Elliptic Problems

in Nonsmooth Domains
P. Grisvard
University of Nice

Pitman Advanced Publishing Program

Boston • London • Melbourne
 PITMAN PUBLISHING INC
1020 Plain Street, Marshfield, Massachusetts 02050

PITMAN PUBLISHING LIMITED

128 Long Acre, London WC2E 9 AN

Associated Companies
Pitman Publishing Pty Ltd, Melbourne
Pitman Publishing New Zealand Ltd, Wellington
Copp Clark Pitman, Toronto

First published 1985

© P. Grisvard 1985
AMS Subject Classifications: 35J, 46E, 65N

ISSN 0743-0329

Library of Congress Cataloging in Publication Data

Grisvard, P. (Pierre)
Elliptic problems in nonsmooth domains.
Bibliography: p.
1. Boundary value problems - numerical solutions.
2. Differential equations, elliptic - numerical
solutions. I. Title.
QA379.G74 1985 515.3'53 84-22827
ISBN 0-273-08647-2

British Library Cataloguing in Publication Data

Grisvard, P.
Elliptic problems in nonsmooth domains.-
(Monographs and studies in mathematics,
ISSN 0743-0329; 24)
1. Differential equations, Elliptic
2. Boundary value problems
I. Title II. Series
515.3'53 QA377
ISBN 0-273-08647-2

All rights reserved. No part of this publication may be reproduced,

stored in a retrieval system, or transmitted, in any form or by any
means, electronic, mechanical, photocopying, recording and/or other-
wise, without the prior written permission of the publishers. This book
may not be lent, resold or hired out or otherwise disposed of by way of
trade in any form of binding or cover other than that in which it is
published, without the prior consent of the publishers

Filmset and printed in Northern Ireland by The Universities Press (Belfast)

and bound at the Pitman Press, Bath, Avon.
Contents

Preface ix
1 Sobolev spaces 1
1.1 Motivation 1
1.2 Boundaries4
1.2.1 Graphs and manifolds 5
1.2.2 Segment property and cone property 10
1.3 Spaces 14
1.3.1 Euclidean space 14
1.3.2 Open subsets of the Euclidean space 16
1.3.3 Manifolds 19
1.4 Basic properties 20
1.4.1 Multiplication and differentiation 20
1.4.2 Density results 24
1.4.3 Continuation, compactness and convexity
inequalities 25
1.4.4 Imbeddings 27
1.4.5 Spaces defined on polygons 33
1.5 Traces 36
1.5.1 Hyperplanes 36
1.5.2 Polygons 42
1.5.3 Maximal domains of elliptic operators 52
1.6 Boundary conditions 62
1.6.1 Normal systems 62
1.7 A model domain with a cut 74
2 Regular second-order elliptic boundary value problems 81
2.1 Foreword 84
2.2 Variational solution of special problems 84
2.2.1 Existence and uniqueness 84
2.2.2 Smoothness 87
V
vi CONTENTS

2.3 A priori estimates 92

2.3.1 An inequality based on the duality mapping 92
2.3.2 An inequality in the half space 97
2.3.3 A general a priori estimate 105
2.4 Existence and uniqueness, the general case 111
2.4.1 The basic result 111
2.4.2 Applications of the Fredholm theory and
the maximum principle 119
2.5 Other kinds of solutions 128
2.5.1 More on smoothness 128
2.5.2 Very weak solution 129
3 Second-order elliptic boundary value problems in convex
domains 132
3.1 A priori estimates and the curvature of the boundary 132
3.1.1 An identity based on integration by parts 133
3.1.2 A priori inequalities for the Laplace operator
revisited 138
3.1.3 A priori inequalities for more general operators 142
3.2 Boundary value problems in convex domains 147
3.2.1 Linear boundary conditions 147
3.2.2 Nonlinear boundary conditions (review) 151
3.2.3 Nonlinear boundary conditions (continued) 156
3.2.4 Oblique boundary conditions 167
3.3 Boundary value problems in domains with turning points 174

4 Second-order elliptic boundary value problems in polygons 182

4.1 Foreword 182
4.2 A priori estimates in an infinite strip 184
4.2.1 Explicit solution by Fourier transform and
consequences 184
4.2.2 Lp bounds for the second derivatives of the
solution 189
4.3 Bounds in a polygon 194
4.3.1 The Lz case 194
4.3.2 The Lp case (p+ 2) 199
4.4 The Fredholm alternative 208
4.4.1 The semi-Fredholm properties 208
4.4.2 The adjoint problem 217
4.4.3 The Fredholm alternative for variational
problems 226
4.4.4 The Fredholm alternative for nonvariational
problems 234
 CONTENTS vii

5 More singular solutions 249

5.1 Behaviour of the derivatives of order higher than two 249
5.1.1 Special data 250
5.1.2 A trace theorem 256
5.1.3 More singular solutions 261
5.2 Operators with variable coefficients 265

6 Results in spaces of Hölder functions 274

6.1 Foreword 274
6.2 A brief review of Hölder spaces 275
6.3 Regular second-order elliptic boundary value problems
revisited 282
6.3.1 The Schauder inequality 282
6.3.2 Smoothness 287
6.4 Second-order elliptic problems in polygons revisited 289
6.4.1 The Schauder inequality in an infinite strip 289
6.4.2 The Schauder inequality in a polygon and its
consequences 295

7 A model fourth-order problem 301

7.1 Introductory results 301
7.2 Singular solutions, the Lz case 305
7.2.1 Kondratiev's method in weighted spaces 305
7.2.2 Getting rid of the weights 321
7.3 Singular solutions, the L, case 328
7.3.1 A priori inequalities 328
7.3.2 Smoothness 335
7.3.3 The related Stokes problem 340
8 Miscellaneous 345
8.1 The Dirichlet problem for a strongly nonlinear equation 345
8.2 Some three-dimensional results (an outline) 356
8.2.1 Edges 357
8.2.2 Conical points and vertices 361
8.3 The heat equation 372
8.4 The numerical solution of elliptic problems with
singularities 384
8.4.1 Weighted spaces and mesh refinements 384
8.4.2 Augmenting the space of trial functions 394
8.4.3 Calculating the stress intensity factor 396
Bibliography 400
Index 409
To Catherine, Olivier, Béatrice and Etienne
Preface

In this book, we focus our attention on elliptic boundary value problems

in domains with nonsmooth boundaries and problems with mixed bound-
ary conditions. So far this topic has been mainly ignored. Indeed most of
the available mathematical theories about elliptic boundary value prob-
lems deal with domains with very smooth boundaries; few of them deal
with mixed boundary conditions. However, the majority of the elliptic
boundary value problems which arise in practice are naturally posed in
domains whose geometry is simple but not smooth. These domains are
very often three-dimensional polyhedra. For the purpose of solving them
numerically these problems are usually reduced to two-dimensional do-
mains. Thus the domains are plane polygons and the boundary conditions
are mixed. Accordingly this book is primarily intended for mathemati-
cians working in the field of elliptic partial differential equations as well as
for numerical analysts and users of such elliptic equations.
Perhaps the main feature of elliptic boundary value problems in a
domain with smooth boundary is the so-called shift theorem'. Let us
describe it on the simplest example, the Dirichlet problem for the Laplace
equation. This will be our model problem throughout this introduction.
Accordingly we consider a function u which is a solution of the equation
Au =f (1)
in a bounded open subset 1 of the two-dimensional Euclidean space R2
Here the function f is given and we assume that u coincides with some
smooth given function g on the boundary I' of M. The shift theorem can
be phrased in the framework of either the Sobolev spaces or the Hölder
spaces. Here, for simplicity, we describe only the Sobolev version.
We denote by WM(R) the space of those functions defined in 1 whose
derivatives up to the order m have their pth power integrable in 2. We
assume that p is strictly greater than 1 and is finite. For the time being,
we also assume that the boundary of 12 is smooth, i.e. is a C°• manifold.
Then when f is given in W(8), the corresponding solution u of the
ix
X PREFACE

problem (1) belongs to Wm2(1). In other words the order of the

Sobolev space is shifted from m to m + 2, by the inverse operator of 4.
The particular case when p = 2 has a simpler proof and is usually the
only one needed by numerical analysts. However, the general case when p
is allowed to differ from 2 (especially p large) is useful when one studies
nonlinear boundary value problems by some kind of linearization or fixed
point method. Most of the current error estimates for the numerical
solution of an elliptic boundary value problem rely on this shift theorem.
Therefore it is particularly important to know whether or not the same
result holds for boundary value problems in a domain with a nonsmooth
boundary.
From now on let us assume that M has one corner. For convenience we
assume that this corner is at the origin of R? and that, in some neighbour-
hood of the corner, & coincides with the sector
G = ¡(r cos 0, r sin e); r > 0,0<0 <a}
in the usual polar coordinates, where w is the size of the angle at the
origin. Otherwise we assume that I' is smooth. For each positive integer
k, we define a function Uk in the following way:
Us = phtlo sin (k+0/w)
when k/w is not an integer and
Uk=pht/wIn r sin (knO/w) + 0 cos (kTO/w)!
when km/w is an integer. It is readily seen that ux is harmonic in & (thus
fk = Aux =0) and that u coincides with a smooth function 8k on I.
Indeed ux vanishes on I near the origin when km/w is not an integer,
while it vanishes on one side of G (for 0 = 0) and coincides with the
polynomial (-1)*corkm/w on the other side of G (for 0 =w) when kri/ is
an integer. Consequently if the shift theorem were valid on &2, Uk ought to
belong to the intersection of all the Sobolev spaces on 2. This would
imply that uk has all its derivatives of all orders continuous in the closure
of S by the well-known Sobolev imbedding theorem. However, it is easy
to check from the explicit formula above for Us that u is I times
continuously differentiable if and only if I is strictly smaller that k/w. A
little extra work shows that u belongs to the Sobolev space W! (O) if and
only if its Sobolev exponent I - 2/p is strictly smaller than kn/w, again.
So much for the shift theorem when & has a corner. Surprisingly
enough, the functions u are all we need to formulate an alternative
statement. Indeed, when f is given in W(M), the corresponding solution
u of the problem (1) has the following property: there exist numbers C
such that
 PREFACE xi

where the k in the summation ranges over all integers such that
T/wEkT/w<m+2-21p,
provided the Sobolev exponent m +2-2/p is not an integer itself. The
limitation on k in the summation means that we exclude the u which
belong to the space Wm+2(1). This result demonstrates that the solution
has the usual regularity far from the corner while it describes accurately
the behaviour near the corner of that part of the solution which does not
belong to the required space.
The terms in the expansion of u above coincide with the terms in the
formal power series derived by Lehman (1959).
The above modified version of the shift theorem does not express a
regularity result in the whole of &. Thus the following question remains
open: under which assumptions of f does the solution u belong to
wm+2(8)? In other words when do the coefficients c vanish? These are
continuous linear functionals of the data f and g. It turns out that they are
local functionals if and only if km/ is an integer. This means that they
only depend on the restriction of the data f and g to any neighbourhood
of the corner. For instance we have
C, =f(0, 0)/t
when w = n/2. On the other hand when k/w is not an integer the
functional c is global; this means that c may not vanish even when the
data f and & are zero near the corner. As a consequence the functional C
depends on the geometry of 1 far from the corner and it is not possible
to make it explicit in such a general case.
Deriving similar modified shift theorems for various boundary value
problems is what this book is about. Let us now proceed with a detailed
description of the various chapters.

Chapter 1

The properties of the Sobolev spaces have been thoroughly investigated

even when they are defined on very rough domains. We review the only
properties we need without proofs and rely on the well-known book by
Necas (1967) for the proofs and references. In dealing with boundary
value problems, one cannot skip a preliminary study of the boundary
values of the functions belonging to Sobolev spaces. Very little is availa-
ble about this question when the boundary is a polygon, although a
complete answer has been given by Nikol'skii (1956, 1958), in the
framework of slightly different spaces more suitable in the approximation
theory. Accordingly we describe completely the boundary properties of
xii PREFACE

functions belonging to Sobolev spaces on domains with polygonal bound-

aries. We include the proofs which turn out to be very similar to
Nikol'skii's proofs. Some extensions of the classical Green formula are also
worked out in the spirit of Lions and Magenes (1963) in the more general
case of nonsmooth domains. This is why Chapter 1 is surprisingly long.

Chapter 2
As a first step toward the generalization of the classical shift theorem, we
attempt to find the minimal assumptions under which one of the classical
methods of proof can be worked out. Our technique is to look at the
problem locally, flatten the boundary by a change of variables, freeze the
coefficients and use partial Fourier transforms. Basically this is the
method followed in Agmon et al. (1959). It turns out that the minimal
assumption under which one obtains solutions in the Sobolev space
Wm (R) is that the boundary I is of class Cm. This means that I can be
locally represented as the graph of a Cm function. Actually one can allow
a boundary of class Cm-1.1 • Consequently a variational solution to a
second-order boundary value problem is shown to belong to W7(8)
provided the boundary is at least of class C'.! This assumption does not
allow a polygonal boundary. We recall that C1.1 denotes the class of the
functions with Lipschitz first derivatives.

Chapter 3

The classical method outlined above includes the proof of an a priori

estimate which looks roughly like this:

[ anlax;" ax;
drsc aul dx + lower-order terms. (2)
Usually we have very poor control of the constant C involved in this
inequality. This is due to the local character of the method of proof.
However in the case when p = 2, an alternative proof based on integra-
tion by parts leads to a very accurate evaluation of the constant C. This is
achieved under very general (possibly nonlinear) boundary conditions on
u, in any n -dimensional domain. Such a proof (for the Dirichlet boundary
condition) goes back to Caccioppoli (1950-51). It turns out that the
constant C depends only on the negative part of the curvature of Г
(roughly speaking). This allows one to take limits with respect to the
domain & and to prove some regularity results in general convex domains
as well as in domains with turning points. Such a technique has been used
for the first time by Kadlec (1964).
 PREFACE xili

Chapters 4 and 5

These chapters are devoted to the proof of a modified shift theorem

similar to the one outlined at the beginning of this introduction for
general boundary value problems for the Laplace equation in a plane
polygon. On each side of the polygon, the condition is either a Dirichlet
or a Neumann or an oblique boundary condition. In Chapter 4 we prove
the regularity of the second derivatives of the solution, while Chapter 5
focuses on the higher derivatives. In addition, some boundary value
problems involving operators with variable coefficients as well as
nonhomogeneous operators are investigated.

Chapter 6
The same boundary value problems as in Chapters 4 and 5 are investi-
gated in the framework of the spaces Cm, (1), i.e. the space of the
functions which are m times continuously differentiable and whose deriva-
tives of order m fulfil a uniform Hölder condition of order ~ throughout
2 (0<0 <1).

Chapter 7
This chapter is focused on the Dirichlet problem for the biharmonic
equation in a plane polygon. We have chosen this particular problem as
our model fourth-order problem because of its importance in several
physical questions (bending of plates, elasticity, fluid dynamics). Again we
prove a suitably modified shift theorem in the Sobolev spaces W. (8).
We follow very closely the method of Kondratiev (1967a) when p = 2.
The shift theorem is also reformulated for the linear Stokes system and
for the stationary Navier-Stokes equations in a plane polygon.

Chapter 8
This chapter includes miscellaneous topics all closely related to the
content of the previous chapters.
First, the Dirichlet problem for a strongly nonlinear elliptic equation in
a convex plane polygon is solved by applying a classical global inversion
theorem following a work by Najmi (1978). The method relies strongly
on the results of Chapters 4 and 5.
xiv PREFACE

The method of Chapter 3 is adapted to the heat equation for a domain

which is not time-like (with only one space variable for simplicity). Here
we follow a work by Sadallah (1976, 1977, 1978).
The third section of Chapter 8 describes without complete proofs the
few results about the behaviour of the solution of a second-order bound-
ary value problem in a three-dimensional polyhedron.
Finally the fourth section is devoted to the consequences of the results
of the previous chapters for the numerical analysis of boundary value
problems.
Singular solutions like the u above have a strong polluting effect on
the classical finite element methods. This difficulty is usually overcome in
two main ways which are described in this section. The first consists (in a
few words) in augmenting the usual spaces of trial functions by the
addition of some of the singular solutions which have been explicitly
calculated here.
The second relies on mesh refinements near the corners. Again the way
the mesh has to be refined is governed by the behaviour of the singular
solutions near the corners. We give here an analysis of the related error
estimates.
In conclusion, let me acknowledge that this book has been strongly
influenced by the outstanding paper by Kondratier about general elliptic
boundary value problems in domains with conical points.
I wish to express my gratitude to the many mathematical colleagues in
the Universities of Algiers, Maryland and Nice, with whom I have had so
many fruitful talks.
Finally I wish to express my sincere appreciation to Pitman Publishing
for their most efficient handling of the publication of this book.

Nice P.G.
August. 1984
1
Sobolev spaces

1.1 Motivation
Why do mathematicians use Sobolev spaces instead of the simpler looking
spaces of continuously differentiable functions?
The most famous Sobolev space is H'(S), the set of all functions u
which are square integrable, together with all their first derivatives, in S,
an open subset of R", the usual n-dimensional Euclidian space. The
derivatives are to be understood in the sense of distributions. It is not
even true that any function in H'(9) is continuous. For instance, the
function
u(x, y)=In #(x2+y2)|1/3

is in H'(Qy, where & is the unit circle in the plane:

However, u is not continuous at (0, 0) and not even bounded. Such spaces
are obviously not easy to handle.
There are several reasons that lead us to use such spaces. The most
significant is perhaps that they appear naturally in the solution of elliptic
boundary value problems by the method of calculus of variations. The
variational approach to the Dirichlet problem in & (with n = 2, say) is the
following. Given a function f in S, we want to find a function u, also
defined in S, a solution of
D›u (x, y) +Du(x, y) = f(x, y) for all (x, y) E S2, (1,1,1)
with the boundary condition
u(x, y) =0 for all (x, y)Eah. (1,1,2)
We now try to view equation (1,1,1) as the equation of a critical point u for
1
2 SOBOLEV SPACES

some functional. One possible functional is obviously

[ID.u?+|D, uff dx dy +I fudrdy. (1,1,3)

If we assume that f is continuous, then F is a differentiable functional
over V, the space of all functions which are continuous together with their
first and second derivatives in S and which vanish on the boundary al.
The Frechet derivative of F at u is

or, after integrating by parts

01* (F'(4), 0)= I-Diu-Diu+fludx dy. (1,1,4)

Consequently, if u is a critical point for F, then u is solution of
equation (1,1,1); u fulfils the boundary condition (1,1,2), simply because
it is an element of V. Now our initial problem is converted into the
problem of finding critical points for F. Obviously F is a convex quadratic
functional on V; its minima are critical points, provided they exist.
Usually a minimum is obtained by considering a minimizing sequence.
This means a sequence Um n = 1, 2, ... in V such that
F(un) Sm (1,1,5)
where
m = WEV
inf F(u).

From (1,1,5), it follows that DUn and DUn n = 1,2, ... are bounded
sequences in Lz(8), the space of all square integrable functions on 8.
Taking in account the boundary condition, an integration shows that Uns
n = 1, 2, ... is also a bounded sequence in L2(8).
We conclude, by using the property of bounded sequences in Hilbert
space, that there exists a subsequence which is weakly convergent.
Consequently, there exist

such that

S Un -> U

(DiM, -543
 1.1 MOTIVATION 3
weakly in L2(9). The theory of distributions shows that U, = Du and
U2 = Du, and therefore u is an element of the Sobolev space H' (8).
Summing up, we have first replaced the original problem (1,1,1) (1,1,2)
by the problem of finding a minimum for the functional F defined by
(1,1,3). This was achieved in the space V, i.e. in the framework of spaces
of twice continuously differentiable functions. Then the construction of a
minimum for F leads to considering a sequence of functions in V (and,
consequently, in C?(8)) which does not converge in C?(8) but which is
convergent in the weak topology of H' (8). Its limit appears naturally as
an element of H'(2).
Actually, it can be proved that there exists a continuous f such that u,
the solution of (1,1,1) (1,1,2), does not belong to C(8). Indeed, assume
the contrary, then the mapping
(1,1,6)
would be a linear mapping from C(R) into itself; here we denote by
C'(R), the space of all continuous functions in & equipped with the
maximum norm. It follows from the closed graph theorem that (1,1,6) is a
continuous mapping. Consequently, there exists a measure du on 1 such
that
(1,1,7)

However, the solution u of problem (1,1,1) (1,1,2) is well known for

some particular domains &2. For instance, when & is the unit circle,
following Courant and Hilbert (1962) we have

where

K(x. y; 5,7) = +27- 08 - 22 108 P

p=V (52+n?)
+2=MI(X-5/83)34(8-1/p7)71.
It follows that

D…D,K(0, 0;5,n) =1, EM

1-p*
T p*
and this is a singular kernel at the origin. Consequently, D,D,u (0, 0) is
4 SOBOLEV SPACES

given by the singular integral

D, D, u(0. 0) =
1-p*- f(5, n) dE dn. (1,1,8)
2 lim
u I fl in p*
f<051
This is in contradiction with (1,1,7).
Now we have at least one good reason for using the space H'(R); but,
what about spaces of functions with more square integrable derivatives?
And, what about spaces of functions of which certain derivatives have pth
power integrable for some p, with 1 < p <0? The former appear in the
variational method for solving equations of order higher than two, while
the latter appear in the solution of nonlinear equations.
There are, of course, several other reasons for using Sobolev spaces in
the solution of partial differential equations and boundary value prob-
lems. One of them is simply the property that the Fourier transform
converts any partial differential equation with constant coefficients into a
division problem. Plancherel's theorem allows one to handle functions
with square integrable derivatives. Unfortunately, there is no counterpart
of Plancherel's theorem for continuous functions. Consequently, the
solutions are built in Sobolev spaces first and their differentiability
properties in the classical sense are obtained through the so-called
imbedding theorems (see Section 1.4.4).
To end this introductory section, let us define the scope of this chapter
about Sobolev spaces. There is a tremendous amount of literature availa-
ble concerning Sobolev spaces. Most of it is quoted in Avantaggiati
(1975) and Triebel (1978), for instance. However, we shall mainly work
with spaces defined on domains whose boundaries are polygons or
polyhedras. On such domains, Sobolev spaces happen to have some
strange properties, which are hard to find in the current literature.
Consequently, the guideline that we shall follow throughout this chapter
is to cite only those properties which are easy to find elsewhere and to
give precise references for their proofs (most of them are to be found in
Necas (1967)). Meanwhile we shall give precise statements together with
complete proofs for all those properties that we need and whose proofs
are too scattered in the present literature. As far as only definitions and
statements of properties are concerned, we attempt to make this chapter
self-contained.

1.2 Boundaries

The properties of functions in a given Sobolev space, H'(8) for instance,

depend very strongly on the properties of the boundary I of the domain
M. Several different points of view have been followed by mathematicians
 1.2 BOUNDARIES 5
for specifying the properties of the boundary I. The purpose of the
present section is to introduce the three main points of view and to
compare them.

1.2.1 Graphs and manifolds

Many authors view (whenever possible) the boundary I of 0 as being
locally the graph of a function . Then the properties of I are specified
through the properties of 4, e.g. continuity, Lipschitz property, differen-
tiability and so on. This is the point of view followed by Aronszjan and
Smith (1961), Adams (1975), Ladyzenskaia and Uralc'eva (1968),
Miranda (1970), Necas (1967) for instance. This last author will be our
usual reference in the present subsection.

Definition 1.2.1.1 Let 2 be an open subset of RM. We say that its

boundary I' is continuous (respectively Lipschitz, continuously differenti-
able, of class Ck.I , m times continuously differentiable†) if for every x € r
there exists a neighbourhood V of x in R" and new orthogonal coordinates
{YI,..., Yn's such that
(a) V is an hypercube in the new coordinates:
V={(y...., Yn)1-a; <y;Sa;15jEn};
(b) there exists a continuous # (resp. Lipschitz,$ continuously differentiable,
of class C.I, m times continuously differentiable) function c. defined
in
V' =<(y., ..., Yn-1)l-a,<y; <a;15jen-13
and such that
le(y') San/2 for every y'=(yi,..., Yn-1)EV'
enV ={y = (y', yn)EV In SQ(y')s, (1,2,1,1)
rnV={y=(y', yn) EV In =Q(y'|}.
In other words, in a neighbourhood of x, & is below the graph of 4 and
consequently the boundary I' is the graph of . We recall that saying that
belongs to the class Ck. means that it is k times continuously
differentiable and its derivatives of order k are Lipschitz continuous.
If an open set & has a continuous boundary I, then & is not on both
sides of I' at any point of F. For instance, R* = R\0} has not a continuous

† in and k are integers #1, possibly equal to +20.

Observe that the word continuous may be omitted there. Indeed, if a function fulfils the
conditions (1,2,1,1), it is easily proved that o has to be continuous.
§ By Lipschitz condition, we always mean uniform Lipschitz condition.
6 SOBOLEV SPACES

Pier.

Figure 1.1

boundary in the sense of Definition 1.2.1.1. Likewise, a domain with a cut

does not fulfil the conditions of Definition 1.2.1.1. However, this defini-
tion allows turning points.
The most important examples in the sequel are the following. A
bounded open subset of R?, whose boundary I' is a polygon, has a
Lipschitz boundary and lacks a continuously differentiable boundary.
Similarly, a bounded open subset of R3, whose boundary I is a polyhed-
ron, has a Lipschitz boundary and lacks a continuously differentiable
boundary.
Many other authors, such as Lions and Magenes (1968) and Hörmander
(1963), prefer to consider (whenever possible) the closure © of the
domain S, as an n-dimensional manifold with boundary, imbedded in RM
They add various regularity assumptions on the manifold.

Definition 1.2.1.2 Let R be an open subset of RM. We say that D is an

-dimensional continuous (respectively Lipschitz, continuously differenti-
 1.2 BOUNDARIES 7
able, of class Ck.', m times continuously differentiable) submanifold † with
boundary in R", if for every x € I there exists a neighbourhood V of « in R"
and a mapping & from V into R" such that
(a) & is injective
(b) & together with ' (defined on (V)) is continuous (respectively
Lipschitz, continuously differentiable, of class Ck.', m times continu-
ously differentiable)
(c) enV =iyER In (y) <O; where In (y) denotes the nth component of
(y).
As a consequence of condition (c), the boundary I of & is defined
locally by the equation In (y) = 0.
In the notations of Definition 1.2.1.1, define & as follows:
(1,2,1,2)
It is easily seen that & fulfils all the conditions in Definition 1.2.1.2 with
the same amount of differentiability for I and is as for 4. In other
words, Definition 1.2.1.1 implies Definition 1.2.1.2 and it is natural to ask
whether the converse is also true. Unfortunately the converse statement is
only partly true. It follows from the implicit functions theorem that
Definition 1.2.1.2 implies Definition 1.2.1.1 provided everything is at
least continuously differentiable. Indeed, we rebuild a function & from a
given , by solving the equation
In (y) =0
with respect to y; where j is chosen in such a way that Din does not
vanish (locally). This is possible when & is continuously differentiable.
Then the chain rule shows that & is continuously differentiable (resp. of
class Ck.I , m times continuously differentiable) when & is continuously
differentiable (resp. of class Ck.!., m times continuously differentiable).
The implicit function theorem does not hold for Lipschitz functions and
the following counterexample will show that Definition 1.2.1.2 does not
imply Definition 1.2.1.1 under the single assumption that & together with
I' is Lipschitz. This counterexample was shown to me by Zerner. We
need some preliminary lemmas.

Lemma 1.2.1.3 The Definition 1.2.1.2 of n -dimensional Lipschitz sub-

manifolds with boundary in R" is invariant under bi-Lipschitz
homeomorphisms.

A homeomorphism n of 1, onto 2 and of a neighbourhood W, of 81

† A continuous manifold is more usually called a topological manifold.
 8 SOBOLEV SPACES

onto a neighbourhood W2 of 82 will be called a bi-Lipschitz

homeomorphism of & onto S if n and n - I are uniformly Lipschitz-
continuous. Lemma 1.2.1.3 is an easy consequence of the chain rule for
the Lipschitz functions due to Rademacher (1919-20).
We now define a bi-Lipschitz homeomorphism from R? onto R? by
n(X) =1X1, 4 (X1) +5623
where
1
311:22k-1 1 1
for52k1511532k
q (t) = 1 1 1
72k+251t
The slope of & is either 3 or -3. Consequently, & is a uniformly Lipschitz
function (with Lipschitz constant equal to 3). This implies that n together
with n ' are uniformly Lipschitz mappings.
Let 2 be defined as follows:

It is clear that M has a Lipschitz boundary according to Definition 1.2.1.1.

A4

-2-2K 12 2k-1
-22k 2-2k

Figure 1.2
 1.2 BOUNDARIES 9
Therefore, § is a two-dimensional Lipschitz submanifold with boundary,
in R?, according to Definition 1.2.1.2, since Definition 1.2.1.1 implies
Definition 1.2.1.2. Next, consider the new domain n(8). This is also a
two-dimensional Lipschitz submanifold with boundary in R?, owing to
Lemma 1.2.1.3. Now we have the following result.

Lemma 1.2.1.4 n(R) has not a continuous boundary according to Defin-

ition 1.2.1.1.
Proof It is obvious from the geometry of n(R) (see Fig. 1.3) that every
linear segment with origin at 0, which cuts I, actually cuts I' at infinitely
many points. This property is true without any restriction on the length of
the segment under consideration. This prevents the existence of a neigh-
bourhood V of 0, together with the existence of new coordinates such
that In V should be the graph of a function as in Definition 1.2.1.1.
Accordingly n(1) lacks a continuous boundary in the sense of Definition
1.2.1.1. &
Summing up, the comparison between Definition 1.2.1.1 and Definition
1.2.1.2 is the following.

In(52).

Figure 1.3
10 SOBOLEV SPACES

Theorem 1.2.1.5 A bounded open subset 1 in Br has a continuously

differentiable (respectively of class (k.', m times continuously differenti-
able) boundary I if and only if M is an -dimensional continuously
differentiable (respectively of class Ck.', m times continuously differenti-
able) submanifold with boundary in R". A bounded open subset 12 in Br
with continuous (respectively Lipschitz) boundary I has a closure D which
is an -dimensional continuous (respectively Lipschitz) submanifold with
boundary in R". The converse statement is not true.

In some special questions, for technical reasons, we shall need uni-

formly Lipschitz unbounded domains of the following kind.

Definition 1.2.1.6 An open subset & of Rn is said to be a uniform

Lipschitz epigraph if there exists new coordinates {Y, ..., n' and an
uniformly Lipschitz continuous function $ of n -1 variables, such that
h={y=(y', Yn) I Un > q (y')}. (1,2,1,3)
Examples of such domains are infinite cones or plane sectors.

1.2.2 Segment property and cone property

In the early stages of the theory of Sobolev spaces, many authors
preferred to describe the boundary properties of the possible domains 2
in a more straightforward fashion. Namely, they required that for each
point * of the boundary I of &, there exists a linear segment C with
origin at x or a cone C with vertex at x, such that C\{x} is contained in 2.
Usually a local uniformity assumption is added (cf. below). This point of
view, adopted by Sobolev, has been followed by Agmon (1965) and
Calderon (1961).

Definition 1.2.2.1 Let & be an open subset of R". We say that D has the
uniform (or restricted) segment property (resp. cone property) if for every
*ET, there exists a neighbourhood V of x in RM and new coordinates
{y...., Yn' such that
(a) V is a hypercube in the new coordinates:
V=fly....., Yn)1-a;<y;<a;15j=n}
(b) y -zES whenever yesn and z€C, where C is the open
segment {(0, ..., O, zn) | 0 <In <h} (resp. C is the open cone
{z = (z', Zn) | (cot e) |z' < In <h} for some 0 € J0, T/2]) for some h > 0.

It is obvious that if M has a continuous boundary according to Defini-

 1.2 BOUNDARIES 11

AY n

an

an/2

y.

Figure 1.4

tion 1.2.1.1, then it has the uniform segment property (just choose h<
an/2). The same way, if & has a Lipschitz boundary, then it has the
uniform cone property. Indeed, this is seen by replacing all the a; by a;/2
in Definition 1.2.1.2 and by choosing h < an/2 together with
1 Un-1
o sint (arctan:Karctan,An
91 ; ... arctan
An
where K is the Lipschitz constant of 4.
The converse statement has been known to be true for a long time by
oral tradition. However, an actual proof has been published only recently
by Chenais (1973). We shall give a transcript of the proof only for
domains having the uniform cone property, because it is slightly simpler
and it is the only one we need in the following sections.
12 SOBOLEV SPACES

Theorem 1.2.2.2 A bounded open subset 2 of R" has the uniform cone
property if and only if its boundary I is Lipschitz.

Proof We have already observed that the condition of having a Lipschitz

boundary is sufficient. Thus, let us consider × € I, assuming that & has the
uniform cone property of Definition 1.2.2.1. We know that x-CeR,
but we can also observe that {x} + C c CO, at least if the distance from x
to CV is greater than h/cos 0; this last condition can always be achieved
by choosing a smaller h. Indeed, if {x}+ CMS is not empty, let y be a
point in the intersection; then yen V since lYn -Xl <h; consequently,
{t-CeR, but this contradicts the fact that {y} - Cx.
From this remark it follows that if we translate the origin of the
coordinates {y. ..., Yn} at * and define a cylinder K by
K={(y', Yn) I -h <Yn Sh, ly'| Shtan el,
then we have
rAKC{(y', Yn) lyn tan A <y' Sh tan Oh;
This means that I cannot 'escape' through the top of K.
We conclude by defining & in the following way
q(y') = sup {yn ((Y', Yn)EIMKI;
4 is defined only for ly'| <h tan 0. Clearly, (y', (y'))e I. Then the cone
property shows that all points (y', Yn) € K with Yn <4 (y') are in I; by
contradiction, as we did previously for x, we show that all points (y', Yn) €
K with Y, > (y') are in CM. Finally, if we consider two points (y', q(y'))
and (z', 4 (z)) on the graph of 4, it follows from the cone property that
Yn -In>- ly'-z'|cot o;
this implies that o is a uniformly Lipshitz function with constant K=
cot A.
We conclude the proof by observing that all the conditions in Defini-
tion 1.2.1.1 are fulfilled when we choose the a; small enough. •

A useful consequence is the following:

Corollary 1.2.2.3 Let I be a bounded open convex subset of R", then S

has a Lipschitz boundary.

Proof Let x, be any point in & and let r >0 be the radius of a ball B
with centre *, contained in &. Since & is convex, all the points ty +
(1-1)z with yeS, ze B, O&t <1, are in M. This shows already that S
has some kind of a cone property but we still need uniformity.
 1.2 BOUNDARIES 13

Figure 1.5

Now fix x € I and choose new coordinates {Y, . .., Yn} with, say, origin
at x and such that xox is parallel to Oyn- Denote by I the distance from K
to x. Then to each y € at a distance less than r/2 from x, we associate a
ball B (y) centred at (y', Yn -1) with radius r/2. Obviously B(y) < B, and
therefore all the points ty + (1 - t)z with z € B (y), 0 st <1 are in 2. The
property in Definition 1.2.2.1 is verified by choosing the a; small enough,
h =1 and sin A = r/21. E
Remark 1.2.2.4 Unfortunately domains with cuts or with turning points
are not well classified by the various previous definitions. Let us consider,
for instance, the following domains in the plane:
#= (* 72)1-13x1511-1 3425 -1x111123
82=((X1, 42) 0<41 521-1582<- (x12) 1123.
The domain & is easily seen to have a continuous (and non-Lipschitz)
boundary according to Definition 1.2.1.1. On the other hand, 2 has not
the segment property of Definition 1.2.2.1; consequently, it lacks a
14 SOBOLEV SPACES

Figure 1.6

continuous boundary. However, Oz is the image of & through the

mapping
4: (*1, 72) +> (*1 +72, 72)
which is a diffeomorphism of class Co of I? onto R?

1.3 Spaces
This section is just a list of the definitions of the Sobolev spaces. We
confine our attention strictly to those spaces which we really need in the
following chapters. Consequently we will not consider any of those
functional spaces such as Besov spaces and Nikolski spaces that are very
closely related to Sobolev spaces and have better properties. The reader
interested in those spaces is referred to Triebel (1978) for instance.
The Sobolev spaces are very easy to define on the whole Euclidean
space. Then a possible definition of Sobolev spaces on a subdomain 2 of
R" with boundary uses restrictions to 2. This is why we treat the spaces
on R" separately and first.

1.3.1 Euclidean space

In what follows, s is any real number and p is a real number such that
1<p <0. We shall denote by m the integer part of s and by o its
fractional part; consequently, s = m to and 0€0 <1.
Definition 1.3.1.1 We denote by W(R") the space of all distributions
(all functions and distributions are complex valued unless otherwise
specified) defined in R", such that
(a) D'UELp (R"), for a &m, when s = m is a nonnegative integer,
 1.3 SPACES 15
(b) ue WE (R") and

(( D°u(x) - D'u(y" dx dy <700

Raxpn
x-yintop
for a =m, when s = m + o is nonnegative and is not an integer.

As usual, Lp (R") is the space of all measurable functions u such that uP

is integrable over R". We define a Banach norm on W(R") by

(1,3,1,1)
in case (a), and by

ule. lal=m
([ D°U(x) - D°u(y)1 dx dy?
x-yintop
RrXRn
(1,3,1,2)
in case (b).
The previous definition is extended to negative values of p by duality as
follows:

Definition 1.3.1.2 For s <0 we denote by W(R") the dual space of

Wa (R"), where p'+q1 = 1.
In the special case where p = 2, we shall use the more common notation
H° (R") instead of W2(R*). The norms defined in (1,3,1,1) and (1,3,1,2)
are Hilbert norms when p = 2.
In some special questions where the use of Fourier transform cannot be
avoided, it is useful to introduce a different kind of spaces as follows.

Definition 1.3.1.3 We denote by H (R) the space of all distributions in

Rh such that

G, * uELp (R"),

where G, is the Bessel potential of order s defined by

(FGs) (4) = (1 +1513)812.

As usual, F is the Fourier transform operator defined by

(Fu) (5) =
1 e-ixtu(x) dx
(277) WIZ
and the star * denotes the convolution product.
16 SOBOLEV SPACES

It is known that HE(R") = W¿(R") (by Plancherel's theorem) for all real
s, and that HP(R") = WD (R") (by Mikhlin's theorem) for all integer m
and 1 <p <∞. Furthermore, it is proved in Taibleson (1964) that
He(R") = W(R"), 1<P=2,
while
WAR")? HA(R"), 2€p<00,
for any real s.
It is also easily checked that W(R") and H$(RM) decrease when s
increases and finally Lions and Peetre (1964) have proved that
W'(R")C H."(Rh)c W'" (RM)

for any real numbers s', s", s' such that

S' >s" > S'''

1.3.2 Open subsets of the Euclidean space

We now deal with S, an open subset of R", possibly different from R"
itself. Our purpose is to extend the definitions given in Section 1.3.1, in
order to define Sobolev spaces on M. In doing that, we can follow
different schools. Here are the three main methods:
(a) We can reproduce Definition 1.3.1.1 by restricting the domain of
integration (replacing R" by 1). This is the point of view in Lions
and Magenes (1960-63) and Necas (1967), for instance.
(b) We can define W(R) as being the set of restrictions to & of all
functions in W(R). This is the point of view in Hörmander (1963).
(c) Finally, following Agmon (1965) and Miranda (1970), we can con-
sider the completion of the space of smooth functions in S, with
respect to the norm in (a).
It turns out that each of these three methods has its advantages. All
three lead to the same spaces when & is smooth enough (we shall give a
precise meaning to this sentence in the next sections). However, for
general domains they may produce three different spaces, which we shall
have to compare.

Definition 1.3.2.1 We denote by W:(M) the space of all distributions u

defined in O, such that
(a) D'UELp (S), for a Sm, when s = m is a nonnegative integer,
 1.3 SPACES 17

(b) u€ Wm (0) and

x-yintop
for a = m, when s = m + o is nonnegative and is not an integer.
We define a Banach norm on W(2) by

(1,3,2,1)
in the case (a), and by

(1,3,2,2)
x-yintop
in the case (b).
We cannot directly reproduce Definition 1.3.1.2 since in general D(S),
the space of all Co functions with compact support in &, is not dense in
W(R). Consequently, the dual space of W(8) cannot be identified to a
space of distributions in M. This is the reason for introducing another
family of Sobolev spaces.

Definition 1.3.2.2 For s > 0, we denote by W(R) the closure of D(R) in

W:(8).
Equivalently, it is the closure in W(8) of all distributions with compact
support in & which belong to W(R).
Then the extension of Definition 1.3.1.2 is the following:

Definition 1.3.2.3 For s <0, we denote by W(1) the dual space of

W= (R), where p'+q1=1.
In the special case when p = 2, we shall also use the common notation,
namely H° (S) instead of W'(R) and H' (R) instead of W2(8). These are
Hilbert spaces.
When s is a negative integer -m, W(R) is also the space of all
distributions T in 2 such that
T=lalEm
& Daf (1,3,2,3)
where fo € Lp (R). The proof can be found in Magenes and Stampacchia
(1958), for instance. An extension of (1,3,2,3) to non integer s is given in
Lions (1961b).
18 SOBOLEV SPACES

For the sake of clarity in the following sections, it will be convenient to

have a specific notation for the space defined by restriction.

Defininions 1.3.2.4 For every s > 0, we denote by W(S) the space of all
distributions in 2 which are restrictions of elements of W(R").

In other words, W(R) is the space of all u la where u€ W(R") and

un is defined by (u In; 4) = (u, $) for all 4 € D(S), where & is the
continuation of ¢ by zero, outside . We define a Banach norm on

inf
uls.p.o=UeWS Ulls.p.ir.
(Rh)
(1,3,2,4)
Uln=u
As obvious consequences of the definition, we have the following
inclusions:
(1,3,2,5)
for every real s>0, and
W… (R) = WM(D)s WI(R) (1,3,2,6)
for every integer m > 0.
Unfortunately we shall need one more kind of Sobolev space whose
technical interest will appear much later.
Definition 1.3.2.5 For every positives, we denote by W(1), the space of
all u € W.(R), where i€ W. (R").

WS (R) is a Banach space for the norm

Nullip.c=Halls.p.som (1.3.2.7)
The only obvious inclusions concerning W(1) are the following
(1,3,2,8)
for all s> 0 and
W. (8) = W(8) (1,3,2,9)
for m integer >O.
The norm defined in (1,3,2,7), although it is the natural one, is
somewhat tricky, because it is the norm induced by W(2) only when s is
an integer.
Lemma 1.3.2.6 Let u belong to W. (8); then
lulim.p.sum.p.r (1,3,2,10)
 1.3 SPACES 19

when s = m is an integer, while

when s=m to is not an integer, where (1,3,2,11)

Pop (x)on=2x-yintop.
dy
It is not easy to describe the weight Pop in general. However, when
is bounded and has a Lipschitz boundary, there exist two constants C1, C2
with O<C, &Cz, such that
Cid(x; r)-OP SPo.p(x) € Cad (x; 5)-00 (1,3,2,12)
where d(x, r) denotes the distance from x to the boundary I of 2. The
same inequalities hold when M is a uniform Lipschitz epigraph (Defini-
tion 1.2.1.6).
1.3.3 Manifolds
In the following sections, we shall need Sobolev spaces on manifolds
which are only (possibly part of) boundaries of open subsets of Rn
fulfilling the assumptions in Definition 1.2.1.1. In other words, they will
be (n -1)-dimensional hypersurfaces in R". Consequently, keeping the
same notations as in Definition 1.2.1.1, the boundary I of 2 is such that
for every x € F, there exists a neighbourhood V of * in R", fulfilling the
condition (a) in Definition 1.2.1.1 and a function fulfilling the condition
(b) such that
rnV={y=(y', yn)eV|n =q(y')}.
Let us define as follows:
@(y) ={y1, . .., Yn-1, Q (Y1, ..., Yn-1)}, (1,3,3,1)
then (Fn V, Ф) is a map of I around x, where we now view I as a
(n -1) - dimensional Lipschitz (respectively continuously differentiable, of
class Ck.I , m times continuously differentiable) submanifold of RM.
The following statement expresses in a precise way the stability of
Sobolev spaces under sufficiently smooth changes of variables. We assume
that 4 is at least a bi-Lipschitz mapping from , onto Oz where & and Sr
are bounded open subsets of R". This hypothesis ensures that Lebesgue
measure is mapped by & or ' to an equivalent image measure.
Lemma 1.3.3.1 Let ue W: (R); assume that if and 4' are of class Ck.
where k is an integer >s-1; then wore W:(R).
20 SOBOLEV SPACES

This property is easy to check with the help of the results in

Rademacher (1919). It is a justification for the following definition.

Definition 1.3.3.2 Let N be a bounded open subset of Rn with a boundary

rof class Ck.1. where k is a nonnegative integer. Let To be an open subset
of I. A distribution u on To belongs to W(To) with Is|<k+1 if uo@€
W(V'N$-'(Ton V) for all possible V and o fulfilling the assumptions in
Definition 1.2.1.1.

Usually distributions are defined only on C manifolds. When To is

only an open subset of a Ck, manifold we consider distributions whose
order is less than or equal to k + 1; those span the dual space of the space
of all Ck. functions with compact support in To. Functions are identified
with distributions by means of the usual injection u -> Tus defined by

where do is the usual hypersurface measure on I (defined provided I is a

Lipschitz hypersurface).
One possible Banach norm on W(F) is

ur>Elluo@illi.p.v;no;cronvi (1,3,3,2)
i=1
where (V; ;)i=1 is any atlas of I such that each couple (Vi, 4;) fulfils the
assumptions of Definition 1.3.3.1 (we recall that ; is defined from 4; by
(1,3,3,1)).
In the particular case when s € J0, 11, one can easily check that any of
the norms defined in (1,3,3,2) is equivalent to

(1,3,3,3)
Г. ГохГо
Ix-yln-1+sp

1.4 Basic properties

This section is only a list of the main properties of the spaces defined
above. We do not report any proof but just indicate easy references
where all the details can be found.

1.4.1 Multiplication and differentiation

The question here is to know sufficient conditions on a function & defined
in O, ensuring that u-> qu is a continuous linear mapping in a given
 1.4 BASIC PROPERTIES 21

We(2). We state here a very simple answer, which is just a straightfor-

ward consequence of the definitions given in Section 1.3. A more com-
plete answer will be given in Section 1.4.4 as a consequence of the
imbedding theorems (see Theorem 1.4.4.2).
We denote by Ck.°(R) (k a nonnegative integer and a €[0, 11) the
space of all functions defined in § which are restrictions to § of functions
of class Ck. defined on the whole of R" which have a compact support.

Theorem 1.4.1.1 Let 4 € Ck."(D) with k+ a ⅞|st when s is an integer

and k+a>s when s is not an integer, then sue W(R) for every
ue W(R), and there exists a constant K = K(4, s, p) such that
puls.p.asKluls.p.o. (1,4,1,1)
An easy consequence is that under the same hypothesis on 4, u -> u is
a continuous linear mapping in W(R) and in W(1). The following
result is also easy to check.

Theorem 1.4.1.2 Let eCk..(1) with k + a >|s| when s is an integer and

k+a> Is| when s is not an integer, then que W; (R) for every u E W. (R).

For a nonnegative integer m, the space Wm(1) is just the space of all
functions defined in &, which are m times differentiable in Lp (R), so to
say. The definition of W(9) for a noninteger s has been stated with the
underlying idea that W(R) should be the space of all functions in S
which are s times differentiable in some sense. Consequently, one could
expect D° to be a continuous linear operator from W(1) into
ws a (R). The extension of the definition of W(8) to negative values of
s was devised with the hope that this rule should hold for every s.
Unfortunately, this is not always true, as we shall begin to show now.
Firstly, D° maps W(R) into W- la (R) provided either lal&s or s €0.
This follows from Definition 1.3.2.1 when a <s. Then, from Definition
1.3.2.2, we see that D° is also a continuous linear operator from W/S (R)
into Ws-lal (1) when lal&s. Transposing this result and remembering
Definition 1.3.2.3, we conclude that Da maps W.(R) into Ws-a (R)
when s €0.
Now it remains to understand how differentiation operates from spaces
with positive order to spaces with negative order. For this purpose it is
enough to consider an elementary differentiation operator D; with respect
to X; with 1 ejen.

Lemma 1.4.1.3 D; is a continuous linear operator from W; (R") into

W.'(R").
22 SOBOLEV SPACES

The only case we have to consider is when 0 <s <1. When p = 2 and
consequently W(R) = H°(R) = H¿(R"), the property is obvious from
Definition 1.3.1.3. Unfortunately, we need another method of proof
when p is not 2. We describe it now. Here R belongs to D(R) and has its
support in the unit ball and integral equal to one.
Lemma 1.4.1.4 Let u€ W(R") and set

U(x, Xn+1) = XER", X+1 > 0;

Kn +1
(1,4, 1,2)
then thiS UPDaUEL, (R'tI), K=1, 2, ..., n+ 1.
Proof We follow the method used in Lemma 5.6, Chapter 2 in Necas
(1967), just adding a weight. We first consider D.U where 1 sk <n. We
have
DRU(X, Xn +1) ==
xh+1
'n+ 1

=- {DeR(! *) Lu(y) -u(X) dy

since obviously

It follows that

and consequently

u(x) -u(y)P dy
00

= (Lu(x) - U(y)° dx dy
sptn x-y+sp
where c is a constant depending on R.
 1.4 BASIC PROPERTIES 23
Now let us consider Dn+, U; we have

Dh+1 U (X, Xn +1) ==

n
yht! xht!
n+1

n
xht X+1

since

{ RI 1
x"=1 Xa+1
by integration by parts. Then each integral in Dn+U is estimated as we
did for DkU. W
Proof of Lemma 1.4.1.3 We consider the bilinear form

and we prove that it is defined and continuous on W(R)x WaS (R"),

where (1/p) + (1/g) = 1. From u and v we construct U and V according to
identity (1,4,1,2). We know from Lemma 1.4.1.4 that
1<k&n+1
Iskent1.
Furthermore, in the topology of Lp (R") we have
lim U= u, lim
Kn+1->0
U=0
X+1-700

while in the topology of La(R) we have

Kn+1
lim V = U, lim V=0.
Xn +1-700

This implies that

== I. ID,U(K. D)D...VIX.D) - D... UM, D)D,WIX, DIdx dr

24 SOBOLEV SPACES

and consequently

perr

k.l= 1

where K is some constant produced by Lemma 1.4.1.4. Taking the limit

when X+1 -> 0, one gets

This proves Lemma 1.4.1.3. -

As a consequence of Lemma 1.4.1.3, it is clear that for u € W. (R), D;u
is the restriction to M of a distribution belonging to W-"(RM). Conse-
quently, a complete answer to the question of whether or not D; maps
W(S) into W'(S) will follow the study of continuation and restriction
properties.

1.4.2 Density results

Here we quote only one basic result proved in Agmon (1959) and Necas
(1967) for instance. We denote by C(R) the space of all functions
defined in which are restrictions to M of C° functions with compact
support in RM

Theorem 1.4.2.1 Let M be an open subset of Rn with a continuous

boundary; then C°(R) is dense in W(R) for all s> 0.

It follows easily from Definition 1.3.2.3 that C°°(S) is dense in W(12)

for all s <0, without any hypothesis on 1.
Moreover, D(R") is dense in W(R") for all s and consequently C°(2)
is dense in W(8) without any assumption on 1.
Another result, closely related to Theorem 1.4.2.1, is the following:

Theorem 1.4.2.2 Let 2 be an open subset of R" with a continuous

boundary, then D(M) is dense in W. (S) for all s > 0.

Together with the identity (1,3,2,10), this shows that when s = m is an

integer and & is a domain with a continuous boundary, then
wm(2) = Wm(S). (1,4,2,1)
 1.4 BASIC PROPERTIES 25

An easy and useful consequence of Theorem 1.4.2.2 is the following:

Proposition 1.4.2.3 Let M be an open subset of R' with a continuous

boundary and let T belong to W(R") with s <0. Then the restriction of T to
§ belongs to the dual space of Wa (8).

Finally we state an improvement of Theorem 1.4.2.1 in the particular

case when s <1/p.

Theorem 1.4.2.4 Let N be a bounded open subset of Rn with a Lipschitz

boundary; then D(R) is dense in W(1) for 0 <s € 1/p.
The same is true when O is uniform Lipschitz epigraph (Definition
1.2.1.6).
In view of Theorem 1.4.2.1 we only have to approximate functions in
C° (R) by functions in D(1) for the norm of W(R). This is easily
achieved by means of a sequence of cut-off functions.
A direct consequence is that under the assumptions of Theorem
1.4.2.4, W$(8) is the same space as W(R), when 0<s $1/p.

1.4.3 Continuation, compactness and convexity inequalities

Now we clarify partly the relation between W(2) and W(M). The
following result is proved in Agmon (1965), Aronszajn and Smith (1961),
Lions (1957), Necas (1967), Stein (1970).

Theorem 1.4.3.1 Let N be a bounded open subset of Br with a Lipschitz

boundary; then for every s> 0 there exists a continuous linear operator Ps
from W(R) into W(R) such that
Paula = u. (1,4,3,1)
The same results hold when 2 is an uniform Lipschitz epigraph or an
infinite strip.
In other words each function u € W(8) is the restriction of a function
Psue W(R"). A counterexample in Lions (1957) shows that the property
may not hold when & has not a Lipschitz boundary. Consequently we
have W$ (R) = WS (S) when M is bounded and has a Lipschitz boundary.
In addition it has been shown in Seeley (1964) and Aronszajn and
Smith (1961) that Ps can be chosen independently of s.
The continuation theorems are powerful tools for extending several
results proved on R" to similar results on a bounded domain with a
Lipschitz boundary. We list some of them now.
26 SOBOLEV SPACES

Theorem 1.4.3.2 Let s' > s">0 and assume that & is a bounded open
subset of R" with a Lipschitz boundary. Then the injection of W. (S) in
Ws" (R) is compact.

(For the sake of convenience here we define W(1) as being Lp (82).)

This result originally due to Kondrasov (1945) is proved in Necas
(1961) for the case in which s' and s" are integers. The extension to
possibly non-integer values of s' and s" may be found in Lions and Peetre
(1964).
The following inequality is closely related to the previous theorem,
through a lemma of Lions (cf. Lemma 2,7, Chapter 1 in Magenes and
Stampacchia (1958)).

Theorem 1.4.3.3 Let s'>S">S'''>0 and assume that N is a bounded

open subset of R" with a Lipschitz boundary. Then there exists a constant C
(depending on So, s', s", s" and p) such that
Mulls.p.re@llulls.p.at KeIs'-s'/6'-s"lulls.p.r (1,4,3,2)
for all ue W: (R).

Such an inequality is also true when 1 = Rh or M is any subset of Rh with

the continuation property of Theorem 1.4.3.1.
This is an interpolation inequality which follows from the similar
inequality on R". See Lions and Magenes (1960-63) for a proof. In the
case when s', s" and s' are integers, this is a particular case of more
general inequalities by Gagliardo (1958) and Nirenberg (1959).

Let us also recall here a related inequality often referred to as Poin-

care's inequality (cf. Necas (1967), for instance).

Theorem 1.4.3.4 Assume that N is a bounded open subset of R". Then

there exists a constant K(S) which depends only on the diameter of I such that

(1,4,3,3)
for all ue W! (1).
Closely related to the inequality in Theorem 1.4.3.3 is the interpolation
theorem (cf. Lions and Peetre (1964)).

Theorem 1.4.3.5 Let II be a continuous linear operator in W(R"),

1<p<+0, sER. Assume that for some t > s the restriction of I to W" (R")
 1.4 BASIC PROPERTIES 27

is continuous in W(R"). Then for every uels, iL, the restriction of Il to

WE(R") is continuous in WU(R").

Due to the continuation property a similar statement holds concerning

the Sobolev spaces on 1 a bounded open subset of R" with a Lipschitz
boundary.
1.4.4 Imbeddings
The most outstanding result about Sobolev spaces is the famous imbed-
ding theorem, derived first by Sobolev himself. The main statement is
this:
Theorem 1.4.4.1 The following inclusions hold:
W$(R") = W'(R") (1,4,4,1)
for t&s and a >p such that s-nlp=t-nlat and
ws(Br)cCk.a(Rn) (1,4,4,2)
for k<s-n/p<k+1, where a=s-k-n/p, k a nonnegative integer.
It is possible to state a weaker result in the limit cases when s-n/p is
an integer, as follows. We have
WH/P(R")cL.(R") (1,4,4,3)
for all q ≥ p, and
Wk+n/P(Rn)cCk-1.a(ph) (1.4.4.4)
for all a €[0, 11, where k is an integer #1.
The proof may be found in any of the references quoted before about
Sobolev spaces.
As a consequence we have the following inclusions
(1,4,4,5)
for t&s, 9≥p such that s-n/p=t-n/g and
W...)cck.a(c) (1,4,4,6)
for k<s-n/p<k+1, a=s-k-n/p, k a non-negative integer. These
inclusions hold without any assumption on 2. As a consequence of
Theorem 1.4.3.1, similar inclusions hold for W(8), when & is a
bounded open subset of R", with a Lipschitz boundary.

† Negative values of t are admitted and W(R) means I, (R").

28 SOBOLEV SPACES

The main interest of these results, in the subsequent sections, is the

following. Assume we are able to build a solution to some given boundary
value problem, which belongs to W(8), with s large enough; then we
know that it is differentiable in the usual sense up to an order (strictly
less) than s - n/p.
A by-product of Theorem 1.4.4.1 is that W(R*) is an algebra for
s> n/p. The more general result which follows has been proved by
Zolesio (1977).

Theorem 1.4.4.2 Let si>s and S2>s be such that either

(1,1-1)20, s; -s> nIp; p, j=1, 2

S1+52-s>n
Or
Si +52-s>n 1 =). 1=1,2,
P2= =0, 5-531
then u, ur> u • v is a continuous bilinear map from We (R") x Wiz (R") into
W(R").

A similar statement holds for Sobolev spaces defined on a bounded

open subset of R", with a Lipschitz boundary. It is a complement to
Theorem 1.4.1.1.
Imbedding results of a different sort deal with weighted spaces. They
are consequences of the well-known Hardy inequality (more precisely
Theorem 330 in Hardy et al. (1952)). Let us recall a convenient statement
of the Hardy inequality. Here we denote by Lp,a (R+) the space of all
measurable functions u defined in R+ such that

Helleran- a lace to de too.

Then, we define two linear operators H and I by
1[
(Hu)(t) =t. U(s) ds

It turns out that H is linear continuous in Lp,a (R+) iff a + 1/p <1, while I
is linear continuous in Lp.a (R) iff a + 1/p> 1. In both cases the norm of
the operator is bounded by la + 1/p-11-1
Theorem 1.4.4.3 Let 1 be a bounded open subset of R" with a Lipschitz
 1.4 BASIC PROPERTIES 29

boundary I and denote by p(x) the distance from a point x to F. Then when
0<s <1/p, we have us eLp (8) for all u € W. (1) and when 1/p<s €1,
we have was eL, (1) for all u € W. (1).

The same result holds (with the same proof) when M is a uniform
Lipschitz epigraph (Definition 1.2.1.6).
This result is proved in Grisvard (1963) for spaces defined on a half
space R4. The result is extended to the case of a Lipschitz domain by
bi-Lipschitz changes of coordinates (use Theorem 1.4.1.1 and Lemma
1.3.3.1).
Iteration of Theorem 1.4.4.3 provides a more complete result concern-
ing the spaces W(1). Since this is not a result easy to find in the current
literature on Sobolev spaces, we give the statement together with a
detailed proof.

Theorem 1.4.4.4 Let I be a bounded open subset of R" with a Lipschitz

boundary F, then for all u € W$(1) such that s-1/p is not an integer, the
following property holds:
pstlaD°uELp (8) (1,4,4,7)
for all ass.
Proof We observe first that by replacing s by s- la, it is enough to prove
the result when a =0.
Let us consider first the case when =R. the nonnegative real axis
and s = m is an integer. Then, for u € D(R) we have
'* (x - y)m -1
u(x)=.
and consequently
u(x) (1,4,4,8)
xm '(m-1)1' xlum'
. (y)ldy,
Hardy's inequality (mentioned above) implies that

By density, this implies the desired result for W(R).

Let us consider then the case when & is still R+ but s=m to is no
longer an integer. We consider now = um, which belongs to W? (R).
We make use of the following strange identity:

0(x) =-W(x)+ (1,4,4,9)

y
30 SOBOLEV SPACES

where

W(x) =- (1,4,4,10)

We first show that x*W€ Lp (R+). Indeed, we have

= 100-15 1040) -06019 de dex

100 W (1) - v (x) |8
- dt dx S/luller. p.re..

Then Hardy's inequality shows that, when o < 1/p,

•y
and consequently * * € Lp (R.). Unfortunately, when o > 1/p, using for-
mula (1,4,4,9) is inconclusive; we therefore use

(1,4,4,11)

with the same w. Now, Hardy's inequality shows that

y
and consequently, again, x € L (R-). Now inequality (1,4,4,8) and one
more application of Hardy's inequality implies that
x M- "u€ Lp (R+).

This is the desired result in W(R.) provided s- 1/p is not an integer.

We conclude by extending this result to a general M. Let us use the
same notation as in Definition 1.2.1.1 and consider a function u whose
support is contained in V. One can always reduce the general case to this
particular case, using a partition of unity. Now for y'€ V' let us set
u. (t) =u(y', q(y')-t).
For almost all y'€ V', we have u, € W(R) and consequently tu, €
L. (R.) with
 1.4 BASIC PROPERTIES 31

where K does not depend on y'. Integrating this inequality in y' leads to

Since 4 is a Lipschitz function, the weight ¢ (y') - In is equivalent to p(y),

the distance from y to I, throughout V. This completes the proof of
Theorem 1.4.4.4.
Corollary 1.4.4.5 Let 1 be a bounded open subset of R" with a Lipschitz
boundary; then when s - 1/p is not an integer we have
(1,4,4,12)
and furthermore, when O<s <1/p we have
W: (8) = W=(8) = W:(R). (1,4,4, 13)
Proof From Lemma 1.3.2.6 and Theorem 1.4.4.4, we know that the
norms of W(8) and of W(1) are equivalent at least on D(2) when
s-1/p is not an integer. Then, from Definition 1.3.2.2 and Lemma
1.4.2.2 we know that D(8) is dense in both spaces W. (2) and W(R).
Consequently, W(R) and W(1) are the completions on D(8) for two
equivalent norms. This proves identity (1,4,4, 12).
We always have W(9)S W$(R). Then when s < 1/p, it follows from
Theorem 1.4.4.3 and Lemma 1.3.2.6 that W. (8) = W(8). This proves
identity (1,4,4,13). &

Another useful consequence of Theorem 1.4.4.4 is the extension of

Lemma 1.4.1.3 to a bounded open domain 2 with a Lipschitz boundary.

Theorem 1.4.4.6 Let S be a bounded open subset of R" with a Lipschitz

boundary. Then D; is a linear continuous operator from W(8) into
ws-'(2) unless s = 1/p.
Proof We have already seen in Subsection 1.4.1 that D; maps W. (1)
into WS-'(S) when either s ≥ 1 or s €0. Let us assume that 0 <s <1. We
know from Theorem 1.4.3.1 that W(8) = W(8). Consequently, for
u€ W;(R), Diu is the restriction to M of a distribution T€ W; '(R").
More precisely, we have
(Diu, v) =(T, O)
for every v € D(R). Furthermore, we have

where (1/p) + (1/g) = 1. This shows that T belongs to the dual space of
32 SOBOLEV SPACES

WA (R). Due to Corollary 1.4.4.5, this last space coincides with WA (8)
provided 1-s+1/g; this means s# 1/p. Therefore Du belongs to
W; '(R) provided s+ 1/p. a
Remark 1.4.4.7 The preceding proof shows that Du belongs to the dual
space of WI(S) when u belongs to W!'(R). This result cannot be
improved as will be shown now. Here, for simplicity, we assume that
p =2.
Proposition 1.4.4.8 The bilinear form (defined for u and o smooth)

(1.4.4,14)

has no continuous extension to H'(10, 1D x H'310, 1D.

This obviously implies that for u€ H 2(J0, ID, w' is not necessarily in
H-13(0, 1D, the dual space of H'2(J0, 1D), since H'200, 1D = H 210, 1D
(see Theorem 1.4.2.4).

Proof Let us assume that (1,4,4,14) is continuous on H'2(0, 10x

H'2(0, 1D; then there exists a constant K such that

for all u, v € D(0, 1]). Now let us assume that v = du, where & is some
cut-off function (de D(0, 11), (0) = 1 and (1) = 0). We have

consequently, there exists a new constant C such that

u(0) <Cul112.2.10.1
for all u € D (0, 1]). By translation we also have

By density, this last inequality implies that all the functions in H'230, 1D)
are continuous near zero. However, the particular function

u(x)=log log*
is an obvious counterexample to this property. Consequently, the form
(1,4,4,14) cannot be continuous. _
 1.4 BASIC PROPERTIES 33

Remark 1.4.4.9 A by-product of the previous proof is that Sobolev's

theorem (1.4.4.1) cannot be improved in the case where s = n/p (here
n =1 and p =2). Indeed, we have
H'200, 1D $ L..010, 1D;
this is the negation of the inclusion (1,4,4,1) in the limit case. The same
way, we have
H'200, 1D ¢ C°10, 11
and this is the negation of (1,4,4,2) in the limit case.
As a last consequence of Theorem 1.4.4.4, we can investigate further
the relations between W(9) and W(R) in the exceptional case when
s-1/p is an integer.
Corollary 1.4.4.10 Let 2 be a bounded open subset of R' with a
Lipschitz boundary; then for all s>0, we have
Ws (8)=. " D'u (1,4,4,15)
p°
where p(x) is the distance from x to the boundary r'of 2 and s= m + 0, m
integer, a €10, 1L.

Proof Let us denote by Z(9) the space on the right-hand side of

(1,4,4,15). From Theorem 1.4.2.2 we know that D(S) is dense in W(1)
for the norm given by (1,3,2,11). This implies the inclusion
W$ (0)= Z(8).
To prove the converse inclusion, first we observe that
D°u = D°u, lal&m (1,4,4, 16)
for all u € Wm (R). This identity is obvious for ue D(M); it is extended to
the whole of W(M) by density. Now let us start with u € Z(82). From
(1,4,4,16) we deduce that u € WM (R"). To prove that u € W. (1), we just
need to check that D°u€ W- m(R"), for lal = m, according to Definitions
1.3.2.5 and 1.3.2.1. This means that
D°uls-m.p,r.
has to be finite, in view of (1,3,2,11). This is obvious from the assumption
that u € Z(2). "
1.4.5 Spaces de ned on polygons
In most of the forthcoming sections, we shall deal with plane domains
whose boundaries are (possibly curvilinear) polygons. First we shall make
fi
34 SOBOLEV SPACES

precise what we mean by curvilinear polygon. Then we shall review

briefly the consequences of the results of the preceding sections, in the
case when & has such a polygonal boundary.
Roughly speaking a curvilinear polygon is a manifold with corners.
More precisely, let us state a definition similar in most respects to Defini-
tion 1.2.1.2.
Definition 1.4.5.1 Let N be a bounded open subset of R?. We say that the
boundary I' is a curvilinear polygon of class Cm, m integer 1 (respectively
(k.a, k integer ≥1, 0 <a €1) if for every x € I there exists a neighbourhood
V of x in R? and a mapping & from V in R? such that
(a) & is injective,
(b) is together with if (defined on (V belongs to the class cm
(respectively Cka),
(c) DMV is either

Or {yER M (y) <0 or 42(y) <0}

where ;(y) denotes the ith component of y.

Any domain & fulfilling the requirements in Definition 1.4.5.1 has a

Lipschitz boundary according to Definition 1.2.1.1. Consequently, the
Sobolev spaces on M will have all the properties already listed for
Sobolev spaces on bounded domains with Lipschitz boundary. However,
the actual advantages of these domains will appear clearly in the next
section dedicated to the trace theorems.

Theorem 1.4.5.2 Let M be a bounded open subset of R? whose boundary

I is a curvilinear polygon. Then we have the following inclusions and
identities:
(a) W. (8) = W: (1) = W: (1) = W. (8) for s > 0.
(b) WS (8) = W(8) for s - 1/p non-integer,
(c) W. (8) = W. (1) for s < 1/P,
(D W: (R) = SUEW.(D)
p°
for s = m + o, m a non-negative integer. Furthermore, C°°(S) is dense in
W'(R) and D(R) is dense in W$(1) for all s > 0. We also have

(e) W: (2) = W' (8), 5-==1- 2

 1.4 BASIC PROPERTIES 35

and

(f) W$(1) c Ck.a (R), k+ a =s - p

2
for s-2/p>0, not an integer.
In practice, we shall often have to check whether or not some concrete
functions belong to a given Sobolev space. For instance, we shall deal
with functions which have an isolated singularity. A criterion for such
functions is the following.

Theorem 1.4.5.3 Let S be a bounded open subset of R7, whose boundary

F is a curvilinear polygon. Assume that 0 € I. Let V be a neighbourhood of
O such that
vnA s{(r cos 6, r sin A) r >0, a s0=b)
with b - a <2m. Finally let u be a function which is smooth in §\0' and
which coincides with
p°p (0)
in VM&, where 4 €Ca, b]). Then
2 (1,4,5,1)
u € W$(1) for Re a >s - p
while

u$ Wi(S) for Re a Ss 2 (1.4.5.2)

p
when Re a is not an integer.

