so-called ^-NEUMANN problem introduced
b y GARABEDIAN and SPENCER for complex exterior differential forms on
8.1. Introduction
317
a compact complex-analytic manifold with strongly pseudo-convex boundary. The problem in its present form was investigated by D. C. SPENCER and J. J. KOHN by means of integral equations. The present author [I3] solved this problem for the special cases of 0-forms and z 1-forms (i.e. forms of the types (0,0) and (0,1) in our current notation) on certain ''tubular" manifolds and used those results to prove that any compact real-analytic manifold can be analytically embedded in a Euclidean space of sufficiently high dimension. Unfortunately there is an error in t h a t paper which is corrected in 8.2 by using the results of KOHN presented in this chapter. These results apply to forms of arbitrary type [p, q) and the solution forms are shown to be of class C^ on the closed manifold provided the metric, boundary, and non-homogeneous term ^ C^ there. Recently HORMANDER [2] has extended these results using Z2-niethods and certain weight functions. He was able to demonstrate existence (in the sense treated in 8.4 below) of forms of type {p, q) in cases where the Levi form ((1.2) below) either has at least q + \ negative eigenvalues or at least n q positive eigenvalues. This is a much less restrictive condition on h M than our condition of pseudoconvexity. Finally, KOHN and L. NIRENBERG have obtained a still greater simplification of this theory. I am presenting here essentially their simplification but confined to the 5-NEUMANN problem, whereas they have applied their method to a more general class of problems. In his recent papers, KOHN sketched applications of his results (a) to the study of the 3-cohomology theory, (b) to the study of deformations of complex structures, and (c) to obtain a new proof of the result of NIRENBERG and NEWLANDER which showed that a complex analytic structure could be introduced on an integrable almost-complex manifold. However, part of the interest in this problem to those working in partial differential equations lies in the fact t h a t the problem is not a regular boundary value problem in the sense of AGMON-DOUGLIS-NIRENBERG ([1] and [2]), BROWDER ([1], [2]) LOPATINSKY, etc.). We shall give an example below after we have introduced the notations and sketched the results; we shall also show the connection with the 5-cohomology. We assume that M = M \J h M is a compact complex-analytic manifold having boundary h M oi class C^. We assume that we are given a hermitian metric
(8.1.1) ds^ =g^pdz''dz^ 1 {gôc=gocp, ^,^5=1,...,^)
which is of class C^ on M. We suppose t h a t the function r ^ C^ {M) and equals the negative of the geodesic distance to b M for points within a distance SQ of b M, SQ <, 0. It is clear that there exists a slightly larger such manifold M' such that M C M' and that the metric r can be ex1 Repeated Greek indices are summed from 1 to v.
318
The ^-NEUMANN problem on strongly pseudo-convex manifolds
tended to ^ C^ {M') so that r is the geodesic distance from b M on M' M. The strong pseudo-convexity of the boundary h M implies that there is a constant co > 0 such that at any point Po on h M (where r = 0) we have (8.1.2) r^rzv T^ Tr > c^g^y T^ Tv for all complex vectors (T^, . . ., T^) such that (8.1.3) rz^T^=.o. I f / i s any other real function of class C^ near b M such that V / " 0 and f = Oonb M, a n d / < 0 on M near b M, then the positiveness of the form / / ^ y T^ fy for T such that/^/S T^ = 0 follows. In the above and throughout this chapter we assume that the operators dldz' and dldz^" are defined by
z<^ =z x^" + i y'^,
z"" = x"" i y"".
We let 91 denote the set of all exterior differential forms of class ^^[M) (i.e. C on M) and denote by 9l2^' the set of all those which are of type [p,q), i.e. which can be expressed in any local analytic coordinate system in the form (8.1.5) (8.1.6) ^^-^'"^^^ (p = Zîjdz^ AdzJ;
We abbreviate the notation to when we use this notation / and / will always stand for increasing sequences as in (8.1.5). However, we shall often wish to have the (pi,ji,. ,jq defined for all sequences of indices î . . . jq; in this case, we assume that the (p's are defined so as to be antisymmetric in the y-indices. We shall at times wish to do the same with the / indices and shall sometimes write (pi,0LR where i? = (ri, . . ., r^_i) with /'i < < rq_\ and oc runs from 1 to V independently oi R. We shall wish to consider M (or M) as a real manifold with metric given in [x, y) = {x^, . . ., A.'', yi, . . ., 3;^) coordinates by (1.1) which becomes, on setting dz^" dx^ + i dy"- and dz^ = dx^ i dy^, giocpidx'^dx^ + dy^dy^) + Ig^^^dx^dy^, (8.1.7 , . Then the dual *(p would be defined by first expressing cp in terms of real differentials dx' and dy"^ and then taking the ordinary real dual of the real and imaginary parts. This procedure introduces a factor 2^+^ into the customary inner product (8.1.8) (99,^) = f(pA*ip
8.1. Introduction
319
of two forms of the same type. Along with most workers in this field we omit these factors. The space S^'^ is the completion of the space %V^ using the inner product (8.1.8) and S is just the HILBERT space sum of all the S^^. For two forms of the same type, it is convenient to define the point function (cp, ip} by ((p,y)ydM = q)A'^yf, <(p,^y = \(p\^ where dM is the element of volume on M, The formulas for ((p, ip), dM, and (^(p, ip} are (8.1.9) {(p,yj)=l{(p,tpydM
M
where in any analytic coordinate system
(8.1.10)
dM = r(x, y) dx dy
^ ^^
if (p is given by (8.1.6) and y) is correspondingly defined. Here F is the V X V determinant of the gx^, g^^ is the p X p determinant of the ^^v^^ and g'^^ is the q X q determinant of the gâ^^. If we use the antisymmetry of the (p's and y)'s in all their indices, we may write <^' ^> ^ ^ ^ * ' ' * ' ^^^**^^^"iî g^^^"^ 9h ivh
XWkl-' 'kph' . Iq-
-'-kX
For forms in 51 we define the operator d as follows: If 99 is of type (p, q) with ^ = r, we define ^(p = 0',iiq <Cv and cp is given by (8.1.6), we define (8.1.11) d(p == 2J(pij-,o^dz^ A dzi A dz^. For forms in 51 we define b^? as follows: If 99^ 51^'^ and ^ = 0, we define b(p = 0; otherwise we define b99 as that form of type {p, q 1) such that
(8.1.12) .
(b(p,f) = {(p,dyj)
for every ip in 9][2>.-i with compact support in M. As is seen by integrating by parts (see 3), this leads to a formula of the form (8.1.13) (8.1.14) {b<p)iR = (-1)^+ir^((^/,ai?.,^ + ^f^^ ((pjtp) = (b(p,f) +
bM
(PS,.T),
for suitable functions A^^^, and to the general formula f^co,y)ydS
where ^ 5 is the invariant surface element on 6 M and (8.1.15) (8.1.16) a}=v(p, (joiR= [^Yg'^(pi,o^Rrz^. dS. From (8.1.15), we may also derive the formula {d(p,^p) = (<p,by)) + j(cp.vxpy
bM
We let 5lo denote the subset of 99 in 51 for which v q) = 0 on b M.
320
It follows immediately from (8.1.11) and the antisymmetry of the exterior product that (8.1.17) dd(p = 0, (p^%. From (8.1.14) and (8.1.16) if follows that if ^ and t^^ 51, then {b(p, dy)) = {(f, ddip) J (v(p, dip} dS (8.1.18) ^^ = (b hep, w) + j i'^'^^y ipydS.
bM
By first letting ^p be arbitrary with compact support in M and then letting it be arbitrary we find that (8.1.19) bbq) = 0 on M and vb(p = 0 on 6 M if cp^'Sio. The 5^-NEUMANN problem is to show the existence and regularity of the solutions of the complex Poisson equation (8.1.20) (8.1.21) nq^^bdf + db(p -= 0 on bM. subject to the boundary conditions vq) = vd(p = 0 This boundary value problem is seen to arise formally from the variational problem of minimizing the integral (8.1.22) d{(p,(p) 2Re(co,(p), d[(p,\p) = {d(p,dip) + [bip^btp) among all 99^ 5Io (^^ 9? = 0 on & M). If co and 99^ C'^{M), we see that 99 satisfies (599, dip) + i^cp, bip) {cD, ip) = 0 = {\I](p co,ip) + f {(v dcp, ip) (Sep, Vyj)} dS
bM
for all tp^ô', the second line follows from (8.1.14) and (8.1.16). Since yip = 0^ we see from (8.1.23) (and known formulas-see end of 3) that the condition vd(p = OonbMisdi natural boundary condition. 8.2. Results. Examples. The analytic embedding theorem In order to get a complete picture of the results, we define (8.2.1) D{(p,ip)=d((p,y^)+{cp,yj) {(p^yj^ô) where d{(p,ip) was defined in (8.1.22). We define ^ as the closure of 5Io with respect to the norm corresponding to the inner product D, The following preliminary theorem is proved in 8.4: Theorem. If (p ^'^, then cp ^ H\{Ms) for each s < 0 and if ip ^% and we form d(p, b(p, dtp and by) as in (8.1.11) and (8.1.13) using the strong derivatives then the inner product D((p, yj) is still given by (8.2.1). Moreover ^ ^ HIQ(M). Finally, if cp^ H\[M) andv(p^H\^{M), then 9 ? ^ ^ .
8.2. Results. Examples. The analytic embedding theorem
321
Then we define the operator L as follows: (p^^(L) there exists an A: in S such that d((p, f) = [a,\p), ip^'i^;
<^q)^%
and
if 99^ (L), we define Lcp = oc. We define as the set of all (p in % such that d(p = b(p = 0. We define ^PQ = ^ PI S^^^, ^^^ = H S^^. In 8.4 and 8.6, we prove the following principal theorem: Theorem, (i) L is self adjoint, ^[L) = S , and is closed (ii) If co^ S 0 $ , d ^ unique (p^'^[L) Pl ( S 0 | ) ) ^Lq) = co. (iii) If we define N w = Ofor co^ and N w as the solution in (ii) if <^ S 0 , i^^'^ ^ ^^ completely continuous. (iv) If q > 1, ^'P^ is finite-dimensional. (v) IfM^%, thenNw^%. (vi) Ifq'>i, ^^C^rParts (i)(iv) are proved, for forms of type (p,q) with ^ > 1, in 8.4. Only parts of the results are proved there for forms of type {p, 0); these are treated completely in ^S.6. The smoothness results in (v) and (vi) are proved in ^S.6. The principal tools used are (a) the important formula (8.3.15) for integration by parts, and (b) the "s-norms" introduced in 8.5 and 8.6. Before proceeding, we introduce some additional notations: The manifold Mg = M(s) for s < 0 consists of all points P on M for which r (P) < s. An analytic coordinate patch with domain G and range 9^ is said to be tangential at some point PQ of b M ^ SL part g of b G contains the origin and corresponds, under the mapping from G to 91, to 9^ H & M, the origin corresponding to PQ, and at the origin gocp{0) = doc^ and the exterior normal to M at PQ corresponds to the positive y^ axis (i.e. ryv{0,0) = 1). In case r = (r^, . . ., r^) or ( n , . . ., tr) is a set of indices, Ty denotes the set (r^, . . ., T^~^, T^+^, . . ., T^). If oc = {oci, . . ., av) is a sequence of non-negative integers, then D^" means D^] . . . D^l. If 99 is a vector function V^^ denotes its gradient. Next, we give an example to illustrate the fact that the 5-Neumann problem is not regular, except in the case where q = v when it reduces to the DiRicHLET problem since dcp ^ 0 and v q) = 0 on b M '^ (p ^=^ 0 on b M in that case. In the case ^ = 0, the problem is obviously not regular since the null space is just the space of holomorphic forms 2^0. To show that the problem is not regular for 1 < ^ < r 1, we take, as an ^ example, v = 2, M the unit ball in i^4, the metric Euclidean, and set cp = (pi dz^-\-(p2 dz'^, 0 = ^ ~ / ' ^ ' A (z),
6
(pi=z^0y z^z^,
(p2 =
z^0,
r^ = z^z^ +
where A (z) ^ H\[M)
and is holomorphic on M(0) but is not in

21
Hl{M).
322
Then r vcp = z^cpi + L^(p2 ^ 0, A(pi=: ZÂ0 0z^,
d(p = CO dz^ A dz'^ + 0zl
(O = 992-1 - 991-2 = Z^ 0-1 + Z^ ^-2 + 20 = (\ - T^) A (z) ^ H\^{M) A(p2 = ZÂ0
A0=--^(zÂzi
+ z^ Az2 +
2A)^L2(M).
I t follows easily t h a t ^^^{L) but 99 does n o t ^ H^{M) as it would if the problem were regular. We now prove a theorem indicating the connection with the dcohomology theory: Theorem. If (p ^ ^^^ with q "> i and if dcp = 0, there is a harmonic field 990 ( ^^^) such that (p cpo =~d 0 for some 0^ S^v,q-i, Proof. First of all, suppose / ^ C^ on Ri, and we define 01 = f{r) {v 99), /(O) - 0, /MO) = 4, f{r) = 0 for y < So < 0 for some such SQ. A computation something like that in (8.4.11) and (8.4.12) below shows that v{(p d0i) =0 on & M . Hence we may assume that 99^3lo^. Then, let 0 = N b(p. Then it follows from Theorem 8.6.3 that 0 , d0, and 10 all ^ 5Io, so that db0 is orthogonal to b{(p d0) (see (8.I.14)(8.I.I6)) and also equal to it and hence zero. Thus (since ^99 = 0) b{(p d0) = d{(p d0) = O or (p d0^ ^^^. The following analog of the KODAIRA decomposition theorem is of some interest: Theorem. S = '^' K where ^ has its usual significance, ' is the totality of forms of the form d(pfor some 99^ , and K is that of forms of the form b\p for ip in%. If the given form in S $ 31, then its projections on , , and g also ^ 31. Proof. I t is clear t h a t if A ^ and 99 and W ^^, then the forms h, d<p, and b!F (see (8.1.14) (8.1.16) and approximation) are mutually orthogonal. Since ^'P^ just consists of the holomorphic forms of degree p and the 2^^ with ' > 0 are finite-dimensional, it follows t h a t is closed. If c o ^ S , let 0 = No), (p = b0, and y) = d0. From Theorem 8.6.3, it follows that 99 and y)^^ 0 ( S ) and that (8.2.2) d(p + bip = 0) and bq) = dip = 0.
It is clear from the first statement above t h a t the decomposition is unique. That and 5)' are closed follows from our principal results and the particular choice of 99 and ^ in (8.2.2).
323
In case co ^ ^ , we first subtract off t h e form coi defined (different notation unfortunately) in the proof of Theorem 8.4-4, so that coo = CO coi^^Q. If we then define 0Q, (po, and ipo as above in terms of coo, we see that (8.2.2) holds for COQ and t h a t 9^0 and ipo^ 5toWe can now present a simplification and correction to the proof given in the paper, MORREY [I3], mentioned above of the possibility of embedding analytically a real-analytic abstract manifold in Euclidean space. The error in that paper was in the proof of Theorem C of t h a t paper which was given in 811. The embedding theorem was proved by BocHNER for compact manifolds in 1937 ([!]) assuming the existence of an analytic metric; this result was extended b y MALGRANGE in 1957 to the case of non-compact manifolds. The result of MORREY [13] was generalized to manifolds with a countable topology b y GRAUERT using methods of the theory of functions of several complex variables and some results of REMMERT which had not been published at that time. We now outline our method of proof. First of all, on account of BOCHNER'S result, it is sufficient to show the following: Theorem A. With each point PQ of the given real analytic compact manifold MQ there are associated v functions Wy which are analytic over the whole of Mo and have linearly independent gradients at PQ. For the gradients will remain linearly independent in some neighborhood of PQ and thus MQ can be covered b y neighborhoods 9^^^, q = 1, . . . , Q, where the functions Wqy, y 1, . . ., r are analytic over MQ and have linearly independent gradients over 9^^, q = \, . . .^ Q. The mapping Wqy = Wqy(P) maps Mo analytically into Euclidean space of Q^ dimensions; the mapping m a y not be 1 1 in the large b u t is locally 1 1 and non-singular and the Euclidean metric induces an analytic metric on MQ. To prove Theorem A, we first embed MQ in an open complex extension
M (see MORREY ([13], 2, SHUTRICK, or W H I T N E Y - B R U H A T where this
embedding is discussed for manifolds with countable topology). Let PQ be any point on MQ, let TQ be a complex-analytic coordinate patch with domain Go containing the origin and range ^0 containing PQ, in which PQ and the origin correspond and the part of GQ in R^ (i.e. for which y = 0) corresponds to Sto H MQ, and choose a Hermitian metric (8.1.1) which is of class C^ (M), which is real on Mo (i.e. goc^ is real on Mo), and for which we have (8.2.3) go:^ (x, y) = docp (x, y) on Go,
with respect to the coordinate system To. For points P on M near Mo, we define r'{P) to be the geodesic distance from P to Mo. I t is easily shown (see MORREY [I3], 3) t h a t the function K(P) = [y'(P)]2/2 is of class C^ in a neighborhood of Mo
21*
324
The d-NEUMAN]sf problem on strongly pseudo-convex manifolds
including all points where r ' (IP) < RQ. We define MR as the complex analytic manifold r' (P) < R. I t is easy to compute that (âjS (x, 0) is real) (8.2.4) 4X,a^^(:^,0) T-T^ = K^-yP\x,0) T-T^ = g,^[x,Q) T-T^ for any complex-analytic patch which carries the points (x, 0) into MQ. Thus if 0 < i? < 2^1, b MR is regular and of class C^ and MR is strongly pseudo-convex; the function r (P) used in this paper reduces to r' (P) R on MR and the pseudo-convexity follows from (8.2.4). In fact, since r{P) =r' (P) R and K{P) = \r'{P)'\'^j2 on MR, we see by following through the proof of Theorem 5.8 (of the cited paper) as far as equation (5.18), that we obtain dR[(p) ^ IR[(P) + j Z^""' ^^"^ g""^
hMiR)
g''r,^',yCpio.SCpK6TdS[P)
(P^^(MR)
> -Ci((p^rp)R+C2R-^l\(p\^dS,
bM{R)
(cf. 8.3.15 below). Thus we obtain l\cp\^dS^CR[dR{(p) + ((p,(p)R], 0<R<Ri<Ro
where C is independent of R. We now sketch the proof, given in 7 of the cited paper (MORREY [13]), of the important inequality (8.2.6) [(p,cp)R<.CR'^dR{cp), 0<R^R2^Ri, (P^^^^HMR), q>\. Incidentally, this shows that ^^^[MR) consists only of the zero element if q > 1 and R is small enough. We conclude first that there is an 7^3, 0 < Rz < P i , such that M (P3) can be covered by a finite number of neighborhoods 3^f each of which C 9^^, the range of a complex analytic coordinate patch n of the type of To (i.e. real on Mo) with domain Gt, and is the range of <3:'-quasi-geodesic" non-analytic (but C^) coordinate system rf with domain of the form Got X P ( 0 , P3), where G^t is a domain of class C ^ in real r-space R^ and Gof C C^ Pi R^. If P $ 9^f, its quasi-geodesic coordinates (f, rj) with respect to rf are determined as follows: There is a unique geodesic through P which is orthogonal to Mo at some point of MQ. Let its equations be
^ ^ ^ (^) ^ ya ^ y a (;,) ^ 0 < r < r' ( P ) ,
in the {x, y) coordinates of T^ We define t{P) = x^{0), C'(P) =- y (0) 0 dyjdr). Since the metric is real along Mo, we see that ds'^ = gio:{x,0) {dx^dx^ + dy'^dy^), along Mo. = gi ^ (x, 0) (di^ d^^ + dC"" dC^),
325
By using the Lagrange method of reducing a quadratic form to a sum of squares, we introduce the rf by where the d"^^ C and the matrix is non-singular so that (8.2.7) (8.2.8) ds^ = îa^(f, 0) ^ | ^ 1 ^ + (5^ drj^" drj^ K(P) = \f,(P)\^l2, r'{P} = \f,{P)\. along MQ. From the construction, it follows that It is shown in MORREY [I3] 3, that the coordinates are C . With each t and each R,0 < R ^ R^, we define an analytic manifold MtR as f_ollows^ Let %R = r* [Got X B(0,R)l and let GtR = r^^ {%R) ; clearly Got = GtR D R^- Choose positive numbers at and At such that if we define Ft=Ff\ Ff:\x''\<At, dc=\,..,,v then GtR,CFlo^xB{0,atl2)^ GotCFl'K Extend the metric gtup to be of class C^ for all (:^, y), to be periodic of period 2 At in each x^" to be real ii y =^ 0, and so that gt<xp(x, y) = d<xp for all X on and near dFt and all y with | y | > 'iatjA- Then, if % is small enough, the quasi-geodesic (|, rj) coordinates can be extended to all (f, Yi) with \rj\ ^ at to be periodic of period At in each f^. We then let MtR be the set of all {x, y) corresponding to the (f, rj) with \rj\-^R, any two points (xi, y) and {x2, y) where each x^ x^ = 2At n'', n"' an integer, being identified. Now, suppose (f^'^'P^ {MR), 0 < i^ < R^. Let {CJ, s = 1, . . ., S be a partition of unity such that each fs G^^H^R^ and has support in MR^ n 91* for some t (Cs need not vanish on h MR^) and let (ps = Cs (fThen, b y approximating 99 by smooth forms, as we may on account of our principal results, we see that each (PS^^^^{MR). But now, we may associate each (ps with a form (pst on MtR by defining the components of (fst on GtR to agree with those of (ps there and to vanish elsewhere on MtR. Then, clearly,
dR{(ps) = dtR[(pst), {(fs, (PS)R = ((Pst> (Pst)tR,
dR[(ps) < C[dR{cp) + (g?,(p)R] since the Cs do not depend on R. Accordingly^ it is sufficient to prove (8.2.6) for iovms^^PQ(MtR), since, if this is done, we would have
(?>, <p)r < 1 (^.. <PsYd' <C-S-R-[dR

s= l
(^) + (9, g,U]W
from which the result follows easily if 0 < i^ < i^2 < ^ 3 So we consider some MtR and drop the i. We first prove (8.2.6) for forms ^0 ^20 (MR) if 0 < i^ < 2^3 < ^ i -
326
We shall hereafter denote the components (pu simply by cpK Theii (gô, (fo) = j (8.2.9) J ((po> (pQ)dS dr
^ '"^/^
^Cfr^-^drfdSf2\9U^>lO)\^d2;(0).
0 F 2 ^'
where ^ = bB{0,\) and 6 denotes coordinates on bB(0,'[), {^,rj) being the quasi-geodesic coordinates, and {r, 6) being polar coordinates in the ?y-space. Since 990 (R, f, 0) = 0, we have
R
(8.2.10)
' B I <pi(r. 1,6) |2 ^ (i? - r) / 1 ,pl{s, 19) |2 ds

r
Substituting (8.2.10) into (8.2.9) and using the conditions s > r, we obtain (^0, (foJR < CR^ ((990, qô))R = CR^{{(po, (PO))R2 < CR^ dEs{(po) = CR^dE{(po) (for the definition of the strong norm ((99, 99))!'2 jn HI, see 8.4) if we define 990 = 0 for jR < y < i?3 and use Theorem 8.4.1 for M(Rs). This is (8.2.6) for 990. From this, it follows t h a t if 99$'5), H a unique 990^ HIQ{MR) which minimizes the integral dR(q)o 99) among all such 990 if 0 < i^ < Rs. I t follows that 990 99 is a harmonic form H and that (8.2.11) dR((p) = dR((po) + dR{H) To prove (8.2.6) for a harmonic H, we shall prove for any harmonic
H, whether in %^Q(MR) or not, t h a t
(8.2.12)
(iJ, H)R<C'RfiH,HydS
bM(R)
(for R small enough)
from which (8.2.6) follows, using (8.2.5). We m a y assume H^ ^^{MR). To prove (8.2.12), we see as in 5, that the real and imaginary parts W of the components satisfy a system of differential equations of the form
in terms of the quasi-geodesic coordinates, where (8.2.14) a'^^dO) =gf(i,0), b^^{i,0) = 0, c'^(|,0) =:d^^. Let us now take spherical coordinates in the ?^-space as above. Then the equations (8.2.13) are seen to be equivalent to (8.2.15) Htr+ {v - i)r-^m + r-ÂêH^ + 2cyH^y + r - i Cy'Hiy,6 + 2rB''Hlôc + ZB'^yH^^^y +A<'^H^^^6 + 2D{^H^â + -1- EiH^; + Ir-ÊiyH^y + Fim = 0 +
327
where Aê denotes the Beltrami operator on the unit sphere and all the coefficients ^ C in {r, f, 6). We define the positive form Q by Q = ^ [(^tf + ^-2 I V, H^ |2 + 2 C^ Hr H,y + r-^ Cr^ H{y H^s + (^-^^^) ' " ' + 2rB^H\ocHl + 2B-yH\o:Hiy+A'^^H\ocHy-\ (8.2.17) F{^)=\i: [H^s, f, e)fdZ{0).
Then we see easily that F'{s)==2l

Fxi:
^ W (s, I 0) Hi(s, f, 6) d^d 2: {0)

?'
(8.2.18)
r'{s) = 2 /
FX2^ ? " Fxi:
?
ZH^HU+[H^:)^]dHZ 0) Hi{0,1 6) dSdZ = 0.
F'(0) = 2 f ZHHi.
Using (8.2.15) to eliminate the Hl^, integrating by parts with respect to the f* and Qy, and using the fact that we find that
E
juA^Qud'^ = f \ ê'^f^'E
E
F"{s) = 2nQ (8.2.19)
+ 2:[-{v-\)r-^WHi+2ClyWHi
+ r-1 C^l W Hiv + 2rB%o. W Hi + 2 B'^y W H\u + + ^pf H^ H\f\ - 2D{^ W H\. - etc.l d^dZ{B),
Using the positivity of Q and the simple device \2ah\ <, e a^ + e-'^ h'^, we conclude from (8.2.19) that F" (s) > - (r - 1) s-i r (s) -X^F (s) where X depends only on bounds for the coefficients and their derivatives (i.e. on the metric). Thus if R is small enough ( < A~i î) (8.2.20) F{s)<2F(R), 0<s<i^. thus (H, H)R < C lf^-^F(r)
0
dr < CR^F{R) < CRJ{H,

hM(R)
H)dS
as desired. The inequality (8.2.6) states that the constant C3 in Theorem 8.4-5 can be replaced by C R^ for the manifolds MR. Thus, from Theorem 8.6.4 we conclude the following theorem: Theorem B'. If w ^"^ [LP^^) on MR and R < R2. then
(8.2.21) \\NR LRW\\<. C5 -R2 \\LRWI
where 1 1 denotes the norm in S^^'^^. 1 1
328
We now show how to prove Theorem A using Theorem B\ We first construct, for each R, functions W2Ry, y = \, . . ., which $^(L^'^), are of class ^^[MR), which are analytic at least in the ball B[PQ, R) with '^2Ryz^{0) = dy^ (8,2.22) (with respect to TQ) (on MR) I LR W2Ry I < Zi JR^
h = [r/2], 0 < i^ < 7^4 < ^ 2 , Zi independent of R. To do this we first define wiy = zy in B {PQ, R2) with respect to To, extend wiy to be of class C^ on MQ U B(PQ, R2). We then extend wiy into some MR2 using Whitney's extension theorem, assigning the various derivatives of order < ^ + 1 with respect to the y' in each complexanalytic patch, real on MQ, in such a way that the Cauchy-Riemann equations and all their derivative of order < h hold along MQ. Thus the second condition in (8.2.22) is satisfied. The functions W2Ry are constructed to ^ "S) (L^R^) by a method like that in the proof of Theorem 8.4.4 which retains the second condition in (8.2.22). Then if we set wsRy = NR LRW2Ry, wc scc from Theorem B' that
II t^3i?y II < Ci^2 11^^ 2^2i?y II <Z2i^^+2+W2.

But also w^Ry = W2Ry w^Ry is analytic on MR, SO wsi^y is analytic on B{PQ, R). From the inequalities of 2.2, it follows that IV wsRy(0) l^ZsRK k= \ + (r/2) + [r/2] - r = % r 1. Hence if R is small enough, the gradients of the wsRy are so small that those of the W4Ry (0) are linearly independent at 0. 8.3. Some important formulas In case q)^^^^, d(p was defined to be 0 if ^ = r and was defined in (8.1.11) otherwise. Starting from that definition, we obtain
= ( ^ ^ ^ 2 ' ^ g 9 ^ / , ; i . . . V ^ 4 ^ ^ ^ A dz- A dzh A . . . A dP'a + + Z ( ^)^ dzi A dzh A . . . A dzh A dz"" A dzh+i A . . . A dz^X
d=i J
= {-\)^
(8.3.1)
I,M y = l
Z^S^-'^y-^
9i,M'-z^ydzi ^dzM{\M\^
q+
\,q<v).
The form (8.3.1) has the advantage that the coefficients are antisymmetric in all the indices mi, . . ., mq_î. We shall be concerned with boundary integrals over the manifolds b Ms. We note that the function r defined in 8.7 is a real-valued function of class C^ {M) such that | V r | (as measured on ikf) = i near
8.3. Some important formulas
329
bM,r=:0 on bM, and ^ < 0 on ilf near b M. If we let d M (P) denote the volume element and dS{P) denote the surface element along some b Ms at P, then dM (P) = r\x, y) dx dy, dM{P) =dS (P) dr (P) '^' dS(P) = \Vr(x,y)\-^r(x,y)dS{x,y), (x, y) corresponding to P in some coordinate patch; here | V r {x, y) \ denotes the gradient of ;' with respect to the coordinates (x, y) and dS (x, y) denotes the surface element of the surface of integration in the (x, y)-space. Let G be the domain of an analytic coordinate patch having range 3J such that "St f) b M is not empty and let g be the part of b G which corresponds to ^ C) b M. Then, if 9 or ^ vanishes on and near 9 b G g, we note that / (p xpzP r{Xy y)dxdy = i (pxp'fzP' \V r[x, y) |-i r{x, y) dS {x, y)
g G
(8.3.3)
(8.3.4)
-fy^[r-i^^rcpyrdxdy
G
and the integral over g can be expressed in the form j (piprzPdS[P).
Next, suppose that (p^ %v<i, yj^ ^l^^-^-i, q >i, (p is given by (8.1.6) and y) is given by a similar formula. Then, using (8.3.1), (8.1.10), and the antisymmetry in the indices, we obtain (7), dyj}=-\2J
(8.3.5) X I
I,K ^' I,K U),{m)
2^g^^ghî
. . . gJa'â X
i-"^)^-^''-^
B,L a,j8
(piJi.,.ky'K,m'yZ^v
If we cover M with coordinate patches and use a partition of unity {Cs}, each with support in one patch, let 99^ = f 99, integrate by parts using (8.3.3), (8.3.4), and (8.3.5), and add up, we obtain (8.3.6) {(pjy^)=l Z Ig'''g''''ojiRyjKLdS{P) + (b(p,f)
where co = 1 9 is given by (8.1.15) and 699 is given by (8.1.13) where the ^ 9 îRjS ^re determined so that
{b(p,yj) ^f^gîgRL{b(p)iRfKLdM{P)
(837)
for allt/;^5(2>.^-i.
^
M
330
Now, let (p and ^ ^ 9l2>Q'; we wish to develop an important formula for (8.3.8) d[(p,xp) ={h(p.hxp)+{^d(p,dxp). Clearly, we may write cp and \p each as sums of forms each having compact support and such that if the supports of (p^'^) and '^(^) intersect, their union lies in one coordinate patch. So we assume 99 and ip have supports in one patch. Suppose cp and tp are given b y (8.1.6), q <C v, and Q = d(p, a dip; then Q and a are given b y formulas like (8.3.1), so
{dq.Jip)=j^~^ I I
(0.3.9)
5;(-l)r+^^^^ X
+ h
ri...r+i = l
X gmm . . . gma+ir^+i cp^^^^-my ipK^/sz'd dM = h
where 7i is the part of the sum where d = y und I2 is the remainder. We obtain^ (8.3.10) Ii = f2;g'''g^^g'^(pij-,ocWKLz^dM. Using t h e antisymmetry of the indices we see t h a t where m^^ denotes the m sequence with my and ms both omitted, etc. Thus, we obtain
M
Using (8.3.7), we thus obtain (interchanging (oc, y) and (/^, d) in (8.3.11)) d{fp,w)-h+flg'''g^''g''^g''\.-Î,.S'zyfK,6Tz0 (8.3.12) + ((piôcSzP + ZAYl^
+ ZÎTV
(pu,ocv) {y^K,6T~zy +
= / i + /g.
Ww,dx)'] rdxdy
Now we define the forms '% and 'co ^ ^2?, ^-1 b y (8.3.13) 'X = I'Xis^^^ A dzs = 'co = {i)Pvy)=-2'cOs:Tdz^
'%IS = g^'^^zP (pieces, 'CO^T =
(~i)Pv^, A dz^
g^^rzyipKdT'
Next, we note t h a t '^(q>i,<xSzyWK,dTzy 9i,o:s'iyWK,dTz^) =j^[(pi,^sWK,dTzy

9l,o^S~zyWK,6T) + -^r:^[fpi,<xSz^WK,6T (pi,ocS fK,dTzP)
(8.3.14)
^^ g^^r^p (pj,s-y = 'xiszy - {g^^^^z^y + gt^zP) cpi,.s gy^r-yipjK^dTz^ = 'coKTzS {g^^r^P-.Y + g^pr-v)ipK,6T'
1 Strictly speaking, the integrand in Ii is not invariant under changes of coordinates; however, the final result in (8.3.15) is invariant.
8.3- Some important formulas
331
We then integrate Is by parts until there are no terms Uke (pu,<xvz^ and Ww^dx'zV ^^ the remaining integral over M. The result is of the form
M
+ 'o)KT[e[h <p)is + BY^^'(pu,ccv] g^^ 'XIS'Z^WK,6T (8.3.15) (^=(-1)^+1).
In the case q = v, dq) = dyj 0; a special computation leads to the result (8.3.15) in this case. In the case q ==:: 0, b(p = btp = 0 = 'x = '^ and (8.3.15) holds without a boundary integral. In the case ^ = r it follows t h a t vq) = 0 on bM^ the components of 99 = 0 on 6 M . From the result (8.3.6) it follows t h a t (p^'$in%<^vq) = 0 on bM. But now if (p and ^ ^ 91 0 , the boundary integral in (8.3.15), is seen to reduce to that of the last term since 'x and 'co vanish on b M and hence (8.3.16) 'Xiszy = îsr-^Y and 'COKTZ^ =p'KTrzP on bM
for suitable functions A/^ and JUKT- From our hypothesis of pseudoconvexity of b M, it follows that ^/gyT^r^" > Cg^yT^ry, C > 0, whenever rgpr^ = 0.
Consequently, ii ip = (p and 99^ 91 Pi , then (8.3.17) fgKIgSTr^p-^gcc^gyS^j^^^^j^.^dS ^^

bM
^ C f g^^ g^"^ g-' (pi,.s^KST dS =
Cf\(p\^ds.
bM
Now, for forms cp having support in the range 5)1 of some coordinate patch, it is sometimes desirable to introduce a new basis f^, . . ., C*', C^, . . , c,^ for the 1-forms, given by (8.3.18) ^ ' ^^^ctd^, C'^^c^dzy, dzy = dyc' dzy^diC^"
and to introduce the corresponding differential operators (8.S.I9)

U y = d%Uz0^y UzOi C'Û^y,
Uy = d'^ûioi,
uôc =
clu-
where of course the matrices {cfj and [d^) are inverses of one another as are the matrices (c*) and (5j;). The exterior multiplication allows us to express
332
any form in terms of the C^ and f^: w = L- 2 ' ^ n . . . i p h . . . k ^ ^ * i A . . . A dzh A dzh A . dzU
^ p\q\ ii...ip
(8.3.20)
'""''
= ^Vq^Î^n...i,h.^.kdl[In case the bases are introduced so that (8.3.21) g-nyc$ = Ay^ and
diid^ .. . i^^fî A . . . A C^^
?i A ... A ?^ .
y,d=h...,v
g,pd^^dl = Ay3, if
I,J
we see that ((p, tp) = 2 "puWiJ^

I,J
We call a basis in which the cl satisfy (8.2.21) an orthogonal basis. Such bases were used by KOHN ([1], [2]). In terms of such a basis, we see that
y=l
QIR = ( - 1 ) ^ " ^ ^ 2 ' 9^'/,i2,a + 2 ; ^ F / > C ^ , F

ex.
where the A's and B's are suitable C ^ functions. Such bases are more useful in boundary neighborhoods 9^ (in which "^ n b M is not empty). In case cr is a tangential analytic coordinate patch with domain G\J g and range 9^, we may, by taking a smaller 3^ if necessary, choose an orthogonal basis C such that (8.3.24) 2ig^Prz^ = d^. It is also possible, choosing ^ smaller if need be, to introduce nonanalytic boundary coordinates {t, r), t = (t^,.. ., ^2'-i) of class C^ which range over some GR U OR and which are such that the metric takes the form (8.3.25) ds^ =î:\At,r) dt'dtf^ + dr^, a,^(0,0) = d,,. Now, since the basis is complex-orthogonal, each ^,y == (^Dû i" D'û)/! and u^- = ['Dy u + i" Dy ^^)/2, where 'Dy and "Dy are real operators which are, in fact, directional derivatives along real unit vectors e^ and e^ in which all 2v are mutually orthogonal and e^ | = iV r- jThus, in terms of the {t, r) coordinates u,y = 2êyUt^,
A
y <v
(Cy complex);
(8.3.26)
u,v = êlutx
(^72) Ur [el real);
8.4. The HiLBERT space results
333
Finally, by using the relations implied by (8.3-20) between the components with respect to the {dz^, dz^") basis and the orthogonal (f*, f) basis, we find easily that (8.3.27) VCp = 2^C07ijC^ AC^, COIR = (~nH-^l2) (Pl,vR provided the (C^, f'^) basis satisfies (8.3.24) on the boundary neighborhood 9^. To see that (8.1.23) for all ip^ô implies that v d(p = 0 on b M in case (p ^ 5Io, we note that if ip has support in a boundary coordinate patch, 0 vdcp, and (C, f) is an orthogonal basis satisfying (8.3.24), and we write then
CO = 2^COIJC^ AC-^, ip = ZfuC^ AC-^*
J <oj, ip)ds = j ^ coijipij

bM b3I ^J
ds.
If (8.1.23) holds, it follows that all the cou = 0 unless y^ = v. But, from (8.3.27), we conclude t h a t
coij =
(-\)^{-il2){^(p)i,,j
and so cou automatically = 0 if y^ = v. 8.4. The Hilbert space results In this section we prove the first four principal results stated in 8.2 for forms of type {p, q) with ^ > 1. Only a few results are proved for {p, 0) forms here, they are treated fully in 8.6. We begin with preliminary results including those stated there. Let M be covered by the ranges ^s of analytic coordinate patches T^ with domains Gg, where each Xs can be extended to a domain FsZ^Gg. For each s, suppose f = 1 on Gg, Cs C"^ (Fs), Cs has support on Fg, and satisfies 0 < ^^ < 1 on Fg. For forms of a definite type (p, q), we define
(8.4.1)
((cp, ip)) = {cp, f) + Z fCs I
' ^ ^
''''
Ija
{9ijx-ij^-
+ (pijy-fiJv-)
^^ dy>
s jV
where (p^/j and ip^/} are the components of 99 and ip with respect to tg. For composite forms, we define these inner products by summing over (p,q). Evidently the s p a c e / f | is the closure of 5t under the strong norm defined by the inner product ((99, y))). Theorem 8.4.1. Hl^ C^.Ifcp^ Hl^, then (8.4.2) (8.4.3) (((^,99))<C((99,(^))-. [{(p,<p))i<CD(<p.cp), ({v(p,v(p))<C(v<p,v(p)i<CD{q>,(p). 7 / 9 9 ^ , then v<p^ H\Q and
334
Proof. The first statment is obvious. If (8.4.2) were not true, there would exist a sequence {q)n} ^(pn^ ^7i^) with
.g . .. {((Pn, (Pn)) = 1 a n d (((pn, (pn)h - > ^
(pn-7(p (8.4.5)
in
HIQ.
Then 99^ -> 99 in S so 99 = 0. Also, for each s ICsI:<p^\j^^dxdy->0. then
But, if we write (p = (pi + i(p2 and assume 99^ Cîr), r (8.4.6)

f ^{(Pz^'l^d^dy
= f C[((pixo^ (p2y^Y + (9^2 a* + Cpiycc)'^] dx dy ;^
= ji^{\(px^\^ + \^yct\^)dxdy + f [Cx<^{(pi (P2y^ (p2 (piy^^) + r r + Cy<x{<Pixo^(p2 (p2x'(pi)]dxdy. Applying the result (8.4.6) to (8.4-5) and using the strong convergence of <pn to 0 in S, we find t h a t ((99^, 99^)) - > 0 which contradicts (8.4.4). If 99^ 9lg^ with ^ > 1, V (p ^ HIQ since it ^ 51 and vanishes on b M. The result (8.4-3) for such 99 follows from (8.3.15) with y) = cp and co' = %' = 0 on b M and from (8.3.17), from which we obtain d{(p,(p) >I((p,(p) + ciJ\(p\^dS, I{cp,(p)>{{(p,(p))-^C[(p,(p), ci>0.
bM
I (99, 99) being the integral over M in (8.3.15)- Moreover if co == i^ 99, then ((co,ft)))< C((co, a>))- < C((99, 9?))<CD((p,(p), The results for 99 in S) follow by an easy limit process. If ^ = 0, then D{(p, (p) and ((9?, 99))- yield equivalent norms. Before proving the next theorem, we conclude from (8.1.13) a n d (8.3.1) and the definition of exterior multiplication that (8 4 7) ^^^ (p) =r]dq) + drj A(p, coiR = {-'i)^g''^VzP(Pi,.Rs (8.4.8) (D==f(r)v(p. b (rj q)) = rjb(p (o rj^W^. then
From (8.4.7) and (8.1.15) we see that if rj =f{r),
Theorem 8.4.2. If cp ^%, then 99 ^ H\{Ms) for each s < 0. / / each <pn^ 5(o <^^<^ (pn -^(p, d(pn - ^ ^ , (^nd b(pn -^X> ^^^^ ip = d(p and ;^ = 699 where d(p and b(p are computed from (8.3.1) and (8.1.13), respectively, the derivatives being the strong derivatives. Finally {{(py (p))s < Ci s-^D((p, cp), So < s < 0 . Here (( )) denotes the strong inner product on M(s). Proof. For each s, choose rj ^ C^ (M) with ?^ = 1 on Ms and rj =f(r) for s < r < 0. Then if 99,^ is a sequence as above, rj(pn^ HIQ for each n
8.4. The HiLBERT space results
335
and, from Theorem 8.4.1, we see that

{(rj (fn, n (pn)) < C{{ri (fny fj (pn)y^ < CD(r] (fn, f] (fn)
Thus we see that rjcpn-^rjip in H\Q SO that the results follow since we may assume that | / ' W | < 2 | s |~i. Theorem 8.4.3. (a) L is self-adjoint. (b) is closed. (c) (p^&Q^(L) < ^ ^ ^ ^ ( L ) andL{(p) = 0<^(p^^. Proof, (a) Since D(q), cp) > (99, cp), it follows that L + / is 1 1 and (L + I)~^ is defined everywhere and has bound < 1. Suppose co and 7}î and let 99 = (L + / ) - i co and ^ = (L + /)~i rj. Then cp and \p^^ and we see from the definitions of L and D that ((L + / ) - i a>, .y) == [cp, (L + / ) r ) = ^ ( ^ , V ) = {{L + I) % y>) ^ = {co,{L + I)-^rj). Thus (L + /)~i is selfradjoint so that L + I and L are also. (b) Suppose cpn^^ for each n and 99^ -^99 in . Since 399^ ^ 0 and 599^ - ^ 0, it follows t h a t 99^ S) and ^99 = 599 = 0. (c) The first statement follows from HILBERT space theory. The second statement follows since 99 ^ ^ (L) with L 99 = 0 <^ ^(99, \p) = 0 for each (8.4.9) j\cp\^dS
hM{s)
\p^%. so<0.
Theorem 8.4.4. Suppose 99 $ ^ ^ ^ and ^ > 1. Then <.C2D{(p,(p), so<s<0,
Proof. It is sufficient to prove this for 99 ^ % . Let co = ^99. From Theorem 8.4.I, we conclude that oy^H\Q and (8.4.10) Next, let us define (8.4.11) Then (8.4.12)
Q - l
((a),co))
yj = (p-4Q, Qjj^^^
<.CD{(p,(p).
--\)y OJIS 4-^"^^^2)8 ^â CO IS ^r-iyO)i,/^. +
r=i
== 0
(near b M]
( - l ) ^ r ^ ^ 2 ^ < ^ / , a s ; , = ^ ^ ^ ' ^ ^ 2 / 5 ^ 2 0 99/,,^,^ = 0
on account of the antisymmetry of 99 in the indices s and oc. Thus i; ^ = 0 near b M. Hence for \s\ sufficiently small, (8.3.I5) and (8.3.17) with (p = ip yields /o / .^ ^s{f,w) =Is['(p>w) + / 2g^^g^^r,nyg^"^g''^Wi,ocsWK,dTdS (8.4.13) bMis) > (^2((w,w)hs C(y),y))s + c j\\p\^ds, c > 0, C2 > 0.
6M(s)
336
From (8.4.10) and (8.4.11), one sees easily that Q^ HIQ and (8.4.14) ({Q^Q))^CD{cp,(p). Since g = 0 on b M, we see that
I \
(8.4.15) bMis) M-M (s)
Q{S,
P)\^ dS (P) < 21

bMis)
\Q{S,
P)\^ d 2 {P)
Si < s < 0 , si < 0.
< C\s\' J\Qr{s,P)\^dM
< C\s\'D((p,(p),
The result follows from (8.4.II), (8.4.13), (8.4.14), and (8.4.15). Theorem 8.4.5. Suppose each q}^^%'PQ with q>\ and suppose (fn 7 q>, ^(pn -7 W, ^'^(^ ^<Pn7 % ^^ S- Then (fn -^(p in Q, (p^'^, y) = dcp, and ^ = hq). Also ^2?a has finite dimensionality. If (p^ S^^^Q^^*, then {(p,(p) < C^d[(p,(p). Proof. Clearly 99 ^ ^ with d(p = y) and hep = Xy since the manifold of triples (99, Sep, b (p) is closed in S S S. To prove the strong convergence, let {;'} be any subsequence of {n}. From Theorems 8.4.1, 8.4.2, and 3-4.4, it follows that there is a further subsequence {s} such that 99s - > some 99', which must be 99, on each Mt with ^ < 0. Now, choose > 0. From Theorem 8.4-4 it follows that there is a ^2 < 0 and so small t h a t (8.4.16) / \(ps\^ dM < C2\t2\ 'D((ps, (ps) < (/3)2.
M-Mit2)
From the weak convergence, it follows that (8.4.16) holds also for 99. Since 99^ ->99 on M{t2), we conclude that (8.4.17) f\(p-qs\^dM<(el})\
Mih)
s> s^.
It follows that 99s -> 99 on M. I t follows that 99^ -> 99 on M. From this, it follows easily that if W is any closed linear manifold in ^TP^y there is a form cp^^'P^ which minimizes d[(p, 99) among all (p^%'P^
n W for which (9?, 99) =:: 1. We m a y let W\ = S>P9 and 991 a minimizing
form in SKi, 9K2 be those forms in S^^ orthogonal to 991 and 992 a minimizing form in ^ 2 , etc. Clearly 0 < d((pi, 991) < d{(p2, 992) < * * *. Now, suppose all the ^(99^;, 99^;) = 0. Since each (99^;, 99^;) = 1, we m a y extract a subsequence {99^} such that (pn 7 (p,^(p -^W {= 0), and 599^ ->;^ ( = 0). But then (pn -xp- But this is impossible since the 99^ form a normal orthogonal set. Hence ^ ^ has finite dimensionality and d{(p,(p) >C((p,(p), C > 0 , if ^^^3?ff n ( S 2 ^ ^ ) from which the last result follows. We can now complete the proofs of our principal results (i)(iv) for forms in S^ Q with $' > 1:
8.5. The local analysis
337
Theorem 8.4.6. (a) 7 / ^ > 1, 'SiiL'PQ) =^VQQ^VQ, (b) If co^&^^Q^^^ and $ ' > 1, there exists a unique
solution
(c) / / we define N^^co as this solution if o)^ S>^^Q ^^^ and N^^co = 0 if w^^'P^, then N^^ is completely continuous [q>\). Proof. By virtue of Theorem 8.4-5, it follows that there is a unique 9? in 2? n (S^^ 2>) which minimizes d[cp, (p) 2Re{co, cp) among all such (p.* If co^ S^^ ^^, then 99 satisfies (8.4.18) d{(p,ip) = {oy,tp), y)^% so that (p^%{LPQ) and L(p = co. From (8.4.I8) and Theorem 8.4.5, we conclude that il 9'!^ < Cz d(f, <p) = Csico. <p) < C3 1 a>l| i| ^^H, so t h a t ||<pi<C3|la>||, d(<p.<p)^CzM\^ SO N is bounded. Finally, suppose con-yco in S ^ ^ ^^^^ Then, if q)^ = Ncofi and q) = Nco, (pn7^, d(pn7^(p, and hcpn^hcp, so t h a t 99^ -> 99 by Theorem 8.4-5. The complete continuity of N^^ follows. 8.5. The local analysis In this section, we consider an equation of the form
(8.5.1)
=IJîi>(t,r)
+ Zf,4i'r)
+Mt,r)+g{t.r)
(^ = ^ 1 , . . . , t^^-i, (p^^ = dipldt^", etc.) in which the coefficients are real and of class C^ for all (t, r) with jR < r < 0 and are periodic of period 2i^ in each t^ and we assume t h a t this is true of each f'^^, /*, / , and g. Actually, we assume t h a t the a'^ depend on R (as do the other functions, and satisfy \S/P e'^^(t,r)\ <KpR^~^, ^^^'^^ ^(0,0) -= 0, ^ = 0, 1, 2, . . . 8^^(t, r) = 3^^ - a^^[t, r).
Notations. In the remainder of this chapter GR denotes the hypercube l^^'l <R, oc = \, . . .,2v 1, and Gij denotes the celUô-ij, R <r < 0 . We shall be interested in functions/^^,/*,/, g, and u which ^ C {GR) and vanish near r = R and / = i^ and we wish to obtain bounds for certain norms which we now define: We define (111^111,^)2 = ( 2 i ? ) 2 - l / 2 ' [ ( l + | m | 2 ) ^ | ^ ^ M | 2 + (8.5.3) + (1 + \m\^Y-^R^\u^(r)\^]
dr
22
338
for any real s, where here and below Um,(r) = {2R)^-^'' f u{t,r) e-^"""^'^^^dt, m = {mi. ,. W2r-i), (8.5.4)
m
The following lemma is an immediate consequence of the definition: Lemma 8.5.1. ||l^|||f+;, = 2 ' <^ IP^^ÎIII' W^)^'''*
(X == (^1, . . ., a2v-l), \(X\ ^ (Xl + ' ' + (X2V-1, Co, = j-^ :,
s being an integer and h being any real number. Moreover the norm ||| ^ l^, i? increases with s. Theorem 8.5.1. Suppose H ^ C^ for 7^ < r <: 0, is periodic of period 2R in each f^, vanishes for r = R, and is harmonic for R <Cr <. 0. Then (8.5.5) \\\H\\\l,^_n<RC{v)l\H(t.0)\2dt.
OR
Proof. Expand H into a Fourier series as in (8.5.4). Then the H^ ^ C^[^, 0] and satisfy
HZ(r)-n^\m\^R-^Hn,(r)=0,
Thus we see that
Hm{-R) = 0.
Ho{r) = coR-Hr + R) Hm (^) = ^w (sinhjr | w|)~i sinhyr | w| (r + R)IR,

/|^(^,o)|2^^ =
OR
m^ 0
2'km|2-(2i^)2-i.
The theorem follows from a straightforward computation. Let us define the *'prime-norm" '||| ^|!|si2 by '|||^|||fB= (22^)2-1/2^(1 + Then |||^||i?E = 1||^|||fE + i^2'|||^rl||f-i,E. Theorem 8.5.2. Suppose that a(t, r) and u{t, r) ^ C"^ for R < ^ < 0 and are periodic with period 2R in each ^. Then, for each real s '\\\au\\\sR <ô'|||î||si2 + C{a,s)'\\\u\\\8î^R (8.5.6) ô = niax 2\an{r)\, C{a,s)^m3LX 2! Ci{n, s)\an{r)\ \m\^)s\um{r)\^dr.
Ci {n, s) = 2'^-il/2 \s\'\n\[i
+ {\ + \n |2)is-ii/2-|.
A corresponding result holds for the full s-norm.
8.5. The local analysis Proof. Expanding a and u in Fourier series, we obtain (1 + \m\^Y'^[au)m[y) (8.57) = 2an{r)vn,m(r),
339
Vn,m(r) = (1 + \m\^rÛm-n{r) and think of the an{r) as scalars and the Vn,m (and others below) as components of a vector function %. We now write Vn,m(r) =Vn,m(r) +Wn,m(r), F^,w = (l + | ^ - ^ |^)'^^ \m-n^Y^^]um-n{r). ^m-n(r) Wn,m(r) = [(1 + \m\^Y^^ - (1 + (8.5.8) Using the inequaUties 2-1(1 + |w|2)-i(i J^\m n\^)î+\m\^<2{\ we find t h a t (8.5.9) \Wn,m(r)\ < C2(;^,s)-[1 + | m - n | 2 ] ( ^ - i ) / 2 | ^ ^ _ ^ ( ; . ) | ^ I C2 (n, s)\ < Ci (n, s). The result follows easily from (8.5.7), (8.5.8), (8.5.9), the definition of the prime norm, and the fact t h a t '|||F,j|||sij = '||| u J\SR for each n. Theorem 8.5.3. Suppose that u, v,f, and g ^ C^ for R-<r<,0, all are periodic with period 2R in the f^, u and v vanish for r = R and 0, and Au =f, Avgr. Then there are absolute constants C3 and C4, such that
1^5 b^\ < | 5 | '{a^-l + 6^-1) \a b\,
a>0,
6 > 0, \mn\^),
+ \n\^)(\ +
III ^ 1 si? < CsR^'\\\f\\\s-2,R, \\\v\lsR < C4i^-'|||^|||si?. 11

Proof. The u^, satisfy the conditions (8.5.10) < ( r ) - 7r2 \m\^R-^Ûm(r) =fm{r), Um(-R) = Um(0) = 0 Multiplying both sides of (8.5.10) by R^(2R)^''-^(i and integrating by parts yields
0
+ \m\^Y-^ Uni(r)
(2i^)2'-i 1 2 ^ (1 + I m |2)-i [71^ I w |2 I um \^ + R^ \ < j ^ ] dr
(8.5.11)
= -i^2(2i^)2v-i /'2^(^
-R m,
+\m^^)s-if,riir)un,(r)dr,
Since Um{R) = Um(0) = 0, one easily concludes t h a t
(8.5.12)
(2i^)2f-i 12^(1 + 1^12)5-11^^(^)12^;-R "^ -^ 0

<ZiR2{2R)îf 2:(i + |m|2)s-iK(r)|2^y,
22*
340
The d-NEUMANN problem, on strongly pseudo-convex manifolds
Combining (8.5.11) and (8.5-12) yields (8.5.13) 1 ^ | | | f ^ < Z i 7^2(22^)2 1 1 /2^(1 + \m\^Y~^fm{r)urr,{r)dr
from which the first result follows using the Schwarz inequality. In the case of V, we obtain the result (8.5.13) with/^^ replaced by ^^. The second result then follows by integrating by parts. Theorem 8.5.4. Suppose that u and the functions Sjf, f'^, /*, /, and g^ C^ for R^r<,0, are periodic of period 2R in each fy, and vanish along r = R. Suppose also that the e%^ satisfy (8.5.2) and that u satisfies (8.5.1). Then there is a number RQ, depending only on v and the Kp and a constant C, depending only on v such that
\MUE<c{s'\\\f^^\\UE + R2'\\\f\
-1/2,
R +
0<R<Ro. Proof. Let H be the unique harmonic function of Theorem 8.5.I for which H(t, 0) = u(t, 0) and let U = u H. Then U satisfies all the conditions with different f'^P, etc., and also vanishes along r = 0. We may write the equation (8,5.1) for C in the form 7 AU = 8%^ U^^p + e\^H,o.^ +f% +f% +fr + g = (/^ + e%Û + 8l^H),o.p + ( / - 2e%%U- lel^^^H),, + fr + + g + et.p U + 8%%^ H = F% + F^,+Fr + G. Clearly (8514) U^V% + V:^ + V + W.
^yoc^^pocp^ ^ya^pu AV = Fr, AW = G
and V'P, F", V, and W all vanish along r = R and r = 0. Now, from Theorem 8.5.2 we conclude that (8.5.15) lllef tî||i/2<^S^'l!|f/|t|i/2+ C-^Ks)'|||C7|||_i/, where A^^ and C'^{a, s) are the constants of Theorem 8.5.2 for 8%^ = a""^ and where we have left out the subscript R on the norms. But, from (8.5.2), (8.5.4), and differentiation, we see that 2\m\^^\8%l{r)\^<ZlKlR^ (8.5.16)
m
,p =
0,i,2,...
I\m\^p\'e'A{r)\^<ZlKl.
m
Thus, from (8.5.15) and (8.5.16), we see that (8.5.17) 'l<fiV\li^^n<A''l>-R-'\lVly^_j,, V= UovH.
8.6. The smoothness results
341
where ^ ^ depends only on the K^. From (8.5.14), (8.5.17) and Theorem 8.5.3 and Lemma 8.5.I, it follows that
11 U 111,/,,;,, < C^ \Z ['||i/^ III,/, + i?^-^C||| U III,/, + 'I ^ III,/,)] + 1 (8.5.18) +2'['ill/11l-i/2 + ^''{'!llf^lil-i/2 +'111 ^1-1/2)] +
a
+ R '\y\Ui^ + R^ ['III g 111-3/2 + R-^ A ('ill [/ i||_3/, + '\\\H \U)]]

where Ci = Ci(r) and the ^ ^ , A"^, and ^ depend only on the Kp. The result follows from (8.5.18), Theorem 8.5.1, and the fact that u = U + H. A similar analysis can be carried through for functions u^ C'îCR) where CR is the full cube CRilt^^l <R, a =- 1, . . ., 2v. We are interested in the solutions of the equations
(8.5.1)'
z a'^nt) ,a? = zf% + if,^ +/.

u\
m
where we have replaced rhyt^^ and the coefficients and functions ^ C^ and are periodic of period 2i? in each P" and the a^"^ satisfy (8.5.2). We define the s-norms by
m
u(t) = Ûme^'''^'^'^,
(m = mi,..
.,.m2v).
Theorem 8.5.5. Lemma 8.5.I holds if (x = [CKI, . . ., a^v). Theorem 8.5.2 carries over without change. In Theorem 8.5.3 ^^ ^^^<^ retain only the case A u =f. In Theorem 8.5.4 we need consider only the cases where u satisfies Au = e^û,ufi+f%+f% + g, where oc, ^ == i, . . ., 2v, in which case the result is
llkHI|i/2<c[2'1!l/'^|||i/2 + i?2^1ll/||U2 + i^2^lk