It is very easy to check these inclusions when s is an integer. However,

when s is not an integer the double integrals which appear in the norm
(1,3,2,2) are so complicated that it is almost impossible to estimate them
directly. The method of proof devised by Babuska consists in proving that
u E W"'(S) for m integer >s and r<p and then using the Sobolev
imbeddings. We get thus all the desired results when p >2. The general
proof for p <2 makes use of weighted Sobolev spaces; we skip it since we
shall mostly need inclusion (1,4,5,1) when p >2.

Proof A derivative of order m of u behaves as a finite sum of functions

p°-m (0), where de Ca, bJ, in VMO (This is true unless a is an
integer where cancellations can occur.) Consequently its rth power is
36 SOBOLEV SPACES

integrable in O iff Re a > m - 2/r. In other words

u€ W'(S) if Re a > m - 2r

By Sobolev imbeddings it follows that

(a) uE W. (12)
provided there exists an integer m #s and an r€]1, p] such that
Re o>m-2/r and m -2/r = s-2/p. This last condition is always
fulfilled when p ≥ 2.
(b) u$ W: (8)
when there exists an integer m&s and an rap such that Re as
m-2/r and m-2/r=s-2/p. This last condition is always fulfilled
when p =2.
So far, we have proved (1,4,5,1) when p >2 and (1,4,5,2) when p 52.
We shall not attempt to extend (1,4,5,1) to all values of p <2 since this
requires the use of weighted spaces as we already mentioned it earlier.
However, the extension of (1,4,5,2) to all p >2 is simple at least when
s-2/p is not an integer. Indeed, a derivative of order m of u is clearly
Hölder continuous with exponent Re a - m when m <Re a €m + 1, and
it is not Hölder continuous with a larger exponent. Consequently, the
second Sobolev imbedding implies (1,4,5,2) in the remaining cases when
Rea<s-2/p.
Remark 1.4.5.4 Similar results hold for the functions p° (In r) ¢ (0).

1.5 Traces
Among the many consequences of Sobolev's imbedding is the continuity
of the functions belonging to W(S) when s > n/p. It is even continuity up
to the boundary, which allows one to consider the values on the bound-
ary, of such functions. This is obviously of the utmost importance in the
study of boundary value problems. However, if we agree to consider
boundary values of functions in a weaker sense, we can relax the
condition on s. This is the purpose of the present section.

1.5.1 Hyperplanes
Here, for the sake of convenience, we denote by n the operator defined
by
(YnU) (X1 ..., X-1) = U (X1, ...., Kn -1, 0)
 1.5 TRACES 37

when u is a smooth function, continuous, say. In other words, we want to

consider the restriction of u on the hyperplane *n =0. The basic fact
about Yn is that Ynu is well defined as soon as u € W(R") when s > 1/p.
We observe that this condition is less restrictive than the condition s > n/p
which is necessary for ensuring the continuity with respect to all variables.
The proof of the following result may be found in Agmon (1965) when
p = 2, in Necas (1967) when s is an integer and in Uspenskiï (1962) in the
general case

Theorem 1.5.1.1 Assume that s - 1/p is not an integer and that s - 1/p =
k+0, 0<0<1, k an integer =0. Then the mapping

which is defined for u € D (R"), has a unique continuous extension as an

operator from

W(Rh) onto I W5-1/0 (R"-1).

i =0

This operator has a right continuous inverse which does not depend on p.

This result is easily extended to the case when mr-1 is replaced by an

(n -1)-dimensional submanifold of R", which is smooth enough. This
simply uses changes of variables. More precisely, when I' is the Lipschitz
boundary of a bounded open subset of R, we define a normal vector field
as follows. Let us keep the same notation as in Definition 1.2.1.1; then a
unit outward normal vector v is defined ae. (for the usual surface
measure on Г) by
{- Did (y'), ..., - Dn-19 (y), 1}

for y'€ V'. This vector field is easily extended to the whole of V by
defining it independently of X- Finally, by a partition of unity, we define
an L* vector field in a neighbourhood of , such that w is the unit
outward normal a.e. on I. Then we observe that when the boundary of 12
is of class Ck. , the vector field w is only of class Ck-1.1. Now we denote
by y the operator defined by (yu) = ur when u is a smooth function.
Theorem 1.5.1.2 Let O be a bounded open subset of in with a Ck,
boundary I. Assume that s - 1/p is not an integer, sSk + 1, s-1/p=1+0,
0<o<1, I an integer ≥0. Then the mapping
ди d'u)
dv
38 SOBOLEV SPACES

which is defined for u € CK. '(D), has a unique continuous extension as an

operator from

W°(R) onto I W5-1/p(F).

i =0
This operator has a right continuous inverse which does not depend on p.

The particular case when s = 1 and k = 0 was proved a long time ago by
Gagliardo (1957).

Theorem 1.5.1.3 Let 1 be a bounded open subset of R" with a Lipschitz

boundary F. Then the mapping u -> yu which is defined for u € CO. ' (D),
has a unique continuous extension as an operator from W! (1) onto
WE"P(F). This operator has a right continuous inverse independent of p.

In the sequel we shall always denote by y the extended operator

defined on the whole of W(1) and we shall call it the trace operator.
In addition it is also possible to characterize the kernel of the trace
operator y and even of the mapping
d'u)
au'S
in several cases.

Theorem 1.5.1.4 Assume that s - 1/p is not an integer and that s - 1/p =
kta, 0<0<1, k an integer >0. Then u € W(R) if and only if u€
W(R') and

Here we denote by R, the half space defined by *n > 0. By changing

variables, we deduce the following result.

Theorem 1.5.1.5 Let M be a bounded open subset of on with a Ckil

boundary I. Assume that s - 1/p is not an integer and that s- 1/p = 1+0,
O<o<1, I an integer >0. Then for s&k +1, ue W; (R) if and only if
u€ WS(R) and

yu =
ди d'u
Yaw

Remembering Corollary 1.4.4.5, we see that Theorem 1.5.1.5 implies

also the following result.
 1.5 TRACES 39

Corollary 1.5.1.6 Let M be a bounded open subset of Dr with a Ck,1

boundary I. Assume that s&k+ 1 and that s-1/p is not an integer. Let
s-1/p=1+0, 0<0<1, I an integer ≥0. Then u € W. (R) if and only if
u € WS (2) and
du d'u
yU =- Y ==...= y.
dv
In some special problems related to the study of mixed boundary
conditions on a regular boundary, it will be convenient to split the
boundary I into pieces and correspondingly to split the trace operator y.
The related trace theorems follow. We first consider functions defined on
R" and define Y+ and y_ by

Theorem 1.5.1.7 Assume that s - 1/p is not an integer and that s - 1/p =
k+o, O<o <1, k an integer >0. Then the mapping u -» (ft -o, (f, 3k-0)
defined by
f*=Y=Dhu, 0sjsk
defined on D(R") has a continuous extension as an operator from W: (R)
on the subspace of

i =0 i =0
defined by the conditions
(a) In-1Dh-ift = Vn-ADh-ifi, 185-j- 2
p
and
(6)
+00

- D'
for 1=s-j-2/p, when s-21p is an integer.

The notation is self explanatory: we denote by R-' the subset of Rh-1

defined by Xn- -120 respectively, Yn-1 is the trace operator on the hyper-
plane X-1 = 0 defined in Theorem 1.5.1.1. This statement is a direct
consequence of Theorem 1.5.1.1 through the following lemma.
40 SOBOLEV SPACES

Lemma 1.5.1.8 Let ft € W-(R'") and de ne f by

f(x) =f* (x) when X-120;
then feW. (R"-") if and only if
1
(a) In-1Dh-ift =Vn-1Dh-if , I<r-p'

- D'
for I= r-1/p when r - 1/p is an integer.
The corresponding results for a domain whose boundary is not a
hyperplane will be detailed only in the case of a plane domain in the next
subsection.
All the results that we have mentioned so far about trace properties are
rather qualitative. It is often useful to have also quantitative results for
traces. Here is a very elementary result in that direction. Before stating it,
we need an auxiliary result about Lipschitz boundaries.

Lemma 1.5.1.9 Let M be a bounded open subset of Rn with a Lipschitz

boundary F. Then there exist 8 > 0 and Me C°(1)" such that
u•v>8 ae. on Г. (1,5,1,1)
Inequality (1,5,1,1) means that u is not very different from the normal
w on I. However, u is much smoother than v.

Proof It is very easy to define u locally. In the notation of Definition

1.2.1.1 we can choose tv (in V) as the unit vector in the direction of Yn-
Indeed, the component of w in the direction of yn is [1 + IV (y'/2]- 12 and
consequently we have

where L? is the Lipschitz constant of o in V*

Then we cover the boundary I of © by the interiors of a finite number
of hypercubes V 15k&K, each of which fulfils the conditions of
Definition 1.2.1.1. To each V corresponds a vector Mv, by the construc-
tion described above. Then we can define u as follows:

u = k=
fi
 1.5 TRACES 41

where Ok, 1SkSK, is a partition of unity on I such that Or € D (RM),

Ok 0 and Ok has its support in the interior of Vk
Obviously u is a smooth vector field, and on I we have

M• w= k=

Theorem 1.5.1.10 Let M be a bounded open subset of R" with a Lipschitz

boundary I. Then there exists a constant K such that

(1,5,1,2)

for all u € W'(R) and & € J0, 1[. In addition K depends only on the norm of
u in C'(R) and on 8 (defined in Lemma 1.5.1.9).

Proof In view of Theorem 1.4.2.1 it is sufficient to prove inequality

(1,5,1,2) for all u€ C'(R). For such a function, we have

• дх;
lup-2u• ax,
du
M; dx
i =1

= p lup-zuwu-udx.
On the other hand, applying the Green theorem (see also Theorem 1.1,
Section 3.1 in Necas (1967), or Theorem 1.5.3.1) we obtain

It follows that

and consequently that

slulderspmaxlallup-1/wuldx+max/div.allapdx.
Then, applying Hölder's inequality, we get
42 SOBOLEV SPACES

where p1+ q-1=1 and then

This inequality, clearly implies (1,5,1,2) when € € j0, 1. •

1.5.2 Polygons
The results stated in Theorem 1.5.1.2 and 1.5.1.3 are not sufficient for
studying the Neumann problem in a domain whose boundary is a
polygon. Indeed such a domain is never of class C'. However, those
theorems give us a hint of what happens.
First let us fix some notation. From now on, we consider a bounded
open subset M of R?, whose boundary is a curvilinear polygon of class
Ck.I. We denote each of the Ck. curves which constitute the boundary by
I; for some j ranging from 1 to N. The curve Fit follows I; according to
the positive orientation, on each connected component of F. We denote
by S, the vertex which is the end point of I; Following the same method
as in Section 1.5.1 we define a Ck-11 vector field v; on a neighbourhood
of M, which is the unit outward normal a.e. on I; (We observe that v; = y
ae. on Ty but in general v;+ w inside S.) Finally we denote by w; the
measure of the angle at S; (toward the interior of 1). For a smooth
function u € D (8) we denote by y;u its restriction to I;. (I; is the interior
of I;, i.e., the set I'; without its endpoints S;- and S;.)

Theorem 1.5.2.1 Let N be a bounded open subset of 1?, whose boundary

is a curvilinear polygon of class Ck.; then for each j, the mapping
ди d'u) 1
I<s--p'
which is defined for u € D(R), has a unique continuous extension as an
operator from

W%.(R) onto [I Ww-i-vegr),

i=0
lEm-15k.
Lemma 1.5.2.2 Let R be a bounded open subset of R? whose boundary is
a curvilinear polygon of class C'. Let fi = y;u; then we have

(1,5,2,1)
x-yi
 1.5 TRACES 43

This is just a consequence of the finiteness of the norm of yu in

WI-1/P(F). From (1,3,3,3), splitting the domain of integration IXI in
Vin I; XIk, we get

§i=1 x-y? - do (x) do (y)

- do (x) do (v) <+00.

i*kГ;хГк x - y?

Since we already knew from Theorem 1.5.2.1 that fi€ WI-1/ (Г.),
15j≤N (and consequently f€ Lp (I), 1 <j < N), the condition (1,5,2, 1)
is automatically fulfilled when the distance from I; to I is strictly
positive. In other words, (1,5,2,1) is an extra condition only when I; and
I have a common end point. By possibly exchanging j and k, we can
assume that k =i + 1. Then let o be the distance along I, starting at S;
and let (or) be the point on I whose distance to S; is o. Consequently
for a small enough, loss;, say, we have X;(o) eI; when a <0 and
x;(0) ET;+1 when o > 0. With these notations, condition (1,5,2,1) may be
rewritten as

(1,5,2,2)

since the angle at S; is not allowed to be 0 or 2 (and therefore

X;(0) -X; (-7)| and o + r are equivalent functions). On the other hand,
the fact that f,€ Wy-l/P(I;) and re WD-1P(I;+1) implies the con-
vergence of the following integrals:

T-OP (1,5,2,3)
•8, E, If: + 1 (x; (0)) - f: + 1 (x; (7) )IP
T-oP - do dr <+00. (1,5,2,4)

From these inequalities, we shall deduce the following result, which is

nothing but a rephrasing of Gagliardo's theorem. For simplicity we
assume that I has only one connected component and agree that IN+1 =
r; the extension of the forthcoming results to non-simply connected
domains is obvious and just leads to complications in the notation.

Theorem 1.5.2.3 Let R be a bounded open subset of R? whose boundary

I is a curvilinear polygon of class C'. Then the mapping u -» If 1, where
f, = viu, is a linear continuous mapping from W(S) onto the subspace of
fi
44 SOBOLEV SPACES

IF, WH-W'(Ty) defined by

(a) no extra condition when 1 <p <2,
(b) f; (S;) = fit 1 (S;), 1 5j S N when 2 < p <00,
- do so, I SjEN
(0) L'VL. (4 (0) - 1641 -ODE
when p=2.
We observe that condition (b) is meaningful since, for p >2, it follows
from Sobolev's imbedding theorem that f, and fi+, are continuous on I;
and I;+ respectively. Furthermore, if for some particular u€ H'(R), fi
and fits are Holder continuous near Si, then it is easily seen that
condition (c) reduces to condition (b). Unfortunately, condition (b) is not
always meaningful when p = 2, since functions in HU(F) are not always
continuous (see Remark 1.4.4.9). This is one of the few cases where
Sobolev spaces related to p = 2 are more complicated to handle than
Sobolev spaces related to p# 2.

Proof We know from Theorem 1.5.2.1, that for u € W; (8),

i=1
Furthermore, when p>2, u is continuous up to the boundary I and
consequently we have f; = u r, for all is in particular, we have
u (St) =f;(S;) = fi+1 (S;)

and this shows that condition (b) is necessary. Finally, in the limit case
p =2, condition (c) will follow from (1,5,2,2). Indeed, we have

10+712
(%,If + 1 (X; (07)) -f: (X; (-7) 12
- do de
10+712
*ILI 10-712
and this is finite in view of (1,5,2,2) and of the fact that f; belongs to
H'(I). We conclude by observing that
(б, dr
 1.5 TRACES 45

and consequently we have proved that

This is why condition (b) is necessary.

We now turn to prove that those conditions are sufficient.

First case 1<p <2. We want to prove that the mapping is onto; in
other words, for every Vi, € IN I WH-W/P(I; we must show that there
exists a u € W(M) such that v;u = f;. For that purpose, we take advantage
of Theorem 1.5.1.3 and it is enough to build up a function f on I' from all
the given f, and to check that f€ WI-1/P(T). Thus we set
f(x) =f; (x), x€T;
Since f, is given in W-(Ti), we know that feLp (I) and that

- do (x) do (v) <+00

x-y?
is finite when i = k and when k+ j-1, j, j+ 1 (so that the distance from I;
to Ik is strictly positive). Remembering identity (1,3,3,2), it remains to
check that

x-yP - do (x) do (v) <+00. (1,5,2,5)

Indeed, we have

IS 165(x)82
LOr,

Since Lit, is a C' curve, the function

do (y)
[x-yl
is equivalent to d(x, I +1)-D+1 and therefore
do (x) do (y) 11/p

gP-1
where K and K' are some constants. The last two integrals are finite as is
46 SOBOLEV SPACES

shown in Theorem 1.4.4.3, since

1 -- 1
DD
when p <2. Consequently, (1,5,2,5) is proved and feW;-"'(r).
Second case 2 €p <0. We follow the same method and eventually we
have to check the finiteness of I; This is equivalent to the finiteness of
o, f: (X: (-0)) - fi+ 1(X; (7)) P 71/p
L do dr .
This last integral is less than or equal to

, 88, 18. (7, (-0)) -1, +1 (4, (0)" do dr?

0P-I
(1,5,2,6)
where K is some constant. This is obviously finite when p= 2 as a
consequence of condition (c). Then when p>2 let us denote by h the
function
0+> 4 (o)lf; (x; (-0)) - fi+1 (x; (0))]
where I is some smooth cut-off function, which is identically equal to 1
near zero and which is zero for o>28; We know that h belongs to
W"- VP(10, 8;D and that h(8) = 0 by construction and that h (0) = 0 in view
of condition (b). Consequently, we have h€ WI-1/P(0, S;D) by Corollary
1.5.1.5. Finally, it follows from Theorem 1.4.4.3 that
h
and this shows the finiteness of the integral which appears in (1,5,2,6).
Consequently we have shown the convergence of I' in all cases; this
means that feW-"P(F) and the proof of Theorem 1.5.2.3 is
complete. -
The remainder of this section is devoted to the extension of Theorem
1.5.2.3 to the spaces W(R), when m > 1. Essentially the method of
proof is the same; however, for the sake of clarity we shall consider
 1.5 TRACES 47

successively the cases where & is a quadrant, then a rectilinear polygon

and finally a curvilinear polygon. First let us denote by R+ XR+ the first
quadrant defined by * >0 and y > 0.

Theorem 1.5.2.4 The mapping u -> (kIk=o, (gir-o's defined by

fk =D'uly=o, gi = DiU x=0 (1,5,2,7)
for u € D(R+ XR.), has a unique continuous extension as an operator from
W-(R. XR.) onto the subspace of
m-1 m-1
T= I Wm-k-Ne(R.)× Il wm-1-1/P(R.)
k=0 1=0

defined by
(a) D'fk(0) = D*g(0), It k <m -21p for all p, and
(1,5,2,8)

for p = 2.

(Here we shall denote by y, the trace operator on * =0 and by Y the

trace operator on y = 0; accordingly fk = Y›Du and 81 = Y, D,u.)
The proof makes use of a simple continuation result which is proved by
applying twice Nikolski's continuation method (once with respect to each
variable). This method is explained in Section 3.6, $3, Chapter 2 of Necas
(1967). See also Theorem 1.4.3.1 since R+ XR+ is a uniform Lipschitz
epigraph.)

Lemma 1.5.2.5 We have W(R. XR.) = WE (R. XR+).

Proof of Theorem 1.5.2.4 Let U be any function in W(R) such that u

is the restriction of U to R+ XR+. Applying Theorem 1.5.1.1, we see that
the traces of u must be in T. Then for each k and I such that k+1€
m - 1, we consider
Uk.! = DEDiu, on 10,1 xJ0, 11.
It is obvious that u. belongs to W00, 1[x]0, 1D and consequently
conditions (a) and (b) in Theorem 1.5.2.4 follow from Theorem 1.5.2.3.
Now we are left with the problem of showing that the trace mapping is
onto. We also need a continuation property on R+ for spaces of fractional
order.
Lemma 1.5.2.6 We have W(R) = W$(R.) for all s > 0.
 48 SOBOLEV SPACES

Proof A short proof is the following. We can apply Theorem 1.4.3.1 to

us, the restriction of a given u € W(R+) to JO, 1[. Let U = Pu, and let 4
be a cut-off function such that q(x) =0 for x ≥} and q(x) = 1 for x 53;
then
@U+ (1-ф)й
is the desired continuation of u.
End of the proof of Theorem 1.5.2.4 Let {falk-d and {gir-o' fulfil all the
conditions in Theorem 1.5.2.4. We must find u € WE (R, XR.) such that
(1,5,2,7) holds. We first reduce our problem to the case when 8 = 0 for
all I. For that purpose, let G, be a continuation of g, with G,€
Wm-1-1/P(R). From Theorem 1.5.1.1 we know that there exists U€
W-(R?), such that y,D<U = G 0 1<m - 1, where v, refers here to the
trace operator on the hyperplane * = 0.
Then, instead of looking for u, we shall look for 0 = W- U In. xB, €
W(R. XR.) such that

O<k&m-1
0<1&m -1
where hk =YD$(U It. xe.). From the direct part of Theorem 1.5.2.4
which we have already proved, we know that
he EWm-k-1/P (R.)
and in addition

(c) D<hk (0) = D$g, (0), I+ k <m - - I for all p and

<+00, 1+k =m -1
for p=2. Let us denote by k the difference fk -h; then ok €
Wa-k-1/P(R.) and from (a)-(d) it follows that

qk(0) =0, I<m -k-=for all p (1,5,2,9)

and

-<+00, 1=m-k-1 for p =2. (1,5,2,10)

At this step our problem is the following. We are seeking U€

 1.5 TRACES 49

WH (R-XR.) such that

0<k=m-1
Ty, D'0 = 0, 051&m -1.
For the time being, let us accept the following result.

Lemma 1.5.2.7 Under assumptions (1,5,2,9) and (1,5,2,10), we have

QK E Wm-k-1/0 (R.), OSk Sm - 1.

This means that k € Wm-k-1/P (R), and applying Theorem 1.5.2.1 we

know that there exists we W(R) such that

where Y refers to the trace operator on the hyperplane y = 0. Then we

obtain v as follows:

i=1
where the 1; are real numbers such that

It is obvious that v€ WE(R. XR.). Then we have

(n, D50) (x) =44 (8) -21, 56 (-ix) = Q4(5)

j=1
for 05k <m -1, since * >0. We also have

(73D4 01 68) = 1-51, 4-19 4173 Diw=0

and consequently v is the desired function. The proof of Theorem 1.5.2.4
is complete provided we check Lemma 1.5.2.7.

Proof of Lemma 1.5.2.7 We just need to extend some of the previous

results valid on a finite open interval to the case of R.. We again use a
cut-off function & which is identically equal to 1 for *⅓ and zero for
* >¾. Then (1 - 4)Qk € Wm-k-1/P(R.) and its support is far from zero; it is
readily seen from Definition 1.3. 1.1 and Definition 1.3.2.5 that (1 - 4) 4k €
Wm-k-1/P(R). On the other hand, we can consider or as belonging
to Wm-k-1/9(10, 1D. Applying Corollary 1.5.1.6 together with identity
(1,4,4,12) when p+2 and Corollary 1.4.4.10 when p = 2, we see that
50 SOBOLEV SPACES

(1,5,2,9) and (1,5,2,10) imply that ok € W-k-1/P(J0, 1D and conse-

quently that Joke Wm-k-1/P(R_). The lemma is proved by addition,
writing ok = dok + (1-4) 44. 1
An extension of Theorem 1.5.2.4 to W(R+ X R+) with a noninteger s
can be found in Grisvard (1966). The method of proof followed here is
close to the method used in Nikolski (1956a,b, 1956-58, 1961) for study-
ing the traces of some slightly different spaces.
The previous results are easily extended to an infinite sector with angle
w € JO, L, by means of a linear change of coordinates. We also observe
that the same results hold for the complement of the first quadrant owing
to Lemma 1.5.2.5. Again, a linear change of coordinate allows one to
extend those results to any infinite sector with angle w € )r, 2 [. Eventu-
ally using a partition of unity, we obtain the corresponding results on a
polygon; for simplicity we assume its boundary to be of class Co

Theorem 1.5.2.8 Let S be a bounded open subset of R? whose boundary

F is a curvilinear polygon of class C°. Then the mapping u -> {vidulavis,
1<≤N, 0<1<m -1 is linear continuous from WE (R) onto the subspace
of
-1
T=П П wm-1-1/P(T.)
j=1 1=0
defined by the following condition: Let L be any linear differential operator
with coefficients of class Co and of order d&m -2/p. Denote by Pi, the
differential operator tangential to I; such that L = Lizo Pi, d'lavi; then

(a) E (Pi. Si. (54) = E (Pit. i+1. )(S,) for d <m -5

2
1=0 p
1=0 2O
do <+00
for d=m -1 and p = 2.
Proof Using a partition of unity, we can restrict ourselves to the study of
one vertex. Then a change of variables of class C° replaces the corres-
ponding vertex by zero, the angle by /2 or 3/2 and the sides by the
coordinate axis. Now the only difference between Theorems 1.5.2.4 and
1.5.2.8 is that in the former we only consider the operators D½D, while
in the latter we consider all the operators with coefficients in C°
However, in the case of a right angle with straight sides, this is equivalent.
Indeed, let the fk and the & fulfil condition (a) of Theorem 1.5.2.4 and
 1.5 TRACES 51

let L be of order d <m -2/p. We can write

(1,5,2,11)
k=0 1=0
then
(1.5.2.12)

(1,5,2,13)
Consequently we have

=>
and this is condition (a) in Theorem 1.5.2.8 (with the necessary change of
notation).
Then in the case p = 2, let us assume further that the &k and the fi fulfil
condition (b) of Theorem 1.5.2.4. We want to check condition (b) of
Theorem 1.5.2.8. A preliminary remark is that we also have

- < +00 (1,5,2,14)

for I+ k <m -1. Indeed in that case we have

DISk - DIg, € H31200, 1D.
From Sobolev's imbedding theorem, we know that Difk - D$g, is Hölder
continuous of order a for every a €0, 1L. Since this function also
vanishes at zero by assumption, there exists a constant K such that
t€J0, 1L.
This implies (1,5,2,14). Then using the same identities (1,5,2,11),
(1,5,2,12) and (1,5,2,13), it is easy to check condition (b) in Theorem
1.5.2.8.
Remark 1.5.2.9 In some questions related to the solution of mixed
boundary value problems, we have to admit the value a as possible value
for the measure of the angles of &. In view of Theorem 1.5.1.7, the
conditions (a) and (b) in Theorem 1.5.2.8 have to be replaced by the
52 SOBOLEV SPACES

following, when the measure of the angle at S; is :

(a) fi(S;) = fi+1. (S;) for I<m -=

2

for I=m -1 and p = 2.

Remark 1.5.2.10 In the particular case when 2 is a rectilinear polygon,
it is enough to consider only those operators L which are homogeneous
and with constant coefficients in the corresponding statement of Theorem
1.5.2.8.
Remark 1.5.2.11 As in Corollary 1.5.1.6 we can characterize the kernel

of the mapping d'u) 1SjEN, 0<15m-1

as being W; (R).

1.5.3 Maximal domains of elliptic operators

So far, wehave defined the trace of a function belonging to some Sobolev
space W(R), under the assumption that s is larger than 1/p. However, it
was shown in Lions and Magenes (1960-63) that when a function u is a
solution, in 8, of an elliptic equation, u has traces on the boundary
provided it belongs to any Sobolev space, without any restriction on s and
p. The purpose of the present subsection is just to extend part of this
result to the case of a domain with a polygonal boundary. A different
approach to this kind of result is presented in Goulaouic and Grisvard
(1970).
The method of proof devised by Lions and Magenes uses Green's
formula. First we recall that Green's formula is valid in any bounded
Lipschitz domain, as is shown in Necas (1967) (Theorem 1.1, $1,
Chapter 3).

Theorem 1.5.3.1 Let 1 be a bounded open subset of R" with a Lipschitz

boundary I. Then for every u € W(9) and we W. (R), with 1/p +1/9=1,
we have

(1,5,3,1)
 1.5 TRACES 53

(n' denotes the ith component of the vector field w which was defined in
Section 1.5.1.)

We shall apply this formula twice to derive the following, where A

denotes a second-order elliptic operator with coefficients smooth enough.

Au = I D, (a,, De,u) + ≥ a, Diu + a, u.

i.k = 1 i=1
Precisely, we assume that a;k and a; are Lipschitz continuous and
ay € La (M). The adjoint operator will be denoted by A*, i.e.,

A*O = È DA (A4, DiN) - ≥ D, (a,0) + 4,0.

ik=1 i=1

The corresponding "conormal derivatives are

ди =
AVA i.k = § aikukDiv.
dUa* ik=1

Lemma 1.5.3.2 Let N be a bounded open subset of Rn with a Lipschitz

boundary I. Then for every u€ W. (2) and ve WE(S) with 1/p+1/9=1,
we have

Awodx-uA*odx
du du
AVA dot | (E v'a, rusador. (1,5,3,2)
=1

When & is a plane bounded domain, whose boundary I' is a C1.1

curvilinear polygon, we can restate this lemma. Using the same notation
as in the previous subsection, we define, for each j, a Lipschitz vector field
v; on S, such that v; is the unit outward normal a.e. on I Accordingly,
we define several 'conormal derivatives'
ди dU 2
dUAi=iK=1
2 aikviDiu, dUAti iK=1
For u € W. (8), vE WE(S), we have
ди-E W. (82) and dU
AvAi AVA*.I
since a;k and v, are all Lipschitz functions. Consequently, y; (au/avA.) and
Vi(dU/AVA*i) are well defined and coincide a.e. on I; with yau/da and
yao/Ava* respectively, as defined previously.
54 SOBOLEV SPACES

Lemma 1.5.3.3 Let M be a bounded open subset of R? whose boundary is

a curvilinear polvgon of class C.. Then for every u€ WE(R) and we
W?(R) with 1/p+1/9=1, we have

ди YU do -
г.Vi avai
(1,5,3,3)
where a denotes the vector with components ay and az.

The first consequence of this Green formula concerns the domain of

the maximal extension of the operator A in Lp (R), which we denote by
D(A; Lp (R)). In other words†
DIA; L. (8)) = {u ELp (R); AuEL. (82)).
This is a Banach space for the norm
urs_llulb.p.nt|lAullb.p.r34/p
Furthermore D(R) is dense in D(A; Lp (S)) when I has a C' boundary.
The same proof as in Lions and Magenes (1960-63) works, although they
only deal with Co boundaries. Then these authors show that the mapping
ди
and fra: you
has a continuous extension as an operator from D(A; L. (8)) into
W."'(F)X Wo'-1P(F). Again, their method of proof allows one to handle
domains with a C1. boundary. However, the similar result for a domain
with a polygonal boundary deserves a proof.

Theorem 1.5.3.4 Let O be a bounded open subset of R?, whose boundary

is a curvilinear polygon of class C'. Then the mapping
ди
which is defined for u € W3(8) has a unique continuous extension as an
operator from D(A; Lp (1)) into

† We observe that for u € Lp (S), we have Diue W. '(R) and consequently a;D;u is well
defined and belongs to W-'(S) too, since a is Lipschitz continuous. Therefore, Au is well
defined as an element of W 2(12).
 1.5 TRACES 55

when p+2 and into

H-(112)-+ (R)X H-(3/2)-* (F.)
for all & >0 when p = 2.

Actually when p=2 we shall prove that the trace mapping defined
above, maps D(A, Lz(1)) into the dual space of
AUZ(E)X A32(T;).
It will also be clear from the proof that it is enough to assume that A is
nowhere characteristic on the boundary I' of 1.
Finally, a different approach (as in Goulaouic and Grisvard (1970))
allows one to show that the trace mapping in Theorem 1.5.3.4 is onto.
However, this is useless for the purpose of the next chapters.

Proof For u € W5(9) and v € WE(5), it follows from (1,5,3,3) that

du
i=1 I,
where K is some constant depending on u. In particular, for a fixed j, we
consider those functions v which belong to

'=O on I's for atil.

For U € V we also have Vk(20/Ava*k) = 0 on Ik for k+ j and consequently
ди - yodo - yu dU
Ir,Vi avAi (1,5,3,4)
On the other hand, we know from Theorem 1.5.2.8 that the mapping

maps V onto the subspace of We-la(r)x WI- la(T) defined by the

following conditions, where ; is the unit tangent vector to I; (following
the positive orientation with respect to 1).
(a) fi.o(S;) = fi.o (S; -1) = 0 for all q

(b) fio(S;) =
д- fio(S.-1) =0 and
At; AT;
fin(S;) = fin(S;-1=0 for q>2
56 SOBOLEV SPACES

12O
do <+00,

12O
do <+00,

- <+00

for q=2.
These conditions show that fine W&- 14(I;) and file WA- la(E;) through
Corollary 1.5.1.6, when q7 2; however, when q = 2, these conditions
show that fro€ H3/(Ty) and fire H'(E;) through Corollary 1.4.4.10. In
other words, v r-> {fio; fi.it maps V onto

when a+2 (and consequently p#2) and onto

when q=p = 2.
Now, since A is non-characteristic on I, we have
dU
Vi ava'i= ali avi
where a is strictly positive on I; Conversely, we have
dU
"au, ros.
It follows that

maps V onto
We-Ma(r)x W!-Ha(T.)
when a+ 2, and onto

This result, together with inequality (1,5,3,4), shows that

ди o do - yuldo
 1.5 TRACES 57

is a continuous bilinear form on

when 972, and on

when q = 2. This defines y; du/ava.; as an element of We'-1/P(I;) and y,u

as an element of W- (F) when p + 2; while this defines v; du/dva. as an
element of the dual space of H3/(T) and y;u as an element of the dual
space of A (I;) when p = 2. This proves Theorem 1.5.3.4.

Remark 1.5.3.5 Actually, since A is non characteristic on I, Y; dulav; is

also defined as an element of W11/P(Ty) (respectively the dual of
H32(T;)) when p+ 2 (respectively p = 2).
We shall also need a Green's formula, extending (1,5,3,3) to u€
D(A, Lp (R)). When M is a bounded open subset of in with a Co
boundary I, it is shown that (1,5,3,2) has a natural extension. Indeed,
Lions and Magenes (1960-63) prove that
/ ди -- ,)
avA LAvA*
(1,5,3,5)
for all u€ D(A; Lp (8)) and ve Wa(R). Here the brackets denote the
duality pairing between W.'-I/P(F) and Wi+/ (I) for the first and
between W-UP(F) and WI/(I) for the second. The same result holds with
the same proof, if we only assume that & has a C' boundary.
Unfortunately, the analogue of (1,5,3,5) no longer holds, if we consider
a bounded plane open set & whose boundary I' is a curvilinear polygon
which actually has corners. The reason is that, in general, for u€
DIA, L. (9)) and ve WA(M), the traces y; duldvai and yu are in the
spaces We'll (I) and W +1/ (I) respectively and these spaces are not
in duality. (This is for p+ 2; the situation is even worse for p = 2.)
Consequently, we shall prove only the following statement.

Theorem 1.5.3.6 Let M be a bounded open subset of R? whose boundary

is a curvilinear polygon of class C.. Then we have

Awodx-LuA-odx
ди _ du (1,5,3,6)
• friarai
58 SOBOLEV SPACES

for u € D(A, Lp (1)) and v€ WE(8), 1/p+ 1/9 = 1, such that

(a) u(S,) = 0, j= 1, 2, ..., N when p > 2
(b) U(S;) = 0, and VU(S;) = 0, j = 1, 2, ..., N when p <2
(c) 0 = 0 in a neighbourhood of S;, j= 1, 2, ..., N when p = 2.

Proof We already know from Lemma 1.5.3.2 that (1,5,3,6) holds for
u€ WE(8) and ve Wa(M). We also know that W(8) is dense in
DA; Lp (S)). So we just have to prove that the right-hand side of
(1,5,3,6) is continuous in u for the norm of D(A; Lp (82)) for those
particular v specified in the statement of Theorem 1.5.3.6.
Now, because of Theorem 1.5.3.4, we just have to check that
dU
yiU E W'+"/P(F;),

at least when p+ 2. It follows from Theorem 1.5.2.1 that

YUEW!+I/P(T.) and du- є WI/P (Г.).

Vi ava*i
Then from the extra hypotheses (a) and (b), we have
y;0 (S;) =0(S;) =0, y;0 (S;-1) =0(S;-1) =0
for all p and

do (S.) =0
aU
Viavat;(S; -1) =AVA*.I
du (S:-1)=0
for p <2. By Corollary 1.5.1.6 we therefore know that
dU
and Vi avati
- € WIP (T.),

and this is enough to prove our Theorem for p+ 2.

In the particular case when p =2, viu and Y; do/dua*; have closed
supports inside I Consequently, it follows from Corollary 1.4.4.10 that

Y;0 є H32(Г.) and du

Vi ava's
This shows that the right-hand side terms in (1,5,3,6) depend continuously
on u E D(A, Lz(82)) in view of the first remark, just after the statement of
Theorem 1.5.3.4). _
 1.5 TRACES 59

In dealing with variational solutions of some boundary value problems,

we shall often need similar results concerning a 'half' Green formula.
Indeed, the following is an easy consequence of Theorem 1.5.3.1. (We
restrict ourselves to the Laplace operator for simplicity, since we shall
only need this result in the coming sections.)

Lemma 1.5.3.7 let 2 be a bounded open subset of R" with a Lipschitz

boundary I. Then for every u eH?(9) and o€ H'(R), we have

(1,5,3,7)
The corresponding statement on polygons is (according to the notation
introduced previously in this subsection) the following:

Lemma 1.5.3.8 Let N be a bounded open subset of B? whose boundary is

a curvilinear polygon of class C'.". Then for every u€ H? (1) and NE
H'(2), we have

- viu do. (1,5,3,8)

Again this can be extended to more functions u. Let us set

ElA; Ip (8)) = {u E H'(R); Au eLp (2)).
This is a Banach space for the obvious norm
ur>llull.z.at|Aullo.p.o.
As before, D(8) is dense in E(A; L, (2)) when S has a Lipschitz
boundary, but this now requires a proof.

Lemma 1.5.3.9 Let M be a bounded open subset of R" with a Lipschitz

boundary; then D(D) is dense in ElA; Lp (M)).

Proof Let P be any continuation operator defined on H'(8). In other

words, P is a continuous linear mapping from H'(1) to H'(R") such that
Pu Is = u,
for every u € H' (8) (see Theorem 1.4.3.1). With the help of P we can
view H'(8) as a closed subspace of H'(R"). Thus for every continuous
linear form I on E(4, Lp (8)) there exists f€ H-'(R") and g € La (R) such
that
(a) =4;Pu)+| sAuds
60 SOBOLEV SPACES

for all u € E(A. Lp (M)). In addition, since I depends only on u and not on
Pu cn, the support of f is contained in S. (See also Theorem 2.3 in
Magenes and Stampacchia (1958), Chapter 1.)
Now, in order to prove the claim of Lemma 1.5.3.9, we just need to
show that any I which vanishes on D(R) is actually zero everywhere.
Indeed for U€ D(R"), we have
(f; U) +18, AU7 = 0,
since we have
(f; U) =(f; Pu)
where u = Ul, due to the properties of the support of f. It follows that
Ag =-f
in the sense of distributions.
The ellipticity of A implies that 8€ H'(R") and consequently that
g€ H' (8). Let us now consider a sequence 8m m = 1, 2, ... of functions
belonging to D(2) and such that
Em -> 8, m -> +00
in H'(R). For every u € E(A, Lp (R)), we have

Ku) = lim S- (Aim Pus+ ( SmAuds?

= 0.
Thus I is identically zero.
The inclusion
ElA; Ip (2)) S DMA; Lp (82))
shows that y; du/dv; is well defined and
du
Vi avi
when p+ 2 and
ди
є Й312(Г.)*
Viavi
when p=2. However, this result can be improved.
 1.5 TRACES 61

Theorem 1.5.3.10 Let M be a bounded open subset of R3, whose bound-

ary is a curvilinear polygon of class C'. Then the mapping
ди

which is defined on D(D), has a unique continuous extension as an

operator from E(A; Lp (8)) into
Al2(.)*.
Proof Consider v € V, where
V=fuEH' (8) YkU=0, k+i}.
Then y;0 € A (I) (see Subsection 1.5.2). Furthermore, for u € D (R), we
have by Lemma 1.5.3.8

r.
It follows that

Vi avi
and consequently there exists a constant C such that

The result follows by density. -

We can now extend the Green formula (1,5,3,7) to u € E(A; L, (8)).
Theorem 1.5.3.11 Let D be a bounded open subset of R?, whose bound-
ary is a curvilinear polygon of class C'". Then we have

(1,5,3,9)
Miao," : 4,0)
for u € E(A; L. (2)) and vE W; (2), r> 2 such that
U(Si) =0,
Proof The identity (1,5,3,9) holds by Lemma 1.5.3.8 for every u € D (R)
and v€ W; (1), r> 2 (since this last space is included in H'(2)). Then
assuming that v is zero at the corners, implies that
62 SOBOLEV SPACES

Consequently, all the terms involved in identity (1,5,3,9) are continuous

in u for the norm of ElA; L (2)). Again, the result follows for
density. •
Finally, let us recall for later reference, the corresponding result on
domains with smooth boundary, due to Lions and Magenes (1960-63).
Here O is a bounded open subset of R" with a C'.' boundary. Then D(S)
is dense in E(A; Lp (8)), the trace operator
ди

is linear continuous from E(8; Lp (8)) to H 12(I) and the Green for-
mula
(1,5,3,10)
holds for every u€ E(N; L (2)) and v€ H' (1).

1.6 Boundary conditions

So far, we have studied the traces on the boundary of a function U,

together with its derivatives in the direction of w up to a certain order.
For the purpose of studying boundary value problems, it is convenient to
replace the powers of d/Av by a more general set of differential operators.
This is the main goal of this section.

1.6.1 Normal systems

From now on, we consider a set of given differential operators
Bk. (x, D.) = 2 a. (x)D*
la|€dk
k=1,.. .. K
with C° coefficients defined in M. For convenience, we assume that these
operators are numbered according to the increasing orders of their
degrees; in other words, we assume that k-> de is a nondecreasing
function of k.
Furthermore, we make the very restrictive assumption that the system
{Bk}k= is 'normal. This means the following

Definition 1.6.1.1 Let 1 be an open subset of RM with a Lipschitz

boundary I. The system {Br}k=, is said to be normal on a subset I' of I if
(a) the degrees de are all different
 1.6 BOUNDARY CONDITIONS 63

(b) the Bk are all uniformly noncharacteristic on F: i.e. there exists m and
M such that 0<m &M and

m<
I/a| =dk
ae. on I'. (As usual v° means (va, ... (wn)%..)
This definition agrees with the usual one given for a bounded 2 with
C° boundary. We observe that k-> de is now a strictly increasing
function.
We shall now investigate the mapping

when u varies in some Sobolev space. We first quote the classical results
of Lions and Magenes (1960-63).

Theorem 1.6.1.2 Let {B.K, be a system of homogeneous differential

operators with constant coefficients in R", which is normal on the hyper-
plane *n =0. Then for s - 1/p non-integer and >dk, the mapping

from W(R") into Ilk= W-d-1/P (R"-') is onto.

This is a consequence of Theorem 1.5.1.1. The following is a conse-

quence of Theorem 1.5.1.2.

Theorem 1.6.1.3 Let 1 be a bounded open subset of Rn with a boundary

of class C'. Let also {BIK=, be a system of differential operators in 1
with coefficients belonging to C°(R), which is normal on the boundary I of
M. Then for s - 1/p non-integer, s-1/p > dk and s l+ 1, the mapping

from W; (R) into IlK=, Ws-d-1/P(F) is onto.

We now restrict our purpose to plane domains whose boundaries are

curvilinear polygons of class C°, for simplicity. We also assume that s = m
is an integer. We use the same notation as in Section 1.5.2. With each of
the curves I; we consider a set of differential operators
B;k (x, Dx) = & a;a (x) Di k = 1,2,...,K;
lated,.k
We assume that the coefficients a;. belong to C°(8) and that the set
{Bik)k-1 is normal on I; for each j= 1, 2,...,N. It follows from
64 SOBOLEV SPACES

Theorem 1.5.2.1 that for u € WE (R) we have

fik =y,B;kUEWm-d, k-1/P(E;).
Then, let us consider for each j, all the possible sets of differential
operatorsPik (x, Dx),k = 1, 2,..., K; and Qi+l.k(x, Dx), k =
1,2, ..., Kit such that Pik is tangential to I; for all k and Oil.k is
tangential to lit for all k. In addition, we assume that
K K
(1,6,1,1)
k=1 k=1
at S; and that the degree of Pi. is d-dik and the degree of Qi+1.k is
§d -d;+ 1k. Consequently, the degree of the operator L in (1,6, 1,1) is <d
and for u€ WE (R) we have
Lue Wm-d(R).
Then, Theorem 1.5.2.8 shows that
K K
(1,6, 1,2)
k =1 k =1
when d<m-2/p, while
<+0 (1,6,1,3)
k=1 k=1
when d= m -1 and p = 2.
Conditions (1,6, 1,2) are obviously necessary conditions on the traces
fik. It turns out that they are also sufficient conditions.

Theorem 1.6.1.4 Let R be a bounded open subset of I?, whose boundary

is a curvilinear polygon of class C*. Let also {Bikk=, be, for each j, a
system of differential operators in M, with coefficients belonging to C°(S),
which is normal on I; Then for p+2, the mapping
UH>{fik = y;B;ku},j=1,...,N, k=1, ...,K;
maps W(R) onto the subspace of
N K
T=П T wm-d,.-1/P(F.)
¡=1 k = 1
defined by the conditions (1,6,1,2) for all possible systems of differential
operators {Pik}k, tangential to I; and {Qj+1.kK-1 tangential to If+1, such
that identity (1,6,1,1) holds.

Proof The conditions (1,6,1,2) have a local character. This allows one,
 1.6 BOUNDARY CONDITIONS 65

through partition of the unity and changes of variables, to reduce the

proof of the sufficiency of (1,6,1,2) to the case when 1 is replaced by
R. X+ and the functions fk have bounded supports. (Indeed, when the
angle at S, is less than a the problem is thus reduced to the case when 2
is replaced by R, X R+. On the other hand, when the angle at S; is more
than , & is replaced by the complement of R. XR+. However, owing to
the continuation theorem, it is equivalent to prove sufficiency in R, XR...)
Since we consider a domain with only one corner, it will be convenient,
throughout the proof, to adopt some slightly different notation. We
replace Pik by Pk (setting K; = K'), Qi+1,k by Ok (setting Kit1 = K"), B,.k
by Bk (setting dik = mk) and finally Bi+1,k by Ck (setting d;+1.k = mik).
Thus, we start from

1<k<K' and 1<1-K", such that

(1,6,1,4)
k=1
for all possible systems of operators Pk and Qi such that

£ P, (D, )B, (D. D,) = £ O,(D, G, ID,, D,) (1,6, 1,5)

k= 1=1

and this sum is an operator of degree d <m - 2/p. With this data, we look
for a function u € W. (R+ XR+) such that
YIBRU =Ok and (1,6,1,6)
Instead of building directly a function u, we shall only look for those
functions fk and & which, through the application of Theorem 1.5.2.4,
allow us to find a function u € W. (R+ X R.), such that
and Y2DKu =fk
for k, l = 1, 2, ..., m - 1. In other words, we have to solve the following
problem: Find
frEWM-k-1/P (R), gIE WM-1-1/9 (R), k, 1=0,..., m -1
such that D'f. (0) = Dkg,(0), I+ k <m -2/p and that
I Be. (D,) gI = OK ISKSK'
1=0 (1,6,1,7)

k=0
66 SOBOLEV SPACES

where the operators Bk. and Ck are defined by

B. (D., D,) = & Bk. (D,) DI

G(D., D,) = & G. (D,) DK.

k=0
Since we have assumed that the systems of operators {Br}k, and (CK"
are normal on x =0 and y = 0 respectively, Bk.mi and Cum are just
nonvanishing functions. Consequently (1,6,1,7) may be rewritten as

11 1=0
(1,6, 1,8)
homiesCim'
In other words, we know 8m and fm, as soon as we know g, for IS mk-1
and fk for k <m'" - 1. However, these identities do not define all the &
and the fk since we never assumed that mk = k-1 and that m"= 1-1.
As a first step in the construction of the missing g, and fk, we first look
for the numbers

ak.! = D\fk (0) = D$g(0), k+|<m -.2

p
These numbers are solution of the linear system of equations which we
obtain by differentiating the first equation in (1,6,1,7) with respect to y
and the second with respect to x and then writing the corresponding equa-
tions at 0 when this makes sense. Namely, we have

1<k<K'. 1≥0
1=0
(1,6,1,9)
1STSK", M≥0.
k=(

We now adopt the following notation:

mi, +1-1
DiB. (D,) ='
a =0
m"+u-k
3=0
 1.6 BOUNDARY CONDITIONS 67

where b^ ki and cit

)E are functions. Then at 0, (1,6,1,9) implies

mi+1<m 2
mi mi+1-1

=0a =0 p
m." m"+u-k
2 (1,6,1,10)
k=0 3=( m'' t u <m - p
This is the system of linear equations in ax. This system is possibly
overdetermined. We must therefore check that it has a solution.
The simplest way to prove that (1,6,1,10) has actually a solution is to
prove that the data are in the kernel of the transposed matrix. In other
words, the data have to annihilate all the linear forms which are zero after
composition with the matrix of the system. We must therefore check that
EPA. 141" (0) = 29444449 (0) (1,6,1,11)
k.1 lau

for all possible numbers Pr.. 1 skSK', 1=0,1, ..., r-mk and 91,us
1<1<K", He = 0, 1,..., r-m" (r is the integral part of m-2/p), such
that
(1,6,1,12)
k.1 L,M

for all possible values of a = K and $ = I. For that purpose, let us consider
the operators
r-
Pa (D,) = L Pr.A (X, y) DIs Qi (Dx) =

where the numbers Pk. = Pk. (0, 0) and 91,4 = 91,. (0, 0) fulfil (1,6, 1, 12).
We then have

• mi +1-1

k=1 1=0 1=0 O=1

As a consequence of (1,6,1,12), this operator is also

m"+u-k
=0 k=0 3 =

§ Q. (D.)G (Dy, Dy).

1=
68 SOBOLEV SPACES

This is exactly identity (1,6,1,5). It follows that (1,6,1,4) holds and this
can be rewritten as
{Ph. 145" (0) =241444(4) (0).
k.u Lace

This proves identity (1,6,1,11). Summing up, we have shown that the
linear system (1,6,1,10) has a solution.
From now on, let us consider any solution of (1,6, 1,10). We must find
(1,6,1,13)
such that
(1,6, 1,14)

gik (0) =ak.h 0<k<m-1-

2 (1,6, 1, 15)

and that (1,6,1,8) holds. We obtain the functions fk such that k+ m"
(1 = 1, 2, ..., K") and the functions & such that 17 mk (k = 1, 2, ..., K')
from the one-dimensional version of Theorem 1.5.1.1. Indeed this
theorem implies that the mapping
hrach (0), ...,h (0)}
from W(R) into Pk+', is onto, when k<s-1/p. We then obtain the
functions fm; (1 = 1,2, ..., K") and 8m (k = 1, 2, ..., R') from (1,6, 1,8).
The functions which have been thus constructed satisfy (1,6,1,13) and
(1,6,1,8). They also satisfy (1.6.1.14) for k+ m' (l= 1, 2,..., K) and
(1,6, 1,15) for 17 mk (k = 1, 2, ..., R'). The last step of the proof consists
in checking (1,6,1,14) and (1,6,1,15) for the remaining k and I. We do
this by induction on m' and m½ separately.
Let us assume that we already know that
fk (0) = ak.u
for k<m'-1 (0<M <m-k-21p) and for k = m', but only for 0 Sus
£ - 1 (possibly with £ = 0; this means no information about fm). We shall
then show that
fla! (0) = Am".a
Indeed we have (1,6,1,8) which is equivalent to (1,6,1,7). Thus
m"'+£-k
k=0 k=0 3=0
m"-1 m' +û -k

3=0 k = () B=
 1.6 BOUNDARY CONDITIONS 69

At zero this implies

m"-I m"+@-k
cil (0) ak.B = 4' (0).
B =( k = ( 3 =(

This is one of the equations of the system (1,6,1,10) with am". n" a replaced
by fla).. Fortunately c§m (0) is not zero because c†m is just the coefficient
of D'" in G(Dy D.) and the axis {y = 0} is not characteristic for this
operator, by assumption. This shows that

Consequently we check (1,6,1,14) by induction. We do the same for

(1,6,1,15). The proof of Theorem 1.6.1.4 is now complete.

In the case when p=2, which we have excluded so far, conditions

(1,6,1,2) and (1,6,1,3) turn out to be also sufficient conditions on the
traces fi.k. However, this is really a result which is more easily proved by
using interpolation methods. For this reason we shall only prove the
sufficiency of (1,6,1,2) and (1,6,1,3) in some particular cases that we need
in the forthcoming chapters.

Theorem 1.6.1.5 Let S be a bounded open subset of R?, whose boundary

is a polygon. Let also {Bikk=, be for each i, a system of homogeneous
linear differential operators, with constant coefficients, which is normal on
I; Then the mapping
uh>{fik = y;B;kus, j=1,-..,N, k=1,2, ..., K;
maps HM (S) onto the subspace of

j=1 k=1
defined by the conditions (1,6,1,2) and (1,6,1,3) for all possible systems of
homogeneous differential operators with constant coefficients {Pink-,
tangential to I; and IQ;+1.kk., such that (1,6,1, 1) holds.

Proof The beginning of the proof is quite similar to that of Theorem

1.6. 1.4. Using a partition of unity and affine changes of coordinates, we
reduce the proof to the case when & is replaced by R. XR.. After this
reduction, we still deal with homogeneous operators with constant coeffi-
cients.
Then we adopt the same simplified notation as in the previous proof.
Thus we are given
MEHm-m"-112(R.)
70 SOBOLEV SPACES

1sksK', 1<l&K" such that

(1,6,1,16)
k =1 1=1

when d <m -1, while

<00 (1,6,1,17)
[=1
when d = m - 1, for all possible systems of operators Pk and O such that

(1,6, 1, 18)

and this sum is an operator of degree d&m - 1. We look for a function

UEH" (R+ XR.) such that
and Y2Gu =44. (1,6,1,19)
Equivalently we look for functions fk and & such that
faEHM-k-112(R.), gIEHm-1-112(.)
for k, 1=1, 2, ..., m - 1, with
D'fk (0) = D$g‹(0), l+k<m-1
<700, 1+k=m -1
and such that

ISKEK'
(1,6,1,20)
1S/SK"
(k =

where the numbers bk and Ck are defined by

G(D. D,) = I GaDmi-KDK.

k=(
The first step will be to find the numbers
ax.! = D\fk (0) = D$8(0), k+1<m-1
 1.6 BOUNDARY CONDITIONS 71

together with those functions ak. € H'2 (R.) such that

< +00
k+l=m -1.
- <+00

The necessary conditions on the ax, are obtained by differentiating

(1,6, 1,20) and then considering the behaviour of +§ and «(w) near zero.
Namely, we have

mk+1<m-1
(1,6,1,21)
0
and in addition

2b. am-1-11 (6)-449-1-mi (0) d'2400, k=1,.,K'

12 dt ,
1=1,..., K".
(1,6,1,22)
Now the interest of the assumption that all the involved operators are
homogeneous is that the systems (1,6,1,21) and (1,6,1,22) are not
coupled. In other words, the unknowns in (1,6,1,21) are only the ax.‹ with
k+<m - 1, while the unknowns in (1,6,1,22) are only the ak. with
k+1=m - 1. This allows one to solve the two systems separately. The
system (1,6,1,21) is the same as (1,6,1,10) and it has a solution since we
assumed (1,6,1,16) which is identical with (1,6,1,4).
We are left with the problem of solving (1,6,1,22). We can consider the
set of functions {ak.ilk+1=m -1 as a vector valued function a in Rm and
consequently (1,6,1,22) can be rewritten as

< +00 (1,6,1,23)

where A is some matrix from Rm into R and b is some given function of

class H'(.) with values in B~. Here N= K'+ K" and b is the function
whose components are the corresponding om-1-mi and Jim-1-m". We
shall use the following auxiliary result

Lemma 1.6.1.6 Let be H'(R.; RM); then there exists a€ HI (R.; Rm)
which is a solution of (1,6,1,23) if and only if So ¢ (b(t))? dilt < +00 for all
linear forms o on BY such that so A = 0.
 72 SOBOLEV SPACES

Applying this to (1,6,1,22), we find a solution if and only if

Ik = 1
for all numbers Pk and a, such that
12 dt
t < +00 (1,6,1,24)

(1,6, 1,25)
k=
for all & and I such that & + = m - 1. Let us now introduce the operators
Pk (Dy)=PhDin-1-mi, Qi (Dx) = qDn-1-mi;
then we have

k=1 i =(

k= 1 1=0
As a consequence of (1,6,1,25), this operator is also
m"

1=1k =0 k=
= 2 Q, (D,) G(D. D,).
1=1
This is exactly identity (1,6,1,18), from which (1,6, 1, 17) follows. This last
inequality is exactly inequality (1,6,1,24), which we wanted to check.
Summing up, we have proved the existence of the numbers ax and the
functions ak. which are solutions of (1,6,1,21) and (1,6,1,22).
Now we are left with the problem of finding the functions fk and gi. We
recall that (1,6,1,20) implies identities similar to (1,6,1,8), namely
m! - 1
8m=
Ok.mi
(1,6, 1,26)
fm" =Cim 1515K"
k =?
We obtain the functions fk such that k+ m' for all I and the functions gi
such that I+ mk for all k by applying the following lemma.

Lemma 1.6.1.7 Let am..., am-zER be given together with am-1€

H'(R.); then there exists fe Hm-12(R.) such that
(fi (0) =ax. Osk&m-2
7 +00.
 1.6 BOUNDARY CONDITIONS 73

The remainder of the proof is similar to that of Theorem 1.6.1.4.

Indeed, so far, we have built fk € Hm-k-1/2 (R) and gi E Hm -1-1/2 (R) such
that (1,6,1,20) holds and such that
(fk (0) = ak1. 0515m-k-1
(1,6,1,27)

for k+ m' (1 = 1, 2,..., K"), and such that

(8(' (0) =ak1 05k5m-1-1
(1,6,1,28)
<+06
(Jo Ig'm-1-» (8) - am-1-1,608 dr.
for If mk (k = 1,.., K'). We still have to check the similar property for
the remaining indexes k=m and I= mi. Proving that fi'(0) = 0k., =
g(k)(0) for k+ I<m -1 is exactly the last step of the proof of Theorem
1.6.1.4. Consequently, let us only check that

7+00
for one I= provided we already know that the property holds for
1<1-1. Indeed we have
f(m-m"-1)
m" = 1( m"-1

C1m" k =0
Near zero this implies that

because of our induction hypothesis. On the other hand, it follows from

(1,6,1,22) that
2 dt
<+00

and since C.m; +0, this shows that

12 dt
It - < +00.

This is the desired result. Consequently (1,6, 1,27) is proved by induction

for k = m' for all 1; (1,6,1,28) is proved the same way.
The existence of u solving our trace problem follows from that of fu
and g through the application of Theorem 1.5.2.4. This completes the
proof of Theorem 1.6.1.5.
74 SOBOLEV SPACES

Proof of Lemma 1.6.1.6 We just write that

a = Mb,
where M is any matrix such that AM = P and P is a projection operator
on the image of A. Fredholm's alternative theorem implies that the image
of A is the orthogonal of all linear forms & such that do A = 'As = 0.

Proof of Lemma 1.6.1.7 We set

2 tk * (+-8) m-2
f(t) =
k =0
where & is any smooth cut-off function which is zero for t7⅔ and
identically equal to 1 for +<}. I
Remark 1.6.1.8 Assume in Theorem 1.6.1.4 that M has a (strictly)
polygonal boundary and the Bi.k are homogeneous and have constant
coefficients (as in Theorem 1.6.1.5); then the necessary and sufficient
conditions (1,6,1,2) in Theorem 1.6.1.4 involve only operators Pi.k and
Qi+1.k homogeneous and with constant coefficients. This is easily checked
by inspecting the proof of Theorem 1.6.1.4.

1.7 A model domain with a cut

Domains with cuts sometimes occur in practice (in fracture mechanics for
instance). We shall not undertake here a comprehensive study of the
properties of Sobolev spaces on such domains. We shall only illustrate, on
the simplest possible example, the basic trick which reduces most of the
proofs for domains with cuts to the more classical proofs of the previous
sections. This relies on the trace theorems.
Our model domain is
1=<(x, y)|x?+ y?<1, x20 when y = 0}.
In other words & is obtained by removing the right half x-axis from the
unit disc. Such a domain does not fulfil the assumptions of any of the
definitions in Section 1.2.
The space W(M) has been defined in Section 1.3, but no property has
been obtained in Sections 1.4-1.6 for this space. However, the trace
theorems in Section 1.5.2 imply many properties. The trick consists in
splitting & into two pieces: let us denote by & the domains

Then 2 are plane open domains whose boundaries are curvilinear

 1.7 A MODEL DOMAIN WITH A CUT 75

Figure 1.7

polygons of class C°. For u € WE (R) we set

U_=u.n.
It is obvious that u. € W() and consequently u have well-defined
traces, up to the order m -1, on {y =0}. To make things more precise we
denote by Y+ the trace operators from W(84) onto W "PO-1, +1D
which are defined by
(Y+ U) (x) = U(x, 0)
for u € D(R) respectively (see Theorem 1.5.2.1). We can reconstruct u
from u and u_ by the following result.

Theorem 1.7.1 Let u belong to Lp (S) and denote by u, its restrictions to

S= respectively; then u€ W. (12) iff us E WE (R=) and
(V. D$U.) (x) = (y-D$U_) (x), Osksm-1, (1,7,1)
for almost every x € ]-1, 0L.

Proof We prove the necessity of (1,7,1) by approximating u by u, €

D(R), v=1,2, ... in the norm of W(R), where

Since & has a continuous boundary we can apply Theorem 1.4.2.1. Now,
on ]-1, OL, we have
(V. DIM,) (x) = (Y-DIU,) (x) = (D$U,) (x, 0).
By continuity the first identity is extended to u.
76 SOBOLEV SPACES

We prove the sufficiency as follows. We can approximate u+ by

u% ED(D.) and u_ by u"E D(R_), v = 1,2, ... in the norm of WHO,
and W: (S_) respectively. We define a distribution u" by setting

for all test functions & € D(R). It is a classical result that

J-1
-(DAY-D'I-'u") (x, 0)]D, Q (x, U) dx.
By continuity, we obtain

¡=0
- (DEv D'-*-'u.) (x, 0)]D, 4 (x, 0) dx

provided k + I&m - 1. This, together with (1,7,1), proves that

(DID;W:4) = DIDiu.qdxdy+| DIDiu a dx dy (1,7,2)

for k+I&m-1. Consequently D<Du is a function and belongs to

L, (R). This completes the proof.
We shall now draw some conclusions from Theorem 1.7.1. First, in
general, there is no reason why Y+u+ and y-u_ should coincide on 10, 1[.
This implies that C°(R) is not dense in WE (9). Indeed 2 is the closed
unit disc. For all u€ C° (S) we have Y+u+ = y-u_ on ]-1, +1[; by
continuity this identity is also valid for all u belonging to the closure of
C° (R) in W' (R). It is also obvious that the norms of Wm (1) and of
W. (R), where M, is the open unit disc, coincide on W. (,). Conse-
quently the closure of C°(R) in W(S) is the space of the restrictions to
of all the functions in Wm(R). Since W(R) = W: (R), this shows
that WM (2) ‡ Wm (R); in other words Wm (1) has no extension property
similar to that in Theorem 1.4.3.1.
In order to obtain some convenient density and imbedding results we
must introduce some other spaces.

Definition 1.7.2 We denote by Ck.a (1) the space of all functions u

 1.7 A MODEL DOMAIN WITH A CUT 77

defined in S, which are uniformly continuous together with their derivatives

up to the order k and such that their derivatives of order k satisfy a uniform
Hölder condition with exponent a, in 2.

It is easily seen that u € Ck.a (sp) iff

and
D'u+ (x, 0) = D'u_(x, 0), -1<x<0
for all I such that 0 51 5k.
An easy consequence of Theorems 1.7.1 and 1.4.5.2 is that
(1,7,3)
provided s-2/p>t -2/g, t Ss and
(1,7,4)
provided k + a &s-2/p and s-2/p is not an integer.
The main consequence of Theorem 1.7.1 is a trace theorem for the
space We (2). We shall state it and derive it carefully since it is funda-
mental for studying boundary value problems in a domain like 22.
For this purpose, besides the trace operators Y+ already defined, we
introduce the trace operator Y on the unit circle. We consider the
subdomains & defined by
& ={(r cos o; r sin 4) 10<+ <1, € <0<21 - 8}
for & > 0. Clearly 1 = 8. = U.so h, and S, has a Lipschitz boundary. We
denote by C the interior of the intersection of the unit circle with I the
boundary of ?.. The function
yulc.
is well defined by Theorem 1.5.2.1 applied to .. In addition it belongs to
WH-W/P(C) when u belongs to W(M). We define a function You a.e. on
Co by setting

for every &>0. (This definition may seem artificial, but it saves the proof
of a density theorem which is not easy.)
From now on we set
d'u 1=0, 1, ..., m -1 (1,7,5)
YEar = 81
78 SOBOLEV SPACES

and
1=0, 1, ..., m -1
for u € WE (8). We consider si as a function of A € ]0, 2mL, and fi. as a
function of x € J0, 11.

Theorem 1.7.3 The mapping

is linear continuous from WE (8) onto the subspace of

m-1
П wm-1-1p00,2mDx Wm-1-1/900, 1Dx Wm-1-1/800, 1D

defined by the following conditions

(a) 81 (0) =14. « (1), 84 (277) =74 (1), #4% (0) =74 (0)
for k+1<m -2/p
(6) {'18 40-89201-08 4.= <+00,

1 1 81 1 277-01-1'441-010 d'-+06,
: <+00

for k+1=m -1, when p = 2.

Proof The necessity of the compatibility conditions between the & and
the f7k follows from Theorem 1.5.2.8 on 4. In the same way Theorem
1.5.2.8 on M_ implies the compatibility conditions between the & and
the f-.k.
On the other hand the functions D' have traces Y-D'u, which
belong to
Wm-1-1/P0-1, +1D
and which coincide for * € J-1, 0L by Definition 1.7.1. Consequently, we

have (+1 -f-1)E WM-1-1/9010, 1D.

The compatibility conditions between f+, and f-, follow from Theorem
1.5.1.5 for k + 1 <m - 2/p and from Lemma 1.3.2.6 for k + 1 = m - 1 (and
p=2).
In order to prove that the above conditions are sufficient, we start from
given functions & and ft.k fulfilling those conditions. Then instead of
 1.7 A MODEL DOMAIN WITH A CUT 79

looking for u € W(M), having such traces, we first look for functions Fo
k =0, 1,..., m-1, the traces of u on the segment {(x, y); -1<x <0}.
We claim that there exists
FLEWm-k-1/PC-1, OD
such that
Fl» (0) = f4(0) (1,7,6)
for k+1<m -21p,
< +00 (1,7,7)
for k+1=m-1 (p=2) and
F'(-1) =(-1)k+'g(k) (tr) (1,7,8)
for k+1<m -21p

t
<+00 (1,7,9)
#lEg+-1+0-1-13443496437-012d4
for k+1=m -1 (p = 2).
The construction of such functions Fk is easy when p+ 2, while their
existence follows from Lemma 1.6.1.7 when p = 2. Then we set
Fk(X) for -1<x<0
F7.k(x) =
for 0<x <1.
Clearly F+k€ Wm-k-1/P(J-1, + 1D and applying Theorem 1.5.2.8 to 12+
and &. we check that there exists

such that

and that
0515m-1
arl
on J0, [ and ]m, 2 [ respectively.

Since we have obviously on ]-1, O

for 0 sk &m -1, it follows that the function on & built up from u+ and
u_ belongs to W(8). In addition it has the required traces.
80 SOBOLEV SPACES

Remark 1.7.4 It is easy to combine the results of Theorems 1.7.3 and

1.5.2.8 to obtain the description of the traces for more general domains
with cuts. It turns out that the statement of Theorem 1.5.2.8 remains
valid if we admit domains with cuts provided we consider both sides of
the cut as two different sides of &. In the same way the Theorems 1.6.1.4
and 1.6.1.5 remain valid for domains with cuts. Also Theorem 1.4.5.3
holds for such domains.
2
Regular second-order elliptic
boundary value problems

2.1 Foreword
In the following chapters, we shall carry out the study of some elliptic
boundary value problems in domains whose boundaries are not smooth:
for example, domains with polygonal boundaries. Throughout this study,
we shall make an extensive use of results concerning the same kind of
boundary value problems in domains with regular boundaries. (We shall
call these problems 'regular.) The theory of such boundary value prob-
lems can be found in Hörmander (1963) and Lions and Magenes (1960-
63), for instance. These authors consider problems of arbitrary order in
domains with a C* boundary. Less general boundary value problems are
solved in domains with less smooth boundaries by Agmon (1965),
Miranda (1970), Necas (1967).
In spite of the great number of possible references on elliptic boundary
value problems, we shall devote this chapter to a self-contained study of
second-order strongly elliptic boundary value problems in regular do-
mains. Apart from the objective of making this book as self-contained
as possible, the purpose of this chapter is two-fold.
Three kinds of methods, at least, have proved to be quite successful in
solving regular elliptic boundary value problems. Namely,
(a) A priori estimates as in Agmon (1965), Lions and Magenes (1960-
63), Miranda (1970) and Necas (1967);
(b) parametrices as in Hörmander (1963);
(c) pseudo-differential operators as in Seeley (1966).
These methods have long been known to allow one to solve elliptic
boundary value problems involving operators with coefficients only a few
times dificrentiable, in domains with boundaries also only a few times
differentiable. However. most of the available references deal only with
the C° case. It is within the scope of this book to try to see to what
extent the assumptions on the coefficients and on the boundary can be
81
82 REGULAR SECOND-ORDER PROBLEMS

weakened when applying those methods. Actually we shall restrict our-

selves to the a priori estimates method, which seems to be more flexible in
this respect. It turns out that the most general domains that one is able to
handle with such methods, have a boundary of class C'". This assumption
clearly excludes polygonal boundaries.
The second purpose of this chapter is to give a brief account of the Lp
theory. The Lp theory of linear elliptic boundary value problems is of the
utmost importance in the study of nonlinear problems. The reason is that,
for a given m and a given domain O, the Sobolev space W(2) is more
likely to be an algebra when p is large. The core of the Lp theory is the
celebrated Lp a priori estimate proved by Agmon et al. (1959). These
authors deal with problems of great generality. Their proofs can hardly be
found, even in simpler particular cases, outside this original reference (but
see Freeman and Schechter (1974)). We give here a simplified proof of
the L estimate in the case of second-order strongly elliptic boundary
value problems. This proof is closer to the L proof, since it uses the
partial Fourier transform, with the Plancherel theorem being replaced by
the famous Lp multiplier theorem of Mih'lin (1956) (see also Hörmander
(1960)). The proof makes use of a technical idea introduced for a
different purpose by Boutet de Monel (1971). The related existence and
uniqueness results will be worked out in domains whose boundary is only
of class C1,This does not seem to be standard material and will be
useful in the next chapters. (Here we attempt to work with the weakest
assumptions on the domain but not on the coefficients of the operators.
Indeed, in most practical cases one deals with simple operators-such as
operators with constant coefficients--in bad domains.)
Let us now introduce the following framework for the remainder of this
chapter. The domain & will be a bounded open subset of R". The
operator A is a second-order strongly elliptic real operator in &, and B is
a real boundary operator of order d (d =0 or 1). In most of the
forthcoming sections, we shall make the following assumptions:

(a) the boundary I of S is of class C. (see Definition 1.2.1.1)

(b) the operator A is in divergence form:

Au = £ D, (a, D, u)
i.i=1

with air=a;;€C. (D) and there exists a >O such that

# a, (1) 485-0183 (2,1,1)

i.i = 1
for all x€S and EE RM.
 2.1 FOREWORD 83

(c) B is either the identity operator (thus d =0) or

Bu = £ b.D.u (2,1,2)
i=1
with b; E C". '(R), i sien (then d=1) and It-
"=, biv' +0 everywhere
on I. (In other words, I' is nowhere characteristic for B.)
For a given function f defined in 1 and a given function g defined on
T, we shall look for u defined in § such that
Au =f in 2 (2,1,3)
Bu =g on T.
For some technical reasons, it will often be convenient to consider the
related problem with an extra real parameter 1 as follows.
in 2
Bu = 8on I.
(2,1,4)
Later on, we shall add lower order terms to A and B and get rid of 1.
In the particular case where B = 1, our problem is just a Dirichlet
problem for the equation Au + 1u = f. Another particular case is when Bu
is the 'co-normal derivative' of u corresponding to A, i.e.

b; =i=1 on F,
where w, 1-jen are the components of the unit outer normal vector
field on I. Then, our problem is a Neumann problem for the equation
Au + lu =f. In the general case when d = 1, we are solving the equation
Au + du =f with an 'oblique' boundary condition.
Actually, we shall pose the problem (2,1,4) in the framework of
Sobolev spaces. Thus we shall look for conditions ensuring that
I.1 : U -> {Au + 1u, yBus
is an isomorphism from W3(8) onto L. (8) X W3-d-1/p(T), 1 < p 200.
Let us conclude this introductory section with some examples of the
results which we will look for in this chapter. These examples are related
to the Laplace operator A. First, Theorem 2.4.2.5 implies that for every
feL, (8) and every g€ W? "P(T), there exists a unique u€ WE(8) which
is a solution of
(Au =f in A
Iyu=g on Г
provided 1 <p <o and M is a bounded open subset of W with a C.
84 REGULAR SECOND-ORDER PROBLEMS

boundary. Then, Theorem 2.4.2.6 implies that for every feLp (8) and
every g € W!- IP(F) there exists a unique u € W2(8) which is a solution of
Au =f in 2
(a4 + b,u) = 8 on F
provided be C' (1) and b.>0 everywhere on I, under the same
assumptions as above on p and M. Similarly, Theorem 2.4.2.7 implies that
for every feL, (R) and every ge W W(I) there exists a unique u€
W2(1) which is a solution of
-Au tau=f in R
on Г
Yar= g
provided agE L* (R) and ay > B > 0 ae. in R, under the same assumptions
as before on p and S. Oblique boundary conditions are also considered in
those theorems.
Unless otherwise indicated, we only consider real-valued functions in
this chapter (with the exception of some proofs in Section 2.3.2 which
require the use of the Fourier transform).

2.2 Variational solution of special problems

The roots of almost all the forthcoming results lie in a basic existence and
uniqueness theorem for solutions in H'(R). This result is proved by the
variational method introduced first by Euler. A much more detailed
description of the extent of this powerful method can be found in
Magenes and Stampacchia (1958), Lions (1956), Necas (1967) and
Agmon (1959) for instance. We quote here the minimal material that we
will need in the following chapters. In particular, we restrict ourselves to
Dirichlet's and Neumann's problems although the variational approach
allows us to solve problems with an oblique boundary condition.
2.2.1 Existence and uniqueness
According to what is said above we are looking for u which is a solution
of
Au + 14 = ⅖ D, (a, D,u) + du =5 in 8 (2,2,1,1)
with either a Dirichlet boundary condition
U = ( on Г (2,2,1,2)
 2.2 VARIATIONAL SOLUTION OF SPECIAL PROBLEMS 85

or a Neumann boundary condition

ди = g on T. (2,2,1,3)
dUA
Euler's variational approach to these problems consists of viewing them
as the equation of critical points for some functional (see Section 1.1).
However, we shall use a slightly different setting based on the famous
Lax-Milgram Lemma. This will allow us also to deal with oblique
boundary conditions later on.
Lemma 2.2.1.1 Let V be a Hilbert space and let a be a continuous
bilinear form on V X V. (a does not need to be symmetric.) Assume that a is
coercive, i.e. that there exists a constant a >0 such that
a(u; u) =a lu
for all u€ V. Then for every continuous linear form I on V, there exists a
unique u € V such that
a (u;v)=/(w) (2,2,1,4)
for every U€ V.
Now the problem is to convert equation (2,2,1,1) and the boundary
condition into a problem of the form (2,2,1,4). This is achieved by
performing integration by parts, using Theorem 1.5.3.1. Let us assume,
for instance, that u€ H°(8) is a solution of (2,2,1,1), (2,2,1,2) and that
vE H'(M). Then we have

(2,2,1,5)

It is therefore natural to define a and I on V = H'(R) as follows:

ur dx
10)= fods.
With this choice of V, a and I, our u is a solution of problem (2,2, 1,4).
Conversely, it is easily seen that a is bilinear, continuous and coercive
on V for 1 ≥0, while I is continuous for f€ L2(82). Applying Lemma
2.2.1.1, we obtain the basic existence and uniqueness result for Dirichlet's
problem.
86 REGULAR SECOND-ORDER PROBLEMS

Theorem 2.2.1.2 For every feL(1) there exists a unique u € H' (1)
solution of equation (2,2,1,1), with the boundary condition yu =0, pro-
vided 1 ≥ 0.

Proof Identity (2,2,1,4) with all u € D(R), means that Au + du =f in the

sense of distributions. This is all the information that we can get from
(2,2,1,4) since D(8) is dense in H'(8). The fulfillment of the boundary
condition yu = 0 follows from Corollary 1.5.1.6. -

We turn now to the Neumann problem. Let us assume, as a starting

point, that u € H7(8) is a solution of (2,2, 1, 1), (2,2, 1,3) and that we have
UH' (N). Then we have

(2,2,1,6)

Accordingly, we define a as above and I as follows on V = H'(2):

140) = I, to dx -1, go der.

It follows again that u is a solution of problem (2,2,1,4).
Conversely, it is easily seen that a is bilinear, continuous and coercive
on V for 1 > 0, while I is continuous provided f€ Lz(8) and g€ La(I').
We again apply Lemma 2.2.1.1 for proving the basic existence and
uniqueness result for Neumann's problem:

Theorem 2.2.1.3 For every f€ L2(1) and g € La(I) there exists a unique
u€ H'(R) such that

È I a, D,uDio dx + a nod = fuds - sudo (2,2,1,7)

for all u € H'(S), provided ^ > 0.

If we restrict identity (2,2,1,7) to U € D(S) only, we check that

Au + du = f in the sense of distributions. Consequently we have u€
E(A, L2(8)) (a space defined in 1.5) and y du/Ava is defined as an
element of H 12(I). This allows one to prove that y dulava = g on I in
the sense of H-12(I) (see details in Lions (1961a)), but we do not need
this in the sequel.
 2.2 VARIATIONAL SOLUTION OF SPECIAL PROBLEMS 87

2.2.2 Smoothness
In this short section, we shall prove that the solutions to the Dirichlet and
Neumann problems that we obtained in 2.2.1 actually belong to H°(8).
The main tool for proving this is the well-known method of tangential
differential quotients due to Nirenberg. We shall use this method only
near a flat boundary, taking advantage of the invariance of our set of
problems under cut offs and C'. changes of coordinates.
Let us begin with the Dirichlet problem. Thus, let u € H'(8) be a solu-
tion of

- (2,2,2,1)

for all v € H'(M). Let A be any function in D(S) and set u, = Au. It is clear
that use H'(R) and that

for all v € H'(Q) with

fi = oft & fa, (D,O)(D, u) + D. La, (D,0)u J.

i.i=1

This function f, is again in L2(8) since a € C' " (5) and u€ H' (1).
Now let V be an open subset of h" and let $ be a C'.' diffeomorphism
of V onto a neighbourhood of the support of 0. Assume that

(We recall that we denote by " the half space defined by x, >0. In
addition, possibly, V does not cut the hyperplane {* =0}.) Then we
consider u2 = u, o d. Again we have uzE H' (U) and setting V = d-1 . We
have

(2,2,2,2)

for all veH'(U), where

1

f2= DO fod.
It is clear that f2€ La(U), akiE C'. (U). In addition, it follows from (2,1,1)
88 REGULAR SECOND-ORDER PROBLEMS

that
# at (1) 555-04157 (2,2,2,3)
k.1 -1
for all geR" and ye U, with some at >0.
The first step is the following.

Lemma 2.2.2.1 Under the above hypotheses, we have use H (U).

Proof We shall use identity (2,2,2,2) with a special test function U

deduced from uz. We observe that the support of uz is contained in the
inverse image of the support of 0, by . Consequently, the support of uz
is compact in V and cuts the boundary of U only on the hyperplane
*n =0. We extend Uz to ü2, a function which is zero outside of V. It is
clear that üzE H' (R).
We define o as follows:

h? üz
where Th is the operator defined by
Tin$ (x) =Q(x+ he;), 1sisn-1, heR,
e; being the unit vector in the direction of x;- We have v€ H'(U) for h
small enough. Writing identity (2,2,2,2) with this particular , we get

k.1=1

This identity implies the following, which is obtained through a discrete

integration by parts, the adjoint of the operator Tin being Ti-4.

k.1=1

h h -й

Then we observe that for any two functions a and • we have

Tin -1(asp) = Din -1
altinet a Tin-14.
 2.2 VARIATIONAL SOLUTION OF SPECIAL PROBLEMS 89

Therefore we find

is dx

- ü, dx.

From this and inequality (2,2,2,3) we deduce the following:

(2,2,2,4)
Here we use the norm of La(U), while M is a bound for all the Lipschitz
constants of the functions at, 1 &k, I &n, N is the maximum of D$ and
finally N' is the Lipschitz constant of D. We already know that
uze H'(U), therefore, from (2,2,2,4) we deduce that there exist two
constants C, and C2 such that

§ De" TIn I is
k=1 h (2,2,2,5)

owing to the following lemma (the proof is easy and left to the reader):

Lemma 2.2.2.2 For » € H' (R') we have

Tik-1 Isisn-1.
h
Next we again apply Lemma 2.2.2.2 to $ = ((Tin - 1) h)ü2; we thus get

k=1
90 REGULAR SECOND-ORDER PROBLEMS

and consequently

(2,2,2,6)
k=1

for 1 sisn - 1.
To conclude, for each i, we consider any sequence h; &0, such that
• Tin - 1
- й2, 1<k=n
h;
converges weakly to some limit k, in L2(R). This is clearly possible, due
to the properties of bounded sequences in a Hilbert space. We have
obviously
- Tih. - 1
Dk h;
in the sense of distributions, and consequently
(2,2,2,7)
for 1 si<n-1, 1-k&n. This shows that all second derivatives of U2
except Dhu are square integrable in U. However, it follows from
(2,2,2,2) that

+ =k.I=Dat, Dius + A | D$| u2 = f2

AnS-a*, so that we can
in U. Furthermore, from (2,2,2,3), we have at.,
write

a# k+152n-1
and this shows that Dhuze La(U). The proof of Lemma 2.2.2.1 is
complete. _
Now we prove the global result corresponding to Lemma 2.2.2.2.

Theorem 2.2.2.3 For every feLz(9) there exists a unique u € H= (8)

solving equation (2,2,1,1) with the boundary condition yu = 0, provided
1>0.
Proof We recall that from the beginning we assume (a), (b) in Section
2.1. Thus M is bounded and has a C' boundary, while a,; € C' (5) for
i, i = 1, ..., n. It is therefore possible to find a finite number of open
subsets Vi ISk&k, of R" together with C'' diffeomorphisms & from
 2.2 VARIATIONAL SOLUTION OF SPECIAL PROBLEMS 91

½k onto Pk(Vk), 1 Sk Sk such that

(a) Ok (Vk), 15k Sk, is a covering of 8

We observe that k need not meet the hyperplane X, = 0, in order that

the Ok (Vk) also cover M. We used here Theorem 1.2.1.5 which allows us
to consider @ as a n-dimensional manifold with boundary, of class C'.!
Rh.
With this covering of M we associate a partition of the unity Oks
1skSk, such that

(d) the support of Ok is included in Ok (Vk)

(e) Ek=1 Ok = 1 on 1.
We apply Theorem 2.2.1.2 to prove the existence of a solution u€
H'(S) to equation (2,2,1,1). Then Lemma 2.2.2.1 shows that for each k
(ORU) O PREH(UK).
We conclude by reconstructing u as follows

4 = 5 0,4 = ½ (04) O GOD' E HR (D)

k =1 k=1
due to Lemma 1.3.3.1. m
Corollary 2.2.2.4 The mapping
u-»{Au + lu; yu}
is invertible from H7(8) onto La (8) XH(T), for 1 > 0.

This is an obvious consequence of Theorem 2.2.2.3 using Theorem

1.5.1.2.
We shall now prove the same kind of results for the Neumann problem.
We start from u€ H'(M) fulfilling the same identity (2,2,2,1) for all
vEH' (8) (instead of H'(8)). Such a solution u exists by Theorem
2.2.1.3 with & =0. Then exactly the same proof as in Lemma 2.2.2.1
shows that

The corresponding global result is this:

Theorem 2.2.2.5 For every fe L, (1) there exists a unique u€ H=(8)

solving equation (2,2,1,1) with the boundary condition y dulava = 0, pro-
vided 1 > 0.
92 REGULAR SECOND-ORDER PROBLEMS

Proof The property that u€ H (2) is proved exactly as in Theorem

2.2.2.3. Then identity (2,2,2,1) shows that Au + 1u = f in the sense of
distributions (this uses o € D(8)). This allows one to rewrite (2,2,2, 1) as
follows:

- 2, 1, 9, DAD, o de = 2, 1, 1D, a, DuDo de

for all o€ H (8). Finally due to Theorem 1.5.3.1, this identity is equival-
ent to

for all woe H'(T). This shows that

du
YavA
in the space H'(F) (since u belongs to H7(8)). a

Corollary 2.2.2.6 The mapping

du
is invertible from H7(8) onto La(8)X HU(F) for 1 > 0.

This follows from Theorems 2.2.2.5 and 1.5.1.2.

2.3 A priori estimates

We now consider the general operators A and B introduced in Section
2.1. We no longer restrict ourselves to Dirichlet or Neumann problems.
We shall prove the basic a priori estimate:
ul2p.asCHAu+dulla.m.o+/yBul2-d-1/p.p.#3 (2,3, 1,1)
for ue W-(8). This estimate holds only for 1 large enough. This is
essentially the inequality in Agmon et al. (1959); however, the proof given
here is slightly different.

2.3.1 An inequality based on the duality mapping

The duality mapping from L (8) into its dual La (8) (with 1/p + 1/g = 1)
is the mapping u-> u* defined by
(lu(x)| 'son u(x) if u(x) ‡0
u* (x) =
10 if u(x) =0.
 2.3 A PRIORI ESTIMATES 93

The reason for introducing u* is that it is the unique function in L, (82),

such that

10.9.8.

The strong ellipticity of A allows us to prove some very useful

estimates for lullo.n.., just by multiplying the equation Au + lu = f, by u*
and integrating by parts. The boundary condition allows one to drop or to
estimate the boundary integrals that appear in the integration by parts.
This is the purpose of this subsection.
The differentiation of u* will be difficult at points where u vanishes,
since the sign of u will be undefined. So we shall approximate u' by u"
defined as follows for & > 0:
u'" (x) =(u (x)7+8)8-2312u(x). (2,3, 1,2)
Assuming that u € C'(8), we can differentiate us as follows:
Du* (x) = (4 (X)2+8)8-2112 Du(x)
+(8-2) (u (X)7+8)(8-41124(X)?D,4(x). (2,3,1,3)
Lemma 2.3.1.1 For u € C'(D), we have

(2,3, 1,4)
i.i -
for all xeS, where a = a int f1, p- 1}.
Proof We have

ii-1
>a int {1, p-1 (u? + E)(p-2112|Vulz.

Lemma 2.3.1.2 Let P be any first-order differential operator, with

Lipschitz coefficients, tangential to I, everywhere on I. Then there exists B
such that

(2,3, 1,5)
+1 14276/19.21214P-d
for all u€ C=(8).
94 REGULAR SECOND-ORDER PROBLEMS

Proof We have
duu* do
+ run" dx. (2,3,1,6)

Using Lemma 2.3.1.1 we deduce the following inequality:

du
• dUA
u; do

• (2,3,1,7)

We then transform the boundary integral. We have

du.it do =(DU + Pu)u, do - Pun" do

• AVA
and

1Pru?+E)pi] do.
D
We use the following auxiliary lemma, which we shall prove later.

Lemma 2.3.1.3 For all 4 € C'(D), we have

(2,3,1,8)

Setting 4 = (u? + 8)2, we finally obtain

p
We now take advantage of inequality (1,5,1,2). This leads to
 2.3 A PRIORI ESTIMATES 95
for all 8 >0. In other words, we have

Pun do
+J (2,3, 1,9)

We can choose & such that (CKp/4) Vs = a' and summing up from
(2,3,1,7) we obtain

If we choose ß large enough, this implies (2,3,1,5).

Proof of Lemma 2.3.1.3 Using a partition of unity and local (C.')
coordinates, it is enough to prove (2,3,1,8) when I is replaced by Rh-1, P
is a first-order operator with Lipschitz coefficients on D- and & has
compact support. Thus we have

and (2,3,1,8) follows. We observe that the constant C depends only on

bounds for the coefficients of P and their first derivatives. •

Lemma 2.3.1.4 Under the assumptions of Lemma 2.3.1.2, we have

+. Home
for all ue W3(8).

Proof We begin with u€ C'(8) and let & -> 0 in (2,3,1,5). It is obvious
that u" -> u* pointwise everywhere and that u* remains uniformly
bounded in S when & -> 0), because u is continuous in 2. Consequently,
by Lebesgue theorem we know that
ut.
96 REGULAR SECOND-ORDER PROBLEMS

strongly in La(8) and in La(F). Thus from (2,3,1,5) we deduce that

Now applying Hölder's inequality, we obtain

This is exactly (2,3,1, 10) when u€ C?(8). However, this inequality does
not involve u, so that it is easily extended to all u€ W7(8) by
density.
We are now able to prove
Theorem 2.3.1.5 Let A and B satisfy the assumptions in Section 2.1; then
there exists 1, such that, for 1 > 1,.

1Au + dullo.p.n (2,3,1,11)

Mallo.p.25-16
for all ue W; (8), such that yBu = 0.
Proof We consider first the case when the order of B is one, i.e. d= 1.
We observe that the boundary condition Bu = 0 may be rewritten as
yAu/Ava + Pu) = 0, for some tangential operator P with Lipschitz coeffi-
cients. Indeed we have:
, ди
Bu = 1-1
£ b,D. =b..Ju
where b, is the component of the vector b = (b,...,b) in the direction
of w and by is the projection of b on the tangent hyperplane to r. We
denote by V the tangential gradient on I. In the same way, we have
ди
duA 1=1
where c, = In
We assumed that I is not characteristic for B and this means that b,
does not vanish on I. Consequently we have

§ Bu=
du
b, AVA b.
 2.3 A PRIORI ESTIMATES 97

We define P by

Pu =
16,
and the boundary condition yBu = 0 implies

y
du'+ Pu) = 0.
LaVA
We observe the Cus bu, br, Cr are all Lipschitz functions since air and b,
are so and I' is of class C'. by assumption.
We now make use of inequality (2,3,1,10) and get the following:

uld.p.o€
1
Inequality (2,3,1,11) follows easily.
So far we have left out the case of a Dirichlet problem (i.e. B = 1 and
d=0). In that case (2,3,1,11) follows obviously from (2,3,1,10), with
P=0, say, because yu = 0 on I. Consequently (2,3,1,11) holds. _

2.3.2 An inequality in the half space

Here we consider a second-order, homogeneous, strongly elliptic
operator L, with constant real coefficients

ik =1
together with a first-order, homogeneous differential operator M, with
constant real coefficients

1=1
We assume that the hyperplane X =0 is not characteristic for M. This
means that mn + 0. The strong ellipticity of -L means that there exists
u >0 such that

£ 4455 5-14162 (2,3,2, 1)

1.k -
for all §eR". The corresponding boundary value problem in I=
{KER" In > O' is
Lu =f in R" (2.3.2.2)
• InMu =g on R° I
where Yn is the trace operator on Kn = 0.
98 REGULAR SECOND-ORDER PROBLEMS

The purpose of this subsection is to prove an estimate for a u € W3(84)

which is a solution of (2,3,2,2). Namely, we shall prove that there exists
some constant C such that
ul2.p.m: SCILullo.p.m.+Y..Mulls-1/p.p.rn-r+/4|11.p.084 ]. (2,3,2,3)
We shall also prove the corresponding estimate for the Dirichlet problem
(i.e. when M is replaced by I). This is the first step of the proof of Agmon
et al.'s inequality. The proof presented here is different.
We shall use two auxiliary results, one of which is the powerful
Lp-multiplier theorem of Mih'lin (1956) (see also Hörmander (1960)).
Here, we denote by R' the set " \0}. (In the rest of this section, we
allow our functions to be complex valued.)

Theorem 2.3.2.1 Let a€ C' (R) be such that there exists a constant C
with
El D'a(E) SC (2,3,2,4)
for all seR4 and lal En. Then the operator
g'> F 'aFg
is continuous in Lp (R") and there exists a function n, p -> K(n, p) such that
IF 'aFgllo.p.men = K(n, p)CHlgllo.p.sen (2.3.2.5)
for all geL, (R") and 1 <p<00.

We emphasize the fact that K(n. p) blows up when p-> 1 and when
p ->∞. The case p = 2 is useless since (2,3,2,5) holds with K(n, 2) = 1 and
C = maxp" al, by Plancherel's theorem.

Lemma 2.3.2.2 The mapping

where S, denotes the Dirac measure in the variable n is continuous from

W. (Br-1) into Ws+1/p-'(R") provided s <0.

Proof This is a very simple consequence of Theorem 1.5.1.1 since Y§ is

obviously the transposed operator of the trace operator y.. •

We shall also use an elementary solution E for L + 1, defined by

EER".
i.k -= 1

The assumption (2.3,2,1) implies the existence of a constant C such that

 2.3 A PRIORI ESTIMATES 99

a = DßFE fulfils (2,3,2,4) when |B|&2. Consequently the convolution

operator by E maps Lp (R") into WE(R").
We now use the elementary solution for reducing the boundary value
problem
(Lu + u =f in R$
(2,3,2,6)
I v.Mu =g on Br-1
to an equation on M-1. For that purpose we set
0=u-E* f. (2.3.2.7)
We then have
LU+v=0 (2,3,2,8)
I y.Mo =h
where h=g-Y.ME #f. We now denote by  the partial Fourier trans-
form of 4, in X1, . .., X-1 (or Fourier transform on R"'), i.e.,

468: 44) - 2r gun- 1721

where x'= (X1,..., X-1). We use the same notation for the Fourier
transform of a function defined on Br-1
1 eix*/(x) dx.
44(8) =12771-112
It follows from (2,3,2,8) that

(2,3,2,9)
Kn =0.
We now solve the differential equation in (2,3,2,9). The corresponding
characteristic equation is

ban +212 hits +11-2444554) =0. (2,3,2, 10)

If we set 5 = (St....,En 1, T/i), this is equivalent to

1-24455=0.
This equation has no real solution, due to the strong ellipticity of - L (see
(2,3,2,1)). Furthermore, since the lik • s are real, the solutions are conju-
100 REGULAR SECOND-ORDER PROBLEMS

gates. It follows that (2,3,2,10) has two symmetric, nonimaginary solu-

tions in r, which are functions of $. We denote them by
p= (E)
with Re p. (E) > 0 and Re p-(E) <0. It follows that ‹ is, for almost every £,
a linear combination of the functions
exp X,.P+(5). exp Xmp-(E).
However, due to the assumption that ^ is a Fourier transform, the fast
increasing component has to be excluded.

Lemma 2.3.2.3 Let vEHZ(R): be a solution of

Lo+0=0 in RI
then
0 (E, x4) = TAN (E) exp xrp-(5), Xn >0
for almost every ¿emph-1.

Proof At first glance we have

Ô (5, X) = a (E) exp xup- (5) + B (E) exp XP . (5), Xn> O
for almost every , where a and B are some functions. From this, it
follows that
a + B = YU a.e.
a.e.
and this shows that a and B are measurable functions since n and
YnDhU are square integrable measurable functions.
Then, from the fact that o belongs to La (R) we deduce that § also
belongs to L2(8) (in the variables (5, X)) and consequently B = 0 a.c. It
follows that a = YnU.

An immediate consequence of (2,3,2,10) is that

a.e.
and thus the boundary condition in (2,3,2,8) may be rewritten as follows,
where Ko = InU:

(m. p- + I im,s, ko = h. (2,3,2,11)

* It is enough to consider here the case p = 2, since later on we shall take advantage of the
density of W?(R)n H?(R4) in W3(R#)
 2.3 A PRIORI ESTIMATES 101

This is the equation on the boundary, which is equivalent to problem

(2,3,2,9).
The equation (2,3,2,11) is obviously uniquely solvable since the func-
tion §-> (MaP + Li=- im;S;) does not vanish on R". Indeed the m; are all
real and Rep (E) <0 everywhere. This leads to the representation for-
mula in Lemma 2.3.2.4.

Lemma 2.3.2.4 Let u€ H (R") be the solution of problem (2,3,2,6); then

we have

(2,3,2,12)
where ko=(mip.+In-lim,s,)-'h, k, =p-Ko and h=g -Y,ME * §.

Proof We first observe that we can apply Lemma 2.3.2.3 to 0 =

u -Es f, since f€ La (R) and consequently E*f€ H? (R").
Let us consider the equation of © (again û is the extension of u defined
by § =0 for X <0):

j.k =

i =1 j.k = 1

i=1
Since © is a tempered distribution we check by Fourier transform that

0=E= (Inly, 085: +7,D, 08811 72 2 41. D9,41 0 885.).

Then we derive ko = Yo from (2,3,2,11) and substitute YnDnU by kn

where k, =p-k, by Lemma 2.3.2.3. m

The representation formula (2,3,2,12) is the key tool for proving the
estimate (2,3,2,3). We need two more auxiliary lemmas.

Lemma 2.3.2.5 Let a fulfil the assumptions in Theorem 2.3.2. 1. Then the
operator
Mag-> F 'aFg
is continuous in W; (R") for all sER.
102 REGULAR SECOND-ORDER PROBLEMS

This result can be found in Triebel (1978), however, we can also obtain
it as a consequence of Theorem 2.3.2.1.

Proof of Lemma 2.3.2.5 Applying Theorem 2.3.2.1 with a replaced by

# (1+1813)m/2 a (E) (1 +|E13)-m/2
and remembering Definition 1.3.1.3, we check that Ma is continuous in
HE(R") for all m eZ. Now, H. (R") is the same space as W(R"). The
result stated in Lemma 2.3.2.5 for a non-integer s, follows by the
interpolation Theorem 1.4.3.5. _

Lemma 2.3.2.6 The functions

are bounded on pr-1 for all a.

Proof This is easily checked on the explicit formula for the roots of
(2,3,2,10)
1/2

It follows from (2,3,2,1) that Inn <0 and that the polynomial in the
bracket is always strictly positive. -
We are now able to prove the basic estimate.
Theorem 2.3.2.7 Let - L be a homogeneous strongly elliptic second-order
operator with constant coefficients and let M be a homogeneous first-order
operator with constant coefficients. Assume that x, = 0 is not characteristic
for M. Then there exists a constant C such that (2,3,2,3) holds for all
u€ WA(R').
Proof It is enough to prove inequality (2,3,2,3) for u € H? (R) n WE(R"),
since this is a dense subspace of W3(R). This allows one to use the
representation formula (2,3,2,12) for u. We shall consider each term
separately.
We start from feL, (R) and he W"- I (R"-"). Since E maps L, (R")
into W(R"), we have
(2,3,2,13)
Then let us set

b(E) = (m. p. (8) +; 'E m.=)

i-1
 2.3 A PRIORI ESTIMATES 103

It follows from Lemma 2.3.2.6 that b, $;b, 1 sjen -1 and p_b, all fulfil
(2,3,2,4). Consequently the mappings
h->ko=F-'OFh
h-> Diko = F'15;6Fh,
hr>k,=F-'p_bFh
are continuous operators in W1-1/P(pr-1) owing to Lemma 2.3.2.5. Thus
we have
Mkoll2-1/p.p.sams+Mlkills-1/p.p.spmCzMlll-1/p.p.spn-1. (2,3,2,14)
From this we deduce that

Then SEW"- " (R"-1) and DisE W, " (R"-1), Isjen-1. Lemma
2.3.2.2 implies that
508,€ W. ' (R') and D,588, EW, '(R'), 15jEn -1.
Consequently

DiE# (588) =E* (D;588n)E Wb(R"), 1si&n-1

by Lemma 2.3.2.5. In other words, we have shown that
E*[OS.)EL. (R")

DiD.E*(588n)ELp (RM), 25j+k52n-1.

We just need to check that DhE * (5 @ 8,)E L. (R") in order to prove that
E # (58 8,)E WE(R"). This is achieved by using the fact that E is an
elementary solution for L + 1. This implies that in R$ we have
DhE # (588) =-
Summing up, we have shown that
E* hank, +2 2 1, D, ko 88, E W. (R")
and in addition, we have

112, p.38".

(2.3.2.15)
owing to the continuity of all the involved operators.
104 REGULAR SECOND-ORDER PROBLEMS

Finally, let us consider E # (k, & Sh). We start from ko€ W3-1P (R"-"),
so that knEW, "P (RM -"), DiKoE W, " (R"-"), 1 5j En - 1 and D, DaKoE
W; "'(R'- 1), 1€j. ken -1. From Lemmas 2.3.2.2 and 2.3.2.5 it follows
that

DiE# (k,88n) ELp (R"), 15j&n-1

DiDkE * (k, 88n) EL(R'), 15jksn-1.
Then we write that
DhE # (k. 08h) = DhE * (k. ASh)
==
since E is an elementary solution for L + 1. It follows that in R4, we have

DrE# (k,&8n) =-

j=1
and that
DiD.E# (k,OSh) =
1
non

ik=
1sisn-1.
Again applying Lemmas 2.3.2.2 and 2.3.2.5, we show that

DiD,E* (k, ASh)ELp (R'), 1sisn-1.

The only derivative missing for proving that E# (ko&Sh) € W: (R) is
DhE * (k, A 8"). Using the property of the elementary solution E, we
derive this last fact as we did for D}E * (5 8 8n). Summing up, we have
proved that

and in addition,
HE #ko@8ndll2.0.021sCallkoll2-1/p.p.pomi (2,3,2,16)
 2.3 A PRIORI ESTIMATES 105

Putting together identity (2,3,2,12) with inequalities (2,3,2,13) to

(2,3,2,16), we obtain the existence of a constant C such that
[ul 2.p.m: S C[| Lu + ullo.p.se: + lYn Mulls -1/p.p.589-1]
for all ue W? (R) M H (R%). By density, the same is true for all u€
W(R#) and inequality (2,3,2,3) follows obviously. This completes the
proof of Theorem 2.3.2.7.
Remark 2.3.2.8 Inspection shows that the constant C (deduced from C
to Ca) is bounded by a continuous function of the lik and the m;
The similar statement concerning Dirichlet's problem is this.

Theorem 2.3.2.9 Let - L be a homogeneous strongly elliptic second-order

operator with constant coefficients. Then there exists a constant C such that
(2,3,2,17)
for all u € W-(R#).

Proof We use here the same representation formula (2,3,2,12) but with
ko=h=g-Y,E*§ and k, = p-ko. The rest of the proof is exactly similar
to that of Theorem 2.3.2.7. -

2.3.3 A general a priori estimate

We consider again the general operators A and B of Section 2.1, in a
general bounded domain M with a C'.' boundary. We shall now extend
inequality (2,3,2,3) to this general case. Namely, we shall prove that there
exists a constant C such that
Mullp.asCIlAulla.p.at/lyBull-d-1/p.p.t+lull.p.r). (2,3,3,1)
Then we shall combine inequalities (2,3,1,1) and (2,3,3,1) to obtain the
basic inequality for the remaining sections of this chapter.
Inequality (2,3,3,1) is very flexible because of the norm of u in W:(1)
that appears on the right-hand side. Indeed, it allows us to 'localize the
inequality. This property is rigorously stated as follows.

Lemma 2.3.3.1 Assume that each point x€S has a neighbourhood Vx

such that (2,3,3,1) holds for all the functions u in W?(1) which have their
support in Vx. Then (2,3,3, 1) holds for all u € W=(S) (with possibly another
constant).

Proof The compactness of M allows us to find a finite number of points

X1, ..., ty in S, such that S is covered by the interiors of Vx, 1 5j&N.
106 REGULAR SECOND-ORDER PROBLEMS

Then we choose a partition of unity corresponding to this covering.

Namely we assume that

1= , 0;
i=1
on R, where O; € D(R) and the support of 4; is contained in the interior of
Vx 15j<N.
The assumption of Lemma 2.3.3.1 is that (2,3,3,1) holds in particular
for all the O;u. It follows that
N
j=1

¡=1
sCIllAullo.p.st/lyBull2-d-1/p.p.r+|ulli.p.pl
(2,3,3,2)
j=1
Here [A; el is a first-order operator with continuous coefficients so that
there exists C2 such that
LA; O;Julo.p.aC2|u|1.p.p. IsjEN (2,3,3,3)
for all u€ W3(8). The same way, [B; 0,] is either O when d=0 (i.e.
B =I) or the multiplication by a Lipschitz-continuous function when
d = 1. In both cases there exists C3 such that
yB;0,Jul2-d-1/p.p./SCs|uli.p.p. (2,3,3,4)
for all u€ W$(M). Inequality (2,3,3,1) follows from (2,3,3,2), (2,3,3,3)
and (2,3,3,4) by addition.
We shall now prove inequality (2,3,3,1).
Theorem 2.3.3.2 Let A and B fulfil the assumptions in Section 2.1; then
there exists a constant C such that (2,3,3,1) holds for all u € W. (R).
Taking advantage of Lemma 2.3.3.1 we shall restrict ourselves to
proving inequality (2,3,3,1) in those two particular cases.
Case (a) the support of u is compact in 1.
Case (b) the support of u is contained in (V) where V is an open
neighbourhood of O and is a C'. diffeomorphism of V onto (V)
such that
D '(DND(V) =U=RIMV.
 2.3 A PRIORI ESTIMATES 107

We observe that the norms involved in (2,3,3,1) are invariant under

C1.1 changes of coordinates. Furthermore the properties of A and B are
also invariant under C1 changes of coordinates. That is why we shall
consider u • instead of u. This reduces the proof to the particular case
where the intersection of the support of u with I' is contained in {Kn = 0}.
The case (a) is solved with the help of this lemma.

Lemma 2.3.3.3 For all y € S, there exists a neighbourhood V of y in 12,

such that (2,3,3,1) holds for all u € W7(8), whose support is contained in
V.
Proof We use the famous perturbation argument known as Korn's
procedure. Freezing the coefficients of A at y, we obtain an operator with
constant coefficients

1= £ 4,D,D;,
where 4; = a;.;(y), which satisfies the assumptions of Section 2.3.2.
We then observe that

in 8. If we denote by a any Lipschitz functions defined everywhere,

such that a; = ai; on S, we have

i.i=
Now we assume that the support of u is contained in V such that V c R.
Using the elementary solution E introduced in Section 2.3.2, we obtain

•1

Since E * is a linear continuous map from L. (R) into WE(R"), it follows

that

Let us now call & the diameter of V and K a bound for the Lipschitz
constants of all the ai; in M. We then have
Mullo.nsCIllAulla.m.atn=K8|lull.a.n)+Callulli.p.p.
We conclude by choosing 8 small enough; indeed, if we assume that 8 is
108 REGULAR SECOND-ORDER PROBLEMS

less than or equal to 1/2C?K, we have

lullap.ss2C|Aullo.p.at2Clulli.p.m.
This is exactly inequality (2,3,3,1) for u, since the support of u is
contained in VC& and consequently Bu =0. The proof of Lemma
2.3.3.3 is complete.
Let us now consider the case (b). It is solved with the help of this last
lemma.
Lemma 2.3.3.4 Let yeT have a neighbourhood W in F, contained in the
hyperplane {*n =0}. Then there exists a neighbourhood U of y in A such
that (2,3,3,1) holds for all u € WE(D), whose support is contained in U.

Proof We shall use the same perturbation argument as in the proof of

the previous lemma. We freeze the coefficients of A and B at y and
obtain operators with constant coefficients:

L=i.i=1
§ L.D,D;
M= either I or £ m,D,
j=1
where 4.; = a.;(y) and m; = b; (y) (when d = 1). These operators satisfy the
assumptions of Section 2.3.2.
We start with an open neighbourhood U of y in & such that Unde
W, and we assume that the support of u is contained in U. We then have

in Und and

in W if d = 1, while Ynu = Y.Bu if d = 0. We again denote by Qi,;

(respectively B;) any Lipshitz functions defined everywhere, such that
Qi = di, in M (respectively B; = b; in 1). We have

in R$ and
Y, Mü = y,Bu - v. E (B, - M,) Diu
 2.3 A PRIORI ESTIMATES 109

on ix, = 03 if d = 1, while V,u = Yn Bu if d = 0. Since &€ WE(R") we can

use estimate (2,3,2,3) proved in Theorems 2.3.2.7 for d = 1 and 2.3.2.9
for d=0. It follows that

when d = 1 (the additional boundary term Er=1 I b; - m;)D;ull2-d-1/p.p.P

does not appear when d = 0).
Let us again denote by 8 a bound for the diameter of U and by K a
bound for the Lipschitz constants of all the as and b;. It follows that
ulz.p.asChAulo.p.stlyBul2-d-w/p.p.rstC,lull.p.s
tn=KC&llull..at&Callb,-m,)Dyulli.p.s (2,3,3,5)
j=1
by the trace theorem (Section 1.5.1). Let us consider separately the last
term of this inequality. We have

(2,3,3,6)
From (2,3,3,5) and (2,3,3,6) we deduce that
lulz.p.o-ChAulo.p.o+yBull2-d-1/p.p.#3+Caulli.p.r
+in'C+nC_K8u|2.p.o.
Choosing & small enough so that In?C+ nCs\K8½, we obtain finally
ulp.as2CMAullo.p.stlyBul2-a-w/p.p.rs+2C4|u|.p.o.
This is inequality (2,3,3, 1).
Proof of Theorem 2.3.3.2 We apply Lemma 2.3.3.1. The existence of V,
follows from Lemma 2.3.3.3 when x€ while it follows from Lemma
2.3.3.4 after change of coordinates, when xer. In this last case we
assume at once that V. C@(V), where (V, $) is a map of the manifold 12,
near * (see notation above). This allows us to 'flatten' the boundary I,
near x, by replacing u by u.?. -
Remark 2.3.3.5 Actually we have proved a little more than Theorem
2.3.3.2 and this will be useful in the next subsection. In the proof of
Lemma 2.3.3.1, it is enough to cover the support of u by the interiors of
110 REGULAR SECOND-ORDER PROBLEMS

V., 15j≤N. Consequently we can release the assumptions on 2. Let us

assume that M is a (possibly unbounded) open subset of pm with a C1.1
boundary. Then for each compact subset K of M, there exists a constant C
(depending on K) such that inequality (2,3,3,1) holds for all u€ W5(8),
with support in K.
We are now able to perform the final step of our search for a priori
estimates.

Theorem 2.3.3.6 Let A and B fulfil the assumptions in Section 2.1, then
there exist C and No such that
lulla.nCIllAu+dullo.p.stlyBull2-d-1/p.p.r] (2,3,3,7)
for all ue W7(1) and 1 > No.

Proof We first improve inequality (2,3,3,1), using Theorem 1.4.3.3. We

have
K (2,3,3,8)

for all u € W⅔(8) and & > 0. Choosing & > 0 small enough and substituting
(2,3,3,8) in (2,3,3,1) we obtain:

lulla.p.asC.IlAullo.p..+yBul2-d-1/p.p.r+|ullo.p.o)
€GilllAu + dullo.p.at|lyBull2-d-1/p.p.r + (1 + 1) llullo.p.02].
(2,3,3,9)
We now take advantage of (2,3,1,11). We have

lullo.p.aS
1-10- Au + Aullo. p.r
for u€ W3(8) such that yBu = 0. From (2,3,3,9) it follows that

1-10
This is exactly (2,3,3,7) in the particular case where Bu = 0.
Let us now consider the general case. From Theorem 1.6.1.3 we know
that there exists a linear continuous operator R from W2-d-1/ (I) into
W⅔(8) such that
yBRg = g
for all g € W3-d-1/P(I). We set o = u - RyBu. It is clear that v € W3(52)
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 111

and that Bu =0 so that (2,3,3,7) holds for v. It follows that

lulla.p.o@0|2.p.o+||RyBul2.p.r
@ClAv+10llo.p.stCslyBull2-d-1/p.p.r
SCillAu+dullo.p.a + | A + 1) RyBullo.p, n) + Cally Bull2-d-1/p.p.r

This is (2,3,3,7). -

2.4 Existence and uniqueness, the general case

In this section we derive a general existence and uniqueness result for

problem (2,1,4) as a consequence of the a priori estimate of Section 2.3.
Then we remove the parameter ^ and attempt to solve problem (2,1,3).

2.4.1 The basic result

We shall show that under the assumptions of Section 2.1, the mapping
Tor: u-> {Au + du; yBus

is an isomorphism from W(8) onto Lp (8)X W3-d-1/P(T), for 1 large

enough. For that purpose, we shall consider successively the three cases
(a) p = 2, (b) p <2, (c) p > 2.
The starting point of the proof consists in observing that T21 is a
semi-Fredholm operator for 1 large enough. Indeed, from (2,3,3,7), it
follows that T2 is one to one and that the image of T, is closed in
L2 (R) X H312-4(F). This allows us to consider the index of T2 which is (in
this particular case):
x (T2,1) =-def R(Tz.1)
where R(Tzn) is the image (range) of T2. and def R(T2.›) is the
codimension (possibly infinite) of R(T2.‹) i.e., the dimension of
{L2 (8) XH31-4(T)/R(I2.1).
It is well known that the index remains constant when one performs a
homotopy from an operator to another, remaining in the set of all
semi-Fredholm operators (see Kato (1966), Chapter IV, $5). We shall
only use the following very simple form of this general principle:

Lemma 2.4.1.1 Let X, Y be a pair of Banach spaces and let t-> T, be a

continuous mapping from a, b] (a and b are any real numbers) into the
space L(X; Y) of all continuous linear operators from X into Y. Assume
112 REGULAR SECOND-ORDER PROBLEMS

that for each t, there exists C, such that

XEX (2,4,1,1)
Assume Ta is an isomorphism; then Th is also an isomorphism.

A direct elementary proof, avoiding the general theory of semi-

Fredholm operators, can be built from the fact that isomorphisms define
an open subset of L(X; Y).
Many of the estimates that we have derived involve a parameter 1
whose lower bound depends on the particular problem which is under
consideration. In performing homotopies from one problem to another,
this will cause problems. This is why we shall use the following technical
lemma.
Lemma 2.4.1.2 Let X, Y be a pair of Banach spaces. Let tr> T, be a
continuous mapping from [a, b] into L (X, Y). Let also S be a fixed element
of L(X; Y. Assume that for each t there exists C, and 1, such that
XEX (2,4,1,2)
for all 1 >1,. Then there exists C and i such that
XEX (2,4,1,3)
for all 1 at and t€ [a, b].
Proof Consider any pair of numbers t and t' in [a, b]. Then we have, for
1>1,
If we assume that t and t are close enough to one another, we have

since the mapping tr T, is continuous. It follows that

for 1 31,. This shows the existence of C and 1 locally. The desired result
follows, since La, bj is compact. _
We now prove our basic result
Theorem 2.4.1.3 Let A, B and 1 fulfil the assumptions in Section 2.1.
Then for 1 <p <o, there exists Ap such that
Th..: ur> {Au + du; yBus
is an isomorphism from W3(8) onto L. (1)X W3-d-119(5) for all 1 3 1,.
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 113

Proof for p = 2 This is nothing but Corollary 2.2.2.4 when B = 1, i.e.

when we are solving Dirichlet's problem. The same way, this is nothing
but Corollary 2.2.2.6 when B = d/Ava on I, i.e. when we are solving
Neumann's problem.
Let us consider now an oblique boundary condition (i.e. d = 1). We
shall perform a homotopy from Neumann's problem to our problem. For
that purpose, we introduce the operators

B.u = (1 -t)
dUA
t€[0, 1].
We observe that I is not characteristic for B, for all t€ [0, 1], provided
bu= En 2, biv; <Ot (if this is not the case we replace B by -B). Accord-
ingly, we can apply Theorem 2.3.3.6 to the mapping
T,:u->{Au; yB,u}
and there exists C, and 1, such that
lul2.2.a SGIlAu+dullo.r.o+|7B,u|1/2.2.1]
for all we H7(8) and 1 3 1,.
Let us now set X = H'(D), Y = L2(9)X H'?(I) and
S:ur>{u, 0}.
It is obvious that tr T, is continuous from [0, 1] into L(X; Y). Applying
Lemma 2.4.1.2, we find 1 and C such that
lull2.2.0 SCIllAu+Aullo.2.52+|lyB,ulh112.2.r] (2,4, 1,4)
for all u€ H? (8), 1 z^ and t€ [0, 1].
A first application of Lemma 2.4.1.1, using homotopy in 1 instead of 1,
shows that To+ 1S is an isomorphism for all 1 z^, since by Corollary
2.2.2.6 we already know that To + AS is an isomorphism for 1 large
enough.
A second application of Lemma 2.4.1.1, using homotopy in t with a
fixed 1 > 1, shows that T, + 1S is an isomorphism, since To + 1S is so. This
proves Theorem 2.4.1.3 when p = 2.

Proof of Theorem 2.4. 1.3 for all p <2 Inequality (2,3,3,7) shows that Tp
is one to one and has a closed range for 1 large enough. On the other
hand, the result already proved for p =2 shows that the range of Tp.
contains L2(9) X H 2(I), since H(D) S W7(8). Consequently, the range
1.1 is also dense; this proves that Tp. is onto.

† We recall that b, does not vanish on I, since I is not characteristic for B.

114 REGULAR SECOND-ORDER PROBLEMS

Proof of Theorem 2.4.1.3 for all p >2 We shall make use of this auxiliary
smoothness result which will be proved later on.

Lemma 2.4.1.4 Let A, B and 1 fulfil the assumptions in Section 2.1. Let
u € W: (8) be a solution of
in O
on Г
where feLp (8), 8€ WE-d-11P(I). Assume that pErn/(n-r) if r <n. Then
u € W=(D).

Exactly as in all the preceding cases, it follows from inequality

(2,3,3,7) that Thp. is one to one and that its range is closed. To prove that
Th. is onto, we start from feL, (8) and g€ W2-d-1/P(I). We have
consequently feL(8) and g € H3/2-d(T). Applying again the result al-
ready proved for p =2, we know that there exists u€ H (8) such that
{Au + du =f in O
yBu = g on Г
for 1 large enough. A (possibly iterated) application of Lemma 2.4.1.4
shows that u€ W$(8). Consequently, Tp. is onto. •
Proof of Lemma 2.4. 1.4 It uses methods very similar to those in Section
2.2.2. Accordingly, we shall first localize our problem with the aid of
cut-off functions. Then we shall use a C. change of coordinates to
flatten the boundary. Finally we shall use Friedrichs' mollifiers method
instead of Nirenberg's tangential differential quotients. This is to prove
smoothness in the case where the boundary is flat.
Thus, let A be any function in D(8) and set us=Ou. It follows that
u, € W(8) and that
f1 = Au, = Of + [A; eJuEL. (8).
Indeed the assumption on p ensures that W; (1) is contained in Lp (1) by
Sobolev's imbedding theorem (see Section 1.4.4). Then we have

when d=0 and

81 =yBu, =0g+y[B;o]u€ WI-1/P(r)
when d= 1.
Now let V be an open subset of ph and let $ be a C'.I diffeomorphism
of W onto a neighbourhood of the support of A. Assume that
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 115

(We do not exclude here the possibility that FAQ(V) = O.) Then we
consider uz = u, ° D. Again we have uze W; (U) and setting V = $-1 . We
have

k.1=1
in U (2,4,1,5)

where

i, =1
fz=|DO|f°Ф.
Clearly we have fzeLp(U), at E CO. (U) and in addition

# at (y) 54515-a* 1513

k.1=1
for all geR" and all ye U, for some a# >0. We also have
(2,4, 1,6)
when d = 0 and

(2,4, 1,7)
¡=1
where

i =1
if d=1. Clearly again, we have 82€ W3-d-10(V(*n =04), b$E CO. (U)
and

i=1
does not vanish on {* = 0} since w and the gradient of Ym are parallel.
A first technical step is:

Lemma 2.4.1.5 Under the above assumptions, we have uzE W(U).

Then, since & is bounded and has a C' boundary, it is possible to find
a finite number of open subsets V 15k&N of R", together with C.I
diffeomorphisms from V Onto Ok (Vi), 1 €k &N, such that
k=1 is a covering of 1
116 REGULAR SECOND-ORDER PROBLEMS

With this covering of M, we associate a partition of unity Ok, 1 sk&N.

such that
(c) OK ED (R)
(d) the support of Ok is included in Ok (Vk)
(e) EX=1 Ok = 1 on 2.
Now Lemma 2.4.1.5 shows that for each k, we have
(OKU) O DIE W3(UK).
Consequently

И=
k=1 k =1

Here the symbol ~ means that the function has been continued by zero in
1 \ k(Vk). Lemma 2.4.1.4 is proved.
Before proving Lemma 2.4.1.5, let us quote and prove one particular
form of the famous mollifiers lemma due to Friedrichs. Here we denote by
Pms m = 1, 2, ... a sequence of functions belonging to D(R-1) such that
Pm (x') =m"-'p (mx')
where pe D(R"-1) is such that Jon 1 p(x) dx' = 1. Therefore, the convolu-
tion by Pm is an approximation of the identity operator when m -> +00.

Lemma 2.4.1.6 Let a be a uniformly Lipschitz function on R4; then

there exists a constant C such that
(2,4, 1,8)
for all m and 1 si&n- 1.

Proof Explicitly we have for v € D(R$):

(apm # Div -Pm * aD;U) (X)

+ Pm (K'-Y)Dia (y', xp)U(J', In) dy'.

man 1

Consequently we have the following estimate where K denotes the

 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 117

Lipschitz constant of a:
(aPm * DiU -Pm * aD;v) (X)I

Applying Young's inequality, we obtain

Mapm * Div -Pm * (aD;u lo.p.Be:

[lp(x)|+|x'||Dip(x)|]dx'.

This is exactly (2,4,1,8) when v € D(R). The general case follows by

density.
Proof of Lemma 2.4.1.5 The main idea is to apply inequality (2,3,3,1) to
a sequence of smooth functions u½, m = 1, 2, ... which approximates U2.
It is convenient to extend uz in , defined, as usual by
uz in U
in R$\U.
Since us has compact support in V, it is clear that ü2€ W7(R$). We extend
fz and 82 in a similar fashion. Then I€ L, (R) and 82€ WE d-11P (R"-").
We extend also the functions at, and bf to the whole of R° in any way
that preserves all the properties of A# and B#. Accordingly, we have
in R$
{ A is =1 On ph-1
We now set
U2 =Pm * U2.
We first show that une W(R). Indeed, we know that u, € W p(R") by the
Sobolev imbedding. It follows that
DiD;unEL. (R'), iti&2n-1
since the effect of Pm n * is to smooth up the functions in the directions of
k; 1 sisn-1. It is a little more tricky to show that Dhun€ Lp (R$).
Indeed, we observe that

"SAMEitis2n-1
añn
-I
118 REGULAR SECOND-ORDER PROBLEMS

Consequently, we have

if we agree to consider , as a vector-valued function of X- Smoothing

with Pm *, we obtain
DiumEL. (R$).
We observe in addition that ur remains bounded in W(R'):
(2,4,1,9)
We now show that A#um = fm remains bounded in Lp (4). We use
Lemma 2.4.1.6 to compare A#uh with Pm * f2. First, since u2 € W7 (R) S
W"(R), we know that

remains bounded in Ip (R%) for 1 sk <n- 1, 1515n. Then we write

Diüz k=1 I=1 ahn

This shows that
•1

Diüz = F- (2,4,1,10)
k =1
where F and G. 1-k &n- 1, belong to Lp (R). Thus we have

k=1

and this is bounded in L (R), owing again to Lemma 2.4.1.6. Adding, we

obtain the boundedness in Lp (R$) of

k.1= k.I=
Clearly, it follows that there exists C such that
(2.4.1.11)
Finally we show that y,B# u# = g" remains bounded in W3-d-1/0 (R"-1).
In the case where d = 0, we simply have ynB*u#= Ynum = Pm * 82 and the
claim is obvious. In the case where d = 1, we again compare y,B# u? with
Pm * 82. First, since zE W(R), it is clear that
bF Pm # D;ü2 -Pm # b Diû2,
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 119

remains bounded in Lp (R) and that

Da (b#Pm # D;Tz-Pm * b#D;42) = (Dab$) Pm * D,#2 - Pm * (Dab t) D,Ur
+[b$Pm # D; DAü2 - Pm * b# D, Daü2)

also remains bounded in Lp (R) provided 1 <j&n -1, 1€k &n, because
of Lemma 2.4.1.6. Then we write
Da (b#Pm '# Diü2 -Pm # b# Dnü2)

+[b#Pm # D7ü2 -Pm # b#Drus]

= (Dub#) (Pm * Dnü2) - Pm * (Debt) Daüz+ b# Pm # F - Pm * b# F

- 2 IbMPm * DiGK - PIm * b# D, Gk]

k=1
owing to (2,4,1,10). Using again Lemma 2.4. 1.6 we show that there exists
C2 such that

Taking the traces, we deduce that

(2,4,1,12)
The conclusion of the proof is now straightforward. The functions u
have their support in a fixed compact set. This allows us to use inequality
(2,3,3,1) (see Remark 2.3.3.5). Accordingly there exists C4 such that

The estimates (2,4,1,9), (2,4,1,11) and (2,4,1,12) imply then that u,

m = 1, 2, ... is a bounded sequence in W(R). On the other hand, we
have
m -> +00

in W(R). This implies that üz€ W-(R). The proof of Lemma 2.4.1.5 is
now complete.

2.4.2 Applications of the Fredholm theory and the maximum

principle
So far, we have dealt with operators A and B respectively fulfilling the
assumptions (b) and (c) introduced in Section 2.1. We are now able to
widen our class of operators by adding lower-order terms. Thus we now
 120 REGULAR SECOND-ORDER PROBLEMS

assume that

Au = ≥ D, (a, Du) + ≥ a, Dut asu

i,j=1 i =1
where ai; = a, € C° (8) fulfil (2,1,1) again and where a; € L° (2), O sis
n. In addition B is either the identity operator (d =0) or

Bu =2b.D.utb.u
i=1

where b; € C'. '(S), Osien and b, = Er-, bu does not vanish on T

(d =1). It will be convenient to assume that b, <0 on I (by possibly
changing B to -B).
Adding lower-order terms to A and B means adding a compact
operator to Tr p,. Indeed, it follows from Theorem 1.4.3.2 that

"+>
is a compact mapping from W7(8) into Lp(0)xW$-1P(F). Adding this
to Tp., which is an isomorphism for 1 large, implies the following lemma:
Lemma 2.4.2.1 The mapping
Tp: ur> {Au, yBul
is a Fredholm operator of index zero from W. (R) into Ip(h)x
W3-d-1/P(F).

In other words, this means that the operator under consideration has a
finite dimensional kernel and a range of finite codimension. In addition,
the codimension u of its range is equal to the dimension of its kernel (see,
for instance, Theorem 5.26, $5, Chapter IV in Kato (1966)).
The problem of showing that the mapping Tp is actually an isomorph-
ism is now reduced to showing that u =0, i.e. that Tp is one to one. This
is a much simpler question since we have some strong smoothness results
for functions in the kernel of Tp.

Lemma 2.4.2.2. Let u € W(D) be a solution of

{ Au=0 in O (2,4,2,1)
(yBu = 0 on r;
then u€ MIcaco WE(R) = C'(M).
By the way, this shows that ker Tp does not depend on p.
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 121

Proof of Lemma 2.4.2.2 The differentiability of u up to the boundary

follows from Lemma 2.4.1.4. This lemma does not apply directly to A
and B since we have weakened our assumptions on these operators.
However, it applies to A and Bo defined as follows:

A, u = I D. (a,; D, u)
i.i = 1

and Bou is either u (d = 0) or

Bou = £ b, Du.
i=1

Indeed it follows from (2,4,2,1) that

Diu E W!(R)
and that yu = 0 when d = 0 and yBou € W2-1/P(I) when d = 1. In all cases
Sobolev's imbedding theorem implies that
{ AguELa (82)
YB.u € W2-d-1/9(F)
where q&pn/(n-p) for p <n and q <0 for p = n. Lemma 2.4. 1.4 shows
that u € W7(1).
Iterating the previous procedure eventually shows that

1<9<00

We conclude by using again Sobolev's imbedding theorem which implies

that u€ C'(R). •
Of course, the result of Lemma 2.4.2.2 is an invitation to use the
maximum principle for showing uniqueness. The proof of uniqueness
would be quite simple if we also knew that u € C'(R). However, this may
not be true under our assumptions on the coefficients a;, 1 <i&n. Thus
we shall make use of the generalized form of the maximum principle for
weak solutions, due to Stampacchia (1965). This author considers general
weak solutions in H'(R). Here we shall take advantage of the smoothness
proved in Lemma 2.4.2.2 to give a simpler proof.

Theorem 2.4.2.3 Let u eMIspso WE(M) be a solution of Au = 0 in 8.

Assume that either a; =O, 1 sisn and a >0 or ao >B > O, then

max (2.4.2.2)
XES u(x) < max (0, max u (x))
122 REGULAR SECOND-ORDER PROBLEMS

The proof of this result will follow after some preliminaries. We set
k = max (0, maxxer U(x)) and
Uk (x) = max (u -k; 0) (2,4,2,3)
Lemma 2.4.2.4 Uk belongs to W(M) for all p.
Proof We can redefine uk as being k° u, where
Фк (t) = max (t - k; 0)
This is a uniformly Lipschitz function with Lipschitz constant equal to
one. Since u € C'(R), it follows that u is also a Lipschitz function and
furthermore, applying a theorem of Rademacher (1919), we see that
Dink =(Qk° u) Diu
almost everywhere. It follows that |Du |Diu and consequently U €
W(8) for all p. In addition, we have shown that

(2.4.2.4)

Finally, to show that yuk = 0, we approximate k by means of a

sequence of functions Ok.ms m = 1, 2, ... such that
(a) Ok, mE C' (R)
(b) Ok.m is uniformly Lipschitz continuous, with Lipschitz constant equal
to one
(c) 054km SOk
(d) 4k.m -> Ok uniformly when m -> +00.
Then we approximate ux by $k.m ° u. It is obvious that 4k,m ° UE C'(D)
and that k,m • u vanishes on I. Then we show that there exists an
increasing sequence m;, ¡ = 1, 2, ... such that k,m, o U -> UK in W(R).
This implies that yuk = lim; so Y (k, m, ° U) = 0, i.e., ur EW | (1).
Actually we have
Фк.m ° 4 (X) -> UK (X)
for all x and in addition
(Qk.m ° 4) (X)| 5 u (X)I
(DiCk.m ° U) (x) | S| Diu (X) I, 1<i<n
for all x. Applying Lebesgue's dominated convergence theorem we find
an increasing sequence m; and functions U;, O<j&n in Lp (8) such that
Ok.mo U-> Vo
DiCk,m°U-=>Vis
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 123

in Lp (R). Applying also Lebesgue's subsequence theorem we can achieve

the choice of the sequence m; in order that
Ok.m OU -> Vo
almost everywhere. We conclude by observing that vo = u almost
everywhere on the one hand and that U; = Divo, si<n in the sense of
distributions on the other hand. Thus UK = vo€ W(8) and Ok.m o U -> UK
in W(M). This completes the proof of Lemma 2.4.2.4. •

Now, as in Section 2.3.1, we shall consider the corresponding function

w through the duality mapping from I (8) into La (R), i.e.,
u# (x) =|Uk (x)18-2 UK (X). (2,4,2,5)
Since we shall only use large values of p, we can view uk as do Un, where
4p (1) =|18-2, (2,4,2,6)
is a continuously differentiable function. Since function u is uniformly
Lipschitz continuous, we can again apply Rademacher's theorem to
differentiate uk. This leads to the following identity
D;us (x) =(p-1) |UK (X)18=2 DiUk (X) (2,4,2,7)
almost everywhere, 1 < <n. This together with identity (2,4,2,4), implies
the following:
fO ae. in <x u(x) =k}
(2,4,2,8)
Du*(x) = (p-1) 14. (x)IP -2 D.u (x) in {x | u (X) > KI

Proof of Theorem 2.4.2.3 We start from the identity

which is obvious, and then we integrate by parts. Since ux € H' (R), we

obtain

This is equivalent to the following, where Ax denotes the set {x€

2 u(x) > k}:

- ¿ (-1)An Ak
124 REGULAR SECOND-ORDER PROBLEMS

We now denote by M an upper bound for lail, 1 sisk in M. It follows

from (2,1,1) that we have:

JAk Ak

A,

Then using Cauchy Schwarz inequality, we get the following inequality

for all & >0
Mn
a(p-1) -ZESdAn
Mn? ( 2 | unP dx €0,
(2,4,2,9)
where B Sao(x) ae. We finally chose & small enough so that B-
(Mn/2) € ≥0. This is possible under the assumptions of Theorem 2.4.2.3,
which mean that either B>0 or M = 0 if B = 0. Once we have chosen &
we can find p large enough such that a (p -1) > Mn/2&. From (2,4,2,9) we
conclude that

a.e. in Ax. Equivalently, we have

(u -k) Wu =0
everywhere in Ax (since u € C'(R)). In other words
W(u-k)2=0
in Ax and consequently u -k is constant in Ax. On the other hand, we
have u =k on the boundary of Ax; thus u = k everywhere in Ax. This
means that usk. _
It is now easy to deduce several uniqueness theorems corresponding to
various kinds of boundary conditions, from Theorem 2.4.2.3. First let us
consider the Dirichlet boundary condition.
Theorem 2.4.2.5 Let 1 be a bounded open subset of in with a C1,1
boundary. Let a,; be uniformly Lipschitz functions and a be bounded
measurable functions such that a;, = aij 1 si, jsn and that there exists
a>0 with
£ a: (x)55.5-a|E|
i. = 1
for all geR" and for almost every xe R. Assume in addition, that either
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 125

a; =0, 1 sisn and a >0 ae. or ao > B > 0 a.e. Then for every f€ Lp (12)
and every ge W3-1/P(F), there exists a unique u € WE(S) solution of

I D.,(a, D,u) + I aD,u tank =f in 2

yu =8 on Г.
Proof According to Lemma 2.4.2.1, we just have to prove that Tp is one
to one. Thus let u eker T.. From Lemma 2.4.2.2, we know that actually
ueker Tp for all p. Then applying Theorem 2.4.2.3, we have
max u(x) = 0
X€S

since maxxer u(x) = 0. The same holds for -u, so that u =0. •
Let us now consider the so-called third boundary value problem and
more generally an operator B of order d = 1, with a nonzero coefficient
bo.
Theorem 2.4.2.6 Let S be a bounded open subset of R", with a C'1
boundary. Let ai,; and b; be uniformly Lipschitz functions and let a be
bounded measurable functions in S. Assume that a, = a;, 1 si, jen and
that there exists a >0 with

§aii (x) 55, 5-a lEl2

i.i=1
for all & eRr and almost every x € M. Assume in addition, that either a; =0,
1sisn and ao=0 ae. or ao> B > O a.e. in M. Assume finally that

on I. Then for every feL(1) and every ge WI-"P(I), there exists a

unique u € W7(8) which is a solution of

Èi.i D,
=1
(a, Du) + ≥
i= a, D, ut a,u =f
in 2
on Г.

Proof Again, owing to Lemma 2.4.2.1, we just have to prove that Tp is

one to one. Thus let u€ Ker Tp. We know from Lemma 2.4.2.2 that
U EMI-pe WE(M). This allows us to apply Theorem 2.4.2.3. We want to
126 REGULAR SECOND-ORDER PROBLEMS

prove that
max u(x) €0. (2,4,2,10)
XEN
Assume the contrary; then necessarily, the maximum of u is attained
on the boundary I. Since the first derivatives of u are continuous up to
the boundary, the boundary condition is fulfilled in the classical sense. In
other words, we have

5 b. (x)D,4 (x) +b. (x) u(x) =0,

i=1
for all x € r.
At the particular point x where u reaches its maximum, the tangential
derivatives of u vanish and the derivative of u in the direction of w is
nonnegative. We can rewrite the boundary condition at x as follows:
(X) + b. (x0) u (X) = 0.

This is contradictory since we assumed that u(x) > 0 and that b, (x) and
b. (x) are both nonzero numbers and have the same sign. This shows that
(2,4,2,10) holds. The same holds for - u, so that u = 0. -
In the next statement we shall allow b, to vanish so as to be able to
consider a Neumann boundary condition for instance. As a counterpart,
we have to assume that a, > B > 0 a.e.
Theorem 2.4.2.7 Let S be a bounded open subset of in, with a C1.1
boundary. Let a,; and b; be uniformly Lipschitz functions in 1 and let a; be
bounded measurable functions in S. Assume that ai; = a;,in 1 Si, jEn and
that there exists a >0 with
£ a (7) 585-0163
i,j=1
for all geR" and almost every x€S. Assume in addition that a, >B > O
ae. in R and that

j=1
b,‡0
on I. Then for every feL, (1) and every ge WHIP(r), there exists a
unique u € WE (1), which is a solution of

≥ D, (a,;, D,u) + I a, Diut aou =f ins

ij =1 i =1
(2.4.2.11)
on T.
 2.4 EXISTENCE AND UNIQUENESS, THE GENERAL CASE 127

Proof We introduce a function p€ C'. (8) such that p > O in 12, p =0 on

r and Ap/Av <0 on I.* Then we define v by setting
u = exp (-Ep)o
and we show that u is a solution of problem (2,4,2,11) if and only if u is a
solution of a problem which fulfils the assumptions of Theorem 2.4.2.6, at
least for & >0 small enough. The result will follow by applying Theorem
2.4.2.6 to v.
Indeed, we have
Diu = exp (-sp) (DU -E[D;p]v),
and consequently

An = È Dila,; exp (-Ep) (D, 0 -E(D,p 10))

+ E exp (-sp) a; (DU - E[D,p]u) + exp (-sp)ago.