[a,/5
a
o<R<Ro.
8.6. The smoothness results Until further notice, we shall be concerned with forms of a given type (p, q) in which ^ > 1. We begin by defining the s-norms for forms (p^WP^. Each interior point P of M is in a neighborhood 3J:p which is the image of a cube CR with respect to a holomorphic coordinate patch T. Each boundary point P of M is in a neighborhood "^p on M VJ h M which is the image of GR D OR under a non-analytic [t, r) boundary coordinate system of the type described at the end of 8.3. Let us choose a covering It of M U b M by means of such neighborhoods 9^1, ., ^T and choose a partition of unity {rjt}, t = 1, . . ., P , of class C^, each having support in the corresponding 9^^ In each 5ft^, let us
342
choose an orthogonal basis (C, f) as in 8.3 which satisfies (8.3.24) if ^t is a boundary neighborhood. We then define
(8-6.1)
lll<?'!llf = 2'2'lh*î*illlf
t IJ
where the 99J*} are the components of 99 with respect to the basis (C, f) in 9^^. Of course the norm depends on the 9^^, the coordinate patches rt, and the bases {Ct, Ct) but any two such norms, for a fixed s, are topologically equivalent. We choose one fixed set of Sl^ f]t, and {Cty Ct)- Then we have the following fact: Lemma 8.6.1. \\\(p\\\s ^'s non-decreasing in s and if rj ^W^'^, \\\'yj(p\\\$ We begin by proving the fundamental inequality: Theorem 8.6.1. If (p ^ 9lg^ {and q > 1), then (8.6.2) \M\m^Cî^>9^)> where C depends on M and the metric and the choices of 91^, rjt, and {Ct, Ct)Proof. It is clear, from Theorem 8.4.1. that, for a given 99 in 5Ig^ there is a unique (po^ HIQ which minimizes D{q) cp,q) (p) among all q) in H\Q. Since this just leads to the Dirichlet problem, 990 and hence H = (p (po^ ^0- Clearly, also, 990 satisfies (8.6.3) (8.6.4) (8.6.5) D((po-(pyip)= [L + I)HÔ, -D{H,y))=0, yj^Hl^ so that we have D(<p,(p)=D((po,(po)-\-D{H,H). From Theorem 8.4.1, the definitions, and Lemma 8.6.1 we conclude that lllô |!!?/2 < ilôlif < Zi ((^0, ^0)) < Z2D((po, (po) < Z2D{(p, <p). Thus we need only prove (8.6.2) for H^SHio and satisfying (8.6.4). From the form of the integral in (8.3.I5) it follows that the components Hij of H in any holomorphic coordinate system satisfy equations of the form [%.6.6) g^P Hjj-x^p + (lower order terms in all the HST) = 0. The components Hjj with respect to an orthogonal basis, being linear combinations of the Hjj (with C^ variable coefficients), satisfy similar equations. If we multiply by 4, write g'^ = g^^ + i g^^, and express the derivatives in terms of those with respect to x^" and y', the equations (5.6.6) take the form
{S.6.7) gf UU ,p + gf Uioc î3 + ^ f ^ ; a ,/? + gf Ui,ccyfi+ =0
where we have denoted the components by uK Now with each boundary point P we choose a {t,r) boundary coordinate patch Tp from GA U CA ô M [J b M so that the origin corresponds to P and the metric has the form (8.6.8) ds^ = ao,^{t,r) dif^ dt^ + (dr)^, a^(0,0) = ^^.
343
We also choose a (C, C) orthogonal basis satisfying (8.3.24). The equations (S.6.7) then take the form
It is clear t h a t the mappings may be chosen so that A and the bounds for the coefficients and all their derivatives are independent of P Having chosen xp, we define <
(8.6.10)
(/, r)^b-^-\-
hn [t, r) [a-^ {t, r) - d-^],

2^-1
{t,
r)^Gn,
hR {t, r)=h[2R-^
r] [J h[IR-^
tr],
where h^ C^(iî), h{t) = 1 for |^| < 5/4, h[t) = 0 for \t\ > 7/4, 0 < ^(/) < 1 for all t\ then we extend a^ to be periodic of period 22^ in each P, We define J^^jj, etc., similarly. Then the a"^ satisfy (8.5.2) with K^ independent of P . Then we can choose an iî, 0 < jRi < c i?o (Theorem 8.5.4) and a finite covering of the strip ~ P i < y < 0 by the parts of the ^P corresponding to Gpi^. In like manner, we may cover M_p^ by the parts corresponding to C7j/2 of holomorphic coordinate patches with 0 < P < Po so the af/ satisfy (8.5.2). Now, suppose H is harmonic and in 9lo and let i] have support in one of these small neighborhoods, say a boundary one. Then if W are the components of K, with respect to (f, f) and W^ = rj W, the u^ satisfy (since the ^^ - al^ in GR,^) / ? = (V îcR + 2(5;jt al^ rjj) HK
(V^JfcR + 2dj]cr}r)r\H^'
fi = {v h'lcR + 2djjc Vr) H^
gj = bl CjjcR + djjc(rjrr + a'lfrj^^^) in ÎkR + 2^3Jc ^lfvp) , From this and Theorem 8.4-4 with cp = H, we conclude that
i:'\mn-î:'mu+'\\\sn\\\i+2f\u^^dt<cD{H,H),
lr,H\\l,,^<.CD{H,H).
where C depends on rj (possibly P ) , and the K>p and the norms may be taken on 501 or on GR as desired. Thus, from Theorem 8.5.4, it follows that A similar result holds for the inner neighborhoods. If the rit form a partition of unity, the result follows. We have already seen that [L + / ) - i is a bounded operator defined everywhere on ^'P^. We shall prove the following theorem concerning the solutions cp of (8.6.11) {L + I)(p = co, (p^'^m, or (p = (L + I)-^ co. We let denote the closure of 51 with respect to the s-norm.
344
Theorem 8.6.2. / / CO^SP^, (8.6.12) Ifo)^ '^^-^l^for

(^6.13)
then cp ^^^^ and
|||9^|||i/2<C||co||. an integer s > 1, then cp^ +i/2 ^^^

|||9^|||5+l/2 < C(s)'|||o>|||5_l/2.
Ifoy^^thencp^S^Q. Proof. We shall show first that if co^ 5t, then (p^%Q and all the inequalities (8.6.12) and (8.6.13) hold. The first statements then follow by approximations. To do this, we assume co^ 51^^ and for each e > 0, we let 99g be the function minimizing (8.6.14) D,{cp,cp)-2Re{m,cp), cp^Sô, D,((p,ip) = e(((p,y^)) + D((p,yj),
This minimum exists and is unique and ^ 5lo as is seen from the theory of 6.4 (as long as > 0). It satisfies (8.6.15) DM>w)-ico,w)> V^3lo. We shall show by induction that if (p^ satisfies (8.6.15), then (8.6.16) e II 9P, lllf+i + II <p, If+y^ ^ C(s) '\\\ co \\\l_ s = 0,1,... We note first that (8.6.16) holds for s = 0, since
* I I % illf + II 9s lllf/2 ^CD^(9? 9?,) = C (m, <p,)
^^^'^^ ^C'WCO
|||_x/2 111 9-. III1/2 < \ II CPe lllf/2 + C 'III , r-V.
(using Theorem 8.6.1) from which the result follows. From the definitions in (8.6.1) and from Lemma 8.5-1, we find that
^ III ^ IIU+i i^ III 9 llls+1/2 Z. Zil^
n
IJ
e III /r. Ill 2 _J_ III /y, III 2 V V fj:" lll-y)
II Vn (Pij llls+1 ^
m(*MI|2
_1_ ||| oo
II Vn (Pij |||s+i/2J
Ar,(w)'||l2
"] I
(8.6.18)
+ 2 2
SC.[s\\\Dtf,n<F'^J\\\l
l\Dtf]n<p?J\il^(Rlnr^''^
n IJ \(x\<^s
where n indexes the coordinate patches and the norms on the right in (8.6.18) are those in GR as defined in 8.5- Our method of showing this will be to use the first inequality in (8.6.17) for the tangential derivatives D'i^n(P) and then use results like (8.6.15) for these derivatives. We begin by considering what happens in one of the coordinate patches which we take as a boundary patch, the treatment for interior patches being similar and simpler. We first replace the subscripts / / , etc., by a single superscript. Then, if 99 or ^ has support in the range of a coordinate system with domain G {= GR), (8.6.19) (8.6.20) ^ A^,t = Aff, Cu = Cii.
where we have put r = t^^ and D^ is hermitian symmetric, i.e.
345
Next we prove by induction that (if cp^ and ^ = 0 outside our patch) (8.6.21) D,[D' f,, xp) = (-\r\DA<Ps, D'f) + 2 " ^ ^ ( 0 " <p,. <p),
where all the D^^ are given by hermitian symmetric integrals like D^, where D^ and Df^ are tangential derivatives, and where (8.6.22) fji <X means 0 <^â < /la, oc = \, . . .,2v i and |//| < | / i | . (8.6.21) is evident for |A| = 1. To complete the induction, we assume (8.6.21) and use that for | ^ | 1 to obtain (8.6.23) D,{D^^' 99, ^0 = - D,{D' 99, D^ xp) + DU[D' cp,yj) cp = cp,. Applying (8.6.21) to the first term on the right in (8.6.23), we obtain
D,{D^+^(p,y)) == (-1)1^1+1 A ( ^ > ^ ' ' ' ' ^ )

(8.6.24) _ ^D',^{D^cp,D^xp) + DU{D'cp, xp).
The result for A + ^ follows by applying (8.6.21) for 1^| = 1 backwards to the terms in the middle sum on the right in (8.6.24). Now, it is easy to see that we may choose a real function Y\ of class C^ and vanishing outside our patch (see (8.6.10)) such t h a t (8.6.25) (8.6.26) 1V^(^)|2<C7^(^), xp ^Yi^D'^cp, t^G. We now assume that 99 (E^ (p^ satisfies (8.6.15) and then set A given in (8.6.21). By using the hermitian symmetry, the reader may verify t h a t ReD^{D^(p,ri^D^(p) (8.6.27) = D^{rjD^ cp^rj D^ cp) - j
G
Alli^,o^Yi,p0^0Ut,
( - 1 ) ' ^ ' D,{(p, D^ xp) = ( - 1)l^l(co, D^ xp) = (D^ (o, > 2 D^ cp), y 0i = D^(p\ (p = (p^^
using (8.6.15). In order to handle the various integrals JD^^, we let / denote any hermitian symmetric integral with coefficients a*f, b'^p Cfj. We notice that, for any vectors Q and a, I(u,r]v) (8.6.28) = I (rj UyV) + ^ [{dj u, v^) (u^, dj v)] '' dju = a'tfrj^û]^ + îYj,oûK
We also note t h a t (8.6.29) djT] u = T] djU + a'lf rj,ocr]j ^^*. Now, suppose ^ is a multi-index with |^| = 1. Then, by applying (8.6.28) twice and (8.6.29) once, we see that (8.6.30) I(xp,rj^xp,Q) =I{rj^xp,xpe) + 2 ^^ [(dj xp, rj x/P^^) - {y)^\ 1^ dj xp ^ g)].
j
Integrating / {yf xp, xp^ g) by parts and transposing one integral, we obtain (8.6.31) 2ReI(xp,rj^xp,g) = Re{-2l{rj rj,gxp,xp) leiv^V,^) + *}
346
where * denotes the sum in (8.6.30) and IQ denotes the integral with coefficients a^^l^, etc. Applying (8,6.28), (8.6.29), etc., we obtain 2ReI [ip, r\^ ^, g) = /^ (j] y},r]y)) + LQ (ip, ip) + (8.6.32) + 2Re(-I(rj^,yj,r]7p)
?
+ Z[{^3V,Qy^>W') - iV^Q^^^Jf)]
where LQ is an integral like that in (8.6.27) with coefficients a^l^. We now complete our proof of (8.6.16) by induction on s: We suppose it is true for all s < ^ and consider the case s = ^ + 1 Suppose | X \ = k+\. From (8.6.17), (8.6.21) w i t h ^ = ^2/)A ^ / a n d from (8.6.27), we conclude that
Using (8.6.25), (8.6.32) and the facts that
(8.6.33) |(/,^)|<Ill/Ilk'llkllk.,
'|||v^|l|,<cj^-illl^||U,
for all T, where Vu is the vector function (uâ), oc \, . . ., 2v, t^^ = r, we find that (taking all norms on Wl and allowing the C ^ to depend on R) (8.6.34) 1 jAl^ri,o.Yi,^D^(piD'pdt\ '<^ <C|||Z)^^^|||_i/2-|||^D^9?l||i/2 '
|(7yZ)^co,7yZ)|<C|||i)â>|||i/2|hZ)^^|||i/2.
Regarding the integrals D^^ with |//| < |A| 2 as sums of appropriate inner products one can easily obtain the bound
i i ? . z)^-^(Z)''(?> ,j2 D^-<p,)! < c [III D " 9,, JII3/2 lih z > V . III1/2 +
(-6-35) +|i|Z)/'9,J||i,2-|l|,yZ)^9,,|||_i/2], \X-^\>2. One may proceed in much the same way for those integrals with | A //1 = 1; however, one starts with (8.6.32). We obtain, for example \I(r]eDf'(p,7]Df'(p)\ < C'\\\\7 fje Df" (p\\\_i,2' '\\\V rj D^ (p\\\ii2 + < C WlrjgDf' (f\\\ii2' \\\rjD^ (p\\\ii2 H \(D^(p^,rjdjD'(p)\^C\\\D^(p\\\il2'\\\rjD'q>\\\il2 + "'" (q) = %)
where the dots denote terms of lower order. The other terms can be handled similarly. By using the common device that \2ab\ <ia^ +i-^b^, ^>0, a,h real, | small and by summing over all X with |A| = ^ + 1 we find that
U|=fc+1
ZQ[eIIIriD'<pll+ 1 nD'<p\\U <: c2"Q ||hz)ô)r_i/, + 1 + CZ\\D''vrm\fi\^k
|A|=fc+l
347
Then by summing over the patches and adding in the lower order terms according to (8.6.1) and Lemma 8.5.1 we obtain (8.6.16) for s = ^ + 1. This completes the induction. By repeated use of the differential equations (8.6.9) we find in each patch that all the derivatives of the q){ satisfy similar inequalities and (8.6.12) and (8.6.13) hold for (p^, independently of e. Thus as ^ 0 it follows easily that q)e->(p together with all its derivatives. Thus (p^ô if C ^ 51 and (8.6.12) and (8.6.13) hold. O We can now prove our desired smoothness theorem: Theorem 8.6.3. If (p ^ "^{LPQ) and 5 > 1, then cp, d(p, and b+^) and (8.6.36) LV'Q+^d(p = dLPQ(p, LP^Q-^hq) = bL^^cp.
^ / ^ d^^ and ^ > 1, thencp^ % . Proof. We first prove the smoothness results. Let L^^ cp = L (p = co^ ^ . Then ^ ^ S2^ and {L + I)(p = CO + (p. Hence 9? and therefore co + (p^ ^^^^ by Theorem 8.6.2. Using induction and Theorem 8.6.2 repeatedly we see that 99^ ^+1^2 for every s. Using the differential equations, we find that all the derivatives of 99^ ^+1^^ so (p^%o (Sobolev lemma, Chapter 3). That dcp^ô follows from the natural boundary condition (8.1.23). That bq)^ô follows from (8.1.19). Since Lip = dbip + bîp for ip and 5 ^ ^ %Q, (8.6.36) follows easily from the facts t h a t dîp = 0 and b b ^ 0 for ^ ^ 51. Now, suppose c o $ S ^ ^ G ^ and L(p ==^w. Let a)n 51^^ 0 ( S ^ ^ ^^), (pn^^{L)y I^(pn = (JOn, and suppose con^co in S^^. For each n,
d{dq)n) = 0 , b(b(pn) = 0, bdcpn + db(pn = (On,
{bd(pn,df)
= {db(pn,hx)
= 0,
yj^Sav^Q-^
x^'ô''^^
Thus dq)n and bq)n satisfy the equations {dd(pn,dip) + {bd(pn-Wn,bip)=^0, (ab99^-co^,5;^) + (bb99^,b;^) = 0, I t follows t h a t
d{d(pn, d(pn) + d(b<pn, bcpn) =- {(^n, 0)n)
'^^ îg'^+l x^'^l'"'^
from which the remaining statements follow (by taking cp^ = N con so t h a t 99^->99). We can now complete the proofs of our principal results in the cases where ^ = 0.
348
Theorem 8.6.4. (a) m (Lv o) = gp o Q |,p o. (b) / / co^ S ^ O 0 i ^ ^ there is a unique (p^% {L^^) Pl [S^^S>^^] such that Lô (p=.o}. (c) / / we define N^^ co as this solution when CJO^S>^^Q $^^ andN^^ co 0 when co^ ^2?o^ ^/^^^ jsivQ is completely continuous. Moreover ||iV^ll<C3, C3 being the constant in Theorem 8.4-5(d) / / C ^ %v^, then m^ w ^ ^v^. O Proof. Suppose co^^^^. Then Sco ^ ^P^ and if ^ 6 ^ ^ (5 0),y)) =^ J <ft), V y))dS = 0
bM
so aco 2^10 ^^. So let 0 = A^^co. Then ( 9 0 $ (L^^s) Q ( S ^ 2 0 |^2>2) (as above) and L^^ d't-' = ddco == 0 so d0 = 0 and hence (8.6.37) d(b0-co) = O, 5 0 - 0 ) ^ 2^0. On the other hand if co$ S^^ 2?o and ip^ ^P^, then (8.6.38) ( 5 0 - CO, ip)s = ( 0 , dip)s - (co, yj)s - f <y0,ip}
bMs
dS,
By averaging this over a small interval (t, 0) and using the fact that the function v0^C^(M) and vanishes on bM, we find that the right side of (8.6.38) -> 0 as ^ - ^ 0~. Thus we must have (8.6.39) (8.6.40) h0 = o), Q = N0, 30 = 0. (p = bQ.
Since 0 {= N dco) ^ S^îQ P I , we may define
Since 5 0 = 0, we again have dQ = 0, so that As in (8.6.38), we see also t h a t ^ ^ S^ÔQ ^JJO. If there were another solution in %[L'P^), the difference (p\ would be in ^[L^^) and we would have
(pi^'^'P^, d{cpi,ip) = {dcpi, dip) = 0, f^ ^2^0
so that 991^ ^P^, since we could take ip = cpi. From (8.6.39) and (8.6.4O) we see t h a t N'P^ is bounded. Finally, if co^ -^ co, it would follow t h a t h0nyM, d0n = 0, and 0n -^ 0, at least in a subsequence, since, by Theorem 8.4-5, D{0n, 0%) is uniformly bounded. Thus, by that theorem 0y^ -> 0 and so (pn -> (fy since [(p, (p) = (bQ, bQ) = d{Q, Q) < C3 \\LQf = Cl\\oyf. = Cs(0, 0) < C | d{0,0)
9.1. Introduction. Parametric integrals
349
Chapter 9 Introduction to parametric Integrals; two dimensional problems

9.1. Introduction. Parametric integrals An integral (1.1.1) is said t o be in parametric form, and we say f is the integrand of a parametric problem, if and only if the value of the integral is unchanged b y an arbitrary diffeomorphism with positive Jacohian from G onto another domain G'; if we assume t h a t / {%, z, p) is defined everywhere, we see that we must have j f[x\ z'{x'), \/z'(x')] z(x) = z'[x'(x)], dx' = j f[x, z{x), Vz{x)] dx, z'p(x) = ^ ^ ^ . - ^ - ^
for any G', and vector z'^C^(G'), and any positive diffeomorphism x' = x'(x) from G onto G\ By taking XQ, ZQ, 'p, and XQ as arbitrary constants and defining '^^ = 4 + ' ^ i ( V - ' ^ S ) , 'x-^'xl + a1(xf^^xi), ^^^^^ (9.1.3) det lla^ll 'f[x^, zo,p[,.. detl|a|l>0 .,pl] - / [ ^ o , ô, aîPa,.. ., ;] and taking G = B (XQ, Q) and then letting ^ -> 0, we obtain for all sets of constants as indicated. Thus, / must be independent of x and we must have iV > r. In case N = v + 1, we now show that / must have the form f(z,p)==F(z,Di,.,.,Drî), (9.1.4) A =
' ^l'"Pi-^PX''-"Pi
where F is positively homogeneous of the first degree in the Di. This is easily seen as follows: (9.1.3) implies that if the p matrix is multiplied on the left b y any r X r matrix with positive determinant, then / is multiplied by that determinant. Thus, if Z)r+i > 0, we may multiply the p matrix on the left by the inverse of the matrix of Dvî. The result is t h a t the new ^-matrix has the identity for its first v columns and the elements in the last column are just ^^DijDvî. Thus, for some F, (9.1.5) f{z,p)^D.^,-F(z,^ , ^ \ ) , Z),î>0.
If D + < 0, we begin by changing the sign o 1 row and obtain î f (9.1.5') /(^.^) = - ^ ^ . i ^ ( ^ . - ^ . - - - . - ^ . - l ) ' ^^.KO
350
Parametric integrals; two dimensional problems
where the two F's join up continuously along Dv+i = 0. I t is not necessarily (and usually not) true that (9.1.6) F{z, -Di, . . ., - D . + i ) = -F(z, Di, . . ., Z),+i). It is clear from this discussion that an integrand of a parametric problem is ordinarily not of class C^ on the set of p for which the ^-matrix has rank <i^ and, moreover, at points w h e r e / i s of class C^, the quadratic form in (1.10.7) and (1.10.8) necessarily has vO eigenvalues and the biquadratic form in (1.5.5) degenerates. We have already seen that the area integral (1.1.4) is in parametric form. Using our current notation (9.1.7) f(z,p) = yni + Dl + D|. Until H. A. Schwarz showed how to inscribe a polyhedron of arbitrarily large area in a portion of a cylindrical surface, it was believed t h a t the area of a surface could be defined as the sup. of the areas of inscribed polyhedra, by analogy with the definition of the length of a curve. This upsetting discovery was the motivation for a lengthy study of surfaces and their areas. The principal surviving definitions are those of the Hausdorff two-dimensional measure of a point set in R^^ N >-_ } , and of the Lebesgue area of a Frechet surface. The Hausdorff measure /12 is generated in the usual sense of measure theory by the outer measure /12 * defined by (9.1.8) yl2* (5) = lim Al * (5),
Q-*-o
A^ * (S) = ini 2J 7t rf
n^Q
for all coverings of 5 by a countable number of balls B[xi, ri). To define the Lebesgue area, we must define a Frechet surface. The idea is to invent a notion of surface which is independent of its parametric representation. In case we have a point set 5 which is the homeomorphic image of a closed circular disc, we would naturally call S a surface and any homeomorphism from a closed Jordan region onto 5 a parametric representation. But there are times when it is desirable to consider more general surfaces. We therefore proceed formally as follows: Definition 9.1.1. Let ^1 and Z2 be continuous vector functions with diffeomorphici compact domains Gi and G2 which must be finite unions of disjoint manifolds, with or without boundary, of class C^ at least. We define the distance (9.1.9) D(zi,Z2) = inf (max Ui (:î) Z2[r{xi)]\\
1 Historically, homeomorphic domains and homeomorphisms were allowed instead of diffeomorphic domains and diffeomorphisms. In the one and two dimensional cases, the restriction which we are making results only in the restriction to domains of class C^; in higher dimensions, Milnor has shown that two manifolds of class Cl may be homeomorphic without being diffeomorphic. Our restriction results in a great simplification of the theory of integrals over Frechet varieties.
351
for all diffeomorphisms T from Gi onto G2. If D (zi, Z2) = 0, we say that zi is equivalent to z^ and write zi ^ ^2Theorem 9.1.1. (a) The distance D so defined is non-negative, symmetric and satisfies the triangle inequality. (b) The equivalence relation defined above is symmetric and transitive. Definition 9.1.2. We define a Frechet variety S as a class of continuous vector functions which consists of all such which are equivalent to some given one. Any vector function in S is called a parametric representation of 5 . The range 9i (5) is the common range of the parametric representations of 5. The topological type ofS is the class of domains of the vectors in 5 and the dimension of S is the common dimension of these domains. Definition 9.1.3. If 5 i and S2 are Frechet varieties of the same topological type, we define the distance D {81,82) =D(zi,Z2)
where Zk is a parametric representation of SA:, ^ = 1, 2. Theorem 9.1.2. (a) With this definition of distance, the Frechet varieties of a given topological type form a metric space. (b) If 8 and 8n are all of the same topological type, z is a representation of 8 with domain G, and D {8, 8n) -> 0, then H ^ representation Zn of 8n, with domain G, such that Zn-^z uniformly on G. (c) If z and Zn are representations of 8 and 8n, respectively, all having domain G, and if Zn z uniformly on G, then D {8, 8n) -> 0. > Proof, (a) and (c) are evident. To prove (b), let Zn be a representation of 8n having domain Gn. From the definition of distance it follows that there exists, for each n, a diffeomorphism Zn from G onto Gn such that \Zn(^) Z(x)\ <.D(8,8n) + \ln, Zn{x) =Zn[rn{x)], X^G. The result follows. Now LEBESGUE [1] was familiar with the following fact: Suppose Zn and z are each piecewise linear homeomorphisms on a polyhedral domain G and suppose Zn converges uniformly to z (nothing being said about derivatives) on G. Then (9.1.10) ^ (^) < lim inf ^ (^^).
A (z) and A (%) being given by the integral (1.1.4) and being the elementary areas of the corresponding polyhedra. In the next section, we shall prove a general lower-semicontinuity theorem from which we can conclude t h a t (9.1.10) holds if G is of class C^, z is Lipschitz, and each Zn^ C^{^)'> in fact z may merely ^ HI (G) in the two dimensional case. Assuming these results, the theory of Lebesgue area proceeds as follows:
352
Definition 9.1.4. Given G of class C^ and z ^ C^ (G), we define (9.1.11) L(z,G) = inf lim inf A (zn, G) uniformly
for all sequences {z^^ in which each z^ ^ C^ (G) and Zn -^z on G. The following theorem follows immediately: Theorem 9.1.3. We suppose that v = 2. (a) L(z,G) is given by (I.I.4) ifz^ H\[G) H C^[G). (b) L {z, G) < lim inf L {zn, G) if Zn-^z uniformly on G.
(c) Given z^C^ [G), there exists a sequence {0^} in which each Zn ^ C^ (G), Zn-^z uniformly on G, and L(zn, G) ->L{z, G). (d) L(zi) = L(Z2) if D(zi, Z2) = 0. Proof, (b) and (c) are evident, (a) follows since the integral is lowersemicontinuous according to Theorem 9-2.1 so t h a t L(z, G) < the integral A (z, G). On the other hand if Zn->z uniformly on G and strongly in H\{G), then A {zn, G) ^ A (z, G). To prove (d), let Zn -^Z2 uniformly on G2 so that A{Zn, G2) ->L{z2, G2) and Z ^ ^ C^(G2). There exists a sequence {xn} of diffeomorphisms from Gi onto G2 such that Z2 [tn (xi)] -^zi{xi) uniformly on Gi. Thus, if Cniî) = ^nltniî)], then Cn->î uniformly on Gi and each Cn ^^ (^1) so that L (zi, Gi) < lim inf A (Cn Gi) = lim A {Zn, G2) = L (^2, ^^2) The argument is symmetric. Remark. The foregoing definition and theorem can be carried over to the case where G is a manifold by using coordinate patches and partitions of unity. Definition 9.1.5. We define the Lebesgue area L(S) of the surface 5 as the common value of L {z, G) for its representations z. Theorem 9.1.4. (a) Parts {b) and (c) of Theorem 9.1-3 ôld with {z, G) and (zn, G). replaced respectively by S and Sn(b) / / 5 possesses the representation z^H\{G) C\ C^[G), then L(S) = A (z, G). Now, there are some anomalies. If we define z = z (x'^) for {x^, x^) on G = 5(0,1), then it is obvious that L{z, G) = 0 even if the range of z fills a cube. But one might hope that if 2: is a homeomorphism with range 5, then L{z, G) would be equal to the Hausdorff measure A^{S). However, this is not the case. In fact, BESICOVITCH has given an example of such a z and S for which L {z, G) is finite and the three dimensional measure of 5 > O! On the other hand from an old theorem of the author [3], it follows that this S possess a ''generalized conformal" representation z = 2:*(A;*) on 5(0,1) = G in which 2:* is continuous and in H\{G)
353
and, moreover, * = G*, F* = 0 a.e. L (z, G) = L (^*, G) = f^* "^ ^* dx*

(9.1.12) G
* = |^*l*|2,
F^ - = 4 l * - 4 2 * ,
G* = 1^*2* |2.
We notice t h a t the area integrand (9.1.13) since (9.1.14) ]/* G* - (F*)2 Ê*+G* ^^^^^^^^ ^ * _ ^*^ F * - 0 B=F* C =G^).
^ C - ^ 2 + | ^ j l ^ \ ^ 4 _ 5 2 _/^1_\^(^ _ * ^
However, the reader can easily prove that if 2: is a diffeomorphism of G onto S, then L {z, G) = A^ (S), In deahng with more general integrals (9.1.15) If{z,G)=
G
jf{z,p)dx
in which / satisfies (9.1.3). it is desirable to introduce the notion of oriented Frechet variety. This is done by defining the distance D (zi, Z2) in Definition 9.1.1 allowing only positive (i.e. with positive Jacobian) diffeomorphisms r and requiring the domains Gi and G2 to be oriented. Then Theorem 9.1.1 holds and we may define an oriented Frechet variety as in Definition 9.1.2 using the ''oriented distance" above. We may then define Z)(Si, S2) for oriented varieties as in Definition 9.1-3 and then note that Theorem 9-1.2 holds. For the special class of integrand functions /(2', ;/>) of (9.1.15) which is described in the next section, we can extend the notion of Lebesgue area to t h a t of an integral over a Frechet variety. Definition 9.1.4'. Suppose / satisfies (9.1.3) and the conditions decsribed in 9.2 below. Given G of class C^ and z^ C^{G) we define S/(^, G) = inf {lim inf If{zn, G)} for all sequences {zn} as in Definition 9.1-4. The following theorem follows as before using the theorem of the next section: Theorem 9.1.3'. If f satisfies the conditions above (with v ^ 2) then Theorem 9.1.3 holds with L replaced hy ^f and A [z, G) by If{z, G), except that we must require z^ C^ [G) if v '> 2 in part (a). Many people have made important contributions to the theory outlined above. We mention only E. J. MCSHANE ([1], [2], [3], [4]), T. RADO
([3], [4]), L. CESARI ([1], [2],'^[3]) and the author ([1], [2]). L . C . Y O U N G
([2], [3]. [4], [5]) introduced the idea of a ''generalized surface" analogous to MCSHANE'S idea of a generalized curve; this idea was developed at
354
some length by YOUNG, W . H . FLEMING and others ([1], [2], [3]). The
idea is suggested by the following: Suppose z^ Hl(G). Then the integral (9.1.16) IF[z(x),Di{x),.
G
.,Drî{x)]dx,
where F is homogeneous of degree 1 in the Di which are defined in (9.1.4), is a linear functional of F. Any linear functional of F, which is > 0 whenever F is, is called a generalized surface. Unfortunately, an arbitrary generalized surface does not arise from a vector function z by means of a formula (9.1.16) even if fairly general vectors z and ''generalized Jacobians" are allowed. The ''integral currents'' introduced by F E D E R E R and FLEMING seem like a more promising tool in the study of the existence and differentiability properties of solutions of variational problems involving such integrals. 9.2. A lower semi-continuity theorem In 4.4, we saw t h a t if/satisfies (9.1.3) and (9.1-4) with N = v + 1, t h e n / i s weakly quasi-convex in p if and only if F [z, X) is convex in the X]c (we shall use Xi instead of Di from now on) which, in turn, implies t h a t / is quasi-convex i n j ^ . In t h a t case, if F^C^ for all [z, X) with X ^ 0, t h e n / satisfies the conditions of Theorem 4.4-5 and hence the integral (9.1.15) is lower-semicontinuous if z^-^z uniformly on G and weakly in Hl{G). This result, however, is not sufficient for the theorems of t h e preceding section. So we shall have t o prove a special lowersemicontinuity theorem for such integrals. We shall use the method of MCSHANE [3]. This method can be extended to certain integrals (9-1.15) in which iV > r -j- 1 as well as to all such integrals if TV = r + 1. Lemma 9.2.1. Suppose f{z, p) ^ C^ for all (z, p) for which the p-matrix has rank v and suppose f satisfies (9.1.3)- Then if (ZQ, po) is such that the rank of po is v, there are unique constants ai(I = ii, . . ., ip with ii < ^2 < < 4) such that
1 i)h . . h^v 1
f (20. zo) =Z<nH:

I
XI = det = 0 for all (^-.a) {d zi = dzh /\ .
(9.2.1)
A-Af{zo.P)-:ZîX'] âidzi
I
= AdI:^ ^...^^f"
for some linearly independent linear functions C'îz), . . ., C^(z). Proof. We first note that if we make a rotation of axes in the ^-space and make the corresponding change in the ^-matrix: zi = c] 'zK PI = cYpi, f(z, p) = f {z\ p')
9.2. A lower semi-continuity theorem
355
then the X^ transform in the corresponding way, (9.2.2) XI = Z^jX'^,

J
Oj = detcj^
and the conditions (9.2.1) are preserved in the new variables. We may therefore assume that (9-2.3) po = \
\pi,...pi,o...ol /i...o^r'---^f\
For p near po, we use (9-1-3) to conclude t h a t f(zo.p)=Xy.--'f[z^,p], (9.2.4) X^ = Xi---''Xi, :/>l=(Zi..-)-idet ' p^{ _ _ \o...opi+K..p^/ ' ^ ^
\pi---pr'Pipr'- -PI J
the XI being formed from the p matrix. From (9.2.4), we conclude that f{zo,Po)-Xl"-^f{zo,Po)
f{ô,po)
(9.2.5) fpi(zo,po) =
dX^'-^
\ <:i
<v
fpi {ô> Po) dpi
+ 1 < ^ < iV '
where po is, of course, the matrix obtained from p by setting all the Pi, with ^ > r, = 0. If we define (9.2.6) fi(zo.p)==^ZaiXi=:^X^--'-2Jî^'>
I I
we obtain, using the formulas (9.2.5) f o r / i , that

fl(ZO,po) = ^l...rô"^ /dX^'"^'
^1.
fliÔ^Po) =
ai,..v
fipi{ô,po) (9.2.7)
dpi dx^dp^
I >
2;(-i)^-^î...<5_i,5+i,...,H
V.
So, from (9.2.5), (9.2.7) and the first two lines of (9.2.1), we conclude t h a t we must have (9.2.8) î...v=/(ô,ô), î....5-i,6+i,...,H = ( - 1 ) " ~ V 2 ? j U , ^ o ) . Now, let us introduce linearly independent C* by (9.2.9)
7 = V+1
23*
356
Then, if we define cj = (5j- for 1 < / < v, the c-matrix has the same form as the p matrix. The third hne of (9-2,1), then requires that (9.2.10) âjdzi =A2J det
^^l ^r
Equating coefficients in (9-2.10) for I = (\, . . .,v) or (1, . . ., ^ 1, (5 + \, . . .,vi) yields (9.2.11) ai,,,r=Â=f{zo,Po), c\ = [f{zo,por^fvi{zo>Po)
1
'v
so that the remaining aj are uniquely determined. The C* are, of course, determined only up to a linear transformation among themselves which affects only the factor ^ . In order to cary over MCSHANE'S proof, we impose the following requirements o n / : General assumptions on / . We assume that (i) / is continuous for all {z, p) and ^ C^ for all [z, p) for which the rank of the p-matrix is v; (ii) f satisfies {9A.}) for all {z,p), etc., (iii) f(z, p)>m \X\, m>0,\X\^ = Z (^^)^ (iv) if the rank of pQ is v and the ai are the constants of Lemma 9-2.1 for [ZQ, po), then /(ô, p)^2:îX'^
I 1
h (^0, P) = f{zo, po) +Iaj
{XI -
XI).
Remark 1. li N = v -j- \, we have seen that these conditions (except for (iii)) are necessary and sufficient for the parametric integrand to be quasi-convex in p. For iV > r + 1, it is not known that the conditions above are necessary for quasi-convexity. Remark 2. We note t h a t i f / m e r e l y satisfies (9-1-3) then (9.2.12) f[z{x),p{x)] dx = (plz(x)Mv(x)] ' dA [z) for any C^ mapping z = z{x), Uv being the oriented tangent r-plane. This fact makes it easier to define I{z, G) where G is a C^ ^-manifold. We must now introduce the notion of the order of a point in v-space with respect to an oriented [v \)-manifold. We shall be concerned principally with such manifolds of the topological type oidG for some domain G in Rv of class C^. However, the generalization to other orient able manifolds will be evident. We allow dG to consist of several disjoint manifolds, each oriented as a part of ^ G in the usual way. liv = 2 and dG is just a Jordan curve, then this order is just the winding number of the point with respect to the curve. Definition 9.2.1. Suppose G is of class C^ and ^ ^ Ci(G), 99 = (99I, . . ., qf) and let r '. z"" = (p'(x), oc = 1, . . ., r. Suppose z fr{dG). Then
357
we define the order 0 [z, r {d G)] b y (9.2.13) 0[z,r(dG)]= f ^

6G
Ko,o.[<p(x) - z]'{-
\r^^dx^^
where KQ is the elementary function for LAPLACE'S equation, as defined in (2.4.2), and
(0 2 id) ^ y g ^ a(yi,...,y-i/r+i,...,<y,r)
Lemma 9.2.2: Suppose r : z^ = (p'^(x) is of class C^ on the open set Q in Rv. Let W he the set of x for which dcpjdx = 0. Then \r(W)\ =0. x^), ii \x \r{x,xo)\ xo\<,d. ê'.\x-xo\, Proof. Suppose XQ $ W. Then we may write T: ^"^ - ^g = [c| + e^(x, xo)] (x^ ^0 = (p'îô)> c^ = (p!^(ô), \e(x,xo)\<e Thus (9 2 15) ^''-4 = <Û^^ - 4 ) + ^ ^ ( ^ , ^ o ) , ii \x xo\ ^d.
The image of B (XQ, r) under the Hnear part of the mapping is an elHpsoid s in some plane of lower dimension. From (9.2.15), it follows that the total image of JB (:ô, ^) C (s, er) ii r < (5. Since the derivative of the 9 * 9^ are uniformaly bounded on any compact subset of Q, the result follows easily. Lemma 9.2.3. (a) 0[z,r{dG)] depends only on the values of the qf on d G and the values of their directional derivatives along d G. Moreover, it is independent of the parametric representation of dG. It is as slimed in this lemma that q)^ C^ (G) and that G is of class C^.
(b) 0 [z, T{dG)] is continuous in z if z i T ( 5 G ) .
(c) If(p(x) (9.2.16)
^ z for any x^ G(i.e. G), then 0[^,T(aG)] = 0 .

K K
(d) / / the closed regions Gi, . . ., GR <^^^ disjoint and C G, all deing of class C^, and if F = G U G^ and z ir(dG) (9.2.17) 0[z,r{dG)] =0[z,r{dr)]+ U r(dGjc), then
2;.0[z,r(dGjc)].
(e) If z ir{dG), (9.2.18)
then 0 [z, T(dG)] is an integer and (almost everywhere),
N [z, r(G)] > 10 [^, r(dG)] \
iV [z, T (G)] being the number of x in G such that (p (x) = z. (f) / / 99 {x, t) ^ C^ (G) in X and is continuous on G x [0, 1] and if {z i xt (dG) for any t on [0,1], then 0 [z, Tt(dG)] is constant with respect to t, xt : z = (p{x, t).
358
(g) The functions N[z, r(G)] and 0 [z, r{dG)'] ^ Li{Rv) and (9.2.19) J0[z,x{dG)]dx=
Rv
f^dx,
G Rv
fN[z,r(G)]dz=f\^\dx.
G
Proof, (a), (b), and (d) are evident. To prove (c), we note t h a t
(9.2.20) since the first term on the right vanishes b y Lemma 4.4-6. To prove (e), we begin by letting W he the set of A; in G for which dcpjdx = 0. Since the (p'^ can be extended, it follows from Lemma 9.2.2
that | T ( P F ) | ^ | T ( ^ G ) | = 0 and t(W) and r{dG) are compact. Now if
z^r{G) - [T(W) U r(dG)] then N[z,r(R)] is finite {R = G(0) - W) and d(pldx 9^ 0 at each x for which (p {x) = z. In fact, if ^ is a small domain of class Ci, its counter image consists oiK = N[z,r{R)'\, small domain Q Gi,, . .,GK and the restriction of 99 to Gjc is a diffeomorphism of Gjc onto Q for each k. If ^ is interior to Q, all of its counter images are in the Gjc, one in each, so t h a t if F is the domain in (d), then (9.2.21) 0 [ ^ , T ( a r ) = 0, 0[z,r(dG)]
k=l
=2;o[z,r{dGjc)].
To evaluate one of the terms on the right in (9.2.21) we begin by replacing KQ in (9.2.20) (G = Gjc) by Koa where (9.2.22) Koa {y) = fKo (y - rj) tp^ [rj) drj,
Bi6,a)
^ * (rj) = g-' ip [q-^ 7]),
where ^ is a moUifier. Thus, for o" > 0 b u t small,

JKOO,.
QGH
[cp (x)-z](-\
r ^ 1 ^ dx'p = f rpt [(p [X)

GJC
~z\^dx
= ^g^ ( S ) / ^ * (^ - ^) ^c = s^^ff
Q
The result (e) follows. Moreover, we evidently have

K
fo[z.r{dG)]dz^^
(9.2.2?)
K
f^dx
^
JN[z,r{G)]dz^
fc = l Tic
j\^^\dx.
359
The results in (g) follow from (9.2.23) by covering almost all of r(G) [r(W) U T{dG)] by disjoint small regions Q. The result (f) is now obvious. Definition 9.2.2. If G is of class O , q)^C^{G), r \ z^" = (p^^ [x) for ^ = 1, . , ., 1^, and z i r[dG), we define 0 [z, T{dG)] = lim 0 [z, rn(dG)], r^ : z- = cpl{x)
for any sequence {^w} CîdG) such that (pn converges uniformly to (p on dG. Lemma 9.2.3'. If G is of class C^ and (p^C^ {G), then all the results of Lemma 9.2.3 hold, except possibly those in (g); (p[x, t) is merely required to ^CO(G X [0,1)]m(f). Lemma 9.2.4. If q) ^C\ (G), G arbitrary, there exist sets Za, ( = 1, X . . ., r, each of \-dimensional measure 0 such that if R is a cell in G, none of whose faces lie along planes x' = c^" where C^-^ZOL, then (9.2.24) Ifv /^^^=/o[..r(ai?)]^..
= 2, this holds ifcp^C^ (G) O HI (G). Proof. We first assume (p ^C\ (G). We begin by extending cp to be LiPSCHiTZ everywhere and then, by using moUifiers we may approximate to (p on G uniformly by vectors cpn whose gradients V <pn converge almost everywhere and boundedly to V 95. Clearly this convergence holds almost everywhere on each plane x"- = c' for c* not in a certain set Zoc of measure zero. For cells R as described, we have
and moreover, if z is not on T(<9 R), we see also that 0 [z, r{dR)]'is given by (9.2.13). Since 99 is Lipschitz, it follows that \r{dR)\ = 0. Thus (9.2.13) holds for almost all 2: and this is also true for each ;^. Thus, for almost all z, 0[z,rn{dR)] -^0[z,r{dR)] and (9.2.26)
e
f\0 [z, rn(dR)]\dz
< (^-^^
-M^-Â'-îdR),
(e measurable),
M being the common Lipschitz constant for the 99^. Thus (9.2.27) lim fo[z,rn(dR)]dz= f 0[z,r(dR)]dz
and the first result follows. In case v = 2 and 99^ C^(G) Pi H\[G), we may approximate to 99 uniformly and strongly in H\ on any domain -D C C G by 99^ ^ C^ {D) so that (pn -^^ on dR strongly in H\{dR) for all i? on a grating as indicated. Clearly (9.2.25) holds. For any such R, cp and 99^ are A.C. and
360
the lengths l[Tn{dR)] and l[r[dR)] are uniformly bounded so t h a t |T^(c?i?)| = |T(a7?)| = 0. Then, d^gîn 0\z,rn{dR)] ^0[z,r(dR)] for almost all z and (9.2.26') j I 0 [z, xn (dR)] i dz < (j-^Y" I [xn [dR)].
e
Thus the result follows. Remark. The only thing that prevents us from allowing q? ^Hl {G) for any v is that we do not know that |T(î^) | = 0 even if 99^ Hl{dR) and is the strong limit in HI {dR) of cpn in C^. This is on account of the example
of BESICOVITCH.
We can now prove MCSHANE'S principal lemma: Lemma 9.2.5. Suppose (p^C\ (G), R is a cell for which (9.2.24) holds, <Pn C^ (D) fo^ some D with R(Z D C C. G, and (pn converges uniformly to (p on R. Then there are measurable subsets VnC R such that
hm
f^dx=f^dx.
liv = 2, the same result holds licp^ C^{G) H Hl{G). Proof. Let n = max \q)n(^) (fi^) \ for x g dR. For each n, let Wn be the subset of R where dcpnidx = 0. Cover almost all of the set of z where 0[z,r(dR)] ^ 0, z irn(Wn), and 2: is at a distance > En from T (dR) by small disjoint domains Qni, for each of which, r^^ (Qni) = U Gntk
k
and Tn \ Gnijc is Si diffeomorphism from GniJc onto Qni- Then (9.2.28) 2' ( ~^dx
^^
GriiJc
= ^
i -^
Qni
(0[z,r[dR)]dz
i,k ^ -'
since 0[z,rn{dR)] = 0[z,r(dR)] for 2:^ U Q!^^. Since the right side of (9.2.28) tends to r 0 [z, r(dR)] dz, we may take Vn = U Gnik to obtain the result. Theorem 9.2.1. Suppose f satisfies the general conditions, G is of class Ci, z = {z'^, . . ., z^)^ î(^)> ^n^ ^-^{Q* ^^^ Zn-^z uniformly on G. Then (9.2.29) //(z, G) < lim inf //[zn, G).
J^-c>o
If V = 2, the result holds if z merely ^ C^{G) 0 H\[G). Proof. We shall prove the first statement; the proof of the second is similar. If the right side of (9.2.29) is + 00, the result follows. Otherwise, by taking a subsequence if necessary, we may assume that If{zn, G) < w M so that (9.2.30) A(zn,G) = j\Xn[x)\dx<.M, #=1,2,...
9-2. A lower semi-continuity theorem
361
From the representation (9.2.12), the differentiabihty condition o n / , and the compactness of the totaHty of n.o. sets in RN, it follows t h a t there exists a number L and a function co(d), depending only o n / a n d A, such that \a\ = \2:ajW^<L, \f(zi,p)-f(z2,p)\ôj(d)\X\,
(9.2.31)
_lL_^
^k^B{0,A)>
pi ^2|<(3,
lim(o(d)
= 0
whenever aj is the set of constants of Lemma 9-2.1 f o r / a t some point (-2^0,iô) where z^s^Bi^, A) and the rank of ô is v, A being a common bound for all the Zn {pc). Next we notice that there are sets Za of measure 0 such that if no face of the cell R lies along a plane x'^ == c'^Zx, then (9.2.24) holds for every set Z^, / = ^'i < < iv. For almost all XQ Urn \R\-^jf[z[x),p{x)] (9.2.32) . lim\R\-'^l\X(x)
R^XQ
dx
=f[z(xo),p{xo)] = 0,
-X{xo)\dx
Let > 0. Then we may cover almost all of G by a countable number of cells Rj, avoiding the Za, having maximum edge < twice minimum edge, so small that (9.2.33) and so that I (9.2.34) 2' f \X{x)-X{xj)\dx<el4L. We may choose a finite number of these Rj so that (9.2.34) Z ff[^{^j)>PM']
? R^
\z(x) - z(xj) I <co(d) < /4M, x,
xj^Rj
f \f[^{^)>P{^)]
-f[zM>P(xj)]\dx<el4
^^ > If{^> Q -
2/4.
Now, if for some j , the rank of p (Xj) < v, then (9.2.35) lim inf f f[Zn {x), pn (^)] dx '^ = j f[z[xi),p{xj)]dx
RJ
>() {X{xj)=0).
Otherwise, if rjn = max | Zn (x)

Rj Rj
z(x)\, dx Co{f]n) A {Zn, Rj)
ff[Zn(x),pn{x)]dx^ff[z{x),pn{x)] (9.2.36) ^lfi,^xi),p{x)]
dx - [co{ri) + e/4M] A (z. Rj).
362
Now, for each fixed j for which X{xj) ^ 0, let ai be the constants of Lemma 9.2.1 and let f i , . . ., C* he the linear functions of z determined * as in that lemma. Then, from our general assumptions and Lemma 9.2.5, it follows t h a t there are sets Vn C Rj such that I f[^(^j)>Pn(x)] dx > If[z(xj),pn{x)'\ dx
(9.2.37)
> / 2 ^^^^W ^^ = / ^ ^ ^ ^ - ^ / ^ ^^^^

= / ^
Rj Vn
I
Vn
Sj
ai X^ {x) dx>
-
Rj
f[z (xj), p (xj)] dx
-L(\X[X)
X{xj)\dx.
From (9.2.30) and (9.3-33)(9.2.37), we obtain lim inf //{zn, G) > lim inf ^ f f[zn (^), pn (^)] dx
^2
ffl^M^Pi^j)] ^^ - 2/4 >//(^, G) - e.

^ Rj
The theorem follows. 9.3. Two dimensional problems; introduction; the conformal mapping of surfaces The simplest possible multidimensional parametric problem is t h e problem of Plateau, i.e. the problem of proving t h e existence (and differentiability) of a surface of least area having a given boundary. This was solved in the one contour case b y J . DOUGLAS and T. RADO [3] in 193031. The next decade saw the solution (see MORREY [8] for
more literature of this period) b y DOUGLAS, R . COURANT, MCSHANE,
and others of this problem among surfaces of higher topological type bounded b y one or more contours. The solution was greatly facilitated b y the use of the theory of conformal mapping which enabled one t o replace the area integral by the DIRICHLET integral; as we have seen in (9.1.13) and (9.1.14) we see that (9.3.I) A{z,G)<'-D(z,G),
the equality holding if and only if z is a conformal representation of the given surface. Thus the minimizing vectors are harmonic. This fact
enabled M. MORSE, ([4]) C. B. TOMPKINS (MORSE and TOMPKINS [1] [4]), COURANT ([1]), M. SHIFFMAN, and others (see MORREY [8]) to apply t h e
MORSE critical point theory to obtain interesting results concerning the
9.3- Two dimensional problems; conformal mapping
363
existence of ''unstable'' minimal surfaces. Finally the author (MORREY [8]) generalized COURANT'S solution in the case of k contours, the surface being of the type of a domain bounded by k circles, to the case of surfaces lying in a RIEMANNIAN manifold. The conformal mapping theorem still holds for such surfaces but the work of BOCHNER ([2]) on ''harmonic'' vectors in RIEMANNIAN spaces was not sufficient to enable the author to carry over COURANT'S work. It was necessary to carry over the theory of functions of class H\ to vectors in a RIEMANNIAN manifold in order to obtain the "in the large" solution of the minimum problem. By assuming that the manifold 501 (of class C^ at least) is "homogeneously regular", i.e. that there exist positive numbers m and M, independent of PQ, such t h a t each point PQ of ^ is in the range of a coordinate patch with domain ^ (0, 1) for which w | f | 2 <^g^^{x)ii^3' < M | | | 2 , x^B(0,i), f arbitrary, the author could show that the minimizing surface is continuous. If Wl is of class Cîgij^ C^), the differentiability results for v 2, N arbitrary, as stated in 1.10 imply t h a t the solution harmonic vectors ^ Q for any //, 0 < /^ < 1; if 501 is of class CJJ, 0 < ^ < 1, for some n > 3, then the harmonic vectors ^ C^; ii Wl is C or analytic, so are the harmonic vectors. This work is presented in 9-4 below. In 9.2, we extended MCSHANE'S proof ([3]) of lower semicontinuity for more general parametric integrals. Shortly after the war, CESARI (W> [2], [3]) extended MCSHANE'S theorem to his more general integrals (we do not discuss these) and also showed that the convexity condition on F (z, X) was- a necessary, as well as a sufficient condition for lowersemicontinuity. But it was not until the concurrent results of CESARI [4], DANSKIN and SIGALOV [2] t h a t an existence theorem for the integral / / of 9.2, with r = 2 and iV == 3> was proved. The difficulty was that the lower semi-continuity theorem requires uniform convergence (at least on interior domains) and even though conformal mapping kept the DIRICHLET integral bounded, this is not enough to ensure equicontinuity. But the implied condition (9.3-2) m\X\ <f{z,p)~F{z,X) <M|X| implies an a priori HOLDER condition, on interior domains, for the solution vector. The author ([17]) was able to give a short proof of this theorem by minimizing a dominating integral / for ^f among all z for which another integral / {z, G) < K, obtaining bounds independent of K, and then letting X -> oo. We present this in 9.5The author is not aware of any general differentiability theorems for the solutions which have been shown to exist. However, if the ratio Mjm above is sufficiently small and if (9.3.3) \X\Fx7x6{z.X)^y^'>mi[\t?-\XmX'm
364
where mi is sufficiently large, then the dominating function/o defined b y (9.3.4) [/o (z, p)f = [/(^. p)]^ + ( ^ ) ' + F^
{E, F, G defined in (9.1.12)) is convex in all the pi{(x == 1,2). To be sure, /o does not ^ C^ for any p for which X ^ 0 b u t the second derivatives are bounded and the analysis of 1.11 and 5.2 5-5 progresses far enough to show that the solution vector z^Cj^ for any ju, 0 < ^ < 1. Since the solution vector is obviously conformal, the higher differentiability follows in the usual way near points x where not all the derivatives z\ {x) are zero; at such points /o ^ C^ and so it is not likely, in general, t h a t derivatives of order > 1 will be continuous a t such points. For the case of minimal surfaces, we shall see t h a t they m a y have branch point singularities. For the general parametric integral with v = 2, the author has constructed a dominating function/**, coinciding w i t h / w h e n E = G arid i^ =: 0, which is regular only if the general sense (see 9-5 below). However, b y making certain a priori assumptions about the solution
(i.e. t h a t it has the form z =f(x,y)^ C^, etc.) J E N K I N S and SERRIN
have obtained interesting results concerning these solutions. Moreover, in 1957, a student T. C. K I P P S , in his thesis, derived some a priori bounds for the HOLDER continuity of the first derivatives of a conformal representation of a solution surface, assuming that these derivatives were continuous on an interior domain and assuming that the problem was regular. We present his results in 9.5- Strangly enough, he was not able to use these results to demonstrate the continuity of the derivatives, in spite of t h e fact that his bounds do not depend on the moduli of continuity of the derivatives. We shall now sketch preliminary results, including LTCHTENSTEIN'S conformal mapping theorem for surfaces, leading to the solution of the problem of Plateau which will be presented in 9.4- We shall confine ourselves here to surfaces of type k, i.e. those having the topological type of a plane domain bounded b y k disjoint JORDAN curves; surfaces of higher topological types are treated in the book [3] by COURANT. We shall not assume any theorems on conformal mapping; these will be proved; we shall, however, use the MOBIUS (linear, fractional) transformations. We shall assume, also, if G is a domain of type k and class CJ, there is a diffeomorphism of class CJ of G onto a circular domain B, i.e. one bounded b y k disjount circles, in such a way that the outer boundaries correspond. Lemma 9.3.1. Suppose {zn} is a sequence of vectors such that each Zn^ C]^(Bn), where Bn is a circular domain of type k. We suppose that Cni = dB{0,\) is the outer boundary of B^ for each n and ^2, , Cnk are its other boundaries, that Fi, . . .,r^ are oriented FRECHET curves
9.3- Two dimensional problems; conformal mapping
365
whose ranges are disjoint JORDAN curves, and that the FRECHET curves Fni, each of which is defined by Zn restricted to Cni, converge respectively to Fi as ^ -> oo. We suppose also that there is an L (<i ^ oa) such that (9.3.5) D(zn,Bn)<L, n=\,2,... Finally, we suppose that the Cni converge to Ci, where each Ci is a circle or a point, and that C\, . . ., Ci, I <,k, are circles and the Ci with I <. i ^ k {if any) are points. Then the Ci, . . ., Ci are disjoint. Proof. If this were not so, two of the Hmiting circles Ci and Q ( 1 0) be the minimum distance from a point on any Fp to one on a different Fq and choose i? > 0 but small enough so that every circle dB{xo, r) with 0 < r < i^ intersects Cf and Cj. Now, choose d, 0 <C d <C R. Then, there is an N such t h a t if ^ > AT then every circle dB{xo, r) with (5 < r < i^ contains an arc yn r which is in Bn and has its endpoints on two distinct Cnp and Cnq (''usually" Cî and Cnj, but there may be circles Cns with / < s < ^ close to XQ). Thus for such r f\ Zn^[r, ^) |2 d^ > (2JT)-1 \ j \ Zn& \ d&
SO t h a t 2 d^
fr-^f\zn4r,
S
ynr
&) |2 d&dr > ^ l o g ( J ) ,
n>N,
Since d is arbitrarily small, we arrive at a contradiction. As a result of this lemma, we easily conclude the following corollary: Corollary 1. Suppose that the sequences {zn}, {Bn], {Cni}, {J^ni}, Ci, and Fi satisfy the conditions of Lemma 9.3-1- Then it may also be assumed that Cni = Cifor each i <,l and, moreover, this assumption may be made without altering the values of Urn inf D{zn, Bn). Now suppose t h a t all these conditions are satisfied. Since the Fi are Jordan curves and the Frechet curves Fni ->-A for each i, it is easy to see t h a t there exist continuous vector functions Cni((p) which are periodic of period 2jr, are not constant on any segment, and which converge uniformly to f^ (99) where ^ = f* (99), 0 < 99 < 27r is a topological representation of Fi, except that CHO) = C'(27r) (Theorem 9.1.2(b)). It follows easily t h a t there are non-decreasing continuous functions (pni such that ^^^^, (9.3.0) (i) Cm[(fniiO)] = ini{0) = Zn [xi + ri COS0, yi -f- ri sinO], Ci = dB{xi,yi,ri), i=\,.,,J; (ii) (pni{^) 6 is periodic of period 2jr; (iii) 7t <(pni{0) <Jr. We may immediately conclude the following corollary:
366
Corollary 2. Suppose the sequences [zî^, etc., of Corollary 1 satisfy the conditions of that corollary and that the ^ni> d, Sni, <^^<^ ^ni ^<^^^ ^^^^ chosen as above. Then a subsequence, still called {n}, may be chosen so that the functions (pni{Q) converge to non-decreasing limit functions cpi{d). Remark. These limit functions need not be continuous but we prove the following lemma: Lemma 9.3.2. Suppose the sequences [zî], {q)ni], etc., satisfy the conditions above. Then, for a given j , 1 <.j^l, and do, either cpj is continuous at do or else (9.37) (pj(do + ) - î(ô - ) = 271:. Proof. Suppose for some i and ô, we have Then dlfpiiOo)] and Ci[(pi(Oo)] are distinct points; let
\Ci[<Pi{e^)]-Ci[cpi{eo)] = d>o,
Let po be the point of Ci corresponding to do- From the uniform convergence of Cni{(p) to Ci((p) and the pointwise convergence of (pni(0) to (fi (9), all being monotone, it follows that H a ^ > 0 and an N such t h a t Ci divides the disc B {po, S), B {po, d) intersects no other Q w i t h ; < I, and if y{r) is the arc of dB{po, r) which contains points of Bn then (9.3.8) Zn [7 {r)] is of length > i / 2 , 0<r <d. It is clear that (9.3-8) certainly holds even if that arc contains points of Cni with ^' > Z if we interpret y{r) = yn{^) as the part of that arc in BnThus, for each e > 0 (9.3.9) L>jr-^j\zn{f)fdy>>^\oî^-y
e
n>N.
Yn (r)
This is impossible. We wish now to prove the following theorem of LICHTENSTEIN [2]: Theorem 9.3.1. Any locally regular^ surfaces of type k and of class Cj^ possesses a representation z which is conformal and of class Cj^ (B) where B is a domain of type k bounded by circles. Three given points on one of the boundary curves of S can be made to correspond, respectively, to three given points on the outer boundary of B which may be taken as dB{0,\). In case the surface is of class C^ for n'>1, C^, or analytic near a point, the representation z has the corresponding class near the corresponding point. If the point is on the boundary of S and d S has the same class near that point, then z has the same class near the corresponding point on dB. Preparatory remarks. As was pointed out by COURANT in [3], this theorem is not necessary if one wishes only to prove the existence of a
1 See (9.3.10) et seq. below.
9.3. Two dimensional problems; conformal mapping
367
minimal surface (in the sense of differentiaLgeometry) with given boundary. But the theorem is interesting in itself and is helpful in proving the existence of a surface of least area having a given boundary which surface is simultaneously a minimal surface. We shall prove this for the general k; LiCHTENSTEiN [2] proved it for ^ = 1. Suppose that (9.3.10) S:z = z{u,v), (u,v)^G, ZuXz^Ô, where G is of type k and of class CJ and z^ Cl{G). Using inversions and simple transformations, we may assume that the outer boundary of G corresponds to that bounding curve of 5 containing the three given points. We first prove several lemmas. Lemma 9.3.3. There exists a mapping (9.3-11) r:u = u{x,y)y v v{Xyy), [x,y)^B from a circular domain B of type k onto G in which u and v ^ H\ (B) n C^{B), and which minimizes D{Z,B)=
(9.3.12)
n[\Z^Y^
B
\Zy\^)dxdy,
among all such mappings which carry the three given {distinct) points on the outer boundary Ci = dB{0,\) of B into the three given points on the outer boundary / \ of G, where dG = FiU . . - v Tjc. Proof. First of all, it is clear that there is a diffeomorphism from some circular domain B' with outer boundary Ci onto G. A map of the desired type is then obtained by preceding this with a Mobius transformation of B (0,1) into itself to establish the three point correspondence; such a transformation carries circular domains into circular domains. We note t h a t D(Z, B) = J J[E(u,
B
v) [ul + îl) + 2F[ux v^ + % Vy) + vl)-\dxdy F ^ Zu ' Z't,, G = z^ ' Z'i^ + Gfji\^M{X'^+ fjil), 0<m<M.
(9.3.13)
+G[vl
E {u, v) = Zu ' Zu,
where E, F, and G^ G^^(G) and there are numbers m and M such t h a t w(A2 + / / 2 ) < A 2 + iFXfjii (9.3.14)
To show the existence of the minimum, let {tn} be a minimizing sequence. From (9.3.II)(9-3.14), it follows that {r^} satisfies the conditions of Lemma 9.3-1 and so, by choosing a subsequence, we may assume that our minimizing sequence satisfies all the conditions of Corollary 2. From the three point condition and Lemma 9-3-2, it follows t h a t the restriction of r^ to Ci converges uniformly to a representation of / \
368
which satisfies the three point condition. There are now three types of degeneracy which can occur with respect to the boundary conditions: case i: ^ > 1, 1 < / < ^, one or more Q with i ^ I are points interior to the Hmiting domain B bounded by Ci, . . ., C^; case ii: some such Ci is a point on some Q with j <,l; case iii: some cpi with ^' < / satisfies (9.3-7). We first consider case i. Let i^ > 0 be chosen small enough so t h a t B{po,R) C B and contains no other Cp than those which reduce to d = pQ. Let 0 < ^ < i^. Then there is an N so large that all the Cwp which are approaching ô lie inside 5 ( ^ 0 , <5). Thus, for each r,d <,r <,R, Tn\_dB[pQ,r)'] encloses all the corresponding F^ and so has length > some ^ > 0 so that
R
(9.3.15)
D{Zn, Bn) >mjr-'^{d^l2n)
dr = {m d'^jln) log{Rjd)
which is impossible since d is arbitrary. The reasoning to eliminate case ii is the same. So we may assume that I = k and all the Bn = B. We now suppose that case iii holds. Let po be the point of Q having coordinate ô- Choose jR > 0 so that B {po, r) H B is bounded by an arc y' of Q and an arc y" of dB(po, r) for each r d. djl, {y U y'" = Q) d<r<R, n>N. But, from Lemma 9.3-2 and the convergence of (pnt to cpi, it follows that diamT^j(y"') < ^ / 2 so l[rn(y")]'> As in (9.3.9) we arrive at a contradiction. Thus for each i, the (pi are continuous so the convergence of Zn is uniform on each Ci to a representation of Fi. Finally, we note that there is an i? > 0, independent of p, such t h a t ii p^B, then either B (p, r) G B or B (P, r) O B is bounded by an arc 7' of some one Ci and an arc y" of dB(P, r) whenever 0 <ir <. R] in the former case, we set y" = dB(P, r). Now, let > 0 and choose ^1 > 0 so small that diam Tn(7') < e/2 whenever r <C di; this is possible on account of the equicontinuity. Now, suppose for some d, 0 <C d <. di, t h a t osc Tn on B{p,d) f) B > s. Then, for ^ < r < di, the length of Tniy') > /2 so that (see (9-3-15)) (9.3.1.6) D(Zn, Bn) > mJr-îe^lSn)
dr = {m e^jSn) log{dild).
The equicontinuity of the Tn now follows from (9.3-16) and a subsequence converges uniformly to a minimizing map r.
9.3 Two dimensional problems; conformal mapping
369
Our aim now is to discuss the minimizing map r. Since a similarity transformation or, more generally, a MoBius transformation in the (x, y) plane carries one circular domain into another and leaves the Dirichlet integral invariant, we see t h a t our r minimizes D (Z, B) among a larger class of mappings from circular domains in which the outer boundary need not be dB{0,'[) and three points do not need to be fixed. So, suppose t h a t
Â '^' = l(^, y;^), y' = r](x,y;X),