It follows that

exp (Ep) Au = E ID, a;, D, 0 - ED, (a4; Dip) 0 - 8a4; D,pD,0

i, =1

- Ea, DiP (D,0 - E[D,p Jo)) + I a, (D, 0 - 8 [Dip]o) + and

and finally that

+63 I a, DipD,p lo= exp (ep)f.

i.i=1

On the other hand, we have

Bu = E b, exp (-Ep) (DU -E[D,p]o) + b. exp (-Ep)o

i=
† It is easy to define p locally near the boundary. In the notations of Definition 1.2.1.1, we
can define Pv as being (y', Yn) -> q(y') - Yn. Then covering I by a finite number of
hypercubes such as V, we build up a function p from the py • s with the help of a partition of
unity.
128 REGULAR SECOND-ORDER PROBLEMS

and consequently

on Г.
We now check that the problem of which v is a solution fulfils the
assumptions of Theorem 2.4.2.6. Indeed we have

do -E La,D,p -E £ D, (a;,;D,p) +8 E a, DipD, p = B12>0

i= 1 ii = 1

ae. in § for & small enough, if a >B>O a.e. in §. Then we have

[8025u Eb=160-6663983863-8639830
since b, does not vanish and Ap/av <0 everywhere on I. m

2.5 Other kinds of solutions

2.5.1 More on smoothness

If we add the same amount of smoothness-so to speak--to the boundary
of M, to the coefficients of the involved operators and to the data of our
problem, we obtain eventually the same amount of extra smoothness for
the solution. More precisely, let k be any positive integer and consider
the operators A and B introduced in Section 2.1. Assume that
(d) the boundary I of 12 is of class Ck+1,1
(e) air = a; i € Ck. " (1), b; € CK. ' (B).
Then we have the following smoothness result.

Theorem 2.5.1.1 Let u€ WE(8) be such that

- yBu = gE W3+k-d-1/0 (F)

then u€ Wk+? (52).
Proof This follows very closely the proof of Lemma 2.4.1.4 to begin
with. That is why we use the same notation. In addition, it is clear that we
can prove by induction on m that u€ W2+m(1) implies that u€
watm+ (R) provided m k-1. Thus we have to prove that uz€
Our additional
W?-m smoothness
* (U), knowing that uzEhypotheses,
WE*' (U). imply that at eCk. (U), b$€
 2.5 OTHER KINDS OF SOLUTIONS 129

Ck. (U), fE Wm* (U) and 82€ Wm+1-d-1/9(U). At this step, instead of
using Friedrichs' mollifier technique, it is possible to go back to Niren-
berg's tangential differential quotients as in the proof of Lemma 2.2.2.1.
However, for the sake of using the notation of Lemma 2.4.1.4, we
proceed with Friedrichs' technique.
With the help of Lemma 2.4.1.6, it is easy to check that
u% =Pm * UZE W3+m+' (R"),
that us remains bounded in W+m (R) and that A#um is bounded in
Wm+'(R), while yB#un is bounded in W?+m+1-d-1/p (pr-1). Then the
inequality (2,3,3,1) shows that um is actually bounded in W3+m+" (R).
Letting m ->∞, this shows that 2€ W3+m + (R*) and consequently usE
W?+m(U).
The remaining steps of the proof are exactly similar to the correspond-
ing steps in the proof of Lemma 2.4.1.4.

Remark 2.5.1.2 To each of the existence and uniqueness results of

Section 2.4.2, corresponds a result with additional smoothness, proved
with the aid of Theorem 2.5.1.1. Briefly, those results are the following.
Under the assumptions of Theorem 2.4.2.5, plus the hypotheses (d) and
(e) above and if a; eCk-1. (1), O sisn, the mapping (notation as in
Section 2.4.2)
U-> {Au; yu}
is an isomorphism from Wk+2(1) onto WK (8)× WK+2-1/0 (T).
Under the assumptions of either Theorem 2.4.2.6 or Theorem 2.4.2.7,
plus the hypotheses (d) and (e) above, and if a; € Ck-1. (5), b; € Ck. ' (5),
the mapping (notations of Section 2.4.2)
UH> {Au; yBu}
is an isomorphism from Wk+2(8) onto W$ (8)X WK+1-1/P(r).

Remark 2.5.1.3 Under the hypotheses of this section we have

ker T. = C'(8).
This makes the proof of Theorem 2.4.2.3 much simpler (see Hopf (1927)
for instance).

2.5.2 Very weak solution

Here, for the sake of later reference, we prove a simple basic result which
is obtained from the results in Section 2.4.2, by applying the transposition
procedure of Lions and Magenes (1960-63).
130 REGULAR SECOND-ORDER PROBLEMS

Let us begin with Dirichlet's problem.

Theorem 2.5.2.1 Let the assumptions of Theorem 2.4.2.5 be fulfilled.
Assume in addition that a; = 0, 1 -isn. Then the mapping
u-> {Au, yul
is an isomorphism from D(A; L. (8)) onto L. (2)X Wo"P(I').
Proof Let us consider the mapping

which is an isomorphism from W2(8) onto L. (R) x W2-1/9(T). The trans-

posed operator T° is also an isomorphism. Assume that p'+ g *= 1 and
consider feLp (8) and g € W. 1/P(I). Define a continuous linear form on
W=(2) by

20).
Here the brackets denote the duality pairing between Wt-la(r) and
W= IP(F).
Since Ta is an isomorphism, there exists a unique u € I (M) and a
unique 4є W.'-1/P(F) such that

10)=wAwdx+468:701

for all o€ Wa(8). In other words, we have dU

(2,5,2,1)

for all o€ W2(8).

If we use this identity with veD(2) only, we check that Au =f.
Consequently u belongs to D(A, L (2)). This allows us to use Green's
formula (see identity (1,5,3,5)): we have
dU
(2,5,2,2)
Mara
for all o€ Wa(R). It follows from (2,5,2,1) and (2,5,2,2) that
dU
Y•ди
Ava AVA
for all o€ W4(2). By the trace theorem 1.5.1.2 this implies that
du-4: 40) =474-8:448
dvA
 2.5 OTHER KINDS OF SOLUTIONS 131

for all No € W2-19(8) and all Me WI-1/a(I). Consequently we have

yu = g
and
ди
© =Y
dUA
This proves the desired result. A
Actually we shall only use this consequence of Theorem 2.5.2.1.

Corollary 2.5.2.2 Let the assumptions of Theorem 2.4.2.5 be fulfilled.

Assume in addition that a; =0, 1 si sn. Let u € D(A, Lp (8)) be a solu-
tion of
{Au = fEL. (82)
I yu = gE W2-110(5)
then u€ W3(1).

This is a straightforward consequence of Theorems 2.4.2.5 and 2.5.2.1.

The corresponding result for Neumann's problem is this

Proposition 2.5.2.3 Let the assumptions of Theorem 2.4.2.5 be fulfilled.

Assume in addition that a =0, 1 si &n and that ao > B > 0 a.e. in M. Let
U ED(A; Lp (R)) be a solution of

S Au-rEL, (8)
Mara
then u € W7(1).

The proof of Proposition 2.5.2.3 is similar to the proof of Corollary

2.5.2.2. The corresponding statement for an oblique boundary condition
requires a little more smoothness on the coefficients.

Proposition 2.5.2.4 Let the assumptions of Theorem 2.4.2.7 be fulfilled.

Assume in addition that a =0, 1 siEn and that b, € C'. ' (D), 1 sjEn.
Let u E D(A; Ip (82)) be a solution of

then u € W3(2).
3
Second-order elliptic
boundary value problems in
convex domains
3.1 A priori estimates and the curvature of the boundary
One of our basic tools throughout Chapter 2 has been the a priori
inequality (2,3,3,7) proved in Theorem 2.3.3.6. In the present chapter we
propose an alternative proof of this inequality in the particular case when
p =2 and when the function u € H?(S) under consideration fulfils the
homogeneous boundary condition

yBu = 0.
We shall mainly consider boundary value problems for the Laplace
operator (in order to avoid some extra technical difficulties). However, we
shall allow some nonlinear boundary conditions.
The main idea of the forthcoming alternative proof is to bypass the use
of local coordinates. These were used in Section 2.3 to reduce the
problem to the case when the boundary I of the domain M is flat. Here
we shall perform straightforward integration by parts to prove the in-
equality

lull2.2.n SC(8)lAullo.2.5 (3,1,1)

for all u € H?(R) such that yBu = 0 on I. The constant C(R) takes into
precise account the curvature of I.
This inequality has various applications, all of which are along the
following lines. We shall consider very rough domains &, such as general
convex domains or domains whose boundary has turning points. We shall
approximate these domains by sequences of domains with a C2 boundary
for which the constant in inequality (3,1,1) can easily be controlled.
Taking the limit will prove smoothness results for the solution of a
boundary value problem in , although M is far from having a C2
boundary.
132
 3.1 CURVATURE OF THE BOUNDARY 133

3.1.1 An identity based on integration by parts

In this section we shall consider a bounded open subset 2 of R" with a
C2 boundary I together with its second fundamental quadratic form,
denoted by B. Let us recall briefly an elementary definition of B. For that
purpose, let P be any point on F. It is possible to find n- 1 curves of class
C2 in a neighbourhood of P, passing through P, and being orthogonal
there. Let us denote by C1, ..., Cn-1 those curves, by T1, ..., Tn-1 the
unit tangent vectors to %1, - .., Cn-1 respectively, and by s1, ..., Sn-1 the
are lengths along %1, ..., Cm-1 respectively. We can assume that
{51, ..., Tn-1} has the direct orientation at P.
Then, at P, Bp is the bilinear form (we shall often drop the subscript P
and write B instead of Bp, whenever this does not lead to any misunder-
standing)

-• THENK
ick as;
where $ and m are the tangent vectors to I at P, whose components are
151 .., Sn-1s and M1, -.., Mn-it, respectively, in the basis {51, ..., Tn-1).
In other words, we have

B (5, n) =-
dv
•M

where d/a$ denotes differentiation in the direction of £. (Actually, we

could also extend the definition of B to sets N with a Cl. boundary with
just a little more extra work. All the subsequent results hold for domains
with a CI. boundary instead of C2.)
Another point of view is this. Let us consider a point P of I and
(according to Definition 1.2.1.1) related new coordinates {Y 1, . .., Yn} with
origin at P as follows: there exists a hypercube
V=<(yr..., Yn)l-a;<y;<a;15j=n}
and a function & of class C2 in V', where

such that
I (y')lsa,/2 for every y'€ i'
enV={y =(y', yn) EVIYn SQ(y'))
rnV=(y=(y', yn)EV| In =Q(y')}.
Let us assume further that W. (0) =0. This means that the new coordi-
nates have been chosen in such a way that the hyperplane Yn = 0 is
134 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

tangent to I at P. Then, it is easily checked that the form B is

98(8,n) = 5 89 (0) 5k Mis

where {£1 .., En-is and {M1, - .., Mn-1 are the components of & and n
respectively in the directions of {Y, - .., Yn-1%. In particular, when & is
convex, the function - is also convex, and consequently the form B is
nonpositive.
Let us observe, finally, that in the general case of a domain & with a
C2 boundary the form B is uniformly bounded on F. In other words,
there exists K such that

for all PeT, where § and m are tangent vectors to I' at P. Indeed, v is a
C' vector field on I. This is why domains with a C2 boundary (or even
Cl.) are said to have a boundary with bounded curvature.
Now we introduce some more notation. Let v be any vector field on Г;
we shall denote by , the component of v in the direction of v, while we
shall denote by vr the projection of v on the tangent hyperplane to F. In
other words
and VT = V-U,v.
In the same way, we shall denote by V the projection of the gradient
operator on the tangent hyperplane:
du
Vru =Vu-v.
dv
We can now state the following.
Theorem 3.1.1.1 Let O be a bounded open subset of Dr with a C2
boundary, and let ve H'(2)". Then, we have

I divoriax-E, J, ax,
Ax; an
Ax; ax

+ (tr B) [(yv) • v]'} do. (3,1,1,1)

Here tr B is the trace of the bilinear form B, i.e.,

tr B = -
=1 ds;
in the above notation.
 3.1 CURVATURE OF THE BOUNDARY 135

Proof First, we apply repeatedly the Green formula of Theorem 1.5.3.1

which holds as soon as I is Lipschitz. We thus get, for ve C'(R)",

, дх; дх;

DUin, do
ii=1

ax; ax; dUi vi do

„ д0; ,v; do.
i,i=1
In other words, we have

1(0) = IdivwFax-E/ 30,20 dx

i.i=1 da Ax; Axi

(3,1,1,2)

We shall now transform the integrand on the boundary. This can be

done locally. Let us consider any point P on I and choose an open
neighbourhood W of P in I small enough to allow the existence of (n -1)
families of C2 curves on W with these properties: a curve of each family
passes through every piont of W and the unit tangent vectors to these
curves form an orthonormal system (which we assume to have the direct
orientation) at every point of W. The lengths S1,..., Sn-1 along each
family of curves, respectively, are a possible system of coordinates in W.
We denote by ..., Tn-1 the unit tangent vectors to each family of
curves, respectively.
With this notation, we have

Vr= k=
where U; = v•; We also have for any € C'(R):
дф
V, Vro=
and consequently

div v =
E1 ds; dv
Consider now the first integrand on I in the right-hand side of
136 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

(3,1,1,2). We have

div v = j=1 ds;

j.k =1 Las; 1.8;

du• "
K=1LdU ask).atlav
and thus

U, div r =00=1
2. ds; i.k =1 {=1 ds;
don,
+ k=
2 OK av au (3,1,1,3)
since (Av/Av) • v = 0.
We then consider the second integrand on I in the right-hand side of

(3,1,1,2). We have dv dv
(v• V)v=
j=1
дт
j,k Dilas;
=1 Ids;
- 1
dv
k=
and thus
n -1 n=1 n -1 d0,
{(v• V)v}• v=
5 =0;0k
i.k 1 37.
as; 0 + 2 ", "s k=1
(3,1,1,4)
because (Aw/ds;) • v = 0.
Subtracting identities (3,1,1,3) and (3,1, 1,4), we finally obtain

U, div v - {(v • V)v} •v=0,

5'29 on дбк.r;
+ 0.ickI =1
=1 As; 35,
n: n-1 do,
+0,=1+ As;
20 .74 -isk2,=1upon054.8-5.91551
as i= 1

(3,1,1,5)
On the other hand, using S1, - .., Sn-1 as coordinates in W, we know
 3.1 CURVATURE OF THE BOUNDARY 137

that B is defined by

док ENK
B3 (5, 1) == 5 ik=1 dS; j,k =1 As;
and consequently we have

tr B = - ET.:dv•
-1

i=1 ds;
It follows that (3,1,1,5) may be rewritten as:

U, divv -{(v. V)r}-V=0,

E1 as; I unark.r;
as;
j.k =1

30,. (3,1,1,6)
(88303-88144547)-2455
Let us finally calculate

dive (U,VT) j=1

=S ik=1 as;
Thus, we have

dive (0.44) = 2 100. are to. as

n -1
(3,1,1,7)
i as 584 70. 2i.kUK
=1 as,
Then, from (3,1,1,6) and (3,1,1,7), we deduce
U, div v - {(v • V)v} • v
-1

(3,1,1,8)
This expression of the integrand on I' in (3,1,1,2) no longer involves the
particular coordinates in W. Varying W, it is consequently true
everywhere on I. Thus, we have

=-21. 11-850. do -J, (47 89103788487:41 do. 13,1,19

Indeed, the integral of divy (u,) is zero since the vector field U, VT is
everywhere tangent to I.
138 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

We have proved (3,1, 1,9) assuming that v€ C2(8)". However, since the
boundary of M is of class C?, the space C(D) is dense in H'(M). From
(3,1,1,9) we deduce (3,1,1,1) by approximating vE H'(1)" by a sequence
Uma m = 1, 2, ... of elements of C'(R)r for which (3,1,1,9) holds. •
Theorem 3.1.1.1 is thus completely proved. Most of the previous proof
can be carried out under weaker assumptions on 1. We shall say that a
bounded open subset of R" with a Lipschitz boundary I' has a piecewise
C2 boundary if F = ToUTs, where
(a) To has zero measure (for the surface measure do).
(b) I is open in I, and each point x € 11 has the property of Definition
1.2.1.1 with a function & of class C°

Theorem 3.1.1.2 Let M be a bounded open subset of R" with a Lipschitz

boundary I. Assume, in addition, that I is piecewise C?. Then for all
VEH?(2)h we have

i,j=1

- Her 88103 + 28(47: v,)I do.

(3,1,1,10)
Proof This is quite similar to the proof of Theorem 3.1.1.1. Indeed,
(3,1,1,2) holds since I' is Lipschitz. Then identity (3,1,1,8) holds at any
point of I. Integrating (3,1,1,8), we obtain (3,1,1,10) since To has
measure zero. Finally, we extend (3,1,1,10) from vEC(D)r tO vE
H°(p)n only, since it is now impossible to give a meaning to the bracket
of we and WrU, when ve H'(R)" and I is only Lipschitz globally.

3.1.2 A priori inequalities for the Laplace operator revisited

We now take advantage of the results of Section 3.1.1 to prove inequality
(3,1, 1). Such an inequality has been proved by Caccioppoli (1950-51) and
Ladyzhenskaia and Ural'ceva (1968) in the case when O has a C1,1
boundary. These latter authors call it the second fundamental a priori
estimate'. All of them make use of local coordinates in order to flatten
the boundary. The proof given below follows Grisvard and looss (1975);
it allows better control of the constant C(R). It also allows one to
consider some nonlinear boundary conditions. A slightly different point of
view is developed in Lewis (1978) for two-dimensional domains.
For the sake of clarity, we shall first consider the particular case when
the operator A is the Laplace operator 4, or the modified Laplace
 3.1 CURVATURE OF THE BOUNDARY 139

operator 4 - 1 with ^ >0. The first inequality concerns a Dirichlet

boundary condition.

Theorem 3.1.2.1 Let M be a convex, bounded open subset of in with a C2

boundary I. Then there exists a constant C(R), which depends only on the
diameter of 1, such that
lull2.2,n =C(8) |Aullo.2.0 (3,1,2,1)
for all u€ H? (8) n' (R).
Proof We first apply identity (3,1,1,1) to = Wu, observing that, since
yu = 0 on I, we also have (yv)r = y VTU = 0 on I. Thus, we have

lax; ax; Ir
Due to the convexity of &, we have tr B €0 and consequently

lax; ax;
dis lAuf dx. (3,1,2,2)
i,=1
So far, we have estimated the second derivatives of u. The estimate for
the first derivatives is well known to be obtainable by the straightforward
integration by parts which follows. We have

lax dx=- su-udx lAulliall

where the norm is the norm of La(9). On the other hand, the Poincaré
inequality implies that

where K(R) depends only on the diameter of M (see Theorem 1.4.3.4). It

follows that

§ дх dx = K(8)2|4ul? (3,1,2,3)
and that
lull<K(8)3||Aull. (3,1,2,4)
Adding up inequalities (3,1,2,2) to (3,1,2,4), we obtain inequality
(3,1,2,1) with C(8)3<1+ K(R)2+ K(8)4
Remark 3.1.2.2 In the case when we assume to be only a bounded
140 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

open subset of " with a C2 boundary without the assumption of convex-

ity, we again obtain inequality (3,1,2,1). This is achieved with the aid of
identity (3,1,1,1), inequality (1,5,1,2) and by using an upper bound for
tr B. Consequently, the constant C(M) depends not only on the diameter
of M but also on an upper bound for tr B. In other words, the constant
C(S) depends on the curvature of I. This is nothing but an alternative
proof of the corresponding inequality in Section 2.3.

Let us now consider a Neumann boundary condition and even some

nonlinear boundary conditions closely related to the "third boundary
problem'. Here we consider a real-valued, nondecreasing function B
defined on the real line. In addition we assume B (0) = 0, and we assume
to be uniformly Lipschitz continuous. We now deal with the following
boundary problem for a function u € H2(1):
in R,
(3,1,2,5)
= B(yu) on T.
-YOu
The corresponding estimate is the following:

Theorem 3.1.2.3 Let S be a convex, bounded open subset of mr with a C2

boundary I, and let B be a uniformly Lipschitz, nondecreasing function
such that B (0) = 0. Then we have

412.2.0 S (3,1,2,6)

for all u€ H?(R) such that -y dulav = B(yu) on I and all 1 > 0.

The particular case when the function B is identically zero is just a

Neumann problem. Obviously the interest of inequality (3,1,2,6) is that
the constant depends neither on M nor on 3. This will allow us to extend
widely the possible Ms and the possible Bs, in the following sections.

Proof We again apply identity (3,1,1,1) to v = Wu. The boundary condi-

tion now means that -(yv) • v = (yu) on I. Thus we get
1 22и[2
ii=1
dx
=+2(V+ (yu); V…B(yu))
 3.1 CURVATURE OF THE BOUNDARY 141

Due to the convexity of &, B is nonpositive. On the other hand,

yUE H32(F), and since B is uniformly Lipschitz, we also have B(yu) €
H'(F). This allows one to rewrite the bracket as an integral. Conse-
quently, we have
| 2Zu 12
lax, ax, dx3+2 Ma(ye) - V, B(ya) do.
The integrand on I is
2W7(yu)VIB (rU) =2B' (yu) IVy(yU)12;

this is a nonnegative function since B is nondecreasing. We conclude that

(3,1,2,7)
# L lax;
lasAx;al
i.i=1
The estimate of the remaining terms in the H (R) norm of u is
obtained, as usual, by integrating (- Au + du)u. Indeed, we have
ди
I, Yau ru do.
Consequently, we have

Since we assume B to be nondecreasing and B (0) =0, it follows that

ß(yu)yu ≥0 a.e.
on I. Then, we have

dx sIl-Au + dulllull
and consequently

(3,1,2,8)
and

дх;
drs!-I- Au + Null?. (3,1,2,9)
The conclusion follows from inequalities (3,1,2,7) to (3,1,2,9).

Remark 3.1.2.4 Again we can drop the convexity assumption on 1 and

let S be any bounded open subset of " with a C2 boundary. Then, we
142 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

deduce from (1,5,1,2) and (3,1,1,1) the inequality

lul22.o SC(1;1)|1-Au+dul0.2.0 (3,1,2,10)
for every u € H?(8) such that -y du/Av = (yu) and every 1 > 0. Here the
constant C(1, S) depends only on 1 and on the curvature of 1 (or more
precisely, on an upper bound for B). It is important to observe that
C(1, 1) does not depend on B.

Remark 3.1.2.5 A priori bounds in H? for solutions of the Laplace

operator under oblique boundary conditions are also proved in Subsec-
tion 3.2.4 in the particular case n = 2.

3.1.3 A priori inequalities for more general operators

The purpose of this subsection is to extend the results of the previous
subsection to the more general operators A that we introduced in
Chapter 2. Here we shall no longer consider nonlinear boundary condi-
tions. This is to avoid some very cumbersome calculations which can be
found in Grisvard and Iooss (1975).
Accordingly, we consider an operator A defined by

Au = I D, (ai;D, u)
i.i= 1
with a, = a;; € CO. (8). We assume again that -A is strongly elliptic; i.e.
there exists a > 0 such that

§ a; (x) 55-5-01512 (3,1,3,1)

i. = 1
for all x€ A and FER".
We only consider here Dirichlet and Neumann boundary conditions.

Theorem 3.1.3.1 Let S be a convex, bounded open subset of Dr with a C2

boundary. Then there exists a constant C(S; A), which depends only on the
diameter of S and on the Lipschitz norms of the coefficients ai,;, 1 si, jen,
such that
lull22n € C(D; A) |lAullo.2,0 (3,1,3,2)
for all u€ H?(R) nH'(R).
Proof We could again use identity (3,1,1,1) with
v = A Vu
where st is the matrix of the ai However, the less natural method of
 3.1 CURVATURE OF THE BOUNDARY 143

proof that we shall follow here is simpler. Namely, we shall deduce

inequality (3,1,3,2) directly from inequality (3,1,2,1) through the same
perturbation procedure that we already used in Section 2.3.3. The main
step is the following:

Lemma 3.1.3.2 Each point y €S has a neighbourhood V, such that

(3,1,3,3)
for all u € H2(9) MH' (R) whose support is contained in V. Furthermore,
the constant Cy depends only on the diameter of & and on the Lipschitz
norm of the aij The neighbourhood V, depends only on the Lipschitz norm
of the dijo

Proof As in Lemma 2.3.3.3 we freeze the coefficients of A at y. Thus,

we set 4,; = a, (y). This defines a strictly negative symmetric matrix L, and
consequently there exists a nonsingular matrix R such that - RLR is the
identity matrix (R is the inverse of the square root of -L.) If we set
U(x) =4 (R-'x),
then the equation

£ 4. D, D,u = f
i.i=
is equivalent to
-40 = g.
We can apply inequality (3,1,2,1) to v.
Precisely, our assumptions on u imply that
VEHZ(RA)MH'(RM).
On the other hand, RM is convex and has a C. boundary. Consequently,
we have
02.2, Rn S C(RM) |1810, 2, RA.
Going back to the original variables, we also have

(3,1,3,4)
10,2.г
where K is a continuous function of the matrix R and of the diameter of

We compare the right-hand side in (3,1,3,4) with the norm of Au in

144 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

L2(8). Actually, we have

£ 4, D, D,u -Au = I D, (4, - an)D, u

i.i= i,j=1

= £ (4 - a,)D,D,u - £ (D, a. )(D,u).

i,j = 1
It follows that

(3,1,3,5)
where M is a common bound for the Lipschitz norms of all the aij
From (3,1,3,4) and (3,1,3,5), it follows that u |2.2.0 < K(R, 12)

The inequality (3,1,3,3) follows by choosing the neighbourhood Vy of y

small enough to ensure that

for all x € V.y•

The proof of Lemma 3.1.3.2 is complete. The claim in Theorem 3.1.3.1
follows easily with the aid of a partition of the unity on 0. •

We turn now to the Neumann problem.

Theorem 3.1.3.3 Let S be a convex, bounded open subset of R" with a
C2 boundary. Then there exists a constant C(1, A), which depends only on
1, at and the CO' norm of the ai; such that
ul22,a SC(1, A) |Au + dullo.2,0 (3,1,3,6)
for all u€ H-(R) such that -y du/dUA = 0 on I' and all 1 > 0.

Proof We first apply identity (3,1,1,1) to

V= A Vu
where A is the matrix of the a We observe that
Au = div v in A
† We recall that o is the ellipticity constant which occurs in (3,1,3,1).
 3.1 CURVATURE OF THE BOUNDARY 145

and that
ди
Y
• AVA § wantduis= (yw) - y on r.
Accordingly, we have

(3,1,3,7)
since O is assumed to be convex.
We then use the following lemma, whose proof is postponed to the
completion of the proof of Theorem 3.1.3.3 (cf. also Lemma 7.1, p. 152
in Ladyzhenskaia and Ural'ceva (1968)).

Lemma 3.1.3.4 The following inequality holds for all u € H? (1):

aZu du (3,1,3,8)
i, =1 lax; ax; i.i,k,1=1

a.e. in D.
From (3,1,3,8) it follows that
ди да. ди
ax; ax; 4=1 ax, ax;
ae. in M. Integrating, we have

where M is a common bound for the CO. norms of all the a;. This,
together with inequality (3,1,3,7), implies:

a 2, la 1lax;
23u 12 dx
ax;

§ In dx
S HAul dxt:a?. 1 22и
i l• list ax
• lax; Ax;
146 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Thus we get

a2ii=1
5 J0lax; дх; •dx.
зах;
(3,1,3,9)
The estimate of the remaining terms in the norm of u in H7(9) is the
classical one. Indeed, we have

(Au + ru)u dx =i.i=1

L
and consequently

i=1
It follows that

(3,1,3,10)

and that
5. "Diull? <
1 (3,1,3,11)
i=1 at Au + ault.
Adding inequalities (3,1,3,9) to (3,1,3,11), we obtain (3,1,3,6). Indeed,
we have

4 n?MA
al a4 • Au + dull?
11 1 4n?MA
9)X
Proof of Lemma 3.1.3.4 By density, it is enough to prove inequality
(3,1,3,8) for u € C'(1). At a particular xed point x, let 11, . .., in be the
eigenvalues of the matrix whose entries are the a;- Also, let Y 1, - .., Yn be
a new system of orthogonal axes which diagonalize this matrix. The
inequality (3,1,3,8) is equivalent to

a2 4=1
51lay;
34dy;13 1 23и 12

This is evident since all the eigenvalues are §-a.

fi
 3.2 PROBLEMS IN CONVEX DOMAINS 147

3.2 Boundary value problems in convex domains

For some of the boundary value problems introduced at the beginning of
Chapter 2, we now have two kinds of results. First an existence and
uniqueness result for a solution in H7(1) provided M is bounded and has
a C'. boundary (see Section 2.4 mainly). On the other hand, we proved
(in Section 3.1) a priori bounds for solutions in H'(8), where the
constants depend very weakly on M provided it is convex and has a C2
boundary. In most of the inequalities the constants do not depend on the
curvature of I, i.e. on the fact that I is C1,1. This will allow us to take
limits with respect to S, i.e. to let & vary among convex domains. Thus
we shall extend our previous results to general bounded convex domains.
The first result of this kind is due to Kadlec (1964) and concerns the
Dirichlet problem. The extension of this result to other boundary condi-
tions has been achieved in Grisvard and looss (1975).

3.2.1 Linear boundary conditions

The possibility of approximating a general convex domain by domains
with C2 boundaries follows easily from the results in Eggleston (1958).
Lemma 3.2.1.1 Let S be a convex, bounded and open subset of R". Then
for every & >0, there exist two convex open subsets ?, and Sez in R" such
that

(6) S. has a C2 boundary I;, i = 1, 2.

(c) d(Is, 12) SE,
where d(Th, Tz) denotes the distance from I, to Iz.
This lemma allows us to approximate a given & either from the inside
or from the outside by a domain with a smoother boundary. The inside
approximation is more convenient for studying the Dirichlet boundary
condition while the outside approximation is more suitable for dealing
with boundary conditions of the Neumann type. By the way, we recall
that we already proved in Section 1.2 that a bounded convex open subset
of R° always has a Lipschitz boundary.
In the following results A denotes the same operator as in 3.1.3,
fulfilling the assumption (3,1,3,1).

Theorem 3.2.1.2 Let O be a convex, bounded and open subset of R"

Then for each f€ Le(9), there exists a unique u € H (S), the solution of
{Au =f in O (3,2,1,1)
? yu =0 on T.
148 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Proof We choose a sequence Sm m = 1, 2, ... of convex open subsets of

R" with C? boundaries Im such that Sm S 8 and dIm: ms I') tends to zero
as m -> +0. We consider the solution Um € H (Sm) of the Dirichlet
problem in Rms i.e.
{Aum=f in Am (3,2,1,2)
on Ims
where Ym denotes the trace operator on I,ms m = 1, 2, .... Such a solution
Um exists by Theorem 2.2.2.3.
It follows from Theorem 1.5.1.5 that um € H'(Om); in other words, we
have um € H'(R"). Then from Theorem 3.1.3.1 we know that there exists
a constant C such that
umi2.2.0m SC. (3,2, 1,3)
This implies that um is a bounded sequence in H'(R"), and in addition
that
Um.i; = (DiD;um), m = 1, 2, ...
are bounded sequences in L2 (R") for 1 € i, j&n. Consequently there exist
UH'(R) and Vi.; € Lz(R") and a suitable increasing sequence of integers
Mks k = 1, 2, ... such that
weakly in H'(R"), k->∞
(Um.i,; -> Vis weakly in Iz(R"), K->∞.
First, we shall check that the restriction u of U to & is solution of the
Dirichlet problem in M. Indeed we have u € H'(1). In addition, all the um
have their support in M; it follows that U also has its support in R, i.e.
U=w. By Definition 1.3.2.5, this means that u€ #' (S) and finally
Corollary 1.5.1.5 implies that yu = 0 on I (here the Corollary is applied
with k=0, which is possible owing to Corollary 1.2.2.3). Finally, let
4 € D (R); then there exists k() such that the support of ¢ is contained
in Sm for all k > k(4). Thus for k > k(4) we have

dImk i,j=1 Semk

Taking the limit in k, we obtain

fie dx =-E anD,uD,p dx.

i,=1
 3.2 PROBLEMS IN CONVEX DOMAINS 149

This identity is valid for all 4 € D(R); it means that

Au = f in 2
in the sense of distributions.
So far we proved the existence of u € H'(M), the solution of (3,2,1, 1).
The uniqueness of u is a classical result by the energy method (see for
instance Necas (1967)). To complete the proof we have to check that the
second derivatives of u are square integrable. We again let q belong to
D(2). Then for k > k(4), we have

Qmk
Taking the limit in k, we get

In other words, the distributional derivative DiD;u is the restriction of Vi;

to M; this is a square integrable function for all i, j= 1, ...,n. S

Theorem 3.2.1.3 Let I be a convex, bounded and open subset of R".

Then for each f€ L2(2) and for each 1 > 0 there exists a unique u € H2(1)
which is the solution of

- (3,2,1,4)
for all v є H'(R).

Identity (3,2,1,4) is the weak form of the Neumann problem for the
equation
Au + lu =f in 12. (3,2,1,5)
As we saw in the proof of Theorem 2.2.2.5, identity (3,2,1,4) is equival-
ent to equation (3,2,1,5) together with the boundary condition

§ viv (a;D;u) =0 a.e. on I. (3,2,1,6)

This makes sense since I is Lipschitz and a;Diu€ H' (8).

Proof This time, we choose a sequence Rms m = 1, 2,..., of bounded

convex open subsets of pM with C2 boundaries Im such that 1 § m and
d(Г.
mo F') tends to zero as m ->∞. We consider the solution um € H7(8,
150 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

of the Neumann problem in Sm no 1.0.

(Aum +1Um=f in Am
(3.2.1.7)
dum =0 on I'm
Im avA
Obviously f€ Lz(Sm) and um exists by Corollary 2.2.2.6.
From Theorem 3.1.3.3, we know that there exists a constant C such
that
(3.2.1.8)
Consequently, restricting the Um to & we obtain a bounded sequence in
H?(9) and then a weakly convergent subsequence. In other words, there
exists an increasing sequence of integers m and a function u € H7(1)
such that
weakly in H7(9), as k->∞.
We now complete the proof by checking that u is a solution of
(3,2,1,4). Indeed let v€ H (M). Since F is Lipschitz, we can apply
Theorem 1.4.3.1 to find a V€ H'(R) such that Vle = v. It is clear that
VIsm , EH' (Rm) and from (3,2,1,7) we deduce that

um.Vdx=I fo dx. (3,2,1,9)

it=1 deems dAmk
We shall now consider the limit of (3,2,1,9) when k ->∞. We have first

Um.Vdx-wo.dr=
and consequently

1Amk

4 - Ullo. 2.02|| Vllo. 2.02.

The right-hand side of this inequality converges to zero due to (3,2,1,8)

and the compactness of the injection of H=(8) in L2(9) (see Theorem
1.4.3.2). In the same way, we prove that

dAmk
owing to the compactness of the injection of H2(8) in H'(R). Summing
up we obtain identity (3,2,1,4) as the limit of (3,2,1,9) when k -›∞. •
 3.2 PROBLEMS IN CONVEX DOMAINS 151

Remark 3.2.1.4 One can prove results similar to those of Theorems

3.2.1.2 and 3.2.1.3 when M is a plane bounded domain with Lipschitz and
piecewise C? boundary whose angles are all convex.

3.2.2 Nonlinear boundary conditions (review)

In the next subsection we shall take advantage of inequality (3,1,2,6),
which concerns the nonlinear boundary condition.
du
- ydu
== B (yu) on Г
where B is a uniformly Lipschitz continuous and nondecreasing function
such that B (0) =0. We shall take limits with respect to & and B. In this
subsection we review some known results about monotone operators.
Here we follow Brezis (1971).
Let H be a Hilbert space and A a mapping from H into the family of all
subsets of H. In other words, A is a (possibly multivalued) mapping from
D(A) = {x€H; Ax+0}
into H. A is said to be monotone if
(91-12;81-42)30,
VxI, X2ED(A) and y, € Axi, YzE Ax. Then A is said to be maximal
monotone if it is maximal in the sense of inclusions of graphs; i.e., it
admits no proper monotone extension. For each 1 >0 we define an
inverse for the multivalued mapping (1A + I) as follows:

(A +D y =(XED(A) !
It turns out that (1A + 1)-1 is univalued and is a contraction in H
provided A is monotone. It was shown by Minty (1962) that A is
maximal if and only if (1A + I) is onto for 1 >0, or equivalently
(XA + I) 1 is defined everywhere.
In what follows we shall only consider monotone operators which are in
some sense the gradient of a convex function. More precisely, let o be a
convex lower semicontinuous function from H into 1-00 + ∞]. We assume
that • is proper, i.e. that 4 # +∞. Let
D(p) ={x€H4(x) <+00}.
For X € D(p) the set

д4 (x) ={yEH|4 (2) -4 (x) = (y; Z-X), DZED (4)}

is called the subdifferential of & at x. It was shown by Minty (1964) that

the operator * -> d. (x) is maximal monotone.
152 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Following Moreau (1965), such a convex function can be approximated

by smooth convex functions 4 defined for 1 > 0, by

-14-23+60(2).. (3,2,2,1)

It turns out that 4 is convex and Frechet differentiable. Thus

2Ф, (x) ={$< (X)}
for every * € H. In addition 4 (x) is a decreasing function of 1 and
41 (X) -> Ф (x)
for every x € H as 1 -> 0. Finally

Ф{(x) == (3,2,2,2)

This is simply the so-called Yosida approximation of A = ., which is

monotone and Lipschitz continuous with Lipschitz constant 1/1.
Not all maximal monotone operators are subdifferentials of convex
functions; however, in the particular case when H is just the real line R,
this does hold. Here are two typical examples of subdifferentials of
convex functions in R. First, if we assume that
+00 x# 0
O (x) = (o x=0
then it is easy to check that
X#0
24 (x)=
x=0.
This is a maximal monotone operator in R. The Moreau approximation of
o is
x2
Pa (X) =
21
and accordingly the Yosida approximation of do is
X
Ф^(x) =

On the other hand, let

X<0
x=0;
 3.2 PROBLEMS IN CONVEX DOMAINS 153

then it is easy to check that

X<0
дф (x) =<[-00, 01 x=0
x>0
(x2 x<0
Qi (x) = (21'
0, x=0

Ф< (x) =
* £0
0, x≥0.
Turning back to the general case, an important existence result is the
following:

Lemma 3.2.2.1 Let 4 be a convex, lower semi-continuous and proper

function on H. Assume that o is coercive, i.e. that
q (x) -> +00 when xll -> +00. (3,2,2,3)
Then • has a minimum in H. The minimum is unique when o is strictly
convex.

Accordingly, if x is such a minimum, we have

ОЕдф (Хо).
This is an existence result for the subdifferential of 4.

Proof of Lemma 3.2.2.1 We denote by m the g.l.b. of . There exists a

sequence X, n = 1, 2, ... of elements of H such that
4 (Xn) -> т
when n -> +∞. Since 4 is proper, we have m < + and condition (3,2,2,3)
implies that the sequence *n h =1, 2, ... is bounded in H.
Consequently, by possibly replacing the original sequence by a suitable
subsequence, we can assume that X n = 1, 2, ... is weakly convergent to
some limit x € H. By the very definition of m we have
4 (x) ≥m.
On the other hand, since 4 is lower semi-continuous (for either the strong
154 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

topology or the weak topology on H), we have

q(x) Slim sup 4 (Kn) &m.
n->00

Summing up we have proved that (x) = m and x is the desired

minimum.
The uniqueness of the minimum, when & is strictly convex is
obvious.
We shall use this existence result as follows. We consider a maximal
operator B in R and the corresponding convex function j on R such that
ß = dj.
Then we build a new convex function on L2(8) by setting

Ф (0) = (3,2,2,4)
if v€ H'(S) and j(yu) E L, (Г)
(+00 otherwise
We are looking for solutions of the following boundary value problem
where c > 0, feLz(8) and S is, say, a bounded open subset of R" with a
Lipschitz boundary T':
1-Au + cu =f in h
a.e. on T (3,2,2,5)

The function & allows a weak formulation for problem (3,2,2,5). Indeed
we have this lemma.

Lemma 3.2.2.2 Let we H-(M) be i solution of (3,2,2,5), then we have

(3,2,2,6)

for all v € L2(1).

Proof The boundary condition in (3,2,2,5) implies that for v € H'(S)

such that j(y) € L¡(T) we have
i(yv)-j(yu)≥-(y Vu)•w(yv -yu)
and consequently that

(i(yo)-¡(ya) doz- (wy Wu) (vo-yu) do.

 3.2 PROBLEMS IN CONVEX DOMAINS 155

Then by the Green formula, we have

Mo-W)dx=[(-Au+culo-w)dx

and consequently

1, 40-wares 4, 84903-16940) do + 1, 54-740-40 da

+c u(v-u) dx.
Finally, observing that
2u(0-4) =02-42
and that
2Vu - V(0 -u) =/Vol2-IVulZ
we conclude that

31 Iva? dx +
This is exactly inequality (3,2,2,6) when v € H'(2) and j(yu) E L, (F'). In
the other cases, we have q(w) = +0 and inequality (3,2,2,6) is
obvious. S
In other words, we have reduced the problem of solving (3,2,2,5) to
that of minimizing the function

(3,2,2,7)
This is easily achieved, owing to the following lemma.
Lemma 3.2.2.3 For c > 0 and feLz(8) the function (3,2,2,7) is convex,
lower semicontinuous, proper and coercive on Iz(8), provided B (0) 3 0.

Proof All but the coerciveness is obvious. To prove (3,2,2,3), we ob-

serve that j(x) ≥0 everywhere. Indeed, the condition that dj(0) = B (0) 30
156 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

means that the graph of j is contained in the upper half-plane. Then it

follows that

This lower bound obviously tends to +0 when fu. •

3.2.3 Nonlinear boundary conditions (continued)
Let us first state the result which is the purpose of this subsection.

Theorem 3.2.3.1 Let N be a bounded convex open subset of R". Let B be a

maximal monotone operator on R such that B(0) a0. Then for each
feI'(N) and for each c >0, there exists a unique u eH? (R) which is the
solution of (3,2,2,5).

Before proving this theorem, let us take a look at some examples. Let
us assume first that
x# 0
¡(x) {+00
10 x=0.
Then obviously we have

= if veH'(5)
@ (w)
(+00 otherwise.
Then, surprisingly enough, problem (3,2,2,5) is just a Dirichlet problem.
Indeed, the boundary condition means that ae. on I,
(yu, -w-yVu)
is a point of R?, which actually lies on the vertical axis. In other words,
yu = 0 a.e. on I With this special choice of j, Theorem 3.2.3.1 is just a
particular case of Theorem 3.2.1.2.
Let us assume now that

i(x) =-10+00 x<0

x≥0.
Then we have

if veH'(2) and yo >0 a.e. on I

Q(w) Ill of de
(+00 otherwise.
 3.2 PROBLEMS IN CONVEX DOMAINS 157

The boundary condition in (3,2,2,5) means that a.. on [

(yu,-v.y Vu)E G
where G is the graph of B = dj, i.e.
G= {(X, y) ER31 x=0, y 50, x-y = 03.
Accordingly we have
yu >0, v•yVu≥0, (yu) (wy Vu) = 0 (3,2,3,1)
a.e. on I' and this is the famous Signorini boundary condition.
Finally let us observe that if we assume that

¡(x) =6=3
where b ≥0, then

if o€ H'(R)
4 (0)=
(+00 otherwise.
Accordingly, we have (x) = j'(x) = bx and the boundary condition in
(3,2,2,5) is just
-w• (y Wu) =byu a.e. on Г.
In particular, when b = 0, this is a Neumann boundary condition and we
have a particular case of Theorem 3.2.1.3.
Before proving Theorem 3.2.3.1, we need some preliminary results on
the approximation of 2.

Lemma 3.2.3.2 Let & be a convex, bounded and open subset of R" and
let Rms m = 1, 2, ... be a sequence of convex, bounded and open subsets of
R" such that I S Sms Sm has a C2 boundary I'm and
d(Г.is F') -> O when m ->0.
Then, for large enough m, there exists a finite number of open subsets Vks
k = 1,2, ..., K in R" with the following properties:
(a) For each k there exist new coordinates {yk, ..., yks in which Vk is the
hypercube
Isien'
(b) For each k there exist Lipschitz functions o* and win defined in
158 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

and such that

6041244.404.241.an
in for every zke Vik

rnVk=lyk=12k.yk)lyk=6*(=k)3
ImhVk=Lyk=/2k.yklyk=ok(=k)}
(c) k=1

In addition -ok and -win are convex functions, in is of class C? for

all large enough m and
(d) I'm converges uniformly to ok on Wk and there exists I such that
1sksK.
(3,2,3,2
Finally, Wok

It follows from Corollary 1.2.2.3 that & and all the m have Lipschitz
boundaries. Accordingly, properties (a) and (b) just refer to the corres-
ponding properties in Definition 1.2.1.1. We actually just have to check
that property (d) holds.

Proof of property (d) in Lemma 3.2.3.2 Since the distance from I' to Im
converges to zero when m->∞, it follows that the distance from the
graph of ok to the graph of oh converges to zero. This means that oh
converges uniformly to ok.
Inequality (3,2,3,2) follows from the geometry. Indeed let us consider a
fixed k and a fixed m. Let yk = (zk, si(zk)) be a point on I'm M Vk, with
-astEsyksak-e, 15jen -1
for a given & >0. The set

is included in Mm. So is the line segment from yk to any point of as. The
slope of such a line has a modulus less than or equal to a*/&. This implies
that

We conclude by replacing all the a, I jen-1 by a"-& with an

 3.2 PROBLEMS IN CONVEX DOMAINS 15

ah

12, 12
.th
P'k

-ah?

Figure 3.1

€ >0 small enough to preserve the condition I'< UK=1 V. Since

d(Г'ms F') -> O as m -> +0, the condition I,"CUK§=1 Vk is also preserved
for m large enough. We can define L as follows:
K
L= max
k=1 an/s.

Let us now complete the proof by looking at the convergence of V4 kn

to Vs*. Since ok is Lipschitz continuous, it has a gradient a.e. in Vk. Let
us consider such a point z€ Vk such that V¢* (z) exists. The tangent
hyperplane at (z, 4 h(z)) to the graph of oh , is above the graph of o*,
since ok and 4m are concave functions and ok in =4*. In other words, we

have ph (2) +94# (2) - (5-7) 74 * (5)

for all g€ Vk. Since ok has a gradient at z in the usual sense, it follows
that
160 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

for all §€ Vi. Then for each j= 1, 2, ..., n- 1, we have

cm (z) + tD,4m (z) 74* (z) + ID; 4 k (z) + 0(141)
for it small enough. If we denote by a; and B; the limits
a; = lim inf D;4mm (z),
M ->ю B; = lim sup D,4 h (z),
m ->do

we easily see that

a, = Dick (z) +0(1), B. <D,4* (z) +0(1).
This shows that
D.ok (z) -> D.ok(z)
when m ->∞ and completes the proof. •
Proof of Theorem 3.2.3.1 We shall approximate & by a sequence of
convex open subsets Oms m = 1,2, ... of R" as in Lemma 3.2.3.2. We
shall also approximate j by its Moreau approximation in, or equivalently B
by its Yosida approximation B. = i. Thus we start from Us,m € La(Sm),
which minimizes the functional

Sm
if v€ H°(Sm)
+00 otherwise.

We observe that since j½ is uniformly Lipschitz, its primitive in does not

grow faster than a quadratic function. Accordingly, when v € H'(Sm), we
have is (YmU)EL./Imг).
There are four main steps in the proof.
1st stepWe check that Ux,m € H?(Rm).
2nd step We prove that Us,m |2.2,0m remains bounded uniformly in m
and 1.
3rd step We take the limit in m.
4th step We take the limit in 1.
In the first step we use the fact that Us,m is the solution of

in Sm
dux.m =Bx (YmUlx," m) on I'm
7-Ym Jum
Indeed, we first observe that since Am.a (Us,m) < +00, Us,m must belong to
 3.2 PROBLEMS IN CONVEX DOMAINS 161

H'(Om). Thus it is the minimum for Um, on H'(Sm). It is easily checked

that Am,) is Frechet differentiable on H'(m) and consequently we have

I'm.a (4x, m) =0.

In other words we have

wmodxtBa (YmUm)yo.dom
Som dOom I'm
- fodx =0 (3,2,3,3)
dOm
for all v € H'(Rm). Making use of (3,2,3,3) with only v € H' (Sm), we
readily see that
-Aux,m + cUx,m=f in Rm:
Then applying the Green formula (1,5,3,10), we rewrite (3,2,3,3) as
follows:

for all YmU € H'(Im). This implies the boundary condition on Ux.ms i.e.
-Im AUx, m/dUm = Ba (YmUx,m) in the sense of H-112(5, n).
Let us now consider Ba (yun,m) as the Neumann data for Us,m• Since
Ux.m EH'(Rm), we have YmUs,m € H'12(Im). Then, taking advantage of the
fact that B is uniformly Lipschitz continuous, we conclude that
Bx (VmUlx.m) EH'(Im).
Now it follows from Corollary 2.2.2.6 that us,m € H? (Rm) since Im is C?.
The second step is just an application of Theorem 3.1.2.3. This theorem
can be used here
is uniformly becausenon-decreasing
Lipschitz, Sm , is convexand
andfulfils
has athe
C2condition
boundary, while B

Ba (0) = 0.
Indeed, we have - B, (0) = (1/1) (1B+1) '0 and (1B +1) 10=0 since
0 € (1ß + 1)(0). Thus, we have

(3,2,3,4)
Ux.ms m = 1, 2, ... is consequently a bounded sequence in H7(1). By
possibly considering a suitable subsequence we can therefore assume that
there exists us € H (8) such that
Un,m -> Un
weakly in H7(2) when m-> +00.
162 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

In the third step we take the limit in m in the equation which expresses
that Um.) has a minimum at us,m: Actually we have

10 Team

(3,2,3,5)
for all Ve H'(R"). It is easy to check that

Then obviously we have

lim inf

since Rm 32. Also, we have

fum. a dx-> fis dx.

Thus the only difficult point in taking the limit in (3,2,3,5), is to prove
that
(3,2,3,6)

(3,2,3,7)

For this purpose, we fix a partition of unity on [' and I'm corresponding to
the covering Vk 1-k K, introduced in Lemma 3.2.3.2, i.e. we consider
Or E D(R"), 1 5k SK, such that Ok has its support in Vk and

1=&k =10. (x)

for all x € r and all x € r.m (for m large enough). We have to prove that

[Oxis(7mV)dom-> Oria(WV)do. (3,2,3,8)

 3.2 PROBLEMS IN CONVEX DOMAINS 163

We drop the index k and set m = Oj. (V). It follows that

Clearly we have
(Vmn) (2, 4m (2)) [1+|74m (2)137112-> (yn) (2, 4 (2)) [17|74 (2)127112
a.e. in V'. In addition, we have
(ymn)(2,4m (2)|[17|74m (2)1271/2

=#1+13713 (m) (2, 4(2)) 174m(2)-462) 12

Thus we have a fixed square integrable bound, and applying Lebesgue's

dominated convergence theorem, we conclude that (3,2,3,8) holds.
To prove (3,2,3,7) we introduce Un € H? (R") such that Uls = Us.
Then, we have
m) dom - is (yu,) do

It follows from (3,2,3,6) that

'I'm
Then we observe that for x, y eR we have
is (x) -is (y) > Bi (y) (x-y)
and thus

(3,2,3,9)
We shall show that the right-hand side of this inequality converges to
zero. Indeed, we have

I'm Sm
164 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

with a constant K which does not depend on m, due to Lemma 3.2.3.3

below and Theorem 1.5.1.10.

Lemma 3.2.3.3 Under the assumptions of Lemma 3.2.3.2, there exists a

Lipschitz vector field u defined on Rh and a constant 8>0 such that
"= 8 on I m for all m.

We postpone the proof of this lemma until the proof of Theorem

3.2.3.1 is completed.
This shows that |B. (Vm Ux)llo,2.rm, remains bounded when m ->∞. Then
let us set

Mm= Ok (Us,m - Ux).

Use local coordinates and drop k. We have

We write

Da Mm (Z, y) dy
Mm 12,4m (2)) = Am (2, 40(2)) +p(z)
('912)

and consequently

IMm (Z, Pm (z))| S|Mm (Z, 4 (z))|

up (z)

Thus we have

+ [1+1231/4 max [4m (2) - 4 (2) 112

*I finles, doGs
 3.2 PROBLEMS IN CONVEX DOMAINS 165

In other words

+ max [4m (2) -4 (2) 7112110(41.m - Ux)ll.2.0m

This shows that

when m -> ∞ since us,m -> Us in H'(8) and Us.ml1.2.0m remains bounded.
Summing up we have proved that

and remembering (3,3,2,9), this implies (3,2,3,7).

Taking the infimum limit in m, of inequality (3,2,3,5), we eventually

obtain da (Us) 54% (0) (3,2,3,10)

for all o€ H'(R), where is is defined by

44 (0) -11 TOPaK+S

In addition, taking the limit in (3,2,3,4) we also have

(3,2,3,11)

We can now perform the last step of the proof. Due to (3,2,3,11) we
can find a sequence A; i = 1, 2, ... such that
1; -> 0, i-> 00
and there exists u€ H?(M) such that
i-> 0
weakly in H7(8) and consequently strongly in H2-(8) for & >0. In
addition, by the Lebesgue theorem on subsequences, we can also assume
166 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

that
yus, -> YU, ¡ -> 0
i-> 0
a.e. on F.
On the other hand, as in the first step of the proof, inequality (3,2,3,10)
implies that
{- Aux + cUs =f in D
(3,2,3,12)
a.e. on F.
It is easy to take the limit in this equation. We thus obviously obtain
-Au + cu = fin 2.
To take the limit in the boundary condition, we use the following trick.

Lemma 3.2.3.4 Let B be a maximal monotone operator in R and let B, be

its Yosida approximation. Let Xm Ima Am be three sequences of real numbers
such that
Xm -> X, Im -> y
Am -> O
(y,m = Brn (Xm).

Then y €B (x).
We consequently obtain
-v• y VuEB(vu)
ae. on F. Summing up, u € H'(S) is the solution of problem (3,2,2,5).
The uniqueness of u follows from Lemma 3.2.2.2 since the functional
which is minimized there is strictly convex (see also Lemma 3.2.2.3). This
completes the proof of Theorem 3.2.3.1. -

Proof of Lemma 3.2.3.3 We fix a point x€ and a ball B of radius

p > 0 and centre at x such that B c 8. Then we fix a function A € D(RM)
such that 0 = 1 on all m for large enough m. Then we can define u by
Me (X) = A (x) (X-X0).

It is clear that the angle of y with Um is less than or equal to

{э
12- arcsinix-x0
 3.2 PROBLEMS IN CONVEX DOMAINS 167

p
at x € I'm. Consequently we have

M•Um

Thus a 8>0 obviously exists and on the other hand, it is easy to check
that u is Lipschitz continuous. &

Proof of Lemma 3.2.3.4 Using the definition of B it is easy to check

that
Im EB(Xm -AmYm).
Since Ym -> y and *m -Am\m -> X, it follows that
y€B(x)
since the graph of B is closed (by maximality). A

3.2.4 Oblique boundary conditions

Here, for the sake of simplicity, we shall restrict our purpose to boundary
value problems in a plane domain 2. Thus let us assume that 2 is a
bounded convex open subset of R?; its boundary is a closed Lipschitz
curve I along which the arc-length s is well defined. We assume, in
addition, that c is a given Lipschitz function in M. We shall solve the
following problem: for a given f€ L2(8) and 1 > 0, find u € H3(8) which
is a solution of
in 2
ди= C д on Г (3,2,4,1)
as yU
The main result below is due to Moussaoui (1974).

Theorem 3.2.4.1 Let S be a bounded convex open subset of R3 and let

cE C' ' (R). Then there exists to such that for each 1 > to and for each
feL2(1) there exists a unique u € H (S) which is a solution of

(3,2,4,2)

for all v€ H' (8).

Of course, identity (3,2,4,2) is a weak form for problem (3,2,4,1).

Indeed, applying the Green formula of Theorem 1.5.3.1, it is easy to
check that for u€ H (0), (3,2,4,1) and (3,2,4,2) are equivalent.
168 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Exactly as we did in the proof of Theorem 3.2.3.1, we shall approxi-

mate & by a decreasing sequence of smoother convex domains Rms
m = 1,2, ... (see Lemma 3.2.3.2). In each of these domains, we shall
solve a problem similar to (3,2,4,1). Then it will be possible to take the
limit in m with the help of an a priori estimate. Let us first prove this
estimate.
Theorem 3.2.4.2 Let 1 be a bounded convex open subset of D2 with a C2
boundary. Let c€ CO' (5) and ME CO. (S)? be such that u•vas >0
everywhere on I; then there exists No and k such that
ul2.2.nsk-Au+dullo.2.n (3,2,4,3)
for all u € H'(R), such that -y dulav = c d(ru) as on I. Moreover, k and No
depend only on 8, the Lipschitz norm of u and max lac/as|.

Note that the existence of u follows from Lemma 1.5.1.9.

Proof We apply again identity (3,1,1,1) to v = Wu. The two-dimensional
version of this identity and the convexity of S, imply that‡

In other words, we have

i,= 1
Due to the boundary condition, we have

Denoting by M an upper bound for lac/as on I, we have

1. =1 da lax; dx;
(3,2,4,4)
Besides this, we have as usual

• Ju- yu dx.
It follows from the boundary condition that
ди
- vU yu ds =-
dr ds
† For any vector field a, a, and am are the normal and tangential components of a on F.
 3.2 PROBLEMS IN CONVEX DOMAINS 169

and consequently

owing to inequality (1,5,1,2) of Lemma 1.5.1.10. Choosing & small

enough (e.g. such that Mk V& ≤½ and 1, ≥ (MK/&) + 1, we finally obtain
(3,2,4,5)
On the other hand, we have

consequently it follows from inequality (3,2,4,4) that

lax; Ax;
MK
- Vult
lax; ax; I VE
by a new application of Theorem 1.5.1.10. Then we have

ax; дх; (3,2,4,6)

where I = MKIe, MK Va =1.
Combining inequalities (3,2,4,5) and (3,2,4,6), we plainly obtain the
desired result.
Remark 3.2.4.3 The same proof can be worked out, without the
convexity assumption on S; of course in that case, the constants k and No
will also depend on the curvature of I.

Remark 3.2.4.4 We can combine the two kinds of computations that

we did in the proof of Theorems 3.1.2.3 and 3.2.4.2 to deal with the more
complicated boundary condition
ди d
-Yas
170 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

assuming that ß is a nondecreasing Lipschitz function. Unfortunately, we

shall not be able to take advantage of such a result in what follows.
We now turn to the

Proof of Theorem 3.2.4.1 We use a sequence of plane convex domains

Rms m = 1,2, ..., as in Lemma 3.2.3.2. We extend c to the whole plane
in such a fashion that
CE CO.' (13).

Consequently c is defined on Im. Then, in each Sm me We consider um €

H°(Rm), a solution of

(-Aum +14m =§ in Sm
a
dum = cdsm- YmUm on Im. (3,2,4,7)

The existence of such a function Um follows from Theorem 2.4.2.7 for

1 > 0. We shall now use inequality (3,2,4,3) to show that Um|2.2,0 remains
bounded when m -> 0.
Here, we take advantage of Lemma 3.2.3.3 again. Obviously
maxrm ac/ Sm remains bounded uniformly in m. This implies the exis-
tence of constants k and No, both independent of m, such that
Um/2.2.0m€kfll0,2..m (3,2,4,8)
provided ^ > No.
Now we proceed as in the proof of Theorem 3.2.3.1. We first observe
that (3,2,4,7) implies this:

ImUm Ym V dSm

(3,2,4,9)
').
for all V€HQ
On the other hand by inequality (3,2,4,8), there exists a subsequence of
the sequence Um m = 1, 2,... which is weakly convergent in H(1) to
some u € H'(8). Let us again denote this subsequence Ums m = 1, 2, ..
for the sake of avoiding further complications in the notation. We shall
show that u is the desired solution of (3,2,4,2); this will be achieved by
taking the limit in identity (3,2,4,9).
First let Ve C'(R?) be fixed. We shall set v = Vlo. Exactly as in the
proof of Theorem 3.2.1.3 we show that

rum-TVdx-Pu.Pudx (3,2,4,10)
 3.2 PROBLEMS IN CONVEX DOMAINS 171

and that

(3,2,4,11)

To complete the proof, we shall show that

д
Casm
I'm
(3,2,4,12)

This requires much more care. First we fix a partition of unity (Oks
1<k<K) on l' and Г n corresponding to the covering Vk 15k < K in
Lemma 3.2.3.2. The limit in (3,2,4, 12) will follow by adding these limits:

(3,2,4,13)
We shall use the local coordinates of Lemma 3.2.3.2 in each V. Let us
first introduce some auxiliary notation:
Wm =VUms n =OncV.
From now on, we shall drop the subscript k to simplify the notation. In
addition, we consider a function U€ H?(R) such that Ula = u (see
Theorem 1.4.3.1) we shall consider separately

and

I'm
where W= VU.
According to the notation in Lemma 3.2.3.2, we have (after dropping
the k):

nlYm (Wm-WHen dsm Im (Wm-W|dsm

XVIto'm(2)3)]dz
172 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

where a does not depend on m. Owing to (3,2,3,2) we have

sblly@Vum-Tulllo.2.rt a V(1+13)

Ju. looks

This implies that

(3,2,4, 14)

since U.1 -> u weakly in H'(S) and Am -> 4 uniformly in V

Next we have

n(YmWIm'
I'm

+ 4'm (2) ImD2 U(2, 4m (Z))]dz. (3,2,4,15)

Obviously n(z, «m (z)) converges uniformly to n(z, q(z)) in V'. On the
other hand, we have

[DaW(z, y)| dy
Local
SIm (2) -4(2)|112

and consequentlv yIm W(Z, Om"(z)) converges to yW(z, q (z)) almost

everywhere in V'. Summing up this shows that the integrand in (3,2,4, 15)
has a limit almost everywhere in V'
 3.2 PROBLEMS IN CONVEX DOMAINS 173

In addition, we have
Im W(z, Am (z) |S/yW(2, 4 (2) |+|@m (2) -4 (2) |112

xf fans
-a/2
[D2W(z, y)|? dy'
1112

Thus the integrand in (3,2,4,15) is bounded by a fixed integrable function

on V'. One of Lebesgue's theorems implies, therefore, that

(3,2,4, 16)
Im
Now (3,2,4,13) follows from (3,2,4,14) and (3,2,4,16). Accordingly, we
have shown that (3,2,4,2) holds for all ve C'(R). This identity is ex-
tended to all o€ H'(M) by density. Finally, the uniqueness of u follows
from (3,2,4,2) by substituting u for v and assuming 1 large enough. •

Remark 3.2.4.5 We can make the lower bound No more precise in the
statement of Theorem 3.2.4.1. Indeed, let us set o = u in (3,2,4,2); thus
we obtain

Then let K be the best possible constant in inequality (1,5,1,2) and set

M = maxдс
г as
it follows that

VE.

Choosing Ve =2/KM, we obtain

f1-(KM
accordingly we have

f1-(KM
and this shows that (KM/2)? is a possible value for No.
174 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Remark 3.2.4.6 In all the previous results the convexity of & can be
replaced by the weaker assumptions that there exists a sequence Rms
m = 1, 2, ... of bounded open subsets of R", with C? boundaries Im such
that d(Im F') -> 0 as u -> +0, the sequence of the corresponding second
fundamental forms Pm m = 1, 2,... is uniformly bounded from above
independently of m, and Am S to solve the Dirichlet problem or
Am 21 to solve the other boundary value problems. This assumption is
obviously fulfilled when M is a bounded open subset of 2 whose
boundary is a curvilinear polygon of class C1,1 , provided all the angles are
strictly convex.

3.3 Boundary value problems in domains with turning points

As we observed in Remark 3.2.4.6, the good domains & for the regular-
ity in H7(8) are those which are piecewise C2 with convex corners. This
is an upper bound on the measure of the possible angles and this leads
naturally to the question whether turning points (i.e. angles with measure
zero) allow the solution of an elliptic boundary value problem to belong
to H?(8). The answer is yes for several boundary conditions as it is shown
in Khelif (1978).
We shall consider here only the simplest problem, namely the Dirichlet
problem for the Laplace equation in a plane domain 2 with a boundary I
which is C? everywhere except in one point, which is a turning point. To
be more precise we assume that this turning point is at O and that there
exists p>0 such that, denoting by V the disc with centre at zero and
radius p, we have
VAn= ((K, y) EV; X>0, 41 (8) <y <42(x)}
where 4, and 42 are a pair of C2 functions such that
(4, (0) =42(0) =0
Ф1(0) =¢2(0) = 0
(Pí(x) <42 (x), 0<X < p.
Thus, near the origin, the boundary of 2 is a pair of C2 curves which
meet at zero and which are tangent there to the positive half x-axis.
We are going to prove the following result of Ibuki (1974) applying the
method of Khelif (1978) which is simpler and more general.

Theorem 3.3.1 Given f€ La (1) there exists a unique u€ H7(8) MH' (S)
such that
Au =f
 3.3 DOMAINS WITH TURNING POINTS 175

01 y=42 (x) X
1=4, (x)

*-Am.

Figure 3.2

in 2 provided

x->0
(One can observe that those conditions are fulfilled in the following exam -
ple: 41(x) =0 and 42(x) = x°', where a is any real number >1.)

Exactly as in the previous section the method consists in approximating

s2 by a sequence knms m = 1, 2,..., of 'better' domains. For this purpose
we consider a decreasing sequence am m = 1, 2,.. of positive real
numbers and we set
Am=8M<(k, y); x>am
Clearly, we have
00

m =1
and each Am has a piecewise C2 boundary with two convex angles.
176 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Consequently there exists a unique

Um EH (Rm) MH'(Sm)
which is a solution of
Aum=flom
in Mm. We are going to show that the sequence
Uml2.2.0m
is bounded as m => +00.
Lemma 3.3.2 Under the assumptions of Theorem 3.3.1 there exists a
constant K such that
Um/2.2.0m€K If 0.2.0 (3,3,1)
for every m and every f€ L2(8).
Proof Integration by parts of (Aum)Um and the Poincaré inequality
(Theorem 1.4.3.4) yield a constant C such that
uml1.2.omSCalfllo.2.n. (3,3,2)
Then we apply Theorem 3.1.1.2 in order to bound the second deriva-
tives of Um• n• For this purpose, we assume that m is fixed and we set
W = DjUm (3,3,3)
The functions v and w only belong to H'(m) and we approximate them
by functions belonging to H?(m).
We observe that v and w fulfil the following boundary conditions:
Syw(Amo y) =0, 91(Am) <y<42(Am)
lAyo+wyw=0 on TMi(x, y) | x> amt,
where 1 and u are the components of the unit tangent vector to r.
Accordingly, there exists a couple of sequences of functions Us Wk
k =1,2, ..., such that
Uk -> U, Wk -> W
in H'(2) as k -> +00 and such that

VWk(Ams Y) =0, Q1(am) <y <42(am)

(AyoR + MyWE = 0 on In{(x, y) | x> amt.
(We skip the proof of this density result due to its similarity to Lemma
4.3.1.3.)
 3.3 DOMAINS WITH TURNING POINTS 177

Applying identity (3,1,1,10) to the vector function {Uk, We}, we obtain

Im
Then, taking the limit in k, we have

ID.0+D.,wFdxdy = TID.oP+/D,wF+2D.WD.0]dx dy

and consequently, using (3,3,3), we have

{IfF dxdy = ID3umP+1D3umP?+2|D.D.umP]dxdy

fem

yum
dv do. (3,3,4)
I'm
Let us now consider the boundary integral in (3,3,4). The second
fundamental form 8 vanishes on the segment
<(ams y), 41 (X) <y <42 (x)}
which is curvature free. On the other hand the form B is bounded on
each C2 curve; in addition B is bounded by 4" (x) | at (x, Q; (x)). Thus we
shall consider differently the points of the boundary I'm according to their
distance to the origin. For this purpose let 8>0 be small enough to
ensure that the points (8, 4, (8)) and (8, 42(8)) lie in V. Assuming that m
is large enough to ensure am1 < 8, there exists a constant M2 such that
m12 dum12do
•8

am

(3,3,5)
dam
Since
82s =8M (x, y): x > 8}
has a Lipschitz boundary, Theorem 1.5.1.10 implies the existence of a
178 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

constant Ks such that

I'm Mi(x, y) : x381 In aunt do

te'laf IumP dxdys (3,3,6)

for every & >0.
Next on the graph of Q; (= 1, 2) we have
rum (x, 4; (x))=0 (3,3,7)
and consequently
V (DAUm) (X, 4; (x)) + 4' (x) + (D, Um) (X, 4; (x)) = 0
for a.e. x. On the other hand we have
- P½VD, Um + VD, Um
Yaw V(1 + 433)
at (x, 42(x)). It follows that

Q4-42
and there exists a constant Ns such that

yдит 1
dyl
for am <X&8 (and Ns -> 1 as 8 -> 0).
The boundary integral corresponding to the graph of 42 in (3,3,5) is
bounded by
Q'(x)|dx
1=Naam o' (x) -05(x)12.
(3,3,8)
Since we have
YD.Um (X, Ф1(X)) + 41(X)YD,Um (X, 41 (x)) = 0,
by (3,3,7), we can write
yD.um (x, 42 (X)) + 41 (x) yD,um (X, 42 (x))
• 42 (x)
= D, {Dium + pi(x) D,um} dy.
41 (x)
 3.3 DOMAINS WITH TURNING POINTS 179

Hence
lyD.um (X, 42(x)) + 41 (x) yD, um (x, 42 (X))|

Ф 1 (x)

D?um? dy
So, (a)
Finally this implies that
(14'(x) | [42(8) -41 (X)1112
112N12 max
0<x=8

+ 0<X=8
max lo'(x) (dAm
is postas (3,3,9)

We have a similar inequality for the boundary integral corresponding to

41. Summing up, there exists a constant M2 such that

I'm I ar do

te 12 Tun.P dx dy?

ID,D.um| dx dy
Sm 1825

+[1+='
Sam 185
(3,3, 10)
for every n>0.
Now we can make precise the choice of 8. We choose & small enough
so that

LIP! (x) + 42(x) 42 (x) - 41 (x)]

A = max
(<x=8
= <2.
180 SECOND-ORDER PROBLEMS IN CONVEX DOMAINS

Then we choose n in such a way that again

A(1+n) <2.
Then we can replace 8 by a smaller one to ensure that for this value of n
we have
NsA (1+ n) <2
and
max Ilyi(x)17+|4½(x)131.
NoA (It:n/ 0<x=8
This is clearly possible since 41(0) = $2(0) = 0 and N. -> 1 as S-> 0.
Finally we choose & small enough so that
MaKgE'2<1.
Then the inequality (3,3,10) may be rewritten merely as

dv ' do
dum

sal ID,um? + D, um? +2|D, D, um PI dx dy

Sm

+B MumFdxdy (3,3,11)
JAm
where a <1 and neither a nor & depend on m.
Let us go back to the identity (3,3,4). Together with (3,3,11) it yields

[TID3umP+|D,umP+21D,D,umP]dx.dy

1 "IAPaxdy+B| Mum-dxdy. (3,3,12)

dAm
This last inequality combined with (3,3,2) implies (3,3,1).

Proof of Theorem 3.3.1 This is very similar to the proof of Theorem

3.2.1.2. By the same technique we find an increasing sequence of integers
Mks k=1, 2, ... and a function u such that
Ümk-> U
weakly in H'(8),
DiD;Uma -> DiD;u
 3.3 DOMAINS WITH TURNING POINTS 181

weakly in L2(2) such that

Au =f
in 2.
Obviously u belongs to H7(8) and we just have to check that u
belongs to H'(R). Here N has no Lipschitz boundary and H'(R) must be
understood as the closure of D(1) in H'(1) (according to Definition
1.3.2.2). Indeed Um , belongs to the closure of D(m ) for the norm of
H'(Sm). Therefore Uma , belongs to the closure of D(1) for the norm of
H'(82) since Sm is a subset of 12. By taking the limit in k it is clear that u
belongs to the closure of D(2) in H'(2). _

Remark 3.3.3 In the work of Khelif (1978) conditions on 4, and 42 are

given for the smoothness of the solution of the Laplace equation under
other boundary conditions. For instance the conditions corresponding to
the Neumann problem are

him sup 1 ' (x) (492(8) - 41 (x) <1.

x->0 [Qá(x) -41(x)]
Mixed boundary conditions are also considered.
4
Second-order elliptic
boundary value problems in
polygons
4.1 Foreword
The purpose of this chapter is to investigate the properties of the second
derivatives of the solutions of boundary value problems for the Laplace
operator in a plane domain with a polygonal boundary. Here, we just
consider classical polygons, i.e. the union of a finite number of linear
segments I;, 1 j&N (it is convenient to assume that I, is an open linear
segment). We also fix a partition of {1, 2, - .., N} into two subsets N and
D. The union of the I; with jeD is going to be the part of the boundary
where we consider a Dirichlet boundary condition. We shall consider
first-order boundary conditions (either Neumann or oblique) on the other
sides. Accordingly, our main problem will be the following. Given f€
Lp (8), we look for u € W3(8), a solution of
Au = f in R
YiU = 0 on I; jeD (4,1,1)
д on I; jEN
Vi al
dv.+Biдт;- Vil =0
where v; denotes the unit normal on I, while ; denotes the unit tangent
vector on I, (following the direct orientation; finally B; jEN are given
real numbers).
The first step in solving (4,1,1) is the proof of a priori bounds for
solutions in W5(8). Actually, we shall prove the existence of a constant C
depending on 2, p, D and B; (jEN) such that
ulz.p.aSChAullo.p.st|ullo.p.as (4,1,2)
for all u€ W(S) fulfilling the boundary conditions in (4,1, 1).
Curiously enough the inequality (4,1,2) always holds when p = 2, while
it does not hold for some exceptional values of the numbers B; (jEN)
when p+ 2 (a detailed investigation of some exceptional cases can be
182
 4.1 FOREWORD 183

found in Fabes et al. (1977)). Actually our methods of proof when p = 2

and when p+2 are quite different. Our proof for p = 2 starts from a
particular case of identity (3,1,1,10) which we shall prove directly by
performing integration by parts. On the other hand, when p+ 2, we shall
use a local method which, together with the same change of variables as
in Kondratiev (1967), reduces our problem to a boundary value problem
in an infinite plane strip. The main advantage is that such a strip has a
smooth boundary. There, we essentially use the same techniques as in
Subsection 2.3.2. That is, we write the solution as a double layer potential
which is estimated by applying Mikhlin's multipliers theorem.
Inequalities like (4, 1,2) when p = 2 for the Dirichlet problem have been
proven in Aronszajn (1951) and Hanna and Smith (1967). General
boundary conditions are dealt with in Grisvard (1972).
The second step in solving (4,1,1) is the following. The a priori bound
(4,1,2) implies that the Laplace operator A has a closed range in Lp (8)
when we look at it, as an unbounded operator whose domain is the
subspace of W7(8) defined by the boundary conditions in (4, 1,1). There-
fore, the annihilator of the range is a space of functions in La (S) (where
p'+ q=1) which are, in some weak sense, solutions of the homogene-
ous adjoint problem. Using separation of variables in polar coordinates,
we shall be able to derive precise expansions of those functions near the
corners. Then it will be easy to calculate the codimension of the range of
the Laplace operator in L (1). This is carried out in Section 4.4.
Such results for p+ 2 were first proven by Merigot (1972), who makes
use of quite different methods. A comprehensive detailed study of prob-
lem (4,1,1) has been carried out independently by Lorenzi (1978) and
Moussaoui (1977). Here is a simplified version of their work.
The reader interested only by the p = 2 case may skip Section 4.2 and
Subsection 4.3.2.
Here are some additional notation. We denote by @; the measure (we
allow w; =m in order to consider also mixed problems along a flat
boundary) of the angle at S; and we set
jEN (4,1,3)
Ml;=
i€ g
Accordingly we have Y; (du/du;) = 0 for all j when u fulfils the boundary
conditions in (4,1,1). Finally, we define ; by
(tan Q; = B; jEN (4,1,4)
14, =1112 jED,
thus ©; is the angle of the vectors v; and y;
Finally, let us mention that all the results in this section hold for
184 SECOND-ORDER PROBLEMS IN POLYGONS

domains with holes. Considering domains with holes (or domains which
are not connected) just leads to more complicated notation.

4.2 A priori estimates for a problem in an infinite strip

This whole section is devoted to the proof of bounds for solutions of the
following boundary value problem in the infinite strip B= R× J0, h
(h > 0). We shall denote by x and y the coordinates in R? and thus we
have
B={(x, y) |XER, O<y Sh}.
We shall deal with a boundary value problem for the operator L defined
by
Lu = Diut Diu+ aDzu + bu,
where a and b are real numbers. The boundary conditions involve the
operators
M;u = a;D,u + B;D.u + 1;4, 1=0, 1.
where a;; B; and 1; are real numbers i =0, 1. (Actually, we shall only
need, in the forthcoming sections, these two special cases: either a; = 1 or
a; = B; =0 and 1; = 1.) Precisely, we look at u € WE(B), a solution of
(Lu =f in B
ly:Mu = 0 on Fir j=0,1 (4,2,1)

where Fo = {(x, O) XER}, Fi = {(x, h) xeR} and y; denotes the trace

operator on Fir i = 0, 1.
We shall look for conditions on the coefficients a, b, c, a; Bi, 1;, j=0, 1,
ensuring the existence of a constant C such that
u2. p.B €C I/flo.p,B- (4,2,2)
For that purpose we shall calculate explicitly a solution u of (4,2,1) by a
Fourier transform in x. Then the explicit solution will be suitably esti-
mated by using Mikhlin's theorem.

4.2.1 Explicit solution by Fourier transform and consequences

As in Subsection 2.3.2 we denote by û (respectively f) the partial Fourier
transform of u (respectively f) with respect to x, i.e.
1 • +00
é-ixtu (x, y) dx,
 4.2 ESTIMATES IN AN INFINITE STRIP 185

for #eR and y€J0, h[. Actually, in order to deal only with Fourier
transforms which are functions, we shall always assume that u € H°(B).
Eventually, we shall take advantage of density theorems for extending
our results to the whole of W(B).
After performing the Fourier transform, problem (4,2,1) becomes a
two-point boundary value problem in the interval J0, h[ depending on the
parameter ⅚. Namely this problem is
(a" +(-82+ias+b)û=f in JO, h[
aot' (5,0) + (iBo5 + 10) a (5, 0) =0 (4,2,1,1)
(ast' (5, h) + (iB, 5 + 1,)û (5, h) = 0
for all §el, where the superscript prime denotes differentiation with
respect to y.
As is well known for two-point boundary value problems, problem
(4,2,1,1) is well posed if and only if the only solution of the corresponding
homogeneous boundary value problem is zero. This possibly leads to
exceptional values of 5. More precisely we have the following.
Theorem 4.2.1.1 Assuming b>0, a+ 0, then the problem (4,2,1,1) is
well posed unless
sin ph(apa,p? -BoB. 52 + 1,1, + i[Bo1, + B, 1.15)
= p cos ph (ad,-a,do]+i aoßs-aBol5), (4,2,1,2)
where p=(b+ias-52)1/2.+
The case when b > 0, a + 0 is the only one that we need in the sequel.
Proof A fundamental system of solutions for the differential equation in
(4,2,1,1) is the couple of functions
U(y) = sin py, U2(y) = cos py
where p=(b +ias-53)12, for § ER. Thus any solution of the homogene-
ous equation

has the form

The boundary conditions in problem (4,2,1,1) are fulfilled if and only if

{a014,010) +430210)) +(Bo5 +10) 1810180) +14202(011=0
la, (M, 01 (h) +44202 (h)) +(iB, 5 +1, ) (44,0, (h) +44202 (h)) = 0.
† We define the square root of a complex number by placing the cut on the negative real
axis.
186 SECOND-ORDER PROBLEMS IN POLYGONS

This is a linear system in (Hi, M2) whose only solution is zero, provided its
determinant d is different from zero. Actually we have
d={a00, (0) + (iBoE + 10)0, (0) Ka, 02 (h) + (iBiE + 1, ) 02 (h)}
-{ap02(0) + (iBoE + 10) 02(0) Ka,vi(h) + (iBE + 1, )U, (h)}
= dop{-op sin ph + (iß¡$ + 1,) cos ph}
- (ißo5 + 10 asp cos ph + (iB,$ + 1,) sin ph}
=-sin phia.a,p? + (iBo5 + 10) (iB, 5 + 1 ,))
+ p cos phia, (iß E + 1,) - a, (iBoE + 10)}.
The condition that d = 0 is exactly (4,2,1,2). It is the one that allows the
homogeneous problem corresponding to problem (4,2,1,1) to admit non-
zero solutions.
We observe that condition (4,2,1,2) involves only analytic functions of
§ Thus, it is fulfilled for countably many exceptional values of §. In
addition, due to the asymptotic behaviour, for large §I, of both members
of equation (4,2,1,2), there is only a finite number of exceptional values
on the real axis. From now on we shall assume that problem (4,2,1,1) is
well posed for all ¢ eR. In other words, we assume that equation (4,2,1,2)
has no real root.
Theorem 4.2.1.2 Assuming b>0, at 0 and that $ is not a root of
(4,2,1,2), the solution of problem (4,2,1,1) is

(4,2,1,3)
where, for zzy,

K(E, y, z) =1 Г ¡Bof + No sin py?

and, for zey,

p
X a, cos plz -h) -¡BiE+M sin p(2-h)?

K (5, y, z)=

X an cos pz-iBof + ho sin pz]

where p=(b+ias-52)112 and

8 = [a (iBos + 10) - ao (ißÈ + 1,)] cos ph
+(a4,0, p= + (iBot + 10) (iB, s+ x, 7 sin ph
 4.2 ESTIMATES IN AN INFINITE STRIP 187

Although these formulas are consequences of general procedures for

solving two-point boundary value problems, it is simpler to check them
directly by verifying that (4,2,1,3) actually gives the solution of problem
(4,2,1,1). This is straightforward.
We shall now use identity (4,2,1,3) to show the existence of a constant
Co such that
Mullo... SColLullo.p.B (4,2,1,4)
for all u € W(B)n H (B) which fulfils the boundary conditions in (4,2, 1),
assuming the problem is well posed. We shall need the following lemma.

Lemma 4.2.1.3 Let 5, Y, z+> K(5, y, z) be a smooth function such that

max ch
max {K(E, y, z) +$I D;K(5, y, z)I} dz < +o (4,2, 1,5)
and
oh

2810,41 do max {lK (5, y, z)|+|5||D K(S, y, z)I}dy < +00;

(4,2,1,6)

then the mapping u -> f defined by

is continuous in Lp (B) for p such that 1 < p <00.

Proof Let us denote by M the function

M(y, z) = max {IK (5, y, z)|+|5||D; K(5, y, z)I}.

Applying Mikhlin's theorem (see Theorem 2.3.2.1) we know that there

exists a constant C such that

We conclude by applying a classical result on integral operators (see for

instance Widom (1965). _
Now we have to check the conditions in Lemma 4.2.1.3 on the kernel
defined in Theorem 4.2.1.2. It is easy to derive the following bounds for
K when the problem is well posed (i.e. when (4,2,1,2) has no real root).
Indeed there exists a constant L such that
L
K(E, y, z) S to exple! (y +2-24)
lElIDAK(E, y, z)|S L exp|p| (y +z-2h),
188 SECOND-ORDER PROBLEMS IN POLYGONS

while the asymptotic behaviour of p is like is, when I&l-> o. It follows

that inequalities (4,2,1,5) and (4,2,1,6) hold and consequently we have
proved inequality (4,2,1,4).
Summing up, we have proved the following result.
Theorem 4.2.1.4 Let us assume that b>0, at 0 and that equation
(4,2,1,2) has no real root, then there exists a constant Co such that (4,2,1,4)
holds for all ue W(B)n H (B) fulfilling the boundary conditions in
(4,2,1).
Actually, with a little more care, we should be able to bound the first
derivatives of u in Lp (B), by using the same Lemma 4.2.1.3. Unfortu-
nately, this does not allow us to bound the second derivatives, which is
our real goal.
Finally in most cases, a simple density argument allows us to extend the
previous result to all u € W(B) fulfilling the boundary conditions.
Corollary 4.2.1.5 Under the assumptions of Theorem 4.2.1.4, inequality
(4,2,1,4) holds for all u€ W(B) fulfilling the boundary conditions in
(4,2,1), provided either a; = 1 or a; = B; = 0, 1; = 1 for each j=0, 1.
This is deduced from Theorem 4.2.1.4 with the help of the following
lemma.
Lemma 4.2.1.6 Assuming a; = 1 or a; = B; = 0, 1; = 1 for each j= 0, 1,
the subspace of W(B)H (B) defined by the boundary conditions
V:M;u =0 on F. is dense in the subspace of W(B) defined by the same
boundary conditions.

Proof Due to the trace theorems in Section 1.5 we can view W(B) as

W(B) X II (W3-1P(E,) X WA-NO(E,))

and H2(B) as

i =0
Since H?(B)n W3(B) is dense in WE(B), it will be enough to prove that
for each i the space Z2M Zp is dense in Zp, where Za is defined for all 9
as follows:
Za = {(k, 1) | kE WE- 119(E;), lE WA- 19(F,), al + B;D,k + 1, k = 03.
In the case when a; = B; =0 and 1; =1, the space Za reduces to
Ox WI- 14(E;) and thus Z2M Zp is obviously dense in Zp.
 4.2 ESTIMATES IN AN INFINITE STRIP 189

In the case when a; =1, the space Za is nothing but

<(k, -B;DAk -1;k) | kE WE- 119(Fr)},
which is isomorphic to W3-1/a(E;). Again Z2M Z, is obviously dense in
Zp. •
4.2.2 Lp bounds for the second derivatives of the solution
In order to be able to bound the second derivatives of u, the solution of
problem (4,2,1), we replace it by a slightly different problem:
in B (4,2,2,1)
ly:Miu =0 on F, j=0,1,
where
Lu = D<u+ D3u-u=(4-1)u.
The reason for doing this is that (1-4) has an elementary solution E
with good properties in the Sobolev spaces. This was not true for L.
We shall prove in this subsection that there exists a constant C, such
that
I|D°ullo..,B < C, l1 -4)ullo.p.B (4,2,2,2)
for all lal =2 and u € WE(B)n H (B) which fulfils the boundary condi-
tions in (4,2,2,1). Then it will be very easy to deduce (4,2,2) from
(4,2,1,4) combined with (4,2,2,2).
Now we proceed exactly as in Subsection 2.3.2. The elementary solu-
tion for 4 - 1, namely
E=-F-11+16127-1
is linear continuous from Lp (R3) into W3(3) according to Theorem
2.3.2.1. Then let us denote by o the function
v=u-E* g; (4,2,2,3)
we obviously have
((1-4)0=0 in B (4,2,2,4)
Cy;M;o =h; on Fir j=0,1
where
h; =-y;M;E *g, i=0,1. (4,2,2,5)
Thus, denoting by d; the order of M; = a,D, + B;Dx + 1;, we obviously

have h, € W3-4-1/9 (E;) M H3-4, -112(E;), 1 =0, 1.

190 SECOND-ORDER PROBLEMS IN POLYGONS

We shall calculate explicitly the partial Fourier transform in x of w. It is

a solution of
(-0" + (1+53)0=0 in 10, h[
908(5, 0) + (Bo 5 + 10)0(5, 0) = ho
(a, 8' (5, h) +(1B, 5+11) 0 (5, h) = M,
for almost all §ER.
Consequently we have ¢(5, y) = a (5) cosh ry + B(E) sinh ry where r =
W(1 + 22) and
1
a (E) = àlaurh,-La,r cost rh + (iB, s+ 1, ) sinh rh lhot
В(E) = it-liBos + 1. h, + la,r sinh rh + (iB, s+ 1,) cosh rh hot
d=rcosh rh[ao(ißE +11) - a (iBoE + 10)]
+ sinh rh[r=a0a1 -(iBo5 + 10) (iß1 5 + 1,)]. (4,2,2,6)
Of course these identities make sense only if we assume that d does not
vanish.
From these formulas, we shall deduce the traces
k; =yU on Fin l;=y;D, on Fi j=0, 1.
Actually we have

Ko (E) =a (5), 10(5) = r (E) B (5)

k, (5) = a (E) cosh rh + B (E) sinh rh
¼, (E) = a (E)r sinh rh + B(E) r cosh rh.

Lemma 4.2.2.1 Assume that d defined by (4,2,2,6) does not vanish for
any #ER, then there exists a constant C such that

j=0 j=0

Proof This is just a repeated application of Lemma 2.3.2.5. -

Then we look at ©, the continuation of v by zero outside B. It is a

solution of
(1-4)0=-k0886+k1884-10880+488491 (4,2,2,7)
where So (respectively S) is the Dirac measure at zero (respectively h).
 4.2 ESTIMATES IN AN INFINITE STRIP 191

Consequently we have
u=E*58-ko@86+k188h-688+1881}.
From this representation of u we shall deduce the following basic result.

Theorem 4.2.2.2 Assume that d defined by (4,2,2,6) does not vanish for
any real 5, then there exists a constant C, such that
ul2.p.B SC, |810.p.B (4,2,2,8)
for all u € W?(B) n H?(B) which are solutions of problem (4,2,2,1).

Proof This is mainly the same proof as for Theorem 2.3.2.7. Indeed we
know that there exists a constant C such that

Then, let us consider one typical term E * (ko&8%). From (4,2,2,5) and
Theorems 2.3.2.1 and 1.5.1.1, we know that there exists a constant C
such that
Thill2-d,-1/p.p.E, SC |1810.p. B.

Then combining this inequality with Lemma 4.2.2.1 we have (with

possibly a larger constant)
koll2-1/p.p.F. €C||8lo.p.B.
Now Lemma 2.3.2.2 shows that
k088ó,
and depend continuously on ko€ W3-1/ (Fo). Then setting
no=E* (k088ó
and applying Lemma 2.3.2.5, we see that

and depend continuously on ko€ W2-1/P (Fo).

Then we write
Duo = D}E* (ko88) =(1-D})E#(k,@So)
in the half plane y > 0. Thus

and
192 SECOND-ORDER PROBLEMS IN POLYGONS

Applying again Lemmas 2.3.2.2 and 2.3.2.5 we show that

DiMo, DxD,up ELp (B)
and depend continuously on ko€ W3-19 (Fo).
Finally we write that
DZUo = uo-D}uo in B.
Consequently DupE Lp (B) and depend continuously on ko€ W3-"P (Fo).
Summing up we have proved that
U0l2.p.B €C||8l0.p.B
for some constant C. The other terms in (4,2,2,7) are estimated by the
same techniques. •
Again using the density result of Lemma 4.2.1.6, we can extend
inequality (4,2,2,8) to all u € W$(B) in most cases.
Corollary 4.2.2.3 Under the assumptions of Theorem 4.2.2.2, inequality
(4,2,2,8) holds for all u € WE(B) which are solutions of problem (4,2,2,1)
provided either a; =1 or a; = B; =0 and 1; = 1 for each j= 0, 1.

Theorem 4.2.2.4 Assume that b > 0, a + 0 and that for each j= 0 or 1 we

have either a; =1 or a; = B; =0 and 1; = 1. Assume in addition that
equation (4,2,1,2) has no real root. Then there exists a constant C such that
inequality (4,2,2) holds for u e WE(B) the solution of problem (4,2,1).

In the proof, we shall need this auxiliary lemma which we will prove
later.
Lemma 4.2.2.5 Assume that b>0, a+ 0 and that for each i we have
either a; = 1 or a; = B; = 0. Then it is possible to find Mi j = 0, 1, such that
rcosh rh[ao (iß,¿ + My) - a, (iBo's + Mo)]
+sinh rh[ra.as-(iBo5 + Mo) (iß, E + Me,)],
where r =V(1 + 52) does not vanish for EEl.

Proof of Theorem 4.2.2.4 Due to the assumption that equation (4,2, 1,2)
has no real root, inequality (4,2,1,4) holds. Next we shall use inequality
(4,2,2,8) with 1; replaced by M;. From Corollary 4.2.2.3 and the trace
theorems in Subsection 1.5, we deduce the existence of a constant C
such that
 4.2 ESTIMATES IN AN INFINITE STRIP 193

for all u € W(B). In particular when u satisfies the boundary condition

vi(a;Du + B;Dxu + 1;u) =0 (4,2,2,9)
on Fi, we have

(4,2,2,10)
We observe that we have no boundary terms in the case when d; =0,
since v;u = 0.
From (4,2,2,10) it follows that

d,=1
In other words, we have
ul2.p.B <C26/Lullo.p.B + lull1.p.B}.
Next we take advantage of the inequality
K
uh.p.Be€ Ul2,p.B + -• ul0.р, в
for all & <1. If we choose & to be 1/2C, we obtain

Mullen 52631/kullan. + KHello set?.

Finally, using inequality (4,2,1,4) we conclude that there exists a constant
Cs such that
/lull... BCCs||Lullo.p.B
for all u€ W-(B) which fulfils the boundary conditions (4,2,3,9).
Proof of Lemma 4.2.2.5 Let us consider first the case when a; = B; =0
for 1 =0, 1. Then the requirement is that equation
MoM, sinh V (1 +52) 1=0
should have no real root &. This is achieved provided M; #0, j=0, 1.
Next we consider the mixed case when a = B. = 0 and o, = 1. Then the
requirement is that equation
MoV (1 + 53) cosh V (1 +53) h+ Mo(ißis + Ms) sinh V (1 +53) h=0
should have no real root. The real part of the equation is
MOV (1 + 53) cosh V (1 + 53) h + Me, sinh V (1 +53) 13=0.
194 SECOND-ORDER PROBLEMS IN POLYGONS

The condition is achieved if we choose No + 0 and f, = 0, for instance. Of

course in the case a, = B, = 0 and a = 1 we choose No = 0 and M1 + 0.
Finally, let us consider the case when a = a, = 1. Then we require that
V(1 + 82) cosh V (1 +53) hi(Bi -Bo)E + (MI - MO)]
+ sinh v(1 +#) h[(1 +53) -(iBoS + Mo) (iB, E + M 1)] =0
should have no real root. The real part of this equation is
(My -Mo)V (1 +53) cosh V (1+53) h
+[1+531+BoB. - MOM, ]sinh V (1 +53) h = 0.
When 1+ BoB, =0, we can choose Mo = u, = 0; then the equation be-
comes
[1+531+ BoB.} sinh V(1 +53) h=0,
which has no real root. On the other hand, when 1+ Boß, <0, we can
choose Mo = us = 2; then the equation becomes
[-3+57(1 + BoB.}] sinh W(1 + 53) h,
which has no real root either. •

4.3 Bounds in a polygon

4.3.1 The Lz case
We prove here inequality (4,1,2) when p = 2. We follow word by word
the proof in Grisvard (1972). The principle is the same as in Theorem
3.1.1.2 plus an additional density result. Again here, it is technically more
convenient to work with
v = DxU, W = D,u.
The boundary conditions for v and w are the following.

Lemma 4.3.1.1 Let u€ H7(5) fulfil the boundary conditions in (4,1,1);

then for all i there exist two real numbers 1; and u; such that
A;v;o + M;V;w =0 on I; (4,3,1,1)
and 17 + p? 70.
Proof The condition (4,3,1,1) means that Wu is orthogonal to some
nonzero vector whose components are 1; and M; Indeed, in the notation
of Section 4.1, we assume that
y; Vu. (wi + B;T;) = 0
 4.3 BOUNDS IN A POLYGON 195

on I; for jeN, and that

y;Wu-T; =0

From now on, we denote by GS (1) the space

{(U, w)EH' (R)X H' (1) V; (1;0 + M;W) = 0 on Fi, 1 5j EN}.

Lemma 4.3.1.2 The identity

(4,3,1,2)
Ja D, 0D, m ax do =1, B,D, w dx do
holds for all (v, w) € G7(1).

Proof Integrating by parts twice, we obtain

owing to the Green formula.

Next we split the boundary integral into pieces. We have

rodow=£
j=1
yiu dy;w.

We assume that the I have been numbered according to the positive

orientation of the boundary. We denote by S; the terminal point of I;;
thus S;- is the origin of I; for j> 1, while S~ is the origin of I. (It is
obviously convenient to set Sv = So, I N+1 = Is.)
When M; 70, we write y;w = -(1;/ u;) YU, and consequently

yody,w= #: ( (70) (5)-(740)7454-135.

2p;
This identity is meaningful since o € H?(R) and consequently v € C(R) by
the Sobolev imbedding theorem.
When M; =0 we just have y; = 0 on I; and accordingly

yody,w=0.
Next we observe that (y;u) (S;) = (vi+1U) (S;) due to Theorem 1.5.2.8. Thus,
196 SECOND-ORDER PROBLEMS IN POLYGONS

we have

1) (403451)
н, н, +1*0 M;

#+#44 (70345) -21 (40)45).

M, *0 M
M, =0 M; +1
M, +170 M, +1=0

We shall now check that (y;)(S;) = 0 for all j such that 1; +1M; + 1;/; +1
and consequently that
rodrw=0.
Indeed at S; the boundary conditions corresponding to I; and Fits hold
together by continuity. Thus, we have
1; (viv) (S;) + M; (V;W) (S;) =0
11; +1 (7,0) (54) + M,; +1 (y;W) (S;) = 0.
This implies that yo (S;) = 0 when A;M; +1 + 1;+1M;. This is the claim.
Summing up, we have proved that

for all (v, W) € G7(8). m

In order to extend the previous result to the whole of G'(R), we need a
density lemma.

Lemma 4.3.1.3 G7(1) is dense in G'(M); consequently (4,3,1,2) holds for

all [u, w} in G'(R).

Proof The trace Theorem 1.5.1.3 allows us to consider H'(1) as

the direct sum of H'(1) with the space H'(F). Thus any continuous
linear form I on G'(R) may be represented as
«1; (U, W}) = (S, 0-pyU) + (T, w-pyw) + (8, yo) + (h, yw)
where S, T€H-'(9) and g, he H 12(F) and where p is a right inverse
for the trace operator Y.
Let us assume that I vanishes on G?(1). Then, in particular, it vanishes
on H?(8) × H=(8) and, therefore, we have
(S, w) +(T, w) = 0
for all o, we H (M). This implies that S= T = 0. We have thus shown that
any continuous linear form on G'(1) which vanishes on G7(8) may be
 4.3 BOUNDS IN A POLYGON 197

represented as
«h; (U, W}) = (8; yo) + (h; yw),
where g and h H-12(r).
Now, in order to prove that G'(S) is dense in G'(R), we have to check
that any such I is identically zero. In view of the above representation
formula, it is therefore enough to prove that the space Z?(F) of the traces
of the functions in G7(8) is dense in the space Z'(F) of the traces of the
functions in G'(R).
A first step is to describe these spaces, taking advantage of Theorem
1.6.1.5. Let us begin with Z'(F). This is a subspace of MIN
An element belonging to Z'(F) will be denoted

where g; stands for y; and h; stands for y;w. According to this notation
Z'(F) is the subspace defined by
1;g; + M;h; =0 on Ti 1<i<N
I' . (410) =41(45 - 098 do ca,
1<i<N (4,3,1,3)
18, 18: +1 (X; (0)) - 8: (x; (-0))12
- do 400, 15jEN.
Now we choose N pairs of real numbers (5;, ni) with 57+n; +0,
1<^N and such that
(a) sir; - nih; 70, 1<i=N
(b) (Sis Mil = (5+1, Mi+1) for all i such that A;M; +1 - 1; +1M; = 0.
In other words, we require the vector v; = (5;; ni) to be linearly inde-
pendent of the vector (1;; M;) and in addition we require v; to be equal to
Vi+1 whenever (1;, M;) and (1; +1, M; +1) are linearly dependent. Such a
choice of vectors v; is obviously possible. Next, let us define
4; =5;8; + nihis 1<i&N. (4,3,1,4)
It is easy to check that Z'(F) is isomorphic to the subspace of those

i=1
such that
(S,
'14: (8, (-0)12 do 200
(4,3,1,5)
<00
198 SECOND-ORDER PROBLEMS IN POLYGONS

when A;k; +1-1; +1M; +0 and such that

<00 (4,3,1,6)

when A;M;+1-1;+14; =0.

In the same way we describe the space Z?(I). It is the subspace of
AIM, H32(E;) defined by
1;8; + Mih; =0 on To I<i<N
h; +1 (S,) = h; (S;), I<i<N (4,3,1,7)
(8;+1 (S;) = 8; (S;), 1<j&N.
Then introducing again ; defined by (4,3,1,4), we check that Z?(I) is
isomorphic to the subspace of those {o. I, EIN , H32(Г:) such that
4; (St) =4;+1 (St) =0 (4,3,1,8)
when A;M; +1-X;+1M, +0 and such that
Q;(S;) =4; +1(S;) (4,3,1,9)
when A;M; +1-1;71M; =0.
Finally let us consider {;}N, eZ'(I). Due to the density of D(I;) in
À'(T;) and of D(I;) in H'2(T;) (see Subsection 1.4.2), we can approxi-
mate 4; by Qi,m € H33(F;), m = 1, 2, ... such that, for each m, (4,3, 1,8)
and (4,3,1,9) hold with 4; replaced by Qi,m. This can be achieved in such a
way that

in the norm of H'(I) and, in addition, that

-> 0
f'ica-ander-consider
- -> 0

when A, Mi+1 - 1;+1M; 70, while

-> 0

when A;M;+1-1; +14; =0. This completes the proof of Lemma 4.3.1.3.

We are now able to prove our main result.

Theorem 4.3.1.4 Let S be a bounded open subset of R? with a (strictly)

 4.3 BOUNDS IN A POLYGON 199

polygonal boundary F. Then there exists a constant C such that

(4,3,1,10)
for all u€ H (S) which are solutions of problem (4,1, 1).

Proof This is proved via a very straightforward calculation. As in

Section 3.1.2 we calculate the following integral
HAul-dxdy = IDiu+ Diul-dixdy

(4,3,1,11)
Then applying Lemma 4.3.1.3 to v = D.u, w = D,u, we obtain

Consequently we have

and the result follows by inequality (1,4,3,2). -

Remark 4.3.1.5 Here we have a very precise control of the constant in
inequality (4,3,1,10). Indeed it depends only on the best constant K in
inequality (1,4,3,2); in this particular case this is
lulli2.0 sallull2.2.0+ Ke'Mullo.2.52.
In most practical cases, given a plane domain M with a polygonal
boundary, the constant K can be determined explicitly as a function of 2.

Remark 4.3.1.6 Let us observe for further reference that we have

proved identity (4,3,1,11) for all functions We H (S) which fulfil the
boundary conditions in (4,1,1).

Remark 4.3.1.7 So far we have excluded the domains 2 with cuts.

However, the inequality (4,3,1,10) remains valid if one allows & to have
cuts. Indeed the only modification of the proof occurs in Lemma 4.3.1.3.
An application of Theorem 1.7.3 must be substituted for the application
of Theorem 1.6.1.5 at the appropriate corners.

4.3.2 The Lp case (p + 2)

We shall now derive inequality (4,1,2) for p + 2. The method that we shall
use here is quite different from the method of Subsection 4.3.1. Curiously
200 SECOND-ORDER PROBLEMS IN POLYGONS

enough, the method used here does not work when p = 2. (See, however,
Section 4 in Kondratiev (1967a) who deals with the case p = 2.) It relies
essentially on the estimates proved in Section 4.2.
We shall need some new weighted spaces similar to those introduced by
Kondratiev (1967a) in the case when p = 2. We shall denote by p(x, y) the
distance from the point (x, y) to the vertices (S; 1 < jEN) of 8.

Definition 4.3.2.1 We denote by P(S) the space of all functions u

defined in 2 such that

for all a Sm.

Obviously we can define a Banach norm on P(S) by setting

la Em
The inclusion of PM (2) into W*(S) is obvious. We shall actually need
a converse statement. Fortunately the converse inclusion holds, up to the
addition of a finite-diensional space, at least when p# 2.
Theorem 4.3.2.2 Let u€ W: (R) be such that

D"u(S.) =0 for lakm-2;

p
j=1,2, ...,N and p+ 2; then u E PM (R).
Proof By induction on k, we shall prove that
pD°u€L. (82) (4,3,2,1)
for a <m -k. Thus we assume that inclusion (4,3,2,1) holds for a given
k and we derive the same inclusion where k is replaced by k+1.
Let a sm-k-1 and set v = D°u. Thus we know that
pho, p* VUEL, (12).
In addition we know that (S;) = 0 when p > 2 or when k ≥ 1. We shall
show that this implies that
pk-'oELp (82).
Let us first look at the case where k =0 and p <2. We observe that the
condition that o k-'y belongs to L. (S) is relevant only near the vertices.

† Let I be the greatest integer <m -2/p; it follows from Sobolev's imbedding theorem,
that u € C'(P) and consequently the condition D°u (S.) = 0 is meaningful for lal < m - 2/p.
 4.3 BOUNDS IN A POLYGON 201

This allows us to localize the problem. Let n; € D (R) be such that n; = 1

near S; and n; =0 outside a small circle centred at S; which contains no
other vertex of M. Let us denote its radius by 8. Now we only need to
show that

for all j. Using polar coordinates centred at S; we can write

(n.0) (pe#) =
•(°'do
a (no) (ore') do.

Equivalently, we have go a (ni0) (ael') do.

p-'(niU) (pe'°) == p do ). do
From the assumptions on v it follows that (A/do)(n;) P is integrable with
respect to the measure o do do. By Hardy's inequality (see Subsection
1.4.4) it follows that p ' no|P is integrable with respect to the measure
p dp de. This is the claim.
Let us now consider the case where either p >2 or k ≥ 1. Then with the
same notation we write

p" (ni0) (pel%) == ps do

and again the inequality mentioned above shows that lo-'n;0| is integra-
ble with respect to the measure p dp de.
The basic result in this subsection is the following

Theorem 4.3.2.3 There exists a constant C such that

Mull..ins@Cillfilo.p.sat/lulli.p.r
for all u € Pp(R) which are solutions of problem (4,1,1), provided

148,11-4, - = a)
is not an integer for any j, where ©; = arc tan B; for jEN and O; = 1/2 for
j€ D.
Proof Again we can consider our problem locally. We fix a partition of
unity In'; j= 0, ..., N on 1 such that ni € D (R?) for each i and
(a) the support of no does not contain any vertex of 12,
(b) the support of n; contains S; and does not contain any other vertex;
in addition the support of Mi does not intersect Ik for k+i and
k+it 1.
 202 SECOND-ORDER PROBLEMS IN POLYGONS

(c) Ani/ark + Bk (anildTk) = 0 on Ik for k = i if jeN and for k =i+ 1 if

it1EN.
It follows that there exists K; such that
4/niu-ni lo.p.asK;llulli.p.r (4.3.2.2)
and
on Fk keD
д д- (nil) = 0 on Is KEN
Lavk
for i = 1, 2, ..., N.
The results in Subsection 2.3.3 imply that
Moulz.p.n-Colli o.p.at/uli.p.o) (4,3,2,3)
since the support of nou is at a strictly positive distance from the corners.
We are now left with estimating n;u for j= 1, ..., N. For that purpose
we use local coordinates as follows.
We fix i once and for all and choose polar coordinates with origin at S;,
and such that 0 = 0 on I;+1, while 0 = w; on I;. We also denote by G the

Ay

0=0

F;+1

Figure 4.1
fl
fl
 4.3 BOUNDS IN A POLYGON 203

infinite sector defined by the half lines with origin at S, and which contain
I; and I+1 respectively.
With this notation the function v = nu is a solution of
A0 = 8 in G
where

with the following boundary conditions. On the line 0=0 we have

v =0
if it 1€D and we have
1 2v
r20- Bitt ar =0
if it 1eN. In the same way, on 0 = c; we have
v=0
if i€D and we have
1 du , dU
r 20 - Biar = 0
if jEN.
Finally we set
w(t, 0) =é12/a»o (e'tit), (4,3,2,4)
where we make the following abuse of notation whose meaning is
obvious: we denote by (e'tie) the value of o at the point whose polar
coordinates are e' and o. Then w is solution of a boundary value problem
in the strip
B = RX]O, w;[.

The equation is

D{w+ D'w+ -D,w+ 4 (4,3,2,5)

qW=k
in B, where we have set
k(t; 0) =€ (2/a)fe? g(e'+io)}. (4,3,2,6)
The boundary condition at 0 = w; is as follows:
w=0 (4,3,2,7a)
204 SECOND-ORDER PROBLEMS IN POLYGONS

if i€D and

D.w -B;D,w =-B;w=0 (4,3,2,7b)

if jeN. In the same way, the boundary condition at 0 = 0 is as follows

W=0 (4,3,2,8a)
if i+leD and
DoW -B., D,W -Z B.,W =0 (4,3,2,8b)

if i+1EN.
The boundary value problem (4,3,2,5) (4,3,2,7) (4,3,2,8) is one of those
that we have studied in Section 4.2. Applying Theorem 4.2.2.4, we know
that inequality (4,2,2) holds provided the following equation has no real
root. When j and i + 1 €N the equation is
sin pa; (1 + B;B; +1) = cos pa; (Bi +1 - B;), (4,3,2,9a)
where p=2/q+ is, §eR. When jeN and j+ 1 € D the equation is
B; sin pa; = cOs pro; (4,3,2,9b)
When jeD and j+ 1 eN the equation is
B; +1 sin pa; =-cos pa; (4,3,2,9c)
Finally, when j and j+ 1 € D, the equation is just
sin pa; =0. (4,3,2,9d)
Separating the real part and the imaginary part in equations (4,3,2,9) we
obtain the following systems of equations:

(CIt B,B, -1) sin 7c, cosh far =(B,.1-B,) cos ?c, cosh fu,

q
|(+ BiB, 1) cos = c, sinh Ew, = (B. - B, 1) sin w; sinh sw;

when j and j+ 1 belong to N and

B; sin- w; cosh tw; = cos-w; cosh tw;

2
B; cos =w; sinh jw; =-sin -
2
9 qa, sinh sw;
 4.3 BOUNDS IN A POLYGON 205

when jEN and i+1€D and

Bit, sin-w; cosh tw; =-cos-w; cosh go,

2
a, sinh so, -sing w, sinh fu,
when jeD and j+ 1€N and finally

sin - w; cosh Ew; =0

2w; sinh Ew; = 0

cOS
when j and j+ 1 belong to I.
In each of the previous systems of equations, $ =0 is a root of the
second equation, while the first equation can be divided by cosh Ew; It
follows that equation (4,3,2,9) has no real root iff
2
(1 + B;B;+1) sin = w;# (B; +1 - B;) cos =

when i and it 1 belong to N and

2 2
B; sin - w;# cos - a;

when i€N and i+1 € D and

B;+1 sin =w;‡-cos= w;

q
when jeD and j+ 1 eN and finally
2
sin - w; ‡0
9
when i and j+ 1 belong to D. If we define ; by the equation
tan Ф; = B;
when jeN and set $; = /2 when j€ D, then all the previous conditions
can be summarized as

#29
Ф; +1 -D; + kTT.
(4,3,2,10)

for all k €I (i.e. k an integer).

When condition (4,3,2,10) is fulfilled then inequality (4,2,2) holds for
206 SECOND-ORDER PROBLEMS IN POLYGONS

our problem and this means the existence of a constant C such that

Finally, performing the inverse change of variables in (4,3,2,4) and

(4,3,2,6), we see that there exists another constant C, such that

and consequently
(4,3,2,11)
These last inequalities together with inequality (4,3,2,3) imply the claim
in Theorem 4.3.2.3. -
An easy consequence of Theorems 4.3.2.2 and 4.3.2.3 is the following
theorem.
Theorem 4.3.2.4 Assume that p+2 and that

1(84 1-8, -341)

is not an integer for any i where $; = arctan B; for jEN and I; = T/2 for
j€D. Then there exists a constant C such that
ul2.p.ssCAullo.p.atull1.p.as (4,3,2,12)
for all u€ W?(S) which are solutions of the problem (4,1,1).
Proof Let us denote by E the space of all u€ WE(2) which fulfil the
boundary conditions of problem (4,1,1). Let us also denote by F the
subspace of E defined by the conditions
u(S;) =0, 1<i<N
(This is not an extra condition when j or ¡+ 1 belongs to D) and
Vu(S;) = 0, 1<j^N when p >2.
(This is not an extra condition when j and i + 1 belong to N and v; + B,r;
and vi+1 + B; +15;+ are linearly independent. It is not an extra condition
either when jeD, i+ 1eN and ; and vi+1 + Bi+ iT; +1 are independent
(and, mutatis mutandis, when jEN and j+ 1 € D).) The codimension of F
in E is finite and due to Theorem 4.3.2.2, F is a subspace of P$(R). Thus
inequality (4,3,2,12) holds for all u € F, by Theorem 4.3.2.3.
Now let us denote by I any projection from E onto F. It is clear that
 4.3 BOUNDS IN A POLYGON 207

when p <2 and

Ta =w-£ tu(5)e, + D,u(S, 14 + D,W(S,)E,)

when p>2 where 4; 4;, 5, belong to W(8). This representation shows
that I is also a linear continuous operator in the norm of W. (8)
provided
2<5€2 when p<2
p
and
1+2-<552 when p >2. (4,3,2,13)
P
We can now prove the desired inequality. Let u € E, then we have
ula.p.ssu-Hulz.p.o+|Nul2.p.s
slu-Mull.p.atCHAIlullo.p.st||/lulli.p.st
since we can apply Theorem 4.3.2.3 to Ilu. It follows that
ul2.p.s.su-Ilulz.p.otCHAulo.p.o+|A(u-Ilu)lo.p.r
+llulli.p.st/lu-Hull.p.st
sClu-Ilulz.p.atCAulo.p.at/ull1.p.rt.
Next, we observe that on the finite dimensional space (1-I)E, the
norms of W?() and of W$(R) are equivalent for s <2. We choose s
such that condition (4,3,2,13) is fulfilled so that I is continuous in the
W$(R) norm. Therefore we have

Is.p.s2+ClAullo.p.rt/lulla.p.r3
€C3|luls.p.at C|Aullo.p.s.
We conclude by taking advantage of inequality (1,4,3,2) which implies
that
Mulls.p.pellull.p.stRE'1-5112-5)Mulli.p.p
for every & € J0, 1. Choosing Cze = ½, we conclude that

This is inequality (4,3,2,12).

Remark 4.3.2.5 The inequality (4, 1,2) follows plainly from inequality
208 SECOND-ORDER PROBLEMS IN POLYGONS

(4,3,2,12) by applying again inequality (1,4,3,2):

Mull.p.aSellulla.p.atKe/lulla.p.r
for every & € J0, 11.

Remark 4.3.2.6 Here we have very poor control of the constant in

inequality (4,3,2,12) in particular because of the abstract functional
analysis procedure that we used for dealing with the equivalences of
norms on finite-dimensional spaces. In that respect we have much less
information in the case p+2 in comparison with the case p =2 (see
Remark 4.3.1.5).

Remark 4.3.2.7 The inequality (4,3,2,12) remains valid for domains

with cuts (i.e. we allow w; =2m for some i). Indeed the results in Section
4.2 have been derived without any limitation on the width h of the strip
B. Thus the only modification lies in the proof of Theorem 4.3.2.4. There
the imbedding (1,7,4) has to replace the usual Sobolev imbedding
(1,4,4,6) in the definition of the space E.

4.4 The Fredholm alternative

In this section we shall derive the consequences of the inequality (4,1,2).

An immediate consequence is that our problem has the semi-Fredholm
property. Then in most cases we shall be able to prove the uniqueness of
the solution by very straightforward arguments. Studying the range of the
Laplace operator under the given boundary conditions will require much
more work. A careful study of the orthogonal of the range will allow us to
calculate exactly the index of our problem.

4.4.1 The semi-Fredholm properties

We first need a classical result of functional analysis.

Lemma 4.4.1.1 Let E, and Ez be two Banach spaces such that E, is

compactly imbedded in Ez. Assume that A is a continuous linear operator
from E, into Ez and that there exists a constant C such that
(4,4,1,1)
for all x € E. Then A has a finite- dimensional kernel and a closed range.

In other words, A is a semi-Fredholm operator. We now apply this

 4.4 THE FREDHOLM ALTERNATIVE 209

result to A = A considered as an operator from

E1 ={u € W;(M); viu = 0 on I; j€ D and
vi(du/av;) + B;(a/aT;)y;u =0 on I; jEN}
into E2 = Lp (M). Due to Theorem 1.4.3.2, E, is compactly imbedded in
Ez and inequality (4, 1,2) is nothing but inequality (4,4,1,1). Thus Lemma
4.4.1.1 shows that the space of the solutions u€ W(8) of problem
(4,1,1) for f=0 is finite-dimensional. In addition, the subspace of all
feLp (8), for which problem (4, 1, 1) has a solution u € W5(12), is closed in
Lp (82).
We shall now investigate the uniqueness of u. We shall state two kinds
of results corresponding to two different methods of proof. We first look
at problems for which uniqueness (possibly up to a constant) follows from
the consideration of

Auu dr dy.
Theorem 4.4.1.2 Assume that B; < B;-1 whenever j-1 and jEN. Then
problem (4,1,1) has at most one solution u € W3(5) defined up to an
additive constant. If in addition, D is non-empty, then problem (4,1,1) has
at most one solution in W(1).

In other words the kernel of 4, considered as an operator from E, to

Ez, is either one-dimensional (when D = 0) or zero (when D+ 0).

Proof Let us assume that u € W5(1) is a solution of problem (3,1, 1) with

f=0. We use the classical identity

- Yiu do. (4,4,1,2)

Such an identity obviously holds for functions in W3(2)n W3(2) with

1/p + 1/g = 1 (see Lemma 1.5.3.3). We therefore consider a sequence Um
m = 1, 2, ... of functions belonging to W5(8)n W(8) and such that
Um -> U
in W$(S) when m -> +0. We have

Vilm do.

for all m, and we can take the limit in m, since by Sobolev's imbedding
210 SECOND-ORDER PROBLEMS IN POLYGONS

theorem (see Subsection 1.4.4), we have

Um -> U in L.(2)
Vum -> Vu in L2(1)
ди in Lp (F;
Vidum
avi a Vi avi
Y;um -> y;u in L.(T.).
This proves identity (4,4, 1,2).
Then since Au =0 and u fulfils the boundary conditions, we have

= Y;uriu do
0=1, Truisands + 82 0 1, an
j=1
This identity is meaningful since u is continuous up to the boundary of 2.
The assumption that B; ≥ B;+1 for all j implies

Wu? dx dy =0
and therefore u is a constant (since M is connected). This constant is zero
if and only if D is nonempty. S

Next we consider problems for which uniqueness follows from the

consideration of

I Hu? dx dy.

Theorem 4.4.1.3 Assume that p >2 and that at least two of the vectors it;
are linearly independent. Then problem (4,1,1) has at most one solution
u€ W=(R) defined up to the addition of a constant. If in addition, D is
nonempty, then problem (4,1,1) has at most one solution in W(R).

In other words, the kernel of 4 as an operator from E, to Ez is either

one-dimensional (when D = D) or zero (when D+ D.)

Proof Let ue W$(8) be a solution of problem (3,1, 1) with f = 0. Then

since we assume that p =2, we have
иє H'(1).
 4.4 THE FREDHOLM ALTERNATIVE 211

It follows that

by Remark 4.3.1.6. This shows that u is a polynomial of degree less than

or equal to 1.
Let us assume from now on that u = Ex + ny + a. The boundary condi-
tion on I means that the vector whose components are & and n, is
orthogonal to u;. Since two of these vectors are linearly independent by
assumption, this implies that u = a, a constant. The constant a is zero if
and only if D is nonempty.
The above investigation of the kernel of 4 as an operator from E, to
Ez is conclusive in most of the practical cases. We turn now to studying
the range of 4. Taking advantage of the fact that it is closed, we shall
instead investigate its annihilator which is a subspace of La(S) with
1/p+1/9 = 1. Naturally this is, in some sense, the space of the solutions of
a homogeneous adjoint problem. This will be stated in a precise way with
the aid of Theorem 1.5.3.6.
From now on we shall denote by Na the subspace of all functions
DE La (8) such that

for all f€ Lp (8) such that there exists u € W(S) satisfying (4,1,1). This is
the annihilator of the image of 4. Obviously N. is a space of harmonic
functions. Indeed, for all u € D(R) we have

and consequently A =0 in the sense of distributions. This implies that

DED(A; La (82)),
a space defined in Subsection 1.5.3. Therefore by Theorem 1.5.3.4 the
traces of u and dolav; are well defined on each of the sides I 1 5j£N.
Precisely, we have
E Wal-la(T;)
Vi avi
when p+ 2 and
doє (H3/2(Г.))*
V;U E (H'2(Ti))*
Vi avi
when p = 2. Accordingly the following statement is meaningful.
212 SECOND-ORDER PROBLEMS IN POLYGONS

Lemma 4.4.1.4 Let v € Na; then v is solution of the following boundary

value problem
40 = 0 in 2,
Y;0 = 0 on I;, j€ D, (4,4, 1,3)
du д- ViV = 0 on li jEN.
(Ju,
For convenience, in what follows, we shall denote by Ma the space of
the solutions of problem (4,4,1,3) which belong to La (2).

Proof Let us first look at the case when p+ 2. Given 4; € W3-1P(Ty),

jEN, and 4, E WH-W(Th), je D, there exists u € W7(1) such that
ди jEN
de;,
=- Bi aT,
ди j€ D.
= Wis

This is a direct application of Theorem 1.5.2.8. Indeed all the conditions

(a) in this theorem are obviously fulfilled since both sides of the desired
identities vanish (see Corollary 1.5.1.6).
We observe that u(S) =0 for all i and in addition that
Vu(S;) = 0 for all i
when p >2. Consequently, we can apply Theorem 1.5.3.6 (the Green
formula) to this function u and o € Na. We obtain

Elian: 20) =0.

i.e.
24)1=0.
: 410) -44, 4.84

If we let 4; vary in W3-VP(Ty) and 4 vary in Wb-1/P(F;), this identity

shows that
y;0 = 0, jED,
. д- y;0 = 0, jEN.
Vi avi- Biati
We have thus checked that v is solution of problem (4,4, 1,3) when p+ 2.
To conclude we must look at the case where p = 2. We just observe
 4.4 THE FREDHOLM ALTERNATIVE 213

that N2S Na for q<2 and consequently v€ N2 is also a solution of

problem (4,4,1,3). -
In the next subsection we shall show that Ma, the space of the solutions
of problem (4,4, 1,3), is a finite-dimensional subspace of La (8). Further-
more we shall be able to calculate its dimension in most cases. This will
show that 4 is a Fredholm operator from E to Ez. Actually, calculating
its index will require some additional work since in many cases Na
happens to be a strict subspace of the space of the solutions of (4,4,1,3) in
La(R). Indeed, let
Ф; € D (R),
be such that:
(a) The supports of 4; and ; do not meet It for It j and it 1 (in
particular they do not contain S, for IF j).
(b) Q; (S;) = 1, Dx a; (S;) = (1, 0), D, 4; (S;) = (0, 1), Tc; (S;) = 0, 43; (S;) = 0.
With this notation we can state the following lemma:

Lemma 4.4.1.5 Let vE Na; then for all ue W-(R), ful lling the bound -
ary conditions in (4,1,1) and all j, we have
fu(S,) Aq; +Tu(S;) •Ad,todx dy

=
- + ВкДТк
kENnil:it1s LAvk

(4,4, 1,4)

when p>2, and

u(S) 44;0 dx dy = - + BRaved
kENNi;i+1}
(4,4, 1,5)
when p <2.
Proof This is again an application of Theorem 1.5.3.6. Let us look at the
case when p > 2 first. Let u € W5(8) and set

W = U-¡=1 i=1

Then obviously we W-(2) and

w(S;)=0, Vw(S;)=0. (4,4, 1,6)
fi
214 SECOND-ORDER PROBLEMS IN POLYGONS

for all j. Since v € Na is also in D(A; La(g)), we can apply identity

(1,5,3,6) to w and U.
Thus we have
dw
Awodxdy = & {(VKAvK
k=1
since 4v = 0. We then observe that (4,4, 1,6) implies that YkW€
We-V/P(Fk). On the other hand, we have proved in Lemma 4.4. 1.4 that
dv д YKU
when keN. It follows that
d
Thus we have

KEN

Since

Auw dx dy=0,
д- VU = 0 on lk for keN
дИ + ВкЭтк
and
YU = 0 on lk for kED,
it follows that

i=1

kEN i= LAvk
dU
KED 1=1
Now if we let u vary, the values of {u(S;), Wu(S;)} for different j are
 4.4 THE FREDHOLM ALTERNATIVE 215

independent. Thus we have

KEN LaU
do\
KED

Due to the assumptions on the supports of 4; and ; the sum in k has to

be extended to k = j and k = j+ 1 only. This proves (4,4, 1,4).
In the case when p <2, we make the same calculations defining was

The last sum

kEDMi.i+1} usiast..or
20)
is always zero since when D{i, i+ 1}+ 0 we must have u(S,) = 0. a

Remark 4.4.1.6 The meaning of these two lemmas is the following: In

addition to being a solution of problem (4,4,1,3), every function U € Na
must fulfil a finite number of linear conditions defined by (4,4,1,4) or
(4,4,1,5) (observe that it is not clear whether or not these conditions are
independent). By the way, this shows that the adjoint problem to (4,1,1) is
not exactly the adjoint boundary value problem as is always the case
when the boundary of & is smooth.
Remark 4.4.1.7 The conditions on ~ expressed by (4,4,1,4) and
(4,4,1,5) can be simplified in most cases. Let us first look at the case when
p <2; we have two possible cases:
(a) If j or i + 1 belongs to D, we always have u(S;) = 0 and (4,4,1,5) is
not an additional condition on v.
(b) If i and j+1 belong to N, then u(S;) is any real number and
consequently condition (4,4,1,5) is nothing but

(,; 440). (4,4, 1,7)

Avk
Then when p >2, there are many more cases.
(a) If j and j+ 1 belong to D and the angle w; is not flat (the case of a
216 SECOND-ORDER PROBLEMS IN POLYGONS

flat angle with Dirichlet boundary conditions on both sides is irrelev-

ant since flat angles are considered only when dealing with mixed
boundary conditions), then we always have u(S;) and Wu(S;) = 0.
Therefore, (4,4,1,4) is not an additional condition on v.
(b) If jeD and j+1€N and if we assume that , and Mi+1 are linearly
independent, we also have u(S;) =0 and Wu(S;) =0 and thus no
additional condition on v.
(c) The same holds when jeN, j+ 1eD and if T; and u; are linearly
independent.
(d) If j€ D and j+ 1€N and if r; and Mi+1 are parallel, we have only
u(S) =0 and Wu(S;) •r; = 0.
Thus condition (4,4,1,4) is equivalent to

(4,4,1,8)
Lav; +1 +14-84:
+ Bi+1 aTit1 41410) -444-845 41842)
(e) A similar result holds mutatis mutandis when jEN and j+ 1€ D.
(f) When j and j+ 1 belong to N and f; and M;+, are linearly indepen-
dent, then we have Wu(S;) =0 and condition (4,4,1,4) reduces to
(4,4,1,7) again.
(g) Finally, when j and j+ 1 belong to N and M; is parallel to M; +1, we
have only
Vu(S;). M; =0
while u(S;) and Wu(S;) • ; are any real numbers. Therefore, condition
(4,4,1,4) is equivalent to condition (4,4,1,7) and the following condition:

(4,4,1,9)
Summing up, we have proved the following theorem.
Theorem 4.4.1.8 Let p+2; then Na is the space of all solutions 0 € La (1)
of problem (4,4,1,3) which in addition fulfil the following conditions:
(a) (4,4,1,7) for all i such that both j and i+ 1 belong to N;
and, when p > 2:
(b) (4,4,1,8) for all i such that j€ D and i+ 1€ N or such that iE N and
j+1eD and Me; is parallel to Mitt;
(c) (4,4,1,9) for all j such that both j and it 1 belong to N and Me; is
parallel to M; +1.
 4.4 THE FREDHOLM ALTERNATIVE 217

Unfortunately we are unable to prove such a precise result when p = 2.

The reason is that, for u€ H7(9), it is not possible to apply the Green
formula of Theorem 1.5.3.6 to the function

W = 4 -5 u(S,)4;
¡=1
Indeed, in general we have y;We H312(T;), and in addition
y;w (S;-1) = y;w(S;)= 0,
but this is not enough to conclude that y;we H312(E.) (see Subsection
1.5.1).
4.4.2 The adjoint problem
In this subsection we shall show that the dimension of Na is finite in most
cases. This will be achieved by studying thoroughly the behaviour of the
solutions of problem (4,4,1,3) which belong to Lg(M). Sometimes, this
will also allow us to calculate exactly the dimension of Na-
Lemma 4.4.2.1 Let v € Ma, then DE C°(DIV), where V is any neigh-
bourhood of the vertices of St.
Proof Actually v is a harmonic function in M and it is well known that it
is smooth inside . We must prove the smoothness of v near any of the
I; For that purpose we fix j and perform a change of coordinate axes
such that the segment I; is on the axis {*2 = 0} and such that & is above
I; Then we introduce a cut-off function 4 € D(R), whose support does
not intersect any of the sides I with k7j (consequently it does not
contain any of the corners) and such that 4 does not depend on X2 for
small values of X2. We shall now investigate the smoothness of qu.
The function w = ∞ belongs to La (R%) where R% = {*2 > 0}. In addition,
w is solution of
-Aw +w=f in R?
YaW
d*z+ BiT axI
on {x2=0} if jeN,
(yw = 0 on {x2=0} if jeD,

where f=590-216-VU-(Ap)u}
and
8 = 109 + Biax if jeN.
218 SECOND-ORDER PROBLEMS IN POLYGONS

At first sight we have feW-'(R7) and g € W- I (R). However, f is

actually a little better than this. Indeed, f is smooth for *2 > 0, while for
small values of X2, we have

1=390-24900
дх, дх,
due to the fact that & does not depend on *2. It follows that
fEL, (R.; W= '(R))
if we agree to view f as a vector-valued function of *2. This will allow us
to show that w € W'(R%) as a first step.
We replace w by Rw, where R is the inverse operator of (1- D3) /2;
i.e.
RW=F; ' (1 +57)-112 F, W,
where F, denotes the Fourier transform in *. It follows from Lemma
2.3.2.5 that RwE La (?) and that
-ARw + Rw = Rf in R3
dRw== Rg on 1x2=03 ifjeN
(aRW + Bi axis
(yRw =0 on {*2 = 0} if jeD,
where Rf€ La (R%) and Rg € W'-1/9 (R). We conclude by applying Proposi-
tion 2.5.2.4 when j€ N and Corollary 2.5.2.2 when j€ D, replacing & by
any domain M, with a smooth boundary containing the support of & and
such that I; cAl. It follows that

and consequently we have we W=(R$) and wE W'(R₴). If we vary 4

and i, we finally show that
vEW(RIV),
where V is any neighbourhood of the vertices of 12.
Now we retrace all the previous steps of the proof. Since we know that
u belongs to Wa(D\V), we also know that f€ L, (R%) and g € WI-1/a (R$).
Thus, applying Corollary 2.5.2.2 and Proposition 2.5.2.4 to w in this case
(instead of Rw) shows that
 4.4 THE FREDHOLM ALTERNATIVE 219

and consequently

where V is any neighbourhood of the corners of 12.

Finally, repeated application of Theorem 2.5.1.1 with & replaced by
R1 as above, shows that

for every positive integer k. The Sobolev imbedding theorem (Subsection

1.4.4) implies that
VEC°(DIV).
The proof of Lemma 4.4.2.1 is complete. W

Now we shall study the behaviour of € Ma near the corners. For

simplicity we begin with those corners S; which correspond to self-adjoint
conditions. In other words, we assume that B; =0 if jeN and that
B; +1 = 0 if i + 1 € N. For technical purposes, we shall need the eigenfunc-
tions of the operator

under various boundary conditions in the interval 10, c;L.

More precisely, let us define the unbounded operator A; in I; =
L2(0, wiL) as follows:

1;4 =-4",
where 4 € D(A;), the domain of 1;, given by

(<4 € H=(0, w;D) I 4 (0) = 4 (w;) =0} if i and i+ 1€ D

{4 E H=00, w;D) I 4'(0) = 4'(w;) =01 if j and j+1EN
D(A;) =
{$ € H=(10, c;D I $ (0) = Q'(w;) = 0} if jeN and j+1€ D
(ic E H=(10, c;D 4'(0) = 4 (w;) =03 if jeD and i+1EN.
This is a nonnegative self-adjoint operator with a discrete spectrum. We
shall denote by Qi.m m = 1, 2, ... the normalized eigenfunctions and by
44.m m = 1,2,... the corresponding eigenvalues in increasing order of
magnitude. We thus have

where Qi.m € D(A;) for every m.

220 SECOND-ORDER PROBLEMS IN POLYGONS

Of course, we have

Pim(0)=V(Z) sin.тто Aj.m = MIT when and it 1€ D

wig

INa, m=1
4im (0) = (m - 1) т0 Aj.m (m -1) t m ≥2
2) cos W; w,
when i and i+1€N
(m -}) т0 (m -=)т
gim(0)=V(2) sin 1;,m =
W; W;
when i€N and i+1 € D
(m -2) I (w; -0) (m -„)т
cim(0)=V/Z sin w; Aim =
W;
when j€D and j+1€N.
Using the polar coordinates with origin at S; (introduced in Subsection
4.3.2), any v € Mg is a solution of

270 120 1220 (4,4,2,1)

r dr 1=302=0,
where p > 0 is small enough (chosen such that the disc whose centre is S;
and radius is p does not cut any side of M except I; and I+1. We set
D. = 2010 <r; <py). In addition it fulfils the following boundary condi-
tions: at A = 0

0=0 if i+1€D and dU

ã60 ifitlEN,
and at A = w;

0=0 if i€ g and до
ae= 0 if jEN.
Since v is smooth for r > 0 by Lemma 4.4.2.1, we have
u(re%) € H300, w;D.

It follows that
U(re) E D(A;)
for each r € 10, p[.
 4.4 THE FREDHOLM ALTERNATIVE 221

Consequently we have
220 120 1
grat:r dr $ 1,0=0, O<r<p. (4,4,2,2)
This implies that v can be expanded in series of the eigenfunctions of A;;
in a very special fashion, which we describe now.

Proposition 4.4.2.2 Let v € C°10, p]; D(Ak)) be a solution of Equation

(4,4,2,2) and assume that u € La (Dp). Then
0(reia) =0, (r)c;.1(0) + E amp.m;m (0) + Bmr-^, mi;m (0),
m=2 0<1m <2/9
where

and
U(r) =as+B,logr if 1;1=0,
and where am and Bm are real numbers such that
lam|SIm'ap^,m (4,4,2,3)
where L is a constant which depends only on v.

Proof Since the sequence Qi;m m = 1, 2, ... is a basis of I;, we have

o(reic) = & Um (r)4;m (0), (4,4,2,4)
m=1
where

(4,4,2,5)

However, since v is differentiable in r with values in D(A;), the differen-

tial equation (4,4,2,2) implies that
1
r
U(r)+-Un(r) -.

Accordingly, we have
Um (r) =AmP'im+BmP^,m
when A;m >0, and
Um (r) =Am+Bm logr
when A;m = 0.
222 SECOND-ORDER PROBLEMS IN POLYGONS

On the other hand, since u belongs to La (D.), it follows from identity

(4,4,2,5) that

and consequently

(4,4,2,6)

This implies that Bm = 0 when 1;m >2/g, and in addition that

for 14,m ≥2/g. This completes the proof of Proposition 4.4.2.2.

We shall now show that v € M has an expansion near each corner
(which looks very much like the expansion in Proposition 4.4.2.2) in the
general case where B;, Bi+1 are possibly nonzero. We shall use here the
eigenfunctions and eigenvalues of a different operator. Let us denote by
1; the unbounded operator in I; = L200, w;L) × L2(0, w;L defined by
1:501,023=502-013
where {us, vale D(A;) and

Icos 4;02 (w;) - sin Q;01 (w;) = 0

It is obvious that A; is a self-adjoint operator and has a discrete
spectrum. The expansions in terms of eigenfunctions are as follows:

Lemma 4.4.2.3 Every {us, UlEI; has an expansion of the following

form
100 1
U, = - am cos (1; mo + 9; +1)
So vw;

U2= Bm sin (1;, mO + Q; +1)

where

Aim =
w;
 4.4 THE FREDHOLM ALTERNATIVE 223

and
Om =
Vw;
provided (Ф; -O;+1)| m is not an integer.
When (Q; -Ф;+1)/t is an integer I, the expansion is
Q-1 1
1
am cOS 1;.imP + Qi+1)
m=, Vw;

.., Jo, am sin (Aim + 9, 41)

where

and am m ‡-I is as before.

Proof It is easy to check that the eigenvalues of A; are the numbers 1; ms

-0 <m < +0, m integer and that the corresponding eigenvectors are

Aim (0) =
1 {cos (1;. mO + D;+1); sin (1;, mo + 0; +1)}
Vw;
for A;m # 0 and

Pi.m (0)=
1
Jw. (1 + tan? p. jil; tan Bit
for 1;.m = 0.
Using again the polar coordinates introduced in Subsection 4.3.2, we
see that each o € Ma has the following features. First, by Lemma 4.4.2.1,
u is a differentiable function of r with values in H3(10, w;]) for r€ 10, p[
where p >0 is suitably small. Then, we have again
270 1201 270
art;r dr 72302=0, (4,4,2,7)
where v fulfils the boundary conditions

=0 (and v =O if j+ 1€ D),
(cos 494, 50 + sin do, 11 O<r<p, 0=0 (4,4,2,8)
=0 (and v = 0 if j€ D),
COS 9, dU+sin
5A Oil ar O<r<p, o=w;
224 SECOND-ORDER PROBLEMS IN POLYGONS

Let us set
W=rдр'. W2 = -20'

then, obviously, w= {W, W} is a differentiable function of r with values in

D(;) for O<r&p and
rdw
dr= A;W. (4,4,2,9)
This implies the following:

Theorem 4.4.2.4 Let 0 € C*° (J0, p]; H3(0, w;D) be a solution of equation
(4,4,2,7) fulfilling the boundary conditions (4,4,2,8). Assume, in addition,
that v € La (D.). Then
u (reit) = SCm rim cos (1;, mo +Ф;+1) + k
1, m =- 21a V c0; 1;.m
where Cm and k are real numbers such that
cm|SIp-*im (4,4,2,10)
for some constant I depending only on u, provided (Ф; -Ф;+1)/T is not an
integer.
When (S; -D;+1)/m is an integer I, the expansion of u is
Cmp.m
u(re') =C-' JIw, (1+ tan? $,)]it
log r - A tan $;
COS 1;me + P: +1) + K
m#-1
with the same growth condition on the sequence Cm.