(9.3.17) i{x,y;0)=x, fA(^%y;0) =v{x,y), r](x,y;0)=y, rj2,{x,y;0)
m <Ao,
=a){x,y)
is a family of diffeomorphisms, of class C^ in (x, y, A), of 5 into a circular domain B;^ for each A. Then v and co must satisfy the following boundary conditions on each Cj, j = \, . . ., k:
{V, co) = {n, COi) + (V2, CO2) + {V3, CO3) ,
(9.3.18)
vi cosO + coi sind = 0,
(v2, 0)2) = (oj, dj) = const. {cj + ifj) = const.
vs + iws = (cj + ifj) (cos0 + i sinO),

Cj : X = Xj + Tj COS0,
y = yj -\- TJ s i n 0 .
Moreover, if we set cpW=D{Z',B,), Z'[x\y')=Z[^'{x\y'-X), rj'[x\ y' K)], al^ :x = ^'{x\y'',X), y = rj'{x\y'; X), then cp' (0) = 0 since 99 (A) takes on its minimum for A = 0, since r (Tj"^ is an admissible map of B;^ onto G. By expressing D (Z\ B^ as an integral over B, we obtain (9 20) ^''^ ^ ^~^ ^^"^ ^^ ~ ^^ ^""^' J = ^x'r]y ^yrjx^^' ^ ^"M - ^ ^ ^^2/ + ^y f J > .
^(A) = / / [ g ( l | + t^,^) - 2 g ( f ^ | ^ + ^^7^^) + (5^(f| + rjl)]J-idxdy (9.3.21) ^ e=|Zr.|^ g-Z^-Z^,,
(l^ = |Z2/|2
, g , and ^ being independent of A and in Li{B), Differentiating with respect to A and setting A (and 1^/ ^.nd ?^^) = 0, we obtain (9.3.22) (p'{0) = f f [U(vx coy) V{vy + Wx)] dxdy -^ i = 0,
t/ = g - ^ ,
F=-2g.
Lemma 9.3.4. Suppose v and co ^ C'^(B) and satisfy the boundary conditions (9.3-18). Then there exists a family a^ of maps satisfying the conditions in and following (9.3-17). Proof. By using a partition of unity of class C^, we may write any such (v, co) as a sum of one which ^ C^ (B) and three for each Cj with
370
y > 1, each of which vanishes except near Q . Corresponding famihes ai in (9.3'17) can be found simply b y setting ^(%^ y] X) = X + Xv{x, y), rj(x, y; 2) = y + Xco(x, y) for the {v, co) ^ C^ (B) or of the types {v2, C02) and (vs, C03) for each Q . The remaining ones are handled b y polar coordinates and these latter each carry B into itself. From this lemma and the duscussion preceding it, we obtain the following corollary: Corollary. / / ( , ^ and @, as defined w (9.3-21), correspond via (9.3-13) to our minimizing map r, then (p' (0) = Ofor all {v, co) ^ C ^ (B) and satisfying (9.} AS). Lemma 9.3.5. If the conclusion of the corollary holds, then (9.3.23) u=v=ô. Remark. The essential idea of the last part of this proof is due to H. L E W Y ; the proof given here follows in a general way the argument given in COURANT'S book [3], pp. 1691/8. Proof. Suppose, first, that v and o) ^^{B), let -y; be a moUifier, and let VQ and OJQ be the ^-mollified functions. Then, defining JJ = V =^ v =z CO = 0 outside B and keeping Q small, we obtain / / [U(vex ojQy) V(vey + ajex)] dxdy
B
^ If II ^? (^"~ ^) ^ (^' ^) (^^"" ^ ^ ~ ^('^' ^) (^^ + ^^^^ ^^ ^y ^^ ^^ ^ ^"

= / / [UQ{VX B
o)y) Ve{vy + ojx)]
dxdy
= fJ[oj{UQy+
B
VQO^) -v(Ue^
Vgy)]dxdy,
(9.3.24) Thus if this = 0 for all such (v, co), we see that UQ + i VQ is analytic for each ^ > 0 and hence we have {9.3.25) 0{z) =: U + iV analytic on B. Now, since U and V^ Li(B), we see t h a t (9.3.26) lim f[U(vdy + codx) V{cody~vdx)] =0
for any (v, co) ^ C^(B) and satisfying (9-3.18) (BQ is the set of (x, y) in B such that B{x,y;Q) C B). Setting (9.3.27) V + ia) + P{r) [ci + i di + (ei + ifi) e^^], 1- A< r < 1 where P{r) = i near r = i and P{r) = 0 near r = i h, v + ico being 0 elsewhere, (9.3.26) for all a, di, ^ i , / i , impHes t h a t (9.3.28) j0(z)dz = Iz0{z)dz = O, 0 < Q< Qo^
371
By choosing the corresponding variations for each Q, we obtain ^^'^^^^ ^^^ ^^-^ 0<^<ô, CjQ being the circle in B concentric with and at a distance Q from Q. From (9.3-29), we conclude that there is an analytic function ^(z) on B such that
W(Z) =0(Z).
Since dW(z)ldx = 0(z) and dW(z)ldy = i0(z), it follows that W^H\{B) and so has strong boundary values in Li(dB). Thus W is continuous on B. Now, let us choose (9.3.30) V = -P(r)A(e) sine, co = P(r) A(6) cos0 where P(r) was defined above (in (9.3.27)). Then (9.3.26) yields (9.3.31) -lim fA(e)Im{W[(\ - Q) e^^]e^^^}dO = 0.
Integrating by parts once, we may then let ^ ->0 in (9.3-31) to obtain (9.3.32) ImjW'{e^^)[ie^Â'{e) - e^Â(e)]de = 0.
If we break this into two integrals, change the variable to 9 in the second 9 integral, write A((p) = A{0)-\- f A'(6) dO, and then interchange the order of integrations, (9.3.32) becomes
f 271
r 2n 0
<p
271
I ml (A [6) i 6'^ W {e^^) - fw'(e^^) e^^d(p dO + iA(0) f-^ W(e^^) dcp lo L e J o (9.3-33) in which the coefficient of A (0) vanishes. Since A' is an arbitrary real function ^ C^ except that (9.3.34) fA'{e)de
0
= o,
it follows from the Lemma of Du Bois RAYMOND (Lemma 2.3.1), that the imaginary part of the quantity in the bracket in (9-3-33) is a constant. Carr^dng out the integration, we see that (9.3.35) Im[iz W\z) i Wiz)] = const, on Ci, so that z W{z) W(z) = ^{z) where ^ is analytic across Ci. Thus we see that W^ C along C\ (since ^îs continuous). Differentiating (9.3.35) with respect to d we obtain (9.3.36) Imliz^JizW -iW]\ = ~-Imz^0{z) = 0. 24*
372
By repeating the argument for each Q , we obtain (9.3.37) Im[z Zj)'^0[z)=::O, z^Cj, j=2,...,k. Accordingly we see that 0 can be extended analytically accross each boundary circle of B. Now, on Ci, let f(6) = z^0(z). Then f (6) is real and it follows from (9.3-28) that
271
2JI 2TC
(9.3.38)
lf{e)de
= lz0(z)'zdO=
-ifz0(z)dz
= O,
Since / is analytic and periodic, it follows t h a t / , and hence 0 has at least two zeros on Ci. By using (9.3.29), we conclude also that 0 has at least two zeros on each Q . Now, either 0{z) ^ 0 on B or else it has a finite and non-negative number of zeros in B and no poles there. The number n of zeros of 0 interior to B is given by
-l^lim
(9.3.39) "'^^B'g
fZ^dz. Hz)
B'^ =
B-U[BnB(Cs,Q)].
the Cs being the zeros of 0ondB and Q being suitably small. But the value of the integral in (9.3.39) is easily seen to yield n <, k since there are at least two Cs- on each Cj,j=\,...,k. Thus we must have 0(z) ^ U + i V -^Ô and the lemma is proved. Lemma 9.3.6. Suppose z = z(w), w = {u, v) ^ G, is a diffeomorphism of class C^, G being of class C^ (z = (z^, . . ., z^)). Suppose S is the [unoriented) Frechet surface having the representation z and suppose z = Z{X), X = {%, y) ^ B, is a representation of S in which Z^C^ (B) n HI {B) where B^ C^. Then ^ a unique mapping r : w = w{X), X^ B, such that (9.3-40) Z{X)=z[w{X)-\, X^B, T ^ C^{B) n Hl(B), and 3 ^ sequence r^ : w = Wn(X) of diffeomorphisms from B onto G which converge uniformly to r. Proof. Let \S\ denote the range of S, i. e. the ranges of z and Z. Since ^ is 1 1, z~^ is a homoemorphism from 151 onto G. Since Z is a representation of 5, H a sequence {xn} as above such that z [xn {X)] ^ Zn {X) converges uniformly to Z(X). If we define r{X) = z~'^[Z(X)], we see t h a t (9.3.40) holds and also that rn(X)=Z-^[Zn{X)] so t h a t Tn converges uniformly t o r on B. Clearly r is uniquely determined by (9.3.40). To show that r^H\[B), let XQ^B, let wo = r{Xo), and ZQ = z{wo) = Z(Xo). It is clear that z(u^, u^) can be extended to a
373
diffeomorphism C('^^> ^^) from a neighborhood in Rjs^ of (u], u^, > 0, . . ., 0) where WQ == (ul, u^) onto a neighborhood of 2-0 in i?^. Then T(Z) = c - i [ Z ( Z ) ] for X in -B near XQ. That r ^ H\{B) follows from Theorem 3-1.9. Lemma 9.3.7. Suppose r \w = w{X) satisfies the conclusions of Lemma 9.3-6. Then the Jacohian Ux Vy Uy Vx has the same sign almost everywhere in B. Proof. From Lemma 9-2.4, it follows t h a t there are sets Ji and J2, of 1-dimensional measure 0 such t h a t if none of the sides of the cell R lie along lines x"- = C^ where c'^^Ja, oc = 1, 2, then (9.2.24) holds. From Lemma 9.2.5, it follows that if i^ is a cell for which (9.2.24) holds, there exists a sequence of measurable sets VnG R such that
Vn
It is clear, since we may confine ourselves to a subsequence that we may assume that all the Jacobians Unx'^ny '^ny'^nx have the same sign throughout B. The result then follows using (9.3.4I). We can now prove the conformal mapping Theorem 9.3-1. Proof of Theorem 9.3.1. We let r be the mapping of Lemma 9.3.3From Lemma 9.3-5, it follows that ^ + 2 ^ 5 = 0 ==: {ux + i % ) ^ + 2F {ux + i Uy) x X {vx + i Vy) + G{vx + i VyP; here (, g , and @ are defined in (9-3-21) and E, F, G in (9-3-13)- It follows from (9-3-42) that
Vx= G-1 (F Ux H Uy) y Vy = G-1 ( i t HUx F Uy),
=^]/EG-F^.
From Lemma 9-3.6 and 9-3.7, it follows that the ^^ sign may be replaced by k, where k = ^ i and is constant on B. Thus (9.3-43) rr. . ..^ (9.3.44) Vx =G-^(Fux % "= k(bux + kHuy), + cvx), Vy = G-^(kHUx + bvx), FUy), or Vy = k(aux
Since T is continuous, the coefficients in (9.3.43) are continuous and (9-3-43) implies (since r ^ ^ K ^ ) also) that
flG~^[Cx{kHux-Fuy)+Cy(Fux
B
+ kHuy)]dxdy
= 0,
^^C\{B),
(9-3.45) From the theorem of 5.5, it follows first that T ^ Hl(E>) for eachp > 2 and each D C CB. This implies that T $ C|J,(Z)), that the coefficients
374
^ Cl(D), and hence that u and v^ CJ(D) for D C C B. The higher different!abihty results on the interior follow as usual. Let {xo,yo)^dB and (uo,vo)=T{xo,yo). A neighborhood '^ of (^0, ô) may be mapped by a diffeomorphism of class C^: U = p{u,v), V = q(u,v), u = P{U,V), v==Q{U,V)
onto a disc B(0,0,R) so (^0,^0) corresponds to (0,0) and ^ddG corresponds to U = 0. Boundary coordinates {t, 6), where t = logr and (r, 6) are polar coordinates with pole at the center of the circle oi dB containing {XQ, JQ), may be chosen and the equations (9-3-43), (9-3-44), and (9.3.45) go over into equations of the same types. Since f/ = 0 when t = const., we conclude from the theory of 5-5, first t h a t U and hence V^ every H\[D), U and F ^ every C^ the coefficients $ C^, and hence U srndV^ C],[D) for D a neighborhood of {x^, yo) on B. Thus T $ C]^{B), Again the higher differentiability results follow as usual. 9.4. The problem of Plateau We now resume our discussion of the problem of PLATEAU. Let F denote the system {Fi, . - ., Fjc), each Fi being an oriented closed FRECHET curve with range 17^.^ |. Definition 9.4.1. We define / ( r ) = inf{liminfL(^^,G^)} d{F)= inf {hm inf D (%, Gn)} for all sequences {zn} where Zn^ C^(Gn), each Gn is of type k and of class C^ with boundary curves Cî, . . ., CnJc, each oriented as a part of dGn, and Zn restricted to Cnt is a representation of Fni where Fni ->/\-, i =^ \, . . ., k; in the case of d(F) we require the Zn also ^ H\{Gn), We also define / * ( r ) - ^ * ( r ) = + 00 if ^ = 1 ^^'^'^^ l*(F) - m i n 2 ' ^ ( r ( * ) ) , d^{F) = mm^d{Fi))
for all possible systems /'(^), . . ., 7^(2?) where each F^'^^ consists of the curves Fj for w h i c h y $ Ti where [J Ti = {\, 2, . . ., k}.
i=\
Lemma 9.4.1. (a) In the definitions of l[F), d{F), l'^(F), and d*{F) we may assume that each G^ is of class C\ and that each Zn ^ CJ [Gn) and is locally regular. [h) d[F) = 2l[F), d*(F) = 2l*(F). (c) d(F) <d*{F). Proof. All the statements in (a) are evident from the definitions except possibly the one concerning local regularity. To see this, let
9.4. The problem of PLATEAU
375
{^n}^ CjîGn). We may define Zn by X = {x^,x^) ^Gn
where Xn is chosen sufficiently small, (b) follows from (a) and Theorem 9.3.1. If ^ = 1, (c) is evident. If ^ > 1, let T == T^) U B^K By reordering the Fi, we may assume Td) = {A, . . ., Fm}, ^(2) = {Fm+i, . ., / \ } . Let ^ 1 and B2 be fixed circular regions of types m and k m, respectively and let z\n^ Gj^(Bi) and Z2n^ Q ( ^ 2 ) for each n and suppose that
L(Zin,Bi) -^/(ra)), L{Z2n,B2)->l{r(^)).
Choose :vi, ;^2, ^ 1 and R2 so t h a t B(xi, '}Ri) C Bi, i = 1, 2. By altering 2:^^ only in B(xi, 3^^), we may obtain z'^^^ G1{Bi) but so t h a t zl^(x) = -2:2^(A;^) for x^ B(xi, 2Ri), i = \, 2. It is clear t h a t we may assume jî = :^2 = 0 and Ri = R2 = R. Next, let ZQ^ R^- We may alter z^^ on ^ ( 0 , 2R) only to obtain 2'^'^, where z^'^ is constant on each circle dB {0, r) with R <r <2R, zl'^{x) = ZQ for x^B{0,R) and ^^^^ C^ (5^), ^ ' = 1 , 2 (of course the local regularity is sacrificed). Now, define B = [Bi B{0, R)] U (r[B2 ~ B{0, R)] where a is the inversion with respect to dB{0, R) followed by a reflection in the x'^ axis. On B, we define I4;[CT-M^)]. Then Zn ^ Q (E) and we have xâ[E2-B(0.R)].
L (z, B)=^L ( 4 ; , Bi) + L ( 4 ; , B2) = L {zin. Bi) + L {z^n, B2). Thus / ( r ) < Hm inf L (z, B) < /(Pd)) + /(r(2)). By induction, we may extend this to p systems for any p <^'k and thus show that I (F) < /* (i^) from which the result follows. Lemma 9.4.2. Suppose lI,^B.^\dB{x^yR)'\ and H is the harmonic junction on B{xo, R) coinciding with C on dB(xo, R). Then
D[H,B(xo,R)]<J\C'{e)\^dd.
0
Proof. This follows by expanding in Fourier series. Theorem 9.4.1. Suppose F = {F\, . . ., F]^, ^ > \, is a set of oriented FRECHET curves whose ranges \Fi\ are disjoint Jordan curves. Suppose also that d {F) < d* {F) {d {F) finite if i * {F) = + oo). Then there exists a vector z^ C^(B), B being a circular region of type k^ where z is harmonic and conformal on B and z restricted to each properly
376
oriented boundary circle Ci of B is a representation of Fi, Moreover z is a representation of a surface of least area', i.e. L(z, B) = l{r). Proof. Let {zn} be a sequence ^ L{zn, G^) -> l{r). We may first assume each Gn^ CJ and each Zn^ Gl{Gn) and locally regular. On account of Theorem 9.3-1> we may assume that each Zn^ C]^[Bn), each Bn being a circular region, in which we have (9.4.?) D{Zn,Bn)^d{r). Finally we may assume t h a t the Zn and the Bn satisfy all the conditions in Lemma 9.3-1 and its two corollaries as well as (9.4.3)- Then there are three possible types of degeneracy as in the proof of Lemma 9-3-3- We shall show t h a t each of the cases considered there implies that d{r) = d*(r) and is therefore impossible. Let M = sup D{zn, Bn). Suppose case i holds and let XQ = \Ci\. Choose 6 > 0 but so small that B (XQ, (5^/2) (2 B and contains no | Cj \ other than those for which | Cj \ = XQ. For n'> N, all the \Cnj\ with | Q | = XQ are inside B(xo, d). For such n, since M > /
6
lr-^\zne{r,0)\^drde,
0
it follows that there is an fn, d ^fn
^ d^l^ such that
f\zne(fn,6)\^d6<2MI\logd\
and Zn(rn, 0) is A.C. in 6. Let Bni = BnU B{xo, fn), B'^^ be the domain outside the Cnj interior to B[x{i,rn), and Bn2 be (yn[B'^<^, where On is the inversion in dCnt followed by the reflection in the x'^ axis. Let Hni be the harmonic function in B{XQ, rn) coinciding on dB(xo, rn) with Zn and let Hn2 be the transform of Hni under the inversion in dB [XQ, rn). Define
{x) , X^Bn B{XQ,fn)
i[x),
x^B(xo,rn)
, ,,
Mx)
x^B'^^nBi
^)> x^B'^^
^^Bn2-
B(xo,fn)
Zn2 {x) = < 2 [O'n ^ (^)] >
Then we see from Lemma 9-4.2 that

D {Zni, Bnl)+D {Zn2, Bn^) < D [Zn, Bn) + 4 M / | l o g ^ | .
Since d can be arbitrarily small, it follows that d*{r) <,d(r). Suppose case ii holds and let XQ = Ci, XQ^ CJ, and let ZQ be the image of XQ under the limiting continuous map of Q onto Fj defined by the function 0j(d) of Lemma 9.3-2 and equation (9-3.6) (recall that \ <j ^ l < i). We may choose a sequence {dn} in which each dn> 0 and
9.4. The problem of PLAUTEAU
377
^n - > 0 , and a sequence fn with dn ^^n ^ ^n^ such that dB(xo, d]l^) intersects only Q and B{xQ,d]l'^) contains only those Cqn for which Cq = XQ, all of these being inside B {XQ, dn) and such t h a t (9.4.4) / I Zne {fn, e) |2 de < 2M/I log^^ |.
Bnf]SB {xo,rn)
Since the Zn converge uniformly along Q we see t h a t mRx\zn{x) ZQ\ ioT x^yn = [Cj n B(xo, r^)] U [dB{xo, fn) Pi 5 ] < eî, (9.4.5) lim^ = 0. Let Rn-^0 with enIRn - > 0 and define the mapping
(9.4.6) (J0n(z) = ZQ + 0n{\z Zo\) ' \z ^ o | - ^ ' {z - ZQ)
where 0n^ C^{Ri)> 0n is non-decreasing, and 0^ (s) = 0 for s < ^ and 0{s) = s for s ^ Rn, s-i 0n{s) and 0;(5) < RnliRn - 2sn) for all s. If we then define (9.4.8) z'^(x) =con[zn(x)] we see t h a t (9.4-3) still holds and z'^[x) = ZQ in an open set D yn- Consequently, for each n, there exists a simple closed curve Cjc+i, n of class CJ which is in this open set, is in B, and encloses the Cpn for which Cp = XQ. Let us define Bî = BnU U CpnU U Bpn where Bpn is the interior of Cpn, and let Bn2 he the exterior of the Cpn and define Zni(x)
(Q.4.Q)
\z^[x), \zo
x^ Bni and outside Cjc+i^n , X on and inside C]c,i nX inside Q + i n and outside and on the Cpn,
, oc outside and on CA:+I,%.
z^{x),
^0
Then, again
(9.4.10) D{Znl, Bnl) + D{Zn2, Bn2) = D {z^ Bn)
and we see t h a t ^* [F] = d(r). Finally, we suppose case iii holds; we assume t h a t 1 <,j <.l and that the function 0j of Lemma 9.3.2 satisfies (9.3.7). L e t XQ be t h e point on Q corresponding to do and let ZQ = 0j (6) for 0o < 6 < ô + + 2jr. Then, as above, we may choose dn and fn so t h a t (9.4-4) holds, dB(xo,dj/^) intersects Q only, and the Cnp for which \Cp\ = XQ are all inside B{xo,dn). This time (9.4.5) holds for all x on yn = [C; B (xo, fn)] U [dB(xo, Yn) Pi B]. We again choose Rn and define o)n{z) by (9.4.6) and (9.4.7) and ^ b y (9.4-8) and see t h a t (9.4.3) still holds and z^ [x) = ZQ on an open set D yn- Consequently for each n, there exists a JORDAN curve Cjcjî^n of class CJ in this open set and in Bn which encloses Q and the Cpn for which Cp = XQ (if any). Then we m a y
378
define Bni BÛ
Q U U CpnU
(their interiors) and 3^2
as the
exterior of Q and Cpn and can define Zni and Zn2 as in (9.4-9) and find again that ^ * ( r ) = ^ ( r ) . Thus none of the degeneracies occur, En = B for every n, and the Zn converge uniformly on dB. The result follows by replacing each Zn by the harmonic function with the same boundary values and using the lower-semicontinuity. The conformality follows from Lemma 9-4.1. As was mentioned above, the author solved the PLATEAU problem, in the ^-contour case, for surfaces situated in a "homogeneously regular" Riemannian manifold (MORREY [8]); these were defined in the first paragraph of 9-3. We now present a simplified proof of this result. We shall assume t h a t this manifold W is of class CJ,, 0 < /^ < 1, at least and that W is connected. I t follows t h a t M is separable. We shall consider vector functions z from various sets having as values points
of m.
Notation. If zi and Z2 are two points in ^R, we let 12:1 ^21 be the geodesic distance between them, i.e. the inf of the lengths of all paths in W joining them. Definition 9.4.2. A vector function z defined on a measure space {E, (j) is said to be measurable-JJL iff for every open set D C 391, the subset of X in E where z[x)^ D is measurable-//. Lemma 9.4.3. If zi and Z2 are ju-measurable on E, then \zi(x) Z2{x) \ is measurable on E. Proof. If z is measurable-// on E and 5 is a BOREL subset of 9K, the subset of X where z(x)^ S is measurable. For each n, divide 9)1 into at most a countable number of disjoint BOREL sets Snj, each of diameter < n~^; it is easy to see how to do this using coordinate patches. In each Snj, choose a point Znj. Let Enjk be the subset of x in E for which Zi (x) ^ Snj and Z2{x)^Snk- For each n, the Enjjc are disjoint and measurable and E = U Enjki'^ fixed). Define 0n(x) = \znj Znk\ ior x^ Enkj- Then each 0n is measurable-// on E and 0n (^) -> I ^1 (^) ^2 (^) | for each xÊ, Definition 9.4.3. A function z (into 9)1) defined on a measure space (,//) is said to be of class Lp[E,[ji) <^-^ is measurable-// on E and \z{x) ZQ\^ Lp{E, fj) (p > 1) for each ZQ in 3K. The following lemma is elementary and we omit its proof. Lemma 9.4.4. / / zi and Z2 ^ Lp(E, //) then \zi{x) Z2(x) \ ^ Lp{E, ju). If we identify equivalent functions and define the distance Ip (^1, Z2) = lf\
liv zi (x) Z2 (x) \P d/A
the resulting space is a complete metric space (^ > 1)
379
Lemma 9.4.5. Suppose z is of class Lp{E) for some measurable set E in Rv, r > 1. Then there exists a function z defined almost everywhere on E and coinciding with z almost everywhere on E such that lim \m (^)]~^ / k (^) ^ (ô) \dx = Q for every XQ where z {XQ) is defined and every regular family of sets e G E about XQ. Proof. Let {zn} be a dense subset of M. Since \z{x) Zn\ ^ Li(E) for each ^, 3 a set Z with m{Z) = Oîi XQÊ Z, then lim{2h)-^ (9.4.11) f \z(x) Zn\dx = \z{xo) Zn\ " ^ ^'^'"'^ lini[m(e n E)lm(e)] = 1
for every n and every regular family of sets about XQ. It follows easily t h a t if Xo^ E Z, then (9.4.11) holds with Zn replaced by any ZQ. Definition 9.4.4. A vector z ^ Cl{G) (G C Rv, 0 < ^ < / / ) <^ for each XQ in G, the components z^ with respect to a coordinate patch having range D z (XQ) are of class CJ in some neighborhood of XQ on G, A vector z is absolutely continuous (A.C.) on [a, b] < for each > 0 , J ad ^ 0 = such that (9.4.12) 2\^K') - ^K) I^ ^ whenever 2^{^7 - ^d ^ ^>
and the intervals (xp x^) are disjoint. We say t h a t z is of class Hl(G) ^ z is continuous on G, ^ is A.C. in each x"" for almost all values of the other variables, and the partial derivatives \zôc\ (i.e. the derivatives of the arc-length along x'^ = const, with respect to x'^) ^ Lp(G). Lemma 9.4.6. z is A. C. on [a, b] {of class Hl(G)) ^ for each XQ on [a, b] (XQ on G), the components z^ with respect to a coordinate patch with range D z(xo) are A.C. {of class Hi) in some neighborhood of XQ on [a, b] {G). If z is of class Hl{G) the quantities (9.4.13) Go.p(x)=gij[z{x)]z%(x)zi^(x), \z,4x)\ = l/Ga(^)
^ Lpi2 {G) and are independent of the local (z)-coordinates used. If z is of class Hl{G) and x = x(y) is a \ - \ J^'-LIPSCHITZ map of F onto G, then w(y) = z [x (y)] is of class HI [F) and the derivatives transform in the ususal way. Notation. For vectors defined on domains in R2, we use the notations E, F, and G instead of Gu, G12, and G22, or perhaps, (9.4.14) \z^\^ = E, Zcc'Zy = F, \zy\^ = G. Definition 9.4.5. We carry over the definitions of D{zi, Z2), oriented and non-oriented FRECHET variety, parametric representation, etc. as
380
given in Definitions 9.1.1, 9.1.2, and 9.1.3- For a vector z^ C'i(G), G of class C^, we define (9.4.15) A(z,G) ^liyWG'^^^J^dxdy,
G
D{z, G) = fj
G
(E + G)
dxdy,
E, F, and G being defined locally by (9.4-13) and (9.4.14). Then, for z^ CO(G), we define L (z, G) as in (9.1.11) and L (S) as in Definition 9.1.5. Lemma 9.4.7. The theorems of 9.1 generalize in the obvious way to vectors z and surfaces S with ranges in W. We interpret the space H^{G) n C^{G) to he our new space H\ [G). Lemmas 9.3.19-3.7 cmd their corollaries and Theorem 9.3-1 < ^ hold for vectors z with ranges in 3Jl. ^^ Definition 9.4.6. We define /(F), d{r), Z*(r), and i * ( r ) as in Definition 9.4.1; in the definitions of d[r) and d*[r)y we require that the Zn^H\[Gn). Lemma 9.4.8. (a) Lemma 9.4.1 holds. (b) Instead of Lemma 9.4.2, the following is true: Suppose ^ is A.C. on dB(xo, R) with
271
(9-4.16)
f\C'{d)\^de<mln,
Then H a vector H of class Hi [B {XQ, R) J which coincides with f on dB [xo, R) for which
2.71
(9.4.17)
D[H,B[xQ,R)]<.K']\C{Q)\^dQ,
K =
Mjm.
(c) Instead of Theorem 9.4-1, the following is true: Suppose that F satisfies the hypotheses of Theorem 9-4.1. Then H a circular region B of type k and a sequence {2'^} of vectors of class H\ (B) such that D {zn, B) ->d(F) and the restrictions of Zn to the properly oriented Ci converge uniformly along Ci to continuous representatives of Fi, i = i, . . ., k. Proof, (a) is evident except for the local regularity. But here again we may embed W in the manifold Tl X R2, the metric in any local coordinate system being given by ds^ = ZgiA^^'
i, ?=1
. - -, ^^) dzi dz^ + m [(^^^+1)2 + (^^^+2)2], m = (w + M ) / 2 ,
the relation between any two coordinate systems being 'zi=f{z^,.. .,z^), i=\,...,N, 'zi==z\ i = N + \,N + 2. the /* being defined as for W. Clearly W is homogeneously regular with the same numbers m and M. (b) Extend f to be periodic in 6 and choose ô- There is a di with \6i - 0o| < 7 r such that |C(l9) - C{6o)\ < \C(6i) - C(0o)\ for all (9.
381
From (9.4.16), it follows t h a t (9.4.18) |C(0i)-C(6>o)|2<jr/|r(<9)P^(9<w.
Now, since W is homogeneously regular, we may choose a special coordinate patch in which f (do) is the origin and (9.4.19) ni\i\^<.gij(z)î& ^M\^\^, z^B(0,\).
From (9.4.19), it follows that for any path z^ = z^{s) (s arc length),
I
I
/ ]/gij [^(s)] 4 4 ^^ ^ y ^ /
0 0
\zs\ds>:l]fm
so that the range of this patch contains all points within a distance ]/w of C(ô) and hence all the C{6). So, if we define H as that vector on 5 (XQ, R) whose components H^ with respect to this patch are the harmonic functions coinciding with C^(6) on dB(xo, R), we see, using Lemma 9.4.2, that D[H,B{xo,R)]=f jgij[H{x,y)]{HiHi
271
HiHi)dxdy
2n
^ ^ 2 j
* B(x,B)
l\VHi\^dxdy<Mj^\Ci\^dO^Kf\Ce\^de.
0 * 0
(c) The proof of Theorem 9-4.1 carries over right down to the last paragraph. The mappings co^ are defined as follows: Since Wl is of class C^, there exists a coordinate patch (Z) in which ZQ corresponds to the origin in the Z-space and gij (0) = dij. Then, for n sufficiently large we define con as the identity outside the range of that patch and define it as in (9.4.6) and (9. 4.7) in terms of the Z coordinates inside the patch. In order to complete the proof of Theorem 9.4.1 for surfaces in W, the author found it expedient to introduce a theory of functions of class HI (G) with values in Wl. Since ^ is not a linear space, some of the theorems of chapter 3 do not make sense and new proofs have to be found for most of the remaining theorems. A useful tool is the well-known device, used extensively by E. SILVERMAN, of mapping W isometrically into the space m of all bounded sequences. Although this space is linear, it is not separable and so this device cannot be used to prove all theorems. Remark. Many years ago, H. W H I T N E Y [1] proved t h a t any manifold W of class C^ could be mapped by a diffeomorphism 0 of class C^ onto an analytic manifold W in some Rp (actually P may be taken 2N + 1). If 5K is a Riemannian manifold and this mapping could be done in such a way that there were numbers mi and Mi with 0 < Wi < Mi such t h a t the geodesic distances along W and W satisfied mi\p-q\ <\0{p) -0{q)\ <Mi\p-q\,
382
it would then follow that a vector z(x) would be A.C. on [a, b] <= X(x) = 0[z(x)] is A.C. on [a, b] and we would have mi\z'(x)\ <\X'{x)\ <.Mi'\z'(x)\. In this case, it would be sufficient to define z to be of class HI (G) <= 0 (z) ^ HI (G); then the whole theory of these functions would be available. Actually, J . NASH ([1], [2]) has proved that 0 can be taken to be isometric (i.e. w i = Ml = 1) if P is sufficiently large. B u t all these facts require proof and the last-mentioned theorem is somewhat difficult. Therefore, in order to keep the exposition relatively self-contained, we present an alternate development of the theory of HI spaces using SILVERMAN'S idea and terminating with the DIRICHLET growth Lemma 9.4.18. Definition 9.4.7. A vector function X(x) (= {X^(x), X^(x), . . .}) into m is A.C. <^ (9-4.12) holds with 2: replaced by X ; here \X(^) X((x)\ = sup \X^^) ~ Xôc)\. X^Lp(G), GC Rv, ^ each X^ is measurable on G and \X{x)\= sup \XHx) \ ^ Lp(G).
i
Lemma 9.4.9. Suppose X is A. C. on m and 0{x) is the length of the arc X = X[t), a <t <, X. Then 0 is A.C. on [a, b] and (9.4.20) 0' {x) = sup I Xi {x) I a.e.
i
If the X^ are A.C. on [a, h] and s\xp\Xl{x) \ ^ L\{\a, 6]), then X is A.C. on [a, b]. Proof, (a) That 0 is A.C. is immediate. Then if a <, x h-^\Xi(x + h) -Xi{x)\, ^'-1,2,... from which it follows that 0' (x) > sup|Xj. (x) |. On the other hand if the
i X^ are A.C. on [a, b] and sup |XJ,(A;) | ^ Li([a, b]), then i
\X^^)
-Xôc)\
< f \Xi(x)\
a
dx < f sup \Xi{x)\

a *
dx
from which the absolute continuity follows easily. Definition 9.4.8. A vector function z from G d R into Wl is of class Hl(G) ^ (i) z^ Lp(G), (ii) z is equivalent to a function ZQ which is A.C. in each variable x"^ for almost all values of the others, and (iii) the partial derivatives \zo,oc(x)\^ Lp(G). A vector X from G into m is of class Hl{G) <^ each X^^HKG) and if \X(x)\ - s u p | X ^ ( : ^ ) | and \X,4x)\ = sup I XJa (^) I ^2? (^) Let {zn} be a countable dense subset oiW; we
i i
z^HliG),
define the mapping r from W into m by r{z) = {\z zi\, \z Z2\,. . .}, Lemma 9.4.10. (a) If z ^ HI (G), then z^Hl (D) for each D C G. (b) / / z^Hl{D) for each DCCG and z and \zo,oc\^ Lp(G), then
9-4. The problem of PLATEAU
383
(c) The mapping r is isometric, (d) A vector z into Tl is A.C. ^ the vector X = r z into m is A.C. A vector z into m^Hl{G) ^ X = r z^ HI(G). (e) If z^Hl(G), X = x(y) is a hi-Lipschitz map from F onto G, and w{y) = z [x (y)], then w^Hl (F). (f) / / z^H\[G), then z is equivalent to a function z which has the absolute continuity properties of the ZQ in definition 9-4.8 and, in addition, if X =^ X [y] is a hi-Lipschitz map ofF onto G and we define w[y) = z [x(y)'], then w also possesses those absolute continuity properties. In fact z may be chosen as the function of Lemma 9.4.5Proof, (c) Clearly \\z' - Zn\ -\z" - Zn\ = \Xn{z') X'^{z")\ ^\z' z"\ for every n. But, by taking Zj^ near z' and zi near z", we see t h a t \X{z') - X{z")\ = sup|X*'(;2') - X^(z")\. (a) and (b) are evident.
(d) follows from (c) and Theorems 3.1.2 (g) and Lemma 3-1-1. To prove (e), it is sufficient to use (d) and then Theorem 3.1.7 for each component X^. If we let Y^{y) = X* [:v(y)], we have ! y,(y) I = sup I .Y> \x[y)-\ xl{y) \ < M2'{sup | X\^ [.^(y)] |},
M = sup I x^^ (y) I.

Clearly each I Y a | ^2)(-/"). Part (f) follows from (e) and Theorem 3-1-8 for each X^. The last statement in (f) must be proved by returning to the vector z in W and repeating part of the proof of Theorem 3.1.8. It is exceedingly important to obtain the usual formulas for the derivatives in the change of variable theorem. We prove what is necessary in the next lemma. Lemma 9.4.11. Let z(x,y) ^Hl(G), let z(x,y) be the equivalent function of Lemma 9.4.10(f), let E,F, and G be defined by (9.4.13) and{9AAA) in terms of the partial derivatives of the z^, let a : x = x{s,t), y = y[s, t) be a bi-Lipschitz map of F onto G, and letw{s, t) z [x{s, t), y(s, t)]. Then \ws{s, t) |2 = E{x, y) xl + 2F Xs, ys + G yf \wt{s, t)\^ = E(x, y) xf + iFxtyt + G y^ (a.e.), \ws\'^ and \wt\'^ being defined similarly in terms of the partial derivatives of the w^. Proof. Since the expressions for E, F, G, \ws\^, and \wt\^ are independent of the coordinate systems (z) on W, we confine ourselves to the range 9^ of an admissible coordinate patch. Let 5 and 5 ' be the respective sets of {x, y) and (s, t) where z(x, y) and w{s,t)^'^; clearly S = a(S'). If z(xo, y) and z{x, yo) are A . C , and (XQ, yo) $ S, then H an oc (XQ, yo) 5 {xo, y) and {x, yo) $ 5 for all x and y 5 | ^ :ô | < ^ and \y ~ yo\ <i(x^' Let e > 0. There exists a compact subset ^ of 5 such that all the functions z^, 4 , and z^ are defined and continuous on ^ , w^, w],
384
'^t> ^s, ys, ^t, and yt are all continuous o n ^ ' = (y~^(2^), ni(S 2^) <Ce, ^^^ z^ (x + h,y) z^ (x, y) = p j {x, y) + s{ (x, y, h)] h z^ {x, 3/ + A) ^* [x, y) = [z^ {x, y) + 4 (^> y^ ^)] ^ whenever the points involved are in ^ ; we also have
\4{^'y*^)\>
\4{^>y'^^)\ < e W > Hmsih) = o,
{x,y)^2!'
Now, further, if (XQ, yo) is not in a subset Z of ^ of measure 0, the linear metric density at (XQ, yo) of the part of ^ on the lines x = XQ and y = yo is 1 and the linear metric density at (SQ, ^0) = or-i (XQ, yo) of the part of ^' on the lines s = SQ and t = to is 1. Consequently the projection of a [^' (to)] on at least one of the lines x xo oi y = yo has metric density 1 at {XQ, yo); here ^ ' (ô) denotes the part of ^' on the line t = to. If that on y = yo has t h a t property, the intersection of ^ ( y o ) and the projection ^2/0 ['^{2' (^0)}] still has metric density unity at {xo, yo) and so the subset A of ^'(^0) such that Py^laiA)] lies in the intersection above still has metric density 1 at (SQ, ^0). The usual relations ^l(5o, ^0) = zi(xo, yo) ^s(so, to) + 4 ( ^ 0 , yo) ys{so, to) follow by letting s ->so on zJ. The formula for Wt and hence those in the theorem follow similarly. The lemma follows since e is arbitrary. Definition 9.4.9. A domain G C i^*' is said to be of class D' <^ it is the union of a finite number of diffeomorphic images of r-simplices which are joined together as are the simplices in a simplicial manifold. Lemma 9.4.12. Suppose z^H\[G), G being of class D\ Then the function z of Lemmas 9A-^ oind 9-4.10(/) is defined almost everywhere on dG and ^ Lp [dG). If G is the cell [a, h], then (9.4.21) lim (\z{x'',x^) z{a'^,x^)\r>dx^ = 0, oc=\,...,v.
If ôoci ^01, ni(Zo:) = 0, then the line x^ = x^^ intersects dG in a finite number of points, at each of which d G has a tangent plane which does not contain the line, and z{x'^, XQ^) is A.C. in x^ on each closed segment of the line in G. Proof. In order to prove the first two statements, it is sufficient to prove them for a cell [a, b], on account of Lemmas 9.4.10(f) and (9.4.11). In this case, choose oc and extend z across the face x"^ a^ by reflection obtaining ZQ with the usual A.C. properties. Then ^0 is defined almost everywhere on x^" = a"^ and fo(<^', â) = lim^(:\;', x^) a.e. Then (9.4-21) follows from equation (3.4.10) which reduces in our case to (9.4.22) j\z{x^,x^)
ai,
- ^ ( a ^ O 2 ? ^ A ; ; < | A ; - a|2>-i| j \z,o.[x)\v

ace cf'
dx.
385
The first part of the last statement (about d G) is well known (see Lemma 9.2.2). The rest follows, since if x^^ is not in Za, a neighborhood % of a point XQ = (XQ, %oa) on ^G can be mapped by a diffeomorphism of the form
yi = < .
y''=--/(^^0.
<K,â]
onto a cell in such a way that dG O'Sl corresponds to a plane y^" const. The following substitution lemma is an immediate consequence of Lemma 9.4.12: Lemma 9.4.13. Suppose G and A are of class D' with A C G, suppose z^Hl{G), w^Hl{A), w{x) =z(x) a.e. on dA, and Z[x) ==w{x) on A and Z{x) = z{x) on G A. Then Z^ Hl(G) and Z{x) z{x) a.e. on dA U dG. Lemma 9.4.14. Suppose {z\ is a family of functions in Hl(G), G Lipschitz, such that Dp (z, G) is uniformly bounded. Then (i) if f \z{x) ZQ\^ dx is uniformly hounded, so is ( \z{x) - ZQ\^ dS)
0 do
(ii) if J \z{x) zo\^ dx or f \z(x) zo\^ dS

X a
is uniformly
bounded
for some cell r dG or some open set G on dG, then (\z[x) zo\^ dx is uniformly bounded. ^ Proof. Since G is the union of a finite number of bi-Lipschitz images of cells, it is sufficient to prove this when G and a are cells. Then (i) follows immediately from (9.4-22). Moreover, if the integral over a (part of a face of dG) is uniformly bounded, (9.4-22) shows that the integral over a cell r adjacent to a is also. If T = [y, d], we see by applying (9.4.22) with a = 1 t h a t the integral over ri : a' <, x' <. b', x[^ [y[, d[] is uniformly bounded. The result (ii) follows by applying this procedure for a = 2, } , . . .,v in turn. Definition 9.4.9. We say that Zn tends weakly to z{zn -^ z) in Hl(G) ^ each Zn and z^ tI\{G), Dp{zn, G) is uniformly bounded and Zn-^z in Lp [G); here we assume G of class C\. In case p =^ 1, we assume also t h a t the set functions Di {z^, e) are uniformly absolutely continuous. Lemma 9.4.15. / / G is of class C\ and Zn 7 z in Hl{G), then Zn-^ z in Lp(dG). Moreover Dp (z, G) < lim inf Dp{zn, G). Proof. I t is clearly sufficient to prove this for G a cell. To prove the convergence in Lp of Znia^", x') to z{a', x^, we let \r) be a subsequence of {yi) and let > 0. We may choose a subsequence {z^ such t h a t Zs{^'^y â) ->^(^*, â) ill ^v\.^'oi^ â] ^^"^ almost all x"^. From (9.4-22) and the weak convergence (and the uniform absolute continuity in the case ^ = 1), it follows that we may choose an x^ such t h a t Zs(x', x'^) ->z(x^, x^ in Lp and the integrals on the left in (9-4-22) (with z ^ z ox
386
Zs and x'^ = %") are all < (e/3)^. I t follows easily that { \\zs{a'',x'^)-z{a'',x'^)\vdx\
<e,
s>S(e).
Thus Zsia"^, x^)->z(a', x^) and, since {r} was any subsequence, it follows that Znia"", < ) ->5(a, < ) To prove the lower semicontinuity, choose a subsequence {zr} so ^p(>2^w, Q - ^ l i m i n f Z)j,(^^, G), and so that ^r(^^ ôa)-^^(^'^> ôJ in
n
Lp [a"^, b^"] for almost all J^Q^^, ^ = 1, . . ., r. In case

J\zr,oc(x'',XQ)\^dx' aoc
is uniformly bounded for some subsequence, the convergence of Zr to z along the line is uniform, so that (9.4.23) / I ^,a(^^ A;;J l ^ ^:^^ < liminf / Ifr,(^^ ^ ^ J |^ ^^^; ^
of course this holds if the right side = + 00. The result follows by integration with respect to XQ^ and summing on oc. Lemma 9.4.16. Suppose Zn Hl(G) for each n,p > \, G is of class CJ, and Dp(zn, G) and f f \^n{Xy y) ô|^ dx dy are uniformly
G
hounded.
Then a subsequence converges weakly on G to some z in H\{G). Proof. I t is sufficient to prove this for the cell G: a'<x<,b,c<,y < d. Let Unic be the set of y such that Zn[x, y) is A.C. and
h
(9.4.24)
/ [\Zn[^. a
y) -
Zo\P +
\-Znx[x> y) | ^ ] dx <
k.
Then m{Unk) '> {d c) Mjk, M being the obvious bound. Let

N=l n=N fc=l
Then Ujc consists of all y such that (9.4.24) holds for infinitely many n and U consists of all y for which (9.4.24) holds for infinitely many n and some k. Clearly m(U) = d c and if y ^ U, then the Zn(x, y) are equicontinuous with |Zn{x, y) zo\ uniformly bounded (for some subsequence). Thus we m a y find a subsequence zis converging uniformly on [a, b] to Ci(^) for y = yi, yi being in the middle third of [c, d]. B y repeating the whole argument on the subsequence î^, we can find a second subsequence Z2 s converging uniformly to C2 {x) ony = y2 in the middle third of [c, 3^1]. B y continuing this and then taking the diagonal sequence, we obtain a sequence Zg such that Zs (x, yq) converges uniformly to some tq[x) for each yq, yq being dense in [c, d].
387
We now show that {zs} is a Cauchy sequence. Let e > 0. From (9.4.22), it follows that there is a finite subset T {s) of the yq such that if c < y 5(),
q^T(e).
Combining (9.4.25) and (9.4.26) for s and t, we find t h a t

b
J \zs(x, y) Zt{x, y)\P dx < eP,

a
c <y
^d
from which the result follows. Lemma 9.4.17. Suppose z ^ H\[B{pQ, a)], po = {xo,yo), with r D[z,B{po,r)] ^k^(r), j Q - ^ k{Q) dq = K(r) < o o , 0 < y
<a.
Then Z(XQ, yo) is defined and if w(r, 6) = z(xo + r cosO, yo + ^ sin0), then w is A.C. in r on [0, a] for almost all 6 and
1
\z[x, y) f (A;O, yo)| < ^ / / [ ( 1 t) X(i + t x, (\ t) y^ + t y] dt,
/(x, y) = [E [x, y) + G [x, y)]i/2,
r^ = [x - xo)^ + {y - yo)^
for almost all [x, y). Proof. From the change of variable theorem, we conclude that if we let L (r, 6) = l{xo + r cos0, yo -\- r sin6), then L(r,d) Iêfine >0, L^{r,e) == \wr{r,e)\^ ^ 2. + r-^ \we{r,d)\^.
h(r)=ffQy^L(Q,e)dgdO,
Then
0 0
h(r) <(2jr)i/2yi/2y^(r)
I f L{Q,e) dgdO = f Q-^l^ h'{g) dg

0 0
r
^ r-il2h(r)
+ ~ f Q-^l^h{o) dg < [2n)^l^ [k{r) + K[r)] -> 0
as r - > 0 , Thus, since L{r, 6) > lwr(?', 0)|, r~^ [weir, d)\, it follows t h a t w is A.C. in y on [0, a] for almost all 0 with w(0,6) in Li and there exists a sequence r^ - ^ 0 5z^6i(^w, ^) - ^ 0 in Li so that w(0,d) is a constant. The result follows easily. 25*
388
Lemma 9.4.18 (DIRICHLET growth lemma). Theorem 3.5.2 holds in the case _^ = r = 2 and u replaced hy z. Sketch of proof. For if {xi,yi) and {x2,y'^ are any points in B{xo,y{i\ R), z{xi,yi) and z{x2,y2) are both defined and the result of Lemma 9.4.17 may be used for (XQ, yo) = {xi, yi) and (X2, ^2) in turn and almost all (x,y). Then the proof of Theorem 3.5.2 carries over with \S7 u\ replaced by L. We can now prove our version of the DIRICHLET problem: Theorem 9.4.2. Suppose -2:* ^ Hl{G) and G is of class CJ. Then there exists a z^H\{G) such that ^ = f* a.e. on dG and minimizes D(Z, G) among all such vectors. If 9K is of class C^, then z ^ Q , {D) for every D CLdG and every /^', 0 < //' < 1. / / 9J^ is of class C^ for some ^ > 3, then z^ C^{D) for such D. IfW is of class C or analytic, so is z on each such D. If G is a circular region and 2:* is continuous along dG, then z is continuous on G. Proof. The existence of a minimizing z is immediate, using a minimizing sequence and Lemmas 9.4.14, 9.4.16, and 9-4.15. Now, suppose B{xo, yo; a) C B and let w{r, 6) = z{xo + r cos0, yo + r sin0). For almost all r, w is A.C. in 0 with \WQ\ in L2. Let 0 (r) = Z) \_z, B (xo, yo; ^)], ^ = max(M/w, n d{z*, B)jm). Then, from Lemmas 9.4.8(b) and 9.4.13. it follows that 0[r) <.Lr0'{r), 0(a) ^d(z*,B) =D{z,B) for almost all r. Hence z satisfies the hypotheses of Lemma 9-4.18 and so z^ C^ (D) for any D G G B, The differentiability results then follow from Theorem 1.10.4 since the continuity allows us to conclude that z is an extremal for the integral I {z, D) whenever D is small enough for z (x, y) to ^ the range of a coordinate patch of Tl for {x, y)^D, where f{x, y, z, p, q) = gij(z^, . ..,zN) (pi pi + qi qJ). Now, let -B be a circular domain and suppose (xo, yo) ^dB. Suppose that R is small enough so that B(xo,yo',R) intersects no part oi dB except the circle C containing (xo, yo) and define D [z,B(xo, yo', R)] == e^(R)l27i, w(r, 6) ^z(xo-rr cosO, yo + r sin(9). It follows that there is an f such t h a t l\we(f,e)\^dd<D[z,B(Po,R)],
BrtdB {PQ, r)
Rle<f<R,
po = (xo, yo) ^
and z is A.C. along B Pi dB{po, f) (so w is A.C. in d up to dB). From the ScHWARZ inequality and the continuity of z* along dB, it follows t h a t if C* = f on 5 n dB(pOy f) and = ^* on B(po,f) n dB, then (9.4.27) osc C* on d[Bn B(Po, r)] <e(R) + oj(R)
389
where CD (R) is the oscillation of z* on B(po, f) H dB. Now, B pi B(po, f) can be mapped b y a conformal map on the closed unit disc B(0, 1) so that a given point pi of B O B (po, r) is carried into the origin. By using maps of the form w WQ = (z zo)' to get rid of the corners and then using Theorem 9.3-^ we see t h a t the resulting map is analytic on the closed domain except at the corners. Let f be the transform of C* and Z be that of z. Then Z is continuous on ^(0,1), Z^HI[B{0,\)], Z = C a.e. on dB(0,\), and Z minimizes D[Z\B(0,\)] among all such functions. If R is small enough, D[z,B(po,R)] ^mjn so that 0{r) <Lr0'(r), 0(r) 0 < y < 1, 0(1) <8^(R)l27i, =D[Z,B(0,r)]. Iju = 1/L, <e(R)(\ +/^-^),
By the method of proof of Lemma 9.4.17, we find that 0(r) < (27r)-i e^(R) (rlR)^f^ - k^(r),
(9.4.28) 2jr _ 1 271
f\C{0)
0
-Z(0,0)\de
<I
jL(r,d)drde
00
Since oscC = oscf* and (9.4.27) and (9.4-28) hold we see that max IZ (0, 0) - C(^) I < (^) (1 + /^-^)l2n + e(R) +a) (R). e Since Z(0, 0) = ^(pi) and pi was arbitrary, the two dimensional continuity of z at {xQy yo) follows. We can now prove our main theorem. Theorem 9.4.3. Under the hypotheses of Theorem 9.4.1, H a vector z of class H\{B), B being a circular region of type k such that D [z, B) = d[r) and L (z, B) = l(r); z is (generalized) conformal and satisfies the boundary conditions as in Theorem 9.4-1. The differentiability results for z are those of Theorem 9.4.2. Proof. Let {zri] be the minimizing sequence of Lemma 9-4-8 (c). Then a subsequence converges weakly in H\ (B) to a vector Z which satisfies the boundary conditions and is continuous along dB. From the lowersemicontinuity, it follows that D(Z,B) <d(r). From Theorem 9.4.2, it follows that H a z^H\{B), z = Z a.e. on dB, and z minimizes D{z', B) among all such z'. Moreover z is continuous on B and so ^ H\[B). Consequently d{r) ^D(z,B) ^D (Z, B) < d(r) so t h a t z is our desired solution. Since d{r) <2l(r) and D{z',B) > 2 L{z\ B) for every z'^ B[\{B), it follows t h a t z is conformal. Remark. Z is also a minimizing function and so has the properties of z.
390
9.5. The general two-dimensional parametric problem We first present the author's simpHfication (MORREY [17]) of the
existence proofs of CESARI, DANSKIN, and SIGALOV referred to above.
We shall assume t h a t the integrand function satisfies the general assumptions o n / g i v e n in 9.2 (just before equation (9.2.12)) and, in addition the condition (9.5.1) M '\X\ >f{z,p) {>m\X\ by (iii), 0 < m < A f ) . We recall t h a t if iV = } (and v = 2) then f{z, p) = 0(z, X) and t h e conditions hold if 0 ^ C^ for X :^ 0, 0 convex in X, and 0 satisfies (9.5.2) m\X\ <0(^,X) < M | X | , X^ = [p^ q^ - p^ q^), Z2 = (^3 ql _ pi qS) ^ x^ = (^1 q^ ~ p^ q^). We shall confine ourselves to the consideration of surfaces of type k with k = 1, i.e. of the type of t h e disc. Definition 9.5.1. If F is an oriented closed FRECHET curve, we define (9.5.?) d(r) = inf jlim inf ^f(zn, Bi)}\ Bi = B(0, 1)
for all sequences {z^} which converge uniformly on ^ ^ i to a representation of r. The following lemma is immediate: Lemma 9.5.1. (a) If Fn-^ F in the sense of Frechet, then
d(F) < l i m i n f ^ ( A ) .
(b) If d[F) <ioo^ ^ a sequence {J!^} 5 ^cich Fn is regular and of class C^, Fn -^F, and d(Fn) -^d(F). If F is a Jordan curve, the Fn ntay be chosen to be regular Jordan curves of class C. (c) / / F is a regular Jordan curve of class C^, d[F) = mi^f{z,B{) for all locally regular z^ C'^(5i) for which the restriction of z to dBi furnishes a representation of F. Remark. In order to prove the local regularity in (c) above without increasing the number of dimensions N, which would require the extension of/, one can begin b y approximating b y polyhedra Fin with boundaries Fny spanning Fn and /^ by a piecewise regular C^ band, and then rounding off the vertices and edges. It is convenient to have at our disposal the ''dominating function'' 0 defined in the following lemma. I t is not necessary for the existence theorem; the rough function W defined by
See Equation (9.1.15) and Definition 9.1.4'.
9.S. The general two-dimensional parametric problem
391
is sufficient for t h a t purpose. However, it yields a simple proof that the solution vector is (generalized) conformal and is also helpful in proving the higher differentiability results once the solution vector is shown to ^Cl(D)ioTDCCG. Let us suppose that f{p, q) is of class C^ when the {p, q) matrix has rank 2 and suppose we perform a rotation of axes in R^ and define 'pi = Ciip. Then we observe t h a t Y = Ciiqi, 'f{p',q')=f{p,q).
(9.5.4)
= 2 ' y^' Jiii^ + ^fîJiiy
+ Ji^P^) + /^' Jm
'^' ' 1 ^
Thus if / is weakly quasi-convex in {pyq), J is also in i^p, 'q). If {pQ, qo) is any point where the {p, q) matrix has rank 2, we may perform a rotation of axes so t h a t (9.5.5) %-'qi = 0 for .>3, 'Pl'ql-'Pl'ql> 0. Let us consider the set of all i^p, 'q) satisfying (9.5-5) and let us drop the primes. The relation (9.L3) becomes (9.5.6) f{^p+pq, yP + Sq)={ocd-^y)f(p,q), ocd-^y>0. Differentiating (9-5.6) with respect to p^ and q^ and solving, we obtain
MP>Q) = dfpi - yU,

fpipi(P,Q) (9-5-7) fpiqj(P>Q) = A-^[d^fpipj = ^~^ [ pSfpipJ
MP, Q)^ yS(fpiqi
^fpi + ocU
+ y'^fqiqi] OCyfqiqj]
+fqipo)
+ Ocdfpiqj + ^yfqipj
fqiqi{P,Q) P = ocp+Pq,
= ^~^W^fpip^
0C^(fpiq7 + fqipi) + OC^fqiqf\ A=ocd-^yy U==fAP>q)> etc.
Q = yp + dq,
Thus, if we let î = (1, 0, . . ., 0), ^2 = (0, 1, . . ., 0) and define