Proof The beginning of the proof is similar to that of Proposition

4.4.2.2. Indeed, the sequence Qi.m is a basis of &;, and thus we have
= (Y

w(reio) =_ Wm(r)@;.m(0) (4,4,2,11)

where
(4.4.2.12)

In other words, we have

до LIL~

r-(reio)
dr = SWim (r)
1
Jo cos (1, mO +9, 41)

- sin (1;, mo + O;+1),

Vw;
 4.4 THE FREDHOLM ALTERNATIVE 225

where

Wm (r) =
Vw;. dr• (reil) cos (1;, mO + ; +1)
ão (re') sin (A,m nO + 9, not do
with the obvious necessary modification when (; -Ф;+1)/m happens to be
an integer. Then the Equation (4,4,2,9) implies that
rw'm (r) =1;,mWm (r),
and, accordingly
Wm (r) =EmP,m.
Thus it follows that

Cm =p^,.m
1
Vw;
- ,3r
до(re'°) cos (1;,mo + Ф; +1)
Ar
al (re') sin (11.0 + 4, 40] do

for every r € J0, p] and therefore there exists a constant L(r) such that

for every r € ]0, p]. This implies the uniform convergence of the following
series:

rdU
ar(re') = Jw;-00
do (reio) = 1 Eam'"imsin (1; mo + Ф; +1)
ae
in the rectangles Or = {(4, 0); € 51 Sp -E, 050 5w;} for E > 0.

Integrating, we obtain p^,.m

cOs (1; mO + Ф; +1) + K.
Sam Jim
This expansion is valid in Us le>o Le 1.e. tor r € JO, pl and o € LO, w;_ (here,
for simplicity, we have assumed that none of the eigenvalues 1; m vanish;
the modifications for covering the general case are obvious). The condi-
tion that o belongs to Lg(D.) imply that cm = 0 when Am 5-2/9. •
Remark 4.4.2.5 The results in Theorem 4.4.2.4 clearly imply those of
Proposition 4.4.2.2.
226 SECOND-ORDER PROBLEMS IN POLYGONS

4.4.3 The Fredholm alternative for variational problems

In this subsection, we restrict our purpose to those problems (4,1, 1) which
are variational. That is why we assume that
B; = 0 unless both j-1 and i + 1 belong to D. (4,4,3,1)
Indeed we have the following statement

Lemma 4.4.3.1Assume that (4,4,3,1) holds. Then for every given f€

Lp (R), problem (4,1,1) has a unique solution u€ H'(R) when D is
nonempty. On the other hand, when D is empty, for every given f€ Lp (82)
such that

faxdy =0,
the problem (4,1,1) has a solution u € H'(R), which is unique up to an
additive constant.

Note that due to (4,4,3,1) this is a pure Neumann problem when

D=0.
Proof As usual, we define a variational solution of problem (4,1,1) as
being any function
u€V={UEH'(R) v;u =0,
such that

alu:;o) =- fo dx dy (4,4,3,2)
for every o € V, where

(4,4,3,3)

We observe that the bilinear form a is continuous on VX V because

the only boundary terms that actually occur (with B;+ 0) are such that
ViU and yU€H ?(Г.)
due to (4,4,3,1) and Theorem 1.5.2.3. On the other hand, by Remark
1.4.4.7 we know that d/at, maps H (I,) into the dual of A 2(T.).
Consequently all brackets in the right-hand side of (4,4,3,3) are continu-
ous on VX V.
Finally we observe that the form a is coercive (see Lemma 2.2.1.1)
 4.4 THE FREDHOLM ALTERNATIVE 227

because we have

a (u; u) = Vul? dx dy. (4,4,3,4)

Indeed, for every 4 € D(I;) we have obviously
дф de- 4 do = 0.
IdT aT;
Then, since D(I;) is dense in H (I,), we have also
дФ.
lat;: 4) =0
for every 4 € HI?(T.).
From identity (4,4,3,4), the coerciveness of a follows with the aid of
Poincaré's inequality when D is not empty. When D is empty, we have
only shown that the form
ü, i r» a (u; v)
is coercive on V = H'(R) C, where C denotes the subspace of the
constant functions.
The existence and uniqueness of a solution u € V to problem (4,4,3,2)
follows now by Lemma 2.2.1.1 when D is not empty. In the case when I
is empty, we have existence and uniqueness in H'(8)/C, provided

in fudidy
is a continuous linear form on H'(1)/C. This means that we have
existence in H'(8) up to an additive constant, provided

J, fax dy =0.
We conclude by showing that our variational solution is actually a sol-
ution of problem (4,1,1). Indeed, restricting identity (4,4,3,2) to 0 € D (82)
shows that 4u = in & in the sense of distributions. Accordingly, u
belongs to E(-A; Lp (82)) (see Subsection 1.5.3) and viu and v; dulav; are
well defined on each I; by Theorem 1.5.3.10. Then the Green formula of
Theorem 1.5.3.11 shows that
ди .д
diavi
on I; for every jEN. A
228 SECOND-ORDER PROBLEMS IN POLYGONS

We shall now try to calculate the dimension of M.. The first technical
step is the following. Here, again, n; is any cut-off function which is I in a
neighborhood of S; whose support does not intersect I for k+ i and
it 1 and such that

du Ani =0
on I; when leNni;i+1}.
Lemma 4.4.3.2 For each j and each A; m E] -219, 01 there exists 0; m € Ma
such that
Oj,m -Mill;,mE H'(D),
where

(Vw;)1;m
-cos (1;, mO; +Ф;+1) if 1;,m" <0
uim (r;e'®,) = log r; - O; tan D
[w,; (1 + tan? Ф,)]
if Air, =0
and i and i + 1 are
not both in D.
Here, again, r, ; denote the polar coordinates with origin at S;.
Proof It is obvious that
An;lim =fim € C° (R)
and that
VIanju;.m-
ди Bi
d on Th led
(Viniuli;m =0 on li lED.
In addition So fm dx dy = 0 when D is empty.
We can therefore apply Lemma 4.4.3.1 to prove the existence of
Vi.m € H' (8), a solution of
(Av;m =fim in 2
д on I lEN
Wim-Brati- ViVi,m = 0
on Th leD.
The conclusion of this lemma follows by setting
Oi,m = Milli,m Ui.mo
We are now able to state the key result of this subsection.
 4.4 THE FREDHOLM ALTERNATIVE 229

Theorem 4.4.3.3 Under the assumption (4,4,3,1) and when

(9; -O; +1+2w;/9)/m
is not an integer for any j, the dimension of the space of all solutions
in La(R) of problem (4,4,1,3) is

4(9) =ji+leD
E card met:
+I card mez
jor i+leN
when D is not empty and u (R) + 1 when D is empty.

Proof Let u € La(R) be solution of problem (4,4,1,3) and consider any

fixed corner S;. We apply Theorem 4.4.2.4 and Lemma 4.4.3.2 in the
related disc D of radius p. It turns out that

Ci.mOm-F Cim p.m - cos (1; mo; + P; +1) € H'(D)

071,.m >-2/a
(4,4,3,5)
with Cm = 0 in the particular case when A;m = 0 and j and i + 1 belong to
D.
We shall now show that the series in (4,4,3,5) belong to H'(D) for
every disc of radius p, < p. Indeed, let us denote this series by w. We have

aw = 7 Gimp'm-1 cos (1; mO, +8, +1)

1 до==
and consequently

k.m >0 vw.

Then due to inequality (4,4,2,10), V w is bounded in D1; indeed we have

Vw(r.ei%,)is &
2
1, 550 Va; p^.m
and this last series is convergent since p, 5 p.
In other words, we have

U- GimO;.m EH'(Di).
071, m>-2/9
230 SECOND-ORDER PROBLEMS IN POLYGONS

Such a smoothness result holds near each of the corners S;. Then, with
the help of Lemma 4.4.2.1, we conclude that
U-
iSN C;mOim E H'(1)
0=1,m>-2/a
(4,4,3,6)

where Ci.m1 = 0 for 1; mг = 0, when both i and i + 1 belong to D.

To end the proof, let us denote by & the function in (4,4,3,6). It is a
solution of problem (4,4,1,3) and in addition it belongs to H'(9). Thus
Lemma 4.4.3.1 shows that ¢ = 0, unless D is empty, where & is a constant
K. In other words, we have
U= Ci.mOjim + K
j=1,..,N
021,m>-2/9
where K=0 unless D is empty. The statement of Theorem 4.4.3.3 is an
easy consequence. I
Then, with the help of Theorem 4.4.1.6, we can derive a bound for the
actual dimension of Na
Corollary 4.4.3.4 Assume that (4,4,3,1) holds and that
(8; -0; +1+2w;/9)/o
is not an integer for any j. Then when p <2, the dimension of Na is less
than or equal to

if D is not empty and (R) + 1 if D is empty. When p>2 the dimension of

Na is less than or equal to
(R) - card {j M; is parallel to Mi+1}
if D is not empty and (R) +1 again if D is empty.

Observe that when D is empty, we are just dealing with a pure

Neumann problem, owing to (4,4,3, 1).

Proof So far, we have shown that the O,m 1 <j&N, -2/9 < 1; m <0 (if j
and i+1€ D), -2/9<1; im SO (if j or i+ 1€N) are a basis of Ma (possibly
up to the constant function).
We shall first show that any Qi.m corresponding to 1;,m =0 does not
belong to N. Due to assumption (4,4,3,1) Am can vanish iff i and
j+1eN (and consequently ; =Ф;+1=0). The corresponding Oj.m is
 4.4 THE FREDHOLM ALTERNATIVE 231

eliminated by condition (4,4,1,7) (see Theorem 4.4.1.8). Indeed we have

(in the polar coordinates related to S;)
Oj.m =nlog r;/ (w;) to,
where u€ H'(8), n(0) =1, n depends only on r; n(r;) = 1 for r; SPio
n(r) = 0 for r; >Pe where O < Pi < Re are chosen in such a way that the
support of n does not meet I for k+ i and i+ 1. Then, in condition
(4,4,1,7), we can choose 4; = n. Accordingly, this condition reduces to

Actually we have

since both n and u belong to H'(R), n has a small support around S, and
both n and & fulfil a Neumann boundary condition on I; and I;+1.
On the other hand, we have

An(n logy)dx dy= An(nlogr) dx dy

'n'
where O' = 2\r, <pi), since An vanishes in M\M'. We can apply again
Green's formula since both n and n log r; are smooth in S'. We thus get

I An(n logr) de dy

where y = d2\d?. It follows that

since O.m is harmonic.

Finally, we have

and this contradicts the condition (4,4,1,7). Accordingly, O.m does not
232 SECOND-ORDER PROBLEMS IN POLYGONS

belong to N.. Consequently, the dimension of Na is less than or equal to

v(R) when D is not empty and to v(R) + 1 when D is empty.
To complete the proof of Corollary 4.4.3.4 we observe that any 0;,m
corresponding to 1; m = -1 is eliminated from Na by condition (4,4,1,8) or
(4,4,1,9) in Theorem 4.4.1.6 when p is greater than 2. The calculations
are very similar to the previous one, so that we do not need to repeat it.
The condition Am =-1 for one integer m is fulfilled iff M; is parallel to
M: +1. 6
This result, together with Lemma 4.4.3.1, allows us to calculate the
index of A as an operator from E, to Ez (these spaces have been defined
in Subsection 4.4.1). We shall also be able to conclude when p = 2, due to
the inclusion N2S Na which holds for g <2.
For that purpose, let us again use the polar coordinates with origin at S;
and let us consider the functions

cos (1; mO; + D; +1) Mi(r, el%,) (4,4,3,7)

with 1; m <0, not an integer. Here are some properties of these functions.

Lemma 4.4.3.5 Sim EH' (8)\ W. (2) for

-= 51; m <0, 1 5jEN, Aim+-1

and in addition
ASin
"an on r, ifleN
(VISi,m =0 on Is if leD.
This is obvious. The following statement deserves a proof.

Lemma 4.4.3.6 Assuming that (4,4,3, 1) holds, AS;, im is not orthogonal to

Na for
2 m <0, 15j&N, 1;m‡-1.
-=51;
Proof This can be proved by contradiction. Thus, if we assume that Si,m
is orthogonal to Na, then there exists Wim € W5(8) fulfilling the boundars
conditions in (4,1,1) such that
Awi.m =AS;.me
 4.4 THE FREDHOLM ALTERNATIVE 233

Therefore Wim , - Sim is a solution of the homogeneous problem and

belongs to H'(1). By the uniqueness result of Lemma 4.4.3.1, this
implies that Sim belongs to W5(8). This contradicts Lemma 4.4.3.5. m

We are now able to conclude.

Theorem 4.4.3.7 We assume that (4,4,3,1) holds and that
(9; -9; +1+2w;/9)/IT
is not an integer for any j, that in addition M; is never parallel to M; +1,
when p=2. Then for each f€ Lp (M), there exist unique real numbers C.m
and a unique u such that
И- CimSi,m€W=(8) (4,4,3,8)
1SiEN
-2/9<1, m CO
1,.m #-1
and u is solution of problem (4,1,1) when D is not empty. Otherwise, when
D is empty u is unique up to an additive constant and exists iff

faxdy=O.
Proof The functions AS; .m corresponding to

_3 <Aim <0, 1;,.m #-1, 1=1, 2, ..., N

9
are in L (R) and are clearly linearly independent. Since they are not
orthogonal to N they do not belong to the image of E, through 4.
Moreover, their number is exactly the upper bound for the dimension of
N. (possibly minus one when D is empty) that we found in Corollary
4.4.3.4. Consequently, Lp (S) is the span of the image of E, through A
and of these functions AS,m The claim follows by Lemma 4.4.3.1.

One could ask why there is a gap in the index of the problem
corresponding to the eigenvalue Aim = -1. Actually, there is no longer
any gap when we consider nonhomogeneous boundary conditions:

Corollary 4.4.3.8 Under the assumptions of Theorem 4.4.3.7, let f€

1p (1) and g; E WE VP (Ti), jED, 8; E WH- HP (F;), jEN be siven such that

g: (S: = à&i+ (S.) if i EN and j+ 1 € 8 or 8;11(5,) = A§i (S;)

A fli Aflit1
ifj€D and j+1 eN, (4,4,3,9)
234 SECOND-ORDER PROBLEMS IN POLYGONS

whenever u; is parallel to u;+1, and p>2. Then assuming that D is not

empty, there exist unique real numbers C,m and a unique u such that
(4,4,3,8) holds and u is solution of
Au =f, (R)
i€D
Yiu = 8j>
du 2
- Vill = Sis jEN.
Vi avi
When D is empty the condition (4,4,3,9) is void and u is unique up to the
addition of a constant and exists iff

Jo rax dy -E J, 8, do =0.
This result follows from Theorem 4.4.3.7 and the trace theorems in
Subsection 1.5.2. We observe that the number of extra conditions that we
have added on the data in (4,4,3,9) is exactly

j=1

4.4.4 The Fredholm alternative for nonvariational problems

Here, we try as far as possible, to deal with problem (4, 1, 1) in most cases.
The existence and uniqueness result of Lemma 4.4.3.1 has been a basic
tool in the study that we carried out in Subsection 4.4.3. Unfortunately, if
we drop the assumption (4,4,3,1), it may happen that problem (4,1,1)
could not be solved uniquely in H'(R). This will make our analysis much
more complicated.
On the one hand, we still have an existence result in H-(S), which is an
application of a lemma in Lions (1956). We recall this result with a
slightly different proof.

Lemma 4.4.4.1 Let W and V be a pair of Hilbert spaces with a continu-

ous injection of W in V and let a be a continuous bilinear form on V× W.
Assume that there exists a constant a >0 such that
(4,4,4,1)
for all we W. Then for every continuous linear form I on V, there exists
u€V, possibly non-unique, such that
a (u; v) =1(v) (4.4.4.2)
for every we W.
 4.4 THE FREDHOLM ALTERNATIVE 235

This lemma is somewhat similar to Lemma 2.2.1.1 and is actually a

consequence of it.
Proof For € >0, we introduce the form
a (4, 0) = a (u; 0) + E (U; U) W. u, rEW.
This is a continuous bilinear form on WX W, which, in addition, is
coercive (with coerciveness constant ≥&). Consequently, by Lemma
2.2.1.1, there exists a unique ug € W such that
a. (Me, w) =1(0) (4,4,4,3)
for every v€ W.
Using the coerciveness assumption on a and setting v = u in identity
(4,4,4,3), we find bounds for u:

Consequently, we have
(luellv Sa'"Hillva,

Due to the famous property of bounded sequences in Hilbert spaces, we

can find a sequence & i = 1, 2, ... converging to zero, together with u € V
(clearly we cannot expect u to be unique in general) and we W such that

S(VE;)U,
", -* u weakly in V
- w weakly in W.
Going back to identity (4,4,4,2), we have
a(us, 0) +VE; (VE; We, ; d)W = 1(0)
for every ve W. Taking the limit in i proves identity (4,4,4,2).

Lemma 4.4.4.1 will be applied as follows. Again, as in Subsection

4.4.3, we set
V={uEH' (8) | viu =0, VjeD}.
Then we set
W={u€H (8)Iv;u =0, VjeD and u(S;) =0, Vil;
this is a Hilbert space for the norm of H3(2). Finally the form a is
defined by

(4,4,4,4)
236 SECOND-ORDER PROBLEMS IN POLYGONS

It is easy to check that a is well defined and continuous on V X W, since

for u€ V and v€ W, we have
д
and
v;u є H"?(Г),
aT;
д
дт;
The coerciveness of a in the sense of (4,4,4,1) follows obviously from
Poincaré's inequality when D is not empty. When D is empty, we must
replace everywhere V by VIC, where C denotes the space of constant
functions in 2.
Consequently, given fe L, (1), there exists at least one u € H' (1) such

that alu; o) =-1 fodxdy (4,4,4,5)

for every v € W (provided Sof dx dy = 0 when D is empty).

We must now make it clear in what sense such a u is the solution of
problem (4,1,1). Obviously, we show that
Au =f in 32,
by writing (4,4,4,5) with U € D(R). Therefore, u belongs to the space
E(A; Lp (8)) defined in Subsection 1.5.3. Consequently, y; du/av; is well
defined as an element of A (F*. Then, applying the Green formula
(1,5,3,9), we deduce from (4,4,4,4) and (4,4,4,5) that
stand;.
iEN дт; jEN

for every ve W (which is a subspace of the space of possible test-

functions in Theorem 1.5.3.11). In other words, this identity holds for
every
jEN.
† Observe that under assumption (4,4,3,1) the forms defined by (4,4,3,3) and (4,4,4,4)
coincide, since
1д def
lat, дт
for every & and We H12(Г.).
 4.4 THE FREDHOLM ALTERNATIVE 237

This is enough to prove that

ди
- V;U = 0, jEN.
Vi avi
Summing up, we have proved the following statement.

Lemma 4.4.4.2 Assume that D is not empty, then for every given f€
Lp (R), problem (4,1,1) has a (possibly nonunique) solution u€ H'(R).
When D is empty, the same result holds provided

fdx dy =0.
Our main trouble now is that we have no uniqueness result in H'(1).
However, we have results in some particular cases, if we assume in
addition that u is slightly more regular, namely u € W(M) with p > 2.
In the first particular case, we assume that D is empty and that
B2≥...≥B;≥B;+1≥... =ß1, (4,4,4,6)
Lemma 4.4.4.3 Let u€ W. (R) with p> 2 be the solution of problem
(4,4,1) with f=0. Assume that (4,4,4,6) holds, then u is a constant.

This will be proved as usual, by calculating the integral of Au against u

On 8. Unfortunately, this cannot be done directly and we must approxi-
mate u by a sequence of smoother functions. This is the purpose of the
following auxiliary lemma, whose proof is similar to that of Lemma
1.5.3.9 and Theorem 1.5.3.10.
Lemma 4.4.4.4 D(R) is dense in the space
FIA; L. (2)) = {u € W(R); Au ELp (8)}
equipped with the norm
ur>lulli.p.st|Aullo.p.s.
In addition ur> y; dulav; has a continuous extension as an operator from
F(A; Ip (8)) into W. "P(T.).
Proof of Lemma 4.4.4.3 We let Um m = 1,2, ... be a sequence of
functions in W5(8) such that
Um -> U in W(8)
" -> Au in L. (1)
where m -> 0.
238 SECOND-ORDER PROBLEMS IN POLYGONS

FOr um € W.(M) the usual Green formula holds. Thus we have

vi avi д
dum +Biari

+58 43 151-431514

2
+ § B, - B, I u= (S.).
We can take the limit in m of this identity, due to the fact that p is strictly
larger than 2. Thus we get

1'ar
du + Biari 1 2
§ Bi - Bi +1 u? (S.).
(4.4.4.7)
We observe that the bracket on I; is meaningful since

and
ди • 2
+ B:
Vi avi
for 1 = 1,. .,N.
Actually the same identity holds with u replaced by u - u(S,). Thus we
get
- sulu-u(S,)) dxdy

+ § B, - B, - Su(S) - u(S, 13.

i=2
 4.4 THE FREDHOLM ALTERNATIVE 239

Since u is harmonic and fulfils the boundary conditions in (4,4,1), we

finally conclude that

j=2
Since by assumption (4,4,4,6) we have
B; -B;+1≥0
for j= 2, ..., N, it follows that u is a constant function.
Another useful particular case is this: We assume that D = {3, ..., N}
and that
B, =B2. (4,4,4,8)
Again we have a uniqueness result for solutions in W(8) with p > 2.
Lemma 4.4.4.5 Let u€ W. (1), with p> 2, be the solution of problem
(4,4,1) with f = 0. Assume that (4,4,4,8) holds, then u is zero.

Proof Again identity (4,4,4,7) holds for u. Thus we have

0= Vuldx dv+ 2
and consequently u is zero.
We shall now study the space Ma. First we observe that the analogue of
Lemma 4.4.3.2 holds in the most general case.

Lemma 4.4.4.6 For each j and each A;m EJ-2/9, 0] there exists 0; m € Ma
such that
Wim =Oj.m -Mi,mU;.mE H' (1)
if 1;.mr <0 or if Mi,m = 0 and j and i+ 1 do not both belong to D.

Proof This is quite similar to the proof of Lemma 4.4.3.2, since there we
only used the existence result in Lemma 4.4.3.1. The corresponding
existence result is now provided by Lemma 4.4.4.2. _

However, we can improve this result due to the fact that O is not a limit
point of the set ;.m I mEl}.
Lemma 4.4.4.7 There exists p > 2 such that
Wim EW:(1)
240 SECOND-ORDER PROBLEMS IN POLYGONS

for each A;.m €J-2/9, 0], provided (Qk - Pk+1 + 2wk/9) / m is not an integer
for any k.
Proof The function Wi,m is one solution of the homogeneous problem
near each corner Sk. In addition it belongs to La(R) and it is smooth in
M\S,. -.., Si} by Lemma 4.4.2.1. Thus it follows from Theorem 4.4.2.4
that
Wim (rk@'°) = - cOS (1 k. 10k + k+1) + K
141 5-219 VAR AR.1

for rk small enough, provided (k - k+1 + 2w/q)/m is not an integer.

Consequently, by Lemma 4.4.4.6, we have

-COS (1k. Ok + Pk+1) E H'(D) (4,4,4,9)

where D is A Mirk < p} for p small enough.
Now inequality (4,4,2,10) implies that

is bounded as I-> +o. It follows that

cOS (1 k. Ok + Pk+1) € W; (D,) (4,4,4,10)

for each D, = AMir <pit, where p, <p. This implies, by difference

between (4,4,4,9) and (4,4,4,10), that
Ck., rake!
- 214 514, 50 VOOK Ak1

Consequently we have
Ck.1 = 0 for Ax,, €0.
Summing up, we have

cOs (1k, 10k + Pk+1)

1450 Var Ak1
near Sk and by (4,4,4,10) this shows that
Wi.m E WA(DI)
for inf {Ark A,k > 01 > 1-2/p. This is true near each corner Sk and
cOnsequently we have
Wi.m € W: (1)
for some p >2.
 4.4 THE FREDHOLM ALTERNATIVE 241

We are now able to calculate the dimension of M in two particular

cases 'adjoint' to the cases considered in Lemmas 4.4.4.3 and 4.4.4.5.

Theorem 4.4.4.8 Assume that D is empty, that (9; -Ф; +1 + 2w;/9) /I is

not an integer for any i and that
tan $25 ... Stan $, Stan Pi+1 S ... Stan P~ Stan P.
Then the dimension of Na is less than or equal to

M(8) = ≥ card {mez|

i=1
Proof Let veMa. Then o€ La(g) and is a solution of the problem
(4,4,1,3). By Lemma 4.4.2.1, we know that v is smooth, far from the
corners. Then near each corner Si, U has an expansion given by Theorem
4.4.2.4. In other words, we have

1= cOS (1; mo; +Ф; +1)

in D= AMir, <p} for some p> 0.

Again here, due to inequality (4,4,2,10), the series

cos (1;, me; + 9; +1)

belongs to W(D). This, together with Lemma 4.4.4.7, implies that

Ci,.mOjim € W.(D).
-2/9<1,m SO
Since dim € W#(M\D), it follows that
W=v- Ci,mO;,m € Wa(82)
1=12 N
-2/9<1,.m 50
for some p > 2.
Now w is solution of the homogeneous problem (4,4,1,3). Applying
Lemma 4.4.4.3 we see that w must be a constant. Finally, the function
O;,m corresponding (possibly) to A;,m n = 0 is eliminated from N. by condi-
tion (4,4,1,7) of Theorem 4.4. 1.8 as in the proof of Corollary 4.4.3.4.

The same method of proof, with Lemma 4.4.4.3 replaced by Lemma

4.4.4.5, leads to the following statement.
Theorem 4.4.4.9 Assume that D = {3,4, ..., Ni, that
(9; -Qi +1+200;/9)/TI
242 SECOND-ORDER PROBLEMS IN POLYGONS

is not an integer for any j and that

tan $, Stan $2.
Then the dimension of Na is less than or equal to

Now, exactly as we did in Subsection 4.4.3, we shall derive existence

results in the space spanned by W(1) and the functions Sim correspond-
ing to -2/q < A,m <0. Indeed, the result of Lemma 4.4.3.5 holds in the
most general case. The analogue of Lemma 4.4.3.6 is the following.

Lemma 4.4.4.10 AS;,m is not orthogonal to Na for - 21q €1;.m <0,1 sjs
N, 1;,.m # - 1.
Proof Actually, we shall prove that
1 (4,4,4,11)
Aj,m

Indeed we have Oj,m = M,m + W;.m where

4i,m =(Vw;) Aj.m cos (1;, mO; + Q; +1) ni

in the polar coordinates with origin at S; In addition, both Sim and Wim
belong to W(8) for some p>2. This allows us to apply the classical
Green formula; thus we get

= E tan $, (Sim Wim) IS., = 0.

1=

This is due to the boundary conditions

d
aSim+tan diatiVIS;m =0
On Th
d
dWim-tan O - YIWim = 0
 4.4 THE FREDHOLM ALTERNATIVE 243

to the properties of the support of Sim and to the obvious fact that
Si,m (S;) =0.
Since Adj.m = 0 we have

(4,4,4,12)
Then, we cannot apply directly the Green formula to Sim and Vi,m
because Vim is singular at S;. Thus, we are led to introduce

where p is chosen such that n;(re) = 1 for r <p. Since the support of
AS;m is contained in S', we have

We can now apply the classical Green formula in S', since both O;,m and
Him are smooth in ¿'.
Let us denote by I' the intersection of I with an' and set
y =Ir, ei%, r; = p,0<0; <0;3.
We have

dim do
1=1 Lam ar;
1=1 )-5
due to the boundary conditions on Sim and on 4, m i.e.
dUi.m-tan dr Villi.m =0 on It.
" Эт.

Here we denote by Ay the origin of I' and by B, the endpoint of Th

according to the positive orientation. We have also set

Ar; dim-SimMW.
[asim ar; m' dor.
244 SECOND-ORDER PROBLEMS IN POLYGONS

Due to the properties of the supports of Sim and Mms it turns out that

AS,milk.m dxdy-{ SimAm dx dy

(4,4,4,13)
00;13,m w0;17m
Finally, we calculate & explicitly. This is elementary, and we get

§= 1 sin P; cos ; - sin O;+1 COS Pj+1

(4.4.4.14)
Aj.m
The identity (4,4,4,11) follows plainly from identities (4,4,4, 12) to
(4.4.4.14). -
Corollary 4.4.4.11Assume that the hypotheses of Theorem 4.4.4.8 are
fulfilled. Then for each feLp (R) such that Safdx dy = 0, there exist real
numbers G.m and a function u such that
И- Gi,m.Si.m € W=(S)
1SjEN
-2/a<1., <0
1,.m #-1
and u is a solution of problem (4,1,1).
Proof This is a simple consequence of the fact that L (R) is the space
spanned by the annihilator of Na, the constant functions and the functions
AS;,m corresponding to the eigenvalues such that

Clearly these functions are linearly independent (this follows from their
explicit definition in identity (4,4,3,7)), do not belong to the annihilator of
Na by Lemma 4.4.4.10 and span a subspace of L (S) whose dimension is
suitable by Theorem 4.4.4.8.
Replacing Theorem 4.4.4.8 by Theorem 4.4.4.9, we obtain the follow-
ing statement.
Corollary 4.4.4.12 Assume that the hypotheses of Theorem 4.4.4.9 are
fulfilled. Then for each feLp (R) there exist real numbers Gim and a
function u such that
И- GimS;.m 1 E WE(R)
ISjEN
-2/9<1.m <0
1, m#-1

and u is a solution of problem (4,1,1).

 4.4 THE FREDHOLM ALTERNATIVE 245

Now with the help of Theorems 4.4.3.7 and Corollaries 4.4.4.11 and
4.4.4.12, we reach our final goal.

Theorem 4.4.4.13 We assume that (P; - P;+1 + 2w;/9) | m is not an integer

for any i. Then for each f€ Lp (1) (such that Infdx dy = 0 when D is
empty) there exist real numbers G,m and a function u (possibly non-unique)
such that
И-
-2/9<1,.r n <0
1,.m #-1
and u is the solution of problem (4, 1,1).
Proof We start from one solution u € H'(8) to problem (4,1,1); such a
solution exists by Lemma 4.4.4.2. Then we study locally the behaviour of
u. Since we have
Au eLp (S),
it follows plainly that nue W. (8) for every n € D(D). This describes the
smoothness of u inside 12.
Then let us look at the behaviour of u near the regular points of the
boundary. For that purpose we let n€ D (R) have a support which does
not meet I; for I+ j. Since u€ H'(R), we have
AnueL, (8) +L2(8)
and

a +Biari lav;
Choosing a plane open subset S' with a CI. boundary such that an' 2 r;
and such that S' contains the support of n, we see that

AMUEL, (R) +13(8)

д
Jut Bi at
Therefore, we have nue H°(8) + W. (1) by Theorem 2.4.1.3. Varying n
this shows that u€ H(R\V) + W5(21 V), where V is any closed neigh-
bourhood of the corners of M. Accordingly, it follows that

(ras
Applying again Theorem 2.4.1.3, we see that nue W(8). Varying n this
shows that
246 SECOND-ORDER PROBLEMS IN POLYGONS

Finally, let us study the behaviour of u near one of the corners, say S,.
We shall use one of the model problems studied before. For that purpose
we introduce new boundary conditions on li 3<j≤N, as follows:
First case 1€ D or 2€ D: we set
L; = I, j=3, ..., N.
Second case 1 and 2€N, tan , > tan 2: we set
д j=3,..,N
L;: du;
with tan $2 Stan $ Stan '+1 S•.. Stan 1, j = 3, ..., N.
Third case 1 and 2€N, tan O, Stan Pz: we set
L; = I, j=3, ...,N.
In all cases, we have nue H'(R) and
AnueL, (12)
v:L;niu =0 on lj 35j<N
) V;MIU = 0 on Fir jeDn<1,2}
д- ViN.U = 0 on Fir jeNn<1,23.
Vii AVi
This is a problem that we have already solved in Theorem 4.4.3.7 in the
first case, Corollary 4.4.4.11 in the second case and in Corollary 4.4.4.12
in the third case. Accordingly, there exists v € H'(8) and constants C,m
such that
U- -2/9<11.m<0
C1.mS1.m E W3(8nV) (4,4,4,15)

where V is a neighbourhood of S, and v is the solution of the same

problem as nu. In other words, we have
MIU-vEH'(R)
4 (Miu-v) =0 in 2
on Ijs j€ DN<1, 2}
д4) (114-0) =0 on li jeNn<1, 2}.
The last step is to apply Theorem 4.4.2.4 to miu-v. This shows that
n. u -v can be expanded as follows near zero:

cos (11 mO, +Ф2) + K,

Van
11m <0^ 11,n
 4.4 THE FREDHOLM ALTERNATIVE 247

where

for some L and p. Consequently, we have

CmS1,m EWARMV). (4,4,4,16)
-2/9<11.m <0
11.m #-1
Adding (4,4,4,15) and (4,4,4,16) we see that
MU- (Cm +CI,m) SI,mEW(AMV).
-2/9<11,m< 1
Aj.m*-1
A similar result holds for niu near S; for each j and this completes the
proof of Theorem 4.4.4.13.

We conclude with a statement concerning the nonhomogeneous bound-

ary value problem.

Corollary 4.4.4.14 Under the assumptions of Theorem 4.4.4.13, let

fEL, (1) and g; E WE-1P (In), jED and gi E WE- 1/P (I;), j EN be given such
that
g;(S;) =agit! (S;) ifieN and i+1€8
du;
Or

gi+1(S;) = agi_ (S,) if jeD and itIEN

Or
8;+1(S;) =M; " Mit1,

whenever M; is parallel to Mi+1 p> 2, and 8; (S;) = 8;+1 (S;) if j€ D and

it 1€ D. Then there exists a function u and numbers C,m (possibly non-
unique) such that
И- GimS;.m € WE(12)
1siSN
-2/9<1,.m <0
Aim #-1
and u is solution of problem (4,1,1).

Remark 4.4.4.15 In this whole chapter we have excluded the domains

with cuts (i.e. w; =2m for some i) for simplicity. However, if we allow
cuts, the basic a priori inequality of Section 4.3 remains valid (see
Remarks 4.3.1.7 and 4.3.2.7). The main tool in Section 4.4 has been the
248 SECOND-ORDER PROBLEMS IN POLYGONS

Green formula in a Lipschitz domain. To handle domains with cuts

requires derivation of the corresponding Green formula. This can be
achieved by using the trick, described at the beginning of Section 1.7, of
considering separately the restrictions of the functions to + and 8_.
Accordingly the results of Theorem 4.4.4.13 and Corollary 4.4.4.14 hold
for domains with cuts (i.e. if we allow w; =2m in the statements).

Remark 4.4.4.16 In the particular case of self-adjoint boundary condi-

tions (i.e. either Dirichlet or Neumann) along a cut the results mentioned
above may be easily deduced from the Theorem 4.4.4.13. Indeed, to
make an example, let us assume that c; =2m and j, j+ 1 € D. Then by
looking at the problem locally one can assume that & is symmetric with
respect to I; A rotation and a translation reduce the problem to the
particular case when S; =0 and I, together with I+1 lie on the positive
x-axis.
Now let us write u as the sum of an even function u and an odd
function lo with respect to y:

4. (I, y) = U(X, y) + W(X, - 12).

Assuming that f€ Lp (M), 8; =0 and 8+1 = 0 imply that uo fulfils a

homogeneous Dirichlet on the axis y =0 in a neighbourhood of O.
Therefore we have
(4,4,4,17)
where V is a suitable neighbourhood of O. In the same way ue fulfils a
homogeneous mixed boundary condition near O:

(Me (x, 0) =0, 0<x<8

ID, 4. (x, 0) =0, -8<x<0
for some 8>0. Accordingly, by Theorem 4.4.4.13 there exist constants
Ci.m such that
Gimt, m+12 sin (m -b)O, E WE(VMS). (4,4,4,18)
1/2-2/q<m <1/2
By adding (4,4,4,17) and (4,4,4,18) we obtain the behaviour of u near 0.
This is an alternative proof of the results stated in Remark 4.4.4.15. The
same method allows one to handle the following cases:
(w; =2T, jED, itIEN or jEN, j+1ED
jEN, i+1EN.
5
More singular solutions

5.1 Behaviour of the derivatives of order higher than two

In this section, we look for u€ Wk+2(2), where k is a nonnegative

integer, which are solutions of the same boundary value problems as in
Chapter 4. However, we shall also consider non-homogeneous boundary
conditions. In other words, we shall try to find necessary and sufficient
conditions on the functions f and 8, 1 j&N, ensuring that the following
problem should have a solution u belonging to Wk+2(5):
Au =f in R
Vil = 8; on I j€D (5,1,1)
ди д
+ 3; - jU = 8; on Fir jEN.
(Var
Here we keep the notation of Chapter 4.
Some necessary conditions are obvious. Indeed, if there exists u€
Wk+2(8) which is a solution of (5,1,1) we must have
fEW: (S),
8; € WK +2-1/P(E;), jEDD and 8; EWK+1-1/P(E;), jEN.
In addition, we must have
(a) 8; (S;) = 8j+1 (S;), if and it 1 € g (5,1,2)

(b)
duit1
88. (S,) = 8,41(5) if jED, 1 + 1EN, Min, is parallel to 7;
and k+1> (2/p)
(c) g;(S;) =agit! (S;) if jeN, j+ 1eD, M; is parallel to T;+1
du; and k+1> (2/p)
(d) 8; (S,) = 1 -'g. +1 (S,) if jand j+1 EN, M; =1 p;+1 and k + 1> (2/p).
249
250 MORE SINGULAR SOLUTIONS

(When p=2 and k + 1 = 1, the pointwise conditions in (b)-(d) must be

replaced by the corresponding conditions with integrals.)
As we saw in Chapter 4 in the particular case when k =0, some
additional orthogonality conditions occur. They were the orthogonality of
f to the space Na. Here we shall find many more orthogonality conditions
of the same nature when k is ≥1. However, for several special values of
the measure of the angles of S2, we shall find additional conditions which
generalize (5,1,2). This makes the study of higher derivatives of u a lot
more difficult than the study of the second derivatives that we carried out
in Chapter 4.
The technique that we shall use here is mainly based on the following
idea. We shall extensively use the trace theorems of Chapter 1 to reduce
the general problem (5,1,1) to the particular case when g; =0 for every j
and f€ Wk (1). In this particular case, the given function f fulfils a lot of
unnatural homogeneous boundary conditions. However, due to these
boundary conditions on f, it turns out that the derivatives up to the order
k of the corresponding solution u are solutions of boundary value
problems for the Laplace operator of the same kind as (4,1,1). Accord-
ingly, we shall take advantage of the results proven in Chapter 4 to find
the behaviour of these derivatives of u near the corners.

5.1.1 Special data

Let us look in a first step, at u € W5(8) which is a solution of (5,1, 1)
under the assumption that g; = 0 for every j and that f€ W(1). In other
words we have
Au =f in 2
Vil = 0 on Is j€D (5,1,1,1)
ди д on F jEN.
• Yiu = 0

We already know that u is smoother far from the corners.

Lemma 5.1.1.1 Let u€ W5(8) be the solution of problem (5,1,1,1) with f

given in W(R); then we have

for every closed neighbourhood V of the corners of 22.

Proof First let n € D(R); we have

- A(nu) +nu =-ñF -[A; nJu + nue WI. (RE)
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 251

and nu € W=(R?). Accordingly

ñu =E*f-nf-[A; nJu+nu}
where E is the elementary solution of -4 + 1 defined by
FE(5) = [1+1517', gER?.
Then Theorem 2.3.2.1 shows that nu € W. (1).
Now let n€ D(R) be such that the support of n meets only the interior
of one side, I; say. Then let w be any plane open set with smooth
boundary such that w contains the support of n and F'Adw C I;. With the
obvious modification when j€D, we have

anú

and it follows from Theorem 2.5.1.1 that

ñú E WE(w).
The claim of Lemma 5.1.1.1 follows by partition of unity on §IV.
Now we consider any unit vector ^= (a; B) and the corresponding
derivative of u, i.e.
ди
1) =as -aD,u + BD, u.

This is obviously a function belonging to W(1) which is a solution of

ar
Av =- DIE Lo (82).
The boundary conditions on v are the following.

Lemma 5.1.1.2 Assume that u€ W(R) is the solution of problem

(5,1,1,1) with f given in W|(1). Let x; be the angle from r to 1; then

do + tan (4, -x,) d- YU = 0 on I; (5,1,1,2)

Vi avi дт;
when (D;-Xi)/t-> is not an integer, and
yo=0 on I; (5,1,1,2')
when (D;-x;)/t-| is an integer.

Proof By possibly performing a rotation of the coordinate axes, we can

252 MORE SINGULAR SOLUTIONS

assume that I; is supported by the axis {y =0} and that 2 is 'above' I;

Accordingly we have
5; = (1,0), V; = (0, -1)
and the boundary condition for u on I; is
- (cos $;) V;D,u + (sin $;) D*V;u = 0. (5,1,1,3)
Differentiating with respect to x, we get
- (cos $;,) Dev;D,u + (sin $;) D}y;u = 0. (5,1,1,4)
On the other hand, the assumption that feW! (52) implies that
(5,1,1,5)
From the definition of x; it follows that
v = (cos X;) Dxu + (sin X;) D,u.
Accordingly, we have
dU
=- (cos Xi) DAY;DyU - (sin x;) y; DZu
Vi avi
= (COS Xi) D< v;u + (sin X;) D. V;D,u
diari
and finally

do + sin (4-x) d
дт;
= - (cos $;) D‹V;D, u - sin x; cos (Ф, -Xi) y;DZu
+ cos X, sin (D; -X;) D<y;u.
Using (5,1,1,5), we deduce that

dU + sin (8, -x) d= YiU

=- (cos $;) DeV;D,u + (sin D,)D? v;u;
this last expression is zero by (5,1,1,4). This is the desired result when
(Ф; -X;)/m -2 is not an integer. Otherwise we have
cOs X; = & sin P;sin X; =-e cos O;
where & is either + 1 or -1. Accordingly, we have
y;0 = E{(sin A;) D+v;U - (cos P;) V;D,u}
and this is zero due to (5,1,1,3).
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 253

We are now able to apply the results of Chapter 4. However, a side

difficulty arises here. We shall not be able to apply Theorem 4.4.4.13
because of the possible non-uniqueness in that statement. Accordingly,
we shall try to make use of Theorem 4.4.3.7, which deals with variational
problems only. This is why we first have to localize our problem.
For this purpose, we introduce a cut-off function related to S, as
follows: n; € D(M), n; is one near S;, the support of n; does not meet I,
for I i, i + 1 and
(5,1,1,6)

if (P;+1-Ф; -w;)/m is not an integer.

Then we look at
ди
U;=др;
in other words, we choose & = u;/ly; in order to study the behaviour
of u near S. By Lemma 5.1.1.2, we have (the boundary condition is
plainly Vi+10; =0 on Ti+1 when ($;+1 -Ф; -w;)/m is an integer)
on Г;
(110, =0
all than (94 1-491-57-40) 574..
Vi+10; =0 on F+1.
(5,1,1,7)
The boundary conditions on the other sides will not matter, since we
finally introduce

By Lemma 5.1.1.1 we know that we W(R) and that

д - f + In; A JU; € Lp (12).
Aw; =Miage;
The boundary conditions on w; follow from (5,1,1,6) and (5,1,1,7).
Accordingly, we have (see the note preceding (5,1,1,7))
on I I+i+1
vw, =0
Yi+ I W; = 0 on Fits.
(5,1,1,8)
Assuming p ≥2 so that w, € H'(M), it follows from Theorem 4.4.3.7
254 MORE SINGULAR SOLUTIONS

that there exist numbers Cm' such that

W;--2/g<1. 1. m' CO
Ci,m'Si.m'€ WA(V;) (5,1,1,9)

where V; is a neighbourhood of S; in & and

wi
Si r;- 1'
(Vw:)1"
This holds provided A.m#-2/9 for all m.
We can now easily derive the following result.

Proposition 5.1.1.3 Let u€ W3(8) be the solution of problem (5,1, 1, 1).

Assume that f€ W"(R) and p =2. Then there exist real numbers Ci.m such
that
И- Ci.mS;.m€ Wa(V;) (5,1,1,10)
-1-2/9<1, m <-1
1.m #-2

where W; is a neighbourhood of S; and 1;.n n= (D;+1 - O; + MIT) /w; (m an

integer),
r. -^,.m
Si.m =(vw;)dj.m cos (1;, mO; + ;+1)

provided 1,.m#-1-219 for all m.

Proof We merely integrate (5,1,1,9). Indeed w; coincides with du/д;

near S;. We can rewrite (5,1,1,9) as follows:

2др;lu--1-2/9<1,. Cimsim) E WE(V,), (5,1,1,11)

1, man 5-1
for some real numbers Cm• Then let us consider any vector §; orthogonal
to M, and having the same length I= IM;. We shall show that

3 (u- CimS,m) E W?(V,). (5,1,1,12)

-1-244579 3 5 1

Indeed, let us set

C.mS;.me
-1-21951, 5-1
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 255

We already know from (5,1,1,11), that

ga+Bi (5,1,1,13)

for a + B 53, provided a > 1. We just need to check that

233 € Lp (Vi),
dS;
since it is clear from the assumptions on u and the Aim that we W(V;)
and accordingly 4, def/as;, 274a53€ In (Vi).
For this purpose we observe that
44 =fEW. (Vi)
since Sim is harmonic near S; It follows that
234 1 af
due to (5,1,1,13). The claim (5,1,1,10) is an obvious consequence of
(5,1,1,11) and (5,1,1,12). -
We are now able to reach the main purpose of this subsection.
Theorem 5.1.1.4 Let f be given in W(9) with p=2 and Infdx dy = O
when D is empty. Assume that (S; -O; +1 + w; + 2w;/9) / I is not an integer
for any j. Then there exists a function u (possibly non -unique) and numbers
Cim such that
Ci.mSi.mE W.(R)
-1-2/9<1, m CO
1,.m #-1,-2
and u is solution of problem (5,1,1,1).

Proof We can choose pi >p such that (P; - Q;+1+2w;/91)/o is not an

integer for any i where 9, is conjugate to Pi. Since f€ Lp, (8), we can
apply Theorem 4.4.4.13 with p replaced by pi. Thus we know the
existence of a function u and of numbers Ci,m such that
CimSim€ WA,(8)
I=jEN
<0
-2149,47.
and u is solution of problem (5,1,1,1). Since the functions Si.m are all
solutions of the homogeneous problem corresponding to (5,1,1,1), we can
apply Proposition 5.1.1.3 to & near each of the corners S;. The smooth-
ness of u far from the corners follows from Lemma 5.1.1.1. &
Iterating the above procedure yields the following result.
256 MORE SINGULAR SOLUTIONS

Theorem 5.1.1.5 Let f be given in WK(5) with p>2. Assume that

Ф; -O; +1 + kw; + 2w19
T
is not an integer for any j. Then there exists a function u (possibly
non-unique) and numbers Cim such that
И- CimSin.€ Wk +2(0)
- k-2/9<1,.m <0
1,.m #- 1,-2,..., -k
and u is solution of problem (5,1,1,1).

5.1.2 A trace theorem

We now want to solve the general non-homogeneous boundary value
problem (5,1,1). This will be achieved by reducing the general case to the
particular case which we solved in Section 5.1.1. For this purpose we
need to find the necessary and sufficient conditions on the functions
fEW'(R) and 8, € WK +2-110 (Ti), jED, 8; € WK +1 - 119(F), jEN which en-
sure the existence of a function u€ Wk 2(8) s.t.
on In j€D
(an,
d
- yiU = gi on Fir jEN
dU +Biati
(5,1,2,1)

and
Ao-fe WK(S).
In other words, we are looking for a function > which fulfils the
boundary conditions (5,1,2,1) and
2' a'f on I; 0515k-1 (5,1,2,2)
diaviAu = Vi au"
(see Remark 1.5.2.11). This trace problem will be solved with the help of
Theorems 1.6.1.4 and 1.6.1.5 (and Remark 1.6.1.8).
In a first step let us define the operators Bi. which we shall need to
apply Theorem 1.6.1.4. The first operator B; will be either I when j€ D
or d/Au; when jeN, while
211-2
B;,= dv; A, 1=2,.., k+1.
Accordingly, the order of B; is either zero or one, while the order of Bit
is 4, 1=2,..., k+ 1. In order to be able to apply Theorem 1.6.1.4 we
must find all the operators Pi. and Qj+1, fulfilling condition (1,6,1,1).
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 257

Since we are dealing with a polygon (in the strict sense) and the operators
Bi. are homogeneous with constant coefficients, we can restrict ourselves
to looking for operators Pi. and Qi+1,1 also homogeneous and with
constant coefficients.
We shall first look for the operators Pi. and Qi+1,1. We must have (by
1,6,1,1))
1-2 1-2
lavit-) -p./2
av
Pi. B;.1-Qi+1,1B;+1,1= 1=2
(5,1,2,3)
We observe that
k+1
1=2
d 11-2
and
1=2
can be any homogeneous differential operators of order d-2 when dis
the order of both sides of (5,1,2,3). Indeed Qi+1, and Pi. are tangential
operators to I'+1 and I', respectively. Accordingly, the identity (5,1,2,3)
means that the symbol of
Pi. B. -Qi+1.,B, +1.1
can be divided by the symbol of 4.
Since P.. and O,+1.1 are tangential to I;- and I, respectively, it turns
out that

(2)"
LaT,
if i€g
P1.1
d)d-1 if jEN
(5,1,2,4)

b.laTit1
d if i+1€9

-'''
Qi+1,1 (5,1,2,5)
if i+lEN.
дт; +1

Lemma 5.1.2.1 There exists real numbers a; and b; s.t. P,., B,1-
Qit1, Bi+1,1 can be divided by A where Pi.1 and Qi+1,1 are defined by
(5,1,2,4) and (5,1,2,5) and such that

a?+b?‡0
iff (P;+1 -Ф; - dw,)/t is an integer.

Proof Let us look for instance at the particular case when both i and
258 MORE SINGULAR SOLUTIONS

j+1 belong to N. Accordingly, we must be able to divide the symbol of

a;(a)d-s 2-bi,/
Aul;
d d-1 a
dui+1
by the symbol of 4. Equivalently we must be able to divide the
polynomial
a; (-1)d-'(x cos w; + y sin w,)d-'(I-x sin w; + y cos w;]
-tan Q, [x cos w; + y sin w;]) - b;xd-"(- y + tan D;+1*) = p (x, y)
by x?+ y?. This means that x = ‡iy are roots of the polynomial p(x, y).
Writing p(tiy, y) = 0 leads to the following system of two equations in the
two unknowns a; and b;:
a; (-1)d-'(sin w; ‡i cos c;)d-" (cos w; Fisin c;]
-tan $;[sin w; #i cos w;D - b; (ti)d(-1 titan $; +1) = 0.
This system is equivalent to the following
a; (-1)d-'eTio,dI1 Fitan $;]+ b;[1 Fitan $; +1] = 0. (5,1,2,6)
The determinant is proportional to
é to, 4(1 -itan $;)(1 + i tan Q;+1) - e, (1 + i tan $;)(1 - itan $; +1)
=e iwd (1 + tan $; tan $;+1 + i[tan $,+1 -tan $; 1)
-ew, (1 + tan $; tan P;+1 - i[tan $;+1 -tan $; 1)
= (1 + tan $; tan $;+1)e to,4(1 + i tan [Ф; +1 -Ф;])
- etiw,d(1-itan [Ф;+1 -Ф;1)}

cos [Ф;+1-Ф;]

= 2i(1 + tan $, tan d, -1) Sin IQ, + 1 - 4, - w,d]

Obviously this determinant is zero iff (P;+1-Ф; -w;d)/t is an integer.
Similar calculations yield to the same result when j or i + 1 belong to D.
The system (5,1,2,6) is replaced by
a;(-1)d-'eFido [ti]-b;[1 Fitan Ф;+1]=0 (5,1,2,7)
when j€D and i+lEN,
a; (-1)d-'eFido, [1 Fi tan $;]Fib; =0 (5,1,2,8)
when jEN and i+1€D and
a;(-1)de Fidw, -b; =0 (5,1,2,9)
when i and i+1€9. H
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 259

Remark 5.1.2.2 When the determinant is zero, the space of the solu-
tions of the systems (5,1,2,6) to (5,1,2,9) is one-dimensional. We shall
completely determine a; and b;, by assuming in addition that b; = 1.
Accordingly, we have (setting m = (Q; +1 - O; - dw;)/t):
cos P;
1-itan $; cOs @j+1
when i and itleN,

cOs @j+1
when jeD and j+1EN,

a; =(-1)deido,
when eN and it 1€D and
a; = (-1) deidw, =(-1)d+m
when i and j+ 1 € D.
Let us now introduce one more notation. We denote by Rid the
differential operator (homogeneous and of order d-2: i.e., Rid = 0 if
d <2) s.t.
d-1 2 _ д (5,1,2,10)
-)
dT; du; laTit
when i and j+1EN,

allaT ) - jd-1 a = R;,aA (5,1,2,11)

-)
when jeD and i+1 EN,

a дтa d-1 a •д -) = R, ad (5,1,2,12)

Aft; (дт;+1
when jeN and j+1€D, and
d •d
(5,1,2,13)
a; aT; laT; +1
when j and i + 1 eD. These identities are nothing but identity (1,6, 1, 1) in
the particular case that we study here.
Accordingly, Theorem 1.6.1.4 implies the following result.

Theorem 5.1.2.3 Assume p+2 and let feWK(8) and g;€

260 MORE SINGULAR SOLUTIONS

Wk+2-1/P(Ty), jeD, gE Wk+1-1/P(Th), jeN be given. Then there exists a sol-

ution U € Wk+2(8) of the boundary conditions (5,1,2,1) and (5,1,2,2) iff
the following equalities hold:

34 ' 9 (59) - (5,1,2,14)

di aTi arti
gd-'Bit! (5) = Riaf (S,)
when j and j+1EN,

28 (S.) -ga-'git= (5.,) = R;,af(S,) (5,1,2,15)

diaTi OT;+1
when jeD and j+1EN,

dud-1
'8'(5)-J"Bit (5.) = R; af (S,) (5,1,2,16)
di aT AT:+1
when iEN and it 1€@ and

(5,1,2,17)
4.53 (5) -3дті+1
°811 (S.) = R; af(S,)
when j and j+ 1€D, for all d€ [0, k + 2/g[ and j s.t. (D;+1 - D; - dw;)/t is
an integer. (d>1 when jor j+1€N).
Proof We just have rewritten the identity
k+1
Pi. 18; (S;) -Qi+1,18;+1(S;) =
2=2f
1=20.. 12 (S.)
'86 a (591 - 2 P. do.
as
Pi.18;(S;) -Qi+1.18;+1(S;) = Ri.af(S;)
for the corresponding value of d.
The similar result when p = 2 follows from Theorem 1.6.1.5.

Theorem 5.1.2.4 Let feH* (8) and 8, E Hk+31(5)), jE D, 8; E Hk +12(5;),

jeN be given. Then there exists a solution o € Hk+2(1) of the boundary
conditions (5,1,2,1) and (5,1,2,2) iff equalities (5,1,2,14) to (5,1,2,17)
hold for all de[0, k+1[ and i st. (S;+1 -O; - dw;)/t is an integer
and provided
Ф; +1-9; -(k +1) W;

is not an integer for any j.

This last provision is made to avoid possible identities (5,1,2,10) to
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 261

(5,1,2,13) corresponding to the order d = k + 1. Such an identity would

yield a condition with an integral similar to (1,6,1,3).

Remark 5.1.2.5 The identities (5,1,2,10) to (5,1,2,13) corresponding to

either d = 0 or d = 1 are just (5,1,2).

5.1.3 More singular solutions

We first infer the consequences of Theorem 5.1.1.5 and Theorem 5.1.2.3
or 5.1.2.4.

Theorem 5.1.3.1 Let feW$(8) and 8; € WK +2-1/P(Fi), jED, g; €

Wk+-1/P(E;), jeN, be given. Assume that (9; +1-9;-Kw; -2w;/9)/T is
not an integer for any j and that the identities (5,1,2,14) to (5,1,2,17)
hold for all de[0, k+1[ and all i such that (S;+1-D; - de;)/t is an
integer. Then there exists a function u (possibly non-unique) and numbers
Gim such that
и- CimS;mEWK+2(82)
- k-2/9<1,.m <0
1, in #-1, -2... ,- k

and u is solution of problem (5,1,1,1).

Indeed, let u be given by Theorem 5.1.2.3 or Theorem 5.1.2.4; we just

have to apply Theorem 5.1.1.5 with f replaced by f-Av, since this
function belongs to Wk(8).

Remark 5.1.3.2 The number of extra conditions (5,1,2,14) to (5.1.2.17)

on the data off and gi, 1 - j&N, is exactly the number of eigenvalues A ;,m
which are excluded from the sum which describes the singular behaviour
of u, by the condition
dim #-1, -2, 0.., -K.
For practical purposes, the identities (5,1,2,14) to (5,1,2,17) are not
easy to check on functions given explicitly. This is due, in particular, to
the fact that we did not attempt to find the operators Rid. We shall rather
try to understand the particularly singular behaviour of the solution u
which occurs when one of these identities is not fulfilled. We need a
preliminary result.

Lemma 5.1.3.3 Lets be the function

s(r; 0) =p-^,. In r cos (1;, mO + D; +1) + A sin (1; mO + D; +1)]
where A;.m = (P; -D;+1+ MIT) w; is assumed to be a negative integer. Then
262 MORE SINGULAR SOLUTIONS

s is harmonic for r > 0 and 0 € JO, w;L. Furthermore, we have

ds= 0 for r >0, 0=0
185 +tan Pith ar
r 2A
Is=o for 120, 0-0 if cost. =0
(1 as for r>0, A=w;
r 20 cos 9;
Is =(-1)'; sin dir-1,m for r >0, 0=w; if cos 9; =0.
This can be verified directly. Now let us consider a cut-off function N;
similar to the one we have used in Chapter 4. In other words, we have
n; € D (R), n; = 1 near S,; the support of n; does not meet I, for It i, j+ 1,
and

ani
Aft;= 0
on I; if jeN
ani =0 on I;+1 if i+IEN.
Let us then set
Gi.m(re'*) =r;^,mIn r; cos (1;,mO; + Q; +1)
+ 0; sin (Ai.mO; + D; +1)]n; (r, ei°,). (5,1,3,1)
This function has the following properties:
AGim=fimEC"(8)
where fi.m is zero near all the corners and
д -Gim = Bi,m.kE C° (Ik)
Ak avk
for keN and

for k € D. In addition gi,m.k is zero for all k, but k = j and k = j+ 1, and we

have

Si.m.j= cos 9; =*m 1 if iEN

if jeD
 5.1 DERIVATIVES OF ORDER HIGHER THAN TWO 263

near S;, while

gi.m,j+1=0

near S;
Accordingly, the necessary condition at S; in Theorem 5.1.2.3 is not
fulfilled. On the other hand, it is easy to check that
Gin
iff s<-Aim+ 1, while S;im$ H-^, m+' (1).
The conclusion of these preliminaries is the following, where for
convenience, we introduce a new definition.

Definition 5.1.3.4 We define the function Sim as follows:

Gim(rei%,) =r;, ..cos (1; mO; + D; +1) n; (r, ei%,)
when Aj.m m is not an integert and
Gim (r,el,) =r,", " In r; cos (1;, mO; + D; +1) + 0; sin (1;, mO, + D; +1)}n; (r, elo,)
when Aim is an integer.

Theorem 5.1.3.5 Let feWK(R) and

8, € WK*2-119(5,), jED,
be given. Assume that (8; +1 - P, - kw; -2w;/g) / m is not an integer for any j
and that g, (S;) = 8;+ (S;) whenever j and j+1 eD. Then there exists a
function u (possibly non-unique) and numbers ki.ms such that
И- kimE;.mEWK+2(8)
-k-219<1,.m CO

and u is solution of problem (5,1,1,1).

Proof We shall apply Theorem 5.1.3.1. For each j,m such that A i,m is an
integer belonging to the interval ]k -2/g, OL, let us define Ki,m as
follows:
ad-1
ki,m yd-'gimi (S,)- Bi.m.it1 (S.)

49 5 1 3 (5) - *dTi+1
84 ' (5) - R, 145,1
† Observe that here G.Jim is just a relabelling for (Va, , mS,.
264 MORE SINGULAR SOLUTIONS

when j and j+1EN, d= (Q; +1 - 8, - MIT) /w; We define Kim in a similar

fashion (mutatis mutandis) when j or it 1 € D.
It is clear that
f' =f- -k -2/9<1,.m <0
kj.mfi.m
A,m integer
and
g! = g; - §
- k-2/9<11m <0
15jEN,
Mi,m integer

fulfil the assumptions of Theorem 5.1.3.1. Consequently, there exist a

function u' and numbers Ci,m such that
"' - - k-219<1, m <0GimS,m EWK+2(12)
1,.m #-1,-2,...,-k
and u' is a solution of
(Au' = f' in 2
v;u' =81 on Th j€ D
du' • d
- yu' = si on I; jEN
We conclude the proof by setting
u=ut-k-219<1, m <0
Am integer

and Kim =Ci.m/(Vw;)1; mo

Remark 5.1.3.6 The solution u in Theorem 5.1.3.5 is unique when

there is uniqueness in the space H'(8) (see Section 4.4.3) or in the space
W; (8) with r > 2 (see Section 4.4.4).

Remark 5.1.3.7 Again one can also handle the domains with cuts by
applying the same techniques (see Remark 1.7.4 in connection with the
trace results of Subsection 5.1.2). Therefore the results in Theorem
5.1.3.5 still hold if we allow w; = 2 for some j. For instance in the case of
a Dirichlet problem on both sides of a cut (i.e. w; = 2m and jED, j+1€
D) the singular solutions are the following:
Gim =r;? sin (m0,/2)n;(r, ei°j)
when m is odd and
Sim =r'/[In r, sin (m0;/2) + 0; cos (m0.12)} n. (r,ei'j)
when m is even.
 5.2 OPERATORS WITH VARIABLE COEFFICIENTS 265

5.2 Operators with variable coef cients

A natural continuation of the study carried out in the previous sections

would be to investigate boundary value problems with variable coeffi-
cients in a plane domain whose boundary is a curvilinear polygon. The
simplest idea is to apply the well-known perturbation method to reduce
such a problem to similar problems involving only homogeneous
operators with constant coefficients. This method will enable us to extend
only part of the preceding results to problems with variable coefficients.
Here just to illustrate such a method, we shall restrict ourselves to the
study of a Dirichlet problem. Thus we will also avoid a lot of side
difficulties which have nothing to do with the specific problem of singular
behaviour of solutions near the corners.
The data are the following. We consider a plane bounded open domain
A whose boundary I' is a curvilinear polygon of class C1.1 (see Definition
1.4.5.1). Thus T= U1=1 1; where I; is an open arc of curve of class C1,1
and I; meets It at S;. The measure of the angle of the tangent vectors
to I; and It at S; (toward the interior of M), will again by denoted by
w;, 1 5j< N. Next, we consider the elliptic operator A defined by

Au = I D, la, D, u) + I a Din+ anu.

i, = 1
where a.; = a;i€ C' (R) (it is sufficient throughout this section to assume
that aj€ W(R) for some p > 2) a; € L °°(M), 0 €i =2. The ellipticity of A
means the existence of a > 0 such that

2 a: (x) 55, 5-a 1512 (5,2,1)

i. = 1
for all x€8 and $€1?.
Our first purpose is to calculate the index of the operator A from
W5(9)n W'(S) into Lp (M). For simplicity, we assume that the corres-
ponding boundary value problem has a unique variational solution in
H'(M). This can be achieved by assuming for instance that
min ao(x)>
1 max a; (x)12. (5,2,2)
XEn
i =1.2
Consequently, applying Lemma 2.2.1.1, it is easily seen that for any given
feLp (9), there exists a unique solution u € H'(R) of the equation
Au =f in 12. (5,2,3)
In addition it follows from the results in Subsection 2.5.1, that
fi
266 MORE SINGULAR SOLUTIONS

for every closed neighbourhood V of the corners. Thus we just have to

investigate the behaviour of u near the corners.
Before stating our main result, we need to introduce one more nota-
tion. We consider the operator A; obtained from A by freezing the
coefficients of its principal part at S;:

A;u =È a, (S,)D, Diu.

k.1 = 1

Then let us consider one matrix J; such that

-J;A;T, = I
where A; is the symmetrix matrix whose entries are the ak, (S;), 1 € k, I€
2. We clearly have
(A;u) (F-'x) =- (40,) (x)
where U;(x) =u(T;'x).
Definition 5.2.1 We denote by w: (A) the measure of the angle at I;S; of
JiR.
We are now able to calculate the index of A.
Theorem 5.2.2 Assume that 2w;(A)/Trq is not an integer for any i; then
the image of WE(R) through the mapping
T: U->{Au; y;u, 1 <jEN}
is a closed subspace of codimension

È card fm -7< mIl

j=1
in the space
Z=I:g; 1=jEMIfEL, (8), 8, EW=- 1P(I), 1 5 j S N, 8; (S;)

This general result will be deduced from a sequence of lemmas of

technical character. In these we are going to deal with the particular case
when M has only one corner. This is possible since we now consider
curvilinear polygons. One can think of the cross section of a wing, for
instance. We shall refer to this particular case as the case when N= 1.
Lemma 5.2.3 Assume that N = 1; then there exists a constant C s.t.
(5,2,4)
for all u € WE(R)n W! (n), provided 2w, (A) / Tra is not an integer.
 5.2 OPERATORS WITH VARIABLE COEFFICIENTS 267

Figure 5.1

Proof This will be derived from inequalities (2,3,3,1) and (4,1,2), with
the help of the same technique that we have already used in Subsection
2.3.3.
First we select a neighbourhood W of S, and a change of variable
U: W-> W with the following properties:
(a) A is of class C1.1,
(b) the Frechet derivative of I at S, is the linear operator defined by
the matrix TI
(c) " (I'M W) is the union of two straight segments with origin at US.
Then we choose a cut-off function m with support in W s.t. n is
identically equal to one near S. We shall look separately at nu and
(1-n)u. Set
0(x) = (nu) (W'x),
and select any plane open domain w with a polygonal boundary such that
(a) @CUW,
(b) w contains the support of U,
(c) dw coincides with U Al near US,.
It is clear that vE W(w)n W(w) and that

- 10 + I b,,D, D,0 + E b,D, 0 +660=8

in w, where
g= (Anu) oU-1
268 MORE SINGULAR SOLUTIONS

and where biE CO. ' (w), b; EL'(w), 0 si =2 and in addition:

bi.r(US.) =0.
We apply inequality (4,1,2) to v. This is possible since the angle of wat
US, is w(A) and we have assumed that 2w(A)/trq is not an integer. It is
always possible to choose the other angles of w so as to avoid the
exceptional cases for inequality (4,1,2). Accordingly, we have

i.i=1 110, p, c

It follows that

i, = 1,2
Xesupp nor -1

Now since bi, vanishes at US, we can choose the support of n small
enough so that
тах
i, = 1,2
bií(x)1€1/8C.
xEsupp nor -1

Accordingly, we have

Going back to u, this implies that

nul2p.asC@Anullo.p.atnulli.p.o) (5,2,5)
with possibly another value for C.
Then we can choose another plane open domain, O' with a C.
boundary, such that the boundary of ' coincides with the boundary of h
out of the set {n = 1} where n is equal to one. Accordingly, we have
(1 - n)uE W$(S') M WH(S') and applying inequality (2,3,3,1), we have
161 - n)ul2.p.aCIlA(1-n)ullo.p..+|11-n)ull.p.p. (5,2,6)
Adding inequalities (5,2,5) and (5,2,6), we obtain the estimate
lulla.p.rsCillAullo.p.nt/lulli.p.st. (5,2,7)
On the other hand, a direct integration by parts shows that
lull.an SC Aullo.p,r. (5,2,8)
The inequality (5,2,4) follows from (5,2,7) and (5,2,8) with the
help of inequality (1,4,3,2). _

The next step is the following.

 5.2 OPERATORS WITH VARIABLE COEFFICIENTS 269

Lemma 5.2.4 Assume that N=1 and that the boundary of 2 is

rectilinear near S,. Then the image of W3(1)n W! (1) through A is a
closed subspace of codimension [2w/malt in Lp (M), provided 2w,/ mq is
not an integer.

Proof Let feL. (g) be given and let u€ H'(R) be the solution of
-Au = f in 2. (5,2,9)
It is clear from the results in Chapter 2 that
uEW=(QI W)
where W is any closed neighbourhood of Si. Again let n be a cut-off
function which is equal to 1 near S. We have
(1 - n)uE W=(1),
and if we choose the support of n small enough, nu is solution of the
Dirichlet problem for the equation
- Anu = nf- [4; n]u = f1
in a plane open domain w, whose boundary dw is a polygon which
coincides with A near S. It is clear that f1 € Lp (w) and, applying
Theorem 4.4.3.7, we know that there exists numbers Cm , such that

nU- miMTIw, sin

W1
0<mT/w, <2/9

where W is a neighbourhood of S, in w.
Adding these results, we have

U- 0<m/w,<2/9Cmimi/w, sin W1
The numbers Cm are continuous linear functionals of f, and consequently
also of f through the Green operator
#:f-> u
defined by (5,2,9). Accordingly, we have u€ W3(8) n W|(8) if f
annihilates [2w/ mq] linear functionals on Lp (R).

Lemma 5.2.5 Assume that N=1 and that the boundary of A is

rectilinear near S,. Then the image of W7(1)n W'(1) through A is a
closed subspace whose codimension is [2w(A)/ mq] provided 2w (A) / Tq is
not an integer.
† [SI denotes the integral part of S.
270 MORE SINGULAR SOLUTIONS

Proof Let us first look at the particular case when A, =-4 or equival-
ently
Ak. (S,)=8k.b 15k, 152.
Then we shall derive the result by homotopy from -4 to A. Let us set
A (1) =1A - (1-1)4, tE [0, 1].
Applying Lemma 5.2.3, we know that for each t € [0, 1] there exists a
constant C, such that
ul2.p.o€CA(t)ulo.p.r
for all u€ W5(0)n WA(8). Accordingly, A(t) is a semi-Fredholm
operator from W5(8)n W'(8) into Lp(8) for every t € [0, 1].
The operator A(t) depends continuously on t. Thus by a theorem in
Kato (1966), the index of A (t) does not depend on t. Consequently, the
index of A(1) = A is equal to the index of A(0) =-4, which is
-[2w/Ilq], by Lemma 5.2.1.4. This completes the proof of Lemma
5.2.5 when A, =-A, since A is one to one inequality (5,2,4)).
The general case is reduced to the particular case when As=-4 by
composition with the matrix It.
Proof of Theorem 5.2.2 Let us start from feLp (8). There exists a
solution u € H'(2) of
Au =f in h (5,2,10)
and in addition u € W5(52 \V), where V is any closed neighbourhood of
the corners. This was observed earlier. For further convenience we
denote by & the continuous linear operator from L (8) into H'(R)
defined by (5,2,10).
Let Mi be a cut-off function whose support is small near S; and such
that n; = 1 in a neighbourhood of S. Obviously, we have
An;u = nif + [A; ni]If = f, ELp (w;)
and niuE W(w;) n W. (wo;), where w; is an open plane domain whose
boundary is a curvilinear polygon of class C'. with only one corner at S;,
which contains the support of n; and such that the boundary of c;
coincides with the boundary of & near S
Finally, we select a change of variable ; defined in a neighbourhood
W, of S; such that
(a) U; is of class C1.1
(b) the Frechet derivative of U; at S, is the identity operator U
(c) A;(FM W) is the union of two straight segments with origin at U;S;.
If the support of n; is small enough, we can also choose w; small enough
 5.2 OPERATORS WITH VARIABLE COEFFICIENTS 271

to be contained in W Under these assumptions, we can apply Lemma

5.2.1.5 in the domain U;w;
The conclusion is that niu € W(w,) iff f, annihilates [2w;(A)/ Ta] con-
tinuous linear forms on L,(w;), 1 5j&N. Accordingly, u€ W3(8) iff I
annihilates
[200, (A)7
V=
i=1 Tra
continuous linear forms on Lp (R), since
ft = nif+ [A; nilaf.
Thus we have shown that A is a one-to-one mapping from W(1) n
W'(R) onto a closed subspace of L. (1), whose codimension is v. This,
together with a trace theorem (Subsection 1.6 implies the claim of
Theorem 5.2.2. -
Remark 5.2.6 The general principle underlying Theorem 5.2.2 is
that we can easily extend the index property of our boundary value
problems from the case when all the operators have constant coefficients
to the general case. However, for practical purposes, one also needs to
know which singular functions must be added to W(R) in order to get
surjectivity. The perturbation method used here will allow us to conclude
only in some particular cases when the singular functions remain the same
when passing from the constant coefficient case to the general case.
Theorem 5.2.7 Assume that 2w; (A)/T is not an integer and that
p <w;(A)/(w;(A) -Tr) (5,2,11)
for all j s.t. w;(A)> tr. Then for every (f; 8; 1 <j"N) given in Z, las
defined in the statement of Theorem 5.2.2), there exists a unique function
u and unique numbers Ci,m such that
И- CimS,mE W7(8),
0 <m <2w, (A)/tg
j= 1,2,...N
and
{Au =f inonRI; 15j≤N,
(vill = g.
where
maO;
S: im
n (I; 'x) = AnTi/W, (A ' sin
c; (A) Mile, eit,). +
+ We identify x with r,e', in J,R. where r, A, are the polar coordinates of the corner J,S,
of J.M; ie., J.S is the point r, =0, while the lines tangent to J;I' at TS, correspond to
0, =0 and 0; = w, (A) respectively.
272 MORE SINGULAR SOLUTIONS

Proof Clearly the functions Sim belong to H'(2) W(8) when O<
mT/w;(A) <2/g. We shall show that
AS;mELp (R), YeS;m E W3-HAIIk), ISKEN,
provided condition (5,2,11) is fulfilled.
Indeed, we have

where bk. E CO. (J,D), be € L. (J,D), O Sk 52 and

bk.'(T;S;) = 0.
From the above definition of Sims it follows that
(AS;m) oF; '(x) = 0(AMI/09, (A) -1)
near J.S, while ASin im is smooth in J; (DIS;). This shows that
AS;.mELp (R).
On the other hand, we have viS,me W3-19(Ik) for all k, because
YKS;m =0 for k# j, j+1, while when k is j or j+1 we can apply the
following Lemma which follows easily from the definition of the space
W3-19(10, aD.
Lemma 5.2.8 Assume that 4 € C. (I0, a]) with a > O and that 4 (0) =
$'(0) =0. Then the function u = ° belongs to W3-1/9(10, aD) provided
a>1-1/p.
Here, choosing the coordinate axes suitably, we can assume that I;r; +1
is the graph of a function o fulfilling the assumption of Lemma 5.2.8.
Then we have
Vit 1S,mI; '(x, y) =1(x27 4 (x)3)Mar/W, (A) sin mt
lo;(A) arctan
near the origin. Accordingly, YitS,m belongs to W5-1/P(I;+) iff p<
w;(A)/(w;(A) -T) when w;(A) > mI, m = 1, 2,.
A similar proof shows that y,S, mE W3-19(I;).
Summing up, we have shown that the mapping T (defined in the
statement of Theorem 5.2.2) maps the space E spanned by W-(8) and
the functions Sim (1 <j&N, 0 <m < 2w;(A) / m9) into Z. The codimension
of WE(S) in E is obviously

V=
[20;(A)
 5.2 OPERATORS WITH VARIABLE COEFFICIENTS 273

Consequenty if follows from Theorem 5.2.2 that T is an isomorphism

from E onto Z, since T is one to one on ES H'(R).
Remark 5.2.9 In the particular case when M is a strict polygon (but A
still has variable coefficients), we can replace the condition (5,2,11) by
the weaker one

P< 20;(A)

Indeed we still have AS; mE L. (1), while IS,m = 0 for all k.

Remark 5.2.10 Starting from the results explained in Remark

4.4.1.14, one can also apply the techniques of this section when w;(A) =
27. This takes care of domains with turning points toward the interior of
M (see Section 3.3 for the treatment of turning points toward the exterior
of 1).
Remark 5.2.11 A technique for calculating singular solutions of other
boundary value problems has been worked out in Mghazli (1983). This
technique allows one to handle second-order elliptic boundary value
problems for nonhomogeneous operators (i.e. including lower-order
terms).
6
Results in spaces of Hölder
functions

6.1 Foreword
All the results in the previous chapters have been stated in the framework
of the Sobolev spaces described in Chapter 1. The basic reason for using
such spaces was explained in Section 1.1. However, one is mainly
interested in statements claiming that the solution of a given boundary
value problem has continuous derivatives up to a certain order. Such a
property cannot be derived directly due to the bad behaviour of the
kernels involved in the maximum norm (see Section 1.1 again). This is the
reason why the classical property of continuity of some derivatives of the
solution has been derived indirectly through the use of the Sobolev
imbedding theorem of Subsection 1.4.4.
Another approach to the continuous differentiability of the solution of
a boundary value problem consists in using spaces of functions with
derivatives up to a certain order (say m a nonnegative integer) which are
Hölder continuous (with exponent o a real number between zero and
one). Of course these spaces are closer to the classical spaces of continu-
ously differentiable functions than the Sobolev spaces. However, they
have few nice properties besides the ones which are obvious from the
definition. Precise statements are given in Section 6.2. Fortunately a very
nice multiplier theorem, similar to Mikhlin's theorem (Theorem 2.3.2.1)
for the Lebesgue space Lp (R"), holds for the Hölder spaces. As it turned
out in Chapters 2 and 4, the multiplier theorem was the basic tool for
proving the a priori inequalities. Accordingly similar a priori inequalities
hold in the framework of Hölder spaces. They will be proved in this
chapter and the corresponding regularity (or singularity) results will be
derived.
To conclude this introductory section, let us mention some references
about regularity in Hölder spaces. A wide set of results is derived in the
classical book Miranda (1970) which deals with second-order problems in
a domain with smooth boundary. Some of these results are extended to
274
 6.2 A BRIEF REVIEW OF HOLDER SPACES 275

problems of higher order in Agmon et al. (1959). Second-order problems

in domains with corners are studied in Volkoff (1965a,b), Azzam
(1979, 1981) and Moussaoui (1971). The first two authors restrict their
purpose to the Dirichlet problem. Their results are included in the
present chapter.

6.2 A brief review of Hölder spaces

In this section, after defining precisely the spaces under consideration, we
shall review their basic properties. In doing this, we shall follow the same
plan as in Chapter 1 for the Sobolev spaces. Here & denotes any open
subset of R".
Definition 6.2.1 Let m be a nonnegative integer and o a real number such
that 0 <o & 1. We denote by C. (R) the space of all functions u defined in
R whose derivatives, up to the order m, are continuous and bounded in S
and whose derivatives of order m are uniformly Hölder continuous with
exponent o.

We define a Banach norm on Cm. (1) by setting

la Em XEN x-yo (6,2, 1)

where s = m + o. Observe that this definition includes the case when
2= Rn
In order to be able to describe the traces on the boundary I' of 1 of
such functions, we need a definition for similar spaces on I. This requires
some smoothness assumption on I. Precisely we assume that 2 is
bounded and that its boundary is of class Ck, with k+17s=m to. We
use the same notation as in Chapter 1; in particular $ is defined by the
identity (1,3,3,1), where & is described in the Definition 1.2.1.1.

Definition 6.2.2 Let h be a bounded open subset of R" with a boundary of

class Ck.!, , where k is a nonnegative integer. Let To be an open subset of F.
A function u defined in To belongs to Cm.°(To), m a nonnegative integer,
0 €J0,1], 5 = m+o sk+ 1, iffu o DECM. (V'N@ '(Ton V)) for all possi -
ble V and & fulfilling the assumptions in Definition 1.2.1.1.

It follows plainly from this definition that u belongs to C. (I) iff u is

Hölder continuous with exponent o in the usual sense. In addition, it is
clear that the trace ul of a function u € Cm. (S) is well defined and
belongs to C. (I'). A converse statement will be derived later.
276 RESULTS IN SPACES OF HOLDER FUNCTIONS

Some of the properties of the Sobolev spaces have analogues which are
just obvious; among them are the following. First Cm.«(R) and Cm.°(F)
are algebras for the usual multiplication. Next the differentiation operator
Di, 1 si&n, is a continuous mapping from Cm. (R) into Cm-1.«(p).
Finally the natural imbedding of Cm.«(R) into Cm." (R) (and of CM.«(Г)
into Cm' ' (I)) is compact provided m' + o' <m + o; this is a simple conse-
quence of the well-known Ascoli theorem.
On the other hand some of the most useful properties of the Sobolev
spaces have no analogue at all for the Hölder spaces. For instance, there
is no convenient density result. Indeed it is easy to check that any
function u in the closure of D(R) for the norm of Cm. (R"), has the
following extra properties:
(a) D°u(x) ->0 when xl->+00, for lasm
(b) D"U(x) - Du(y) _>0 when |x-yl->0, for la| = m.
The main reason for introducing these Hölder spaces is the following
multiplier theorem.

Theorem 6.2.3 Let a € Ch (R) be such that there exists a constant C with
D'a(E) SC(1 + El)-lal (6,2,2)
for all geR" and la|&n. Then the operator
g->F-'aFg
is continuous in CM.° (R") for all m (a nonnegative integer) and o € 10, 1L.

A short outline of the proof may be found in Meyer (1978) while a

detailed proof is given in Triebel (1978).
Let us now focus our attention on the continuation property. The case
of a Lipschitz condition, i.e. the case when o = 1, is very peculiar. For
instance, if M is bounded, any function u € Cm, '(S) is the restriction to D
of some function U€ Cm. (R"). This very strong result may be found in
Schwarz (1965) for instance. Unfortunately, since Theorem 6.2.3 ex-
cludes the case when o = 1, we shall mainly use the spaces with 0<0 <1
in studying boundary value problems. There exists a Hölder version of
Theorem 1.4.3.1: under the same assumptions, Ps maps Cm. (R) into
Cm. (R), where s = m + o. It is hard to give a precise reference for this
result. However, the proof is just the same as the corresponding proof for
the Sobolev spaces with Theorem 2.3.2.1 replaced by Theorem 6.2.3. The
following statements, whose proofs are easy, are sufficient for our pur-
pose.
 6.2 A BRIEF REVIEW OF HÖLDER SPACES 277

Theorem 6.2.4 Let M be a bounded open subset of R" with a boundary I

of class CM.!. Then there exists a continuous linear operator Pm+1 from
cm.o (5) into Cm.° (RM) for every o€ J0, 1], such that
(6,2,3)
Outline of proof This is quite similar to the proof of Theorem 3.9,
Section 3, Chapter 2, in Necas (1967). The first step is a reduction to the
case when & is replaced by a half space $ defined by X >0. This is
achieved through the use of local coordinates near the boundary and a
partition of unity. The second step consists in defining U by
u(Xy..., Xm), In =0
U(Xy,..., Xn) =3m+1
1
assuming that
m+1
1 =§ (-i) *Nisa = 0,1, ..., m. (6,2,4)
It is very easy to check that U€ CM,«(R"), when u is given in CM. (R#),
and that Un" = u.
Theorem 6.2.5 Let S be a bounded open subset of R? whose boundary I
is a curvilinear polygon of class Cm.'. There exists a continuous linear
operator Pm+1 from Cm. (5) into CM.° (R3), for every o€ JO, 11, such that
Pm+1ula =u
for every u€ Cm." (1).

Proof This is basically a repeat of the previous proof. Again we use a

partition of unity and local coordinates. According to Definition 1.4.5.1,
the problem is reduced to one of the following cases:
(a) M is a half space and we proceed as in the proof of Theorem 6.2.4;
(b) 8 is a quadrant, defined by x, > 0, *2 > 0 and we define U as
follows:
V(XI, X2), X2≥0
m +1
U(xy, X2) = 2 1; V (x1 -ix2), x250
1
where
u(Xy, x2), *, ≥0,*2≥0
V(Xy, X2) =
21;4 (-ixs, +2), *150, 8230.
278 RESULTS IN SPACES OF HÖLDER FUNCTIONS

assuming again that (6,2,4) holds. Clearly U€ Cm.« (R'), when u is

given in Cm. (R) and Ula = u.
(c) & is the complement of a quadrant, defined by x, ≥0 or x2 ≥ 0. Here
we proceed by steps. We start from u € Cm. (R) and we denote by u,
the restriction of u to the half plane defined by x, ≥0. Then we

define V, by (4, (X1, x2), *1=0

m+1
Vi(x1, X2) =
2 1;4, (-ixs, X2), 4150

still assuming that (6,2,4) holds. Clearly Vi belongs to CM. (R?) and
V, =u for x1=0.
Next we set w=u-V, in the half plane defined by x2≥0. We have
wE CM. (R) and w = 0 in the first quadrant (x, ≥ 0 and X2 ≥ 0). We finally
define a continuation W for w by
W(X1, 42),
m+1
*2=0
W(X1, X2) =
# AW (x15 -iX2), *250
i=1
again. The function W belongs to Cm, (R?) and vanishes for ×1≥0.
Consequently the function
U=Vi+ W
is a suitable continuation for u. •
As a consequence of the above continuation property, it is easy to
derive the following inequality. Assuming that s' > s" > s'" > O and that S is
a bounded open subset of R" with a Lipschitz boundary, there exists a
constant K (depending on s', s", s'" and M) such that
(6,2,5)
for all & >0 and all u € CM," (R), s' = mt o.
We shall conclude this section by a brief survey of trace results for the
Hölder spaces. We shall obviously need them in the study of boundary
value problems. Their proofs are much easier than the corresponding
proofs for Sobolev spaces.

Theorem 6.2.6 Let N be either a half space or a bounded open subset of

R" with a Ck. boundary I. Then the mapping
du d'u) (6,2,6)

maps CM. (R) onto I'-o Cm-1. (F), provided I Sm and m+o sk+1.
 6.2 A BRIEF REVIEW OF HOLDER SPACES 279

As in the previous chapters y denotes the mapping

Here all the functions we consider the traces of are continuous and
accordingly no extension is needed to define y (compare with Subsection
1.5.1).
Theorem 6.2.6 is proved by reduction to the case when & is a half
space using local coordinates and a partition of unity. Assume that the
half-space is defined by *n >0, then the claim is that
Ur> Irnu, yn Dau, ...., YnDhu}
maps Cm. (R) onto I'=o Cm-i«(R"-1). Indeed starting from u € Cm. (R'),
it is clear that Dhul, =o belongs to Cm-1. (R"-"). Therefore the mapping
in (6,2,6) is well defined. To show that it is onto one can follow the
method of proof of Theorem 5.8, Chapter 2 in Necas (1967). (Observe
that the same kind of proof works when & is an infinite strip Ja, b[R
(with a, beR; a <b). This will be useful later.)
We also need trace results for domains with corners. The model result
(corresponding to Theorem 1.5.2.4) is the following:

Theorem 6.2.7 The mapping ur> (k}k-o; (81}i-o) defined by

fk =DRuly=o, 81 = D'ulx=o (6,2,7)
is a continuous mapping from Cm.°(R, XB.) onto the subspace of

k =0 1=0
defined by
Difk (0) = D$81(0), I+k sm. (6,2,8)
Proof The direct part of the statement is easy. Indeed when u is given in
CM. (R+ XR+), it is obvious from the definitions that

belongs to T and satisfies the compatibility conditions (6,2,8). We just

have to prove the converse, i.e. that the mapping is onto.
For this purpose we start from fk €Cm-k(R), Osk&m and giE
Cm-lo (R+), 0€1&m, given such that the conditions (6,2,8) are fulfilled.
We have to find u € CM. R+ X R+) such that (6,2,7) holds. As we did in
the proof of Theorem 1.5.2.4, we first select functions G, € Cm-1. (R) such
that
Give. = 86 0515m.
280 RESULTS IN SPACES OF HÖLDER FUNCTIONS

Then, by the trace theorem for Hölder functions in a half space, there
exists U€ Cm. (R?) such that
DiUl=o= Gh
Next, instead of looking for u, we look for v = u - U, i.e. for a function
U E CM. (R+ X R+) such that
O<k&m
DiUlx=0=0, Osl&m.
In other words, we have reduced our problem to the particular case when
§ is replaced by zero for all 1. It is clear from (6,2,8) that
Dihk (0) =0, l+k=m
and consequently h € Cm. (R). (Again, ~ denotes the continuation by
zero out of the domain of definition of the function.)
Applying again the trace results concerning a half space, we can find
WE Cm. (R?) such that
Osk&m.
We obtain v as follows:
m+1
0 (X, y) =W(X, y) -2 1,W(-jx, y), *70, 4>0,
i =1
assuming that
m+1
2 (-i)'1, =1, 0515m. m
This theorem implies a similar result on curvilinear polygons whose
proof uses the same techniques as the proof of Theorem 1.5.2.8.

Corollary 6.2.8 Let 12 be a bounded open subset of R? whose boundary I

is a curvilinear polygon of class C°. Then the mapping
d'u) 15iSN, Osl5m,
is linear continuous from Cm. (R) onto the subspace of

¡=1 1=0
defined by the following conditions: Let L be any linear differential operator
with coefficients of class C° and order d Sm. Denote by Pit the differential
 6.2 A BRIEF REVIEW OF HOLDER SPACES 281

operators tangential to I; such that

120
then
I (P, fiN) (S,) = E (P, -1.1,+ 1.)(S,). (6,2,9)
120 1≥0
The notation is the same as in the previous chapters. Namely I; is the
ith side of F, v; the corresponding outward normal vector field and Y; is
the corresponding trace operator, 1 - j< N. Following the direct orientation,
I; ends at S;.
Another consequence of Theorem 6.2.7 through Corollary 6.2.8, is a
statement similar to Theorem 1.6.1.4. Here we keep the same notation.

Theorem 6.2.9 Let S be a bounded open subset of R? whose boundary is

a curvilinear polygon of class C°. Let {B.k}k=1 be for each j, a system of
differential operators in S, with coefficients belonging to C°(R), which is
normal on I. Then the mapping
Ur»(fi.k = y; Biku}, 1sjEN, 1sksK;
maps Cm. (R) onto the subspace of
K

i=1 k=1
defined by the conditions (1, 6, 1, 2) for d&m and all possible systems of
differential operators {P;.k}k=, tangential to I; and {Q;+1.k}k-1 tangential to
Titi, such that identity (1,6,1,1) holds.
The proof is quite similar to that of Theorem 1.6.1.4.
Now we conclude this section with one technical result which is useful
in the remainder of this chapter. It is an extension of Theorem 1.4.5.3.
Theorem 6.2.10 Let R be a bounded open subset of R?, whose boundary
I is a curvilinear polygon. Assume that 0€ F. Let V be a neighbourhood of
O such that
vnA si(r cos 6, r sin 6); r=0, a 50=b)
with b- a <2m. Finally let u be a function which is smooth in § \50 and
which is equal to
r°p (0)
in VMS, where seC a: b]). Then
u €Cm.o (52) (6,2,10)
282 RESULTS IN SPACES OF HOLDER FUNCTIONS

for a >m to, while

usCm.o (R) (6,2,11)
for a <m to, when a is not an integer.

Observe that (6,2,10) follows from (1,4,5,1) with the help of the
Sobolev imbedding theorem, when a is strictly larger than m + o. Other-
wise it is a matter of direct elementary proof.

6.3 Regular second-order elliptic boundary value problems

revisited

6.3.1 The Schauder inequality

Here we shall derive a Hölder version of the a priori estimate proved in
Section 2.3. Let us first briefly recall the notation (which we keep
consistent with that in Chapter 2). The domain & is a bounded open
subset of R" with a C2.1 boundary; - A is a strongly elliptic real second
order operator in 1 and B is a real boundary operator of order dd = 0
or 1) which is nowhere characteristic on F. The reader is referred to
Section 2.1 for the detailed assumptions on the coefficients. Our aim is to
prove that there exists a constant C such that
lulla+2.0.5€ClAullo.0.52t|lyBullo+2-d.0,5+|/ullo+1,00,53] (6,3,1,1)
for all u € C2.9(8), 0<0<1.
Exactly as in Section 2.3 the first step is the proof of an inequality in
the half space. Here we make use of the notation of Subsection 2.3.2.
Namely, L is a strongly elliptic, real, homogeneous second-order operator
which has constant coefficients, while M is a real, homogeneous, first-
order operator with constant coefficients, noncharacteristic on the
hyperplane *n = 0. We now intend to prove that there exists a constant C
such that
(6,3,1,2)
for all u € C2.0 (R) whose support is bounded.
Observing that such a function u also belongs to H (R), we can use
the representation formula (2,3,2,12). Thus we have

(6,3,1,3)
Here F is a suitable continuation of f = Lu + u such that F€ CO. (R) and
has compact support; E is the elementary solution for L + 1 defined in
 6.3 REGULAR SECOND-ORDER PROBLEMS 283

Subsection 2.3.2. In addition we have

ko= (m, p- +i=1I im,s) in

(6,3,1,4)
k1=p-ko
h=g-Y.ME*F
where g = YmMu.
Now it is very tempting to proceed with the same proofs as in
Subsection 2.3.2, just substituting Theorem 6.2.3 in Mikhlin's multiplier
theorem 2.3.2.1. Unfortunately this method of proof requires using
spaces W$ corresponding to a negative order of differentiation s. This is
in particular necessary for Lemma 2.3.2.2. A theory of Hölder spaces
with negative order of differentiation is not yet well established. We shall
not attempt to define such spaces and accordingly we shall not be able to
derive any property of the operator n in the framework of Hölder
spaces. However, we shall be able to conclude by deriving directly the
properties of the Poison operator
(6,3, 1,5)

Lemma 6.3.1.1 When u belongs to H=(R) then

u = E* (F+ 48Sh), In >0, (6,3,1,6)
where

(6,3,1,7)
i=1
In other words u is the solution of a Dirichlet problem
Lu +u =f, Kn >0
(6,3,1,8)
Ink =9, In=0.
Proof This is a consequence of identity (6,3,1,3), observing that

Ê(E, x.) =-
Inn P+ (E) -p-(E)
for xn > 0.
From identities (6,3,1,4) and (6,3,1,7) and from the multiplier theorem
6.2.3, it follows that there exists a constant C such that
CP 12+09,0,1M 15 Cilfllor, 0,1584 + 118|1+0,00,894-15. (6,3,1,9)
284 RESULTS IN SPACES OF HOLDER FUNCTIONS

Next we focus our attention on problem (6,3,1,8). A suitable linear

transformation of coordinates reduces L+ 1 to -4 + 1, and accordingly
we can assume now that I is simply -4. The basic estimate is the
following.

Lemma 6.3.1.2 For 0<o <1, there exists a constant C such that
(6,3,1,10)
for all u € C2.0 (R$).

Proof We first reduce the general case to the particular one when both
Yal and nf vanish. Indeed, by the trace theorem 6.2.6, we know that
and InfECD."(R"-1),
and consequently there exists u € C2. (R) such that

In addition, o depends continuously on Ynu and f, i.e. there exists C such

that
(6,3,1,11)
Then we look at w = u - v; this function is a solution of a homogenous
Dirichlet problem:

{-Aw+w=8, Kn =0,
where g=f-(-Av +w). Therefore g€ CO. (R) and Ing =0. We now
perform an odd reflection through the hyperplane X = 0, i.e. we define W
as follows:

*n =0
-W(x', -Xn), In <0.
We define G in a similar fashion. Since In8 = 0, Y.W =0 and In Diw = 0,
it follows that Ge C.«(RM) and WE C2.0 (RM). In addition we have
(-4+1)W = Gin R".
Consequently the multiplier Theorem 6.2.3 shows that there exists a
constant C2 such that
 6.3 REGULAR SECOND-ORDER PROBLEMS 285

By restriction to R$ we derive the existence of another constant C3 such

that
W|2+0.09.515C31181105,0,589 (6,3,1,12)
Finally inequality (6,3,1,10) follows from (6,3,1,11) and (6,3,1,12).

With the help of inequality (6,3,1,9) we conclude:

Theorem 6.3.1.3 Let - L be a homogeneous strongly elliptic second-order

operator with real constant coefficients and let M be either the identity
operator or a homogeneous first-order operator with constant coefficients for
which the hyperplane *n =0 is not characteristic. Then for 0 <o <1 there
exists a constant C such that

for all u € C2.° (R4) with bounded support (0 <o <1).

This statement is quite similar to Theorem 2.3.2.7 in the framework of

Hölder spaces. Then by applying exactly the same technique as in the
proof of Theorem 2.3.3.2, we derive a statement concerning operators
with variable coefficients:

Theorem 6.3.1.4 Let A, B and & fulfil the assumptions in Section 2.1
Assume in addition that
(a) the boundary I of 12 is of class C2,1
(b) ai, E C". ' (R), 1 si, jen
(c) b; ECHO (5).
Then for 0 <o <1 there exists a constant C such that (6,3,1,1) holds for all
u€C2.0(5).
We first prove that it is enough that each * € has a neighbourhood V
so that (6,3,1,1) holds for all u € C2. (8) whose support is contained in
V This is a statement similar to Lemma 2.3.3.1. The proof is exactly the
same in the Hölder norms due to the extra assumptions that are made on
the coefficients of A and B.
Then the existence of V is checked in the two particular cases (a) and
(b) similar to those in the proof of Theorem 2.3.3.2. Accordingly we have
two lemmas:

Lemma 6.3.1.5 For every y € M, there exists a neighbourhood V of y in 12

such that (6,3,1,1) holds for all u € C2.5 (R) whose support is contained in
V.
286 RESULTS IN SPACES OF HÖLDER FUNCTIONS

Some minor modifications of the proof of Lemma 2.3.3.3 are neces-

sary. This is why we shall detail the proof of Lemma 6.3.1.5.

Proof We freeze the coefficients of A at y and thus we obtain an

operator L with constant coefficients such that - L is strongly elliptic:

1= £i, =41 D,D;

where hi;=ai,¡(y).
Then if the support of u is contained in V and We we have

where o, are functions belonging to C' (R") such that a. = a,; in M. Let
E be the elementary solution for L + 1 introduced in Section 2.3.2. It
follows that

ii = 1

By Theorem 6.2.3, E * is a linear continuous mapping from CO. (RM) into

C2. (R) since the Fourier transform of any derivative of E up to order 2
fulfils the assumptions of that theorem. Therefore there exists a constant
C, such that

(6,3,1,14)
Handling the last term is slightly more tricky than in the case of the
Sobolev norms. Actually we have

§max la;; - Wil| D;D;ula,o,n + lai,/ - Vi,illo, o, n max DiD;u,

provided u has its support in V. If we assume that the diameter of V is
58, then we have
max
X€V
a;.;(x) - li. sK8°
since ate CO. (@) and we also have

max |DiD;u (x) | 58° |14|12+0,00,57.

Accordingly there exists a constant C2 such that

(6,3,1,15)
 6.3 REGULAR SECOND-ORDER PROBLEMS 287

We conclude by choosing & small enough to ensure that C28 is less

than 1. Then (6,3,1,13) holds for all u€ C2.9(8) with support in V,
provided V is contained in M and the diameter of V is less than 8.

The technical lemma corresponding to Lemma 2.3.3.4 is the following:

Lemma 6.3.1.6 Let ye I have a neighbourhood W in F, contained in the

hyperplane {* =03. Then there exists a neighbourhood U of y in S such
that (6,3,1,13) holds for all u € C2.0 (R), whose support is contained in U.

Translating the proof of Lemma 2.3.3.4 into the framework of Hölder

spaces requires the same kind of modifications as for Lemma 6.3.1.5. It is
not worth detailing a theorem.

6.3.2 Smoothness
We shall now derive some regularity results similar to those in Section
2.5.1. We assume again that A and B fulfil the assumptions in Section
2.1. In addition we assume that
(d) the boundary I of M is of class C2,1
(e) ajj E C'. '(8), bi E C. ' (R).
These requirements are slightly more restrictive than those in Theorem
6.3.1.4 (about the a priori inequalities). These extra assumptions could
be avoided, but they will save many boring technicalities.
Theorem 6.3.2.1 Let u€ W=(D), with p> n, be such that
(Au =fEC"..(D)

with 0<0<1; then u € C2.0(12).

Basically we shall approximate the data f and g by better ones to which
the regularity result of Theorem 2.5.1.1 may be applied. Then we shall be
able to take limits with the help of the a priori inequalities of Subsections
2.3.3 and 6.3.1. However, the lack of convenient density results in the
Hölder spaces introduces an additional complication. The following state-
ment is a possible substitute for a density result.

Lemma 6.3.2.2 Let N be a bounded open subset of Rn with a Lipschitz

boundary F. Then each u € Cm. (1) can be approximated by a sequence
Um, V=1, ... such that
(a) u, ED (12),
 288 RESULTS IN SPACES OF HÖLDER FUNCTIONS

(b) 4, lmto, a, n is bounded when v => +00,

(c) Hu, -ullm+o: 0, ->0 when v-> too for every o' <0.

Proof This is straightforward. One can define u, as follows:

where U€ Cm. (RM) is a continuation of u (i.e. Us = u) and Pus V = 1,.

[
is an approximation of identity; in other words

P, ED(R"), P, dx =1
and the support of p, converges to {0} when w->o.

Proof of Theorem 6.3.2.1 We first choose ^ large enough for inequality

(2,3,3,7) to hold. Then we set h =f+ lu; it is clear that h€ C"."(R).
Let h, € C'(R) and 8. € C(F) be such that Uh lIl. a.se and Ilsull2-d+0,0, r
remain bounded and
y->00
y->00
for all o' <o. Such sequences can be found by Lemma 6.3.2.2. (In order
to approximate g, we can consider g as the trace of a function G
belonging to C2-d. (R) and apply Lemma 6.3.2.2 to G.) Finally let u, be
solution of
Au.+14,=h, in O
Bu,=8u on Г.
We know that u, exists by Theorem 2.4.1.3. Then Theorem 2.5.1.1
shows that u, € W.(8) for every p < +0. Consequently u, € C2.0(8).
In order to take the limit in , we make use of the two a priori
inequalities (2,3,3,7) and (6,3,1,1). We first have
w-unza.asClih.-ho.p.nt|80-8112-d-1/p.p.F3
and consequently u, -> u in W5(1), when v-> +00.
On the other hand, we have
u.-U/llo'72.00,5_CLh.-hallo.s.nt|8v-8v'lo'+2-d,%,r
+46-46-No'+1,00.581
and since W?(M) is continuously imbedded in Cl.o'(M) for p large
enough, this shows that u, v = 1, 2, ... is also a Cauchy sequence in
C2.4' (5) for all o' So. It follows that u € C2.0' (R) and that D,D, u, (x) -
DiD; u(x) for all x€S, i, j = 1, 2, ..., n.
 6.4 SECOND-ORDER PROBLEMS IN POLYGONS 289

Finally, applying again inequality (6,3, 1, 1), we see that|U, l2.o.§ remains
bounded as v->+∞. Accordingly, there exists a constant K such that
ID;D;u. (x) - DiD;u, (y) SK x-y°

for all i, j= 1, 2, ..., n, v = 1, 2..., and x, ye. Taking the limit shows
that all the second derivatives of u are Hölder continuous with exponent

Remark 6.3.2.3 A proof of the same result assuming that a; and b;

belong to C. (R) requires application of the same technique as in the
proof of Lemma 2.4.1.4 (i.e. locally flattening the boundary and mollify-
ing tangentially).

Remark 6.3.2.4 As a consequence of Theorem 6.3.2.1, we can restate

each of the Theorems 2.4.2.5-2.4.2.7 in the framework of the Hölder
spaces. For instance the result corresponding to Theorem 2.4.2.5 reads as
follows. Assume that the hypotheses of Theorem 2.4.2.5 are fulfilled and
that in addition the boundary I of M is of class C21 and that ai,€
C'. (8), 151, jEn; a, € CO." (1), 0 si sn. Then, for every f€ C°." (S) and
every g€ C2. (I), there exists a unique solution u € C2.5(1) of

& D, (aiD,u) +£ a,D,utaou =f in D

i.=1
yu = g on Г.

6.4 Second-order elliptic problems in polygons revisited

6.4.1 The Schauder inequality in an in nite strip

We now look again at the boundary value problems of Section 4.2.
Keeping the same notation, we are going to find sufficient conditions on
the coefficients a, b, a;, B;, 1; i = 0, 1 for the existence of a constant C
such that
(6,4, 1,1)
for all u€ C2. (B) such that v;M;u = 0 on Fi, i = 0, 1.
A first step will be the proof of the weaker inequality.
max u SC||LUlla,a.B. (6,4,1,2)

The technique is quite similar to the one we used in Subsection 4.2.2, just
replacing the multiplier theorem 2.3.2. 1 by Theorem 6.2.3. Then inequal-
fi
290 RESULTS IN SPACES OF HOLDER FUNCTIONS

ity (6,4,1,2) makes it possible to replace I by 4-1 to estimate the

second derivatives of u in CO. (B).

Theorem 6.4.1.1 Assume that b > 0, a + 0 and that for each j= 0 or 1,

we have either a; = 1 or a; = B; =0 and 1; = 1. Assume in addition that the
characteristic equation (4,2,1,2) has no real root. Then for 0 <o <1 there
exists a constant C such that inequality (6,4, 1,2) holds for u € C2.°(B) such
that y;M;u = 0 on Fi, j = 0, 1.

Proof We first consider the particular case when u € C2. (B) and has a
bounded support. Thus u also belongs to H'(B) and we can apply
Theorem 4.2.1.2. In other words identity (4,2,1,3) holds.

Then a Hölder version of Lemma 4.2.1.3 is this one:

Lemma 6.4.1.2 Let E, Y, z-> K(E, y, z) be a smooth function such that

max {|K(5, y, z)| + (1 + |$|) |D K (5, y, z)1} dz < +00;

then the mapping u->f defined by

is continuous from CU. (B) into the space of continuous bounded functions
in B.
Applying this lemma we obtain inequality (6,4,1,2) but only for u€
C2.0 (B) n H? (B) such that y;M;u = 0 on F;, j = 0, 1.
We shall reduce the general case to the previous particular one with the
help of cut-off functions. Accordingly let 0 € D(R) be such that
0(x) =1,X ≤1
0(x) =0, x ₴2
and let for & >0, 0 be defined by
O (x) =0 (EX).
Clearly 0 (x) -> 1 when & ->0. Thus let us start from u € C2. (B) such that
V;M;u = 0 on Fj, i = 0, 1. We have
Y;M;Oçu = B;dev;u on Fj, j=0, 1.
By the trace theorem (Section 6.2) we know that there exists v. € C2.0 (B)
 6.4 SECOND-ORDER PROBLEMS IN POLYGONS 291

such that

on Fi, j=0,1
Ty;M, ve = B;0 = Y;U on Fir j=0,1
and there exists a constant C, such that

(6,4,1,3)
i=0
If we replace v. by 02 U, we can assume in addition that v. has a bounded
support. Consequently we have
0.U -v. €C2.9(B) MH?(B)
and
V;M; (OU-v) =0 on Fir j=0, 1.
Applying inequality (6,4,1,2) to 0 U - ve, we get
max 0u - v. | = C||I (0,W) - LUella, %., B.

Therefore it follows that

+2|10Dyullo, 00. Bs + |al llO's ullor, o., B + Call Uell2 + 0,00, B3}

Finally with the help of (6,4,1,3) we get

.00, B + 104 U lo.00, B + 2 |10" DU lo.00. B

+|a||l@sullo.c.Bt Cs|l@@ulli+0,00,B}
and taking the limit in & -> O yields plainly
max uS||Lula,o, B;

this is the desired result. S

Proof of Lemma 6.4.1.2 Again, as in the proof of Lemma 4.2.1.3, we

denote by M the function
M(y; z) = l.u.b. {K(S, y, z) + (1 + 5D) |D K(S, y, z)I}.
292 RESULTS IN SPACES OF HÖLDER FUNCTIONS

It follows from Theorem 6.2.3 that there exists C such that

u (x, y)-u(x', y)
1.u.s.lu(I, y)1+1.u.b.
x*× Ix-x'°
• h

SC M(y, 2)K1.u.b. If(x, z)1+1.u.b.

5. F(X, Z) - f(x', 2) I dz.
x*x' [x-x'o
It follows that

We are now able to prove the stronger inequality (6,4,1,1).

Theorem 6.4.1.3 Under the assumptions of Theorem 6.4.1.1, there exists

a constant C such that inequality (6,4,1, 1) holds for u € C2." (B) such that
ViM;u =0 on Fj; 1 = 0, 1.

Proof Exactly as in the proof of Theorem 6.4.1.1, we begin with the

particular case when u has a bounded support and therefore belongs to
H°(B). We choose M; i = 0, 1 as in Lemma 4.2.2.5 and set
{8=4и-u
W. = y; (aD,u + B;Dyu + p;u), j=0,1.
Clearly we have
(8=Lu -aDxu + (b-1)u
(6,4,1,4)
74: = (M; - 1; ) viu, j=0,1.
Then we set 0 = u -E*G where G is a continuation of g from B (i.e.
GIs = g) such that G€ CO.° (R?) has bounded support and
Glo, 00, 182 C, glo, 0, B,
and where E is the elementary solution of (4 - 1) defined by
E=-F-4(1+1512)-1.
By Theorem 6.2.3 the convolution by E maps continuously CO. (R?) into
C2. (R%). Consequently there exists a constant C2 such that
lull2+0,0,BSCallUll2+0,00,8 +18llo,0,B). (6,4, 1,5)
Now, we have to estimate > which is a solution of
540+0=0 in B
(6,4, 1,6)
1=0,1,
 6.4 SECOND-ORDER PROBLEMS IN POLYGONS 293

where h; is defined by
h; =Y; -Vila;D, +B;Dx + M;) E* G, 1=0, 1. (6,4, 1,7)
Since G has a bounded support, it belongs to L2(R?) and E* G belongs to
H°(R°). It follows that v also belongs to H°(B). This is why we can use
the calculations in Subsection 4.2.2.
This means that, setting k; = y;u and 4; = y;DU, j = 0, 1, we have:
ko= a, to =rB,
k, = a cosh (rh) + B sinh (rh)
1 =ar sinh (rh) + Br cosh (rh),
where
a =àlaorh,- La,r cosh (rh) + (iB,& + M,) sinh (rh) lhot
#I-GiBott Math, + laur sinh (rh) + (B, & + M41) cosh (rho) lid!

d =r cosh (rh)[ao(iß¡E + Ms) - as (iBoS + Mo)]

+sinh (rh) [77a0a1 -(iBo5 + Mo) (iß15 + pa)I
r=V (1+53).
It is worth recalling here that due to Lemma 4.2.2.5 d does not vanish for
any real &. Applying Theorem 6.2.3 we get the following inequality:

Then (6,4, 1,4) and (6,4,1,7) imply that

(6,4, 1,8)
¡=0
To conclude we write the equation of i:
(-4+1)0=-40886+k1884-1688074@81
Accordingly we have
0=-E#4-K088+k884-688+1883.
And consequently (as in Lemma 6.3.1.11)
(6,4, 1,9)
+ Here
exp (-1*2|/ (1+53))
£(E, x,) =
2V(1 + 82)
294 RESULTS IN SPACES OF HOLDER FUNCTIONS

where

80 =-kot JI + 53)° 41 = Kit J=73 (6,4,1,10)

Thus, in B, is the sum of two functions vo and v, which are solutions of a
Dirichlet problem in the half planes x2 > 0 and * <h respectively
00=-E*40886 01=-E*p,@Sh
1-40, +00=0,82>0
14600=-240
1-40, +01=0, 82 Ch
19.0, =401.
Applying Lemma 6.3.1.2 twice, we get the inequality

i =0

This implies, by (6,4,1,10) and Theorem 6.2.3, that

(6,4,1,11)
i =(
Summing up, from the inequalities (6,4,1,5), (6,4,1,8) and (6,4,1,11) it
follows that
U|2+o.x. B SCRILullo. (6.4.1.12)

Now we take advantage of the classical inequality (6,2,5) in the case of

an infinite strip:

Lemma 6.4.1.4 There exists a constant K such that

lullita.a.B.Sallull+a.a,B+Ke-(Ito) max |ul

for all & € J0, 1] and all u € C2.° (B).

Choosing & small enough, inequality (6,4,1,1) follows from (6,4,1,12)

and (6,4, 1,2).
So far we have always assumed that u had a compact support. In order
to remove this extra assumption, we approximate a general u by 0 u
exactly as in the proof of Theorem 6.4.1.1.

Proof of Lemma 6.4.1.4 One first extends the function u into a function
 6.4 SECOND-ORDER PROBLEMS IN POLYGONS 295

U€ C2. (R). Then the corresponding inequality on the whole plane, is an

easy consequence of the Taylor formula.

6.4.2 The Schauder inequality in a polygon and its consequences

The notation is now that of Section 4.4. Thus & is a bounded open subset
of R? whose boundary I is a polygon. To each side I; corresponds a
boundary condition which is either the Dirichlet condition or an oblique
condition (in the direction of M; = v; + B;r;). Given fe CO. (8), 0<0<1,
we look for a solution u € C2.0(1) of

Au =f in 2
in F; jeD (6,4,2, 1)
du - Vil = 0 on li jEN
1 + Bi aTi

We begin with the proof of an estimate of the Schauder type namely

for all u € C2. (S) which fulfil the boundary conditions in (6,4,2,1). Here
we shall use the same local method of proof as in Subsection 4.3.2, using
suitable weighted spaces (depending on p, the distance to the corners):

Definition 6.4.2.1 We denote by Pm. (1) the space of all functions u

defined in S such that
(a)plal-m-"D°u is continuous and bounded in 1 for all a with a Sm,
(b) D'uE C". ' (12) for la! = m.

A Banach norm for this space

Hullpm-co, = E 1.4.b.plat-m-aDau
a Em
+ I lu.b. ID°U(x) x-y°
- D°W(y)
(al=m XER.YES)

The space pm. (R) is a subspace of Cm. (R) while a converse inclusion is
given by the following result.

Theorem 6.4.2.2 Let u€ Cm." (R) be such that

D'u(S)=0 for a &m
j=1, 2, ...,N; then u € Pm." (1).
296 RESULTS IN SPACES OF HÖLDER FUNCTIONS

This is an elementary application of Taylor's formula using polar

coordinates near each corner (see the proof of Theorem 4.3.2.2.
The basic estimate here is the following.

Theorem 6.4.2.3 For 0 <o <1 there exists a constant C such that
ulpasca@Cilfla.a.stullito.a.st (6,4,2,2)
for all solutions u € P3.°(52) of problem (6,4,2,1), provided

(441-8-1270101)
is not an integer for any j, where $; = arctan B; jEN and $; = T/2, jED.

Sketch of the proof We just outline this proof since it is very similar to
the proof of Theorem 4.3.2.3. We consider the problem locally with the
help of a partition of unity fn,)-o....N on & such that n; € (3) for each j
and such that
(a) the support of no does not meet any vertex of 1.
(b) the support of ni contains S; and none of the Sk with k+ j, j=
1, 2,..., N.
(c) we have
Ani =0 on Ik
duk + BkATk
AMi
for k = i if jeN and for k=j+1 if j+1EN.
Thus we have
(6,4,2,3)
and
on Ik keg
д
[Yk (n;4) =0
Vk- AVk
д- VK(n;u) = 0 on Fir KEN,
1=1, 2, ..., N.
We can apply inequality (6,3,1,1) to nou and this yields the following
inequality:
(6,4,2,4)
We now consider the functions niu, 1 < j€N.
Using the polar coordinates with origin at S; and setting (for a given j)
w(t, 0) =e-(o+2) (niu) (e'tie),
 6.4 SECOND-ORDER PROBLEMS IN POLYGONS 297

we obtain a function belonging to C2. (B) where B = RX]0, w;L. In

addition w is solution of a boundary value problem in B to which
inequality (6,4,1,1) may be applied. This implies the following
(6,4,2,5)
Finally inequality (6,4,2,2) is a plain consequence of (6,4,2,3), (6,4,2,4)
and (6,4,2,5). -
Now Theorem 6.4.2.2 shows that P3, (R) contains a subspace of
C2.9(1) whose codimension is finite in C2.9(8). Therefore a proof similar
to that of Theorem 4.3.2.4 yields the following:

Theorem 6.4.2.4 Assume that 0<0<1 and that 1/ (Q;+1-9; -

[2+o Jw;) is not an integer for any j. Then there exists a constant C such
that
(6,4,2,6)
for all u € C2.0(1), a solution of the problem (6,4,2,1).
We are now going to draw the consequences of Theorem 6.4.2.4. The
following existence result is closely related to the content of Subsection
5.1.3. Let us recall briefly some notation (taken from Definition 5.1.3.4):

dim =. j=1,2, ...,N, me,

W;
Gim(re'%,) =r;^mcos (1; mO;+ Dits)n; (r,@'%,)
where Aim is not an integer and
Gim(re'',) = r;", miln r; cos (1;, mO; + d; +1)
+ 0; sin (A; mO; + d; +1))n; (r, ei%,)
when Aim n is an integer. Here again r and o; are the polar coordinates
with origin at S;
The basic result is the following where, for the sake of simplicity, we
assume the uniqueness of the solution in H°(8). Sufficient conditions for
this uniqueness have been found in Theorem 4.4.1.3.
Theorem 6.4.2.5 Assume that D is not empty and that at least two
of the vectors u; are linearly independent. Assume in addition that
(1/7) (Q;+1 + Ф; -12+0 w;) is not an integer for any j. Then for each
f€CO. (R), with O<o <1, there exists a function u and numbers Gim
such that
И- Cim Gi;m € C2.0(8)
- (0 + 2) <1;.m <0
and u is solution of the problem (6,4,2,1).
 298 RESULTS IN SPACES OF HÖLDER FUNCTIONS

Proof We choose p >2 such that (ф; - d;+1 + 2w;/g) / m is not an integer
for any j and apply Theorem 4.4.4.11. Since f€ Lp (82), there exists a
function u and numbers C,m such that
W = U- CimGi.mE W=(12)
1SiEN
-2/9<1,.m <0
1.m #-1
and u is a solution of the problem (6,4,2,1). Now we study w. Clearly we
have
Aw =f- Gi.mAG;m =8 in D
ISiEN
-2/9<1,m <0
1.m #-1
Y;w = 0 on I; jED
Aw on Fi jeN
+Biar,- Y;W = 0
because the functions Sim , fulfil all the homogeneous boundary condi-
tions in (6,4,2,1). In addition we have g € C°° (1) since G;m is harmonic
near S; and smooth far from S;
On the other hand 4 is one to one (Theorem 4.4.1.3) and has finite
index (Theorem 4.4.4.11) from
dw д
into Lp (8). Let us denote by U1 ..., Um a basis of the annihilator
(formerly denoted by Na in Chapter 4) of 4E. Necessarily we have
(8;0k) =0, k = 1, 2, ….., M. (6,4,2,7)
We shall now approximate w by smoother solutions. For this purpose we
shall approximate g by smoother functions, trying to keep (6,4,2,7).
We apply Lemma 6.3.2.2. This provides us with a sequence 8w, V =
1, 2, ... such that
(a) 8. € D (R),
(b) |8ullo, 0,5. is bounded when v-> +00,
(c) Il8u -gllo: o.5-> 0 when v->o for every o' <0.
Unfortunately there is no reason why & should be orthogonal to
Ok, k = 1, 2, ..., M. Thus we introduce k, k = 1, 2, ..., M belonging to
D(2) and such that
(Uks (1) = 8k.10 k, I = 1, 2, ..., M.
If we replace 8, by 8. - L-, (8u; Uk) k, we get a new sequence that we
 6.4 SECOND-ORDER PROBLEMS IN POLYGONS 299

still denote by & v = 1, 2, . .., such that the properties (a)-(c) above still
hold and that in addition
(80;0k) =0, k=1,2, ..., M, V = 1,2 ....
Accordingly there exists for each v a unique w, € E such that 4w, = 8.
In addition w, ->w in W5(8) when v=>+00.
Next Theorem 5.1.3.5 shows there exist real sequences Cm, such that
Cim. G;.m € WE, (12)
-1-2/91<1, m <-219

where ps has been chosen such that none of the [4;+1-8;-00;-2003/91)7

is an integer, 1/p. + 1/91 = 1. If pi is large enough the Sobolev imbedding
theorem implies that
W:, (8) = C2.9(8).
Consequently we have
W. -(2+0) <1,.m <-219
Ci.m, Gim € C2.0(12).

In other words w, belongs to S, the space

dw д
It Bi aT;
augmented with the span of the functions Si,m corresponding to
-(2+0) <1im1 <-2/9, j=1,2,...,N. If p is also large enough, the
Sobolev imbedding theorem implies that

and consequently we have Sc C1. (8). A natural Banach norm on S is

Kimly
-(2+0) <1,.m <-219

where the infimum is taken over all the possible functions o€ & and
numbers Cm such that
W=o+-(2+0) <1, m<-219CimG;.me
In addition, & has a finite codimension in S. It follows from Theorem
6.4.2.4 that there exists a new constant C such that
wils SCIAwlo,o.s.+w|lito.o.rit.
We apply this last inequality to w,. It shows that Wu, V = 1, 2, ..., is a
bounded sequence in S. Remembering that w, ->w in W(8) when
w->+o, it is now easy to conclude that we S. Going back to u we see that
300 RESULTS IN SPACES OF HÖLDER FUNCTIONS

it belongs to C2. (R) augmented with the span of all the functions G;.m
corresponding (with the exception of Aim =-1) to -(2+0) <1; m <0,
j=1,2, ..,N. •
Theorem 6.4.2.5 is a Hölder version of Theorem 4.4.4.11. In Section
5.1, we proved a wide extension of Theorem 4.4.4.11, namely Theorem
5.1.3.5, where the behaviour of higher-order derivatives of the solution is
investigated. This was achieved by differentiating the original problem
(4,1,1) and taking advantage of the very general trace theorem 1.6.1.4.
Replacing Theorem 1.6.1.4 by its Hölder version, Theorem 6.2.9 yields
the following extension of Theorem 6.4.2.5.

Theorem 6.4.2.6 Assume that D is not empty and that at least two of the
vectors M; are linearly independent. Assume in addition that
(8; +1-84-k +2+0)00;)
is not an integer for any j. Then for each fe Ck."(5) and g; € Ck+2.9(Th), j€
D, 8; ECktl."(Ty) jeN with 0<0 <1 such that g; (S;) = 8i+1(S;) when i
and j+1 eD, there exists a solution u of
Au =f in 2
ViU = 8; on lj j€D
du - ViU = 8; on F jeN
and there exist real numbers Cim such that

И- Ci.mE;mE Ck+2.9(5).
-(k +2+0) <1, m <0

Remark 6.4.2.7 This statement is also valid when one allows

cuts in M (i.e. w; =2m for some j).
7
A model fourth-order
problem

7.1 Introductory results

In this chapter we shall study the properties of the solutions of the first
boundary value problem for the biharmonic operator in a plane domain
with a polygonal boundary. All the notation concerning the domain 2
will be the same as in Chapter 4. Our main goal is this: Given f€ WK(1)
with 1<p<+0, k integer ≥-1, we look for a solution u € Wk+4(1) of

Azu=f in O
| Viu = 0 on I; j=1, 2,...,N (7,1,1)
ди = 0
on Is j=1, 2,..., N.
"Jus
The reason why it is useful to consider f given in W-'(R) will appear
clearly in Section 7.4, which is devoted to the related Stokes problem.
We shall start from a variational solution u € H?(1) to problem (7,1, 1).
Then localizing the problem near one corner, we shall apply the method
introduced in Kondratiev (1967a), to study the behaviour of u near the
corners. This is done in the framework of the weighted Sobolev spaces
which we defined in Subsection 4.3.2. Then the trace theorems of
Subsection 1.5.2 allow one to get rid of the weights in a very simple way.
Finally we shall extend all the results to the case p+ 2, by a technique
using a priori estimates, very similar to those of Subsection 4.3.2.
First let us recall briefly the classical variational approach to the
problem (7,1,1). We apply the Lax-Milgram lemma (see Lemma 2.2.1.1)
with the following choice of V and a:
ди= 0 on I; j=1,2,...,N.
and
a (u: v) =
27и 270
dx.

301
302 A MODEL FOURTH-ORDER PROBLEM

It is obvious that
"1 2zu112
a (u; u) iF,
= 5lax; ax; #allull2.2,0
for a > 0, due to Poincaré's inequality. Therefore a is a continuous and
coercive bilinear form on V. Applying Lemma 2.2.1.1 we get the follow-
ing result.

Lemma 7.1.1 For any feW. ' (R) given, there exists a unique u € H? (8)
solution of problem (7,1,1).

Indeed we solve the variational problem:

a (u; v) =(f;0)
for all v € H? (R). The Sobolev imbedding of H?(1) into W(R) for any
q€ ]1, ∞[, implies that f is a continuous linear form on V.
In order to be able to study the behaviour of u only near the corners
we need a smoothness result away from the corners.

Theorem 7.1.2 Let u eH?(R) be the solution of the problem (7,1,1) with
fEW$(R). Then ue WK+ (DIV) for any neighbourhood V of the corners.

The smoothness of u inside & is well known. Indeed we have

Que Wk+A (1) for every € D (R). The corresponding smoothness result
near the sides I; of , deserves a proof. Here, we denote by R3- the half
plane defined by X2 > 0; y is the trace operator on {*2=0}.

Lemma 7.1.3 Letv € W'(R#), 1 = 2 be a solution with bounded support of

aU
= 0.

Then v € Wit' (R₴)

Proof We shall prove that yAu € W'-1-1/P(R). The claim will follow from
the known results for the Dirichlet problem for the Laplace equation (see
Subsection 2.5.1) applied to 1 = 4v and then to v.
First we shall prove a representation formula for w in terms of
(-4+1)u, yo and y 20/2xz. For this purpose, we approximate o by a
sequence of functions Ums m = 1, 2, . .., such that
Um € W'(RZ) MH* (R=)
 7.1 INTRODUCTORY RESULTS 303

and such that Um -> in W(R%). Then we write

Um =Wm+ E*E* Phm
where E is the elementary solution for -4 + 1 introduced in Subsection 2,
hm =(=4+1)7Ums
P is a continuation operator form W -3(7) into WI-3(R) and from
I?(R%) into I?(R?). Obviously we have

WWm =NUm -YE*E*Phm EWI-1/P(R) MH712(0)

=dXz
AWm = dUm-Yaxz
Vax2
I EXE + Phan E WE 1-11P (R) M HSIZ(R).

Let us now perform a Fourier transform in X. We get

(1+53-D27im=0 a.e. in R7
Wm ($1, 0) =60.m (51) ae. in R
DoWn (51, 0) =41m(51) a.e. in R,
where do.m = YWm and 41,m = yD2Wm. It follows that
wm (51, *2) = exp (-V (1 +57)X2140,m (51) 41 + 824 (1+55)} + 41.m (51)x2).
Accordingly we have
Awn (510) =-(1+253) 80,m -21(1+57) 61.m.
In other words, if we denote by T the operator defined as follows:
(75) (6) =4(1+53) 6(81),
we have
yAWm=Go.m-2T41.m-27740.m
Equivalently, we have
yAUm =y AWm + y AE*E*Phm
=(7-2TID2-277y)(Um-E*E#Phm)-YAE*E*Phme
Taking the limit in m, we obtain finally
740=(7-27D2-2TZy)(0-E*E*Ph)-YAE*E*Ph,
where h=(-4 + 1)3. Due to the assumptions on , we conclude that
y40 =-(7-277D2-2IZy+yA)E*E*PIg-240 + u}.
304 A MODEL FOURTH-ORDER PROBLEM

The right-hand side of the identity belongs to W-1-1/P(R) due to

Theorem 2.3.2.1. -
Proof of Theorem 7.1.2 It is a step by step proof using Lemma 7.1.3 at
each step.
Assuming first that p is larger than two and that feW. '(1) we show
that u € W(R\ V). Otherwise there is nothing to prove since H3(1)c
W3(8) when p €2. Let us consider one of the sides I. After translation
and rotation we can assume for convenience that I; lies in {x2=0}. Let us
consider n€ D(8) a cut-off function whose support is contained in
{x2≥0} and does not meet I for l7 j. Finally let us define o by
v = фи.
Therefore o€ H? (R₴), yo = y 20/2x2 = 0 and
430=of+ (42;o]u)EH-'(R7).
Applying Lemma 7.1.3 shows that ve H°(R?). Then varying & and i
shows that u€ H (RIV) C WE(DIV).
Now we assume that we know that u€ W. (2\ V), 251<k+3 for
every neighbourhood V of the corners and we show that this implies that
u € W'*'(RI V). Indeed we go through the same steps as before. Setting
v = qu,
we have ve W" (R'), yo=y 20/2x2=0 and
430=of+ (142; 414) E WI 3(83).
Lemma 7.1.3 shows that v € W'*'(R%) and therefore varying 4 and j
shows that u€ W!* ' (81V).
We conclude by induction. -
In order to study the behaviour of u the solution of the problem
(7,1,1), near one of the corners, say S, we use the related polar coordi-
nates (r;, O;) as in Subsection 4.3.2. We also use a cut-off function Ni
which is equal to one near S, and has a bounded support which does not
intersect any of the I, but I, and Fir. Therefore u; = niu is the solution
of a boundary value problem in the infinite sector
G, = {re, r; > 0, 0 <0; <c;).
Precisely we have u, € H?(G,) and
434; = (n; 474) + (Ini: 4334) = fE WI(G;)

and in addition u; has a bounded support. Dropping, for convenience, the

subscript i, we are left with the problem of investigating the behaviour of
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 305

u € H?(G) a solution with bounded support of the problem

4Zu =f€ WK(G) (7,1,2)
in an infinite plane sector G with angle w.

7.2 Singular solutions, the L case

7.2.1 Kondratiev's method in weighted spaces
In this section we shall study the problem (7,1,2) in the framework of the
spaces P'(G) defined in Subsection 4.3.2. Briefly, we recall that a
function u belongs to P(G) iff
r-kHlalD°u€L. (G), ask.
Theorem 4.3.2.2 has been useful in comparing the weighted space
P(G) with the usual Sobolev space W(G); unfortunately it excluded the
case when p = 2. A corresponding weaker statement, when p = 2, is the
following:

Theorem 7.2.1.1 Let ue Hk (G), then u € P°(G).

The proof of this result is quite similar to the corresponding part of the
proof of Theorem 1.4.4.4.
For technical reasons which will become obvious later, it is also
convenient to introduce a weighted space of order - 1:

Definition 7.2.1.2 We denote by P-'(G) the space of all the distributions

T=!50 + DIgi + D282

r (7,2,1,1)
where g; € L. (G), 05j≤2.
A Banach norm on P-'(G) is the following.
2
Trog.l.b. Lllsillo.p.G
i=1
where the g.l.b. is taken with respect to all the functions 8; 05j<2
belonging to Lp (G) and such that (7,2,1,1) holds.
Now let us go back to the problem (7,1,2). Accordingly we consider
u € P…(G)
(remember Theorem 7.2.1.1) such that u has a bounded support and such
306 A MODEL FOURTH-ORDER PROBLEM

that
AZu =f€P$(G)
with k>-1 and such that yu = y du/av = 0. We shall denote by R a
number such that
u(r, e) =0 for r≥ R.
The Kondratier method consists in performing the same change of
variable r=e' as in Subsection 4.3.2 and then solving the problem by
Fourier transform with respect to t. Thus we replace the equation 4Zu = f
in the infinite sector G by a similar equation in an infinite strip
B= RXJ0, cL.
This change of variable also replaces the weighted Sobolev spaces by
ordinary Sobolev spaces. More precisely, the rule is the following.

Lemma 7.2.1.3 Assume that 4 € P(G) with k ≥ -1 and define is by

W(I, A) = q (e' cos O, e' sin 0)e(-k+21p»,
then be W$(B).

Proof This result is rather obvious when k is nonnegative. Consequently

we leave its proof to the reader. However the case when k is -1 is less
obvious and deserves a detailed proof.
From Definition 7.2.1.2 we know that

r
@=.80 + D, 81 + D282
where g; € L. (G), 0 ≤j<2. Using polar coordinates this means that

r• dk,+-rIdk,
© =KOx'dr 20'
where k; € Lp (G), j=0, 1, 2. Consequently we have

+ Die /P"k (e' cos O, e' sin A)}

+ Dote 2/D' k¿(e' cos 0, e' sin 0)}.
Here, each function
4, 0 r-> e2/P"k; (e' cos o, e' sin 0), j=0, 1, 2
belongs to L. (B). This clearly implies that y belongs to W. '(B). •
 7.2 SINGULAR SOLUTIONS, THE L2 CASE 307

Since u belongs to P2(G) by assumption, we set

U(t, 0) =é 'u (e' cos O; e' sin A) (7,2,1,2)
in accordance with Lemma 7.2.1.3. Hence
vEHZ(B) (7,2,1,3)
since, in addition, o and 20/20 have zero traces on Fo= RX{0} and
F=RX<w}. The equation of v is
(D4-2D?+1)0+2(D?+1)D50 + D$0=8 (7,2,1,4)
in B where
8(4, 0) = e'fle' cos O; e' sin 0). (7,2,1,5)
We observe that the assumption that f belongs to P(G) implies that
é (k+2)g є HK (B). (7,2,1,6)
Finally, since u vanishes for r ≥ R, it follows that o vanishes for t ≥ log R.
We recall that we define the partial Fourier transform of v with respect
tot by
1 (**
é it u(t, 4) dt. (7,2,1,7)

Here a is the dual variable of t and is possibly a complex number. From

(7,2,1,3) and from the fact that o vanishes for t > log R, we derive that ¢
is defined for Im + >0, analytic in Im + > 0 and that

(7,2,1,8)
for every 2>0. Here we have applied the Paley-Wiener and Plancherel
theorems.
In addition the function O-> ¢(r, 0) is, for almost every 71(7=71 + iTz,
Tz>0), an element of the space #7(10, wl). This implies four boundary
conditions on v (valid for every + with Im + > 0 since ‹ is analytic).
Ô (7, 0) = ¢ (7, w) = Doû (+, 0) = Do¢ (T, w) = 0. (7,2,1,9)
From (7,2,1,4), we derive the equation of ¢:
(74+27271)8+12-272) D38+Do =8. (7,2,1,10)
For each + (7,2,1,10) and (7,2,1,9) define a two-point boundary value
problem for a fourth-order differential equation in 10, [. Solving such a
problem is easy. It is uniquely solvable away from a discrete set of
characteristic values for r. This will make ^ an analytic function of - in
308 A MODEL FOURTH-ORDER PROBLEM

any subset E of the complex plane where & is an analytic function of T,

away from the characteristic values. Eventually we take advantage of
(7,2,1,6), from which we derive that & is defined for Im + ≥- (k + 2). This
will provide us with an analytic continuation for \ which implies more
regularity for v.
Now we shall make all the previous outline more precise, step by step.
First we investigate how well posed the problem (7,2,1,9), (7,2,1,10) is.
The characteristic equation for the differential equation is
p+ +2(1-42)p2+(57+272+1) =0
and its roots are p= ‡ #i.t Accordingly, a fundamental system of
solutions for the fourth-order equation is
sin A sinh rO, sin A cosh tO, cos 0 sinh tO, cos 0 cosh TO
when - is different from 0 and #i. In the particular case when =0, a
fundamental system of solutions is
sin 0, cos 0, 0 sin 0, 0 cos o
while in the particular case when r = ti, a fundamental system of solu-
tions is
1, 0, sin 20, cos 20.
For each , a Fredholm alternative holds; namely, for a given & the
problem (7,2,1,9) (7,2,1,10) has a unique solution iff the corresponding
homogeneous problem has only the zero solution. In other words we are
reduced to checking whether the following problem has only the zero
solution:
((74+252+1) 6+241-57)6+44443=0 in JO, w[
(7,2,1,11)
14(0) =4(00) =44'(0) =4'(w) =0.
Later we shall call regular the values of r for which (7,2,1,11) has only
the zero solution.

Lemma 7.2.1.4 The problem (7,2,1,11) has only the zero solution in the
following cases
(a) + is not a root of the characteristic equation
sinh? (tw) =72 sin? c (7,2,1,12)
(b) T=0,
(a) + = #i if w+ tan a and c#2m.

† When 7 = 0, p = i are double roots and when r = #i, p = 0 is double root

 7.2 SINGULAR SOLUTIONS, THE Lz CASE 309

Consequently the characteristic values are the roots of (7,2,1,12) except

0 and #i in the general case and except 0 when tan w = w.

Proof When r is neither zero nor #i, & solution of the equation in
(7,2,1,11) is of the following form
I = sin ALa sinh tO + B cosh tO]+ cos A[y sinh tO + 8 cosh TO]
where a, B, y, 8 are complex numbers. Substituting & in the boundary
conditions, one finds an homogeneous system of four equations in the
four unknowns a, ß, Y, 8. The corresponding determinant is
sinh? (tw)-72 sin? w.
When r =0, & is of the particular form
4 = sin O[a + B0]+ cos 0[y + 80].
The determinant is now
sin?w-w?,
which is not zero.
Finally when r = ti, & is of the particular form
4 = a + BO + y sin 20 + 8 cos 20.
The corresponding determinant is proportional to
sin w sin w -w cos w.
This is not zero unless @ = tan w or 2 Tr.
In each case, a, B, v, & are all zero and thus & vanishes unless the
determinant is zero.
Let us denote by D the set of all the regular values (i.e. the noncharac-
teristic values) for the problem (7,2,1,10) (7,2,1,9). D is the complement
of a discrete set in the complex plane. When € D the problem has a
unique solution

provided & is given in H 200, wD. In addition, let E be any open subset
of the complex plane such that
Th g
is analvtic from E into H 2(0, w[), then

is analytic from DE into H3(10, ∞D). Let us now find a subset E.

310 A MODEL FOURTH-ORDER PROBLEM

We recall that &, like u, vanishes for t> log R. Applying the Paley-
Wiener and Plancherel theorems, we derive from (7,2,1,6) that, when k is
nonnegative g is defined for Im + # - (k + 2), analytic for Im + > - (k + 2)
and in addition that

§R'a+(k+2) Me- (k +2) gllk. 2,.B. (7,2,1,13)

A similar result when k = -1 deserves a detailed proof.