/^ r- O N fp^pj(^ly ^2) = îj, fpiqj{ei, ^2) = hjy fqiq^(^l> ^2) = Cfj,
fpi{ei, 62) = di, f{p.q) = h^{PH^-pH^),
fqi{ei, 62) = e\, fpi=dtq^--e[q\
f(ex, ^2) = ^0 f^,= -d^p^ + e,p\
we find that, on the space (9-5-5)
Qif) = K-îiaiAq^^-P^H)^
(9.5.9) K - ]lE G-F^ =.piq^
- (hj + hi) {q^^-P^fA

ii iJ ~ ^2 ^ 1 .
X ( ^ u - p^ fi) + cij (^1 A-p^/^)2]
Q{f) being the form in (9-5.4). Thus the form Q necessarily degenerates and has rank < iV 2 in f for all (A, //) with X^ -^^ fjfi = 1.
392
Definition 9.5.2. We say that a n / o f the type discussed is the integrand of a regular parametric problem <^ the form Q (/) is non-negative definite with rank = iV 2 f or all (A, fji) with X^ + [JL^ = \ and all [p, q) of rank 2. Remarks. In the case N ='},this corresponds to the existence of numbers mi and Mi such t h a t 0 <mi < M i . Since the set of all normal orthogonal pairs of vectors (ei, ^2) is compact, the condition is equivalent to the existence of mi and Mi such that Ml (22 + ^2) | | | 2 < J [a^,. A2 + (bij + bji) 111 + C^,//2] | ^ | y (9.5.10) ^ < M i ( A 2 + / . 2 ) | | | 2 , |l|2 = i ( f O ^ miK'HGX^ - iFXf, + /^2) | | | 2 < Q^f) + EiJi^)\i\^.
or, at the general point in the space (9-5.5) <.MiK-^(GX^-2FXiJi
By multiplying by a constant we may assume t h a t / satisfies the m M condition with 2 = m < M . Lemma 9.5.2. / / / is the integrand of a regular parametric problem with V = 2 and iV > 3 and if f^ CJJ [or C^) whenever the {p, q) matrix has rank 2, then there is a function 0 which is homogeneous of degree 2 and of class C'^(or C^) for all (p, q) ^ (0, 0), which is the integrand of a regular nonparametric problem (i.e. Q{0) has rank N), and which satisfies (9.5.12) D^0(z,p^q)^MD^ ^-^^>
<^{z,p.q)>f[z,p,q),
the equality holding <^ E = G and F = 0, M being the bound for f Proof. Since z will act as a parameter in the following construction, we shall suppress it. Since / satisfies the w M-condition of (9.5-1), we m a y suppose (by multiplying by a constant) that m = 2. We define 0 to have the form (9.5.13) ^{P.q)=D[i+co{r,h)],
T ( ^ , q) = D-^K,
D =^^,
q),
E = \p\^ etc.
K = ]/EG-F^,
h{p, q) = K-^f{p,
Clearly we m a y apply the rotation procedure to D, K, r, and h as well as to / . Thus it is sufficient to compute all the derivatives on the space (9-5-5)- AH the requisite derivatives o f / c a n be computed from (9-5-7), (9.5.8), and (9-5-9)- The derivatives of K are easily found on our space from the equation K^ = E G F^] we obtain K=piq^^p2qi, Kj,i=Kqi=0, ^ > 3 .
9.S- The general two-dimensional parametric problem
393
A straightforward computation leads tothe result that Q(0) = (1 + C - TO),) (A2 + /<2) If |2 + T-2(Tft), - h CO^) X O
X D-i(GX^-2F?./A (9.5.14)
+ Efi^)\d\^ + (oX^ +
-f 2 T - l K f t - T-1 COft) X y + T-2 Wftft r 2 + 7-2 o)^ D-1 X
if X 2" [jj e^ - (^ + *j) e < + c; (T2] I< f^ ^ Q = Xq^ jup^, a = Xq^ [xp^.
X and y are seen to be bounded and defined ii D ^ 0, x i=^ 0, and the last term in the expansion for Q (0) is between (9.5.15) miT-^conW and Mir-^wuW, W = D~^(GX^ - 2FXju +
Clearly we m a y assume (9.5.16) (9.5-17) (9.5.18) 0<mi<1. l+co Tco^>m2>0, 0 = D{\ + co) ^f=rhD, r co^ hcofi'> ~ niicoji. co(\,h) = h \, Accordingly, we want there to exist an W2 > 0 such that We must also have the equality holding only if T = 1 (corresponding to E = G, F = 0), We notice also t h a t
(9.5.19) (l - ] / l - T 2 ) | | | 2 < T F < ( I + 1 / I -T2)|f|2.
We define CO(T, h) = [\ + {h -2)T](p{u), (9.5.20) 0<T<1, u = {h1) T/[1 +(h 2) r ] , 2 < A , T = fi/[(A 1) (A 2) w],
Then it follows t h a t u increases from 0 to 1 as T does, for each h, and co^=[[h \) {h 2) u] cp'(u) + (h 2) f [u) r co^ o) = u (p' (u) (p (u), (9 21) (2 wi) 99 [u]] 0)^^ (Dnh ~ {corh T-i co^)2 = 0 0)^^ = [{h 1) (A - 2) w] (p"(u) (duldr) > 0 <^ ^ " ( ^ ) > 0. For T = 1,(1 + c o ) T h = 0 and ^ [(1 + co) T A] =- co^ /j < 0 ^/ ^ " {u) > 0 coh = r[(p (u) -\- (i u) cp' (w)] "^<^r (^ ^ i ) oiji = T{[(2 wi) ^ (1 mi)] 99'[u)
A ^ ^ = [ ( / , _ 1) _ ( / , _ 2) ^] (^"(^) > 0
394
SO 0)^. takes its max. for r = \ where its value is co,{\,h) = q)'{i) But for T = 1, we must have ((9.5-17)) (jo^{\, h) co(1, A) < 1 m2^ (p'('i) +h 2<h m2 ^^^^^^ ^(p'{^) < 2 - m 2 . Thus a>^ <,h and so (9.5-18) holds. The first inequality in (9-5-17) holds if the last one in (9-5-22) does. If 99"(^) -= 0 for ^ < (1 wi)/(2 wi) after which (p" (u) > 0, all the above hold and r co^ {h mi) cohÔ so we see t h a t (9.5.23) (p'W > 2 - m i . So we take 99 = 0 for 0 < w < (1 wi)/(2 mi), choose cp^C^ [0, 1] and analytic with q)" [u) > 0 for (1 wi)/(2 m-i) < w < 1 and so that 99(1) = 1, 2 wi < 99'(1) < 2 ^ 2 where 0 < m2 < mi. Definition 9.5.3. We define the integral
G
+ h - 2 .
J(Z, Bi) =fj(\zaâ^\^

By.
+ 2 1^^2,12 + \zyy\^)
dx dy,
Remark. We note t h a t I(z, G) = ^f{z, G) ^ z^ Hl(G) and z is conformal. Lemma 9.5.3. Suppose F is a regular Jordan curve of class C^ and z = C(p), p^dBi is a regular representation of F of class C^. Then, for K sufficiently large, the class g (K) of vectors z^H\ {Bi) O HI {Bi) for which (9.5.24) J{z. Bi) < K, z{p) = C (P) for p = (1,0), ( - i, ^ ,
is not empty. For each such K, there is a vector ZK ifi 3 {K) which minimizes I (z, Bi) among all z in ^(K). Moreover (9.5-25) d{F) =\im I(zK^Bi).
Proof. The first statement is evident since there is a locally regular Z $ ^ ^ ( ^ i ) such that z = Z(p), p^dBi, gives a representation of F; the three point condition may be secured by performing a Mobius transformation, the transform of z being in C^ (Bi). Now let {zp} be a minimizing sequence for / in Q(K). Then the norms of the Zp in Hl(Bi) are uniformly bounded so that the Zp are equicontinuous on ^ i by Sobolev's lemmas (see 3-5)- Thus, a subsequence converges weakly in Hl(Bi) and uniformly on 5 i to some function ZK in HI {Bi). Since this convergence implies weak convergence of the second derivatives and strong con-
9.5- The general two-dimensional parametric problem
395
vergence of the first derivatives in L2(Bi), we see that ZK is admissible. Since / and / are both lower-semicontinuous (/ involves only first derivatives), ZK is a minimizing function. The last statement follows from the fact that we m a y select a sequence {zp} of regular surfaces of class C^, each bounded by F, such that ^f{zp,Bi) -^d(r); each of these m a y be represented conformally by some admissible vector z*. Lemma 9.5.4. For each K, ZK satisfies the condition
(9.5.26) D [zK,B(Po. R)] <.D[ZK,B{P,, a)] - g ) \ X= ^ ,
0 <R
<a,
for every circle B (Po, a) C Bi. Fhus the ZR satisfy (if K is large enough) a uniform Holder condition on EA, which depends only on A, m, M, and d{F), for each A <. \ and are equicontinuous along dBi. There exists a vector z^ Hl(Bi) such that z = z(p), p^ dBi, gives a representation of F which satisfies the three point condition in (9.5-24) and for which ^/(z, Bi) = d(F); the vector z satisfies the Holder condition satisfied by the ZK on each BA' Moreover z is [generalized) conformal. Proof. The equicontinuity on each BA follows from the first result and the writer's ''Dirichlet growth theorem" (Theorem 3-5.2). The equicontinuity along dB\ follows from a well known lemma of Courant (Lemma 9.3.2) since we have D(ZK, BI) uniformly bounded and have a three point condition. To prove (9.5-26), we use (9.5-12) and the minimizing property of ZR to show that
'-mD [ZK, B (Po, R)] < / (ZK. B (PQ, R)] <I[Z,B (PQ, R)]
(9-5.27)
. <^^-MD[Z,B(Po,R)],
(m = 2)
where Z is the biharmonic function such t h a t Z z^^ HIQ[B(PO, R)], We shall omit the routine justification of the following formal calculations. Let (r, 6) be polar coordinates with pole at Po and let ZK (r, 0) = ^ + 2J [^n (r) cosnO + bn(r) sinnO],
0 < r < ^, 0<r<R.
n=l
Z(r,Q) =^~-
+ 2^[An(r)
cosne+
Bn(r) sinnQ],
Since Z is biharmonic and has the same Dirichlet data as 2: on ^ 5 (Po, R), we obtain
/ r \n ,y\n+2 /yxn /y\n +2
396
where Cn, dn, %, and fn are constants defined by

2Cn = (n + 2) ocn ^n, 2dn = ^n nan,
and similar formulas hold for en and fn for ^ > 0. Thus if we set W{R)=D[z,B{Po,R)], equation (9.5.27) and a computation of D [Z, B(Po, R)] shows that W{R) <LD[Z,B{Po,R)]
R
oo
n-l
<2LRW{R),
1 = .
The result (9.5-26) follows. Thus from the ZK, we may extract a subsequence {zp} which converges weakly in Hl(Bi) and uniformly along dBi and on each EA with ^ < 1 to some z^H\ (EA) for each A <i \. Since both / and 3 / îre lower semicontinuous with respect to this type of convergence, (Theorem 4.4-5) we conclude that ^f{z,Bi)Î{z,Bi)<d{r). But suppose we define ZR = z on ER and ZR to be the harmonic function on Bi - ER which coincides with z on dBiU SBR. Since each ZR is continuous and ^/(ZR, BI) -> Sf(z, Bi), etc., we find that d{r) < ^f(z, Bi) < I(z, Bi) < d{r). From this we conclude that z is conformal (on the interior). From the lower-semicontinuity of / on each domain D of type k, we obtain I {z, G) = lim / (zp, G). But now, from the lower semicontinuity of Z), we have (9.5.28) -mD(z,G) ^
1
< - m lim inf D (zp, G) < lim /(zp, G) ^ ^'"^'^ ^"''^ = / ( ^ , G) <I(H,G) <^MD{H,G),
H being the usual harmonic function. The continuity of z at points of dBi now follows as in the proof of Theorem 9.4-2. The result (9.5-28) gives the Holder continuity of z on interior domains. We can now prove the existence theorem: Theorem 9.5.1. Suppose F is a Jordan curve for which diT) < 00. Then H a continuous vector z^H\ (9.5-29) Moreover z is conformal. (Ei) such that = d(r). ^f{z^Bi)
9.5- The general two-dimensional parametric problem
397
Proof. Choose {Fn} so that each Fn is regular and of class C^ and so t h a t Fn ^ r and d(Fn) ->d{F) (Lemma 9.5-1). Choose regular representations z = ^n(p), P^ dBi, of class C ^ of the Fn so that the Cn(p) converge uniformly on dBi to C{p) where ^ = f (^) is a topological representation of F. For each n, let Zn be the minimizing vector of Lemma 9.5-4. From Lemmas 9-4-2 and 9-5-4, we conclude that a subsequence, still called {zn}, converges uniformly along dBi and on each BA as in the proof of Lemma 9.5-4. B y repeating the argument in the last part of that proof, we find that z{= limzn)^ Hl(Bi), that (9.5-29) holds, and the z is conformal. We conclude this section with the results of K I P P S mentioned above. In fact, we generalize his results somewhat. Lemma 9.5.5. Suppose p and q^H\ [^(Po, <^)]. (a) Then, for almost all r, 0 <Cr <i a, p and q [Lemma 9-4-5) <^^^ A.C. on dB (PQ, r) with pe and qe^L^y <^^^ 2/ /
B(Po,r)
ipxqy pyqx)dxdy
= J {pdq
SBiPo.r)
qdp).
(b) If, -also, jr and x^Hl[B[PQ, K-q^H\^[B{Po,r)-\,then J J

B{Po,r)
r)] for such an r with jt p and =J J

B{Po,r)
{jtx^y JtyXx) dxdy
(px^y
pyqx)dxdy.
(c) If p =z Zx and q = Zy for some z^ Hl[B(Po, a)], then qx = Py a.e. on B{Po, a). Proof, (a) and (c) are proved b y approximating b y the mollified functions. To prove (b) write n = p -\- TIQ, K = q + XQ', then TCQ and XQ may be approximated strongly in HI [B (PQ, r)] b y functions ^ C^ [B (PQ, r)] and p and q may be approximated similarly b y functions ^ C [B(P,r)]. Lemma 9.5.6. (a) The conformal minimizing vector z of Theorem 9.5-1 satisfies (9.5.30) fJiCifv^ + Clfai + C'fzi)dxdy = 0, C^Hl,(B,).
" Bx
(b) / / also, p and q^ H\\B{PQ, a)] for some B{PQ, a) C C Bi, then fpi, fqi, and fzi ^ HI [B [PQ, a)] and z satisfies (9.5.31) Lr^f^fpr + ^jar-fzr=0 a.e. o 5 (PQ, ^). n
(c) For almost all [x, y) where p = q = 0, we have fp==fq=fz = ^ and the derivatives of p, q, fp, and fq vanish a.e. on that set. Moreover p and q satisfy the relations (9532) P ' P^ ^ ' ^^ = q ' Px + P ' q^ = P ' py q ' qy = (l'Py+p-<ly = ^ a.e.
398
(d) If N = } and we define X = p x q and f{z, p, q) 0[z, X), then ^fpr + -^-U+fzr = kr[{Ap^ + B Py + B q^ + C qy)' k B = HD],
(9.5.33)
k = \X\-^X,
C = 0XQXO q^q\
A=0xQXop^p'.
H = 0XQ ZQ
0xQXop^q^,
for almost all (x, y) where X ^ 0, i.e. {p, q) ^ (0, 0). Proof, (a) follows since z minimizes the non-parametric integral / and so satisfies (9.5-30) with / replaced by 0. Since 0 > / everywhere with the equality holding along our solution, (9.5.34) <>pi= fpi, 0qi = fqi, 0zi - fzi a.e.
along our solution. Since 0 is homogeneous of degree 2 and the derivatives 0ppy 0pq, and 0qq are continuous and bounded if {p, q) ^ (0, 0) and the derivatives 0pz and 0qz are continuous everywhere, we easily see that fp = 0p and fq = 0q^ HI [B(Po, a)]. We have seen elsewhere (. }A) that the derivatives of ^ , q,fp,fq, a n d / ^ = 0 almost everywhere on the set w h e r e ^ = q = 0, The relations (9-5.32) follow by differentiating the conformality relations. The relations in (9.5-33) follow by simply carrying out the differentiations in (9-5-31) and using the form of / : Since qx = py, we see, for almost all {x, y) where X i^ 0, that
Lr = %rsp% + "^rsiPl ^ r s = (pXQX<^ Xlr X^s , 5)r5 = q^XQzs Xlr , + q%) + ^rsty + "^rsp' + ^rsq' - 0ZT
S&rs = (fXQX*^ X^r X^s , rs = (fXQzs X^r,
Krs = (pXQX^ X^r XJs , ^ (p (z, X)) .
{f(z, p,q)
Using the conformality relations, one easily sees t h a t k^X^r^ Since

prXlr = qrXlT=X^, pr Xlr = ^^ Z g r = 0 ,
-p^,
krXlr=
- qK
the results in (9.5-33) now follow easily from the homogeneity of 99 as a function of X, Our method is to set up a non-parametric integral of the type discussed in 5.4 with unknown 3-vector functions n and x which is minimized by taking [n, K) = (p, q). The Holder continuity then follows from the results of that section. Definition 9.5.4. Given a conformal vector z ^H\ (Bi) with p and q also $ HI [J5(Po, ^)], B{Po,a) C ^ 1 . We define (9.5.35) I* [^, ^, B(Po, a)] = j
J ( o BP . )
J W(x, y, n, x, Jix, ^x, ^y> >Cy) dx dy
9.5- The general two-dimensional parametric problem where, for X iÔ, i.e. [p, q) i=- 0, W=\ny+ {P'TlyK:c\^ + 1X|-1 {[p-n^-qq- HyY + {q'ny+p' K^)^ + [q ' n^: + p ' x^)^ +
399
Hy)'^] + J^ [(A ' Tlx + B ' Jty +
+ B ' Kx + C ' Ky) ' k H(p ' n -\- q- >c)j2\'^ 2l{7ix ' Ky Tiy - Kx) + \n-p\^ + \K-q\^, k = \X\-^X, / = (2 - ]/2)/(1 + W ^ i ) (9.5.36) where A, B, C, and H are defined in (9.5-33) and mi and Mi refer to (9.5-9). If ^ -= ^ == 0, we define W = m2{\7Zx\^ + \7iy\^ + \xx\^ + \xy\'-) + ITT -P\^ + \X(9.5.37) m = (W2 + M2)/2, m2 = I mijMi, M2 = larger of 2 + ]/2 / and / Mijmi. q\^,
Theorem 9.5.2. HWQ denotes the sum of all the terms in W which are qtiadratic in jix, 7iy, Xx, d'yid Xy, then (9-5-38) msd ^ ^ P + \7iy\^ + \xx\' + \xy\^) < ^ô < M2{\7ix\^ + + \ny\^ + \^x\^ + \>cy\^) > m2 + M2^2. Suppose that z is the minimizing vector of Theorem 9-5A and that p and q^Hl[B(PQ, a)], [B(PQ, a)] C Bi. Moreover the pair of vectors (7c,x) ==(p,q) minimizes I'^[7i,x, B(Po, a)] among all such pairs (7z,x)
^H\[B{PQ, a)]^n p and x q^H\Q[B{PQ, a)]. Thus, p and q
^ C^ [B (PQ, R)] for each R < a. Moreover the Holder condition depends only on the bounds m, M, mi, and Mi and not on any bounds or moduli of continuity for z, p, and q other than those holding by virtue of the minimizing property of z. If 0 is of class Q ( C ' ^ , or analytic) for \X\ ^ 0; then so is z away from the locally compact subset where p ==z q = 0, Proof. At points where {p, q) ^ 0, let us define the n.o. set {i,j, k) b y i=\P\-lf, j=\q\-U. and let us set Ttx = ocii + îj + yik ny = a2i + ^2 j + y^k Then, for such points. (9.5.39) k=\X\-^X
Xx = ocsi + PsJ + 73^ Xy = aî + ^d + 74^. +
Wo = {oc2 - ocs)^ + ih - iffs)^ {h -73)2 ' + (^1 ~N'
+ (^3 + i5l)2 + (^2 - i^4)2 - (^4 + hY ^ AC (^71 + + By2 + Bys+ C y^)^ + I [(<X2 ocs -Oil ^4)+(i^2/?3- -PiN ai 0C4) + + (y2 73 7174)]. A surprisingly simple computation shows that t h e characteristic roots are (some are multiple roots) 1,1 + 2, l + 2y2, 2 1, lAjC, and I CIA,
400
The higher dimensional plateau problems
Since the latter two are between ImijMi and I Mijmi, (9.5-38) follows. In case TZ = p and x q, we conclude from Lemmas 9.5-5 and 9-5.6 that (9.5.40) 'P=-2l{Pcc'qy-py'qx) (A^)^(0,0). For other {n, x) we see that
(9 5 4^) ^ > -2l[nx'Ky B(Po,a) -Tly'Kx),
/ * [TZ, X, B(Po, a)\> 21 j
J (px ' qy py' qx)
dxdy
since (9.5-40) and the first inequality in (9.5-41) hold almost everywhere where [p, q) = (0, 0) or, respectively, (jr, x) ^ (0, 0). Hence the minimizing statement follows. Finally, if z is the minimizing vector, we know that {p, q) ^ L2 (Bi) and also satisfy a Dirichlet growth condition of the form / / {\P\^ + \q\^)dxdy^D[z,Bi)'[rjh)^-i^\
B{Fi,r)
5 = 1 - |OPi|.
Thus the integrand W satisfies the conditions in 5.4 whether p and q are continuous or not. The remaining results follow.
Chapter 10
The higher dimensional plateau problems

10.1. Introduction Until recently, no general results had been obtained concerning the existence and/or differentiability of the solutions of parametric problems in more than two variables. The greatest single stumbling block was the non-existence of a useful generalization of a conformal map t o higher dimensions. Now, by imitating the proof of the author's old conformal mapping theorem (MORREY [3]), one can prove that a "nondegenerate'' Frechet variety of the topological type of the I'-ball (i.e. a Frechet variety which possesses a representation on ^(0,1) in which no continuum is carried into a point) which possesses a representation of class Hl[B{0,i)] possesses such a representation which minimizes f \v z\^ dx among all such. However, one can not conclude that
5(0,1)
the value of this integral <, C L [z,B{0,\)'] or even that L [z,B[0,\)'] is given b y t h e area integral for such a representation. So the methods which had been successful in the two dimensional problems did not lead to results in the higher dimensional cases.
10.1. Introduction
401
Almost simultaneously, results on the r-dimensional PLATEAU problem

with V > 2 were obtained b y DE GIORGI [2], REIFENBERG [1], F E D E R E R and FLEMING and FLEMING [2]. D E GIORGI proved t h a t a portion of
minimum area of a part of the ^^-dimensional boundary of an open set in (v + 1)-space is a regular analytic manifold. REIFENBERG proved that if A is any compact point set in R;^, there is a compact point set X in R^ which is bounded by ^ in a certain sense (see Definition 10.2.6) which minimizes A*{X A) among all such X and which has the additional property that each point ^ of ^ ^ , not in a relatively compact subset Z oi X A with /[*' (Z) = 0 , is in a neighborhood on X A which is a topological r-disc. Recently (REIFENBERG [2] and [3]) he has proved t h a t these topological i^-discs are in fact analytic. Very recently, since delivering the Colloquium Lectures before the American Mathematical Society in August, 1964, the author (MORREY [21]) has found that these results m a y be carried over to sets X and ^ in a Riemannian manifold of considerable generality (see below). Almost concurrently, F E D E R E R
and F L E M I N G ( F E D E R E R [2], FLEMING [2], F E D E R E R and FLEMING) have
approached this problem using their integral currents and have obtained results more or less comparable with those of REIFENBERG. Since F E D E RER is writing up their results in book form, we shall present the author's extension to Riemannian manifolds of his simplification of the work of
REIFENBERG.
Suppose that 9Jl is a Riemannian manifold without boundary of class C**, w > 2. Let P o ^ - I t is well known that there is a (non-unique) coordinate patch r of class C^ having a domain containing the origin such that (10.1.1) hi^{0) = dij, hiJy;Jc(0) = 0, the hij being the components of the metric tensor with respect to T. T O obtain such a mapping, let co be any coordinate patch of class C^ with range containing PQ ; we may clearly assume that its domain D in the y-space contains the origin. By letting T be a properly chosen linear transformation from the y-space to the <2:-space, we may arrange that gij(0) = dij, the gij being the components of the metric tensor in the z coordinate system. We obtain r by letting (10.1.2) wi = zi + a)T^z^z^ where the a)^ are constants to be determined. Differentiating the relation
with respect t o z^, setting 2 = 0, and using (10.1.1) and (10.1.2), we see t h a t the a|j. must satisfy
(10.1.3)
5ft + b = ^M(0).
26
402
The higher dimensional PLATEAU problems
The reader may easily verify that we may take
Definition 10.1.1. A coordinate system co of class C^ (at least) is a normal coordinate system centered at a point qonWl^ the domain G of 0} contains the origin, a> (0) = q, gij (0) = dij, the gij being the components of the metric tensor with respect to co, the arcs on 50^ which correspond to segments in G through 0 are arcs of geodesies, and the distance along such segments equals the corresponding distance along t h e geodesic arcs. Lemma 10.1.1. Suppose W. is of class C^. Then (a) Each point q of W is the range of a normal coordinate system of class C^ centered at q, (b) If 0)1 and co2 are two such systems both centered at the same point q and both having the domain B{0,R), then OJ^^ coi is the restriction to B (0, R) of an orthogonal transformation. (c) If oy is such a system and gij(z) are the components of the metric tensor with respect to oj, then
(10.1.4)
gij(tX) ^n^ = \ if 2* W^ = ^ ^^^

^=1
0<t<R;
gijzJc(0) = 0 for all i,j, k. (d) Given a point PQ and a mapping r of class C^ satisfying (10.1.1), there exists a number Ro'> 0 and a family cop of normal coordinate systems, one o)p being centered at each point p in B (PQ, iô/3) [on W) and having the domain B(0, RQ), such that the vectors in the tangent space to at p which correspond under ojp to the unit vectors ei, . . ., e^ in RN CLY^ obtained from those corresponding to î, . . ., ^jv under x by the GramSchmidt process. If p and q^B(Po, i?o/3), the mapping (10.1.5) ^=U{y;p,q)ôj-ncop{y)] is of class C2 for {y, p, q) ^G x 31 X "SI where G = i5 (0, Roj}) C RN ^^^ W=B{Po,Rol3). Proof. Clearly part (a) follows from part (d). Part (b) is evident since there is a unique geodesic passing through a given point and having a given direction. The first equality in (10.1.4) follows since the lines z^ =z X^t correspond to geodesies and t is the arc length along the geodesic and (10.1.6) (dsldt)^=gij'{tX)Xn^^ 1. By differentiating this with respect t o t and 2fi we obtain
gijzic{o)xnn^ = 0 (aiu)
(^^-^'^^
gijzk (0) A*' A* + 2 lim^-i [gjcj (t X) - djcj] X^ = 0 ^
for all A. The second result in (10.1.4) follows from (10.1.7).
10.1. Introduction
403
To prove (d), we first note that theEuLER equations for minimizing the length integral become, if t is a parameter proportional to arc length
(10.1.8)
"^
"^
Û == :^^^^[h}wTc + hikwj hjkwi]' So, let PQ and r be given as stated. From the existence theorem for systems of ordinary differential equations as applied to the system (10.1.8), we see t h a t the solution function IX (^; y, z) of the system nU(t;y,z)+HUmt;y,z)]nin^ = 0 U**(0; y, z) = z\ Ui(0; y, z) == a}(z) yj where the vectors a}{z) ei are the vectors in RN corresponding under r to the vectors in the tangent space at ^ = T (Z) which are obtained from those corresponding under r to ei, . . ., e-t^hy the Gram-Schmidt process. That is, the a}{z) are uniquely determined by the conditions that (10.1.9) hij{z) ai(z) al(z) = djcu 4(^) = 0 for i > k. I t is easy to see t h a t IX ^ C^{Q) for some open set Q containing (0; 0, 0) and t h a t (10.1.10) (10.1.11) Vi{t',ry,z) == Vi[tr]y,z) so Vi{t]y,z) = XX(l;/y, ^). r-^p)]} If we define U{y, z) = 1X(1; y, ^), a>^(y) = r{Uly, we see that 17 ^ C^ in a domain of the type desired and that the cop satisfy the conditions stated. Remark. Of course IX is of class C^ in t and, since the 7^^ ^ C^ (and, in general, nothing more can be said if 3)1 is only of class C^), we see that the first and second partial derivatives of X , X ^ and XI ^^ with respect to X X, y and z are of class C^ in (t, y, z). Since the hfj are of class C^, so are the a}{z). We now state our general assumptions on Tl: General assumptions on 501. We assume that W is a separable Riemannian manifold without boundary of class C^ and that there are positive numbers RQ, CQ, f]o, KQ, and Ki, which are independent of PQ, such that each point PQ of Wt is in the range of a coordinate system r of class C^ which satisfies (10.1.1), has domain B(0, 4Ro), is such that T(0) = PQ, and for which 2 (\V^ hij(w)\ + \V^ hij(w)\) < Ko. We suppose also that RQ, CO, and KQ are so related that (1 + Cor^)-^<hij(w)2,n^ r=\w\, 2'^^=''. + Co/'2)2,
(10.1.12)
<(i
(1 + l 6 C o i R g ) 2 < 5 / 4 .
26*
404
With each PQ, we associate a definite r as above, and we associate a family {cojj} of normal coordinate systems related to T(PO) as in Lemma 10.1.1; we assume that the number RQ mentioned there is the same as that mentioned above. Instead 0/(10.1.12) we assume that, for each fixed p, cop satisfies (1 +r]or)-^^gij{y;p)?i'X3'^(\ ^^^'^'^^^ |r;fe(y;^)AV^>^l < ^ o | A | - | / ^ | ^ +r)or)^, 2" W = ^
0^\y\Lr^Ro,
the gij {y; p) being the ojp components of the metric tensor and the Fjj^ {y, p) being obtained from the gij by the corresponding formulas (10.1.8). Finally, we assume that the second gradients of the function U (y, p, q) in (10.1.5) are uniformly bounded by Ki for each PQ and all {yyp,q) in Remarks 1. It is probable that many of the results can be carried over to manifolds of class C^ with n <i 4 ( 0 < / / < l ) . However, a relaxation of the differentiability requirements appears to necessitate a different method of proof. It is shown in 10.7 t h a t the topological ^-discs (on the minimizing set) are of class C^ for any //, 0 < // < 1, if 3}l is of class C4, are of class CJJ if W is of class Q with n '> 4, and are of class C ^ or analytic if M is. Remarks 2. It is seen that any compact manifold of class C^ satisfies all the conditions. Lemma 10.1.2. Let co be a normal coordinate system with domain B(0, RQ) and suppose that the gij {z) are the components of the metric tensor. Then, writing z = [z^, % ) , we obtain
(a) gNm{z^, 0)=dNm, ^ ^ | < iÔ i
(b) the angle on W at p between an arc y through p and the geodesic from q = (0(0) to p is the same as the angle at y = CJO~^ (p) between co~^ (y) and the line 0 y. Proof. Setting z^ = tX^ in (10.1.6) and using the fact that lines through the origin are solutions of (10.1.8), we obtain (10.1.14)
(10.1.15)
gij(z) zizJ=
\z\^,
r;j^{z)z3z^~0,
2gNizN[z^,
i=
\,. . ., iV.
0) == 0 .
Setting 2:* = 0 for ^ < iV and using the formulas for F^^, we obtain *
gNN[z^> 0) = 1, 0) - gNNzi[z^,
Differentiating (10.1.14) with respect to z'^, setting z^ =^ 0 for i < and then using (10.1.15), we arrive at the equation
N,
^^^[^^?i^m(^^,0)] = 0
which leads to (a), (b) along the z^ axis follows from (a) and along any line through the origin follows by a rotation of axes. Lemma 10.1.3. For each ? > 0, there is an RQ with 0 <, RQ <. Rol3, y which depends only on W (i.e. RQ, CO, TJO, KQ, and Ki) and rj and which has
10.1. Introduction
405
the following property: Suppose 0 < R <, RQ, B{p,r) C B(q, R), W is a normal coordinate system centered at q, andy' is any point on a>~i [dB{p, r)], then BiCco-^[B(p,r)]GB2 where Bi and B^ are the halls in R^ of respective radii r{\ + ^)~-^ (^^^^ r[\ + rj) whose boundaries are tangent to co~i [dB{p, r)] at y'. Proof. 'Letp' = co {y'). Since our results are unaltered by orthogonal transformations in R^, we may suppose that co = coq, and that ojq, Wp, and a>^, are all related to a single r as in Lemma 10.1.1 and our general assumptions. Let a be the non-homogeneous linear transformation on RN which carries y' into the origin and which osculates a)~} MQ at y'; we write a in the form z = (T{y), a{y') = 0. The angle at the origin in the ^-space between two arcs which intersect there is the same as the angle at p' between the arcs which correspond under the transformation coq a~^. Also lines in the y-space correspond to those in the 2:-space. Let y be the directed arc in B (0, r) which starts at the point WQ oj~^(p') ^dB(0, r) and is such that G a)~'â)p(y) is a segment starting at 0 and making an angle 6 < 7r/2 with the inner normal to a a)~^ cop[dB{0,r)]. From Lemma 10.1.2, it follows that y makes the angle 6 with the radius vector from WQ to 0. Since U(y; q, q) = y and the second gradients of U (including mixed ones) are bounded by Ki, we see that
(10.1.16)
(i+i^^^)-2<|iM<^(1+if,7?)2,
w being the coordinates in the cop system. Now, let us introduce the Euclidean metric in the ^-space and let
/ N
X7
J V
(>2^
^z^
Then the arc y is a solution of the equations
(10.1.17)
- _ + r'o,,(t.) ^
= 0
where the F^jj^ are defined as usual in terms of the gQjcj and t is the distance in the 2:-space. From our assumptions it follows that
(10.1.18)
Now define
\ri,,(w)xi,xii^'^\
v{t) = [d{Pt,pW
<z^{m)\x\-\i,\\
= Z(wi)^.
406
Letting dots denote differentiation with respect to t and using (10.1.17), we obtain v{0) = r^, v{0) = 2 2Jw^{0) wÔ) = 2r \w(0)\ cos^
N
(10.1.19)
'^^
v(t) = 2 2! (ze')2 2^! rijj, (w) w^ w3 w^,
Since t is distance in the ,2:-space, (10.1.16), (10.1.18), and (10.1.19) yield r^-2{\ (10.1.20) +Ki R)^tr cosd + {\ + Ki R)-^ (1 - Zi R) t^ + (1 + J ^ i 7^)4(1 + ZiR>Q. 1 <v(t) ZiR)P, < ; ' 2 _ 2(1 +KiR)-^trcosd
By letting ^* be the first positive value of t for which v{t) r^, we find from (10.1.20) that (10.1.21) 2(1 + K i J R ) - 6 ( 1 + ZiR)-^rcose <t* ^ ^ ^ ^ +KiR)^{\ -ZiR)-^rcosO.
This proves the theorem with a)~^ [B(p, r)] replaced by a co~i [B(p, r)]. Call the balls in the ^-space B[ and B^^ Their respective radii are (10.1.22) (\ + Ki R)-^ (\ + Zi R)-^ ' r and (\ + KiR)^ and we have a-HB',)Cco-nB{p.r)]Ca-HB',) where a~^ (B[) and a~^ (^g) ^^^ ellipsoids with min. and max. radii a-^(B[): (1 + Ki R)- (1 + Zi R)-^ r and (1 + Kx R)-^ (1 + Zi JR)-ir, cy-i(^2): (^ + Ki R)^ (1 - Zi i?)-i y and (1 + Ki RY (1 - Zi i?)-i r, respectively; these ellipsoids are tangent to co~^[dB{p, r)] at co~'^{p') = y', If R is sufficiently small, the ellipsoids B\ and ^g ^^^ nearly balls and balls B\ and B^ can be chosen with B\ C ^ i and B^ D ^g^ which have respective radii (1 \- r\)~^Y and (1 -^ Y()Y and which are tangent to o\^\dB(p, r)] at y'. The last statement is proved similarly. Definition 10.1.2. Given a set 5 , we define (5, ^) as the set of all points within a distance ^ of S; i.e. (5,e) = U S ( P , e ) . If S\ and 52 are compact sets, we define their point set distance D (Si, S2) as the smallest number Q such that Si C (5*2, Q) and 52 C (5i, ^). Definition 10.1.3. ^ geodesic k-plane centered at a point P 0/ 9K is a locus of the form colliD B(0, RQ)] where 77is a ^-plane in R^ through 0 and ft) is a normal coordinate system with domain B (0, RQ) for which co(0)=P. (\ ZiR)-^-r,
10.2. V surfaces, their boundaries, and their HAUSDORFF measures
407
Lemina 10.1.4. Suppose co is a normal coordinate system with domain B (0, iô) ^^^ range B (PQ, RQ) for which co (0) = PQ. Let P be a point in B(0, R), let Si and S2 he compact subsets of B{0, R), and let a){P) = Q and w (Sjc) = Tjc, k = 1,2, 0 < J R < RQ. Then (1 + i?o R)-^ d{P, Si) < d(Q, Ti) < (1 + T o R) d{P, Si) y (IO.I.23) (1 + rjo R)-^D(Si, (1 + rjo R)-^Q{Si) S2) < Z)(ri, T2) < (1 + >;o R) D{Si, S2) <Q(Ti) < (1 + 7^0 R) Q{SI)
where Q(S) denotes the radius of the smallest sphere containing Si [compact). Proof. If Si consists of a single point, the first line of (10.1.23) follows from our assumption (10.1.13). There is a point P i in 5 i such that d{P, Pi) = d{P, Si). If we let Qi = co(Pi), then d{Q, Ti) < d(Q, Qi) <.{\+f)o R) d(P, Pi) = {\ + rjo R) d{P, Si). The other inequalities are proved similarly. We conclude this section with the following lemma: Lemma 10.1.5. There are constants RQ = RQ (W) <, Roj} and K2 = K2 (W) with the following property: Suppose that ^ is a geodesic k-plane centered at a point p^ B{q, R) where 0 < P < RQ. Suppose co is a normal coordinate system with domain B (0, PQ) for which co (0) = q and suppose ^0 is the k-plane in P^v which is tangent at 0)~^{p) to co~^(^). Then co-i [ZnB (q, P)] C (Zo, K2 P2) co-^[i:r\B{p,r)](Z[Zo,K2r^), r ^ RQ . Proof. If the coordinates in t h e system co are w and t h e F^j^ are defined as usual, the arcs in B (0, PQ) corresponding to the geodesies are solutions of the system (10.1.17) with F^^j^ replaced b y F^j^. Since the geodesies making u p ^ H P (q, R) correspond to arcs tangent to ^ 0 at co-îp) and since ^ fl P ( ^ , P ) C ^ Pi P ( ^ , 2P) so that no such arc is of length > 2 P , the result follows from the assumptions (10.1.13). The last statement follows since the second derivatives of the transformations co^^ cooq, etc., are uniformly bounded. 10.2. V surfaces, their boundaries, and their Hausdorff measures In this section, we define the r-dimensional Hausdorff measures of sets in a metric space and prove certain theorems about these measures. Most of the less standard theorems presented here are due to R E I F E N BERG [1]. We also introduce our notions of r-surfaces and their ''algebraic boundaries''. Most of the topological results needed for the remainder of this chapter are presented in the following section. The presentation there given makes extensive use of the general theory as presented in the
book ''Foundations of Algebraic Topology'' b y E I L E N B E R G and S T E E N -
ROD, especially Chapter 1, 1 14, Chapter 9, 7 , and Chapter 10, 2 and 5.
408
Definition 10.2.1. Suppose 5 is a set in a metric space and t h a t v > 1 and 6 > 0. If S is empty, we define *^^ (S) == 0; otherwise, we define *Al (S) as the inf. of ^ yv rl for all coverings of 5 by finite or countable
i
families of balls {B (Pi, r^)} in which each n < d. We then define the Hausdorff outer measure *yl*' by (10.2.1) *yl''(S) = l i m * / l S ( 5 ) .
f5--0+
A'-measurahle sets are then defined as usual. If r = 0, we define */lg (S) as the inf. of the number of balls of radii < d required to cover S and then define *ylo(5) by (10.2.1). The following lemma is well-known and we omit the proof: Lemma 10.2.1. (a) Borel sets are measurable A^ for each v. (b) Any set is measurable A^ and A^{S) is the number of points in S. (c) / / 5 is a set in a metric space ^ and r : S -> ^' is a mapping in which W is a metric space and d[x{p), r[q)'] <.2.d{p,q) for p and q on S, then *A'[r{S)] <A'*/L^(S). Definition 10.2.2. Suppose G is a domain on a manifold X of class C^ and suppose r ^ C'i(G) where r is a mapping into the Riemannian manifold Tl of class C^. We define
L,(T)
G
=jF{p;r)dS(p)
where if cr : F ->G and co : Q -^Tl are coordinate patches with domains FG Rp and Q C RN and ranges containing p and T {p), respectively, we have F{p\T)dS{p) = [y[x)']-î'^dx, y{x) =det(yap{x))
dS(p) = fe(^)]i/2 dx,
g{x) = de t(g.^{x))
where Gij(z) and gocp{x) are the components of the metric tensors on W and X with respect to the coordinate patches w and cr, respectively, and z(x) = o)~^T a(x). Lemma 10.2.2. Suppose that G, X, 90^, and r have their significance as in Definition 10.2.2 and suppose that r is a diffeomorphism. Then Lp(r)Â-[r{G)]. This is proved by first proving the formula lf[T(P)]
G
F[j>;r] dSip)
= ff{q)
r(G)
dA'iq)
for / ^ C'^[r{G)]. This is proved first for functions with support in a small neighborhood of a given point. The proof in that case is elementar5\
409
Definition 10.2.3. A family g^ of sets T is said to cover a set S in the sense of Vitali iff each point P of 5 is in a set T of g^ of arbitrarily small diameter. Definition 10.2.4. Let g be a family of sets T. By (7(g), we mean the union of all the sets T for T ^ g^. In case f is countable and g = {B{Pi, ri)} ^ we still use the notation am = UB{Pi,ri).
Lemma 10.2.3. (cf. MORSE, A. P. [2]) Suppose that W satisfies the conditions o/ 10.1. Suppose that S is a bounded set and that % is a family of closed halls in 9}l which covers S in the sense of Vitali. Then there is a countable disjoint sub-family {B(Pi, ri), i = 1 , 2 , . . . } such that S -U B(Pi,ri) CU B(Pi, Sn), k = \,2,...
Proof. We suppose that 5 C B{PQ, ro) and let ^' be the family of all balls B(P, r)^^ such that B{P, r) C B(Po, ro) and B(P, r) 0 S is not empty. Then g ' still covers 5 in the sense of Vitali. We now define the radii Ri, the balls B {Pi, ri), and the families î by induction as follows: iî sup r for B (P, r) ^ ^\ B (Pi, ri) is a ball of g ' in which ri > iî/2, and g i consists of those balls of g ' which do not intersect B(Pi, ri). Having defined Ri, . . ., Rjc, B[Pi, Vi) for ^ = 1, . . ., k, * and g i , . . ., %jc, we define Rjcî = snpr for B(P, r) ^ ^jc, B(Pjc+i, rjc+i) as a ball in ^jc for which rfc+i > Rjc+il^, and ^k+i as those balls in Â; which do not intersect B{Pjc+i, rjcî). We see by induction that any ball in g ' g i must lie in B(Pi, Sri) and, in general, any ball in ^jc ^jc+i must lie in B(Pjc+i, Srjc+ij^ By induction, we conclude that the B{Pi, r^) are disjoint and that a(^jc)CUB(Pi,Sri)GUB(Pi,Sri)
i=Jc+l i=j
if
j < k + \ ,
Now, suppose P^S
\J B(Pi,ri).
i=l
Then, since the finite union is
closed P ^ some ball B{P',r) which does not intersect any B[Pi,ri) with i < k. Thus B{P', r) ^ %k and hence P ^ a{%ic). The result follows since
5 - u5(P^,n) = n S-\jB{Pi,ri)
Definition 10.2.5. Suppose 5 is a set, P is a point, and / 7 is a ^-plane, all being in Rjsf. We define the cone C (P, S) to consist of all the segments P Q ioT Q in S. We define C(77, S) to consist of all segments P Q where P ^S and Q is its projection, on 77. If M satisfies the conditions of 10.1, C is a normal coordinate system with origin at P and domain containing O
410
C(P,S), we define the geodesic cone C(P', S') == (o[C (P, S)], where P ' = co(P) ^nd S' = a)(S). Theorem 10.2.1. Suppose S C RN, S is a Borel set, and A^~^(S) < oo. Then, for any P, ^^ [C (P, 5)] < r - i r / l " - ! (5) if S C B{P,r). The proof is similar to but simpler than that of the following theorem and is left to the reader. Theorem 10.2.2. Suppose S C RN> A^~'^{S) < oo, and IT is a p-plane with p <. N. Then ^ U B{P,r). pen Proof. The theorem is evident ii v = \. So we suppose r > 1. We cover 5 with a countable family of balls {B (Pi, r^)} such that (10.2.2) 2yv-i^r^<^'"'^{S) + ' n<d, i=\,2,...,
i
AnC(n,S)}^Ci{v)'r'A^-HS)
if
S C {n,r)
e and 6 being arbitrary positive numbers. Let P^y, 0 < y < (r ri)lri be the point on the segment I (Pi), joining Pi to its projection Qi on / 7 , which is at a distance j Ti from P^. li P^B (Pi, fi), we see that /(P)C UB(Pij,2ri). Accordingly, since j < rJTi, we obtain Al, [C (77,5)] < 2 ^ yî^î)' ^ ''(2" Y'lY'-i) ^ y - i rV'
1
<c[(v)-r-[Ar'{S)+e]. The result follows from the arbitrariness of d and e. Theorem 10.2.3. Suppose 9)1 satisfies the conditions o/ 10.1, S is A^-measurable with yl*'(5) < oo, 5 C G C 501 where G is open, and U^C^(G), S/U ^ 0 on G and U satisfies a Lipschitz condition with Lipschitz constant M. Then
o o
/ y l ^ - M ^ n Ch)
o o
dh<MA''(S),
Cji being the subset of G where U (P) = h. Proof. It is sufficient to prove this for 5 compact; then \VU(P)\ > c > 0 for P on 5. Let e and 6 be small positive numbers and cover S with a family {B (Pi, ri)} of balls such that 2r^n<A',{S) + s, B{Pi,n)(lG. {ri<d). Now, for each i,
i
Ar' [Sh n B (Pi, n)] < y._i [Qi (A)]-i,
Su==sncn,
where Qi(h) is the radius of the smallest closed ball containing Ch OB (Pi, ri). This intersection is non empty only for U(Pi) -[ri'\V U(Pi) I + 0(d)] < A < U(Pi) + [n I V U(Pi) I + 0(^)].
411
For such h, Qi{h) = ir^,-h'\ Thus A' = [| V U{Pi) i + 0(a)]-i [h U[Pi)].
JAr^Su)
^MZyvrl
i
dh<Z
jl^^ U(Pi) I y,_i [Qtih)]"--^ dh' + 0(d)

+ 0{d) + e.
+ 0(a) < MAl(S)
The result follows. Definition 10.2.6. A 7^-surface is merely a compact set (in W) X. In case ^ is a compact subset of X and v ^ \, we define the algebraic boundary b{X, ^ ) of X with respect to A (more properly b[X, A,G), G being a group of coefficients; but we shall suppress G) as the kernel of the homomorphism i^: Hv-i (A) -> Hv_i {X) (i.e. Hv-i {A,G) ->Hvî (X, G)); here i is the inclusion mapping from A into X and i^ is the corresponding homomorphism. The Cech homology theory is used. Remarks. In case X is a compact orientable r-manifold of class C^ with ordinary boundary the connected (v 1)-manifold A of class C^, then b{X, A) = Hv-i(A) (for any G). Of course our definition allows X and A to be arbitrary compact sets with A G X.li A = X, then b {X, A) = 0, an uninteresting case. But if L is a non-zero subgroup of Hv-i{A), the class {{A, L) of sets X for which b((X, A) Z> L is an interesting class and, as we shall see, large enough to contain an X such t h a t X A has minimum HAUSDORFF /l*'-measure. The topological theorems which ensure this are the following: (i) Each class ^{{A,L) is closed under point set convergence (see Definition 10.1.2) of X^ to Z (i.e. if each Z ^ $ e : ( ( ^ , L ) , then X $ e : ( ( ^ , L ) ) . (ii) Suppose X^{i[A, L),_G is open (in Tl),GnAis empty, XDdG = B, U is compact, U CG, Ur]dG = B, and b(U, B) Z) B(Xi, B) where X\ = {X r\G). Then the surface X\ obtained from X by replacing XnGbyUnG, also ^^(A, L), (i) is just Theorem 10.3-16 and (ii) is Theorem 10.2.5 below. The theorems in 10.3 and many others were proved by J. F. ADAMS in the Appendix to the paper [1] of REIFENBERG mentioned above. That Appendix contains many further examples illustrating the notion of algebraic boundary. Lemma 10.2.4. Suppose A is a compact set in R]>^, P is a point in R^, and X = C{P, A). Then b(X, A) = H^_i{A) if v > 1. This follows from Theorem 10.3-2 since X is contractible. The following isoperimetric inequality of REIFENBERG is important: Theorem 10.2.4. Suppose W and RQ satisfy the conditions o/ 10.1. Then there are constants C2[v,W) and C^{v,W), 2 < r < iV, with the following property: If A is compact, Ad some ball B(Po,Ro), and yl''-i(^) = / " - i < + oo, there exists a surface X such that b(X,A)
412
D Hv-.i(A), X is in the geodesically convex hull of A and within a distance < C2I of A, and (10.2.3) A^X) ^Cs'l'Proof. By virtue of the results of 10.1 and Lemma 10.2.1 (c), it is sufficient to prove this for 90^ = i^jsr in which case RQ may be arbitrary. Accordingly we assume that A C RN anxi carry out the construction in
RN-
We prove this by induction on v. We define

00
From Theorem 10.2.3, it follows that

i.iz
0
J(p(a)da <A'^{A) = I.
Accordingly there is a value of a for which (p(a) < 1 and hence 0. By repeating this construction along each axis, we see that we may divide R^ into cubes of side 1.1 I, no one of which contains a point of A on its boundary. Of course only a finite number contain points of A ; call them iî, . . ., Rg. In each Rs, select a Ps interior to the convex hull of A n Rs and define X = UC(Ps,As), As = AnRs^
That b{X, A) = Hv_i{A) follows from Theorem 10.3.3. The other properties are evident. Suppose, now that v> 2, N '>v, and that the theorem has been proved for all (v', N') in which 2 < 1^' < v and N' > v\ Let A be given and define
Then there is a value ^ i , 0 < ^^ < /, such that 99I (a^) < l^-^. Let n]i = m. 1 and D\x = An n\i. Then
From our induction hypothesis it follows that we can find {v 1)surfaces B\i C n\i such that h{B\i, D\i) = J^,_2(Dii), ^î is in the convex cover of D\i and within a distance C /î < C Z of D\i and satisfies (10.2.3) for r 1. If we let ^^î be the part of A for which a^ + /(s^ 1) < :x;i < a^ + / si and define Ci =- ^ U U B\i,
si
C\i = B\i_^ U A\i U B\i,
(/ii)-2 _ ^.-2 p i ^ ) ,
413
then the hypotheses of Theorem 10.3.11 are satisfied (with r = s^ + + some integer) and also
cÎA^-HDi^y--l)/('-2) <
C/^-1.
For each s^, we replace A by C\i and repeat the construction above along the x^ axis obtaining the surfaces D ^ 2 spanned as above by B^^ ^ satisfying (IO.2.3). We let A^^l^^ be the part of C\\ for which ^ 1 / + + Z(s2 - 1) < ::t;2 < ^12 + / s^'and define
Again, the hypotheses of Theorem 10.3.11 are satisfied (with r = s^ + + some integer) and also
Then, for each (s^, 5^), we perform the same construction along the x^ axis with A replaced by O-^ ^. This process is continued until all the axes are exhausted. Then each CJi - T ^ is in the closed cube ^1 + Z(sl 1) <::t:l < a i + Zs^, a^f + / ( s 2 1) < ^ 2 < ^^1^2 ^ / ^ g , . . ., of side I, For each Qi--^j^, select a point Pgi'V.^'^ interior to the convex cover of C,V::5^ and define Z,V;;5^ = C(P,V;:5^', CJ,-;;.^^)
S^ Si
Then, by using Lemma 10.2.4 first and then Theorems IO.3.9, IO.3.IO, and 10.3.11 repeatedly, with A = Qi*;;.^s& and Agk+i = Clxii%t}u ^s+i = Xli;;^gt}i, we see in turn that
b{X,A)=H,_i{^)
and that all the other conditions are satisfied. Theorem 10.2.5. Suppose L is a subgroup of Hv-i{A), b{X, A) D L, and G is an open set such that G C\ A is empty. Define Xi=^ [X r\G), Ai = Xi G, and y = Y1 U ^2, where Y2 =^ X G and Y\ is any surface such that 6(Yi, ^1) D 6(Xi, ^1) and YiCW^G A^. Then b{Y,A):DL, Proof. Define Z 2 - y2, A2 = A{jAi = B, U = b{Xi,Ai), U ==b(X2,A2), L[ = b(YiMi), and L^ = L2 = ^ ( ^ 2 , ^2). Then the hypotheses of both Theorems IO.3.9 and 10.3.10 are satisfied by {X, Xi,
414
X2) and (Y, Yi, Y2). Thus b(X, A) = i{B, A)^ni{B, Ai)^ Li + i{B, A2U U] h{Y, A) = i{B, A)-^ [i{B, Ai)^ L[ + i(B, A2U L',]. Since L[ D Li and V^ = L2, it follows t h a t b{Y, A) DbiX, 10.3. A).
The topological results of Adams (see REIFENBERG [1])
We shall assume that all sets are subsets of a HAUSDORFF space. {X, A) is a pair (of compact sets in which A G X). bA (X) or b{X; A) Ker i^: : Hv-.i(A) -^Hv-i{X) where i : A -^X is the inclusion map. We shall denote the inclusion map of A into X by i{X, A) : i{X, A) (a) = a if a^ A. Let PQ denote a ''base point" and also the set {Po}- Given X, we let fx'.X ->Po be defined by fx(x) = PQ for x^X. We let ex (or s(X)) : Ho{X) - > G be the map y / x * where y is the isomorphism from HQ(PQ) onto G (Actually, in the book by EILENBERG-STEENROD, G is defined as Ho(Po), in which case y is just the identity). x is called the ''augmentation homomorphism" by ADAMS (REIFENBERG [1]). Theorems 18 below are just the respective Lemmas 1A8A of the paper
R E I F E N B E R G [1].
Theorem 10.3.1. If X = A, then b{X, A) = 0. Proof. For i{X,A)^: Hv_i{A)-> Hv_i(X) is the identity so no elements a ^ 0 in A are carried into 0. Definition 10.3.1. A set X is said to be contractible (oti itself to a point) iff H a continuous map h \ X x I -^X such that h {x, 0) = A; and h{x, \) = X{i for all x on X, XQ being a fixed point. Theorem 10.3.2. IfX is contractible, then b {X, A) = HP_I(A) if r > 1, and b(X, A) = Ker SAifv= 1. Proof. From EILENBERG-STEENROD, Chapter 1, Theorem 11.5, it follows that Hq(X) = 0 for every q ^ 0 and HQ{X) = 0 (see Chap. 1, 7) and Ho {X) is isomorphic with G. By definition b{X, A) = Ker i(X, A)^ = Hv.i{A) since Hp_i {X) = 0 if r > 1. In case v = \ b (X, A) = Ker i^ (X,A)
n
ii
v>i
= Ker SA
since SA = y / x * H {^> ^ ) ^^^ y / y * is an isomorphism. Theorem 10.3.3. Suppose that X = U Xr where the Xr are disjoint
r=l
and contractible. Define Ar = A 0 Xr, Sr = e(Ar), and Ko = SH^> ^ r ) *

r
KersrCHo(A), 1.
Then b(X, A) = Hp_i(A) if v > 1 and b(X, A) =Koitv=
10.3- T h e t o p o l o g i c a l r e s u l t s of A D A M S
415
Proof. From ES^, Chap. 1, Theorem I3.2, it follows, by replacing the general pair {X, ^4) by (^, 0) and (X, 0), that
r=l r
and, in fact, each a in -ô(^) is uniquely representable in the form (10.3.1) a = 2^ i(A,Ar)âr, ar^Ho(Ar)
and the correspoinding result holds for X and Xr. If r > 1, it follows as in Theorem 10.3.2 that each Hvî(Xr) = 0 so that Hvî(X) = 0 and the theorem follows in that case as before. Now, let (5 and 6 be the isomorphisms above from ^ o ( ^ ) onto ^ Ho(Ar) and from Ho(X) onto ^ Ho(Xr) and let / be the homomorphism from ^ Ho{Ar) to 2! Ho(Xr) in which Xr ~ i[Xr, Ar)^(ar). Then, it foUows that i{X, A)J=d-Îd. T h u s a ^ h { X , A ) ^ i ( X , A)^ a = 0 4 > i{Xry Ar)^ Ur = 0 for r = i, . . .,n. But since r = y fxri^ i{Xry = Ar)^ and y fxr^ is an isomorphism, it follows t h a t a^ hiXÂ) ^ Sr(ar) = 0, r = i, . , .,n. But this holds ^ a^2
r r r
i(A, ^ r ) * Ker Sr =
KQ.
Theorem 10.3.4. Suppose that X is an v-disc with boundary A. Let L denote Hp_i(A) ifv'>\ and Ker SAifv= 1. Then b{X, A) Z) L. Proof. This follows from Theorem 10.3.2 since X is contractible. Theorem 10.3.5. If A is a Jordan arc, then Hi{A) == 0. Proof. This follows since A is contractible. Theorem 10.3.6. Suppose that f: (X, A) -> (Y, B) and LA is a subgroup of Hv_i{A). Let LB = {f\A)^LA and suppose b{X, A) D LA. Then b{Y,B)DLB. Proof. From Theorem 4-1, Chapter I of ES and its proof, it follows that the following diagram is commutative: Hv.i{A)
/l*
^Hr.ii^)
, ^*
/2*
where we have used the notation of that theorem. In particular We are given that i^(LA) = 0 and that LB = / 2 * ( ^ ^ ) . It follows t h a t i^ (LB) = fu (0) = 0 which is to say that LB C Ker i(Y, B)^ = b{Y, B). Theorem 10.3.7. Suppose L is a subgroup of Hv_i{A), thatL C b(X, A), and that Y D Z . Then LCb(Y,A).
^ i. e. EILENBERG-STEENROD.
416
Proof. This follows from Theorem 10.3.6 by t a k i n g / as the inclusion map. Theorem 10.3.8. Suppose N = v and A is the unit {v \)-sphere in Rr{A =dB{0, 1)). Then (a) IfXD^^), b{X,A) = H,_i{A) if v > \ or b{X,A):^Ker SA if V = \. (b) If X does not contain B{OA),h{X, A) = 0, Proof. Part (a) follows from Theorems 10.3.4 and IO.3.7. In part (b), there is a point in ^(0,1) X, since X is compact. Let XQ be such a point and let Y consist of all points (1 t) x + t^, 0 < / < 1 , x^X, where f is the intersection of the ray [x^s x) with A. Then {A, A) is a strong deformation retract of {Y,A). By Theorem 11.8, Chapter 1 of ES, the homology sequences for (Y, A) and (A, A) are isomorphic. Accordingly, in considering the ^'j,.-parts of these sequences which are, respectively, HmMY) < - ^ m - i ( ^ ) and Hm-i{A) ^^^H^_i(A), we conclude t h a t Hm-i{A) :^ Hm-i(Y) so b{Y, A) = 0. Hence b(X,A) = 0 by Theorem 10.3.7. Theorem 10.3.9. {R \^, Lemma 11 A). Suppose that X = U^ Xr. Suppose ArdXr and A (Z X and we define B = A\J U Ar. Suppose
r
L and Lr are subgroups ofHv-i (A) and Hv-i [Ar), respectively, and suppose b(Xr, Ar) Z) Lr. Suppose also that (10.3.2) i(B, A)^LCZ ^ {B. Ar)^ Lr. Thenb(X,A)DL. Proof. Suppose h^L. Then
i(X, A)^ h = i(X, B)^ i{B, A)^ h^ i(X, B)^ Z HB, ^ r ) * Lr r = 2 " i(X. Ar)^ Lr = Z i{X. Xr)^ i{Xr, Ar),. L , r r = Zi(X,XrUO = 0 r since b {Xr, Ar) D Lr. Thus h^ Ker i(X, A)^ = b(X, A). Lemma 10.3.1. Suppose that X = U^Xr, B = UÂr, Ar C Xr for each r and Xr C[ Xg =^ Arf\ Ag whenever r ^ s. Then, for each q, each u ^ Hq {X, B) can be represented uniquely in the form (10.3.3) U^îrÛr, Ur^Hq(Xr,Ar), ir : (Xr, Ar) C (X, B). r Proof. We shall prove this ior n = 2; the theorem can be proved for any n b y induction. To do this we consider the triad (Yi U Y2; Yi, Y2) and the inclusion maps i^2:(Yi, Yi 0 Y2) C (Yi U Y2, Y2) and h: (Y2, Yi 0 Y2) C (Yi
1 i. e. REIFENBERG [1].
10.3- The topological results of ADAMS
417
U Y2, Yi), where Yi = X^D B and Y2 = ^ 2 U ^ ( s e e E-S, Chapter 1, 14). Since each kr is an identity map, Yi (Yi fl Y2) = (Yi U Y2) Y2 = Yi Y2 and is open in Yi U Y2, and ^2 maps Yi (Yi n Y2) in a 1 1 way onto (Yi U Y2) - Y2, it follows that ^2 is a relative homeomorphism; the same is true of ki (see ES, Chapter 10, 5, p. 266). From Theorem 5.4 on that page (generalized excision), it follows that the maps ki and ^2 induce isomorphisms of the homology groups in all dimensions, so that the triad above is proper. Thus it follows from S , Chapter 1, Theorem 142 t h a t each w ^ Hq{Yi U Y2, Yi n Y2) is uniquely expressible in the form u = ii^ ui + ^'2^ U2, where ur^ Hq{Yr, Yi n Y2) and ?, : (Y Yi n Y2) C (Yi U Y2, Yi H Y2). From the hypotheses and the definitions of Y\ and Y2, it follows that YiUYs^X, Y i n Y 2 = 5, Hq{YiDY2,Y^r\Y2)=H^(X,B), B - (Yr - Xr) = Ar,
Yr - {Yr - Xr) =Xr,
and Yr Xr is open (in X) and C Yr 0 B. Thus, from the generalized excision (ES, Chapter 10, Theorem 5.4), it follows that the inclusion maps Tr : (Xr, Ar) G (Yr, Yi n Y2) induce isomorphisms of the homology groups in all dimensions. The result follows (^V* = ^V* "^r*, ^r = Tr* Ur). Theorem 10.3.10. (R 1. Lemma 12 A) Using the notation of Theorem 10.3.9, suppose further that A 0 XrC Ar and that Xr H XgC ArH As whenever r ^ s. Let Kr = h(Xr, Ar) d Hv_i(Ar). Then b(X,A)=i(B,A)^'\ZHB,ArUKr] Proof. If we let L denote the right side, then (10.3.2) holds, so t h a t b (X, A) D L. To show that Lob(X, A) we must show that if h ^ Hv_i(A) and i(X, A)^ h = 0, then i(B,AUh^2:HB,ArUKr,
r
i.e. there are elements Ar Hv_i(Ar) such t h a t (10.3.4) i(Xr, Ar)^ hr = 0 and i(B, A)^ h = 2^ i(B, Ar)^ hr.
r
Since i(X, A)^h = i(X, B)^ - i(B, A)^ h, this will follow if we can show t h a t any k^ Hv-i(B) with i(X, B)^ k = 0 can be represented by the sum on the right in (10.3-4) in which each hr satisfies i(Xr, ^ r ) * K = 0. This is proved by diagram chasing in the following diagram: Hr(X,B) ^-^ HrM^) - ^
Ui
Hrî(X)
r Morrey, Multiple Integrals
i~ i:d 2; H, (Xr, Ar)~^-^2:
Hv.l [Ar) - J i/._l (Xr).

r 27
418
The marked isomorphism follows from Lemma IO.3.I and 2J ^ denotes the transformation defined by * " 12 ^] (î> '^n) = {dui, . . ., dun), > etc. If ^ ^ Hv_i(B) and i(X, B)^ ^ =- 0, then k^Ker i{X, B)^ = dHy{X, B). This follows from the exactness of the Cech homology theory over the category ^ c of compact pairs and maps of such {ES, Chapter 9, Theorem 7.6). Thus H a # in Hv{X, B) ^du = k. From the marked isomorphism, it follows that u is uniquely representable in the form (10.3.3) with q replaced by v. Then Then r %, Ur^ Hv[Xr, Ar), ir ' (Xr, Ar) -> (X, B). hr = dur
du = ^ diri^ Ur = 2 i{B, Ar)^ hr, r r as desired and since each of the rows
^ H, (Xr, Ar) -> H,_i (Ar) -^^ H,_i (Xr) > is exact, it follows that i(Xr, Ar)^ hr = 0 SLS required. Theorem 10.3.11. Suppose that n A = U Ar, Ar n Ar+i = Dr for r = \, . . .,n ~ \, ArH As = 0 if \r s\> \. Let Kr denote Hv-2 (Dr) if v > 2 and Kr = Ker s : HQ (Dr) ->Gifv = 2. Suppose that b(Br, Dr) 0)Kr, r = 1, . . ., n; BQ = Bn+i = 0 C := AUU Br, Cr = Brî U ArU Br, r = \, . . .,n+ i, Then AnBr 2:i(C, = Dr, r = 0, . . ,,n+ \. H,_i{A).
CrU H,_i{Cr) D i{C, AU
Proof. We prove this for w = 2; the proof for the general n is by induction. In this case we change our notation as follows: AÂiUA2, AinA2 = D, Ci = AiUB, C^Â^^JB, C = Ci\J C2 = A\JB, AnB:=D, b(B,D)DK where K = H^_2(D) ii v > 2 smd K = Ker s : Ho(D) -> G if r = 2. The result is proved by diagram chasing in the following diagram: Hv-i (Ai) -r-> H,_i (Ci) > H,_i (Ci, B) ^ F . _ i (^1, D)
Hrî(A)
- > i / . _ i ( C ) >Hv-i(C,B)
^* A
^'i
i
A
-I'
''^
A
1
A
j / *
^H,_i(A,D)-^-^H,_2(D) 1
Hv-1 (A2) > H,_i (C2) - ^

^2*
J2*
H,_i (C2, B) <- H,_i (A2, D) ~

T2
10.3. The topological results of ADAMS
419
Since C A = B D =^ Ci Ai = C2 A2 and is open on C, it follows from the strong excision theorem (ES, Theorem 5-4, p . 266) t h a t the maps T, TI, and T2 are isomorphisms (onto). Clearly the images in Hvî(A, D) of the Hv_i(Ar) are included in those of the Hv_i{Cr), i.e.
Or r~^ jr^ ir^ [Hv-1 ( ^ r ) ] = T - i y * Qr ir^ [ i ^ v _ l ( ^ r ) ] C T'^ j^{Qr[Hv-l{Cr)]]
Next, the sequences ^ . - 1 {Ar) ^-^^^-t Hrî {Ar, D) -^^ H,_2 {D), r=\,2,
H,_i (A) :^:lhl%
H,_i (A, D) ^ H,_2 (D),
are exact, so the image of Hv-.i(Ar) in Hv_i{Ar, D) consists of exactly those elements Ur of the latter group for which dur = 0, r = \, 2; and the corresponding statement holds for the image of Hv-i {A) in Hv_i(A, D). Finally, from Lemma 10.3.1, it follows that each element u of Hv-i(A, D) is uniquely representable in the form u = aiui -{- 02^2, Ur^ Hvî (Ar, D). Taking inverse images under j'^^ x, it follows that i(C, A)^ H,_i{A) C i(C^ Ci)* H,_i{G) + i(C, C2)* H,_i(C2) which is the desired result. Theorem 10.3.12. Suppose that X = 1 xY and ^ = (0 X Y) U (1 X y ) . Let lo and I\ he the natural embeddings of Y in A as [0 X Y) and (1 X Y), respectively. Then h{X,A) D the totality of elements in Hvî (A) of the form If h-I^hforh^ H^-\{Y). Proof. Let / o = i{Xy A) IQ and / i = i(X, A) Ii. Then / o and Ji are homotopically equivalent in X. Thus / i * = / o * [ES, Chapter 1, Axiom 5), so that i(X, A)^ (h^ h - / o * h) - / u A - J o * ^ = 0 for all h^ H,_i(Y), A[ and A\, subthat This is the result. Theorem 10.3.13. Suppose that X' = I xY, AQ=0 xY, = 1 X Y, A' := AQKJ A\ and suppose that f: X' -^X is continuous that fiA^) = Ao, f{A[) = Ai. Suppose that fo=f\ A^, that fi=f\ and that fQ is a homeomorphism. Suppose K = h[X, A) and LQ is a group of Hvî(Ao). Then there is a siibgroup of Li of Hv_i(Ai) such (10.3.5) K + i(A, Ao)^ Lo = K + i{A, Ai)^ Li.
Proof. Let io: Y - > AQ and iii Y -^ A[ be the homeomorphisms defined by io{y) = (0, y) and ii{y) = (1, y), respectively. From ES, Chapter 1, 5, it follows that/o^jt, ^'0*, and ^'i,^ are isomorphisms. Thus we define Li as the totality of elements of the form (10.3.6) hi =fi^ ii^ h where h = V*Vo* 3^0 and ho^ LQ.
27*
420
Now, suppose k^K and h^^ LQ. There exists a unique h in defined by (IO.3.6); let h be defined by (10.3.6). Then
Hvî{Y)
i{A, ^ 0 ) * ho =i{A, ô)*/o* ^'0* h = Fo^h, i{A, Ai)^ hi = i{A, ^ i ) * / i * ^1* h = Fi^ h, Fo:Y-Â, Fo = i(A,Ao)foio. Fi'.Y-Â, Fi = i(A,Ai)fiii. But now the maps Go = i{X, A) Fo and Gi = i{X, A) F\ are seen to be homotopically equivalent in X since we may define Gt'.Y^X by Gt{y)=f{t.y), 0<^<1. Thus from ES, Chapter 1, Axiom 5, it follows that i{X,A)^[i[A,Ao)^ho~i[A,Ai)^h{\ = 0, or (10.3.7) i[A,Ao)^ho=-k' + i{A,Ai)^hi, k'^K = b{X,A), or k + i(A, AoU ho = k + k' + i(A, Ai)^ hi^K + i{A, A^) U. In like manner ii k^K and hi^Li, there is at least one h and hence one ho satisfying (10.3.6) and hence (10.3.7). Thus k + i{A, Ai)^ hi = k-k' + i[A, Ao)^ ho^K + i[A, Ao)^ U. Theorem 10.3.14. Suppose that IJ is a v-plane [in R^) through P, that A =nndB(P,r), and that D = U r\ B{P,r). Then h (D, A) ~ Hvî(A) and if X is any surface I) A for which b(X, A) = L where L is a non-zero subgroup of Hv-i(A), then A'{X) >yl'(Z)) =yrrr Proof. That b{D,A) = Hv_i(A) follows from Theorem IO.3.8. Let f = Pn ^nd Y = / ( Z ) . Then / : (X, A)-^{Y, A) and /2 - / U is the identity. From ES, Chapter 1, 4, the diagram H,_i(A)~^*^H,_i(X) I/2* l/i*
Hr_i(A)^H,_i{Y)
is commutative and /25J- is just the identity (Axiom 1). Now, suppose y^ b(X, A). Then y^ Hp_i{A) and i^(X, A) y = 0. So
n(y) =fuHf2iy
= ^'
Thus b(Y, A) D b(X, A) ^ 0. The theorem then follows from Theorem 10.3.8 and the fact thatyl''(Z) >/l'(Y). Theorem 10.3.15 {R \, Lemma 21 A). Suppose Xr D Xr+i and b(Xr, A) D Lfor r = \,2, . . ., L being a subgroup of Hv-i{A). Let X = C\Xr. Then b {X, A) D L. Proof. Suppose h^L, Then i{Xr, A)^ h = 0. By the continuity of the Cech homology {E~S, pp. 260261) the injections i{Xr,X) yield
10.4. The minimizing sequence; the minimizing set
421
an isomorphism of Hvî{X) onto the inverse Hmit group l.imHv_i(Xr). This isomorphism maps the element i(X,A)'^h into the sequence {i {Xr, ^ ) * ^} which is a sequence of zeros and is the zero element of the inverse limit group. Thus i(X, A):^ h = 0. Since h is arbitrary b(B,A) Z)L. Theorem 10.3.16. Suppose b{Xr, A) Z) L for r = \,2, . . ., L being a subgroup of Hv_i{A), and suppose Xr -^X in the point set sense. Then b{X,A) D L . Proof. From Theorem 10.3.7, it follows t h a t b[Yr, A) D L for each ;', where Yf = \J Xg. Clearly Yr D Yr+i and pi Yr = X so the result follows from Theorem IO.3.15. 10.4. The minimizing sequence; the minimizing set In this section we construct a minimizing sequence and a minimizing set and prove a first smoothness peoperty of the minimizing set. We consider compact sets X and ^ on 50^ in which (10.4.1) (10.4.2)
(10.4.3)
r
ACX,
b{X,A)DL,
L a subgroup of ^ r _ i ( ^ ) . S ^X A.
We define, for a given compact A and subgroup L, d(A, L) = iniA'îS) for all X with b{X,A)DL,
d(A,L) < + 00.
We shall assume that 9 ^ satisfies the conditions of 10.1 and D We now define R(P) = min[^(P, A), RQ]. li b(X, A) ^ L, we define (p{r, P]X) (10.4.4) y){r,P]X) =Â''[Xr\B{P,r)-\ = fA'-^ [X n dB{P, t)] dt ,R{P)>0, 0<r<R(P).
Lemma 10.4.1. Suppose b{X, A) D L. Then (IO.4.5)

10.4.6)
cp{r^, P; X) - (^(n, P ; X) < ^ ( r 2 , P ; X) - y>{r,, P ; X ) , w{r\P']X)<w(r' ^^ / o<n<^2<P(P), + I \PP'\,P]X),, / , rv i h y + \PP'\<R{P), r' <R(P'), V y.
and (p and ip are non-decreasing on r for each P i A, Proof. (10.4.5) follows from Theorem IO.2.3. with U(Q) = d(P, Q) taking S = Xr\[B{P,r2) -BiP.n)'}. (IO.4.6) follows since B{P\r') GB(P,r') + \PP'\). Lemma 10.4.2. Suppose b(X, A) Z) L and (10.4.7) A^(S) =d(A,L) + s.
422
For 0 <r <R(P),
define
^{r, P; X) = max[0, (p{r, P; X) - e],

(10.4.8)
wir,. P;, X); = msix[0, >w(r,' P;>X)J s], ^ '^v I w\ J^ Q {P, X) = sup r for those r for which ^ (r, P; X) = 0 Q*(P, X) = supr for those r for which y){r, P; X) = 0.
Then, for almost allr, 0 <,r < R(P), (10.4.9) Cs[(pr(r,P;X)rli^^) + 8. , ,^ ^ . ^ / p VN i \y-ir{i + hr) (pr(r, P; X) + e, where C3 denotes the constant of Theorem 10.2.4 and h is chosen so that (1 + rjo r)^"-^ < (1 + /^ ^) for 0 < r < Jô. Also y^(r,P;X)< g*(P)<Q(P), (10.4.10) (10.4.11)
(10.4.12)
^r{r,P;X)^^r{r,P;X) -^{ri,P;X) <y^(r2,P;X) and ^-"(1 + hryipir,

Q(P, X) <r
{riZ{P))) -y^{ri,P;X), 0<.ri<r2<R(P); P; X)

< R{P) ,
^{r2,P;X)
^-"(1 + hrY^(r,P;X)
^{r,P;X)^H[r -Q(P,X)]\
are non-decreasing; X = !;-'Q-" > 0 ; (10.4.13) w{r,P]X) <e + v-^\ -k)-^[i +Q(P,X)]S, 0<k<\. 0 <r ^kQ{P,X), Proof. For almost all r < R{P) (pr{r, P; X) = A'-^ [X 0 dB[P, r)]. From Theorem 10.2.5 with G = B{P, r), Ai = X H dB(P, r), and Yi any surface with b(Yi,Ai) D Hv_i(Ai). I t follows that [X B (P, r)] U Yi = Y is a surface with b{Y, A) D L. B y Theorem 10.2.4 we m a y take for Yi a surface with A^'iYi) < C3[(pr(r, P] X)fl(''-^). Thus if the first inequality in (10.4.9) did not hold, we would have A^(S) > A'(Yi) + A^'IS -B(P, r)] + > d(A, L) + s {S = X-A). The other inequality follows similarly since we m a y also take Yi = C [C (0, ^10)] where to is a normal coordinate system with domain O ^ ( 0 , RQ) such that a>(0) = P , ^10 = co~i(î). Then we would arrive at t h e same contradiction, since in this case A'{Yi) < ( 1 +7yo^)M'[C(0,^10)] = ( 1 <(1 +r}or)^''-^v-'^r(pr{r,P;X), +r^^ry'v-'^rA'-'^{Aô)
as is seen from our assumptions (10.1.13), Theorem 10.2.1, and Lemma 10.2.1(c). The results in (10.4-10) follow from (10.4-5) and (IO.4.8).
423
From (10.4.5) and (10.4.9), it follows t h a t (10.4.14) ^(r,P;Z)<| {riZ{P))
from which (10.4.12) and the first inequality in (IO.4.11) follow. To prove (10.4.13), we assume 0 < ^ < 1 , 0<r<ks, S<Q{P,X) and use (10.4.9) to obtain ^(r, P ; X) <ip(k s, P ; X) < - - ^ - / % - ! ^(1 +ht) < [1 +hQ{P, X)] r - M l - k)-^(p[s. P)X) cpr{t, P ; X) dt-^ e + e
< e + r - i ( 1 - k)-"^ [1 + Q{P. X)] S from which (IO.4.I3) follows. To prove the second inequality in (10.4.11), we define X(r. P ; X)=Jipr(t^
0
P ; X) dt,
co(r, P ; X) =ip{r, P ; X) - f (r, P ; J^).
From (10.4.10) and the definitions it follows t h a t co(r2) -win) >j[q>r{r, P, X) -y>r(r, P, X)] dr.
ri
0<n<n
<R(P)-
But since co is singular, i.e. co' (r) = 0 a.e., it follows that co is non-decreasing. From (10.4.9) and (10.4.10) and the definitions, we obtain y^[h(r)]-îp(r,P;X) A [h{r)]-ix{r. > 0, h{r) =yvr^\ +hr)''', or
P; X) > A [h(r)]-^co(r, P ; X) = - {[h (r)]-^y co{r,P; X) (r ? Z2 (P)) d[h(r)]~i
[h{r2)]-^X(r2) - [h{ri)r^x{ri)
> -fco(r,
P ; X)
h~^ (^2) CO (^2) h-^ (ri) CO (ri) f h'^ (r) dco (r) {co{r) =o(r,P;X))
from which the desired inequality follows, since co is non-decreasing. Now, suppose t h a t ^ C S ^ and t h a t {Xn} is a minimizing sequence; O i.e. b{Xn, A) Z^ L and A^[Xn A) d{A, L) + en for each n, where 7i -> 0. Let {^^} be a countable dense subset of 9fi. There is a subsequence, still called {X^, such t h a t the functions -y; (r, Qf.Xrij converge for each i
424
and each r, 0 < r <R(Qi) to functions ^ (r, Qi), Next, we define (for t h a t subsequence {Xn}) y)+ {r, Q) = Hm s u p ^ (r, Qi), ip~ {r, Q) = h m inf ^ (;', Qi) 9^" i^y Q) = 1 sup99 (f, Q]Xn), (p- {r, Q) = hm inf99 {r, Q;Xn).
Then we have the following results: Theorem 10.4.1. (a) (p^ and ip^ are all non-decreasing in r and y)'^{r, P) is upper-semicontinuous in (r, P). (b) y)-{r, P) < lim miy)(r, P\Xn) <. lim sup^(r, P\Xn) <. y)+{r, P ) . (c) f- (^2, P) > yj+ {ri, P) whenever 0 ^ ri <i r^ <i R (P). Thus yj+ (r, P) ip- (r, P) whenever either y)+ or ip~ is continuous in r. (d) (p-(r, P) < ^+(r, P) < \p+[r, P), 0 < r < R{P). (e) ^+(V', P') <y)-(r, P) if r' <R{P') and r' + d[P, P') < r <R{P). (f) For each P in ffl^ - ^ , [h(r)]-îp (r,) and [h{r)]-^ (p{r, P) are all non-decresing in r and y)(P) = lim [h(r)]-i ^+ (r, P)
r->0+
(h{r) = y^ r"" {\ + h r)-")
Upper-semicontinuous. (g) There is a positive number ^ such that tp (P) > ^ whenever ip (P) > 0, so this latter set is closed in'^ A, (h) (p-{r, P) >ip{P) ' h[r), 0<r< R(P). (i) If y}(P)=0, then y)+(r,P) = 0 for 0 <r <Q{P) for some positive number Q ( P ) . (j) / / we define S as the set of P in W A for which ip(P) > 0, then S is bounded and XQ = S U A is compact. (k) An altered sequence, still called [Xri], exists, which has the same functions \p^, for which the new functions 'cp^ are still non-decreasing in r and satisfy the conditions in (a), (f), and (h), and for which D {Xn, XQ) -> 0. (1) IfP^S,0<,r<: R(P) and {B(Pi, ri)} is a finite or countable disjoint family in which each Pi^S and each B{Pi, ri) C B{P, r), then
i
is
Proof, (a) If / i and f^ are continuous and non-decreasing on an interval I, then the functions msix[fi(x), f2(x)] and min[/i(:\;),/2(:\:)] are easily seen to be non-decreasing. This result is easily extended to imply that the sup and inf of a countable family of non-decreasing functions (continuous or not) are non-decreasing. Thus (a) follows easily. The proof of (b) uses the same ideas as that of (c). So we prove (c). Choose ^ > 0 so 4^ < ^2 ^1- We have ip-(r2, P) > infv^(/, Qi),
r'>r2Q d(P,Qi)<Q
y)+(ri, P) < supip{/\

r"<ri+e d{P,Qj)<Q
Qj).
425
B u t now, if r' >r2 g, d{P, ft) < Q, r" <ri ^ ( f t . Qj) < 2 e a n d
+ Q, d{P, Qj) < Q, then ft-),
y^{r\ Qi) >ip[r' - d{Qi, Qj), ft] îp(r2 - }Q, QJ) >ip[r",
using (10.4.6) and a limit process. Parts (d) and (e) follow from (10.4-5) and(10.4.6) and limit processes. TliatA~^99 and/^~i^ are non-decreasing follows from (10.4.11) and the argument in (a). To prove (h), we first note that (pr{r,P\Xn) '>vr-^{\ +hr)-^[f(r,P;Xn) Sn]. By integrating with respect to r between TQ and r, 0 < ro < ^, and then letting ^ -> 00 first and then TQ - ^ 0 , we obtain
(p-(r,P)>:fvt-^(\+h
0
t)-^ yr [t, P) dt
from which the result follows, using parts (f), (b), and (c). To prove (g) and (i), we consider two cases: Case 1 l i m ^ ( P , Xn) = 0. Then it follows from (10.4.12) that (p-{r,P)^^yvr'', 0<r<R(P), ^ = Ky-\ lim Q ( P , Xn) = Q(P) > 0 as n runs through a sub-
Case 2
W>o >o sequence. In t h a t case we see from (10.4.13) by letting n-ôa that y)~ (r, P) = 0 for 0 < f < Q(P) SO that y){P) -= 0. To prove (j), let us suppose that S is not bounded. Then there exists a disjoint countable family of balls B{Pi, RQ) with Pt^S and B(Pi, RQ) n A empty for each i. But then d{A,L)= limA^{Sn) > lim inf Z^' l^n 0 B{Pi, RQ)] > lim inf 2 " y> {Ro, Pi ,X)>Z <p- (Ro. Pi) > ^ N h(Ro)
for every N. This is impossible. To prove (k), choose for each p a finite number of regular domains of class C2 whose union Gp contains (XQ, \j2p) and is contained in (Xo, \lp). Let G^ be the set (G^, Qp) where Qp is so small t h a t G^ and rp = G'^ Gp are of class C\ and define Up on Fp by defining Up [x) = t for x^dGpt the parallel surface *7 of the way" from dGp to dG'^. Then Up^ Cî^p) and \J Up^ 0 there. Clearly it follows from part (i) and the compactness of Fp that /l*'(Pp Pi Xn) -> 0. Thus, for each py we may choose an ftp and a ^^, 0 < ^^ < 1, so thatyl""! [dGpt^ C\ Xn^ < \lp^-'^. From Theorem 10.2.4 it follows that the part of Xn^ outside Gptj, can be replaced by a set Yp such that b(Yp,Ap) D Hv_i(Ap), Ap = dGpt^ n Xn^, Yp C {Ap, C i ^ - i ) , and A^Yp) < C^p-r Clearly also, if P ^ 5, there are points of the new (and of the previous) Xn^ converging to P as ^ ^cxD. The result follows.
426
To prove (/), we note t h a t ip+ {r, P) > lim s u p ^ (r, P;Xn)> lim inf/l' [Xn D B (P, r)]
> Urn inf 2 " A' [Xn O B {Pi, n)] > Z {lim iniA" [X^ D B (Pi, n)]] from which the result follows. Lemma 10.4.3 (cf. REIFENBERG [1], Lemma 5*). Suppose Xn, XQ, S, and all the various other functions have their significance above. Suppose p '> V >}. For each ri'> 0, there are positive numbers e (rj, ^, p, v, W), r\ (rj, ^, p, V, W), and A(r], /5, p, v, W) with the following property. If P' ^ 5 , 0<r'<R[P'), and K[P', r') = S r\ B(P\ r') C {U', ri'r') for some geodesic p-plane U' centered at P' and if P < [h(r)]-^(p-(r, P) < [h{r)]-'îp-^{r, P) < ^ + e' [h[r) =y^r'[\ + hr)-'') for every B(P, r) C B{P\ r') for which P^S and 0 <r <R{P), then there exists a P * in S such that K(P*, 1 r') C [11, rj X r') for some p \ dimensional geodesic plane U centered at P * and B(P*, X r') C B{P', r'). In fact we may take e, rj', and X all of the form C{^, p,v,W) rj^, (10.4.15)
h = h{p,p,v,m).
Proof. We first show that there is an R' = c{p,v,Wl)r\ c > 0, such t h a t there is SL Q in LP such that B [Q, R') C B [P', r'jl) and K(Q, R') is empty. If this were not so, every sphere B[Q,R') with QÛ' would contain a point Q' of S and so every sphere B{Q, 2R') would contain a sphere B{Q\R'). But there are at most Zi{p,m) - (r'jRy disjoint spheres B(Qi, 2R') C B(P',r'l2) (QtÎI'), so t h a t (recall Theorem 10.4.1(1) and the fact that i?(P) < RQ for every P^m~ A).
ip + 8') h{r'l2) ^ ^+ g , P') > 2^W-{2R\ Qi) > Zi. [^Y ' p h{2R').
i
Thus, if we take e < /5, we conclude that (10.4.16) (r'IRy-'' < 21-2*' Z^;! (1 + 2hRy. Now, let B(Q, R) be the largest sphere in B(P\ r'l2) with Q on 11' such that K{Q,R) is empty. Let P^S ndB{Q,R) and let PQ be a point on Q P produced beyond P and let ^ = | P P o | - Let In ^ ln(P, r) ^ Xn D dB{P,r), let Ce be the geodesic half-cone consisting of the geodesic rays through P which make an angle < 0 at P with the geodesic QPPQ, and let Ine^ lne{P,r) ^ ln(P,r) f) Ce. For almost all r and 6 (10.4.17) AnC(PoJn)]Â-[C(PoJne)]+AnC(PoJn-lne)]^ Let T be a normal coordinate patch centered at P . Then /.^..ox (10.4.18) lneC:B(Po,g), ln-lneCm-B(Q,R), r-Hln6)CB{Po,Qo), Ql = r^+ ln-lneCB(Po,a) x^-2rxcosd.
427
From Lemma 10.1.3, it follows that B[T-^(Q'), Rjl] CT-^[B(Q, R)] C B[T~^(Q"), 2R] where Q' and Q'' are on the ray P ^ at respective distances Rl2 and 2R from P (this assumes that RQ is small enough). Accordingly, if we let Q and a be the smallest radii satisfying (10.4.18), we must have ^2 < (1 + Zi r)2 (y2 j ^ x^ ^2rx cosO), (10.4.19) (72 < (1 + zi rf (r2 -\- %^ -\- 2r^ xjR) 0-2 > (1 + zi r)-^(r^ + x^), Zi < 2rjo ii x <, r{i2 l) on account of Lemma 10.1.3. From the minimizing property of the Xn, we conclude, as in the proof of Lemma 10.4.2, that
A^{Kn) < / t ^ [C(Po, In)] + Sn. Kn ^ Kn[P. f) ^ Xn f^ B{P, r ) ,
A^{Kn) < v-ife(1 + hQ) A^'-^lne) + a(i + ha) A'-^{ln - ke)] + en. Solving for A^-'^[lne), we obtain
(10.4.20) yl'-l(4e) < ( ( T - ^ ) - l [ o r ( 1 +ha)A'-^{ln) - V A^ {Kn) + V Sn].
Now, let us choose ro, ri, x, and cos^ as follows: MO A 21^ 0 < r o = (ro/i^)i/i2; ' = min[^,(ro/ie)]. We shall consider only values of r with
^1 < ^ < ^0.
Then, using (10.4.19) and (10.4.21), we obtain (using ]/l y < 1 db y/2, etc.)
(T2 - ^2 > _ [(I _!_ ^^ ;.)2 _ (1 4_ 7^ ;.)-2] (^2 _^ ;^2) _|_
+ 2r x{\ + Z i r ) 2 c o s e > ( 1 + Z i r ) [AZir{r^ + x^) + + 2r :v COS0] > (1 + Zi r) 2r ro{{rolR)^^f^^ - 2Zi ro [1 + (ro/i?)]} (Zi < 27^0) ( + ^<(1+Zi.).[2 + |-^cos0 + 5](J<^cos0) T <(1 +Zir).2r[l+i(Vi^)ii/6]. . . ( - ^ > i ro (ro/i^)ii/i2 if 2fjQR^ h T and /-o/i^ < Ao.
Now, since yl" (X^) - s^ < T^-I r (1 + /^ ^) A"-^ (In) and yl*' (iC) > P h[r) by Lemma 10.4.2, we may use the inequalities above to conclude that (if TQIR < ^0, etc.) 0â{\ +ha) / t ' ' - i ( y -vA'(Kn) + + vsn(a<.r(\+Zir)(\ + 2rolR) < 2ro) ,.^ . ... < ^ ( 1 + Ziro) (1 + 2ro/i^) (1 + 2hro)A--Hln) (10.4.22) _ ^ ^ ^ ^ ^ . ( ^ j^kr)-' + ven < ^0 [(1 + 3^0/i^) ^''"MM - /^ ^'W] + ^^n, {h{r) =y^f^{\ + Ay)-").
428
Using (10.4.22), integrating (10.4.20), and letting n -^^^, lim s u p / y l ' ' - i ( 4 e ) dr<.{a~P + ^h(n)lh{ro)]
we obtain -
Q)-^ roh{ro) [(1 + }rolR) (^ + a)
< Z2(ro/i^)i/i2 h(ro)
since r > 3 so h(ri)lh{ro) < Z3 {roIR)- Using a similar procedure we obtain lim sup A^'iKn (P, ro) H Ce] < lim sup A^ [Kn (P, ^0)] - lim inf A" [Kn (P, ro) n (2)1 - Cg)]
To
(10.4.23) < (i^ + e') h (ro) - lim inf / A""^ [In) dr +
ro
ri
+ lim sup JA''-^ n
(In d) dr + lim sup J yl'-i (4) dr 0
<Z4(ro/ie)i/i2A(ro). Let ITi be the geodesic (AT' 1)-plane through PQP. Let A = ro/2y'. Then (10.4.24) r o / / - 2A < ro/i^ < (ro/f') (r'lR') < Z5(i^, r, 501) A. We now show that if A, or ro/i^, is chosen small enough, then K (P, roll) = K(P, Xr') ~SnB(P, Xr') C (77i, T^ A//2). For suppose P'^^KIP, roll) and is at a distance > rjrolA = rj Xr'll from 77i. Let P^ = T ~ I ( P * ) , (Jo = '^~^(Q')> nio =r-^(IIi), and Ceo = r-^(Ce). Since r is centered at P , T-i(P) = 0, T-i [ 5 ( P , f)] =B(0, r) if r < iô, /Tîo is an actual (A^ 1)-plane through 0, and Ceo is an actual half-cone whose generators make an angle 6 with the normal OPQO to i7i. Poo = 't~^(Po)' From the definition of Q' given just below equation (10.4.18) it follows t h a t B(QQ,RI2) is tangent to Uio at the origin. We have remarked above and can also conclude from Lemma 10.1.3 t h a t B(QQ,RI1)CT-^[B(Q,R)]. Finally P * is at a distance Q from 77io where (Lemma 10.1.4) (1 + r]o roll)-^ fj rolA < ^ < (1 + ô ^ 2 ) r] fo/4. Since P * is not mB(Q,R), P^ is not in B(QQ, Rll) and so, since it is in B (0, roll), it must be on the same side of Uio as is Ceo if ô/^ is small enough (<rjl(i +7joRI6)). By passing a 2-plane through 0, Poo, and P ^ , we see t h a t the distance from P^ to Ceo is d = Q sec(p sin(0 (p) > (f = angle Poo 0 P * . It is easy to see that d takes on its minimum with Q given when P ^ $ 5 5 ( 0 , roll). For t h a t value of (p. d = Q sin0 COS0 ]/(rg/4) ^^.
10.4. The minimizing sequenze; the minimizing set
429
Using Lemma 10.1.4 again, we see that B[P*,{r] TQ }rocosd) 16] GB(P, TO) n Ce which implies (see (10.4.23)) that Ph[{r) ro - }ro cos6>)/6] < Z^h{ro) {roIR)^/^^ which is false if ro/R is small enough. Now, suppose K(P', r') C {U', rj' r'), let P be the (geodesic) projection of P on IT', and let 77* be the geodesic ^-plane through P such that T-i(77*) is parallel to 77^, the tangent plane at PQ = T~^{P) to r-îU'). Since B[P,Xr')(ZB{P',r') it follows that K[P,}.r') C.{n',rj'r') so that d[P,P) <,rj' r\ From Lemmas 10.1.4 and 10.1.5 we conclude that
T-1[K(P,AO]C(T-I(/IX(1 +r]o^r')rj'r')
C ( T I ; , (1 +rjo^ r') rj' r' + K21^ r'^) C (/7o*, 2{\ +r]o?i r') rj' r' + K^ l^ r'^). Consequently K(P, Xr') C {IT'^, }rj'r') if X is small enough. Since d (P, P)ld{Q P) < ri' r'lR' <.ZQr]' < 1/2, it follows t h a t 77* and 77i are almost orthogonal so that K (P, A r') C (77, rj X r') if ^'<Â/4 and ? ^ ' < 1/2Z6, and 11 = 11'^ Dili. The fact, that e' ty', and A m a y be taken in the form C(p,p,v, W) r]^ follows from an inspection of the proof. Theorem 10.4.2. (cf. REIFENBERG [1], Lemma 6*). For each | > 0, there exists an so(S, ^,v,M) > 0 and a v{$, p,v,Wl) > 0 such that if P'^S and r' <:R [P') and (10.4.25) p < [h{r)]-'^cp-{r, P) < [A(r)]-iy;+(r, P) < / 5 + 0 for every (r, P) with P^S and B(P,r) G B{P\r'), then to each such B (P, r) will correspond a P*^S and a geodesic v-plane IT through P * such that (10.4.26) B(P*,vr)CB{P,r) and K(P^, v r) C {IT, ^ v r). Moreover so and v may he taken to have the form C {^, v, W) f ^,
h = h(^,v,m).
Proof. K(P,r) lies a t a distance 0 from the whole of W. Then apply Lemma 10.4.3Lemma 10.4.4. Suppose G is open and does not intersect A and suppose S C] G is not empty and y = the inf 'y;(P) for P^ S O G. Then if SQ > 0, there exists a sphere B (Pi, ri) such that W'-(r>P)^(y + eo)h(r), P^S, r<R{P), B{P,r) C B(Pi,ri).
(10.4.27)
Moreover, for all B{P, r) ^B{P,r)CG, that q)-{r, P) ^y P^ S, and r <R{P), ' h{r). it follows
430
Proof. Choose P i in S fl G so t h a t 'if (Pi) <y + eo/4, choose ^2 > 0 so that B(Pi, 72} C G and y)+(r2, Pi) <(y + eojl) h(r2), and choose n , 0 < /'I < r2y so that (y + o) h{r2 - ri) >{y + o/2) h{r2). Then, if for some B(P,r) in B{Pi,ri), with P^S and r<R{P), (10.4.27) does not hold it follows from Theorem IO.4.I that (since r < ri
-lî^l)
{y + eo/2) h{r2) > ip+{r2, Pi) '>ip~{r + r2 n, P) > ip+{P, ^2 n) > (r + eo) h{r2 - ri) >{y + eo/2) ^(^2) which is impossible. The second statement follows from Theoremi0.4.1(h). Theorem 10.4.3. v^(P) > 1 for all P^S. Proof. Let ^ = iniy)(P) for P ^S. Choose f > 0 and define s and v as in Theorem 10.4.2. There is a sphere B{Pi,ri) with P i ^ S and n < it! (Pi) such that (IO.4.25) holds. We may apply Theorem 10.4.2 to conclude t h a t each B{P, r) G B(Pi, ri) with P^S contains a point P'^^S and that H a (geodesic) i^-plane 77 through P * so that (10.4.26) holds. We shall assume that the minimizing sequence, called {Xn}, satisfies the additional condition in Theorem IO.4.I (k). Then, for large n and fixed P and r Kn (P*, vr)C.(n,2^vr) (10.4.28) (A- [Kn(P^
(10.4.29)
, A' [Kn (P*, v yjl)] >^'h(v etc.).

h(vr).
r/2)
(p(vr, P * , Xn) <(p + 2o) ' h(vr), r)] = y)(r, P * ; Xn), <,qn ^vr such that
Hence, for each n, there is a Qn, vrjl
A'--^ [Xn n a ^ ( P * , Qn)] < 2 ( ^ + 2o) (?^ ^ ) - l
Let us choose a normal coordinate system r centered at P * and define
yô - T-i [x^ n ÎP*7^)], Ano = T-1 [x^ n a^(P*, ^^)],

//Q A^ =r-Hn), =rA^Q, rnO=fn[C(no,Ano)], Yn='TYnO, An=^TAnQ, A^ = r(A'^^ A^^^fnTzAnO-rnoOno, In^=TInO, ^nO ^=ÂnOÂ^Q,
where n denotes the projection (in RN) into 77o, a n d / ^ denotes the radial projection onto dB(0,Qn) outside B{o,Qn'^\ 25^2) and the expansion f r o m P ( o , Qn^\ 2512) o n t o P ( 0 , ^^) otherwise. Since Kn(P'^,vr) C(n,2ivr),it follows that Yno C (77o, 2(1 +riovr)Svr)C (10 4 30) (/7o, S^Qn). From Lemma 10.2.2 and Theorem 10.2.2, it follows that A''(rn)<Z,{[i+2eo)-h(vr)-i,
if we assume that 1 + rjoQn < 5/4.
431
Now, suppose t h a t 7i(Yno) doesn't contain the point QnÛo nB{o,Qnp - 2 5 f 2 ) , let Qn=fn{Qn)^nonB{0,Qn), and let gn{Q) for Q ^ Q'^ be the radial projection from Q'^ onto dB(0,Qn)' Let 92^ = gnfnTt and let Zno = (fniyno)- We note also t h a t A^Q = (pni^no)In the next paragraph, we shall apply Theorems 10.3.9 and 10.3.10 with /.^ . ..^ ^l = ZnO, (10.4.31) ^2=^w0, X = ZnoUrnO, AÂnO From this, it follows Ai = A^Q , A2 = AnoU A*Q,
to show t h a t b{ZnoU Fno, Ano) D b(Yno, ^no)from Theorem 10.3.6 that (10.4.32) b{ZnUrn,An)Db(Yn,An),
Zn = T{Zno)-
From (10.4.32) it follows that we may replace Yn by Zn U Fn and obtain another surface in (^(A, L). But, since Zô C H^ D dB(0, Qn), we see that A^Zno) = A''{Zn) = 0 and so A-[Zn U Fn) = AîFn) <. Zî^ + 2o) ' h[v r) ' ^ A^[{Xn Yn) UZnU Fn] < d{A, L) + En - 2-"-! ^'h{vr) 2fo) h{v r)'^< d(A, L) + + Z2'(p+
if I and 0 are sufficiently small and n is sufficiently large. Thus, for such I , 0, and n, n{YnQ) D FTQ D B{O, Qn ]/l 25|2), so /l''(Yo)>r.^:(l-25|2)-W^ A-{Yn) > r . ^ ; ( 1 - 25|2)-W2(i _|_ ô?;r)-^ The theorem then follows from the arbitrariness of | and SQ by letting n -^<=>o through a subsequence so Qn -^Q and then using the arbitrariness of r. We notice first, using Theorem 10.3.12 that b(I X Xno,Dno) D{h'\ K (h) Dno = (0 X Ao) U (1 X Ao) where 7o and Ii are defined in that theorem. Then, using Theorem 10.3.6 with F:(I X Ano, Dno) ^ ( A o , Ao U A*), where Fn(i,Q)=fn[{^-t)Q we see t h a t b(rno.AoDA*o)D{k\k=Jni^(h)-Jno*{h). Jnl(Q)={<Pn\Ano){Q), Jno{Q) = Q , <?40, QAo. heHr_i(Ano)} Jnl{Q) AnoU A^. Jo(Q)ÂnoUA*^. = L2, + tn{Q)], QÂo,
Also, using the same theorem with F replaced by (p, we obtain b{Xi. Ai) = b(Zno, A*^) D (fn I Ano)* [b(Yn, Ano)] = Lx.
432
In order t o prove (10.4-32), we conclude from Theorems 10.3-9 and 10.3.10 that it is sufficient to show that [b(YnO,Ano)] (10.4-33) C i{Ano U ^*o> ^*o)*(9^^ I ^no) [b(Yno, Ano)] + L2,
or /wO*[^(^7iO, ^wo)] C Jnl^[h[YnQ,AnQ)] + L^.
But if h^h{Yno,
An^) (C Hvî[Ano)), we see that
so that (10.4.33), and hence (IO.4.32) holds. Theorem 10.4.4. The set X minimizes A^{X' A) among all X' ^^{Ay L). Moreover ip (P) 1 almost everywhere on S and xp-{r,P) -:yl''[5n5(P,r)] for all r <C R (P) not in a countable set. Proof. We have seen, using Theorems 10.4-1 (k) and Theorem 10.3-16, that Xo^^(A, L). For each ^ > 0, we can find a disjoint family {B (Pi, ri)}, where 5 ^^ < ^, Pi S> and B {Pi, ri) f] A is empty for each i, such that (10.4.34) S(Z\JB{Pi,ri)\J \jB[Pi,3ri), k= Q,\,2,...
(recall Lemma 10.2.3). From Theorem 10.4-3, we obtain
l r . ^ ^ < ( 1 +hQy2:h[ri)^{\
i=l i=\
00
+ hQYZw-{n,Pi)
i^l
<{^ +hQyZ
(10.4-35) '^'
lim inf ^(r^-, P^; Xn)

* 00 * n i=l
<{\
+hQy
Hm inf 2 ' Win. Pi', Xn)

n
<(i +h QY \imA''(Xn) = (i + h QY d(A, L).

From (10.4-35), we see that the series y^rj + . . . converges. From this and (10.4-34) we see t h a t (10.4.36) Al[S) <Zr'rl<(\+hQYd{A,L). The first result follows by letting ^ - ^ 0 in (10.4-36). By replacing 5 by 5 H P (P, r) and requiring the B {Pi, ri) C B (P, r), we conclude b y the argument above t h a t (10.4.37) A'[Sr\B (P, r)] < ^ - {r, P) < lim inf A' [Xn H B (P, r)].
n
If, now, P^S, r<R{P), and A'[S 0 dB{P,r)] = 0 and A^[Sn ndB{P, r)] = 0 for every n, we find by repeating the argument for the set S - B{P,r) t h a t (10.4.38) A''[S - B (P, r)] < lim inf A' [Sn - B (P, r)].
433
Since A^'iXn A) -Â^'iXo A) = /l*'(5), the equality must hold in (10.4.37) and (10.4.38). Consequently ip(P) =lim{y,r^)-Â'[SnB(p,r)],
r->0
P^S.
Suppose AîZjc) > 0 and choose s with 0 < e < A''{Zjc)l2k. Cover Zjc by an open set G so t h a t A^ [G H (S Zjc)] < a. Again we may find a disjoint family {B(Pi, rf)} in which P ^ $ Zjc, ri < R(Pi), 5n <Q, and B(Pi, n) C G such t h a t (10.4-34) holds with S replaced by Zjc. Clearly the B{Pi, Yi) cover all of Zjc except possibly a set ^]c with /1''(CA;) = 0Then A^[Zi,) + 8 > i A - [ S n 5 ( P ^ , n ) ] > (1 +hQ)-^{\ >(l+^-l)(l+A^)-4(Z;,). Since this holds for each e > 0 and ^ > 0, it follows that /l" (Z^) > (1 + k-^) A^'iZjc), so t h a t A^'iZjc) = 0. Since this is true for each ^, the result follows from Theorem IO.4.3. Our aim now is to show that S = XQ A satisfies Reifenberg's {e, Ri) condition at a n y point PQ where ip{Po) < 1 + 0 where eo depends only on v, N, W, and s (of course e, 0 > 0). Definition 10.4.1. A locally compact set 5 is said to satisfy Reifenberg's (e, Ri) condition at PQ <^ there is a number P i > 0 5 to each point P $ P ( P o , 2Pi) and each P R) n P ( P , P)] 0, there is associated a positive number RQ, with PQ < Po, which has the following property: If 0 < / < P Q , 0 <Q^d<.do,andB {Q, Q) C B (P, r) and P i B {Q, g), then ArHS)<{\+e)ArHS'), S = B{Q,Q)ndB{PJ), S' = S n n n being any geodesic v-plane centered at P and containing Q. Also (10.4.39) l^'-^\nr\B{Q, Q) n a P ( P , Ol dt -AnnnB{Q,Q)nB{P,ri)].
= A^[nnB(Q,Q)nB(P,r2)]
Remark. lfTl = RN, Ar^{S) = A^^S') without restriction on r. Proof. The last statement is evident since IT O B(Q,Q) f) dB(P, t) moves perpendicular to itself on 30^ as ^ varies. To prove the first statement, let C be a normal coordinate system centered at P with -2:-^ axis O through QQ = co-^{Q). Let us write coordinates of points on the z^ axis in the form {a, 0). Let 77 > 0 and choose PQ to satisfy the condition
434
of Lemma 10.1.3 for ^ instead oi sirjisto be chosen later. Let So = co^^ (S), SQ = co-i(S'), and i7o = co~^(n). Then ITo is a r-plane through the z^' axis and 5Q = 5o Pi UQ. Since ^ < (5, it is clear t h a t (10.4.40) (10.4.41) Ar' (S) = yr.i [Q ( S ) ] - i , Ar' {S') = r._i fc (S')?" ^ where ^(S) was defined in Lemma 10.1.4. From that lemma, we obtain Q(S) < (1 + rjor) ^(So), S-^- = UB[c Q{S',) < (1 + rjor) Q(S') If we write Qo (c, 0), c > ^, we see from Lemma 10.1.3 that S^ C 5o C S+, + kQ, 0; (1 - TJ) Q] n dB(0, t) +ri) Q\C\ B[c-riQ,0;{\ + rf) Q]. S+ = dB{0,t)r\B[c (10.4.42) + riQ,0;{\
From (10.4.41), it is possible to conclude, for ri small enough, that Q[S^] <Q(5o) <.Q{St), Q{S^)^[\+Zri)Q[S'^). The desired result follows from this, (10.4-40), (10.4-41) and proper choices of 7] and RQ . Lemma 10.4.6. Suppose A*{X) < 00 and > 0, suppose RQ satisfies the conditions of Lemma 10.4.5> <^^^ suppose that, r <,RQ, 0 > 0, le > 0, and 0 < i{\ -i-rjodo) < 1 cos0. Suppose that for arbitrarily small ^ < ^0 that there exists a disjoint family of halls {B (Pf, ri)} such that B{Pi,ri)CB(P,r), h{ri) Â'[K[Pi,ri)-] P^B{Pi, n<d <.Mh[ri), (for each ^) ^r^K
i
Q{Si)<.{\+Zri)Q(So-)
e(So- n/7o) = e(5o-) < e ( s ' ) <.Q{st n TIO) = e(5o+)
> ^0
ri) for some i, Pi and P^X
for each i
K[Pi, n) C (i7^, f ri) (K{Q, Q)^XnB{Q, Q)) where ITi is a geodesic v-plane centered at Pi and making an angle > 0 there with the geodesic P Pi. Then cp[r,P)X) < ( 1 + e)[\ +r]or)^tp{r,P;X) -2-^(\ +r]or)-^[\ -cosO -i(\ +>7oô)?'^0. Proof. Let us cover K(P,r) (10.4.43) gj<d, UK{Pi,ri) by a countable family
{^ {Qh Qj)} of balls C B (P, r) and such t h a t

0
2:y^Q'i^^nK{P,r)-UK{Pi,ri)]+d.
i
Now, using Lemma 10.4.5 with 11 j centered at P and passing through Qj, we obtain
(10.4.44)
/ Ar^ [B (Qi, Qi) n 1{P. t)] dt^(\+ < (1 + e) (1 + riorYy.Q'i
s) A' [77; n B (Qi, Q))] t)).
(1{P. t) = X n dB(P,
435
If P iB{Pi, n), we note first that B{Pi, n) 0 l[P, t) is empty for t < Si, where (10.4.45) Si = d(PPi) -ncosO ~rii(\ -\-rjod). This expression for Si is arrived at by choosing a geodesic coordinate system centered at Pi and using the result of Lemma 10.1.4 in this system. Using the same idea, we conclude that B [Pi, tj) C B {Pi, Vi) r\B[P ,t) for a suitable point P^ on PPi, where / . . . .^^ (10.4.46) 2^z =^Si-{-IriO,
r
{\PPi\ - n) =-ri[\t>Si.
cose - f (1 + riod)] +
Thus, if 77* is centered at P and passes through Pi, we conclude that jAr"- [B{Pi, n) n 1{P. t)]dt<(i 0 + e)x
X / / I ' - i [/7f n B {Pi n) n / (P, <)] << ^ S i (10.4.47) = (1 + e) yl- [ilf n B (-Pi. n)] - (1 + e) ^'([ilf n B (P,, n) n 5 ( P , s)] < (1 + e) (1 + rjQr)'Y,r\ - ( 1 + e)(1 +J7o>')-''(T''r.rf using (10.4.46) and Lemma 10.1.4. Since the B{Pi, ti) are disjoint, at most one contains P and (10.4.48) / A'f^ [B {Pio, n^) n I (P, t)]df< A' [K {Pi,, ri,)] <Mh (no). 0 Adding the results in (10.4.44), (10.4-47), and (10.4-48) and using (10.4-43), we obtain (since y, rj < (1 -{- hSy h (vi)) jAr' [l{P,t)]dt<(i+e)(i+rjo ry A^ \K [P, r) - U K {Pi, n)] + +hdyA^\UK{Pi,ri)] -
+ (1 + ) {^+rjory{\
- (1 + e) (1 + Vo yy o^ ee from which the lemma follows by letting ^ -^ 0. Definition 10.4.2. Suppose ZT is a geodesic _/>-plane centered a t P and suppose f/ is a set C B{P, RQ). We define the projection pn{U) of U on n as follows: Let r be a normal coordinate system centered at P. Then we define pn{S) =T^pon^r-^{S) where po denotes the ordinary orthogonal projection into r~'^{n) (which is an actual ^-plane in Rjsf). From Lemma 10.1.1(b), it follows that pn{S) is independent of the particular r chosen.
28*
436
The method of proof of Theorem 10.4.3 leads to the following lemma: Lemma 10.4.7. Suppose that X ^ S ( ^ , L),A''{X A) < d{A, L) + rj, Q < RQ, B(P, Q) D A is empty, JJ is a geodesic v-plane centered at P , K(P, Q) =XnB{P, Q), 1{P, Q)=Xr\ dB[P, Q), and K(P,Q)C{n,iQ), A-[_K[P,Q)]>:^h{Q), Z|<l/2, Z==l+7yo^, 71 + C i | ^ ( 1 +rioQ)^'{\ -ZHYÎÂ^-^[l(P,Q)]<ph{Q)
Ci being the constant of Theorem 10.2.2. Then Pn[K[P, Q)] Dnr)B{P, Q l/l - Z 2 f 2 ) . Proof. For, otherwise Y = K{P,Q) U 1{P, Q) could be replaced by FU Z, constructed as in the proof of Theorem 10.4-3, in which yl" (Z) = 0 and
/i" (r u z) + /t^(r) < Ci I e (1 + 7^0 ^)2' (1 - Z21^2)-W2 x

XA^-^[1(P,Q)]<A^{Y) -7J.
Theorem 10.4.5. There are positive constants ci, C2, h'> 1, and RQ with the following property: Suppose XQ is a minimizing set as in Theorem 10.4.4 and suppose PQ^S ^ XQ A. Then, if 0<r2<.Ro, (10.4.49) B{P,r)c:B(Po.r2), S satisfies an (e*, Ri) condition at Pofor any Ri < ^2/32. / / jp (Po) < 1 + 2, there is an ^2 such that ^2 < RQ and r2 <. R (Po) such that the second line holds in (10.4.49). Proof. The proof of the last statement is like t h a t of Lemma 10.4.4. Suppose 2 > 0, I > 0, and RQ is chosen so t h a t Lemma 10.4.5 holds with e = 1/4, say. We suppose that f2 < PQ- From Theorem 10.4.2, it follows that if 2 < 2{v, Wl), there is a t* > 0 such that each B[P, r) C B (Po, r2) contains a point P * in 5 through which passes a geodesic I'-plane ^ ( P * ) , centered at P * , such t h a t B[P^,vr) P'^S, C P ( P , r), K(P^, v r) ~ S D B{P^,vr) Qj^K{P\ri4), C (2{P'')> ^ ^ ^ ) j=\,...,v + 2. Suppose, now, that P(P',/)CP(Po,^2), We suppose that 0 < ^ < (5o, where A ^ ( 1 + 7 y o ^ o ) < (16/15)2. From Lemma 10.2.3, it follows t h a t we m a y find a disjoint family of balls {B {Pi, ri)} such that K{r, r'l4) C U K(Pi, Qi) U U K{Pi, SQi), PiS, 5ei<d. ^1, 2, . . . r2<R (Po), ^2 < oi *, 2 < ^2 (e*)^, and + 2^2) for all h{r) < y)+{r, P) <h(r){\
B(Pi.5ei)CB{P',r'j4).
437
Since A^[K(Pi,Qi)] '>h{Qi), it follows that the series of /['-measures converges. Hence we may choose a finite subfamily such that Â'K{Pi,Qi)>\A'K{P',r'lA) (10.4.50) Zh{Qi) > 2-1(1 + 2e2)-Â'K{P, r'jA)
' > 2 - i ( l + 22)-iA(//4). For each / < v + 2 and each i, B {Pi, Qi) C B {P\ r'lA) C B (ft-, / / 2 ) C B {P\ / ) . Thus we have a set of balls {B[Pfy v QI)} such that Pi"^ S and there exists a geodesic r-plane ^ (Pf) centered at Pf such that K[Pf, Qi) C iZ(Pf), SvQi), Now suppose 0 > 0 and suppose B {Pf, VQt)GB {Pi, Qi).
(10.4.51) (1-cos6>) : - 3 A f , X=\+rjodo. Let Oj be the set of i 5 the geodesic Pf Qj makes an angle > 0 at P f with ^ ( P f ) . Suppose for some j and some // > 0 that, for arbitrarily small d, we would have
(10.4.52) 2JyvV-Qlêe=jLcv^(r'l4Y,
ieOj
JLOO.
Then, from Lemma 10.4.6 and (10.4.51), it would follow that (10.4.53) (p{r'l2,Qj) < v ^ ( / / 2 , ft) -eeX'^'{\ ^p[r'l2, ft-) - 99(//2, ft-) < 2^2(1 + i) (1 + ^ o / / 2 ) - ^ +hy'l2)-'y,[y'l2Y hr'liy. But, from (10.4-49), we have which contradicts (10.4-53) if (10.4.54) 2 < 2-i-'/i(A?;|)'(1 + i) (1 + co/2/4)-v(i +
Thus t h e reverse inequality must hold in (10.4-52) for all sufficiently small d. Since there are only (r + 2) ft-, we m a y choose a ^ so small that there is an io which is not in any Cj, provided merely that [ji is chosen small enough so t h a t (10.4.55) [v + 2)fi< 2-1(1 + 22)-Ml +^r'lA)-''yv\ this follows b y comparing (10.4-52) (with the inequality reversed) and (10.4.50). Thus for some i^ the geodesic Pf^ ft- makes an angle < 0 at P^* with ^{Pf^ for eachy. Thus, using this fact, (10.4-51), and the proof of Lemma 10.1.5 (the curvature of arcs corresponding to geodesies), we have shown that if ft, . - -, ft+2 are any {v + 2) points in B[P\ r'jA), there is a geodesic r-plane ^ centered at a point of K{P', r'jA) such t h a t each ft- is within a distance (10.4.56) ^ '^[svae + \K, g ) ] ^ ^ [ | [U^l^ + \K, (-)] ^ ar'jl
otZ(=I{Pt).
438
Let T be a geodesic coordinate system with domain B (0, RQ) for which T(0) = P'. Choose (?*, . . ., (?*+! in K{P',r'l4) so that yl''(zl*) is a maximum, J * being the r-simplex in R^^ having the r~^ [Qf) as vertices. Let Q be any other point of K{P', r'jA), let ^ be the geodesic 7^-plane of the preceding paragraph determined by Q and the Qf, suppose ^ is centered at P{^K[P', r'jA))y and suppose ^ o is the tangent r-plane in Rjsf to T~^ ( ^ ) at T~i (P). Then it is easy to see, using Lemmas 10.1.4 and 10.1.5 that the distances from ^ o of T ~ ^ ( 0 and the r~'^{Qf) are all < ( 1 + ^0 ^74) cr//2 +7^2/2/16 = a ' / / 4 , o'' = 2(1 +ô//4)(T + i^2//4. Let AQ be the [v + 1)-simplex having Zl* as one face and T~'^[Q) as opposite vertex. From our construction it follows that the A* measure of each face of Zlo is < JD = yl*'(Zl*). Consequently the A^ measure of the projection of Zio on ^ o is < (r + 2) D and hence (10.4.57)
^*'+Mô) < 2 Z ) ( T ' / / 4 .
Since Zl* is one face of Zlo, it follows that the distance of r~^ {Q) from the r-plane of zl* is < (r + 2) a' r'l2. Since Q is any point of K[P', r'l4), it follows that T"i {P') = 0 is at a distance < (r + 2) a' r'j2 from that plane. Thus T ~ ^ ( 0 is at a distance < (i^ + 2) o*'/ from the plane through 0 parallel to that of Zl*. The image of this plane under r is a geodesic v-plane ^ through P ' and we have shown that (10.4.58) i^(P',//4)C(2^,'//4) provided that (10.4.59) 4(1 + ^0^74) [v + 2) or' < '. Now, from (10.4.49) and (10.4.58), it follows that there exists an x between ;'78 and ^74 such that ^' ZL"-! [/ (P', x)]<{\ A'[K(P', +2e2)h [r'jA) or Zl^-i [I {P\ x)] < Zi x-^ h {x), x)] >h{x), K[P\ x) G (2;, 2e x).
Since Xo is minimizing, it follows from Lemma 10.4.7 that the projection of K[P', x) on 2^ contains 2^r\B (P', x j/l 4^2 e'2)^ 7 being the number (near 1) mentioned there, provided merely that f is small. We now show that 5 satisfies an (e*, R{) condition at PQ if (10.4.60) Pi <f2/32. Let P^K(Po, 2 Pi) and R < Pi. Then B (P, P) C P (Po, ^2/8). If we take P' = P and / = 8P, then B(P\ r') C B(PQ, r2) and we may take 2 ' ( P , P) = 2'- Using (10.4.58) we find that K{P, P) = K(P\ //8) C (Z (P> ^ ) . 2e' P).
10.5- The local topological disc property
439
From the preceding paragraph we conclude t h a t the projection of K{P, R) on^ = ^{P,R)z:>2^r\ B{P, R ]/i - 4Z2 e'^), so t h a t
Z{P, R) n B{P, R) C[K{P, R), 2e' R + R{i - ] / i 4 ^ 2 s'^)]
(10.4.61)
[K(P,R),'}s'R]
(l - ]/l - 4^2 e'2) < '. The cases where P = PQ are taken care of in a similar way b y setting P ' = Po, / = \6Ri. Now, b y following back through equations (10.4-61), (10.4-59), (10.4.58.), (10.4.57), (10.4.56), and b y choosing do and RQ SO that A = 1 +770(9o< (16/15)2, we see t h a t it is sufficient to have (v + 2) (100fi/2 + 105i^2 ^ 8 ) < *, and from Theorem 10.4-2, we may take 2 of the form Z ^^. 10.5. The local topological disc property In this section, we prove what is probaly the most interesting theorem
in R E I F E N B E R G ' S first paper:
7yoi^;'<1,
Ro<Ro>
Theorem 10.5.1. There are positive numbers EQ SO{N,W) and RQ = RQ (N, W) < RQ such that if 0 <, s <, SQ and if the locally compact set S satisfies the {e, Ri) condition at PQ for some positive number Ri < RQ, then a neighborhood of PQ on S is a topological disc. Definition 10.5.1. If ^ and ^ ' are r-planes in R^ we define where ^ and ^ ' are the respectively parallel r-planes through 0. If c and c' are orthogonal matrices, we define \c-c'\=\Z\ci-'cim^ {c = {ci),etc.).
We notice the following facts: Lemma 10.5.1. (a) If c is any orthogonal matrix, | c | = ^N. (b) / / Si and S2 are compact sets, P is a point [in W), R, rj, r\ > 0, and SiHB [P, R) C (52, r)') f) B {P, R)> then (5i, ri)nB [P, R) C (S2, r) + r|')f^B (P, R). (*^) U2 <^^^ .Z" ^'^^ v-planes through P in RN and if D [^ H B (P, R), 2;'nB (P, R)] =rjR, then do {Z> I') = V(d) If c and c' are orthogonal matrices and ^c ^^^ 2c' ^^^ v-planes in RN spanned respectively b y ci, . . ., c^ and 'ci, . . ., 'cv, then do{2e,Zc')<\o-c'\ {c = (4), c^ = {cl,...,c^).
440
Proof. Parts (a), (b), and (c) are easily verified by the reader. To prove (d), we first note that if ^ c and ^ g , pass through 0, thenô(^c> 2^) i^ just the maximum distance from^c of any point of 2Jc' ^ dB{0, 1). Let us now write (10.5.1) c^ = cj + 2eôjc.
k=l
A point P will be on ^Jc^ C) dB{0, A) <^
(10.5.2)
o^p^;^^xic]
9= 1
^tVo
3^1
L^.
+ il^î o^ '4,
P= l J
I'% I = 1.
If we let d = d ('xp) denote the distance of the corresponding point P from ^, we see that
(10.5.3)
d^ = ;^{xlY,
^p = 2 ' 4 H '
^ > ^ + l.
From (10.5.1) and (IO.5.3), we obtain (by taking the max.)

j=l fc=v+l 1,k=l
from which the result follows. Lemma 10.5.2 (Extension lemma), (a) There is a constant C^{^) > 1 with the following property: Suppose {j)i\ is a finite set in 3JI, r > 0, > > 0, ^ the fi are vectors in some RQ and we have d{Pi,p,,)>r if i'î \fifi,\<.rj if i'î and \pi pi.\ <.6r.
i
Then there is a vector function f (into RQ) ^ C^ on U B{pi, 5^/2) such that fiPi) = fi /<^^ ^^^^ ^' \^f{p)\^C^Vlr, |V2/(^)| <C4-77/r2, pÛB(Pi,Srl2), If \fi I < M for all i, we may take f so that \f{p) \ < M. (b) There exist constants C5 (M) > 1 and di (M) > 0 such that the extension theorem of part {a) holds for orthogonal matrix functions provided that T) -< di (M) and we replace C4 by C5. Proof, (a) Let ^(5) ^ CîRi) with (p'(s) < 0, 9?(s) = 1 for s < 0, and ^ (5) 0 for s > 1, and define
(10.5.4)
It follows that
h{p>) = ^[
j^
^J,
i
Mp)=Up)n[\-%[p)'\,
i+i
k{p)=k{p)llMp)'
}^{p) = 1 if d(p, Pi) < rl4. k{p) = 0 if d{p, PJ) < r/4, j i- i, (10.5.5) k(p) = 0 if dip.Pi) > i3r/4, \v^MP)\ <î(5K) -*. Sk{p) = 'i, 2'V*Aj(^)=0, pÛB(Pi,5rl2), k = i,2.
10.5. The local topological disc property
441
The bound for |VA^(^)| does not depend on the number of points pi since the number of non-zero terms in the expressions for S7Xi(p) and VXi[p) for a given p is < the maximum number of pj in the ball B{p, 13W4). Then we define (10.5.6) mÎfi'kiP),
i
p^S=UB(Pi,Srl2).
Now, suppose p^S and callpo the nearest pi a n d / o = f(po). Then since 2! V 2.i{p) ^ 0, we have Vm - K f i i
/o) V kiP) = 2:{fi-
d(2>.2>i)'>13r/4
/o) V
k(p).
Since po is the nearest pi to p, we have d(p, po) < 5 ^/2 so that the number of terms in the latter sum is again < maximum number of pj in B(p, 13^/4) and this depends only on 30^. Clearly, if d(p,pi) < 13;'/4, then ^{po,pi) < 6^ so that each \fi ~ / o | < ' ^ . The first result follows and the last statement follows from (10.5.6). To prove (b), we merely extend each of the vectors Cj as in (a). HT] is sufficiently small, the ci, . . ., cjsf as extended are linearly independent on 5 and may be replaced by an orthonormal sequence by applying the Gram-Schmidt process. Definition 10.5.2. Given an n.o. set c = (ci,..., C;^) and a r - p l a n e ^ in jRiv- Let 11 be an {N ~ r)-plane J_ ^ and suppose t h a t cf, . . ., c* are the respective projections of ci, . . ., Cv on ^ and c^+i, . . ., c^ are those of Cvî, . . ., Cjsf on U. For those (c, ^) for which c*, . . ., c* span ^ and c*+i, . . ., Cj^ span i7, we define where c^, . . ., c^ and c^+i, . . ., c^ are obtained from the respective sets cf, . . ., c* and c*+i,. . ., c^ by the Gram-Schmidt process. Definition 10.5.3. A finite set (^^}is onr-setfor asetS ^ d(pi, p^,) > r if ^'^ 9^ ^' and 5 C UB{pi, r). Lemma 10.5.3. There are numbers d^lW) > 0 and CQ{^) > 1 SWC/J that \A{c.Z) -A{c', 2 " ) i < C6[|c - c'l + <ô(2'.2")]. provided that
Ic c'I < ^2, ô(X m ^ ^2, ô(X -Z'c) < ^2, ô(2",2';')<^2
ZÂ^;'^ ^ c OLnd ^ ^ , <^re if/^^ v-planes through 0 spanned hy (ci, . . ., Cy) a^t^
( q , . . ., c^). In fact, A is analytic in (c, ^ ) zÂ^;'^ ^'^ is defined. Proof. This last statement means that if we suppose that c" = A [c, ^ ) , where ^ is spanned by the vectors c^\ . . ., cJ', then the vectors c/ are analytic functions of the components of the vectors (ci, . . ., cjsf, c{', . . .,
442
4"). Since the domain of definition of these functions is open and contains the compact subset of vector sequence (ci, . . ., c^, c^\ , c'J') in which ci, - . ., Civ is an n.o. set and c^' z= Cj, j = 1, . . ., v, the lemma follows. Lemma 10.5.4. There exists a positive number RQ < RQ with the following property. Suppose S is a locally compact set on ^ , YQ^W, S C B(YO,RQ), and r is a normal coordinate system centered at YQ with domain B(0, RQ) {andr(0) = YQ). Then (i) if S satisfies an {s, Ri) condition at a point P and B (P, 3 Ri) C B{YQ, RQ), then r~^(S) satisfies an [e', iî/2) condition at r~^{P), where e' = 2e + 4KR; (ii) if So ^ T~^ (S) satisfies an (e, Ri) condition at a point PQ and B (PQ, 3iî) C B (0, RQ), then S satisfies an (s', iî/2) condition at P = T(PO). Proof. We choose RQ < RQ so t h a t :^d{p, q) < d[T{p), r{q)] < ]/2 d(p, q)
whenever p and q are in B (p, RQ) and T is a normal coordinate system. Suppose t h a t So satisfies an (e, Ri) condition at PQ and t h a t B {PQ, 3i?i) C -6(0, RQ). Then r[B(Po, 3i?i)] C B(Yo, RQ). Choose R, 0 < R < RxjZ, anAQ^B (P, Ri) where P = T (PQ) Then
SnB{Q,R)cSn r[B{Qo.Ry2)] n B(Q, R) (Qo = T-I(0)
= CrlZoiQo,
r[SonB{Qo.R]/2)]nB(Q.R) R]f^),sRy2\nB{Q.R)
(z{r[Zo{Qo.R]f2)].2sR}nB{Q,R) CiSiQ.R). = 2sR + 4KR^]nB(Q,R) Ry^)] (see [S(Q.R).e'R]
2^(Q,R) being the geodesic plane tangent at Q to r\2o{Qo, Lemma 10.1.5). Likewise, C r{i;o{Qo.Rf2) nB{QoRf2).
2'(a R) nB{Q, R) c r{T-HS(Q, R)] n BiQo.Rfi)] n B(Q, R)

2KR^} 0 B(Q, R) C r{So n B{Qo,R]f2), C{Sns'R)nB(Q,R), 2KR^ + 8R']f2} C\B{Q, R) 0<R<Ri!2.
The remainder of the proof of (ii) and that of (i) is similar. We now recall the developments of 10.1. We suppose that P o 501, that T is the associated coordinate system of class C^ and satisfying (10.1.1), and t h a t coop is a, normal coordinate system centered at p, t h e coop and r being related as in Lemma 10.1.1 (d) and satisfying the bounds
443
in our general conditions on 3R. The various Zi will depend only on v, N, and Wl, unless otherwise specified, and may denote different constants in different proofs. Lemma 10.5.5. There are positive numbers rjQ{Wl), RQ(TI\), Kz(Wj, and K4(W), with RQ < iô ^fid rjQ<, 1/2, with the following property: Suppose p and q^ B(0, Roj}), d{p, q) == ri < RQ, 0<r2<.Ro, ^^^ ^ {py 2Q) = ^3 ^ ^ 0 ' ^2* Suppose olso that ô is a v-plane Rj:^ and ^p = coop(2o)' ^^^ 2Q = o)oq (2Jo)- Then (10.57) D [Zv n B[p, r^), Z^nBip, rg)] < Ks{ri + r2) r^ + K^ r^. Proof. L e t ^ o p be the r-plane in Rjsi tangent a t ^ o = ^ôqiP) to ^ J ^ = o)Qq{Zp)' Clearly V 0)^^ coo^(O) differs from the identity matrix by a matrix of norm ^ Zid(p, q). Thus (10.5.8) d{2;op.Io)<=Z2ri, d(Po. ô) <. (^ + Von) rs. Let us let F = COQ^ [B(p, ^2)] and let ^Q be the i^-plane | | ^ o through po. Then, we conclude from (10.5.8) and Lemma 10.1.5, etc., t h a t
D[2:op nr,Zor\r]<Zznr^, D[Z' nr,2*0n rj <z^r^ D[Zopnr,z$^nr]<K2ri

I t follows that
D[ISvr)r,Zonr]<Zsnr2 + ^2^1 + 74^3.
The result follows since ojoq{2^^^ 0 F) = 21P H B{p, ^2), 0)0^(2^^0 0 F) = lQnB(P,r2). Definition 10.5.4. The geodesic r-planes ^ ^ and ^ ^ are said to be parallel with respect to r and the family {coop}. Lemma 10.5.6 (The c-lemma). There are positive numbers QO(M) < 1/2 and ds {W) such that to each pair of numbers {rji, Q)for which 0 < ^ < qî^) and 0 <,rji < dz (W) Q, there correspond positive numbers ei [rji, Q, Tl) and RQ' {fji, Q, W) such that if the locally compact set S satisfies an {e, Ri)condition at a point PQ, with 0 <i e <, si and 0 <C iî < RQ', and the coop have their significance above, then there exists a sequence {cj[p)] of orthogonal matrix functions with Cj^ C^(S'^, 3^;/2) such that (10.5.9) (10.5.10) (10.5.11) (i) \^cj{p)\^^7]irr^;
(ii) \cj(p) - Cj_i{p)\ < Ce(2e + ^1/4^) + (1 + Q)r]ilS; (iii) ifp^S^\ D[Sn B(p, 6rj), y > 1, and IHP) = a)op[2;i{p)]> ZKP) ZHP) n B{p, 6rj)] <rji rj. 6ro).
being the v-plane in RN through 0 spanned by Cji{p), . . ., Cjv(p), then Here, we define (10.5.12) ro = RilS, rj = roQ^, S* = S n B{Po, 2Ri -
444
Proof. We define CQ{P) as a constant matrix such t h a t coi, . . ., cor spans O)QP^{^). Since ^l[p) = ^^PO(2J)> ^^ follows from the (e, Ri) condition and Lemma 10.5.5 (with ri = d(p, PQ) <, \0ro, ^2 = 6^0, ^3 = ^(P>2) < 8 ? ' o ) that D[Sr\ + D[2;n B(p, 6ro), 2nB{p, 6ro)] + B{p, 6ro), ZîP) n B[p, 6ro)] < X3 l6ro 6^0 + K^ - Sero
which implies (iii) for y = 0, provided that (10.5.13) 96K2ro + SK4 e <7yi. Obviously (i) holds. Now suppose that Cj_i(p), / > 1, has been defined to satisfy (i) and (iii) above. We define Cj(p) as follows: Let {pjt} be an TJ set (Def. 10.5-3) for S' = Sr]B{Po,2Ri), For each i, select a geodesic r-plane ^ | = coop.^ (Zlo) through pji such that (10.5.14) D[Sn B(Pji, Srj), Zi n B{pî, 8r;)] < Ssr,-1. It follows t h a t Zl n B[Pn, Srj) C (5, 86 rj) D B{Pji, Srj) (10.5.15) C (S, Ssrj) n B{pji, 6rj_i) D B[pî, 8rj) C il^-HPji), Serj + rjirjî) H B{Pji, 8rj).
Applying COQ^.^ and Lemma IO.5.I to (IO.5.15), we find t h a t (10.5.16) doiZio^ ItHPji)) < (1 + 8 7^0^,-) {s + ml^Q) <2e 87^0^0 < 1, 28 + r]il4g<d2. + ml'^Q-
iZio = (ôpu{Zi)> etc.) provided that (10.5.17) so t h a t we must have (10.5.18)
If (10.5.18) holds, we may then define (10.5.19) (10.5.20) Cj-{Pji)=A[c^_i(Pji),Zio\' oj_i(Pj'i) = A[cj_i(pji), ZtrîPji)], From Lemma 10.5.3 and the fact t h a t
we conclude, using (10.5.16) also, t h a t (10.5.21) \c3{pH)-oj_i(Pji)\ <Ce(2s + rjil4Q).
1 We shall omit the bars over the various sets.
445
Using (10.5.14) and Lemma 10.5.5 with î < 6^^, r^ = 2rj, rs = SeTj, we find that if d(pji, pji') < 6^;, then 2^1 n B[Pji, 2rj) C (5, Serj) H 5 ( ^ n , 2^,-) C (S, 8erj) O ^ ( ^ ^ ^ 8^,-) 0 5 ( ^ ^ , 2;^;) C(2'l'.'l6^r;)n^(^^,2r,.) C (2'ls ^6^; + X3 I6r? + X4 8r,-)
and
^ f being the geodesic r-plane centered at pji and parallel to ^ j ' (with respect to {coop}). Proceeding as in (10.5.16), we obtain (10.5.23) (do 2'lo> Ih) < (^6 + SKi) s + I6i^3ry provided that (IO.5.17) holds. From (IO.5.9) fory 1, we find that (10.5.24) \cj_i(pji) cj__i{pji')\ < ^ ^ ^ i if d{pji,pji') < 6rj.
Therefore we conclude from Lemma 10.5.3 and equations (10.5.19), (10.5.23), and (10.5.24) t h a t (10.5.25) IcîPu) -Cj[Pn')\ < C 6 [ ^ ^ ^ i + \6s + %Kê+\6Kzr,^
provided that (10.5.26) ^ ^ ^ 1 + ^^^ + 8 i ^ 4 e + 16X3^0 < ^2.
If (10.5.26) holds we conclude from Lemma 10.5.2(b) that there exists an orthogonal matrix function of class C'^[S', '}rjl2) which is an extension of Cj as so far defined with (10.5.27) IV c^{p)\ < C5 Cî^^Qrii ^ 1 _i + 16 + 8X4 + l6i^3ro) rr^
if we choose (10.5.28) C5 C^{6Qrii + 3 2 0 + 1 6 0 ^ 4 ^ + 320X3^0) <>7i (C5, C 6 > 1 ) . We must now prove (iii). Let p^S^, Then there is a ^;^ such t h a t ^(^ A'^) ^ ^j- Using Lemma 10.5.5 with ri < r^, ^2 = 6^;, andr3 = Se r^, we find t h a t ( ^ | being {coo p}-parallel to ^ \ through p) (10.5.29) D[2\ (10.5.30) n ^ ( ^ 6r,-), Z\ n ^ ( ^ 6r,-)] < K4 88;^; + 42i^3^,?. 6rj), Znp) n B(p, 6r;)] ^rj^rj(\+6rjorj)^ Using Lemma 10.5.1(d) and (i), we conclude that D[2\^B{p,
Using (10.5.29), (10.5.30), (10.5.17) and (IO.5.28), we see t h a t (10.5.31) D[Zi n B{p, 6rj)^Z^(p) n B(p, 6rj)] < { ^ ^ i ^ / .
446
Then Sr\B{j>, 6rj) C S n B{pî, 7Yi) n B[p, 6rj) (10.5.32) ^ ^^^^ ^^^^ ^ ^^^ ^^^^ ^ (inp),nirj) n B(p, 6rj), In like manner, we see that (10.5.33) ZHP) n B{p, 6rj) C (2^1,{^^1 rj) H B{p, 6r^) 0 ^ t o ^ , Jr^) ^^^
Part (iii) follows from (IO.5.32) and (10.5.33). Since d{p,pji) <.-rj, we obtain, from (i) and (10.5-21),
(10.5.34) \cj{p) - Cj_i{p)\ < Ce{28 + rjil4Q) + (1 + ^) ml^ from which (ii) follows. All the conditions in equations (IO.5.I2), (10.5.17), (IO.5.I8), (IO.5.26), and (10.5.28) can be satisfied b y requiring Tyo iî < 1, I8C5 Ce ^ < 1, (960 + 48OK4) C5 CQ e^rji (^^5.35) rj^^d2Q. \20Kz C5 CQRi<.rji since Q < 1/2. Proof of Theorem 10.5.1. We suppose that the locally compact set S satisfies an (a, i?i)-condition at PQ and that Q, rj2, and 61 are positive numbers satisfying (10.5.36) 0<Q<Qo{m)<^, 0<i^2<~, O<0i<:7r/12. We suppose t h a t T and the family {coop} have been chosen to correspond to PQ as in Lemma 10.5.6 with cooPo(ô) = 2> 2ô being the i^-plane in Rx^ spanned by ei, . . ., ev. We suppose t h a t 0 < 171 < d3(W) Q, 0 <. e < i(>yi, Q, 9Jl), 0 < iî < i?o(î, ^, ^ ) , and that {cy(i^)} is a sequence of orthogonal matrix functions satisfying the conditions of Lemma 10.5.6, the Tj being defined in (10.5.12). We retain the notations of that lemma. For each j , we define (10.5.37) A; = r]2roZQ^> Sj = S"^ 0 B(PQ, 2RI - 6ro - Aj), ; > 0. for 5* of Lemma 10.5.6 and define
For eachy, we choose the rj-set{pji} the mappings aj^ C^(S'^, 3 r;/2) by (10.5.38) ajip) = coj^lrjjip)],
rjjip) = ZhiiJ>) S [ ^ ^ ; ( ^ ^ ) ? ^^> i ^="+1 cojpiy) =a}op [cj{p) y ] , ; > 1, where, for eachy, the ^ji{p) are defined by (10.5.4) in terms of the^^^. We then define the mappings TJ^ C^{TQ) and the Tj by ro = o)oPo\Tl (10.5.39) rjâjTjî, T,- = rjm) = Tg = I ' D H ^ ( 0 , 2i?i - 6ro - Ao) aj{T^_i),
We divide the remainder of the proof into several parts.
10.5. The local topological disc property
447
Remarks. To orient the reader, we notice that if 9K == RN cind SG 2 ! = ^Q, then we may take COOP, = I, cooz{y) =y + z, coQ^(y)=y z, Then, since the pji == Zji^^^
i
Cj{z)=I.
all the ^^^ = 0 if ^ > v, so that

k=v+l
N k=v+l V fc=l
k=v+l
since 2 ^Jii^) =" '^' Thus
aj(z) = z 2J zê}c = 2! zêjc and is the projection of 2: on 5. Part 1. Ifrji and hence e and Ri are small enough, we verify the following facts hy induction on j (i) d[aj{p), p] < ri2 r^ for p^ T^_i, ; > 1; (ii) TjC(Sj,r]srjî), rjs = rj^H < ^196, y > 0 (10.5.40) (iii) the angle at p between ^^(p) and Tj is < 0i if p^ Tj (iv) Uip, q) <d[aj{p), Ojiq)] < 2d(p, q) if
Proof of part 1. For j = 0, TQ = 2J C) B (PQ, 2Ri 6ro Ao) and So = S n B{Po, 2R1 6ro Ao) so (ii) follows provided merely that (10.5.41) SsQ-^<rj3, To prove (iii), we use the normal coordinate system a>oo = cooPo- Clearly (10.5.42) cOo-oMô) = T, coôniHP)] = coîmp{T'o)Thus O)OQ[Z^(P)] is part of the locus of the equation (10.5.43) w=U(y;p,Po), From 10.1, we conclude that yÎl
<:Zid[p,Po),
I l=v+l m=l I
\X\ = \^\ = 1
S that (iii) follows provided merely that O (10.5.45) Z2d{p,Po)êi. Suppose now that (10.5.40) is true for y 1. Suppose _^^ Tj_i. From (ii) for y 1, we conclude that there is a point p' ^ Sj C S* within a distance < rjs rj oip and hence one of the^^-^-, call it q, is within a distance (1 + ^3) ^j of P' From the (s, iî)-condition and the construction in Lemma 10.5.6, we see that (10.5.46) 5* n B{q, Srj) C (2'^'(^), Ssrj) H B{q, Sr^)
448
SO that all the pî in B {q, STJ) ^ (^^' {q),Se Vj) and (since p^{S*,
r}^ TJ))
(10.5.47)
PîU'iq)^ êrj + ri^fj).

îai^P) = ^> ^no[z) = hi{P)^ ^Jo{z) = Cj(p) _ N V tj{z) = co^'qâj[cojq{z)], tjo = z 2 zêjc = Z ^^ ^Jck=v+l fc=l
So we use the normal coordinate system cDjq and define (10.5.48)
We shall show in part 2 below that (10.5.49) i9i^)=i3o{z)+tf{z), tf {0) = 0, \^tf{z)\
From (10.5.47), it follows that the distance (10.5.50) d{z, Zo) < (1 + ^srj) (Ssrj + rjsrj) < 2r]sr^ \tj{z) -z\<{r]4 + 2rj3)rj if the last inequality holds. Thus \tjo(z) z\ = d{z, ^ g ) , so t h a t (10.5.51) ^ ^ d[p, (Tjip)] < (1 + Zsrj) {r)4 + 27^3) r^ <rj2rj
if the last inequality holds. Since d[aj(p), q] < (1 + ?^2 + ^3 + Se) r-; < Svj, we see from (10.5-46) and the construction t h a t d[(yj{p)> S*] <D[S^n (10.5.52) + d[a^{p),2nq)] B{q, Srj),2;nq) r\B[q, Sr^) + 2'g] < 8srj + (1 + Zsrj)d[tr{z),
< 8ry + (1 + ^3 rj) rj^ Yj < ^3 ^;+i (since tj{)[z) ^ ^ g ) provided that
(10.5.53)
r H ^ + ( i + ^3^0)^4] < ^ 3 .
Thus (i) and (ii) are verified. To verify (iii), we note first t h a t the angle at ^ b e t w e e n 2j^~'^{P) ^ ^ ^ Tj_i is < 01 so the angle at z between the tangent planes to coj'^ [^^~^ {p)] and o)Y^(Tj_i) at 2: is < 20i. Now, evidently, coj'q [2^^~^ {p)] is the locus of the equations (see (10.5.38) and 10.1) (10.5.54) w = a),-/co;_i,2,(y) = [oj{q)]-Û[c^_i{p) ei[cj_i(p) y',p, q], y^ZoFrom 10.1, we see first that U\m[cj_i(p) y;p, q]=di+ y;p, q], \8\ < Z^r^. Letting S/w and e denote the matrices dw^jdy^ and e^, respectively, we obtain Vw = fe-(^)]-i[/ + e\ cj_i(p) = [cj(q)]-Uj.i{P) + 1(1^11 <^5^;-) = ^+ [(0j{q)]-Hcj{p) - cj(q)] + fe-(^)]-i [c,'_i{p) - Cj(p)] + si. Since, on our locus coj'^ [2'^~-^{P)^' J ranges only over ^ g , it follows t h a t the angle between that locus and ^ g is (10.5.55) < ^6 \r^ + (W20) + C6(28 + 1^1/4^) + (1 + ^) ^ 8 ] < 6)1
449
if the last inequality holds. Thus the angle between 0)^^ (Tj_i) and ^ g is < 3 ^1. Now let dy be a vector tangent at z to co^^(Tj_i) and let dw he the corresponding vector tangent to o)J^{Tj) at tj[z). Let us write dy = dyi + dy2 where dyi is the projection of ^y on ^Q. Then dw = dyi + V tf (z) dy = dyi + dw"^, \dw*\ <?^4 \dy\ (^0.5.56) 2-1/2 | ^ y | <^cos}di'\dy\ <\dyi\ <\dy\ since 3 0i < jr/4 and the angle between ^ g and the tangent plane to cjoJg^(Tj_i) at ^ is < 3 ^ i - An easy argument, using (10.5.56) shows that 2-1 \dy\ < |dfz^| < 2 \dy\, so that (iv) holds, and the angle between ^ g and the tangent plane to coY^(Tj) at tj{z) is < arcsin rj4^ ]/2. Now, co,^^[^^*(^')] is seen, as in (10.5.54), to be the locus of the equations
(10.5.57)
t^ = [cj{q)rÛ[cj{p')y;p\q]^
y^Zl
P'-ÂP)-
Thus the angle b e t w e e n ^ g and the tangent plane at tj[z) to ooy^ [.21KP')\ is < Z7 rji so t h a t the angle at p' between T; and 2HP') i^ (10.5.58) < ( 1 + ZzTj) ' Z^rji + arcsmrj^]/! <ei if the last inequality holds. This proves (iii). Part 2. Proo/0/(10.5.49). F o r ^ and q fixed, we define u(y, z) = oyT^ o)jp(y) = cr^Ô) U [cjo{z) y\ cojq(z), cojq(0)] y := V(w, z) <^ w == u(y, z), v[w, z) = coJ-^[a)jq (w)] = cj^{z) U [cjQj{p) w] cojq(0), a)jq{z)] (see (10.5.38), (10.5.54), and 10.1). Then we see t h a t
N
rjjo(z) = ^ 2.jio(z) 2J '^^(^ji> ^) ek (10.5.59)

i k=v+l
( = m{P)) v(w,0)^w
^(yô)=y,
u(0,z)=z,
v{z,z)=0,
tj(z) = (oJ-^^Mjplrjjoiz)] = u[r]jo{z), z], Zji = oyJ^{pn)From (10.5.59) and the uniform bounds for the second derivatives of U(Y;p,q) with respect to its arguments, and from the bounds in (10.5.46) and Lemma 10.5.6 and the extension lemma (Lemma IO.5.2), we find t h a t r]jo(0) = 1 2 " ^^(0, 0) Â; = 0
fc=v + l N
(zji = 0 for Pji == q)

fc=v+l
VJo(z) = Z^*o(^)
i
2 [v^[zji, z) - v^{0, z)] eic + Z ^ ^ 0 , ^) eic

fc=v+l
^-iz'êu^
fc=v+l
vfo W ,
Vfo(^) = i { 5 | ( z ) + S k 3 i o ( ^ ) [î +
k=v+l i 29
450
dl(z) = [v^{z, z) v^(z, 0) v^{0, z) + v^{0, 0)], d\^,{z) = [v^ZJ^, z) - v^zji, i^^.l ^Zserj; Ve tj(z) =1+ 0) - v^0, z) + vÔ, 0)] ^y(O) = ^(0, 0) = 0, 1^21 <Ziirj < Z i 2 y y + Zi3e |V tyfo(^) I < ^9 ^ + ^lo^^ VzTijo {z) + e2{z), \Vtf{z)\
tj(z)=tjo(z)+tf(z),
which proves the result. Part 3. The maps tj, defined in (10.5.39), converge uniformly to a homeomorphism T from T^ onto a topological v-disc S (Z S*. From (10.5.39) and (10.5.40), it follows t h a t ^ ^ ; (y) > T^j-i (y)] = ^ [(^3 b^j-i (y)]. T^J-I (y)] < v^ ô ^* so that the TJ converge uniformly on 7^ to a map T and (IO.5.6O) d[rj{x), d[Tj(y),r{y)]<Pij(=rj2roZQ'')' riiy)] < 2Ay = 2r]2roQ^+^l{\ - Q).
Suppose, now, that T{X) T(y). Then, from (10.5.60), we obtain But then it follows from (10,5-40) (iv) t h a t d[Mx),ro(y)] ^7j2ro(2Qy+W -9) for every y. Thus x = y, since ^ < 1/2. Part 4. If e and R\ are small enough, S D K{PQ, 4 ô)- From Lemma 10.5.4, we conclude t h a t caô (5) satisfies an (e', R'^ condition at 0. By assuming Ri small enough, we may assume that CO^Q (5) satisfies an [e', i?i)-condition at 0, where we assume 36e'<1, ^<ô. We shall drop the prime on e and assume that all the sets are in RN ; we assume ^ ^ g and TQ T^ and use the simpler notation. In order to prove the result in this part, we select a y ^ S Pi -S (Po, 4ô) and define a continuous map cp with domain To as follows:
Since 99{x) is on the segment from x tor[x), \(p{x) r[x)\<^rQ,
't{x) if \r{x) y| < 2^0 X if \r{x) y| > 3^0

^^^ { [ k W y| 2^0] ^ + [3^0 \t[x) yl] T(:V)},- otherwise.
we see t h a t x[^ro, =^ Z ^ dB{Po, 2Ri y\ \'t[x) x\
\(p{x)
on account of (IO.5.6O). Accordingly if x^dTo 6ro Ao), then \x(x) y\= \\x(x) ~ x'\ -\- {x y)\'>\x
> iRx 7^0 4^0 ^0 = 4^0
451
S that (p{x) = X on ^TQ. Since \(p{x) r(x)\ < r o , it follows that if O \(p(x) y\ < ^0 then \r{x) y\ < 2ro so (p(x) =r(x). Thus (p{To)nB(y,ro)CS. Now, suppose y i S. Then, for some n, (10.5.61) (p(To) n B(y, 2-^ro) is empty. If this holds for n = 0, then B (y, TQ) 0 (p (TQ) is empty. But since (p (x) is between x and r (x) and T (:^) ^ 5* C ( ^ , 8 ro), it follows that (p (To) C ( X Seô) C (^/'^ l6ero), 2 " being n i ; through y. Since I6e < 1/2, it follows that if ji denotes the projection on ^, then 990 = ^ 9 is a continuous ^ map from TQ into 2J which is the identity on d TQ and whose range contains no points in B (yo, ro/2) Pi ^ C the disc TQ (yo = ny). But this is impossible. If (10.5.62) holds for some w > 0, then (p(To) n B{y, 2t) = (p{To f) [B(y, 2t) -B(y,t)], t = 2-^ro. But, if 2 ( y , 2t) denotes a r-plane through y whose existence is guaranteed by the (e, JRi)-condition, then (p(To) n B(y, 2t)CSn B(y, 2t) C (Z{y> 2^), 2t s) a[B^t - Bt) and this set may be mapped continuously onto a subset of (^(y, 2^), 2t E) ndB2t by a map con which is the identity outside and on dB2t' Thus (setting Br = B(y, r), etc.)
((On (f) (To) nB4t=^ {(p(To) n (^4^ ^2^)} U Vn-l
Vn-i = con{(p(To) n (^2^ - ^ C (Z(y^ 2t), 2t e) 0 dB^t. But now, using the (e, iî)-condition, we find that (Z(y> 2^), 2t e) n dB2t C (5, At s) 0 dB^t = (S, At s) H B^t O dB^t C(Z(y>4t),St8)ndB2t. (p(To) n B^t C (2(y, 4t), At e) n B^t for an appropriate ^(y, At). Since s is small, we see that that the set (con cp) (To) n B4t can be mapped continuously into a subset of (^(y, At), At e)ndB4t by a map co^_i which is the identity outside and on dB^f This procedure may be repeated and we conclude finally that (coi. . . con cp) (To) n B(y, ro) C (2/(^-^0), ^0 ) Pi dB(y, ro) for an appropriate 2^(y, ro). Now (2'(y> ^0), ro s) n î5(3;, ro) C (S, 2ro e) O B(Po, 8ro) 0 ^^(y, ro) C ( X Ôro e) n B(Po, Sro) H aJ5(y, ro) C (2", ^Sero) H aJ5(y, ro), 2J' being | | 2 through y. Since coi . . . co^j is the identity outside and on dB(y, ro), it follows that ncoi. . . (J0n(p has the. same property as the former mapping 990 = JT 99. Since this is impossible, 5 D 5 fl B(Po, 4ro).
29*
452
In order to prove further smoothness properties of these discs, it is necessary to consider their r-dimensional Lebesgue areas. Lemma 10.5.7. Let us suppose that S, S, the Tj, the TJ and T, the Cj, the pji, the 2^{p), and the positive numbers , Q, di, RQ, rji, and rj2 have their significance above and that e, Q, 6i, RQ, TJI and rj2 are sufficiently small. Then for each rj6'> 0, there is a d ^ 0 such that (10.5.62) hm miLv(rj | r ) < (1 + ê) ^*'[T(r)] if A^[r(dr)] = 0
provided that all the numbers above are <id, F being any open set for which r C, To, Lv denotes the Lebesgue area. Proof. For a fixed y and each i, we define Dji as the set of all p ^ (S*, ^rjl2) such that d[p,pji) < d{p,pjr) for all i' and let 2 ' ; ^ = SHPji)Each Dji is bounded by portions of ^ ( 5 * , }rjl2) and loci of equations of the form d(p, pji) = d(p, pj^). These latter loci are approximately planes which are approximately normal (depending on rji and rj) to the geodesic r-plane ^ji', we call these loci the projecting faces. Since {pji} is an rj-set for S*, B(pji, rjj2) G Dji. On the other hand since any p ^ Dji O SÛ B(pjfc, rj) with d{p, pji) < every d(p, pjjc), we obtain (10.5.63) B[Pji, rjjl) C Dji, Dji Pi 5 C B[Pji, rj).
From (10.5.63) and (IO.5.40) (iii), etc., it follows that Tj 0 Dji is almost parallel (depending on di, rj2, and rj) to 2Jji C\Dji. Since we have seen that the projecting faces are almost normal to 21H> we conclude that (10.5.64) L,{Tj n Dji) = A'(Tj n Dji) < (1 + r]e)îÂ'(Dji 0 Zn)
if Oi, Yji, rj2, and jRi, and hence e (if Q is chosen < ^0 (9K)) are small enough. Moreover if B(pji, rj) H 5 is not empty and B(pji, rj) f) dS is empty, it follows from Lemma 10.4-7 and Lemma 10.5.6 t h a t the projection of 5 n B(pji, rj) = the projection of 5 pl B{pji, rj) (as defined in Definition 10.4.2) D ZH n B{pji, rj 1/1 - Z 2 | ^ f / 3 6 ) Z being the constant in Lemma 10.4.7- Thus, since 2!H H B(pji, rj) D 2Jji n Dji and the projecting faces are nearly normal to 21ji, we conclude that (10.5.65) A^(S n Dji) > (1 + rje)-^'Â-(Dji n ^Ji)-
From (10.5.64) and (IO.5.65) we deduce t h a t (10.5.66) Lr{Tj n Dji) < (1 + ^6) A'{5 n Dji).
Let rt = rJ^{TjC\Dji), Fji = T-^(S D Dji). Since r and TJ are homeomorphisms, it follows that all the Fji, for those i for which /^^ H F
10.5' The local topological disc property
453
is not empty, lie in (F, Sj), where sj - ^ 0 . Thus, summing over such i and using (10.5.66), we obtain
i i
<{i+7je)A'{r(r,sj)}
from which the lemma follows by letting y - > o o . We wish to conclude this section with the following theorem: Theorem 10.5.2. Suppose that e, Q, 61, RQ, rji, and rj2 dye- fixed sufficiently small numbers, that S,S,r, and the tj have their previous significance, and that (10.5.62) holds for some fixed (finite) TJQ > 0. Then Lv{r I r) < y l ' ' [ T ( r ) ] , if T i s open and T c Tg. We shall prove this theorem after stating several definitions and proving several lemmas. The methods of proof are suggested by those of some deep results of F E D E R E R ([1], [2]). We present simplified proofs of some additional results, which we don't require here, in 10.8. Definition 10.5.4. Given a ^ Rjsi and > 0. We let C^(a, e) denote the totality of iV-cubes of the form \z^ a^ e d^) \ < e, where ^ is a sequence of even integers. We let C^_i(^, E) be the totality of the [N 1)-cells on the boundaries of the iV-cells of C^(^, e), and denote by C^-îa, e) the totality of [N 2)-cells on the boundaries of the [N 1)-cells of ^'N-I(^> ), etc. Similarly we let C'^{a, e) denote the totality of A/'-cells of the form \z^ a^ e6^\ < e where 0 is a sequence of odd integers, and let C^_i(<3!, e) denote the totality of {N 1)-cells on the boundaries of the cells of C^(a, e), etc. We sometimes allow C'j^(ci, s) or Cj^' (a, e) to denote the union of the point sets corresponding to the collections. If a == 0 and e = 1, we denote C^(0, 1) by C^, etc. Lemma 10.5.8. (a) yjcq is a cell of Q <^ it is of the form I-s:* (5Î < 1, i^sjcq and z^ = 6^ if i ? Sjcq sjcQ being some set of k distinct integers < N; here the d^ are even and the 6^ odd. The corresponding statement holds for the cells of each C^. (b) For each k,v-\- 1 < ^ < iV, and each q, the center ofy^g^ C^_^_i. If k = V + 1, yjcqC) C"^-!,-! is the center of yjcq. If k <^v + \, then ykqC\ C^_v-i ^'s empty. (c) / / we define rjc : C^ C^_y_i -> C^_i by defining it on each ykq C^_y-i as the radial projection from the center ofyjcq onto its boundary, then rjc(C^ C ^ - ^ . J C C;_i Q % - i , k>v + i. Proof, (a) is easily proved by induction downward. To prove (b) and (c) we may assume that yj^q = y^, where ^fc : |>^*| < 1 for i = \,. . .,k\ z^ = \ for k + i <,i <,N. Now if y^-v-i ^N-v-v YN-V-I li^s along a plane 2:* = d^, d^ even, for V -\- i values of i. If y ^ - ^ - i O y^. is not to be empty, these i must all be
454
< ^ . For each such set of i, there are 2^"*'-i cells of C^_y_i which intersect yjc, in which the z^ must satisfy 1^^' 0 ^ ' | < 1 , e^=^\, 1 < ; < ^ , y^^ any^'. ThusyJJ. n C^_y_i consists of all the (^ r 1)-cells for which 2:^+1 = =: z^ = 1, -2:*' = 0 f ori^ + 1 values of i < k, and 1 : * | < 1 for the remain2^ ing j . Parts (b) and (c) now follow easily. Definition 10.5.5. We define Qv : RN C';^-v-i -^RN by v = ^f+i 0 . . .or^ where the rjc are defined in Lemma IO.5.8 (c). We define B as the unit cube in Rj^, i.e. B == {y : \y^\ < \, i = \, . . ., N, and define I4v(y) = (\yh Sh\, . . ., \yU+i (3v+i|) where d is t h a t sequence of even integers such t h a t y d^B and (^'i, . . ., ^V) is a permutation of (1, . . ., iV) for which I yh _ ^h I < I /yi2 _ ^^2 I < * * * < | y% S^N \ . Given a and e, we define the maps co, ta, and r^ by 0}{y) = ay, ra(y) =y + a, ra = raOa)OQpO o)-^ 'r_a. Lemma 10.5.9. (a) If rj ^Ri C^_y_i, then Qvirj) ^ C^ a^^ ^*s 0^ the boundary of the N-cell of the form rj d^B in which rj lies. (b) J \u,(v - r])\-^ dA^ (rj) = C7 (r, N) A^ {B). (c) ^Tj^R^C^_,î ^ w.(i^) 7^ 0. (d) Ify = a-\-d-{-srj, where rj^B and 6 is (as usual) an sequence of even integers, then ra [y) is the point on C [a, e) given by ra{y) = a + ed + eQv{r]), and \Dra[y)\<.C^{v)'\u,{ri)\-^ where D denotes the maximum directional derivative. Proof, (a) This follows from Lemma 10.5.8 and its proof. To prove (b), we note that u is periodic of period 2 in each Y}'^. Thus we have j\nv[v-Yi)
B
N-
Qv(r))^dB,
Y dA^(7]) = f Iuvirj) \-'

B
dA^(r]).
The result follows by direct evaluation which is made easy by the symmetry, (c) follows from the definitions. To prove (d), we note that ra{y) = a + [d + Qv(r])'], From the homogeneity, we see that \Dyra(y)\ = \D^Q47])\. To calculate a bound for the latter, it is sufficient, because of the symmetry, to assume that 0 < ?^^ < T^^ < < iy-^ < 1. Then r]^B.
455
we must have rf^^ '> 0 ii rj i C^_y_i. In our case and the bound Cg \ Uv (rj) \~^ is easily obtained for \D o^(rj)\. Lemma 10.5.10. Suppose G is a hounded open set in Rp, r is a homeomorphism from G into W, A''\T:[G)] < O O , A'\T{dG)'\ = 0, r[G) C B (Po, R), R < RQ, and r = co o z where co is a normal coordinate system with domain B(0, R) and co (0) == PQ. Suppose also that there exists a sequence {r^}, each r^ = co o Zn^ C^{G), such that Xn ->t uniformly on G and such that, for some rj '> 0 UmZv (r^ I T) < (1 + rj) A' [r (P)], if T (ZG and A' [r (^P)] = 0. Let 9 : 9K ^ 501 he a piecewise smooth mapping with domain D T {G) such 9 that \D (p(p)\ <.%(p))y where % is continuous near r(G). Then Lr{(por)^(\+rj)f[x{p)rdA{e^),
r{G)
Proof. Let & be those cells r in Pjv such that A^lziG) H dr] = 0. For each or > 0, we may cover z {G) with a finite number of closed cells Ti of @ of diameter < cr, the open cells r^f^ being disjoint. For each i for which rt H z(dG) is empty, let Ft = z-^ {r^f^). Since z and co are homeomorphisms, the diameters of the P^ and of the T(P^) - > 0 as a->0. Moreover, for each such i, A^'lridTi)] = A^'lcoidrt D z{G)} = 0. For such i and all n
U ((pOTnl ri) =A'[(pO Xn (P*)] < IX {pi)r A- \Xn (P^)]
(10.5.67)
- \x[pi)'rU{rn I Ti) < (1 + 7y) [;^(^)]M'[T(P,)] + ^n.
where x(Pi) denotes the maximum value oi x(P) for ^ ^ T ( A ) . The result follows by adding the results in (10.5.67) for those i for which rf n z(dG) is empty, taking the lim inf and then letting cr -> 0. Lemma 10.5.11. Suppose G is a hounded open set in Rv, r is continuous on G, Lv (T) < 00, r{G) C B (Po, R), R < Po, ^'^d r = co 0 z where co is a normal coordinate system with domain B (0, R) and co (0) = PQ. Suppose Z is the union of a finite numher of{v \)-planes in R^, and Q = z~^ [Z O z(G)]. Then Lr{r\G-Q)=L,(r). Proof. It is sufficient to prove this for Z a single (v 1)-plane. Let cooo be the normal coordinate system related to co by a rotation of axes such that Z is part of the {N -]- \ v) plane z^ = z^, i = v, . . ., N. By approximations from the interior, we may also assume that G is the union of a finite number of domains of class C^ which have disjoint closures. Moreover, for each e and each o*, 0 < cr < 1, we can find an open set
456
r (ZG Q such that F consists of a finite number of domains of class C^ with disjoint closures and I v ( T | r ) > Z V ( T | G - ( ? ) - e/2, r[T{x)']<al2 for x^G-F,
r(p) being the geodesic distance from^ toW = (Ooo{Z), We may assume o) = coooi for each q^ B{Po, iô/3)> we let COQQ be the normal coordinate system centered at q and related to cooo as in 10.1. To points p near W we may assign coordinates w as follows: Let q = nw{p), the (geodesic) projection oi p on W\ then oy^Q (q) {w'^, . . ., w^-^, z^, . . ., z^). Then p must lie on t h e geodesic (iV + 1 r)-plane W through q so that p = a)Qq{0, . . .,0,w^, . . ., w^). We define oy^ hy p co* {w); clearly to* is of class C2 near Z. Now, we define a sequence {C;}, each C;$ C^(^) ^^^ ^ i ( ^ ) (i-^Lipschitz) such t h a t Cy -> T uniformly on G and (10.5.68) Lr(Cj\F)-^L,(r,F).
To do this, we first approximate uniformly t o z on F b y {C*}, each C* Ci(^)> so that (10.5.68) holds with (^y = co C* and then approximate uniformly t o ^ on G by any sequence {rjj}, each rjj^ ^^(G). Now, for each y, f * ^; I ^ can be extended to satisfy a Lipschitz condition on G, Let ^f be such an extension and define
w=
co^ ==
if \if{x)\
<2co,-
\-îf(x),
max 1 If
xeF (A;)
if |ff(^)|>2a>,-
Qj (^) = % (^) + f; (^) ^ > C; (:^) = C 0 ^; (:\;). O

Clearly QJ [X) = Cf (A;) on Z' and each QJ is LIPSCHITZ on G. Since so coy - > 0 and I;/ - > 0 , so that ly; | r - > ^ | r , it follows that ^f\F->0 Qj -> 2: (all convergence uniform) so [l,^] satisfies the conditions. Finally, let us define 99a : 9}l -> 9)1 by if r{p) < (T, if c 7 < ; ' ( ^ ) < 2 ^ , p' being the point A of the way from nw (P) along the geodesic to p where h == {^ or)~^ [^(P) o"]- With the aid of the co"^ coordinate system, it is easy to see that (fa is piecewise smooth and that \Dq)a(p)\ <.ioy %a = ^ const., lim;^cT = 1
a->0
^[ô(p), p]<.o,
1 I n case Cf iii\r
both for ^ ^ 9K.
=0 on JT, set | ; (;ir) = |*(;t^) = 0 on G.
457
We now define T; = (fa O Cj and we see that if j is large enough for I^j{P) ~ '^(P)\ ^ <^/2> then tjis L i p s c h i t z a n d r j ( x ) ^ W i o r a l l x ^ G F. Therefore, for such j ,
L,{Tj)^L4ri\r)<xlLr{Ci\r)
The lemma follows from (10.5.69) and (10.5.68) and the arbitrariness of a and e. Proof of the theorem. Since 5 minimizes A^ (X) among all compact X^{{dS,L), where L = Hv_i(dS) (Theorems IO.3.4 and 10.4.4), it follows t h a t ip{P) = 1 almost everywhere on 5. From Theorem IO.4.I, it follows t h a t ip is upper semicontinuous. Let rj > 0 and choose positive numbers Q, di, RQ, rji, r]2, and e so small t h a t the rjQ of Lemma 10.5-7 is < 7]. From Theorem 10.4-5 it follows t h a t we may take eo > 0 but so small t h a t if y^(P) < 1 + 0, then 5 satisfies an (e, JR2)-condition at P with R2<, RQ, if P is interior to 5. Let W be the subset of S dS where ^ ( P ) > 1 + o/2; then W is closed in S a 5 , AîW) = 0, and X = r-^{W) is closed in B(0, R) ~ (T^Y^\ R = 2Ri ~ 6ro - AQ. We assume t h a t Po^S and that 5 C B(Po, 2 Pi) and that coo 0 is a normal coordinate system with domain B(0,Ro) and cooo(O) = Po- We let Z = 0)^0 (W) z(X). Let G be any open subset of T^Y^K We cover each XQ oi G X by Si neighborhood "^ with "^ d G X such that there exists a sequence {tn}, each Zn ^ C^ {^) with Tn = coooo Zn, which converges uniformly to T | 9^, and for which (10.5-70) l i m Z . ( T ^ i r ) < ( 1 +fj)A^[r(r)], ^-^^ if
PC^,
Anr(dr)]
=0.
li xo^X, we take 9^ = G. Let F be any compact subset of G. We may cover z(F) by a finite number of closed cells rf with disjoint interiors such that A^ [z{G) H dri] == 0 and Vi f] z(dG) is empty for each i, each also being of diameter < a. Since r and z are homeomorphisms, we may take a so small that each of the sets Ff = z'^ (ri) is in some one neighborhood 9^, provided F^ f) F is not empty. So we let G' be the union of these Ff; clearly G' D P . It is sufficient to show that Zr(T, G') = A^[r(G')] for such G\ Let Y be the union of these dfi. The Fi have the property (10.5-70) possessed by the 9^. Let d be an arbitrary positive number and choose a in Rjsf with | a* | < h, where ^ > 0 but so small that f [Xo(z)rdA^(e,)<d, Vo={Y,SNh), \}q2(z) + 1, Xo = Xa,h C9 = Cg + 3-
Xo{z) = {C9 \u,[h-^z
- a)]|-i -
458
q2[z) = 1 on (y, ANh) and qîz) = 0 on RN - (Y, SNh) 0 < ^ 2 ( 2 ) < 1 on (Y,SNh) -{Y,4Nh), and q2 has Lipschitz constant (Nh)~^. To see that this choice is possible, we note that (using the notation of Lemma 10.5.9) / dA^(a) I [xo(z)rdAê^)
= h^ f dAN (b) I Cl\u, (/z-i z~h) \-^ dA' [e^) ^ VonziGn = A^ Q C7(v, N) A^^B) A'[Vo 0 z(G')] = A^ [CO (B)] 'C'Cl'AnVoDz (G')] using Lemma 10.5.9. Since /l*'+i [^ (C)] = 0, it follows also that yl^+i [P*^ 0 z(G')] = 0 for each (i^ + 1)-sequence / . Thus the A''-measure of the totality of (iV r 1)-planes of the form z^ == a^ which intersect the set z [G') is zero. Hence we may also choose a so that z [G') C RN C^_y_i [Uy h). Moreover by choosing a small enough, we may ensure that each Fi which intersects X is such that r^ C a neighborhood U of z[X 0 G'] such that (10.5.71) f^\xo{^)\'dAê^)<d,
UnriGn
Having chosen e and a as above, we define qi (z) as we defined q2 (z) except with 4N h and SN h replaced, respectively, by }N h and 4N h. Then we define
niy) = y + î(y) [^(y) y]> C = (poo z, r' -= a>oo o c.

We note that (foiy) =y ^^^^^^^ for y^RN (Y, 4iVh) l9ô(y) y| < A ' A , y^RN \U^)~-z{x)\<Nh -^\Dra{y)\ < 1^0(^)1, y^RN^
\D(p^{y)\ < 1 +{Nh)-^'Nh+\
Now, let Q be the collection of components P of G' ^-'^ [C^.^ {a, h)\ Then we conclude from Lemma 10.5.11 that (10.5.73) U[r'\G')=ZLv{T'\r).
If r does not intersect z-^(Y) nor X then it is part of one of the Ff C some ^ with "^G G X, and we conclude from Lemma 10.5.10 that (10.5.74) L,(r'\r)<L(\ + rj) f \xo(z)\^ dA^(e^), z{r)
10.6. The REIFENBERG cone inequality
459
since T' =(por, (p = COOO(POWQ^, \I^(p{P)\ < (1 + ^Vo î)^ \I^ (po[cooi(p)]\, piV, V = cooo{Vo), L==:(\ +2rjoRi)^^.
\D(p{p)\ = i,
If r does intersect z-^{Y), let xo^m z-^{Y). Then z(xo)^ Y, so that (by (10.5.72)) C(xo)^(y,Nh) and hence ^ a component r cube (of side 2h) E of Cl(a, h) C^_i (a, h). Let xi be any point of F and join XQ to xi by a path xu 0 <,t <,\. Ks long as C (^^) ^ (Y, 3iV A), t (^^) E, and as long as ^{xt) ^ E, it $ ( Y , '}Nh). Consequently ^[E] C E C.{Y,'i N h) and z {F) (Z [Y, AN h) G FQ. For such a F, we can conclude only that (10.5.75) L,{x' \F)^L(\+r]Q)f \xo(z) \' dA'(e,).
If F intersects X but not z-^(Y), then ^(J*) C C/ and we can conclude t h a t (10.5.75) holds. Substituting the results (10.574) and (10.575) in (10.573), we obtain LAr' I G') <(^+v) (10.576) ^ ^ ^ ( G ' ) ] + {2 + 7j+rje)'L -fjxoiz) VonziGn l^^dAîe^) +
+ (1 +V^)J\xo(z)'rdA'(e,)
Unz(Gn
< (1 + rje) A^riG')] + (3 + i? + 2rje) ^ L . max \r'{x) r(x)\ <,Nh. Since, for a given G, rj, and d, h may be arbitrarily small, we obtain (10.5.76) with T' replaced b y T. Then the result follows for G' from the arbitrariness of 7] and d. 10.6. The Reifenberg cone inequality
Suppose 2: is a polynomial A^-vector which defines a mapping from B (0, R) onto S, suppose t h a t ZQ = Po^ RN> and R(PQ) is the minimum value of \z(s) zo\ for s on dB(0, R). Let U(s) = \z(s)-zo\^; then t / is a polynomial. From recent results by W H I T N E Y [3], it follows t h a t t h e locus of Uôc(s) = 0, oc = \, . . .,v, has only finitely many components. From a theorem of A. P. MORSE [1], it follows that U is constant on each of these components. Thus, except for a finite number of values of r, 0 < / <R{Po), the locus l(r, PQ, Z) = S H dB(Po, r) consists of a finite number of regular orientable analytic manifolds without boundary. The remainder of this section is devoted t o the proof of the following theorem: Theorem 10.6.1 (The REIFENBERG cone inequality). The reexistpositive numbers QQ, eo, and k, k <, \, with the following property: If z, r, and
460
I (y, Po, z) have their significance above, r is not exceptional, and there is a v-plane ^ through PQ such that / l " - ! [/ (r, Po, ^)] < 1 (1 + o) yv r'-^, ^ D[l{r,Po,z),ZndB(Po,r)]<Qor then there is a set Y* with b(Y'^, A) D Hvî(A), A == l(r, PQ, Z), such that (10.6.2) A^{Y^) < (1 - ^) v-^rA^-^ [l(r, PQ, Z)] + ky^/^ Suppose we let Yr be the cone with vertex Po and base l{r, PQ, Z). Then the inequaHty (10.6.2) becomes (10.6.2') /["(Y*) < ( 1 -k)A^(Yr) +kyrrr On account of the homogeneity, we may assume that r = \. We may let SRi be the collection of regular analytic manifolds, without boundary, in B (0, R) which is the counterimage of / ( I , Po, z). We choose an (x, y) coordinate system with origin at Po and x"^, . . ., x^ axes i n ^ . We may then represent Yi parametrically on / x Tli, I [0, 1], by (10.6.3) xr=rl{p), y = rf]{p), 0 < r < l , \i\2 + \r]\^ = \, p^Wi. and
It is convenient to let Y denote the infinite cone obtained by extending the rays of Yi to infinity and to represent Y b y x = ri(p), y=:rrj{p), rÔ,
We may approximate to f and rj (or ^ and rj) uniformly together with their first derivatives b y piece wise analytic functions ^^ and rjn, each pair of which is such t h a t the cone Yn is composed of a finite number of "simplicial cones'', each of which is the union of all rays joining Po, hereafter called 0, to the points of a (v 1)-simplex in a (r 1)-plane not passing through 0. We call such a locus Yn a polyhedral cone. Our method of proof of the theorem will consist in constructing a set Y* as desired in the theorem for each Yn and then showing that the Y* tend to Y* in the point-set sense, in such a smooth way that yl*'(Y*) ->/l*'(Y*). We may assume that each simplicial cone in each Yn has a 1 1 projection on the ;^;-plane ^ . On a polyhedral cone Y, we shall denote by x{P) and y{P) the x and y coordinates of the point P on Y. We have seen that each of the manifolds composing Y is orientable; we suppose that each such manifold is oriented. This induces an orientation on each simplicial cone. If P is interior to such a simplicial cone, we define y (P) = 1 according as the simplicial cone and its projection on ^ are similarly or oppositely oriented. On. any simplicial cone of Y, we may consider the y^ as linear functions of the x^ and we define \Jy{P) ^ as the sum of the squares of
461
the Jacobians of all orders < min(r, N v) of the y^ with respect to the xK The area element on Y at P is then (10.6.5) where
|Â;(P)
]/^ + \Jy(P)\^d2:{P),
N-
d2:(P) = \dx(P)\
| is the v-area element on ^ . We also define \Dy[P)\^= W ^ ^ ' ^ ^ ^ 2J \bx0j

?=i fc=i
(10.6.6)
= / Vl + \Jv{P) f dZ(P) - y. = A'{Y^)

Yi
where Yi is the part of Y in the cylinder |^(P)| < 1^. Comparing this notation with that in (10.6.4), we see that (10.6.7) x{P)=ri(p), y(P)=r7](P). We shall prove the theorem for all Q and e which are sufficiently small, Q being a bound for | y (P) \ on Yi; we may also assume that (10.6.8) ^<^6^ We first notice the following fact: Lemma 10.6.1. Suppose Y is a polyhedral cone of the type described and suppose we define y{x) = Zr{P)y{P) x{P)=x for points x not on the image of the boundary of any simplicial cone of Y. Then y coincides with a single-valued function which is continuous on Rp and linear on each of a finite number of simplicial cones. Moreover (10.6.10) Zy{P)=>^y a.e. on Y,K=\. x{P) = Xi Proof. The projections on the ^v-space of the simplicial boundaries of all the simplicial cones of Y divide that space into infinite conical regions which can be further subdivided into a finite number of simplicial cones. Let us consider two of these simplicial cones Zli and A^ which have an {v 1)-dimensional simplicial cone J in common. The points P of Y whose projections lie in A\ form a finite number ^ly, / = 1, . . ., / i , of simplicial cones, each of which is a part of one of the original ones of Y. The same is true of ZI2; let d^j, j = 1, . ., J2, be those correspoinding to Zl2- From the nature of Y, it follows that each dij abuts on a unique ^2; or another dij and the same is true of each ^2;. Thus, we may assume (by reordering if necessary) that each dij abuts on ^2;' for y = 1, . . .,/3 and
^ The e here has no connection with the e's used in other sections.
(10.6.9)
462
then that ^1,^3+22>-i abuts on ^1,^3+22) for p = \, . . ., {Ji Js)l2, and (d2,Jz+2q-i abuts on d2,j^+2q, ^ = 1, . . ., (73 72)/2. Then
-y(Pj2+2p-i)]>
xÂ,
XÂ2,
y(x) 2:y(P})y{Pi)i[y{Pjz^2,)? = 1
y{Pjz2q-l)],
Q=l
(10.6.11)
PJB+2P
x{Pj)=x,
2A=Js-Ji, 2B =
Js-j2.
Clearly y(Pj) has the same value on î^ and 62 j but has opposite sign on and Pj^^2p-i' Consequently as A;-> a point xo on A, the second sums -> 0 and the first sums approach the same value. Since y (x) is obviously linear interior to Zli and A2, the continuity of y follows. Since y{Pj) has the same value on a dij and ^2^ and has the opposite sign on Pj2+2p and Pj^+2p-i and on Pjs+2q and Pj^+2q-i, we see that (10.6.12) Iy(Pj) = Iy{Pj)^ ^^2. It follows that the sum in (10.6.10) has the same value K on Ai and A2 and hence for almost all x in R^ From the hypothesis (10.6.1) (which holds for Yn on account of the uniform convergence of the derivatives) it follows from Lemma 10.4.7, etc., that ;><;=: -j^ 1. Remark. Hereafter we assume that Y (and the Yn) cire oriented so that (10.6.10') 2y(P) = +'^ a.e.
For Y a polyhedral cone as above we define u(0) = 0 and (10.6.13) u{x) =f(pî,i(t
B{X,Q\X\)
x)y{t)dt
where 9 is a non-negative molHfier ^ C^ [B(0, 1)] C C^(i?v) and 9 (10.6.14) ^tiy) = Q~"îQ-^y) (see Theorem 3.1.3). If we represent Y by (10.6.15) (10.6.16) x = X(s), y=Y{s), s $ ^ = i?+xMi, Y(s)dX(s). we see from (10.6.9) and the definition of y that u(x) =f<p^^^^[X(s)-x]
In the form (10.6.16), we see that as Yn ^Y, Un-û uniformly with all its derivatives on any bounded closed set in Rv which does not contain the origin. Clearly u is homogeneous of the first degree in x and $C^(7?,-{0}).
463
For each Yn and for Y, we construct our desired Y^ = Vnl U Vn2 U Un3 U F^4, Y* = F i U F2 U F3 U F4 where Vi is the part of Yi outside the cyHnder \x\ = 1/2 and F2 can be represented on [1/4, 1/2T x ^ 1 by (10.6.17) v^-.r rl(^), Z{r,p) 1/4<^<1/2, ^{Ar i)rri(p) p^Wi + {2 - 4r) u[ri{p)].
It is shown in Lemma 10.6.12 below that if | V ^ | is sufficiently small (this is shown in Lemma 10.6.8 below) then there is a nearby Cartesian coordinate system ('x, 'y), which depends smoothly on V u (evaluated on \x\ = 1, say, since u is homogeneous of the first degree) such that if U denotes the cone (10.6.18) U:y = u(x) or 'y ='u('x) then
fl
^p^dx'
= 0,
^ 3 = 5 ( 0 , 1/8),
i<.i<v,
1<^<A/'
(10.6.19) Then F3 is the part of U for which | ^ ' | > 1/8 and \x\ < 1/4 and F4 is defined b y ^ . n ^ .^ ^ 4 : 'y =^v('r,'e) ^'u + (8y)2['^(1/8, '6) - 'u], (10.6.20) ^ ^ 0 < V = |':;t;| < 1 / 8 , 'd=^yx\-^'x. Corresponding formulas hold for Vni, etc. The parametric representations of the Vnj converge uniformly on their respective domains together with their first derivatives to those of the Vj so that /1*'(Y*) ^ / l * ' ( y * ) . It is shown in Lemma 10.6.11 below t h a t A^{Un n TF) < Ai(^, ) + A^{Yn n W), \imhi{Q, e) 0,
for any bounded measurable W. Using this, it is shown in Lemma 10.6.10 below t h a t /1*'(F^2) < ^^[Q, S) 6 + /t' (corresponding part of Yn). Finally, in Lemma 10.6.14 below, it is shown t h a t (10.6.21) yl'(F^4) < J^^ ^ A'' (corresponding part of Un).
Thus, whereas replacing the part of Yni outside \'x\ = 1/8 by Vni U Fw2 U F^3 may increase the area by h{Q,s) s where h(g, s) ->0 as Q,e -^0, replacing the part of Yn inside \'x\ = 1/8 reduces the area by an amount A (v, N) - s, A > 0. This leads to the theorem. However, Wi ^ F3 U F4 is evidently a r-cell with boundary (10.6.22) (10.6.23) Wo:x= Ci:y X(a,p), = u(x), y=Y(a,p), |A;| = I/4, a ^ [0, 1], p^ Mi and Wo = F i U F2 may be represented on [0, 1] x Wi by equations like
464
in such a way that a = 0 yields a homeomorphism of Wi onto / {r, PQ, Z) ^ Co and a = 1 yields a mapping onto Ci which, however, need not be a homeomorphism. We wish to show t h a t (10.6.24) b(Y^, Co) D H,_i{Co), Y^ = WoUWi. To do this, we first apply Theorems 10.3-9 and 10.3.10 with A = Co, B=Âi.
Those theorems yield the result (since b(Wi, Ci) = Hv_i{Ci) by Theorem
10.3.4)
, , , (10.6.25) ^ ^ X' = IxW, ^'(Co U Ci, Co)* b(Y^, Co) = b(Wo, Co U Ci) + + ^ ' ( C o U C i , Ci)*i/._i(Ci). A'^ôxmu A[=:ixWi, A'==A'ûA[,
If we now apply Theorem IO.3.13 with Ao = Co, Ai = Ci, X=Wo, the mapping / : X' ->Wo being given by (10.6.23), we conclude that if LQ == Hv_i(Co) there is a subgroup Li of Hv_i(Ci) such t h a t = K + i(Co U Ci, Co)* Lo = K + i{Co U Ci, Ci)* L i , K==b(Wo,CoUCi). That is, a h^ Hv_i(Co) and /î$ i^, there exists SL k2 in K and an I in ^'_i(Ci) such t h a t ^(Co U Ci, Co)* h = (k2ki) + t(Co U Ci, Ci)* / which proves (10.6.24) by showing (cf. (10.6.25)) t h a t i{Co U Ci, Co)* ^v_i(Co) C b{Wo, Co U Ci) + ^*(Co U Ci, Ci)* H,_i(Ci) the right side being that of (10.6.25). Of course this is also true for each n. We now prove the inequalities which we have mentioned for a polyhedral cone Yi. We use the notation of (10.6.5)(10.6.7). Lemma 10.6.2. (a) ]l\ +af + h< 1 ^ 2 ' [ l / 1 + ^f - 1j
(b) |/i + [Aa + (1 ~ A) b]^ < A |/l + ^2 + (1 - A) 1/1 + ^^2, 0 <;c < 1, (c) 1/1 + A2 a2 - 1 ^ A[]/I + a2 - 1], 0 < A < 1 .
(d) yi + (i + . . . + a)2<2:1/1 + ?.
Proof, (a) For ^ > 0 and J5 > 0, we define f(A, B) =(\ + ^)l/2 + (1 + 5)1/2 _ (1 _|_ ^ _j_ J5)l/2 _ 1 and show by differentiation that / i > 0 so / > 0. The result for more terms follows by induction, (b) just follows from the fact t h a t ]/l + x^
465
is a convex function, (c) follows b y setting b = 0 in (h). To prove (d) set f{ai, ...,an)=Z
i
y^ + af - j / l + (ai + + a^)2.
It follows easily t h a t each fai < 0 and t h a t f(ai, . . ., Un) -> 0 as all
Ui -> -\- CO.
Lemma 10.6.3. Suppose f is homogeneous of degree zero. Then
Bi
r l{ I 1/' l ^ r l / L l ^ -:^)^ + ?j]\dA dt\dx J

Proof. Let g[x, r, ?.) = \f{x + XQ\x\r)\. dr}dx
Bi U i L O
{B{X,Q\X\)IO
< CioW r \f{x) I dx,
Then the integral above is
= / { / [ jimldrjUAdr.
Bx lo [^(T.A) J J
G{r,^)CB(0.i+kQ),
where, for r and X fixed, G (r, A) is the image of Bi under the 1 1 transformation 7} ^ X -\- XQT \X\. That this transformation is 1 1 follows since 1^2 ^11 -{1 ^ ^
1T|)
< |?y2
: ^ I | ^\X2
xi\ (1 + A^ | T | ) .
The result follows easily. Lemma 10.6.4. Let us define [almost everywhere)
V{X)=Z\Y{P)\< X{P)=X
define Xi as the subset of Bi where v[x) =^ \ and X^ as the subset of Bi where v{x) > 1. Then Jv(x) dx < 2e,
^2
f [v{x) \]dx = ( [r(:v) \'\dx <, e.

Bi Xz
Proof. Clearly yr + 8= A^'iYi) > jv{x) dx = yr + f [v{x) 1] dx.

Bi Xz
The result follows from the fact t h a t v{x) < 2[v(x) 1] for v(x) > 2. Lemma 10.6.5. / / \x\ < 1, then \u{x)\ < Cii(r, N) Q [if q < 1/2). Proof. From the homogeneity, it follows t h a t is is sufficient to prove this for |;t:| = 1 ^, if ^ < 1/2. From (10.6.9) and the definition oiv[x) we see t h a t (10.6.26) \y[t)\^qv[t). From (10.6.13), it then follows that \u{x)\ <Q + ZiQ'-*'f Q'[v{t) - i]dt<:Q[\
Xz Morrey, Multiple Integrals 30
+ ZiQ-^'e]
466
from which the result follows, since e < Q^^. Z\ depends also on the moUifier ^. Lemma 10.6.6. Suppose fx is a measure over a set E, iLt{E) is finite, and f is non-negative and integrable (/j) over E. Then
min I C (E)]^ [ ffdnf] M
< 3 [/* ()] / (Vl +P
-i)dfi.
Proof. From the Schwarz inequality with / = {fiyi\ + / 2 + 7 ) Vl'l + / 2 + 1, we obtain
ffdf,f < IJiYT+P - l)^/^]. [/(l/l +/2 - 1 + 2)df,

<A^ + 2AB, A=: f{-]/\+p-\)dft,
E
B=fx(E). }AB, then
li A^ + 2AB^'}AB, the lemma is true. liA^ + 2AB> ^ > ^ a n d 5 2 < }AB. Lemma 10.6.7. f\u[x{P)]
- y[P)\^ dj;[P) < Ci2{v,N) - Q^ e.

+ Z2 {V, N) Q^ 8,
Proof. Let us denote the integral by / . Then

/ < / l + / (ClO + 1)2 Q^V{X) dx<^h
(10.6.27)
, ^' h = J \u{x) y{x)\^ dx,
using Lemmas 10.6.4 and 10.6.5 and (10.6.26). Now ii x^Xi

B(X,Q\X\)\'\U(X) y(x)\
<
Z3'Q\X\'J
IZs{(p,v,N)'
\f\Dy[{i
Q\X\'\B{X,
-X)x
Q\X\)\,
^t]dA\dt
on account of Lemma 10.6.5. Using Lemma 10.6.6 with E the set {t, X) with t^B{x, Q\X\) and 0 < A < 1 and d/bt = dt dX, we obtain /.n^..^ (10.6.28) B[x,Q\x\)\'\u[x)-y{x)\^ < 3Zf^2|^|2| U\^J\^-\Dy[[\
U
-K)x + Xt]\^-
\\d?\dt.
J
B{X,Q\X\)
Using Lemma 10.6.3, we therefore conclude that (10.6.29) f\u{x)

Xi
y{x)\^dx^Z4.Q^f{]/i
+ \Dy{x)'f - l)^^
But, from (10.6.9) it follows that (10.6.30) Dy{x)\<2:\Dy(P)\.
467
Thus, from Lemma 10.6.2 (d), we conclude t h a t

Bi Yi
(10.6.31) The result follows from (10.6.2/)(IO.6.3I). Lemma 10.6.8. (i) \Du(x)\ < Cisiv, N) e^l^, if Q < 1/2; (ii) \u(x)\ <\x\ '\Du{x)\. Proof, (ii) follows from (i) and the homogeneity. From (10.6.13), it follows t h a t
u{x) = f(p(X) y[x +

so that (10.6.32)
Bx
XQ\X\)
dl
(A = (^ \A)~^{^ - A)
Bi
^y
x^-^B^y
I 1 ^ dX
from which it follows, by setting t =^ x -\- Q\X\X, (10.6.33) \Du{x)\^{\+ZiQ)j

B{X,Q\X\)
that
99*,,,(^ - x) \Dy[t) \ dt.
Since D u and D y are homogeneous of degree 0, it is sufficient to prove (i) for \x\ \ ~ Q. Then, since B[X,Q\X\) <Z Bi, we conclude from the Schwarz inequality as in the proof of Lemma 10.6.6 that
j^%\if-^)\i)y{i)\dtf
BI,X,Q\X\)
(10.6.34)
/9^*i.,(^ - x)[^J\ + |Z)3^(^)|2 B{X,Q\X\)
1)^^] X
J
/ vtU^ - *)(yi + l-DyW^ - 1 + 2) dt]

iB{x,Q\x\) J
<.ZQ-ê{2
ZQ-ê). x
The result follows from (IO.6.34), (10.6.8), and (IO.6.3I). Lemma 10.6.9. /* I/1 + \Du{x)\^v{x) dx < f ^ l/l + \Dy{P)\^
^ ^ x{F)=x
Xdx + C Qe, B = Ba Bi,, 0 <b <a<^\. Proof. Since D u, v, and Dy are homogeneous of degree 0, it is sufficient to prove this with B replaced by Bi. From (10.6.33) and the definition of y, we obtain
\Du{X) I < (1 + Zi ^) l\Dy(P) I df,(P;X)

Y{X,Q\X\)
(10.6.35) {d/,(P;X) where Y{X,Q\X\)
- (^ + ZiQ) f <Ptui{i - X)
B{X,Q\X\) x{P)=t
I\Dy{P)\dt
= cptui[^{P)
X]dZ{P)) X[P)^B{X,Q\X\).
30*
consists of all P on Y such t h a t
468
Using the device of (10.6.34) with 5 ( Z , ^ | Z | ) replaced and IZ) y (^) I replaced by | JD y (P) |, we find t h a t [ ( ] / l + \Du^~
Bi
hyY{X,Q\X\)
^)dx<^f\Du(X)\^dX
Bi
< i (1 + Z i ^)2 fw{X) [w{X) + 2co{X)] dX,

Bx
(10.6.36)
w{X) = f <ptui(t-
X) J V [l/l + l-Dy(P)|2 _ i]]dt
B(x.eix|)
B{X,Q\X\)
Clearly w and co are homogeneous of degree 0 so we may estimate them assuming t h a t \X\ = \ ~ Q. Then, since \D y{P)\ < \Jy{P)\, we see first, using (IO.6.36). Lemma IO.6.4, and (10.6.8), that w {X) < Z2 o-*' e, ^ so t h a t (10.6.38)
(10.6.37)
o) (X) < 1 + Z2 p-" 2s

V / 1 t^
w{X) + 2co{X) < 2 ( 1 + Z 3 ^ ) / ^ | ( l / l + \Du\^ ~ \)dx<

BX
(1 +Z4Q)
BX
fw{X)dX.
If, in (10.6.36), we set t = X + Q\X\r, order of integration, we obtain

Bx
integrate, and change t h e
lw(X)dX
= fcp{r)\f/
Bx
\BX \^^P)=X+Q\X\r
\^\ +
\Dy{P)\^-\KdX\dr.
/ J
(10.6.39) If we set, for each r in i5i, rj ^= X -\- Q\X\r,we see t h a t the transformation and its inverse are Lipschitz and the Jacobian \dXjdYi | < 1 + Z5 ^; moreover, for each r the image of Bi C î+e- Accordingly we obtain (since D y is homogeneous of degree 0) (10.6.40) /<(1+Z6^)r Z [l/l + | Z ) y ( P ) | 2 - l ] ^ ^ .
Now, using Lemma 10.6.8, we may write J y i + \Du[x)\^v[x)dx<, j ^ \ + \Du{x)\^dx \[\+Z,Q)j[v{x)-\]dx^y,

Bx
+ l/l + | D y ( P ) | 2 | x
J
+j\
Bx U(P)-n
X drj f v(x)dx
Bx
+ J [v(x) \]dx -{Bx
+ ^ 6 ^ / 1 2 ' [^J^ + \Dy[F)f-\]\d7l

Bx V<^)=? J
+
Bx
Z,Q([v{x)-\\dx
from which the result follows {\Dy[P)\
< | / y ( P ) | , Lemma IO.6.4).
469
Lemma 10.6.10. Suppose Z = F2 and is defined by (10.6.17). Then A^'iZ) <yl^(Y2) +(p(Q,s)' , lim(p{Q, s) = 0 where Y% is the part of Y for which 1/4 < \x\ < 1/2. Proof. On each simplicial cone of Y, we may consider 3/ and z as function of X. From (10.6.17) and the notation of (10.6.5), etc., we obtain z[x) = (4^ 1) y{x) + (2 4^) i^[x) g
b(x\
= (4^ - 1) f j + (2 - 4r) U
XO) ~ ^^' ^^ 'l b (Xi, XO) ^ '
+ d\ d\, = A{y^~ u^) [x^jr)

^^^^' ~^ 31'
d(Ar*i, xH, xh)
d (;;*!, ;ir*2, x^^)
where each d^p is a sum of Jacobians, of order <^p, oi the y^ with respect to the x^ multiphed by products of components of D u and y u. Regarding the Jacobians, etc., as components of a vector, we obtain \Jz(P)\^{4r-\) \Jy{P)\+{2-4r)\Du[x(P)]\ + \d{P)\. Then, for each ? > 0, we have y \Jz{P)\^ <{i+ri) {{4r - 1) \Jy{P)\ + (2 - 4r) \Du[x(P)]\}^ + -\-{i+r)-i)\d{P)\K Hence (refer to Lemma 10.6.2) /[I/1 + \jz{P)|2
Yz
- i]d2:</[l/i
Fa
+ (1 + v-^)\d(P)|2
-i]d2
+ \]dZ
+ / [ y i + (1 + rj) {{Ar - 1) \Jy[P) I + (2 - 4y)\Du[x[P)-\ \f < (1 + j-i)i/2 / [yi + i^(P)|2 - 1 ] ^2" + + {^+V?'^l[]/^+{{4r-i)\Jy{P)\
Y2
(2-4r)\Du[x(P)]\}^-i]dS
+ ( 4 r - l)/[fl + \Jy{P)\^ - \]dZ +

Y2
< (1 + ?-l)l'2 / []/l + |i(P)|2 - 1] (^2" + ^2 + [(1 + ?)i'2 - 1]/(4r - 1) / [1/1 + 1/y (^) P - 1] ^2" + I r^ + (2 - 4;') / l/[l + IZ) [^ (P)] |2 - 1] <f2l + y^ J
+ (2 - 4r) f [|/l + \Du[x(P)]\2 - 1 ] ^ 2 ../]/l+|70(P)|2c^2^(4r-1)/]/l+|7y(P)|2^2 + + (2 - 4^) /1/1 + ID M [^(P)] |2 c?2 + 9'i(e. ) from which the result follows, using Lemma 10.6.9.
Fa
470
Lemma 10.6.11. / / U is the cone given by (10.6.18), then AÛ n W) </l'(Y nW) + h{Q, 8) e, lim^(^, e) = 0 for any bounded Borel set W. Proof. It is sufficient to prove this for W the cylinder \x\ < 1. Since \J u\^ = \D u\^ plus the squares of all the Jacobians of the u^ with respect to the x^ of order two or higher, we see from Lemma 10.6.8 that A^{Ui) = / l / l + \Ju{x)\2dx<f]/\ < f 2^ Y^ + \Dy{P)\^dx + \Du(x)\^dx + Zs'^l^ + CQe
+ Z^l^ + CQs<A''{Yi)+Ze^^^
using Lemma 10.6.9. Let us now consider the cone U given in the {x, y) system by (10.6.18) where u^ C'^{Rp {o} and u is homogeneous of the first degree. We wish to prove the following lemma: Lemma 10.6.12. / / |Z)^(%)| ^5 sufficiently small, there is a nearby {'x, 'y) Cartesian coordinate system which depends smoothly on \7u{x), evaluated on \x\ = \, such that (10.6.41)
Bi
fy^^dx'Ô,
i = i,...p
= N -V,
k=i,...,v,
Remark. It is clear that (IO.6.41) will then hold with Bi replaced by Ba = B (0, a) for any a> 0. Proof. Suppose we let (10.6.42) 'x^ = c^x^ + d^yr, V = elx^ +/jy'* be a positive orthogonal transformation. Then U is given parametrically in the ('x, 'y)-system by (10.6.43) 'x^ = c^x^ + dûr{x), 'ui = elx^+fiur(x). Then (10644) -^^^^ V ( ^.,^rd'ui^dr.\,../.^-i/.^î,,../x^)[d.-.-i Thus, the equations (IO.6.4I) become / (10.6.45) _ r y i .
~JjiLj^
B*
^,,^-..V^rî,.../^^-i/^^î,.../^v)
^ d'xi d{x^,...,xi-^,xi+^,...,x^) ^^-^
where the derivatives d' u^dx^, d' x^jdx^, etc., may be found from (10.6.43) and 5 * is the region in the ^-space correspoinding to the unit sphere B in the ';v-space by the first equations in (10.6.43). To determine the matrices c, d, e, and f, we have the iV(iV + l)/2 equations which require the matrix (10.6.46) (^ ^ ) (iV = i . + ^ )
471
to be orthogonal. Next we have the vp equations (10.6.45), which are of the form (10.6.47) '^l{c,d,ej)u) =Q, i = \,...,p, k=\,...,v where g | is a polynomial in the c, d, e, f with coefficients which are integrals over 5 * of various jacobians of the u'^ with respect to the xK This leaves {v + p)^ V p , v^ V p^ p Vp \ equations which we can use to prescribe the cf with / > ^ and the / j with j '> i. By replacing uhy u -\- Xv.itis easy to see that the Frechet differentials %i{c,d,e,f]u,v) (10.6.48) _ ^.^^_^ ^^^^^^ ^^ ^^^. u^Xv)-%i[c,.. / ; u)] are continuous for || u || sufficiently small, where 1^11 = mdiK\Du[x)\, \x\=\ We have already noted the dependence of ^^ on {c, d, e,f). The integrand in (10.6.45) is /I + cj + d]iii^ cl + dfui,...el + {6} +f})ui^... c\ + d]u\\ (ietf ^2 + ^Û ^ + ^1 + d]u\^...e' + (^j +/j)^!2--. c\-\-d}u\, (10.6.50) we have replaced 4 by 1 + c\ and f) by d) + f]. If we now regard all the c, d, e, f as small and neglect products of two or more of them and two or more of the ^-derivatives, the determinant in (10.6.50) reduces to that of the matrix (10.6.51) (10.6.52) (4 + ^ : f c + / > y Ui = \B\-^Ju%(x)
B
(10.6.49)
^2 + ^>;2 j + d)u[J)
cl + dû\,...el+[d]+f;)ui,..A+cl
If we also replace 5 * by B and define dx
the equations for the c, d, e, f become
4 + ui+f}Ui = o, c|=/| = o
(10.6.53) 4 = - cf if ki^l, 4^-î' i,j=i,....p, fi = - / j if ; V i k,l=\,...,v.
472
It is clear that we may prescribe the f) with/ > i and the cf with / > ^ and the remaining c\ and /} are determined. Then the e\, are determined from the U^ and f] and then the d\ are determined. Since the equations (10.6.53) 9.re obtained by linearizing the actual equations, and since the derivatives and FRECHET differentials are continuous, it follows that the equations (10.6.47) together with the orthogonality relations yield a unique solution in which the matrix (10.6.46) is sufficiently near the identity provided that the prescribed ei and f) are sufficiently near 0 as is I ^||. I We now prove: Lemma 10.6.13. Suppose that u ^ C^(Rv {O}) and is homogeneous of degree 1. Suppose also that (10.6.54) fu%{x)dx
B
= 0,
i=i,.,.,p,
k=\,.,.,v.
Let V he defined by (10.6.55) Then (10.6.56)

B
v(r,d) =u(0) +r^[u(\,d) u(0)], r u{0) = \Z\-^fu(\,d)d2;{e), f\Dv\^dx< ^^l.^

B
f\Du\^dx.
Proof. Vr = 2r [u{\, 6) u(p)] ur.^r.^ (IO.0.57) From homogeneity, u{r,d) =ru(i,e), Ur{r,e) = ^ ( 1 , 6 ) , r-^V6u{r,e) = \7eu{\,d). Hence j\Dv \^dx = - L _ | ( 4 1^(1, 6) - f^(o) |2 + I V6u(\, 6) |2}^2'(^) B /'|Z)^|2^:^ = 1/'{|^(1,0)|2 + iV0^(l,6l)|2}^2'{6>).
B
2:
r-^Vev(r,e)=rVeu(\,e) \Dv\^== 4r^ 1^(1, 6) -u (0) |2 + ^21 Veu{\, d)\^.
If we now introduce the spherical harmonics qnk (see after Lemma 10.6.14 below) with ^01 = 12^1-1/2, qijc=C,jc'\x\-^xK our hypotheses and definitions imply that
n,]c n^2
k=
\,...,v;
473
It follows (see below) that

B \ n,k I
j\Dv\^dx
B
V + 2-
I = ^{^\anjc\n4
\ n,k U^2
w^2 + n(n +
J v-2)]\
I J
U^2
(v + 2) L
J
2v 4- 1J j I I 2v + 1 j 1
<7-f-ri +--^1./'|DW|2Â; = - ^
flDul^dx.
I
Lemma 10.6.14. TA^r^ ^'s <; number r]{v,N) '> 0 such^that if u ^ 3 C^(R {O}), \D u\ ^rj,u satisfies (10.6.54), and v is defined by (10.6.55) then (10.6.58) f{\/\
B
+ \Jv\^ - i ) ^ ^ < - ^ | ( i / i + \Ju\^

B
~\)dx,
Proof. For if \Du\ <r], it follows from (10.6.57) that \Dv\ < 2r} on B. Thus Ay^ + 17^1^- ^)dx<^^f\Jv\^dx<^(\
B 2V+1 j ' B ' B ( 2 X ' + 1 ) ( 1 - Z 2 i y 2 ) J Vy B
+Zirj^)
B ri
f\Dv\^dx
I J
since l/l+f2_1>if2_^|4 for all f and |Z) ^| < | / ^|. The result (10.6.58) follows easily. The spherical harmonics. We begin by noticing that iin ^ p and Hn and Hj, are homogeneous harmonic polynomials of respective degrees n and py then 0 = J HnAHpdx
(10.6.59) = 5 / {HnHpr
dB
= f Hn,ocHp,ocdx + J
6^ dS.
HnHprdS
B HjpHnr)
From the homogeneity it follows that Hnr = rHnr = nHn, Thus we see that (10.6.60) JiP n) Hn HpdS = 0 or J HnHpdS
dB dB
Hpr == r Hpr = r Hpr = p Hp if = 0.
r=\.
474
So we define the qnk on dB = 2^ ^^^ each n, to be a complete n.o. set, each of which is the restriction to ^ of a (homogeneous) harmonic polynomial of degree n. Then all fhe qnk for all n form a complete n.o. set for
^2(2')Let us suppose t h a t u is representable as a finite sum of the form (10.6.61) u{r, 6) = Z^nk(r) qnJc{6), anjc{r) = r^bnic(r) where each hnTc[y) is an even polynomial in r. Since r'^ qnk ^^ a harmonic polynomial, ^ is a polynomial. Then J \Du\^dx
B
= J {u^ + r~2 l"^ Q u\^) dx = J u Ur d^

B
f uAu
B
dx.
(10.6.62) If ^ is given by (10.6.61), it follows t h a t ,_^^^^

(10.6.63) n.fc
dB
l^r = 2 <lc{^) ^nk(6)
A U = Urr + (^ 1) ^~-^ ^r + ^~^ Zl2(9 ^ whereZI20 is the Beltrami operator o n ^ . Applying (10.6.63) to u = Hnk = ^^ qnky we obtain ,, ^ ^ , , 0=^n[n + v~2)r^-^qnk + r'^-Â2eqnk. (10.6.64) Aêqnk = n{n + v 2) qnkThus (10.6.65) ^ w - = 2 ' K f c + ( ^ - i)r-âlj,-n{n + v 2)r-ânk]qnkSubstituting (10.6.61), (IO.6.63), and (IO.6.65) in (10.6.62), we obtain
1
n,k
f\Du\^dx
B
= fr^-^f = fr'-^
0 1
(u^ +
r-^\Veu\^)d2^dr 4,^] dr. (1).
S K l + n{n+v-2)r-^
n,k
From the arbitrariness of n, it follows t h a t f\Veu{\,e)\^d2 = In(n + v-2)alj,
wf .c
10.7. The local differentiability In this section we prove the following main theorem: Theorem 10.7.1. Suppose XQ = S \J ^ ^s a minimizing set as in Theorem 10.4.4- There exists an 0 > 0 with the following property: If PQ^S =^ XQ A andip{Po) < 1 + 0, there exists for each/bc, 0 < // < 1, a neighborhood of PQ on S which is a regular v-cell of class C^. Thus the subset of points P^ S where ip[P) = 1 is open on S and ip(PQ) = \. If W is of class C^for some n > 4 and some [x, 0 <^ [x <C \,the v-disc may be taken to be of class CJJ. If W is of class C or analytic, the v-disc may be taken to be of class C"^ or analytic, respectively.
10.7- The local differentiability
475
From Theorem 10.4.4, it follows that y){P) (see 10.4) = 1 everywhere on S. From Theorem 10.4-5, it follows that if-y; (PQ) < (for i = ei(*,r, SJ^) sufficiently small), then 5 satisfies a n condition a t PQ. If * and Ri are sufficiently small, it follows neighborhood of PQ on S is a topological r-disc 5. From Lemma it follows t h a t iir]o'> 0 and si and RQ are small enough, we may so small t h a t y)(P) < 1 + TÔ for P on 5. Let (10.7.1) r:B(0,R) onto 5 , r = cooooz,
almost 1 + i (6*,Ri) that a 10.4.4, take 5
where COQ o is a normal coordinate system centered at PQ. Then we conclude from Theorem 10.5-2 t h a t (10.7.2) Lv(r)Â^{S), U{r\r) <A''[r{r)], ifTis o p e n a n d r c 5 ( 0 , ie). Lemma 10.7.1. The equality holds in (10.7.2). Proof. For suppose for some F that the inequality holds. Then H a sequence {r^}, where Tn 6 C'^ [P) and rn->r uniformly on P such t h a t limL, (rn\r)=d< A'' [r{P)];
we may assume that each Zn is locally regular. Then it follows t h a t Lvlxn I P] = A^[rn(r)]- We may now set up the functions y){r, P , T%), q){r, P , Tn) and their limits just as in IO.4. The developments of t h a t section may be repeated and lead to a set XQ for which A^ (S') = d, where S' =^ XQ r{dP). We now modify the tn as follows: We select a sequence of analytic domains {Gn} ^GnG Gn+i for each n a n d U Gn P. B y going along the inner normals to ^G^ a short distance Qn, we form another analytic domain G'^dGn. We define r'^[x) = rn(x) for x^G'^,r^(x) = r{x) ioT x^P ~ Gn, and
T*[(1 -t)x' + tx] =COoo[(1 -t)Zn(x') 'Vn = COQoZn, +tz{x)], 0 < ^ < 1 ,
where x' and x are extremities of a normal segment to dGn and cooo is a normal coordinate system centered at PQ. This alteration does not affect the functions ip (r, P , Zn) and q? (r, P , Zn) if we merely restrict r <, R (P) an where On - > 0 . But the sets T * ( P ) and r(P) now all have the same boundary. Thus XQ has t h e same boundary as r(P). But, since d <A^[r(P)], this contradicts the minimizing character of XQ and hence r(P). The proof for the whole surface is the same. Proof of the theorem. Suppose that PQ 6 S> y^(ô) < 1 + 2, and 5 is a r-disc containing PQ for which (10.7-2) holds. Since ip is upper semicontinuous, we may find an ^3 > 0 (see Lemma 10.4-4) such t h a t (10 7 3) ^ ^ ( ' ' ' -P) = ^ ' ' [-5 ^ -S (P, r)] < (1 + 2s2) h{r).
476
From Theorem 10.4.5> it follows t h a t there is an. r^, 0 < ^4 < ^3 such t h a t with each P^B(PQ, ^4) and each ?', 0 < r < ^4, there is a geodesic ^'-plane 2J (P> ^) through P such that (10.7.4) D[SnB(P,r), 2'(^^^)n5(P,r)] <ôr/3, ^0 being the constant in the cone inequality. Let us suppose that 5 is represented as in (lO./.l). Then we may approximate uniformly to z on B(0,R) b y polynomials Zn so that
limLv{Tn) = Lv(T) = A^(5), rn = COQQO Zn.
Choose P interior to 5 , let OJP be a normal coordinate system centered at P , and let us define Sn = rn[B[0, R)], l(r, P , Sn) =5nn dB{P, r), lo{r.P,Sn)=cojni{r.P;Sn]. From the cone inequality it follows that there are positive numbers ^0 (above) and 3 with the following property: For each r for which there is a I'-plane ^ in RN through 0 such that D[lo [r, P , Sn). 2 ^ 0 dB(0, r)] <Qor, and A'-^ [lo (/, P , Sn)] < (1 + 3) r 7. r^-^, there exists a set Y^ with b{Y'^, An) D Hv_i{An), where such t h a t A-{Y^) An = l(r,P,Sn), Y*=cop(y*), kyrr\
< (1 - ^) r - i r / t - i [ / o ( r , P , Sn)] +
From the uniform convergence and (10.7.4), it follows that if we choose ri and ^2, with 0 < î < ^2 < H, then (if 1 + ^yo ^ < 3/2) D[l{r, P , 5^), 2^(P, r) n a 5 ( P , ^)] < 2^0^/3 n <r <r2
Z)[Zo(r,P,5^), 2 ' n a 5 ( o , r ) ] < ^ o r
^>Ar.
If we let dn be the inf. of A" {X) for all sets X for which b {X, An) D Hv_i{An), it then follows, as in the proof of the lemma above t h a t dn ->A''{S). Then, we conclude from the cone inequality, the inequalities in 10.1, and the argument in the proof of Lemma 10.4.2 t h a t
^(r,P,S^)<f
(IO.7.5) (a)
jrjn + t-^r(\ + hr) cpr{r. P , 5^) + kyvf''[\ \r]n + v~'^r[\+hr) (pr(r, P , Sn), in all cases P^B
+ >yo^)^ if (a) lim?7^ = 0
rî < r < ^2, ^ = (1 k)-^ V, n>N, [\ +hr->(\ +r]or)^''-^.
(Po, n),
(fr {r, P , 5 ^ ) < (1 + ss) V yr r""! {\ + rjo r)^~%
For n and P fixed, define kn {r) = max [0, y) (r, P , Sn) rjn]-
10.7- The local differentiability
477
Since cpr and % (below) are > 0 and (pr < ipr (Lemma 10.4.1, 2, etc.) and 99 is A . C , we notice first t h a t (10.7.6) kn[r) = qn{r) + Sn{r), s^{r) = 0 a.e. where qn is A.C. and % is non-negative and non-decreasing. Then, from (10.7.5)> we conclude that (a.e.) (10.7.7) ^ ^^^ - 1^-1 r{i +hr)k:^ ^ (r) = y, 7- (1 + T o r)\ y (10.7.8) {r),
(D {r)
in aU cases = {i + 3) v yy r^-i (1 + ^0 ^ ) ^ - ^ + hr)a){r).
Since (1 + TJQ r)2''-i < 1 -f /^ r, it follows that t-'^r[\ -f hr)co{r) + kx<v~^r(i We now define J^7^(/') as the unique solution in C^ of the equation (obtained b y requiring the equality in (10.7-7)) itr-^(\ Pn= lvr-^(i (10.7.9.) +hr)-^lpn + hr)-^pn,
kn(r2).
kx), if pn<t-^r(\ if pn >v-^r(\
+ hr)co
+kx
+ hr)a) + hr) o)
\cjo(r), if t-^r{\ + hr)co + kx <pn <v-^r{\

Pn{r2) =
By considering all possible cases, we conclude t h a t (10.7.10) k'^{r) ^p'^{r) (10.7.11) whenever kn{r) >pn{r), ri^r<r2, n>N. n < r < ^^2, So, if we define u^ {r) = kn (r) pn (^)> we have Unirî) = 0 and u^{r) > 0 whenever Un{r) ^0, (n > iV). I t follows that Un(r) <. 0, ri <^ r <, r2, n ^ N; for if, for some n and r with ^'i < r < ^2, we would have Un (r) > 0, then Un {r) would be > 0 for r in an interval [f, r + 6] (on account of (10.7.6)) in which i^n{r) > 0 ^'^' From (10.7.6), it would follow t h a t r + ^ = ^2 which would be impossible since Un (^2) = 0. If the first line of (10.7.9) holds for ri < r < ^2, then (10.7.12) Pn-X-t-^r(i+hr) {f, - Xn) + [\ - k) y. ^^^^ z{r), z{r) = (1 -f T o rY'"^ (^ + ^0 + 2hrjo r). y Integrating (10.7.12), we obtain Pn ir) -x{r)< (10.7.13) +Zi ( 7 ^ ) ' ( ^ y [^ (''2) - X (^2)] + "' r>+i- if i < v + 1. Now, since 1 -\- hr >{i (10.7.14) + rjor)^'~^, -k)x^(i -k)ezx{r):f-i>-(1 +hr)co-{\
478
Now, the right side of (10.7.13) will be < that of (10.7.14) iff f{r) =: [ht (/-)]-! X{r), Now (10.7.16) = _ tr-i{i m=^^-'{' + i f b ) ^-Mi+^^)<Oiir<Z^. ht(r)=rt{i+h r)-K
+ A , ) - i [1 _ (1 _ k) (1 + hr) ( ^ ^ ^ ^ ) ]
So f(r) takes on its minimum for r = r2. Then (IO.7.15) holds for all ^ < ^2 if (10.7.17) Pn{r2) < [1 + ( 1 -k)sz]x(r2) -Zirl+HA +hr2)-K Since (Lemma 10.4.2), [h(r)]~^ kn(r) is non-decreasing for each n and kn{r)-^y){r,P) and [h(r)]-^ ip{r, P) N\ Accordingly kn{r) <=pn[y), wherepn is bounded as in (10.7.13). Passing to the limit, we see t h a t (since ri was arbitrary) (10.7.18) yj+{r, P) < x{r) + {i - k 3) [ht{r2)]-^[ht{r)], 0<r<r2. From (10.7.18) and (10.7.5), we conclude that there are numbers Zi and ^, depending only on v and 9)1 such that (10.7.19) h{r) < ip+[r, P) < (1 + Zi fP) h[r) for all ^ ^ p ^ ^^ ^ ^^p^^ ^2;ô), 32/0 < min (rg, ^4)
and ro is sufficiently small to satisfy the further conditions in (10.7.21) and (10.7.22) below. From Theorem 10.4.5, it follows t h a t S satisfies an (^0' ^0) condition at P Q with e"^ = larger of 32 cf ^ ro and Z2 r^, (10.7.20) (10.7.21) (10.7.22) (10 7 23) Z2 = ( 2 - i - c ^ i - Z i - 3 2 0 i / ^ , 32ro<Ro\ 32fo<Cl^ 2 = oc = Plh>0 provided that TQ satisfies the following conditions (^2 = 32?'o) 2-i-Zi-32^-rg<C2-(4)^. We see t h a t ej = Z2 r^ provided that yJ-'^<32-iciZ2. Now, suppose 0 < A < 1. For each k, we conclude from (10.7.19) t h a t ^(''^ - ^ ^ ( ' ' ' P) < (1 + Z i . 32Â^^/g) ^(/) for all B(P,r)cB[Pjc, 32A^ro) where Pjc^B(PQ, 32ro - 32A*ro). Thus, from Theorem 10.4.5 and equations (10.7.21) and (10.7.22) it follows t h a t S satisfies an (e*, A^ TQ) condition at any Pjc^B(Po, }2ro 32A^ro), where (10.7.24) s^ = Z2^^''r^.
10.7. The local differentiability
479
Using the definition, we see that for each k, the following is true: For each P in any B {Pjc, 2X^ ro) and each R '<X^ TQ. There exists a geodesic r-plane ^ (P, R) centered at P, such t h a t (10.7.25) D[SnB(P,R), 2:{P>R)nB(P,R)]<8^'R. For a given k, U B{Pjc, 2^^ ro) = B{Po, 32ro 30A^ro) and the intersection of these for all ^ > 0 is just B(Po, 2ro). If P^B(Po, 2ro), the conclusions above hold for all k. So, if we assume that ^k+i we see t h a t E^<ZzR\ (10726) Zz = l-^Z^. Since this is true for each h, we conclude that ^[5n5(P,i^),2'(^>^)n5(p,ie)]<Z3iî+,
P ^ J B ( P O , 2ro), 0<i^<ro.
ro<R<^^ro
From the proof of Lemma 10.5.4, it follows that (10.7.26) holds with S and ^{P, r) replaced b y CO^Q (5) and COQQ [ ^ ( P , r]), respectively. Accordingly, for the remainder of the proof, we shall assume t h a t 5 stands for o)^^ (S) and that ^ ( P , r), etc., are actual r-planes in Rj^. Next, we observe, using (10.7.26), that 2;{P, rl2) n B(P, rj2) C [5 n ^ ( / / 2 ) , Z3(r/2)i+] n B{P, r/2) (10.7.27) C [5 r\B{P, r), Zs(rl2)^+-] H 5 ( P , r/2) C IZ (P. r)nB (P, r), Zsr- + Z3(r/2)-] 0 B (P, r/2). Accordingly (10.7.28) d[2(P,r),2{P,rl2)]<^Z4r-, Z^ = 2Zs{i + 2~^-). Replacing r b y 2-'^ r for n = \, 2, . . ., in (10.7.21), we conclude that there is a )^-plane ^ (P, 0) through P such that (10.7.29) ^ [ 2 ; ( P , ^ ) , 2 ' ( ^ > 0 ) ] < ^ 5 ^ ^ Z5==Z4(1 + 2- + 2 - 2 - + . . . ) , so t h a t
z)[5n ^(P,^), ^ ( P , 0) n 5(P,r)] <Z6rl+^ Ze ^ Zs + Zs

0<r^r2, P^B{Po,r2) provided that ^2 is small enough. Next, if P and P ' $ 5 ( P o , ^2) and | P P ' | = r/2, we see, b y proceeding as in (10.7.20) and using (IO.7.23) ZiP', 0) nB{P\ rl2) C [5 n B{P\ r/2), Z6(r/2)i+] fl ^ ( P ' , ^/2)
c Lr(P. 0) n 5(P, r), Ze ri+- + Z6(r/2)i+-] n P(P', ;'/2)

C [2;'(P> 0) n -B(P, r)l 2Z6 fi+- -f 2Z6(r/2)i+-) fl P ( P ' , r/2) since P ' ^ [^/(P, 0) H 5 (P, r), Ze ri+'^ + Ze (^/2)i+^; here 2 " (^. 0) is the r-plane through P ' \\2(P, 0). Thus (10.7.31) ^ [I(P, 0), 2 ' ( ^ ' , 0)] < Z^(rl2Y.
480
From (10.7.23), it follows t h a t if ^2 is small enough and lip is the (N v)plane through P J _ ^ ( P , 0) and 77' is another {N r)-plane through P with ô(7J', Up) < 1/2, say, then P is the only intersection in B{P, r^) of 11' and S. By taking 11' as the plane through P||i7po> we see t h a t S n B {PQ, r2l2) has a 1 1 projection on ^ ( P o , 0). I t follows easily t h a t 2J (P> 0) is the actual tangent r-plane to S at P . If we choose a (f, rj) coordinate system with origin at PQ and f axes in ^ ( P o > 0), we may represent the part of 5 near PQ in nonparametric form rj =r](i). From (10.7.24) it follows that rj^ CI. The higher differentiability follows from the results of 6.4 and 6.6 since the EULER equations form an elliptic system. 10.8. Additional results of Federer concerning Lebesgue i^-area
In this section we present an account of those results of F E D E R E R which the writer originally thought were relevant for our problem but which he later found not to be essential. Since the results are important and not easily accessible, we include them there together with the necessary background material. Our presentation leans heavily on the notion of the order of a point with respect to a mapping which was introduced in 9-2. The use of algebraic topology has been reduced to a minimum. We begin by extending the definition of the order function and introducing FEDERER'S essential multiplicity function. Definition 10.8.1. We suppose t h a t W is an oriented i^-dimensional manifold of class C^ which is diffeomorphic to an analytic manifold, G is a domain on 3K the closure of which, G, C 3Jl, and we suppose t h a t J : z ^^ z(x) is a continuous mapping from G into P^, and .2:0 ^ Pv- If ô^TîdG), we define 0 [ZQ, T{dG)] == 0; otherwise, we define (10.8.1) 0[zo,r(dG)] = limO[zo,r{dGn)]
where {Gn} is any sequence of domains of class C^ such that Gn C Gn+i for each n and U Gn = G; itis clear t h a t the limit is independent of the sequence. We define M (zo. r) = sup 2 ; 10 [^0, T {d Gi)] \
i
for all finite or countable sequences {Gi} of disjoint subdomains of G. Lemma 10.8.1. Suppose G, T, and each T satisfy the conditions oj Definition 10.8.1 and Xn converges uniformly to r on G. Then (a) 0 [z, r(dG)] is constant on each component of Ry r{dG). (b) 0[z,r(dG)] =limO[z,rn(dG)], zir(dG) (c) M{z, r) < lim inf ikf (2:, Xn)
10.8. Additional results of FEDERER concerning LEBESGUE I'-area
481
(d) L,(z,G)
>JM(z,r)dz is a homotopy, z ^ Up, and {% \ xt (x) = z} M{z, TQ). Lemmas 9.2.3 and 9.2.3'. To prove (c), sequence of disjoint sub-domains of G. z^r(dD),
, =-{x\ro (x) = z}, then M[z, n) = Proof, (a) and (b) follow from we let Z)i, . . ., Dm, be any finite Then, since 0 [z, T{dD)] = 0 ii
m
Z\0[z,r(aA)]
I < lim inf 2" 10 [2, T(5A)]
<liminfM(^, T^). The result follows. To prove (d), we choose {z^ of class C^ on a domain D G (on W) so that Zn^^z uniformly and Lv[zny G) ->Lv{z, G). For each n, the inequality holds (use (9.2.19)) and it holds in the limit by lowersemicontinuity. To prove (e), let Di, . . ,, D^ be disjoint subdomains of G such that z i To (dDi) for any i. Then r^'îz) OU dDfis empty and 2\0[z,rt{dDi)]\
i
is continuous in t. Thus M (z, TI) < M{z, To). The argument is symmetric. Lemma 10.8.2. Let aiy = C{^) ^^ ^ continuous map from dB{0,i) into dB{0, 1), B{0, 1) C Rjc, k^ 2, let yo and yi^dB{0, 1) with yi =^ yo, let y = a~^(yo), and let F he a domain on dB(0, 1) such that y C. F hut yiia{r). Suppose also that r:z rj(y) is a diffeomorphism from dB{0, 1) {yi} onto Rjcî. Finally, suppose that 0 [0, a{dB{0, 1)}] = 0. Then (10.8.2) 0[ryo,ra{dF)] -0. Proof. Suppose Cn converges uniformly to f on dB{0,\), each C w being a simpUcial map (see below). If n is large enough the hypotheses hold for {an, Cn) It is clear, then, that we may as well assume that C is already simplicial. This means that the domain and range spheres are each subdivided into small curvilinear simplices Ai and Gj, each of which is the radial projection on ^5(0, 1) of the actual simplex having the given vertices. The simplices are supposed to fit together in the usual way as do the simplices of a polyhedron; the subdivision is then completely determined by its vertices. Finally the map C maps each simplex Ai of its domain berycentrically onto some simplex Gj on the range sphere (several different At may be mapped onto the same Gj); that is the points of the flat simplices corresponding by radial projection correspond in that way. If we introduce local coordinates 6}, . . ., 6^-^ on each domain simplex, the formula for the order becomes (since |C| = 1)
0[o,<y{dB{o, m=rj;^Sf^'i-^r-Hdadet)ddt.
^ At Morrey, Multiple Integrals 31
482
The quantity ('^)'~^(dCuld6i) is proportional to the normal to the (k 1)-surface ^ = C(6^). Since this lies on the (k 1)-sphere dB(0, 1), the normal either points towards or away from the origin; the direction does not change within any simplex Ai. Thus 0 [0, a{dB (0, 1)}] =rj^^Z f î {dlldS) dS, ^ = 1
where dS and d^ are the [k 1)-area elements on the domain and range spheres, respectively. Clearly, we may write 0 [0, a{dB(0. 1)}] = r,-i 2" ^*-^ {G}) <i
for each y interior to Gj, coj is the number of x such that C(^) = y, each point being counted with its multiplicity. We shall now show that each 0)^ = 0 [0, a{dB (0, 1)}]. To do this, let G and & be two adjacent range simplices having the (k 2)-simplex y in common. Let Zl^, . . . be those At carried into G and Zl'i, . . . be those carried into G\ From the nature of the mapping, we see that we may order these simplices so that Zl^? and Zl'^ are adjacent for _ = 1, . . ., Pi, say; the remaining Z|2> can be arranged in adjacent pairs, / > the same being true of the remaining A'^. It is clear that e^ = e'^ for each p '<P\ and that the e'^ and e^ for each of the remaining adjacent pair have opposite signs. Thus
Px Pi
Hence all the coj are equal and hence equal to 0 [0, a(S)], S =^ dB(0, \), In our case, the order vanishes. Now, if we consider the piecewise analytic map rcr, we see again that if yo i T(dGj) for any y, the number of times that r a(x) = yo with positive multiplicity equals the number of times this happens with negative multiphcity. Approximating to F by smooth regions and using Lemma 9.2.3 and its proof, we obtain the result (10.8.2). If 3^0 some r(dGj)y the same result holds by continuity. We now prove the following special case of HOPF'S extension theorem
(ALEXANDROFF-HOPF, pp. 498508):
Theorem 10.8.1. Suppose that G and r satisfy the conditions of Definition 10.8.1 with r > 2, and suppose that yoiridG) andO{yo,T(dG)] = 0, Then 3 ^ map r"^ : G ~> Ry, such that ^* (x) = z(x), x^dG, and yo i T* (5). Proof. We shall prove this and the following statement A by induction on r > 2:
10.8. Additional results of FEDERER concerning LEBESGUE v-area
483
A-Suppose a: y = C(x) is a map from dG into S = S*'-i = dB(0, i) (B(0, \) CRv,v > 2) such that 0 [0, a{dG)] = 0, G being diffeomorphic to B(0, 1). Then H a homotopy atiy = h{x,t), where h{x,0) =C(x), h(xJ)^S, h(x,\)=y, x^dG, 0<^<1, y being a fixed {independent of x) point of S. We first show that, for each given v, statement A implies the theorem. Clearly, we may assume that the manifold 9)1 of Definition 10.8.1 is analytic and in some R^. Let us choose analytic domains Gt where GQG G and d Gt is obtained by proceeding a distance Q t along the normals to dGo, Q being small, 0 < ^ < 1, GiG GQ. We may choose z^ analytic and near z on ^ i and then define zi(x) = z(x) on G Go, zi{x) =(\ Zi(x) = z* {x) on Gi, on Gt, 0 <t<i\ t) z*(x) + tz(x)
G Go is supposed to be so close to dG and zf so close to z that yo t r(G Go) U r(dGi). If dzijdx^ 0 on Gi, then ri(Gi) is of measure 0 and we may find a yi arbitrarily close to yo which is not in n (Ci) and T* may be formed by following TI by a radial projection of B{yi, QI) {yi} onto dB[yi,Qi) where yo^B{yi,Qi) and ^ ( y i , î) fi T ( ^ G ) is empty. Otherwise we can find a yi arbitrarily close to yo such that Tf ^ (yi) is a finite set of points C G\. We may then choose a domain F of class Ci 5T5:^(yi) C - T c ^ C Gî, and P is diffeomorphic with 5 ( 0 , 1) (begin by passing an arc of class C^ through the points, going slightly beyond the first and last points, then begin with a thin tube about t h a t arc). On Z', we introduce polar coordinates r, 0 < r < 1, and p on dB{0, \) by mapping on J5(0, 1). Choosing î as above, we define 2-* [x) = zi (x) outside P and z* (r, p) = yi -\- Qi h[p, \ r). We now prove A for r = 2. Clearly we may assume that G = B(0, 1). Then, if we introduce angular coordinates 6 and <p on the domain and range circles, respectively, we may write
a:(p = (p{d),
O<0<2ji;.
li (p^ Ci, then the proof of Lemma 10.8.2 shows that (10.8.3) ^[^'^l'^)] = ^ / (p'(0)de=[q){27t) dB{0,\)), - (p{0)]l27z = 0
{S =
If (p(6) is merely continuous, we see that (10.8.3) holds by approximating. We may then define (rt:(p = (^ -t)<f[Q)+tcp, ^ = ^
271
I cp{d)dd = <p(d).
0
3' 1
484
Now, let us assume that statement A and our theorem are true for 2 < r < y^ and let us consider statement ^ for r = ^ + 1. Again, we may assume that dG = S dB{0, i). We first show that there is a map ci* y = Ci W from 5 to 5 which is homotopic to r and which is such that some point yo >5 is not in R(ri). Then it is clear that ci is homotopic to the map 0*2: y = y where y is on 5 and diametrically opposite to yo- If a already omits some point yo, we may take cxi = a. Otherwise, let <T* ( y = C* (^)) be an analytic map from S to 5 such that |f* {x) C{x)\ < 1/3 for A;^ 5. The homotopy
shows that f* is homotopic to C. If dC^^jdx^ 0, the range of or* has measure 0 and we may take fi = C*- Otherwise, a'^(Z) has measure 0, Z being the set where d^^ldx = 0, and we may choose yo and yi in 5 with yo 9^ yi so that (T*~i(yo) is a finite set. Then we may choose a smooth domain F Z) <y*~-^(yo) such that a* {F) does not contain yi. If we now map vS {yi} diffeomorphically onto Rjc by a map CJO: z = rj(y), it follows from Lemma 10.8.2 that 0[co yo, co a{dF)] = 0 so that there exists a map n : F -^Rjc such that n and co a coincide on dF and co(yo) i Ti(/^). If we define ai(x) = o)-^ri{x), a{xj) a(xj) =a*{x), x^S x^F, ~F +tTi{x)], ai{x) = a'îx), x^SF, we see that yo iai (5). If we define -=co-i[(1 --t)coa'^(x) x^Fj
io</<i,
we see that ai is homotopic to a* and hence to or. Lemma 10.8.3. Suppose G satisfies the conditions of Definition 10.8.1, U CRv, r > 2, U = {y\ | y"'| < 1, ^ = 1, . . ., r} and: G -Û is a mapping 5 0 ^r(dG). (a) Suppose 0 [0, r{dG)] = 0. Then H ^ w<3î> î: Q >dU such that ri[x) = r(x) for x^dG n r-^(dU). (b) Suppose O[0,r{dG)] = xm, K = \, 0 < w < + 00. Then H disjoint compact sets Qi, . . ,, Qm ^^ G ^^<^ ^ mapping TI: G -> 7 $ TI (A;) ^ = T(J%:) O# ( ^ G n r~^{dU), (TI | ft) ^'^ ^ diffeomorphism from Qi onto U with Jacohian having the sign of x, i = i, . . ., m, and ri[G-UQf^]<ZdU.
i
Proof, (a) The mapping n may be obtained from that of Theorem 10.8.1 by following it with a radial projection from 0 onto U; clearly if r(x)^d U, n (x) = r(x). (b) We choose the analytic domains Gt as in the first part of the proof of Theorem 10.8.1 and approximate closely on Gi to r by an analy-
485
tic T ' ; as before we construct r * = r on G Gi, r * = T' on GQ and r*{x) = {\t)r(x) + tr'{x) on dGt By translating r ' slightly, if necessary, we may ensure that 0 ^ T * (Z), Z being the subset of Go where dr'^ldx 0; since m ^ 0, dr'^/dx =^ 0. Then, T * ~ ^ ( 0 ) = {x[, . . ., x'^ and dr'^jdx ^ 0 s each x^. There must be at least m of these, call them ^1, ' -/^m* Q-t each of which ^ T * / ^ A ; has the sign of x. If we choose Q, 0 < Q < 1, small enough and set UQ = {y\ {y^^l <:Q, oc = i, . . ., r}, then T*~^{i[7e) DU Qi where T*(ft) = UQ and A;^$ (2^, ^' = 1, . . ., m. If we let r = G U ft, we see that 0 [0, r * ( 5 r ) ] = 0 so we conclude from Theorem 10.8.1 t h a t there is a map r^'- T -^R^ {ff) which coincides with T* on dF) by following this with a map which is a radial projection from 0 onto dUqiiy^ UQ and is the identity outside UQ, we may ensure t h a t r2'- T -^Rv UQ. Let T3 = T2 on /^ and T3 = r * on each Qi. We then obtain n by following T3 by o" where a is the radial expansion from UQ onto U' for JÛQ and (7(3/) is the radial projection of y on ^ C7 iiyÛ - UQ. Definition 10.8.2. For each sequence I = {ii, . , ., iv) in which the 4 are integers and 1 < n < ^2 < * * * < ^ < A^, we define the projection P^ from Rx^ (N > r) onto Rv by (10.8.4) ze^'(2:) = z^x, oc ="[,.. .,v, w h e r e w = P^[z),
We now recall the notations introduced in Definition 10.5.4 and Lemma 10.5.8. In addition to the statements proved in Lemma 10.5.8, we prove those in the following lemma: Lemma 10.8.4. Suppose G satisfies the usual conditions (a) If w: G -^ C^. C^_^_i andy for some XQ and I, P^ 0 a){xQ) = 0, P^ 0 rjc [cjo {xo), t] = 0 for 0 < ^ < 1, where (10.8.5) rk(coJ)=^(\ -t)co + trjc{a)),
fjc being defined in Lemma 10.5.8. {h) If co: G -^ C^ C^_y_i and, for some XQ and I, P^ 0 co(xo) ^ 0, then Pi 0 r]c[(jo[xQ), /] 9^ 0, 0 < ^ < 1. Proof, (a) We assume that co{xo) ^ some yjcq, where (10 8 6) r ^ f f ' l ^ ^ ' - ^ ^ ' l < 1> j=ji,---Jk, o}3 = ei, d^ even, 6^ odd, 1 < 71 < y2 < * < jk < N, j:^^nyjp,
Since co*"(ô) = 0, the 4 are among the jp. We may, without loss of generality, assume t h a t the d^' = 0 and the 6^ = 1. By renumbering the axes, we may assume t h a t yjcq yk, as defined in the proof of Lemma 10.5.8(a), and assume that ioc = oc. But, since a){xo) i G^_^_;^, we must have co^{xo) ^ 0, i = v + \, . . .,k. Clearly the first v coordinates of ^k [co [XQ)] will still be 0 and the others will not.
486
To prove (b), we again assume that a){xo)^ some yjcq as defined in (10.8.6). If t h e ioc are not included in the jp, then coôc(^xo) = 6^ and rk[o)(ô), Q has ^-coordinate = 6^ ^ 0 for 0 < ^ < 1. If the 4 are included in the jp but all the dh are not 0, the result still holds. But, if all the dj = 0, we may again assume t h a t 4 = oc for 1 < ^ < r and jp = p^ \ <,p '<k. Since P^ o (^{XQ) i^ 0, not all the co*a(:vo) are zero and this still holds after applying r]c{(Dy t). Lemma 10.8.5. Suppose G satisfies the conditions of Definition 10.8.1. Suppose a^Rjsf, e > 0, z: G->Rx^ C^_y^^{a, e) is a mapping ^z(dG) C RN C'jl^_^,(a, s), and A denotes the totality of v-termed sequences d of even integers. Then ^ a Lipschitz map u: G ->RN such that
(10.8.7)
(10.8.8)
l^(^) ^(^)| < 3ey^^, x^G,

Ly (u, Q < 2^ 2^ 2" " M [Pi 0 z, Pi [a) + 86].
I 6eA
Proof. If the right side is + oo, there is nothing to prove. In view of an obvious reduction b y homothetic transformations, we m a y also assume that a = 0 and = 1. By virtue of Lemma 10.8.4 and our hypothesis, it follows that the maps my. G -> C'j^_^ C^_v_i given by coj = r^-u^ 0 ' ' ' oTjsf 0 z, j = i, . . .^ N V are all defined and continuous. If we define w co^_^, it follows from Lemma 10.8.4 and from Lemma 10.8.1 (e) that (10.8.9) M(Pi ow,d) = M{Pi 0 z; 6), w{G)(Z C; for all / and 6. Let % be the family of all components F' of w-^{C'^ C^_i). Since the intersection of C^-^ with one of the open i^-cells i^ of C^ C^_i is just the center of that cell and since w is continuous on a compact set, it follows from the Heine-Borel theorem that the sub-family ^ i of those F' which contain points of t h e compact set w''^ [C'^-v H (C^ C'^-i)] is finite. Each component F' of %i is open on & but it can happen t h a t F' r\dG IS not empty, in which case w[F' C\dG) C some K\ otherwise w[dF') G dK. Let us define F = F' dG] then F is also a domain ( = J " if JT' n ^G is empty). Since 2:(^G) H C^_v is empty, it follows from Lemma 10.8.4 that the center CR of K is not on w(dG). So let us choose a F, let K be t h e open r-cell of Q C^_i containing w{F), let / be t h e unique r-sequence ^P^(K) is a r-cell Ud (= {co I Ico*^ S'^l < 1, 6" even, oc= \,. . ., v}), d = P^ (CR) and let cor be defined b y cor (x) = P^ 0 co(x) P^ CR ; clearly wp'-T ->tJ, U = Uo. Then we note first of all (see 10.8.9) that (10.8.10) Z\0[0,cor(dF)]\ ^M{Pioz,d)
487
where 5(/, d) is the set of all r^w(r)CK for some K^Pi K= Us. From Lemma 10.8.3, it follows that we may choose a map vr : T -^ K ^ (a) if 0 [0, cor(dr)] = 0, then Vr(r) C dK, and (b) if 10 [0, coridF)] \ = m, 0 < m < + oo, there are domains <2i, . . ., Qm with the Qi disjoint and in r such that Vr{r) cR, Vr\ Qi is a diffeomorphism of Qi onto K^ and Vr{r \J Qi)CdK, In both cases, Vr{x) = w{x) for A ; $ ^ / ' n z;-i(îL). Let us define v'{x) = Vr(x) for x on each /^ of gi and i;'(A;) w{x) elsewhere on G. Then, since if F^ g gî, it is clear that w[r) does not contain the center CR of any cell K in C^ Ci_i, it follows that v'[G) contains no such CR- SO, let us define v{x) =rv[v'[x)]y n being defined on each J? cx as the radial projection from CR onto dK. Then (10.8.11) \v{x) -z[x)\<.2]lN, x^G, since if z{x) ^ some K^^ C'j^, v{x) is on ^Xjv"- Moreover there are a finite number i^^, ^ = 1, . . ., P, of domains such that the Ri are disjoint * and C G, and v\Ri is a, diffeomorphism from Ri onto some cell i^ of C^. Finally v{x)^C,_^ for : ^ ; $ G - U i ^ ^ (10.8.12) * Unfortunately v is not Lipschitz (probably). To remedy this, we begin by extending ?:; to a slightly larger domain D Z) G. For each sufficiently small Q, we form the ''mollified" functions VQ by defining a function KQ (X, I) on M for x and | near D so that iê(:t:,f) > 0 , jKg(xJ)di = \, Kg{x,i) = 0 if |;^-||>^,
where KQ^ C^ in (A;, |) and C^ in %, and |^ l | denotes the geodesic distance along W. We then define
vi{x)=fK,(xJ)vî)di, m
x^G,
Then, for ^ > 0, each v^ ^ C^, VQ converges uniformly to v on G, and V VQ converges uniformly to V t" on each compact subset of each Rf. Moreover, for each ;^ > 0, we can choose a sufficiently small ^ > 0 such that Vg > furnishes a diffeomorphism from some RI C Ri onto the corresponding Ki^, the cell of side 2 2;<; concentric with Ki{= v{Ri)), and carries all points outside the R^ into the set (Cl^^,x). We then define u = rorpî 0 . . . 0 r^ 0 VQ where r is defined on each K^C'^ as the expansion from K^ onto K for y on R^ and define r (y) as the radial projection of y from CK into dK, a y^K ^ R^. Then u has all the preceding properties of v and is also Lipschitz. That u is the desired vector follows from its form, (10.8.12), and Theorems 9.1.3, 9.1.4, and 9.2.1.
488
Lemma 10.8.6. Suppose G satisfies the conditions of Definition 10.8.1, z : G -^ Rx^, A has its preceding significance, > 0, and R is the set of a^Rx^^ (10.8.13) ZZ^''^"M\.Pôz,Pia + ed]<.Z (M[Pi oz,y)dy.
ThenA^(R) > 0. Proof. Letting 5 and T be the cubes in i?jv and Rv, respectively, with center at 0 and side 2 s, we obtain f2j2^''M[Pôz,Pi{a)
I s ^
ed]dA^(a) + sd]dA^{a)
- 2^2'' f ^M\Pioz,Pi{a) = 2:2^6^

I y
f 2^M[Pioz,y
a
+ sd]dA^'(y)
= Hz
f M[Pi
oz,y]dy]dA^'{a).
The result follows. Theorem 10.8.2. Suppose G satisfies the usual conditions, z:G-^ R^^, N '>v, and (10.8.14) Then (10.8.15) If N = V, then (10.8.16) /l'+i[^(G)] = 0 and A^lzidG)] = 0, L,{z,G) <2Lr{Pôz,G). M{z,y)dy.
L4z,G) = f
Proof. Let e>0. From (10.8.14) we conclude that A^'+ÎPJ0 z{G)] = 0 for each (v + 1)-sequence / , so that the AT-measure of the set of all jV r 1 dimensional planes of the form x*^ = a^ which intersect z (G) is zero. Consequently the set of all a ^ R^^ ^^(Q O C^-^-i {a, e) is not empty has measure zero. In like manner, the set of ^ in R]sf^z[dG) n C'^-v {(^, e) is not empty has N measure 0. From Lemma 10.8.6, it then follows that there is an a in RN such that (10.8.13) holds and z{G) D C^_y_i [a, e) and z[dG) H C'^-v (^, fi) are both empty. From Lemma 10.8.5 and (10.8.13), it follows that H a Lipschitz map u : G -^RN such that (10.8.7) holds and
U{u,G)
<Zf^\Pôz,y]dy.
489
Since this u exists for each e > 0, it follows that (10.8.17) U{z, G)<2^lM{Pioz, y) dy.
I n case N = v, (10.8.16) follows from (10.8.17) (/ must be (1, 2, . . ., v)) and Lemma 10.8.1 (d). Then (10.8.15) follows from (10.8.16) and (10.8.17). Definition 10.8.3. By an admissible measure 0 we mean one which is the difference of two non-negative measures, for each of which Borel sets are measurable; a set 5 is measurable ^ for each e > 0, 3 a compact
set F a n d a n open set O^FGSCO a n d 0 (0) - 0 {F) < e, 0+
and 0~ being defined as follows: \m{e)= sup 2 mei)\

i
for all measurable partitions e = \J ef. We then define 0+ (e) = [|| 0 il (e) + 0 (e)]l2, 0- {e) = \\\ 0 \\ (e) - 0 (e)]/2. Definition 10.8.4. A continuous mapping / : X -> Y is> said t o be monotone 4 ^ / - i ( y ) is a continuum or a point for each y^f{X). A continuous mapping f: X -^Y is said to be light ^ f-^ {y) is completely disconnected for each y^f(X). Suppose z:X->Z is continuous, X being a connected metric space. Let Y be the collection of continua of constancy of z and, if yi and y^ are two such, define dz (71, 72) = inf diam [z (7)] for every continuum y which intersects both yi and 72Lemma 10.8.7. With this definition, Y is a metric space, the mapping m : X onto Y, defined by the condition that m(x) be the continuum of constancy containing x is monotone, and the mapping ^ :Y ->Z defined by ^(rj) == z [m~^ (rj)] is light. If ^ : Y ~^Z is a light mapping, Y and Z being compact and metric, then H a function 99 (^) - ^ 0 as Q -^0^ if y is a continuum in Z of diameter < g, then any component of f "^ (y) has diameter <=(P{Q).
Proof. The first statements are evident. To prove the last, let {yn} be a sequence of continua in Z whose diameters -> 0 and, for each n, let rjn be a component of C~'^ (yn) and suppose diam ly^j > Co > 0. By choosing a subsequence, still called {n}, we m a y assume that yn ->y and 'i]n -^rj' Clearly rj is R continuum of diameter > 0*0 and, b y continuity, ^{rj) = y. B u t since diam ynÔ we see that y is a point. B u t this contradicts the lightness of ^. Definition 10.8.5. We suppose that G satisfies the conditions of Definition 10.8.1 and that Z:G->RN, LV{Z,G)<C^, A''+^[Z[G)'] = 0,
A^ [z[dG)'\ 0, and z C 0 m where m : G onto X is monotone, and ^ : X -^RN is light. Let (3 be the totality of cells r in Rjs^ such that A^[z(G) C) dr] = 0; evidently all cells r, none of whose faces lie along a certain countable number of hyperplanes belong to (S. Let us define Qo as the totality of components of f-^ (r) and Uo as the totality of comMorrey, Multiple Integrals 3^ ^
490
ponents of z-'^(r) dG for all r ^ . We define Q{ as the totality of components of (Pôf)~i(r) and Ui as the totality of components of {Pi oz)-^{r) f o r a l l r C i ^ v . Lemma 10.8.8. For each / , there is a unique measure 0^, defined on X, such that, for each open set co d X, (10.8.18) 0J(a>) = jO\y,Pioz{dU)'\dy, U = m-^{a)) -dG.
We assume that G, X, z, m, and f satisfy all the conditions above. Proof. We begin by defining 0l{o)) by (10.8.18) for co ^ i ^ o so t h a t U^ Ho- Now if CO, coi, . . .,con allîô, U = m-^((jo) ~ dG, Ui = m-^((jOi) dG, the coi and Uf are disjoint, andco = L)coi, U = U ' Ui, then we
i
see by choosing a smaller smooth domain in each Ui that, ii y i P^ o z(dU) yjpioz{dUi),

n
0[y,Pioz{dU)-\ (10.8.19) *;'
^ZO[y,P'oz{dUi)],
From this, it follows that, if we define
i<îM = sup2'l*SWI
i
for generalized partitions of co like those above, ||ôl| satisfies (10.8.19) for the same types of partitions. Next we may define 0lHw) = \\0i\\ (CO) + 0l{w)]l2,
n
0i-{co) = [im
(>) -
0l{w)]
and then define (co^ disjoint and ^ i3o) 0i (oj) = sup ^ 0 J (co^),
n
U6)iCo), a C Uwi,
CO open, a closed.
0l (a) =. inf 2J 00^ (oJi),
Since C is light and there are arbitrarily small cells r^ ^, there are CO^QQ of arbitrarily small diameter. Thus if a is closed (and hence compact) and CO is open, it follows easily that 0l{co ~ a) ^ 0l(co) 0i{a). We then define the exterior measure 0^^ (5) as the inf of 01^ (co) for open CO D S and the interior measures 0^f (S) as the sup of 0^^ [a) for closed a C S and the measures can be developed classically. Lemma 10.8.9. Suppose G, z, m, C, ^'^d X satisfy all the conditions above. For each I, let Z^ be the set of ^^X'^^^ a continuum y of positive diameter such that P^ o ^ (y) is a point. Then Z^ is of Borel type Fa and (10.8.20) A'[{Plot) [ZI)-] = 0, \\<l>i\\[{C'rH^''ZI)}] = 0.
491
Proof. If we let Z^ be the set of | ^ J^ 5 f 3- continuum y of diameter > l/yfe such that P^ 0 ^(y) is a point, then Z]c is seen to be closed. For a given / and a given y^[P^ o C) (Z^), we see that C(^^) 0 {P^)-^(y) D a continuum of positive diameter, so that A'^[I^{X) pi [P^)~'^{y)] > 0. Hence A^ [[P^ o f) (Z^)] = 0, for otherwise it would follow from Theorem 10.2.3 that yl'+i[f (Z)] = /L^+i[^(G)] > 0. Let Zi be defined as above, suppose C is open on X and co D (f^)~-^ {C"^ (î)} ^-nd suppose U m~^ (co) O dG. Then from the definitions, we find that I n I m-^ I I \J:' {Zi)]} <:\\ll{oy)<JM[y,{Pioz)\U]dy
R
<jM[y,Pioz]dy.
{Pioz){U)
From Theorem 10.8.2, it follows that M{y, P^ oz)^ Li{Rv) so the last result follows by expressing (C^)~^{C^(Z|)} as the intersection of a decreasing sequence {co^^} of open sets. Since our goal in this section is really Theorem 10.8.3 below, we content ourselves in what follows immediately with situations in which G = G is a compact manifold of class C^ without boundary. Lemma 10.8.10. Suppose G = G is a compact manifold of class C^ without boundary, z : G ->Rv is continuous with Lv(z, G) < oo, and otherwise suppose, z, X, f, and m satisfy their usual conditions. Suppose Zn ->z uniformly on G with Zn^ C^ (G) and Lv [zn, G) < Kfor all n and suppose (p is continuous on X. Then (10.8.21) lim [ (p[m\x)-\^
G
dx = f (p(^) d0o{ei),
X
00 being the measure 0^, I = (\, 2, . . ., v), of Lemma 10.8.8. Proof. Since the limit depends only on the limit vector z, it is sufficient to prove this for some subsequence. We may therefore assume that the measures Wn defined by
"Pnie) = 1bzn dx
dx
converge weakly to some measure S^. Let % denote the cells of continuity oiW. li r^(^ and > 0, we may choose r' and r" with f' Crdf Cr" such that (10.8.22) f M{y,z)dy <s, Wn{r'' r') < s for all #.
Let C be an open subset of C^îr) which consists of components of O C-^{r), and let U = m-^{w) = z-^{r). Then z{dU)Cdr ^nd z{U)Cr.
31a*
492
Also, since Zn(O) C r" and Zn(d U) C r" r' for n large f^dx=f0[y,z{dU)]dy+l0[y,z{dU)]dy, 0o(ct>) = jO[y. z[dU)] dy + fOly, z{dU)] dy. n>N.
Since 0 [y, ;^^(^ 17)] = 0 [y, ^(^C7)] for y^r\

OX
it follows from (10.8.22) that for n > N.
^X 0o(co) < 2e
^ '
Since C is light it follows from Lemma 10.8.7 that we may write X = U co^, this being a finite union of co^ where the cot have arbitrarily small diameter and the cof are disjoint and of the type discussed above. The lemma follows. Theorem 10.8.3 (FEDERER'S convergence property [2]). Suppose G = G is a compact mamfold of class C^ without boundary, suppose z : G -^ Rjsf is continuous with Lp (z, G) < oo and, suppose z, X, m, and C satisfy their usual conditions. Suppose that Zn->z uniformly on G with Lv {zn, G) ^K for all n, each Zn ^ C^ [G). Suppose that the (pi are continuous on X, Then (10.8.23) lim fy<pi[m{x)]^^dx=-Jycpj(^)d0l{e,),
where the 0^ are the measures of Lemma 10.8.8. Proof. Let us factor P^ o ^ = r]^ o hj and let hi{X) = Xj. Then hjom is monotone from G to Xj and t]^ is light from Xj to Rv and P^ o z = r}-^ 0 { 0 m). Let 0^^ be the measure of Lemma 10.8.10 on Xj. Then, since Lv (P^ o z, G) < K for each / , it follows from that lemma that, if the (pf are continuous on Xj, (10.8.24) lim f ^
'^~*'^G I
cpf [(hi 0 m) [x)] ^
dx = ^
f <pf (f) ^ 0 * ' ( ^ l )

I Xi
Moreover, for a subsequence still called Zny we have some measures 0^ on X such that, if the q)i are continuous on X, (10.8.25) lim f y cpi[m{x)]^-^dx = f y cpj{i)d0i{e,).
Next we note that if co is a component in Xj of rjj^ (r) where r^î and to Pi (^ (so to speak), then hj'^ (co) consists of a finite number of disjoint open sets coi in X, and a comparison of the definitions shows t h a t 0^^w) =Zô{î)' I^= follows that
(10.8.26) ^0*^ W = ^0 lÎ^ W] > ^ a BOREL set CXj.
493
In addition, b y choosing the 99/(f) =(pi\hi{^)] in (10.8.25), we conclude that (10.8.26) holds with 0 J replaced by 0^. But now, since ||0J|| (Z^) ~ 0 and hi is topological off of Z^, it follows rather easily that 0^ = 0^. Since the limit in (10.8.23) is independent of the subsequence {zn}y the theorem follows. In the paper [2], F E D E R E R proves the following two theorems which we shall merely state, since we haven't found a way to simplify their proofs and do not need them in their full generality. Theorem 10.8.4. Suppose G = G, z, X, M, and ^ satisfy their usual conditions and that the 0Q are the measures of Lemma 10.8.8. Then H a simple v-vector v^ (f) defined for \\ 0 \\-almost all i on X for which \v{S)\ is an integer and 01 {A)^fvi(i)dS
A
{es).
S{e)= A' [C (e)],
where A and e are Borel subsets of X. Theorem 10.8.5. Suppose G = G, z, X, m, C, and 0Q satisfy the conditions above. Then, if A is any open subset of X, \\0[A)\\ = That is, Lebesgue area is additive. Lr[z\m-^[A)],
32
Bibliography
AGMON, S.: The Lp approach to the Dirichlet problem. Ann. Scuola Norm. Sup. Pisa 13, 405448 (1959).
AGMON, S., DOUGLIS, A., and L. NIRENBERG: [l] Estimates near the boundary for
the solutions of elliptic differential equations satisfying general boundary values I, Comm. Pure Appl. Math. 12, 623 727 (1959). [2] Estimates near the boundary for the solutions of elliptic differential equations satisfying general boundary values II, Comm. Pure Appl. Math. 17, 3 5 - 9 2 (1964). AKHIEZER, N . T . : The calculus of variations (trans, from Russian by ALINE H. FRINK), New York: Blaisdell Publishing Co. 1962. ALEXANDROFF, P . , and H. H O P F : Topologie. Berlin: Springer 1935.
ARONSZAJN, N . , KRZYWICKI, A . , and J. SZARSKI: A unique continuation theorem
for exterior differential forms on Riemannian manifolds. Arkiv for Math. 4, 417453 (1962). ARONSZAJN, N . , and K. T. SMITH: [1] Functional spaces and functional completion. Ann. Inst. Fourier (Grenoble) 6, 125 185 (1956). [2] Theory of Bessel potentials I, Studies in eigenvalue problems. Technical Report No. 22, University of Kansas, 1959. BERNSTEIN, S.: [1] Sur la nature analytique des solutions des equations aux derive^s partielles du second ordre. Math. Ann. 59, 2076 (1904). [2] Demonstration du theoreme de M. HILBERT sur la nature analytique des solutions des equations du type elliptique sans I'emploi des series normales. Math. Zeit 28, 330348 (1928). BESICOVITCH, A . S. : Parametric surfaces. Bull. Amer. Math. Soc. 56, 288296 (1950). BiCADZE, A. V.: On the uniqueness of the solution of the Dirichlet problem for elliptic partial differential equations. Uspehi Mat. Nauk (N. S.) 3, 211212 (1948). BLISS, G . A.: Lectures on the calculus of variations. University of Chicago Press, 1946. BocHER, M.: Introduction to Higher Algebra. New York,The Macmillan Co. 1931. BOCHNER, S.: [1] Analytic mapping of compact Riemann spaces into Euclidean space. Duke Math. J. 3, 339354 (1937). [2] Harmonic surfaces in Riemann metric. Trans. Amer. Math. Soc. 47, 146154 (1940). BOERNER, H . : tJber die Legendresche Bedingung und die Feldtheorien in die Variationsrechnung der mehrfachen Integrale. Math. Z. 46, 720742 (1940) (and two earlier papers). BoLZA, O.: Lectures on the calculus of variations. New York: Chelsea Publishing Co.
Bibliography
495
BROWDER, F. E.: [1] On the spectral theory of elliptic differential operators, I. Math. Ann. 142, 22130 (196I). [2] On the spectral theory of elliptic differential operators, II. Math. Ann. 145, 81226 (1962) (and other papers). [3] Non-linear elliptic boundary value problems. Bull. Amer. Math. Soc. 69, 862874 (1963). CALDERON, A. P.: Lebesgue spaces of differentiable functions and distributions. Proc. Symp. Pure Math. IV, Amer. Math. Soc, Providence, R. I., 3349 (196I). , and A. ZYGMUND: [1] On the existence of certain singular integrals. Acta Math. 88, 85 139 (1952). [2] Singular integral operators and differential equations. Amer. J. Math. 79, 901921 (1957). CALKIN, J. W.: Functions of several variables and absolute continuity I. Duke Math. J. 6, 170185 (1940). CARATHEODORY, C : [l jtjber die Variationsrechnung bei mehrfachen Integralen, Acta Szeged 4, 193216, (1929) [2] Variationsrechnung und partielle Differentialgleichungen erster Ordnung. Leipzig: Teubner 1935. CESARI, L . : [1] La nozione di integrale sopra una superficie in forma parametrica. Ann. Scuola Norm. Sup. Pisa 13, 73117 (1947). [2] Condizioni sufficienti per la semicontinuita degli integrali sopra una superficie in forma parametrica. Ann. Scuola Norm. Sup. Pisa 14, 4779 (1948). [3] Condizioni necessarie per la semicontinuita degli integrali sopra una superficie in forma parametrica. Ann. Mat. Pura Appl. (4) 29, 199224 (1949). [4] An existence theorem of the calculus of variations for integrals on parametric surfaces. Amer. J. Math. 74, 265292 (1952). CONNER, P. E.: [1] The Green's and Neumanns' problems for differential forms on Riemannian manifolds. Proc. Nat. Acad. Sci., U. S. A., 40, 11511155 (1954). [2] The Neumann's problem for differential forms on Riemannian manifolds. Mem. Amer. Math. Soc. no. 20 (1956), 56 pp. CoRDES, H.: [1] tJber die erste Randwertaufgabe bei quasilinearen Differentialgleichungen zweiter Ordnung in mehr als zwei Variablen. Math. Ann. 131, 278 312(1956). [2] Vereinfachter Beweis der Existenz einer a priori Holder-konstanten. Math. Ann. 138, 155 178 (1959). [3] Zero order a priori estimates for solutions of elliptic differential equations. Proc. Sympos. Pure Math. 4,157166, Amer. Math. Soc, Providence, R.I., 1962. CouRANT, R.: [1] Plateau's problem and Dirichlet's principle. Ann. of Math. 38, 6 7 9 - 7 2 4 (1937). [2] The existence of minimal surfaces of given topological structure under prescribed boundary conditions. Acta Math. 72, 5198 (1940). [3] Dirichlet's principle, conformal mapping, and minimal surfaces. Interscience, New York, 1950. DANSKIN, J. M.: On the existence of minimizing surfaces in parametric double integral problems in the calculus of variations, Riv. Mat. Univ. Parma 3, 4363 (1952). DEDECKER, P . : Calcul des variations et topologie algebrique, Th^se. University de Liege, Faculte des Sciences, 1957. D E DONDER, T . : Theorie invariantive du calcul des variations. Brussels: Hayez 1935. D E GIORGI, E . : [l] Sulla differenziabihta e I'anahticita delle estremali degli integrali multipli regolari, Mem. Accad. Sci. Torino 3, 2543 (1957).
32*
496
Bibliography
D E GIORGI, E . : [2] Frontiere orientate die misura minima. Seminario Mat. Scuola Norm. Pisa, 196O6l. DEMERS, M . R . , and H. FEDERER: On Lebesgue area, II, Trans. Amer. Math. Soc. 90, 499522 (1959). DENY, J.: Les potentiels d'energie fini. Acta Math. 82, 107183 (1950). D E RHAM, G., and K. KODAIRA: Harmonic integrals (mimeographed notes). Institute for Advanced Stud}^ 1950. DOUGLAS, J.: Solution of the problem of Plateau. Trans. Amer, Math. Soc. 33, 263 321 (1931). DouGLis, A., and L. NIRENBERG: Interior estimates for elliptic systems of partial differential equations. Comm. Pure Appl. Math. 8, 503 538 (1955). D U F F , G . F . D . , and D. C. SPENCER: Harmonic tensors on Riemannian manifolds with boundary, Ann. of Math. 56, 128156 (1952). EELLS, J., and J. H. SAMPSON: Harmonic mappings of Riemannian manifolds. Amer. J. Math. 86, 109160 (1964). EiLENBERG, S., and N. E. STEENROD: Foundations of Algebraic Topology. Princeton; N. J.: Princeton University Press, 1952. EVANS, G . C : [1] Fundamental points of potential theory. Rice Inst. Pamphlets No. 7. Rice Institute, Houston, Texas, 252359 (1920). [2] Potentials of positive mass, I. Trans. Amer. Math. Soc. 37, 226253 (1935). FEDERER, H . : [1] On Lebesgue area. Ann. of Math. 61, 289353 (1955). [2] Currents and area. Trans. Amer. Math. Soc. 98, 204233 (196I). , and W. H. FLEMING: Normal and integral currents, Ann. of Math. 72, 458520 (i960). FLEMING, W. H.: [1] Irreducible generalized surfaces. Riv. Mat. Univ. Parma 8, 251281 (1957). [2] On the oriented Plateau problem. Rend. Circ. Mat. Palermo (2) 11, 6990 (1962). FLEMING, W . H . , and L. C. YOUNG: [1] A generalized notion of boundary, Trans. Amer. Math. Soc. 76, 457484 (1954). [2] Representation of generalized surfaces as mixtures. Rend. Circ. Mat. Palermo (2) 5, 117 144 (1956). [3] Generalized surfaces with prescribed elementary boundary. Rend. Circ. Mat. Palermo (2) 5, 320340 (1956). FRIEDMAN, A.: [1] On the regularity of the solutions of non-linear elliptic and parabolic systems of partial differential equations. Jour. Math. Mech, 7, 4360 (1958). [2] Generalized functions and partial differential equations. Englewood Cliffs, N. J. Prentice Hall Publishing Co.: 1963. FRIEDRICHS, K . O . : [1] On the identity of weak and strong extensions of differential operators. Trans. Amer. Math. Soc. 55, 132151 (1944). [2] On the differentiability of the solutions of linear elliptic equations. Comm. Pure Appl. Math. 6, 299326 (1953). [3] On differential forms on Riemannian manifolds. Comm. Pure Appl. Math. 8, 551-558(1955). FUBINI, G. : II principio di minimo e i teoremi di esistenza per i problemi di contorno relativi alle equazioni alle derivate parziali di ordine pari. Rend. Circ. Mat. Palermo 23, 5884 (1907). FUNK, P.: Variationsrechnung und ihre Anwendung in Physik und Technik. Berlin Gottingen Heidelberg: Springer 1962. GAFFNEY, M . P . : [1] The harmonic operator for exterior differential forms. Proc. Nat. Acad. ScL, U.S.A. 37, 4850 (1951).
Bibliography
497
GAFFNEY, M . P . : [2] The heat equation method of Milgram and Rosenbloom for open Riemannian manifolds. Ann. of Math. 60, 458466 (1954). [3] Hilbert space methods in the theory of harmonic integrals. Trans. Amer. Math. Soc. 78, 426444 (1955). GARABEDIAN, P . R . , and D. C. SPENCER: Complex boundary value problems. Trans. Amer. Math. Soc. 73, 223242 (1952). GARDING, L.: [i] Dirichlet's problem for linear elliptic partial differential equations. Math. Scand. 1, 55 72 (1953). [2] Some trends and problems in hnear differential equations. Proc. Int. Congr. Math., 1958. GEL'MAN, I. V . : The minimum problem for a non-linear functional, Leningrad Gos. Ped. Inst. Uc. Zap. 166, 255263 (1958). GEVREY, M . : Demonstration du theoreme de Picard-Bernstein par la methode des contours successifs; prolongement analytique. Bull, des Sciences Math., (2) 50, 113 128 (1926). GiLBARG, D.: Boundary value problems for non-linear elliptic equations in n variables. Proc. Sympos. Non-linear Problems, Madison, Wis., 1962. GouRSAT, E.: Legons sur Fintegration des equations aux derivees partielles du premier ordre. Paris: Librairie scientifique J. Hermann, 1921. GRAUERT, L . : On Levi's problem and the imbedding of real analytic manifolds, Ann. of Math. 68, 460472 (1958). GRAVES, L . M . : The Weierstrass condition for multiple integral variation problems. Duke Math. J. 5, 656660 (1939). HAAR, A . : tJber das Plateausche Problem. Math. Ann. 97, 127 158 (1927). HADAMARD, J . : Le9ons sur la propagation des ondes. Paris, p. 253 (1903). HESTENES, M . R . : Applications of the theory of quadratic forms in Hilbert space to the calculus of variations. Pacific J. Math. 1, 525 581 (1951). , and E. J. MCSHANE: Theorem on quadratic forms and its application in the calculus of variations. Trans. Amer. Math. Soc. 47, 501 512 (1940). HILBERT, D . : tJber das Dirichletsche Prinzip. Jber. Deutsch. Math. Verein. 8, 184188 (1900). HIRSCHFELD, O . : Uber eine Transformation von Variations- und Randwertproblemen, S.-B. Berlin, Math. Ges. 110122, 33. Jahrgang. HODGE, W. V. D.: [1] A Dirichlet problem for harmonic functionals with applications to analytic varieties, Proc. London, Math. Soc. (2), 36, 257303 (1934). [2] The theory and applications of harmonic integrals, second edition. Cambridge University Press, 1952. HOLDER, E . : Beweise einiger Ergebnisse aus der Theorie der 2 Variation mehrfacher Extremalintegrale, Math. Ann. 148, 214225 (1962). HOLMGREN, E.: tJber Systeme von linearen partiellen Differentialgleichungen. Ofversigt af kongl. Vetenskaps-Akademiens Forhandlingar 58, 91 103 (1901). HoPF, E.: [1] Elementare Bemerkungen iiber die Losungen partieller Differentialgleichungen zweiter Ordnung von elliptischen Typen, S.-B. Preuss. Akad. Wiss. 19, 147152 (1927). [2] Zum analytischen Charakter der Losungen regularer zwei dimensionaler Variationsprobleme. Math. Zeit. 30, 404413 (1929) [3] tJber den Funktionalen, insbesondere den analytischen Charakter der Losungen elliptischer Differentialgleichungen zweiter Ordnung, Math. Zeit. 34, 194233 (1932). HoRMANDER, L.: [1] Linear differential operators. Berlin Gottingen Heidelberg: Springer 1963.
498
Bibliography
HoRMANDER, L.: [2] L^ estimates and existence theorems for the d operator. Acta Math. 113, 89152 (1965). JENKINS, H . : On quasi-hnear elhptic equations which arise from variational problems: J. Math. Mech. 10, 705 727 (1961). , and J. SERRIN: Variational problems of minimal surfaces type I. Arch. Rational Mech. Anal. 12, 185212 (1963). JOHN, F . : [1] On linear partial differential equations with analytic coefficients. Comm. Pure Appl. Math. 2, 209253 (1949). [2] Derivatives of continuous weak solutions of linear elliptic equations. Comm. Pure Appl. Math. 6, 327 335 (1953). [3] Plane waves and spherical means applied to partial differential equations. New York: Interscience Publishers 1955. Kipps, T. C.: The parametric problem for double integrals in the calculus of variations, Univ. of California, Berkeley, Dept. of Math., Tech. Rep. No. 23, ONR Contract Nonr-222 (37). KoDAiRA, K.: Harmonic fields in Riemannian manifolds. Ann. of Math. 50, 587 665 (1949). KOHN, J. J.: [1] Harmonic integrals on strongly pseudoconvex manifolds I. Annals of Math. 78, 112148 (1963). [2] Harmonic integrals on strongly pseudo-convex manifolds II. Annals of Math. 79, 450472 (1964). , and L. NIRENBERG: Non-coercive boundary value problems. Comm. PiTre Appl. Math. 18, 443491 (1965). , and D. C. SPENCER : Complex Neumann problems. Ann. of Math. 66,89140(195 7). LADYZENSKAYA, O . A . , and N. URAL-'TSEVA: [1] Quasi-linear elliptic equations and variational problems with many independent variables, Russian Math. Surveys, London Math. Soc. 16, 1791 (196I). [2] On the smoothness of weak solutions of quasi-linear equations in several variables and of variational problems. Comm. Pure Appl. Math. 14,481 495(1961). [3] Linear and quasi-linear differential equations of elliptic type. Publishing House "Science", Moscow, 1964 (Russian). LAX, P . : On Cauchy's problem for hyperbolic equations and the differentiability of the solutions of elliptic equations, Comm. Pure Appl. Math. 8, 615633 (1955). LEBESGUE, H . : [1] Integrale, longueur, aire (thesis), Ann. Mat, Pura Appl. 7, 231 359(1902). , [2] Sur le probleme de Dirichlet, Ren. Circ. Mat. Palermo 24, 371 402 (1907). L E PAGE, J. T.: Sur les charrips geodesiques des integrales multiples. Bull. Acad. Roy. Belg. CI. Sci. Vs 27, 2746 (1941).
LERAY, J., and J. L. L I O N S : Unpublished.
LEVI, B . : Sul principio di Dirichlet, Rend. Circ. Mat. Palermo 22, 293 359 (1906). L E V I , E . E . : SuUe equazioni lineari totalement elliptiche alle derivate parziali. Rend. Palermo 24, 275317 (1907). LEWY, H . : [1] Neuer Beweis des Analytischen Charakters der Losungen elliptischer Differentialgleichungen. Math. Ann. 101, 609619 (1929). [2] On minimal surfaces with partially free boundary. Comm. Pure Appl. Math. 4, 113 (1951). LicHTENSTEiN, L.: [l] Ubcr den analytischen Charakter der Losungen Zweidimensionaler Variationsproblem. Bull. Acad. Sci. Cracovie, CI. Sci. Mat. Nat. A, 915941 (1912). [2] Zur Theorie der konformen Abbildung. Konforme Abbildung nichtanalytischer singularitatenfreier Flachenstiicke auf ebene Gebiete, Bull. Acad. Sci. Cracovie, CI. Sci. Mat. Nat. A, 192217 (1916).
Bibliography
499
LiCHTENSTEiN, L.: [3] Ncuerc Entwicklung der Potentialtheorie, Encyklopadie der mathematischen Wissenschaften. Vol. 2:3:1, pp. 177377, particularly p. 209 (1918). LIONS, J. L.: [l] Theoremes de traces et d'interpolation I, Ann. Sc. Norm. Pisa 13, 3 8 9 - 4 0 3 (1959). [2] liquations differentielles operationelles et problemes aux limites. Berlin Gottingen Heidelberg: Springer 196I. , and MAGENES: Problemes aux limites non homogenes VII. Ann. Mat. Pura Appl. (4) 63, 201224 (1963). LoPATiNSKii, Y. B . : On a method of reducing boundary problems for a system of differential equations of elliptic type to regular equations. Ukrain. Mat. Zurnal 5, 123-151 (1953). McsHANE, E. J.: [1] On the necessary condition of Weierstrass in the multiple integral problem of the calculus of variations I ; Ann. of Math. 32, 578 590 (1931). [2] On the necessary condition of Weierstrass in the multiple integral problem of the calculus of variations, II. Ann. of Math. 32, 723 733 (1931). [3] On the semi-continuity of d uble integrals in the calculus of variations. Ann. of Math. 33, 460484 (1932). [4] Integrals over surfaces in parame:).ric form, Ann. of Math. 34, 815838 (1933). [5] Parametrizations of saddle surfaces with applications to the problem of Plateau. Trans. Amer. Math. Soc. 35, 718733 (1934). MALGRANGE, B.: Plongement des varietes analytiques reelles, Bull. Soc. Math. France 85, 101113 (1957). MANDELBROJT, S.: Series de Fourier et classes quasi-analytique de fonctions: Collection Borel, Paris, 1935MEYERS, N . : Quasi-convexity and lower semicontinuity of multiple variational integrals of any order. Trans. Amer. Math. Soc. 119, 125149 (1965). MiLGRAM, A., and P. ROSENBLOOM: [l] Harmonic forms and heat conduction I: Proc. Nat. Acad. Sci. U.S.A. 37, 18O184 (1951). [2] Harmonic forms and heat conduction, II. Proc. Nat. Acad. Sci. U.S.A. 37, 4 3 5 - 4 3 8 (1951). MiNTY, G. J.: On the solvability of non-linear functional equations of "montonic" type, Pac. J. Math. 14, 249255 (1964). MIRANDA, C : [1] Sui sistemi di tipo ellittico di equazioni hneari a derivate parziali del primo ordine, in n variabili independente. Atti Accad. Naz. Lincei. Mem. CI. Sci. Fis. Mat. Nat. Ser I (8) 3, 85 121 (1952). [2] Equazioni alle derivate parziali di tipo ellittico. Ergeb. der Math. N.F.H. 2 Berlin. Gottingen. Heidelberg. Springer (1955). MoRREY, C. B., JR.: [1] A class of representations of manifolds, I. Amer. J. Math. 55, 6 8 3 - 7 0 7 (1933). [2] A class of representations of manifolds, I I . Amer. J. Math. 56, 275293 (1934). [3] An analytic characterization of surfaces of finite Lebesgue area, I. Amer. J. Math. 57, 692702 (1935). [4] On the solutions of quasi-linear elliptic partial differential equations. Trans. Amer. Math. Soc. 43, 126166 (1938). [5] Functions of several variables and absolute continuity I I . Duke Math. J. 6, 187215 (1940). [6] Existence and differentiability theorems for the solutions of variational problems for multiple integrals. Bull. Amer. Math. Soc. 46, 439458 (1940)
500
Bibliography
MoRREY, C. B., J R . : [7] Multiple integral problems in the calculus of variations and related topics. Univ. of California Publ. in Math., new ser. 1,1 130 (1943). [8] The problem of Plateau on a Riemannian manifold, Ann. of Math. 49, 807851 (1948). [9] Quasi-convexity and the lower semicontinuity of multiple integrals. Pacific J. Math. 2, 25 53 (1952). [10] Second order elliptic systems of differential equations. Ann. of Math. Studies No. 33, Princeton Univ. Press, 101 159, 1954. [11] A variational method in the theory of harmonic integrals II, Amer. J, Math. 78, 137170 (1956). [12] On the analyticity of the solutions of analytic non-linear elliptic systems of partial differential equations, I. Amer. Jour. Math. 80, 198218, II. 219234 (1958). [13] The analytic embedding of abstract real analytic manifolds, Ann. of Math. 68, 159201 (1958). [14] Second order elliptic equations in several variables and Holder continuity. Math. Z. 72, 146164 (1959). [15] Multiple integral problems in the calculus of variations and related topics. Ann. Scuola Norm. Pisa (III) 14, l61 (I96O). [16] Existence and differentiability theorems for variational problems for multiple integrals. Partial Differential Equations and Continuum Mechanics. Univ. of Wisconsin Press, Madison, 241270, I96I. [17] The parametric variational problem for double integrals, Comm. Pure Appl. Math. 14, 569575 (1961). [18] Some recent developments in the theory of partial differential equations, Bull. Amer. Math. Soc. 68, 279297 (1962). [19] The d-Neumann problem on strongly pseudo-convex manifolds. Outlines of the Joint Soviet-American Symposium on Partial Differential Equations, Novosibirsk, 171 178 (1963). [20] The î-Neumann problem on strongly pseudo-convex manifolds. Differential Analysis. The Tata Institute, Bombay, 1964. [21] The higher dimensional PLATEAU problem on a Riemannian manifold. Proc. Nat. Acad. Sci., U.S.A. 54, 10291035 (1965). , and EELLS, JAMES, JR.: A variational method in the theory of harmonic integrals. Ann. of Math. 63, 91128 (1956). , and L. NIRENBERG: On the analyticity of the solutions of linear elliptic systems of partial differential equations. Comm. Pure Appl. Math. 10, 271290 (1957). MORSE, A. P.: [1] The behavior of a function on its cricical set, Ann. of Math. 40, 6 2 - 7 0 (1939). [2] A theory of covering and differentiation. Trans. Amer. Math. Soc. 55, 205235 (1944). MORSE, M . : [l] The calculus of variations in the large. Amer. Math. Soc. CoUoq. Pub., vol. 18, New York (1934). [2] Functional topology and abstract variational theory. Ann. of Math. 38, 386-449(1937). [3] Functional topology and abstract variational theory, Memor. Sci. Math. no. 92(1939). [4] The first variation in minimal surface theory, Duke Math. J. 6,263289 (1940). , and C. B. TOMPKINS: [1] Existence of minimal surfaces of general critical type. Ann. of Math. 40, 443472 (1939) [2] Unstable minimal surfaces of higher topological types, Proc. Nat. Acad. Sci. U.S.A. 26, 713 716 (1940).
Bibliography
501
MORSE, and C. B. TOMPKINS: [3] Minimal surfaces not of minimum type by a new mode of approximation. Ann. of Math. 42, 6272 (1941). [4] Unstable minimal surfaces of higher topological structure. Duke Math. J. 8, 3 5 0 - 3 7 5 (1941). MosER, J.: A new proof of De Giorgi's theorem concerning the regularity problem for elliptic differential equations, Comm. Pure Appl. Math. 13, 457468 (i960). NASH, J.: [i] C^ isometric imbeddings. Ann. of Math. 60, 383 396 (1954). [2] The imbedding problem for Riemannian manifolds. Ann. of Math. 63, 2063 (1956) . [3] Continuity of the solutions of parabolic and elliptic equations. Amer. J. Math. 80, 931954 (1958). NEWLANDER, A . , and L. NIRENBERG: Complex analytic coordinates on almost complex manifolds. Ann. of Math. 65, 391404 (1957). NiKODYM, O.: Sur une classe de fonctions considerees dans I'etude du probleme de Dirichlet, Fund. Math. 21, 129150 (1933). NIRENBERG, L . : [1] On nonlinear elliptic partial differential equations and Holder continuity, Comm. Pure Appl. Math. 6, 103 156 (1953) [2] Remarks on strongly elliptic partial differential equations. Comm. Pure Appl. Math. 8, 648674 (1955). [3] On elliptic partial differential equations. Ann. Scuola Norm. Pisa (III) 13, 1 - 4 8 (1959). PALAIS, R . , and S. SMALE: A generalized Morse theory. Bull. Amer. Math. Soc. 70, 165 172 (1964). PARS, L . : An introduction to the calculus of variations. London: Heinemann 1962. PAUC, C. Y . : La methode metrique en calcul des variations. Paris: Hermann 1941. Especially p. 54. PETROWSKY, I.: Sur Fanalyticite des solutions des systemes d'equations differentielles. Rec. Math. N. S. Mat. Sbornik 5 (47), 370 (1939)RADEMACHER, H . : tJber partielle und totale Differenzierbarkeit von Funktionen mehrerer Variablen, und iiber die Transformation der Doppelintegrale. Math. Ann. 79, 340359 (1918). RADO, T . : [1] Das Hilbertsche Theorem iiber den analytischen Charakter der Losungen der partiellen Differentialgleichungen zweiter Ordnung, Math. Zeit. 25, 514589 (1926). [2] tJber zweidimensionale regulare Variationsprobleme. Math. Ann. 101, 620 632 (1929). [3] On the problem of least area and the problem of Plateau, Math. Z. 32, 7 6 3 - 7 9 6 (1930). [4] Length and area, Amer. Math. Soc. Colloq. Publ. Vol. 30, Providence, R. I., 1948. RADO, T . , and P. V. REICHELDERFER: Continuous transformations in analysis. Berlin Gottingen Heidelberg: Springer 1955. REIFENBERG, E . R . : [1] Solution of the Plateau problem for w-dimensional surfaces of varying topological type. Acta Math. 104, 192 (I960). [2] An epiperimetric inequality related to the analyticity of minimal surfaces. Ann. of Math. 80, 1 14 (1964). [3] On the analyticity of minimal surfaces, Ann. of Math. 80, 1521 (1964). RELLICH, R . : Ein Satz iiber mittlere Konvergenz, Nachr. Akad. Wiss. Gottingen, Math.-Phys. Kl., 3035 (1930). ROSENBLOOM, L . C. : Linear partial differential equations. Surveys in Applied Math. 5, New York, 1958.
502
Bibliography
SAKS, S.: Theory of the integral, 2nd revised ed., Eng. translation by L. C. Young. Hafner Pub. Co., New York. ScHAUDER, J.: Uber lineare elliptische Differentialgleichungen zweiter Ordnung. Math. Z. 38, 257282 (1934). ScHECHTER, M. I On LP estimates and regularity, Am.r. J. Math. 85, 1 13 (1963). SEELEY, R . T . : [1] Singular integrals on compact manifolds. Amer. J. Math. 81, 658690(1959). [2] Regularisation of singular integral operators on compact manifolds. Amer. J. Math. 83, 265275 (196l). [3] Unpublished. SERRIN, J.: [1] On a fundamental theorem of the calculus of variations, etc.. Acta Math. 102, 1 32 (1959). [2] On the definition and properties of certain variational integrals. Trans. Amer. Math. Soc. 101, 139167 (1961). SHIFFMAN, M . : Differentiability and anlyticity of solutions of double integral variational problems. Ann. of Math. 48, 274284 (1947). SHUTRICK, H . B . : Complex extensions: Quart. J. Math. Oxford, ser. (2) 9, 189201 (1958). SiGALOV, A. G.: [1] Regular double integrals of the calculus of variations in nonparametric form. Doklady Akad. Nauk SSSR 73, 891894 (1950) (Russian). [2] Two dimensional problems of the calculus of variations. Uspehi Mat. Nauk (N. S.) 6, 16101 (1951). Amer. Math. Soc. Translation No. 83. [3] Two-dimensional problems of the calculus of variations in non-parametric form, transformed into parametric form. Mat. Sb. N. S. 38 (80), 183202 (1956). Amer. Math. Soc. Translations, Ser. 2, 10, 319340 (1958). SILVERMAN, E . : Definitions of Lebesgue area for surfaces in metric spaces, Rivista Mat. Univ. Parma 2, 4776 (1951). SMALE, S. : Morse theory and a non-linear generalization of the Dirichlet problem, mimeographed notes. Columbia Univ., New York. SoBOLEV, S. L.: [1] On a theorem of functional analysis. Mat. Sb. (N. S.) 4, 471497 (1938). [2] Applications of functional analysis in mathematical physics. Transl. Math. Mon. Vol 7, Amer. Math. Soc, Procvidene, R. I., 1963. STAMPACCHIA, G. : [1] I problemi al contorno per le equazioni differenziali di tipo ellittico, Atti VI Congr. Un. Mat. Ital., Naples, 2144 (1959); Edition Cremonense, Rome, 196O. [2] Problemi al contorno ellittici, con dati discontinui, dotati die soluzioni holderiane. Ann. Mat. Pura Appl. (IV) 51, 138 (196O). [3] On some regular multiple integral problems in the calculus of variations. Comm. Pure Appl. Math. 16, 383421 (1963). ToNELLi, L.: [1] Sui massimi e minimi assoluti del calcolo delle variazione. Rend. Circ. Mat. Palermo 32, 297337 (1911). [2] Sul caso regolare nel calcolo delle variazioni. Rend. Circ. Mat. Palermo 35, 4973 (1913). [3] Sur une methode directe du calcul des variations. Rend. Circ. Mat. Palermo 39,233264(1915). [4] La semicontinuita nel calcolo delle variazioni. Rend. Circ. Mat. Palermo 44, 167249 (1920). [5] Fondamenti del calcolo delle variazioni. Vols. 1~3, Zanichelli, Bologna. [6] Sulla quadratura delle superficie. Atti Reale Accad. Lincei(6)3,633638(1926). [7] Sur la semi-continuite des integrales double du calcul des variations. Acta Math. 53, 325346 (1929).
Bibliography
503
ToNELLi, L.: [8] L'estremo assoluto degli integral! doppi. Ann. Scuola Norm. Sup. Pisa (2) 3, 89130 (1933). VAN HOVE, L . : Sur Textension de la condition de Legendre du calcul des variations aux integrales multiples a plusieurs fonctions inconnues. Nederl. Akad. Wetensch. 50, 1823 (1947). VisiK, I. M.: Simultaneous quasi-linear elliptic equations with lower order terms Doklady Akad. Nauk 144, 13 16 (1962). , and S. L. SOBOLEV: Nouvelle formulation generale des problemes aux limites. . . Doklady Akad. Nauk 111, 521523 (1956). VoLEViCH, L . R.: On the theory of boundary value problems for general elliptic systems. Eng. translation. Soviet Math. 4, 97100 (1963). W E Y L , H . : [1] Geodesic fields, Ann. of Math. 36, 607629 (1935). [2] The method of orthogonal projection in potential theory, Duke Math. J. 7, 411444 (1940). WHITNEY, H . : [1] Differentiable manifolds. Ann. of Math. 37, 645680 (1936). [2] Geometric integration theory, Princeton University Press, Princeton, N. J., 1957. [3] Elementary structure of real algebraic varieties. Ann. of Math. 66, 545 556 (1957). , and F. BRUHAT: Quelques proprietes fondamentales des ensembles analytiques reelles. Comment. Math. Helv. 33, 132160 (1959). YOUNG, L . C : [l] Some applications of the Dirichlet integral to the theory of surfaces. Trans. Amer. Math. Soc. 64, 317335 (1948). [2] A variational algorithm. Riv. Mat. Univ. Parma 5, 255268 (1954). [3] Generalized surfaces of finite topological type. Memoirs Amer. Math. Soc. No. 17 (1955). [4] Some new methods in two dimensional variational problems with special reference to minimal surfaces, Comm. Pure Appl. Math. 9, 625632 (1956). [5] Contours on generalized and extremal varieties. J. Math. Mech. 11, 6l5646 (1962). ZYGMUND, A . : Trigonometrical series. 2nd ed. Chelsea Pub. Co., New York, 1952.
Index
Admissible boundary coordinate system, 30O Admissible coordinate system, 300 Admissible measure, 489 Algebraic boundary b{X, A), 411 Area integral, 1 Basis (for complex 1 forms), 331 Basis, orthogonal, 332 Boundary b M oi M, 30O Boundary operator, 319 Boundary values, 76 Bounded slope condition, 98 CALDERON'S extension theorem, 74
CALDERON-ZYGMUND inequalities, 56, 58 CAUCHY inequality, 3 5 DIRICHLET'S principle, 5
DiRiCHLET problem, 44 Distance between FRECHET varieties, 351 between mappings [D {zi, ^2)], 350 do{i:,U'), 439 geodesic, 378 point set, 406 Distribution derivatives, 20 Domain BfiR. 174 BohR> 181 Gfir, 181 Lipschitz (of class Cj), 25, 77 of class C^, CJJ, C, analytic, 4 of class D\ 384 strongly Lipschitz, 72 with regular boundary 72, Elementary function, 43 Elliptic systems (of differential equations), 210
EULER'S equations, 2, 7, 8
Class 6:(^, L), 411 Closed forms (p{d(p O), 299 Coerciveness inequality, 253 d cohomology, 322 Complementing condition, 212 Complex-analytic manifold, 317 Complex DiRiCHLET integral d{q),xp), 320 Complex improper integral, 174 Contractible, 414 Convex function, 21 Convex set, 21 Cover (in the sense of Vitali), 409 d{r),d*{r), 374, 380 DiNi condition, 54 Direct methods, 16 DiRiCHLET data, 252 DiRiCHLET growth condition, 32 DiRiCHLET growth lemma, 79 DiRiCHLET integral, 1 DiRiCHLET integral for forms, 291, 320
Exterior co-differential, 290 Exterior derivative, 29O Exterior differential r-forms, 288 Extremal, 32 FEDERER'S convergence property, 492 FEDERER'S multiplicity function, 480 Field of extremals, 14 First differential, 31 First variation, 7.8 Form, closed, 299 complex, of type (p, q), 318 even, 288 exterior differential, 288 in H^, 2p, 288 22-flat, 299 normal part of a, 301 302 odd, 288
Index Form, potential of a, 295, 314 tangential part of a, 301, 302

FRECHET variety, 351
505
Integrand, of type I, 96 of type II, 97

JACOBI condition, 15 JACOBI'S equation, 13 JENSEN'S inequality, 21
on a manifold, 379 oriented, 353

FRIEDRICHS moUifier, 20
Function, absolutely continuous in

the sense of TONELLI, 19, 67
admissible, 8 essentially homogeneous of degree p, 47 mollified, 20 monotone in the sense of

LEBESGUE, 16, 17
KoDAiRA decomposition, 286, 298, 312, 322 / ( r ) , / * ( r ) , 374,380

LAPLACE'S equation, 43 LEBESGUE area of a FRECHET surface,
352
LEGENDRE condition, 2, 10 LEGENDRE-HADAMARD condition, 11
of class C]J, 4 of class if 1(G), H'^{G), 20 quasi-convex, 112 strongly in ifl(di?), 114 strongly quasi-convex, 114
GAFFNEY-GARDING inequality, 293 GARDING'S inequality, 253
Generalized derivatives, 62 Generalized surface, 354 Geodesic cones C{P, S), C{n, S), 409, 410 Geodesic ^-plane, 406 centered at P , 406 Gradient, V-s^, 1
GREEN'S function, 44
Linear functionals on H'^{G), 70 Linear sets (S,^, 312 LiPSCHiTz condition, 4 Lipschitz convergence, 113 LiPSCHiTZ domain (of class Cj), 25 Localized inner product, 301 of complex forms, 319 Manifold with boundary (of class C^, etc,), 300 Mapping, absolutely continuous, 379, 382 light, 489 monotone, 489 of class i i i , 382 of class ^ 1 379 rjc: C^ C ^ - v - i -> C^-v 453 Qv ' ^N C^^-r-l -> ^N, 454 u^{y), 454 Maximum principle, 40, 61 Mean value properties of harmonic functions, 40, 41 Measurable mappings, 378 Minimizing sequences, 16 Minimum conditions (on a general elliptic system), 215 Minus potential (of a form), 310 Multinomial coefficient CQJ, 63 n-th gradient V ^ ^, 4
d -NEUMANN problem, 320
Harmonic field, 294 Harmonic form, 316 Harmonic function, 40 HAUSDORFF outer measure, 350, 408 HERMiTEan metric, 317 HILBERT'S invariant integral, 14
HILBERT transform, 55 HOLDER condition, 4
HOPE'S extension theorem, 482, 483 Inner product, localized (of forms), 302 of complex forms, 319 of complex forms, localized, 319 of forms in S2, 288 of forms in if|, 289 D{(p, tp), 320 {{(P> W))> i{^> W))z' 333 Integral in parametric form, 2, 349 Integrals over oriented FRECHET varieties, 353 Integrand, normal, 92 of a parametric problem, 349
Norm III w |Li?, 337
'llwillsi?, 338
III 9? Ills for complex forms, 342 Normal coordinate system, 402 centered at a point q on S!Ol, 402 Normal part (of a form), 301, 302
506
Operator b, for complex forms, 319 d, for complex forms, 319 L, for complex forms, 321 N, for complex forms, 321 , for complex forms, 320 Order 0[^, T{dG)], 356, 357, 359, 480 Oriented FRECHET variety, 353 Orthogonal basis, 332 Parallel geodesic planes, 443 Parametric representation, 351 Plus potential (of a form), 310 Polyhedral cone, 460
POINCARE'S inequality, 69
Index Singular integrals, 50 Slope functions, 14 SoBOLEV lemmas, 78, 80 SoBOLEv's embedding theorem, 80 SoBOLEV spaces, 19, 20 Space 51, 318 9lo,5l2>, 319 ^jc{G;L;y;d), 194 2), of complex forms, 320 '^PQ, 21 ^ of complex harmonic fields 0(p ='&q) = 0), 321 ^P9, 321 H^(G), 20, 288 ^'. 343 2p, 288 2P, L, 319 Lp{E, fx) (of mappings), 378 ^ + and ^ - , 309 Spherical harmonics, 73, 474 Stokes' theorem, 290 Strongly elliptic system, 252 Strongly LIPSCHITZ domain, 72 Strongly pseudo-convex boundary, 318 Strong relative minimum, 12 Supplementary conditions, 211 Tangential coordinate system, 321 Tangential part (of a form), 301, 302 Topological type (of a FRECHET variety), 351 Transversality conditions, 2 Uniformly elliptic system, 211 Weak convergence in i/^(G), 70 Weak convergence of maps in H^, 385 Weakly differentiable functions, 97 Weak solutions, 33
WEIERSTRASS condition, 2, 12 WEIERSTRASS E- function, 12 W E Y L ' S lemma, 41, 42
Point set distance D {Si, S2), 406

POISSON'S equation, 47
^?o(G) 68
POISSON'S integral formula, 45, 46 Potential, 47 Potential of a form, 259, 314 Principal part (of a differential operator), 210
Problem of PLATEAU in R^, 375, 376
Projection on a geodesic ;^-plane, 435 Projection P^, 485 Properly elliptic system, 211 Quasi-linear equations, 8 Quasi-potential, 86 Reflection principle, 54 Regular mapping, 64 Regular parametric p r o b l e m , 32 Regular variational problem, 8, 11 REIFENBERG cone inequality, 459, 460
REIFENBERG'S (e, Ri) condition, 433 RELLICH'S theorem, 75
RiEMANNian manifold of class C^ Cf 288 Root condition, 211 Second variation, 10 Set (S, Q), 406 Sets Cj^{a, e), Cj^'{a, e), 453 Simplicial cone, 460

Multiple Integrals in The Calculus of Variations Oct 2008

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Multiple Integrals in The Calculus of Variations Oct 2008

Transféré par

Droits d'auteur :

Formats disponibles

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

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

Die Grundlehren der mathematischen Wissenschaften

B. Eckmann und B. L. van der Waerden

Multiple Integrals in the Calculus of Variations

Charles B. Morrey, Jr.

Springer-Verlag Berlin Heidelberg New York 1966

Prof. Dr. B. Eckmann

Prof. Dr. B. L. van der Waerden

Title No. 5113

4.1. 4.2. 4.3. 4.4.

5.1. 5.2. 5.3. 5.4.

126 128 134 143

5.6. 5.7. 5.8. 5.9. 5.10.

157 i64 170 186 191

5.12. A class of non-linear equations Chapter 6

The general two-dimensional parametric problem Chapter 10

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

Multiple Integrals in the Calculus of Variations

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

Another example is the area integral

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

1.2. The plan of the book: notation

1.3- Very brief historical remarks

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

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

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

1.4. The EuLER equations

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

1.4. The EuLER equations

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

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

1.5. Other classical necessary conditions

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

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

0 < y i < Ai/2,

1.6. Classical sufficient conditions

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

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

Letting fi = coy, we see

1.7. The direct methods

1.7. The direct methods

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

1.8. Lower semicontinuity

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

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

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

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

1.8. Lower semicontinuity

Let us now define

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

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

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

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

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

1.10. The differentiability theoy. Introduction

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

1.10. The differentiability theory. Introduction

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

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