Lemma 7.2.1.5 Assume that e 'g€ H '(B) and that g vanishes for
t> R. Then § is defined in Im r >-1, analytic in Im r>-1 with values in
H-'(0, wD. Furthermore for each R'> R, there exists &1 and 82 such that
8=81+8z
where &1, respectively 82, is defined in Im r>-1, analytic in Imt>-1
with values in L2(10, D), respectively H '(J0, wD). In addition there exists a
constant C such that

(7,2,1,14)
for every 123-1.
Proof The assumption that e 'g€ H-'(B) implies that
e'g=fo+D.fit Dofa (7,2,1,15)
where f;€ Lz(B), O<j€2. Now we fix R'> R; then with the help of a
cut-off function we can modify the f, in order that they vanish for t≥ R'.
Taking the Fourier transform in (7,2,1,15) we derive that
§(7-1, 0) = folt, 0) + irfi (7, 0) + D.$2 (7, 0),
where the f, are defined in Im + ≥0, analytic in Im + > 0 and there exists a
constant C, such that:
0<i<2
for every T2>0. One obtains the desired result by setting
§i (r, 0) =i(r+i)fi(r + i, o)
82(7, 0) = fo(r + i, 0) + Dofz(r + i, 0).
Observe that this method of proof lets & and £2 depend on R'. •
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 311

Going back to the problem (7,2,1,9) (7,2,1,10), we know that § is

analytic in Im + > 0, while & is analytic in Im + > - (k + 2). Consequently 0
has an analytic continuation to the domain
{-(k+2) < Im TInD
where D is the set of the regular values defined above. We still denote
this continuation by . Furthermore, from (7,2,1,13), (7,2,1,14) we shall
derive some growth condition on 0.

Lemma 7.2.1.6 Assume that (7,2,1,3), (7,2,1,6), (7,2,1,9) and

(7,2,1,10) hold and that o vanishes for t > log R. Then there exists K such
that
k+4
1.u.b.
-(k +2) 57250 [laK i=0
(7,2,1,16)
Proof The main step is to find a bound for & in term of § at least for
large values of Til. This is straightforward for k =-1 and k = 0. Indeed we
calculate (DZ'0; § for 1 = 0, 1, 2, where the brackets denote the pairing
between distributions and functions in JO, [. Integrating by parts, we find
a constant C and a number K such that
k +4
(7,2,1,17)
i=0
for k=-2, -1, 0, In > K, - (k +2) €72€0.
It is not possible to estimate further derivatives of ^ by mere integra-
tion by parts. We shall prove the corresponding inequality later.

Lemma 7.2.1.7 For every nonnegative integer k, there exists a constant C

and a number K such that the solution ^ of (7,2,1,9) (7,2,1,10) verifies
k+4
i=0 (7,2,1,18)
i=0
for lil> K, -(k +2) €72€0.
It is clear that (7,2,1,16) follows from (7,2,1,13) and (7,2,1,18) when k
is nonnegative. Now let us look at the case when k is -1. The inequality
(7,2,1,17) implies in particular that the problem (7,2,1,9) (7,2,1,10) is
well posed for til > K, - 1€72 €0. Thus we can write
312 A MODEL FOURTH-ORDER PROBLEM

where
(+4+272+1)8: +(2-272) D30; + D60; =8;
j=1, 2, with &1 and 82 given by Lemma 7.2.1.5 and ¢1, ¢2 fulfilling the
boundary conditions (7,2,1,9). The inequality (7,2,1,16) follows by apply-
ing (7,2,1,17) with k =0 to ⅙, and with k=-1 to dz. •
Proof of Lemma 7.2.1.7 First we consider an auxiliary problem on the
half-line R4 = J0, ∞[: wE H? (R.) is a solution of
TiW-277w"+w=h in R+. (7,2,1,19)
It is clear that we Hk+ (R) when g is given in HK (R). In addition, for
TI=1, there exists a constant Kk such that
(7,2,1,20)
Then we observe that replacing A by Or reduces the equation
(7,2,1,19) to a similar one where = 1. Performing this change of
variable in (7,2,1,20) leads to
k +4
(7,2,1,21)
for all TiER.
Next it is easy to check that the same inequality holds for w€
H=(0, D solution of the same equation in the interval 10, w[. (Use a
cut-off function and continuation by zero, then apply inequality
(7,2,1,21)).
Finally we can consider ^ solution of (7,2,1,9) (7.2.1.10). We observe
that w = ^ is solution of

+10-2750" + @liv) =h in JO, c[

where
h=8+(+1-74-272-1)8+(272-2-273) D38.
Consequently we can apply the previous inequality to d. We get
k +4
2 IT, /K+4-i/lOll; 2.10,01 € K'k
i =(

j=0
since we assume that -(k + 2) < T2 € 0. It is now clear that we can choose
K large enough such that (7,2,1,18) holds. •
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 313

Remark 7.2.1.8 The inequality (7,2,1,18) is also a particular case of

some more general inequalities proved in Agranovitch and Visik (1964).

We shall now consider the continuation \ of the solution @ of problem

(7,2,1,9) (7,2,1,10) on the horizontal line
72=-(k + 2).
The function \ is well defined almost everywhere on this line provided
there is no characteristic value (for the problem (7,2,1,11)). In addition, it
follows from (7,2,1,16) that
• +00 k+4

L. S
This inequality implies that

is the Fourier transform of a function

1 +00

which belongs to Hk+4(B).

It is easy to compare v with w. Indeed from (7,2,1,7) it follows that

0(5, 0) =J27). - 00

and by Cauchy's formula that

0 (1, 0) =e'k+2) w(t, 0)t Sm (t, e) (7,2,1,22)
-(k+2) SIm Im 50
where Tm m = 1, 2, ... denotes the sequence of the characteristic values
and Sm is the residue of

at T = Tm.
We shall now calculate these residues. Some additional understanding
of the characteristic values is necessary for this purpose.

Lemma 7.2.1.9 Let Tm be any characteristic value of the problem

(7,2,1,11); then for + = Tm the solutions of the problem (7,2,1,11) span a
one-dimensional space. In addition let § be any analytic function in a
neighbourhood of Tm with values in H-(J0, wD, then the corresponding
solution of the problem (7,2,1,9) (7,2,1,10) has the following Laurent
314 A MODEL FOURTH-ORDER PROBLEM

expansion near Tm:

(a) Ô (т, 0) = Um (0) + We. (7, 0) (7,2,1,23)

T- Im
where Im is a solution of (7,2,1,11) with + = Tm and Wm is an analytic
function near Tm with values in H300, wD, provided m is a simple
root of (7,2,1,12).

(b) Ô (T, e)=

- 4. 2 + 9mT-0)Tm
+ W. (7, 0)
(7,2,1,24)

with similar properties for Im and we

'ms While Im m is a solution of

(7,2, 1,25)
in 10, [, with the boundary conditions (7,2,1,9) provided Tm is a
double root of (7,2,1,12).
Finally, the equation (7,2,1,12) has no root with a multiplicity
larger than two.

Proof It is readily seen that any solution of the equation is (7,2,1,11)

which fulfils the boundary conditions at zero, is a linear combination of u
and uz defined below:
sin A sinh +0
U (r, 0) = T †‡ 0
u (0, 4) = A sin A
Uz (t, A) ==
cos A sinh re
T - sin o cosh to. r‡0, ti
42(0, 0) = A cos 0 - sin O
(sin 20
u/(=1, 01=312 - 0.
These functions are entire analytic functions of r. Then it follows from
the general results about the two-point boundary value problems that ¢,
the solution of (7,2,1,9) (7,2,1,10), is such that dû is analytic near Tm
where d is the determinant
d(7) =U1 (7, CU) DOUz(T,W0) -42(7, W) DOMA (I, C0)

= sinh? tw
The zeros of d are described in Lemma 7.2.1.4. They are either the
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 315

solutions of the equation (7,2,1,12) with +70, ti or i in the particular

case when w = tan w or w = 2m. In addition the order of the zeros of d is
the multiplicity of the solutions of (7,2,1,12) when + + 0, #i, while #i is a
simple zero of d when tan w=w Or w = 2 Tr.
Differentiating the identity (7,2,1,12) with respect to r shows that the
multiplicity of the solutions is at most two.
If we assume that d has a simple zero at Tm then § has a simple pole at
Tm and consequently (7,2,1,23) holds. Applying the differential operator
(+* +272+1) +(2-272) D6+ D$ = I(T, Do)
to both sides of this identity multiplied by (r- Tm) yields
(T-Tm) 8=I(T, Do)Um + (7-Tm) I (T, Do) Mm.

It follows obviously that L(Tma De)m = 0. The boundary conditions on m

are obvious and thus m is a solution of the homogeneous problem
(7,2,1,11) at T = Tm-
Let us now assume that d has a double zero at Tm .. then ^ has a double
pole at Tm and consequently (7,2,1,24) holds. Applying the differential
operator L(r, D.) and multiplying by (7 - Tm)?, we obtain
(T-Tm) 8 =I(T, Do)Im + (T-Tm) I (T, Do) Am + (7 - Tm) I (T, Do) Mm.
Again it is obvious that L(Tm Do) m = 0 and that Um fulfils the boundary
conditions in (7,2,1,11). Next we have

(T-Tm) 8 =
T -Im
and consequently

L(T; Do) Am =

Taking the limit when + -> m implies the equation for Ims namely

Again, the boundary conditions on 4m are obvious. -

Remark 7.2.1.10 The existence of m solution of (7,2,1,25) with the

boundary conditions (7,2,1,9) is not obvious since Tm is a characteristic
value. Accordingly we must check that the right-hand side

is orthogonal to the kernel of the transposed problem. In other words we

must check that this function is orthogonal to every function n which is a
316 A MODEL FOURTH-ORDER PROBLEM

solution of
L(im; Do)n=0 in JO, c[
n(0) =n(c) =n'(0) =n'(c) =0.
An auxiliary result for this verification, is the following.

Lemma 7.2.1.11 The double solutions of the equation (7,2,1,12) are all
imaginary.

Proof All the solutions of (7,2,1,12) are solutions of either

sinh (Tw) = + sin w (7,2,1,26)
Or
sinh (7w) =-+ sin w (7,2,1,27)
Let us consider the first of these equations, for instance, and assume
that Tm is a double root. We have
sinh (Tm() = Im SiN W,
together with the differentiated equation
w cosh (Tmw) = sin c.
If we denote by 'm and Mm the real part and imaginary part of Tm
respectively, we derive
sinh (Em) cos (Mmw) =5m sin c
cosh (Emw) sin (NmW) =Mm sin c
w cosh (Emw) cOS (Nmw) =sin c
sinh (5mw) sin (Amw) =0.
From the last equation, it follows that we have either 5m =0 Or Mm =
km/w, where k is an integer. Assuming that §m + 0 and accordingly that
Mm = kar/w, it follows from the second equation that k = 0, i.e. Mm = 0.
Then the first equation yields
sinh (Emw) sin w

This equation is impossible since we have (sinh t)/t|≥ 1 for every t, while
we have (sin w/w| 1.
In conclusion, all the double roots are such that 5m = 0. -
Since Tm a double root, is imaginary, we have im =-m and conse-
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 317

quently L(Im; D.) = L (im; De). Going back to the existence condition on
Im we must check that L (Tm Do) m is orthogonal to all the solutions of
the problem (7,2,1,11) with 7 = Tm.
Let u(r, A) be the solution of
( L(r; Do)u =0 in 10, col,
u (т, 0) =0,
Dou (7, 0) =0,
u (T,w) =0,
(Dóu (7, 0) = 1.
It is easily seen that u exists and is unique near each characteristic value
n. The function 4m (0) is a scalar multiple of u(Tm e), since the space of
Tm:
the solutions of the problem (7,2,1,11) is one-dimensional. Let us
differentiate with respect to ~ the identity

LIT: Do)lu(r: Oru(r: 0) do=0.

This yields

At + = Tm, We obtain

Consequently L'(Im De)u (Im; A) is orthogonal to u(Tins 0) and the same

way La(Ima De) m (e) is orthogonal to Im in L2(10, wL). This shows that
the solution m n of equation of (7,2,1,25) with the boundary conditions of
(7,2,1,9) actually exists.
Now going back to identity (7,2,1,22), we have
u(t, 4) = e'k+2) w(t, A) +
-(k + 2) <Im rm <0

+ iv (21) e'TinIAm (A) +it/m (0)},

-(k +2) <Im Th CO

where we denote by in the characteristic values which are simple and by

'm the double ones.
Summing up, we have proved the following statement where we have
performed the change of variable r = e' and used Lemma 7.2.1.3.
318 A MODEL FOURTH-ORDER PROBLEM

Theorem 7.2.1.12 We assume that u € P-(G) is a solution with bounded

support of
4Zu =f in G
with the boundary conditions
ди
=0 on AG

We assume in addition, that fe P'(G) with k >-1 and that the problem
(7,2,1,11) has no characteristic value on the line
Im r =-(k +2).
Then u =u, + us, where u, E PX*(G) and
u, (r; A)=

+
- (k +2) <Im T' <0

- (k + 2) < Im T." <0

n (0) +in r) m (e)s (7,2,1,28)

where Th, m = 1, 2,.. denotes the sequence of the simple roots of the
characteristic equation (7,2,1,12) (augmented with #i when tan w = w),
" m = 1,2, ... denotes the sequence of the double roots of equation
(7,2,1,12), Um is a solution of (7,2,1,11) with + = I'm Or i'm and finally
Am is a solution of (7,2,1,25) and (7,2,1,9) with +=rm

This result holds in particular when f is given in Hk (G) by Theorem

7.2.1.1. It implies that u belongs to Hk+4(G).
In what follows it will be convenient to restate the expansion (7,2,1,28)
in a slightly different way. We define the function Sm by

(In +25m+1)Sin +2(1-55)S'+Sm =0 in 10, w[

15m (0) =Sim (W) =5in (0) =5'm (c) =0 (7.2,1,29)
with the normalization condition

[Is. (0)F do =1 (7,2,1,30)

and we define the function Om by

(rm* +25m2+1)Om +241-am2)omtom
in 10, wl
Om (0) =Om (W) =0'm (0) =Om (W) =0 (7,2,1,31)
 7.2 SINGULAR SOLUTIONS, THE L2 CASE 319

with the orthogonality condition

Sm (0)Om (0) de =0. (7,2,1,32)

The functions Sm and Om are uniquely determined and Um is a multiple

of Sms say,

for some complex number Am; then obviously iN (27) mn- AmOm is a
multiple of Sm say,
iN (27) Am =AmOm+MmSm
for another complex number Um Accordingly we have

t
Us (r, 0) = Am''tit,'''Sm (A)
- (k +2) <Im Th <0

Amiltiti.Om(0) +i(In r)Sm (0)}+ Ump'tiriSm (0).

- (k + 2) <Im rm <0
(7,2,1,33)

Remark 7.2.1.13 When the assumption on the characteristic values of

the problem (7,2,1,11) is not satisfied, one can prove some partial results.
Indeed if there is some characteristic value on the line Im 1 =-(k + 2),
there exists & >0 arbitrarily small such that there is no characteristic
value on the line Im 1 = - (k + 2) + €. Again

is the Fourier transform of a function belonging to Hk+3(B). Conse-

quently we can replace (7,2,1,22) by

u(t, e) =e k+2-8)p(t, A)t Sm (t, e)

-(k + 2) += <Im Im <0

where p€ Hk+3(B). Then the corresponding expansion in Theorem

7.2.1.12 implies that u = u, + Us, where u € Pk*3(G) and u, is given again
by (7,2,1,28). This is not the best possible result but this will be techni-
cally convenient in Subsection 7.3.2.

Remark 7.2.1.14 It is easily checked that Sm is proportional to the

function
- sinh Tmw{sinh ImO sin (0 -w)} + Im sin c{sinh Tm (0 - w) sin e}
and that Om is a linear combination of the function above and its
320 A MODEL FOURTH-ORDER PROBLEM

derivative (with respect to m), i.e. the function

-w cosh rmw<sinh rmA sin (A - w)} + sin c{sinh m(0 - w) sin 0}
- sinh rmw{0 cosh r mO sin (0 -c)}
+ I'm sin w{(0-w) cosh i'm (0 - w) sin 0}.

Remark 7.2.1.15 One can derive similar results for the boundary value
problem (7,2,1,3) (7,2,1,4) in a strip B whose width is & = 2m. This takes
care of fracture problems. The characteristic values are the numbers
1=-im/2 with m an integer (m + 0). The multiplicities are 2 unless
m =2. It is readily seen that the solutions of the homogeneous problem
(7,2,1,11) span the two-dimensional space generated by the functions

8m (0) =sin (1 + m/2)0-

m- sin(-1+ m/210

and hm (0) =cOS (1+m/2)0-cos (-1 + m/2) 0.

This implies that the function &m defined by (7,2,1,23) is a linear

combination of 8m and hm• On the other hand, when Tm is a double root,
Um defined by (7,2,1,24) vanishes while m n is again a linear combination
of 8m and hm (this follows from the solvability condition for the problem
(7,2,1,25)). Consequently the expansion (7,2,1,28) is simply:

Us=0<m <2(k +2)p'+m/2@am8m (0)+Bmhm (0)),

where am and Bm are constants. Actually only the terms corresponding to
odd values of m are relevant in this expansion (since the other terms are
simply polynomial). Thus it will be convenient to relabel everything by
replacing m by 2m - 1. The expansion (7,2,1,33) now has the following
form:
Us- Am"m+1/2sm' (0) +Umim+1/25M(0),
1=m <k+512
where
2m+1
sm' (0) = sin (m + ½)0
1-2m -3sin (m -3)8

and sin' (0) = cos (m +½)0-cos (m -2) 0.

Again we have u - UsE Pk*3(G) according to Remark 7.2.1.13.
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 321

7.2.2 Getting rid of the weights

First we go back to the original problem (7,1,1) in a polygon 2. This will
be merely a matter of notation. Again we denote by I 15j&N, the
sides of M, S; being the final point of I;. We denote by w; the measure of
the angle at S; and finally we use again the polar coordinates with origin
at S;.
Now we define the singular functions corresponding to each corner.
The sequence Aj.ms m = 1,2, ... denotes the set of all the roots of the
characteristic equation
sink? (Aw;) = 12 sin? c; (7,2,2,1)
excluding 0 and +i when tan w; + w; and excluding only zero when
tan w; =w; Then the sequence A j,m m = 1, 2, ... denotes the set of the
simple roots of (7,2,2,1) including +i when tan @; =@;. Finally A',mo
m = 1, 2, ... denotes the set of the double roots of (7,2,2,1). Accordingly
we have
(Ai.m'm=1.2=51.m)m=1.2URA.m)m=1.2
and the numbers I'm m are all imaginary. Next we set
Sim (ris 0;) = r| iN, mi,m (0;) n; (r; ei%.,), (7,2,2,2)
where n; is a cut-off function which is equal to one near S; and vanishes
near all the other corners and near all the sides but I; and I'+1 and where
Si.m is a solution of the equation.
(1'4 Pim + 1) Sim +2(1 -17.m)s",.m + sim = 0 (7,2,2,3)
in ]0, w;[ with the boundary conditions
Si.m (0) = Si,m (w;)=54.m(0) =S.m (60;) =0 (7,2,2,4)
and the normalization condition

simn (0)|? do = 1.
We also set
Jim (ris 0;) =r! tin,. tim (0;) ni(r; ei%,), (7,2,2,5)
where tim is a solution of the same problem as Sim with Am replaced by
I'm and
Aim (ris 0;) =r|inimfu;m (0;) +i(In r;)4,m (O,)In;(r;ei%,), (7,2,2,6)
where u;m is a solution of the equation

=-41", m (1", m +1) tim +41", m'",. m (7,2,2,7)

 322 A MODEL FOURTH-ORDER PROBLEM

in JO, w;L with the boundary conditions

Ui,m (0) =4j.m(w;) =4.m (0) =u'j.m(w;) =0 (7,2,2,8)
and the orthogonality condition
f°,Uj.m (0)t;.n m (A) de =0.
Our starting point in this section is the following statement.

Theorem 7.2.2.1 We assume that u € 12(1) is a solution of

AZu =f in h
with feik (1), k=-1. (Let us agree for convenience that I '(1) =
H-'(1).) We assume in addition that the equations (7,2,2,1) for i=
1, 2, ...,N, have no roots (other than -i) on the line
Im 1 =- (k + 2)
and we exclude the case k =-1 when tan w; = w; for at least one value of j,
15j&N. Then u belongs to the space spanned by Hk+4(1), the functions
Sim which correspond to
-(k +2) < Im !..
i.m <0
and the functions Tim and U;m Which correspond to
- (k +2) <Im d".rmo,<0.

Proof It follows from Theorem 7.1.2 that u€ Hk+(M\V) for any

neighbourhood V of the corners. Then we proceed as we did at the end
of the Subsection 7.1, considering u; = niu in the infinite sector G;
corresponding to the corner S; (after a rotation and a translation, possi-
bly). The function u; belongs to H=(G;), has a bounded support and
43u; =fi
where f; € Hk (G;) has a bounded support and coincides with f near S;. By
Theorem 7.2.1.1, this implies that
f; € P≤(G,).
We can conclude by applying Theorem 7.2.1.12 to u; 1 <j&N.
The purpose of the remainder of this subsection is to eliminate the very
unnatural assumption that f belongs to Hk (S) instead of HK (R) (this is
an actual assumption only when k≥1). This will be achieved with the
help of a new trace theorem. Later in this chapter we shall also need a
similar result in the framework of the Sobolev spaces related to Lp with
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 323

p 72. This is the reason why we state and prove this trace theorem in the
general case, at once.
Theorem 7.2.2.2 Assume that k #-1 and 1 <p <00 and let feWK(12)
and o; E WK +4-11P (Ty), 4, E WK+3-1/P(Ty), 1 5j SN; then there exists
U€ Wk+4 (12)
such that
1SjEN (7,2,2,9)
dU
=V 1SjEN (7,2,2,10)

1430-16194109 (7,2,2,11)
iff
Ф;(S;) = 4;11(S;)
24;+1 (S.) + sin w, V4 mi (S.)
дт; (5) =-COS Wi AT; +1
дФі (7,2,2,12)

dip; ' (5,) -cos c, 4441(5)

for i = 1, 2, ..., N, and iff in addition

244+1 (5.)
-cOs Wi ad2 74;# (Si) + sin wi arith
(7,2,2,13)
when p>2 or k=0 and

34, + (8, (01)?

Jo coscar дт
12 do
<+00 (7,2,2,14)
+ sin of 34 (4, (-01) -314141
ATi+1
(4, (61))) !
O
for i = 1, 2, . .., N when p = 2 and k =-1.
(We again identify N + 1 with 1; the notation 8;, *; has been introduced
in Subsection 1.5.2.)
Proof First, we must view the property (7,2,2,11) as a trace property.
Indeed 42y-f belongs to Wk (R) iff
al d'f on T; (7,2,2,15)
Vi au!+430 = Miao.
324 A MODEL FOURTH-ORDER PROBLEM

for 0€15k -1, 15j SN. Consequently we look for a function o which
satisfies (7,2,2,9), (7,2,2,10) and (7,2,2,15).
We shall solve this problem by applying Theorems 1.6.1.4 and 1.6.1.5.
Accordingly, we have to define the operators Bi. which are involved in
these statements. We set

B;.1=1
д (7.2.2.16)
B;2= du;
21-3
B;=do! = = 43, 3515k+2.
The degree of B, is clearly dit = 1-1, when I = 1, 2 and d;. = 1 + 1 when
1>3. The corresponding functions fu are

(7,2,2,17)
21-3f 3515k+2.
1-39

Next we have to find the operators P. and Qi+1, such that (1,6, 1,1)
holds. Due to Remark 1.6.1.8, we look for operators which are
homogeneous and have constant coefficients. Consequently we have
d-d.i
Pit = a, Qi, = b..(aT; +1
дт;
where for simplicity we do not make explicit the dependence of a and b,
also on i and d. The corresponding identity (1,6,1,1) reads as follows
(when d ≥4).

+azaT;
d d-1du;a -k+2 121 d-1-1 at-3

д
ay dT. 1=3 laT;
d d-1 k+2
245542
-+ > b. d
d-1-1 21-3
=b, laT; +1
+6,10
(AT:+1 AV;+11=3 laT; +1
24, 4, 4?.
This implies the identity

ay дт: +a2 at;

д d-1 a d-1 a = R43,
du; laT; +1 - 13657., Avit1
(7,2,2,18)
where R is a homogeneous differential operator of order d - 4. Consider-
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 325

ing the corresponding symbols, this identity implies that the polynomial
a (-x cos w; -y sin w;)d + az(-x cos c; - y sin w;)d-1
Xy cos w;-x sin w;)-byxd+b2xd-1y =S(x, y)
can be divided by (x2 + y2)7.
Equivalently, this means that x = #iy are double roots of S. This yields
the following system of equations for a1, a2, b1, bz:
a, (-1)d(=i cos w; + sin w;)d + az(-1)d-'(ticos w;+ sin c;)d-1
X (cos w; Fi sin w;)-b. (‡i)d + b, (ti)d-1=0

and a, d(-1)d cos co; (ti cos w, + sin w;)d -1

+asi (d-1) (-1)d-1 cos w; (=i cos w; + sin w;)d-2(cos w; Fisin co;)
+(-1) (ti cos w; + sin w;)d-' sin w;}
-bad (ti)d-1+b/(d-1) (‡i)d-2=0.
An easy but lengthy calculation shows that the corresponding deter-
minant is proportional to
d(d-2) sin? w; -sin (d -2) c; sin dw;.
Consequently the determinant does not vanish and the only solution of
the system is the null solution.
Summing up, in the particular case under consideration here, there
exists no nonzero operators Pit and Q;+1,, such that (1,6,1,1) holds with
d≥4. Let us now consider the cases when 0d <4.
First, when d =0, we look for numbers a and b, such that a, = bi; the
corresponding relation (1,6,1,2) is
Ф; (S;) = 4;+1 (S;). (7,2,2,19)

a a
Then, when d = 1 we look for numbers a,, az, b1, b2 such that

ati a2 debratits
as @+ avi +62 AVitt
This can be any homogeneous first-order operator; the corresponding
relations (1,6,1,2) are

дт;
(7,2,2,20)

44(5) = - sin a 97 ' (5) - cos a, M4. 1651

since Aldt, and a/av; generate ail the first-order operators.
326 A MODEL FOURTH-ORDER PROBLEM

Next, when d = 2 we look for numbers ay, az, bi, b, such that

a,ддт; 2+a2 ar;22dvi=b, 2 22

) +628794104451
Equivalently, we look for numbers a1, 02, B1, B2 such that
\2 22 2 22
ay = B. (D
IdT; IdT; +1
+a2 any atitt
Obviously, we have a, = B, =0 and @2 = B2. The corresponding relation
(1,6,1,2) is
Ali (s.).
- cOS Wi39i+
aT; +1(S,) + sin asdil: + (51) =-cos wi841
ariti at? (5)-sin wiari
(7,2,2,21)
Finally, when d = 3 we look for numbers ay, a2, b, b, such that
†az(2 12 2
\дт,/ ди;bilarita dVit1
ідт;
Obviously the only solution is a, = a2 = b, = b2=0, and there is no
corresponding relation (1,6, 1,2).
In conclusion the image of Wk+4(8) by the mapping
{y;BiiisisN. 1515k+2
is the subspace of

¡=1 \1=1
defined by the conditions (7,2,2,19) (7,2,2,20) and (7,2,2,21), when k=0
or when k =-1 and p >2. When k=-1 and p <2, only the conditions
(7,2,2,19) and (7,2,2,20) occur, while, in the limit case k =-1 and p = 2
(7,2,2,21) is replaced by the corresponding integral condition (the pattern
being that (1,6,1,3) replaces (1,6,1,2)). This implies the claim of Theorem
7.2.2.2. _
Going back to the particular case when p =2, we get the following
consequence of Theorems 7.2.2.1 and 7.2.2.2:

Theorem 7.2.2.3 Assume that u € H2(2) is a solution of

AZu =f in a
Yiu = 4; on lj 15j<N (7,2,2,22)
ди
 7.2 SINGULAR SOLUTIONS, THE Lz CASE 327

with feH" (8), Q;E HK+712(Ti), 4; E HK+SIZ(E;), k=-1 that (7,2,2,12)

holds and that (7,2,2,13) holds when k≥0. In addition we assume that
(7,2,2,14) holds when k=-1. Finally we assume that the equations
(7,2,2,1) for i = 1, 2, ...,N have no root (other than -i) on the line
Im 1 =-(k + 2) and we exclude the case k =-1, when tan w; = w; for some
¡. Then u belongs to the space spanned by Hkt4(9), the functions Sim
which correspond to Im Min E]-(k + 2), OL and the functions Tim and
W; m which correspond to Im A'", m € J-(k +2), OL.

Proof We merely apply Theorem 7.2.2.1 to

W= U-v
where U€Hk+4(8) is a solution of (7,2,2,9) to (7,2,2,11) given by
Theorem 7.2.2.2.
Remark 7.2.2.4 When the condition that no root of the equations
(7,2,2,1) lies on the line Im=-(k + 2) is not fulfilled, then from Remark
7.2.1.13 we conclude that u belongs to span of Hk+3(1) and the
functions Sims Jim and U;m corresponding respectively to Im 1',.m €
[-(k + 1), OL and to Im A,m € [- (k + 1), O[ (see also Theorem 1.4.5.3).

Remark 7.2.2.5 If we allow cuts i.e. w; =2m for some j, then the
assumptions (7,2,2,12) to (7,2,2,14) must be replaced by the following:

(OP: (5) = (-14 849, + (S.). 12k+4-3

, дт; ) aT:+1 p (7.2.2.22)
944 (51) ==6-19 2441. (5,, 14+3-
(дт; p
for every p and by

" log (41 -01) -4-1' do ' 44601 dare tea.

1=k+3
[.9 1 999 x: (-01) +4-19-8'44+4(84601)
athi 1°Odo <+00, 1=k+2
(7,2,2,23)
for p =2. This result follows from Section 1.7.
Then we set 1;,m =-i(m -2) (m an integer) when w; = 2 and
Sim (r;, 0,) =pm+113s m (0;) ni(r, ei%,), i=1,2
where sm, i = 1, 2 have been defined in the Remark 7.2.1.15.
The statement corresponding to the Theorem 7.2.2.3 is now that u, the
328 A MODEL FOURTH-ORDER PROBLEM

solution of the problem (7,2,2,22), belongs to the span of Hk+3(1) and

the functions
Sims Tim and A;m corresponding to Im 1'.m (Or M',m) €] - (k + 2), OL
when w; <2m and the functions
Sim, Sim corresponding to m <k+3
when c; =2m. This holds provided feIk (1), y; eHk+712(I;), 44, €
Hk+s/(Th), (7,2,2,12) and (7,2,2,13) hold when w; <2m, and (7,2,2,22)
and (7,2,2,23) hold when w; = 2t.

7.3 Singular solutions, the L case

7.3.1 A priori inequalities

Assuming that 1 <p <+0 and p+ 2, we shall prove the existence of a
constant C such that
(7,3,1,1)
for all u € Wk+* (8) with v;u = 0 and y; du/dv; = 0 on I;, k ≥-1, provided
some conditions are satisfied by the angles. This inequality is similar to
inequality (4,3,2,12) and we shall follow the same method of proof.
Consequently the first step is the proof of the corresponding inequality for
u belonging to the weighted space Pk+4(1) (see Definition 4.3.2.1). This
is done locally by considering first the equation AZu = f in an infinite
sector G.
Thus we consider
u € Pk+4(G)
such that
AZu =f€ Pk(G)
with the boundary conditions yu = y du/d = 0 on the boundary of G,
which is the infinite sector defined, in polar coordinates, by
G = {(r cos A, r sin e); r > 0, 0<0 <w}.
We use the change of variable r =e' in order to obtain an equation in the
strip B = RX]0, w[. Setting
(u(r cos l, r sin 4) = 0(t, e)
f(r cos e, r sin e) = 8(4, 0)
 7.3 SINGULAR SOLUTIONS, THE Lp CASE 329

we obtain the equation

(D4-4D3+4D2+2D{D3-4D,D3+D'+4D3)U = eH.
However, due to Lemma 7.2.1.3, it is more natural to consider

Th=e-k+2/D g€ WK (B)
and the corresponding equation is
[(D, -p)4-4(D, -p)3 + (4+ 2D3)(D,-p)?
- 4D}(D, - p) + DA+ 4D⅔]w = h (7,3,1,2)
in B, where p=-k -4+2/p (i.e. w = e').
Now we use again the method of Section 4.2 We study the well
posedness of the problem by performing a partial Fourier transform in t.
Thus w(r, e) is a solution of
([(i7-p)7-4(ir-p)3+4(ir-p)?]w
+12(17-0)2-4(7-0) +41w' + wliv)=h in 10, c[ (7,3,1,3)

for every TeR, where the superscript ' denotes the differentiation in 0.
Here we assume in addition that w has a bounded support (in t) in order
to give a meaning to its Fourier transform everywhere in .
In the particular case when p = 2 and k =-1, the problem (7,3,1,3)
coincides with the problem (7,2,1,9) (7,2,1,10). In the general case let us
set
8=+r+i(p +1),
then the equation in (7,3,1,3) is just
(84+282+1)n +(2-283) W"+ ww = h.
Consequently the well posedness of the problem (7,3,1,3) has been
investigated in Lemma 7.2.1.4.

Lemma 7.3.1.1 The problem (7,3,1,3) has a unique solution for every
real riff the characteristic equation
sinh? (Aw)=12 sin? c (7,3,1,4)
has no solution on the line Im^ =-(k+1+2/g) (assuming p72 and
k»-1).
From now on, we assume that this condition is fulfilled. The solution of
the problem (7,3,1,3) can be written down explicitly through the use of a
 330 A MODEL FOURTH-ORDER PROBLEM

Green function:

where the kernel K is smooth in r and three times continuously differen-

tiable in e and o'.
More precisely, let us set
a (e) = sin A sinh 80
B (A) = sinh 80 cos 0 - 8 cosh 80 sin A
Y(e) = sinh 80 cos A + 8 cosh 80 sin A
and
8=82 sin? w-sinh? Ew;
then we have
K(8-i[p +1]; 0, 0)
=8-15B (W)
128 a (A)a (w-e')-28a (A)B(w-A')
a (w)

y(w)
-2841 + 83 B(0)B (w-0) - 28
(al) 8(0)a (00-038
for 0<058=w
and g-11B (0)
128 a (w-A)a (0) - a 28
(W) B(00 -0) a (0)
y (w) a(w)
-284+83) B (W-0) B109-28
for 0<6'sA=w.
It is not hard to check that the kernels K, K and D.K fulfil the
conditions of Lemma 4.2.1.3. This yields the following inequality.

Theorem 7.3.1.2 Assume that p+2, k>-1 and that the characteristic
equation (7,3,1,4) has no solution on the line Im 1 =-(k +1+2/q). Then
there exists a constant C such that
wllo.p.B <C|hllk.p.B (7,3,1,5)
for every we Wk+4 (B)n W…(B) such that (7,3,1,2) holds.
Proof When k>0, we apply directly Lemma 4.2.1.3 and actually show

that wo.p.B S|/hllo.p.B.

 7.3 SINGULAR SOLUTIONS, THE Lp CASE 331

In the particular case when k=-1, we write

h = got Digit D.82
with g; € Lp(B), 0-i =2. Consequently we have

since K(r; 0, 0) = K(r; 0, w) = 0. Then applying Lemma 4.2.1.3 three

times, we obtain

and consequently
wllo.p. B < C' Thll-1.p.B
for some other constant C'.
Actually these estimates have been derived only for a w which has a
bounded support in t, but it is easy to deduce the general case by taking
limits. Indeed the functions with bounded support are dense in
wk+4(B)n W?(B) (use cut-off functions).
Inequality (7,3,1,5) is just the analogue of inequality (4,2,1,4). Now we
must find bounds for the derivatives of order k + 4 of w. We proceed as in
Subsection 4.2.2, namely we will neglect some non leading terms in the
equation (7,3,1,2). For this purpose we rewrite the equation as follows
(7,3,1,6)
where clearly there exists some constant such that
Th.lik.p.BShlk.p.B+C|WIlk+3.p.B. (7,3,1,7)
Lemma 7.3.1.3 There exists a constant C such that
/wll+a.p.B SCIllAwllk.p.B t|/wll+3.p.B}
for every we Wk+4 (B)n W?(B).
Before proving this lemma, let us take some preliminary steps. We shall
start from the inequality
412.p,B = CollAullo.p. + lulla.p.B} (7,3,1,8)
which follows from (4,2,2) (see Theorem 4.2.2.4) for every u€
WE(B)n W(B). It implies the following.
332 A MODEL FOURTH-ORDER PROBLEM

Lemma 7.3.1.4 For each integer k >O there exists a constant Ck such that
lullk+2.p.BSCklAullk.p.B+|/ullk+1.p.Bs (7,3,1,9)
for every u € Wk*2(B) n W|(B).

Proof Since we already know this result for k =0, we can proceed by
induction. Thus we consider u € Wk+(B)n W(B), assuming that Ck-1
exists. We have to estimate the Lp norm of the derivatives of order k + 2
of u. Applying (7,3,1,9) to
D.U € Wk+ (B)n W!(B),
we get the desired bound for all these derivatives but D&+Zu. We
conclude by writing
D×*Zu = D\Au - DìD…u.
It follows that
MIDk*Zullo.p.Bs||Aulk.p.Bt||D.ullk+1.p.B.E=

Now using the trace theorem it is easy to deduce the following result:

Corollary 7.3.1.5 For each integer k >0 there exists a constant Lk such
that
(7,3,1,10)
for every u € Wk+2(B).

We recall that Fo and F, denote the two components of the boundary

of the strip B, i.e.
E=RX{ja}, j=0,1
We are now going to apply the inequality (7,3,1,10) to 4w and then to
w. Clearly we have
Wilk+4.p.B€Ix+214wllk+2.p.Bt|w|k+3.p.B}. (7,3,1,11)
Then we have to estimate |4 wIlk +2.p. B. We will apply (7,3, 1, 10) to w now.
We get

(7,3,1,12)
The last step is the proof of an estimate for y; Aw = y;D{w (since
Y;D;w=D=y;w=0) for j=0, 1.
 7.3 SINGULAR SOLUTIONS, THE Lp CASE 333

Lemma 7.3.1.6 Let 4 € H3(B) be such that

(-4+174=0 in B
then we have
(VODEW) (7) = p? (VoG) (7) + 2p(sinh? po-p'c73-1

+ [sinh pw -pw cosh pa](y Do$"(r)

-p sinh po (yo$" (r)
+ [pw-sinh pw cosh pw] (yoD.$)"(7)}
where p=W(1+73).

Proof It is a simple calculation: we write that ¢, the Fourier transform of

q in t, is solution of the differential equation
(-DZ+1+7376=0 in 10, c[.
Consequently we have
@(r, e) = (a (r) + b (7)0) sinh pl + (c(r) + d (r) e) cosh pl.
The explicit value of the functions a, b, c and d is obtained by substitut-
ing the above expression for 4 in Yo, v14, Yo Da4 and Y, D.4. -
Obviously we have a similar formula (mutatis mutandis) for
Actually we need a consequence of Lemma 7.3.1.6:

Corollary 7.3.1.7 For every k ≥-1, there exists a constant Kk such that

Ily; Awll +2-11p.p.F, SKkI-4 +1)wilk.p.B (7,3,1,13)

1=0, 1, for every we WK+ (B) n W}(B), p=2.

Proof Let us set a = (-4 + 1) w and assume in addition that w has a

bounded support. Thus we H '(B)n W*(B). Let lE H-'(R3) n W$ (R3) be
a continuation of & out of B such that
(7,3,1,14)
for some constant C.
Then consider the elementary solution E for -4 + 1 defined by
FE(E) = (E}2+1)1, ÉER?.
We write w = E* E*ll + 4. By the multiplier theorem 2.3.2.1, we know
 334 A MODEL FOURTH-ORDER PROBLEM

that
E*E*llk+4.p.m=€Cz|k,p.523. (7,3,1,15)
Then we have o € H3(B)n Wk+4 (B) and in addition
(-4+1)2=0 in B.
Therefore we can use the explicit formula for y; A given by Lemma
7.3.1.6. Applying again the multiplier theorem (or rather its corollary,
Lemma 2.3.2.5) we obtain

Ily; Aqlik+2-1/0.8.F.€Cs
Vi a0 llk+3-1/p.P.F
(7,3,1,16)
On the other hand we have
y;4 =-y;(E*E*l)

Vião
and accordingly

§CAllE*E*llk+4..,B. (7,3,1,17)
Vi 20 lx +3-1/P.P.Es
Summing up the inequalities (7,3,1,14) to (7,3,1,17) imply that
Tv; Awlk+2-1/p.p.E, SlY; AE*E*llk+2-1/p.p.E, + ly; Adllk+2-1/P.B.E,
§CslE*ExIlk+1.p.BSC2Cs|llk,p.583

This is the desired result when w has a bounded support. The general
case follows by a density argument. S

We are now able to derive the Lemma 7.3.1.3.

Proof of Lemma 7.3.1.3 From (7,3,1,11) and (7,3,1,12) we easily find a

constant C such that

Then applying inequality (7,3,1,13) to estimate y; 4w, we obtain the

desired result.

We conclude this subsection with a result which summarizes Theorem

7.3.1.2 and Lemma 7.3.1.3:
 7.3 SINGULAR SOLUTIONS, THE L, CASE 335

Theorem 7.3.1.8 Assume that 1 <p < +0 and p+ 2, k>-1 and that the
characteristic equation (7,3,1,4) has no solution on the line Im1 =
-(k +1+2/q). Then there exists a constant C such that
wlk+4.p.B €C lIhllk.p,B (7,3,1,18)
for every we Wk**(B)n W(B) such that (7,3,1,2) holds.

Proof This is a direct consequence of the inequalities (7,3,1,5) (7,3,1,7)

(1,4,3,2) and Lemma 7.3.1.3.
Going back to the original coordinates (r= In t) we have proved that

provided u€ Pk+ (G) and yu = y du/av = 0 on 2G. Finally with the help
of Theorem 7.1.2 and a partition of unity we deduce the following:

Corollary 7.3.1.9 Assume that 1<p<00, p+2, k>-1 and that the
equations
sinh? (Aw;) = 12 sin? w;
have no solution on the line Im 1 =-(k + 1+2/g) for any i = 1, 2,...,N.
Then there exists a constant C such that the inequality (7,3,1,1) holds for
every

Proof First we obtain directly the inequality

Then by Theorem 4.3.2.2 we know that Pk+*(2) has just a finite

codimension in Wk+4(1). This implies the inequality (7,3,1,1) with,
possibly, another constant (see the method of proof of Theorem
4.3.2.4). -
7.3.2 Smoothness
We extend now the results in Theorem 7.2.2.3 to the general case
1<p<+00.
Theorem 7.3.2.1 Assume that u € H? (1) is a solution of
AZu =f in Q
Vil = 9; (7,3,2,1)
ди
= U;
336 A MODEL FOURTH-ORDER PROBLEM

with FEW: (R), 4; € WK +4-119 (55), 4; € WK+3-11P(Th), K=-1 such that

(7,2,2,12) holds in any case and such that (7,2,2,13) holds when either
k=0 or p>2 and (7,2,2,14) holds when k=-1 and p =2. Assume in
addition that the equations (7,2,2,1) for j = 1, 2, . .., N have no root (other
than -i) on the line
Im 1 =-(k +1+2/g)
and exclude finally the case k=-1 when tan w; =w; for some j. Then
u belongs to the space spanned by Wk+ (1), the functionsSim Which
correspond to

(7,3,2,2)

and the functions Tim and U;i.m which correspond to

imE - (k+1+5), 0 (7,3,2,3)

Proof First we approximate the data of the problem (7,3,2,1) by better

ones. Indeed Theorem 7.2.2.2 shows that there exists v € Wk+4(1) such
that
(y,0 = 9; 1<jEN
dU
(ar =Vis
Then Theorem 1.4.2.1 implies that there exists a sequence vi, l = 1, 2, ...
such that ye C°(R) and v, -> o in Wk+(M); the corresponding traces
(which are smooth)
4:1=7;06 1sjEN, 1>1
du, 15jEN, 1>1
converge respectively to 4; and 4; in Wk+4-1/P(I;) and Wk+3-1/P(F;). In
addition Theorem 7.2.2.2 implies that they fulfil the conditions (7,2,2, 12)
and (7,2,2,13) (7,2,2,14) when suitable. We also approximate fin Wk(R)
by a sequence fiE C° (R), 1 = 1, 2,..
Clearly we have f, € Hk+2(8). Qi1 E Hk+1112(51), 4, 1 E Hk +913(1,) and all
the conditions for applying Theorem 7.2.2.3 with k replaced by k + 2 are
satisfied provided the equations (7,2,2,1) have no root on the line
 7.3 SINGULAR SOLUTIONS, THE Lp CASE 337

Im 1 =-(k +4). Let u, € H7(8) be the solution of

AZu, =fi in A
Ville =4il on I; 1<j≤N
Al = dirt on lj 1<j^N.
Then u belongs to the span of Hk+ (S) and the functions Sims J,mo
, corresponding respectively to -(k + 4) <Im 1m <0 and - (k + 4) <
Im
only1",m <0.that
claim When the conditions
u belongs on of
to the span theHk+5(1)
roots areand
notthe
satisfied, one can
same functions
Sims Tim n and ;m; this follows from Remark 7.2.2.4.
The Sobolev theorem and Theorem 1.4.5.2 imply in both cases that
ü€E
where E is the span of Wk+4(1) and of the singular functions Sims Jim
and Am corresponding to the conditions (7,3,2,2) and (7,3,2,3).
To conclude we shall take advantage of the inequality in the following
lemma, of which we postpone the proof.

Lemma 7.3.2.2 There is a constant C such that

lines flat element I leaders one, + Elliallen vans.

(7,3,2,4)
for all u € E, where E is equipped with the natural norm

aj;.mit- (k+1-279) CA.m

- (k+1+219) <1, m <0 <0
where

-(k+1+2/9) <1. m
aj,mPi.mt-(k +1-2/9) <1.m <0
.m TO

Let us apply this inequality to up- u; this yields

N

Consequertly u, I = 1, 2, . .., is a Cauchy sequence in E; its limit u is the

solution of problem (7,3,2,1). This shows that u € E.
338 A MODEL FOURTH-ORDER PROBLEM

Proof of Lemma 7.3.2.2 First we observe that 42S; ms Vs Sim and

Ys ASi.m/avs are smooth functions. The functions Jim and m have the
same property.
Consequently

actually maps E into

N
W: (O)xII (Wk+4-1/P(F;)x Wk+3-1/0(F;).
i=1
Consequently the inequality (7,3,2,4) is meaningful.
Now we proceed by steps starting from inequality (7,3,1,1). Combined
with the obvious estimate

and Theorem 1.4.3.3, it yields the existence of C2 such that

Mullk+4.p.0€CallAZullk.p.r
for every u€ WK*4 (1)n W3(8), under the assumptions of Corollary
7.3.1.9.
Then the trace theorems imply the existence of Cs such that

for every u€ Wk+ (1). Finally the inequality (7,3,2,4) follows by aug-
menting Wk+4(M) with the finite-dimensional space spanned by the
functions Sims Jim and m n corresponding to the conditions (7,3,2,2)
and (7,3,2,3) (see the method of proof of Theorem 4.3.2.4).

Remark 7.3.2.3 Theorem 7.3.2.1 does not express a regularity result in

general, since the solution does not belong to Wk+4(1). However, this is
a regularity result when the equations (7,2,2,1) have no roots except -i in
the strip -(k + 1+21q) SIm ^ <0. Then the solution u of problem
(7,3,2,1) belongs to Wk+ (1). Therefore the behaviour of the solution of
the biharmonic equation is reduced just to the behaviour of the roots of
equation (7,2,2,1). We now mention a very useful result in this direction.

Lemma 7.3.2.4 Assume that 1 = & + in is solution of

sinh? (Aw)=12 sin? w
and assume that O <w <m; then In is strictly larger than 1, unless $ =0
and In. =1.
 7.3 SINGULAR SOLUTIONS, THE Lp CASE 339

Proof The equation is equivalent to

sinh (Aw) = 11 sin w.
Taking the imaginary part of this equation yields
cosh (Ew) sin (nw) = In sin w. (7,3,2,5)
Since the function
†->
t
sin t

is decreasing in 10, J, it follows that

sin no sin w
I no
for n€ [-1, +1]. Consequently, we have
sIn nO>1
In sin w
and identity (7,3,2,5) is impossible unless $ =0, since cosh (Ew) > 1 for
£‡0.
This lemma together with Theorem 7.2.2.1 imply the following general
principle.

Corollary 7.3.2.5 Assume that M is a convex plane polygon; then 42 is

an isomorphism from H3(8) nH?(8) onto H-'(S).

Remark 7.3.2.6 The whole Subsection 7.3.1 is valid for a strip with
width w =2m. Taking advantage of the remarks 7.2.2.5 and 7.2.1.15, and
applying the techniques of the proof of Theorem 7.3.2.1, we derive the
following statement where, for simplicity, we assume that k ≥ 0 and p+ 2.
The solution of the problem (7,3,2,1) belongs to the span of Wk+4(52)
and the functions Sims Tim and U;.m corresponding to ImM'm (or
1', m) EJ-(k+1+2/9), OL, w; <2m and S,m j,m , corresponding to m <
im, Sp(2)
2(k+1+2/g) and w; =2m. This holds provided feW (8), 4;€
WK+4-1/P(Th), 4; € WK+3-1/P(I;) and the conditions
(7,2,2,12) and (7,2,2,13) hold when w; <2t,
(7,2,2,22) hold when w; =2t,
and the equations (7,2,2,1) for w; <2m have no root (other than -i) on
the line
Im 1 =-(k +1+21q),
and p‡4, 2, ¾.
340 A MODEL FOURTH-ORDER PROBLEM

When the conditions (7,2,2,22) are not fulfilled the following additional
singular solutions must be introduced:

Fin, (r, cir, (ln r, sin 10, + 0, cos 10,)

1-2 "hr, sin (1-2)0, +0, cos (1-210) -11=25 sin (1-210,
together with
rin(r, ei?, (In r; cos l0; - O, sin 10,)-(In r. cos (1-2) 0. - 0, sin (1-2)0,)}
for I<k +4-21p.
7.3.3 The related Stokes problem
We consider here a given vector function
f= (f1, f2)
in H-'(8)3 and
V = (01, Uz)
the solution in H'(8)? of the system of equations
(-Av+p=f
v.v=0 (7,3,3,1)
in M, where p is a scalar function in 12.
The existence and uniqueness of v is well known. One can apply the
variational method, i.e. Lemma 2.2.1.1, choosing H, V and a as follows:
V is the subspace of H'(R)? spanned by the divergence-free vector
functions; H is the closure of V in L2(8)? and finally

al, w) =E Du-Tu, dx.

the space H defined above is characterized in Teman (1977). We are not
going to detail this proof of existence and uniqueness here but we rather
focus our attention on the regularity of u. Thus, assuming that f is given
in W$(8)?, we ask whether v belongs to Wk+2(1)? or not.
We reduce this problem to the corresponding one for the biharmonic
equation by considering as usual the stream function u € H7(2) defined
by
du du
(7,3,3,2)
U.= dy' U2= ax'
This function u is well defined since v is divergence-free and & is simply
 7.3 SINGULAR SOLUTIONS, THE Lp CASE 341

connected (an assumption). It follows that u is a solution of

424 =g=afs afz in A.

ду дх (7,3,3,3)
This function g is given in Wk-'(S) and we can apply Theorem 7.3.2.1.
provided k is nonnegative. This shows why we have always included the
case k = -1 in previous subsections.
It follows that there exist a function u, and numbers a ms bim and C.m
such that
U =Up t a;.mSi.m t (bimJim + Cim';.m)
-(k+2/9) <Im N,, m <0 -(k+2/g) <Im ^"

where u, € Wk*3(8), provided the equations (7,2,2,1) have no root on the

line Im1=-(k+7).
д
Consequently we have the following expansions for u and U2:

t
0, =01. +1-171 aj,mP.m
) 2x3-1 -s <Im 1. m <0

<(
(bi.mJi.m + Gi, mU;,.m ).
1=1,2, where r belongs to Wk+2(1) and where we set

s=k+?
9
If we go back to the identities defining Sim Jim and in13 1.C. (1,2,2,2),
(7,2,2,5) and (7,2,2,6), we obtain directly the singular solutions corres-
ponding to the Stokes problem. We shall use the following notation:
S2.im (r; 0,) = pitin' (iN';.m + 1) cos 0; Sim (0;) - sin 0; Sim (0,)} n; (r, ei%.,)

Slim (r;; 0,) =-rim (iN';m + 1) sin 0; Sim (0;) + cos 0; Sim (0,)]n; (r, eid,)

T2.im (%;, 0,) = piN.m (iA'"m + 1) cos 0;4m (0,) - sin 0,8,.m(0,)In; (r,@'%,)

Thim (r, 0,) =-r' (iN "),m + 1) sin 0,4;.m (0,) + cos 0,8,.m (O,)In; (r,eio,)
U2.im (ris 04) =MI(1 +in'", m)(4;.m(0,) +iln r;4.m(0;)) +itim (0,)]
Xcos 0; - [uím(0;) + iln r; tim(O;)]sin 0; }n; (r/°)

Xsin 0; + [W'.m(0) + i ln r; t'.m(0;)] cos 0; }n; (r,e'o,).

(7,3,3,4)
342 A MODEL FOURTH-ORDER PROBLEM

It is clear that {Sl.j.ms -S2.im), {TI.j.ms - Is.j.m} and {Ul.j,m U2.i.m) coincide
respectively with Vins VT;m and VA,m.
Theorem 7.3.3.1 Let ve H'(R)? be the solution of the problem (7,3,3, 1)
with f given in W(1). Assume that the equations (7,2,2,1) j=
1, 2, ...,N have no root (other than -i) on the line

ims=-(k+2)
and exclude the case k=0 when tan w; = w; for some j. Then , belongs to
the span of Wk+2(9), the functions Slim which correspond to

Im d!.
imE - (k+7), 0. (7,3,3,5)
and the functions Tim and Ulim Which correspond to

(7,3,3,6)

1=1,2.
Remark 7.3.3.2 This theorem implies that ve Wk+2(1) in the particu-
lar case when the equations (7,2,2,1) have no root (except -i) in the strip

= k+q.- <Im 1 <0. (7,3,3,7)

In particular, due to Lemma 7.3.2.4 (or the Corollary 7.3.2.5), if f is given

in L?(8)? and S is a convex plane polygon, then the solution ve H' (12)?
of (7,3,3,1) actually belongs to H?(8)3. This is the result proved by
Kellogg and Osborn (1976). In Chapter 3 we proved a similar result for
many boundary value problems for a single Laplace equation in any
convex domain and in any dimension. It would be very tempting to try to
extend the previous result on the Stokes problem to a general plane
convex domain. The technique in Chapter 3 was to take limits with
respect to M. However, here, we are unable to achieve such an extension
because we have no method of proof for an inequality similar to (3, 1,2,1)
providing good control of the constant C(S) (as a function of 12).

Remark 7.3.3.3 Writing

Vp=f-Av
and applying Theorem 7.3.3.1, yields that Vp belongs to the span of
Wk(8)? and the functions (AS1.i.ms AS2.i.m), (ATh,i,ms ATz.i.m) and
 7.3 SINGULAR SOLUTIONS, THE Lp CASE 343

(AUL.i.me AU2.i,m) corresponding to (7,3,3,5) and (7,3,3,6). By integrating,

one finds easily the singular part of p which does not belong to Wk* (1).
In particular p€ Wk+ (M) when there are no roots (except -i) of the
equations (7,2,2,1) in the strip (7,3,3,7). When feI?(8) and 1 is
convex, then p€ H'(R).
We shall conclude this subsection with a few remarks concerning the
Navier-Stokes equations. Let us first recall a now classical result (see
Temam (1977) for instance). Let f (a force) be given in L2(8)?; then there
exists a unique solution v (a velocity in H' (1)3) of
- Av + (v. V)v =1- Vp
7 .v=0 (7,3,3,8)

in S, where p (a pressure) belongs to Lz(8) and is unique up to the

addition of some constant.
We shall derive some smoothness results for v, just by rewriting this
problem as a linear Stokes system
(Av+Wp=f-(v:V)v (7,3,3,9)
I v.v=0
considering the components of f-(v-Wv as the data of our problem.

Theorem 7.3.3.4 Let ve H'(R)? be the solution of the problem (7,3,3,8)

with f given in Lz(1)7. Assume that the equations (7,2,2,1), j= 1, 2, ..., N
have no root (other than -i) on the line Im 1 =-1, and assume that
w;+ tan w; for every j. Then u, belongs to the span of H'(M), the functions
Slim Thi,m and Uni,r n Which correspond to - 1 <Im ! <0 and the functions
Thim and Whim Which correspond to -1 < Im 1', m <0, 1=1, 2.

Proof Knowing that v belongs to H' (9)? implies that ve Lp (9)? for every
p by Sobolev's imbedding theorem. Consequently
(V •V)VEL, (12)3
for every r <2, by Hölder's inequality. We choose r> 1, such that the
equations (7,2,2,1) have no root on the line
2 1
Im ^ = -
r-+-=1
S
and we apply Theorem 7.3.3.1 with p = r. Thus , belongs to the span of
W?(M) and the functions Slims Tim and Uhim corresponding to
344 A MODEL FOURTH-ORDER PROBLEM

and
2
-= SIm A'",m m <0.

Now we observe that the functions in WE(S) and Slims Tim and Ulim
are all bounded functions. Thus ve L..(1)? and consequently
(V • V) VELa (12)3.

We apply again Theorem 7.3.3.1 with p=2 and get the desired
conclusion.

Corollary 7.3.3.5 Let ve H' (R)? be the solution of problem (7,3,3,8)

with f given in Lz(R)? and assume that M is a convex plane polygon. Then
VEHZ (8)3 (and consequently p€ H'(R)).

Now applying the same procedure as before, one can obtain further
results when & is convex.

Theorem 7.3.3.6 Let ve H' (1)? be the solution of the problem (7,3,3,8)
with f given in L, (1)7, 2 <p <0. Assume that M is convex and that the
equations (7,2,2,1), ¡ = 1, 2,..., N have no root on the line

Im ^ = -2 1-+1- = 1
pq
Then , belongs to the span of W3(9), the functions Slim Which correspond
to - (2/9) <Im 1'm <-1 and the functions Thim and Unim which
correspond to -(2/g) <Im 1' jim <-1.

Proof We already know that we H?(8)? and consequently veL..(1)?

and Wu, EL, (1), 1=1,2, by the imbedding theorem. Consequently we
have
(V •V)vEIp (8)}
and we can apply Theorem 7.3.3.1.
8
Miscellaneous

8.1 The Dirichlet problem for a strongly non-linear equation

Let M be a plane domain with a strictly polygonal boundary as in the

previous chapters. We keep the same notation as in Section 4.4. We are
looking for a function u in 2 which is the solution of
- V. p (Vu) Vutu=f (8,1,1)
in M, with the Dirichlet boundary condition i.e. u = 0 on I. In practice the
equation -V • (Vul) Wu = f is more often found. The zero-order term in
equation (8,1,1) has been added just for technical convenience. See
Remark 8.1.8, however. Here f is given in M and + is a positive
nondecreasing real function defined on R+ = 10, ∞L. Therefore we have

q(Vul) =4 (0) > 0

and the equation (8,1,1) is elliptic.
Such a problem has been solved by Caccioppoli (1950-51) when M is a
plane domain with a smooth boundary. This author does not make any
assumption on the rate of growth of & at infinity. In more dimensions, a
similar equation has been solved by Ladyzhenskaia and Uralc'eva (1968)
under the assumption that the growth of • is of polynomial type at
infinity. This assumption allows one to use an optimization method.
Namely one minimizes a functional related to (8,1,1) in a suitable
Sobolev space W(R). The exponent p is given by the rate of growth of 4
at infinity. This method provides a weak solution, while the more classical
method of Caccioppoli leads to strong smooth solutions.
Here we want to allow the domain M to have corners. The correspond-
ing problem has been solved by Najmi (1978) under the assumption that
R is convex. The method is very close to Caccioppoli's and requires some
smoothness for the solution of the linearized problem. This is the reason
why it is assumed that 1 is convex (see Chapters 4 and 5).
345
346 MISCELLANEOUS

For technical purposes it will be convenient to introduce the function

a (x) =4 (Vx), x*0.
The function a is positive and nondecreasing. We shall assume in addition
that a is C2. Thus equation (8,1,1) is equivalent to
-V•a(Vu?)Vutu=f (8,1,2)
and we shall prove the following result (due to Najmi (1978)):

Theorem 8.1.1 Let & be a plane open and convex set with a strictly
polygonal boundary. Let a be any positive non decreasing C? function. Then
for every fe C'."(D), there exists a unique

solution of equation (8,1,2) where 2 <p <2/(2 - 1/w) and w is the measure
of the largest angle of 2.
Let us first outline the method of proof. Basically we want to globally
invert the non-linear mapping
F:ur> -V-a(Vu2) Vut u
between suitably chosen functional spaces. We observe that

F(u) = - (8,1,3)
k.1=1

where ak. (Vu?) =a(Vu?)8k.+Za'(Vu?)DkuD,u (8,1,4)

for k, l= 1, 2, where 8k. is the Kronecker delta. Consequently F is a

well-defined mapping from into Y= Lp(R), (8,1,5)

provided Wu is continuous for every u€ W(2). This is achieved by
assuming p >2 (see Subsection 1.4.5).
A simple criterion for a non-linear mapping to be onto is the following
(see e.g. Ambrosetti and Prodi (1973)):
(a) F is locally invertible (this can be proved by checking that the
Frechet derivative of F is invertible everywhere in X).
(b) F is proper i.e. the inverse image of every compact subset of Y is
relatively compact in X (this is usually obtained by proving an a
priori estimate).
In the particular case of F defined by (8,1,3) it will be easy to prove the
 8.1 A STRONGLY NON-LINEAR EQUATION 347

property (a). Unfortunately we will not be able to prove the full a priori
estimate which implies the property (b). Thus the above principle will
only be a guideline.

Lemma 8.1.2 F is locally invertible from X into Y provided 2<p<

2/(2-T/w).
Proof The Frechet derivative of F at u€ X is the operator

(8,1,6)
k.1=1
where the functions ak, are defined by (8,1,4). We shall apply Theorem
5.2.1.2 to this operator. Let us check the assumptions of this theorem.
(Setting A =-F'(u)).
We have
, ak. (Vul7) 5k51 = a (Vul7) |E|2 + 2a' (Vul)(E • Vu)2= a (0) |813.
k,1=1
Consequently inequality (5,2,1,1) holds with a = a (0).
It is also easy to check that the coefficients ak, (Vul?) belong to W. (2).
Indeed we have
D,a k. (V u|3)

i =1
+Za' (Vul?) DiuD,Dru + 2a' (Vul?) DRUD;Dius.
Each of the terms here is the product of a second derivative of u (i.e. a
function belonging to L (M)) and continuous function of the first deriva-
tives of u (i.e. a bounded function, by the Sobolev imbedding Theorem
1.4.5.2). Consequently we have
D,ak. (Vul)ELp (8), 1552.
Finally inequality (5,2,1,2) is obvious in the particular case under
consideration here (we have a; =0, i Si <2).
In order to apply Theorem 5.2.1.2, we must calculate w, (A). Since
u€ W5(8)M WH(S), Vu is continuous up to the boundary and vanishes at
the corners S,. Accordingly, we have
ak. (V u(S,)17) = ak.‹(0) = a(0)8k.t
In the notation of Subsection 5.2.1, this implies that
A, =- a (0)4
348 MISCELLANEOUS

and I; is just the multiplication by a(0)-112. This implies that

w;(A) =Wis
Theorem 5.2.1.2 shows that F'(u) maps W3(0)n W! (R) onto Ip (8),

2<mar<01-0.
provided

Card {mez. w;
This is achieved when 2/q < m/w; for every j when p <21(2-T/w). I
A first step toward the proof of an a-priori inequality is the follow-
ing simple form of the maximum principle (see for instance Protter and
Weinberger (1967)).
Lemma 8.1.3 Let u€ C2(1)NC (1) be a solution of the equation

-É I a, Dout u =fin 2
an, D, Dutk=1
k.I = 1
where the ak1 and the a are continuous functions such that
Ak,i(x) = ark(x), k,1=1,2
§ ak. (x) 5451 > a |513
k,1=1
for every gel? and xeR with a > O. Then we have
max u|-max lu) + max f. (8,1,7)

Proof Let x€D be a point where u reaches its maximum. There are
two possible cases: either X, € I, or Xo € M. If xo€ I we obviously have
max u -max u. (8,1,8)
On the other hand, if Xo € S, the differential equation at x reduces to

ak. (x) (DEDiu) (x0) + u (x0) =f(x0).

k,1=1
The operator

k.1=1
is nothing but the Laplace operator in different coordinates, thus we have

É ax. (KU) (D, Du) (x) 50.

k.l=1
 8.1 A STRONGLY NON-LINEAR EQUATION 349

Consequently we have
max u = u(x) < max fl. (8,1,9)
Inequalities similar to (8,1,8) (8,1,9) hold for the minimum of u on 8
and this implies (8,1,7).
Let us now go back to equation (8,1,2) i.e.
F(u) = f,
with u € W?(8)n W! (1) and f€ CO. (R). The classical interior regularity
results imply that u € C2.0(1) (see for instance Miranda (1970); this can
also be easily deduced from results in Section 6.3). In particular, we have
u€C°(D)NC'(D).
Consequently Lemma 8.1.3 implies that
max \u|€max (8,1,10)
Next we consider the equations obtained from (8,1,2) by differentiat-
ing. Let v; = D;u, i = 1, 2; then we have

- k.!=

k.1=1 i =1

k.l=
(8,1,11)
Assuming that fe C' (8) implies that
U; EC° (R)NC(8).
Applying again Lemma 8.1.3 yields that
max D;u <max D;u + max Difi. (8,1,12)

We want to find a bound for Wu| in M. The last inequality shows that it
is enough to find a bound for Vul = du/awl on I. A classical tool is the
use of 'barrier functions' (see Oleinik and Radkievitz (1971) for instance).

Lemma 8.1.4 Let u € C'(8)N C'(8) be a solution of the equation

§ ak, DkDiu + u =f in 2
kI
(8,1,13)
fl
350 MISCELLANEOUS

with the boundary condition u =0 on I. Assume that the ak. are continu-
ous in h such that
ak. (x) =01. k(x), k=1,2

k.l= 1
for every tel? and xES, with a > O. Then we have
ди exp d
max as€. max fl, (8,1,14)
where d is the diameter of 12.

Proof Let us consider one of the sides I; of M. We recall that I, is a

linear segment and & is convex. Let I; be defined by the equation r = 0

where r=ax, +bxz

for some real numbers a and b and assume that I lies in r > 0. We define
a barrier function w by
w(x, y) =C(1-e%.
The function w is nonnegative in , provided C > 0. In addition W = 0
on I;. Consequently we have
U-w=0 on Г (8,1,15)

and u-w=0 on I; (8,1,16)

On the other hand we have

k,1=1 k.l=1

In addition, it follows from the assumptions on the ak, that

2 ax.,DarDirza|Vr$=a(a=+b3).
k.l= 1
We can assume that a? + b? = 1; consequently we have

- k.l- 1
where d is the diameter of M. The right-hand side of this last inequality is
 8.1 A STRONGLY NON-LINEAR EQUATION 351

nonpositive provided
• ad- max f . (8,1,17)

It follows that u - w cannot have a strictly positive maximum inside I.

Together with inequality (8,1,15) this implies that

everywhere in &. Then from (8,1,16) it follows that

(u -w)≥O on I;
du;
In the same way we show that

=(-4-w)ZO on F.
du;
just by replacing u by -u in the above considerations. This yields

du aw1dv;=C. (8,1,18)

Summing up inequality (8,1,14) follows from (8,1,17) and (8,1,18). -

The above result can be applied to u solution of F(u) = f, and with the
help of (8,1,10) and (8,1,12) we conclude that

max ul + max Vul = Kamax l/ + max VAS (8,1,19)

provided u€ W3(8)n W! (R), fE CL."(5). Here K depends only on a

and d.
The above inequality will not be enough for our purpose. Actually we
shall need a bound for a Hölder norm of the gradient of u. This will be
achieved with the help of a deep regularity result due to Caccioppoli
(1950-51). This result concerns the smoothness of u inside 2; however,
it will be easy to extend it into a smoothness result up to the boundary,
when & is a convex polygon.
Let us recall Caccioppoli's result. A reasonably simple proof can be
found in Talenti (1966).

Lemma 8.1.5 Let 12 be a bounded open subset of R2. Let

352 MISCELLANEOUS

be the solution of

2 ax.,D.Diut u =f in D
k,1=1
where the aki are bounded measurable functions in I such that
Ak.l=Al,k ae. in D
and such that there exists A1, 12 with O <1, $12 < +00 such that

1,15355
k.1=1
for every gER? and ae. in M. Assume that feLp (R) with p> 2. Then for
every compact subset K in & there exists u € J0, 1[t which depends only on
11, 12 and p, and there exists C such that
Vulla.a,kSCilfllo.p.o+lull.2.o). (8,1,20)
We shall improve this result as follows.

Corollary 8.1.6 Assume that 2 is a convex plane polygon and assume all
the hypotheses of Lemma 8.1.5. Then Wu € CO.# (R) and
Vuln lu, a. 5 SChliflo.p.a + |/ulli.2,n). (8,1,21)
Proof We already know that Wu€ C'# (1). We must investigate the
behaviour of Wu near the boundary.
Let us consider one of the sides I; After rotation and translation we
can assume that I; lies on the axis {* = 0} and that & lies above the axis.
We perform a reflection with respect to x2 = 0 by setting
u (X1, X2), *2=0
X2 CO.
Then U is the solution of the equation

& AK, DAD,U + U=F in w, (8,1,22)

k. = 1
where F is defined from fin the same way as U was defined from u and
where
*2=0 (8,1,23)
Ax.! (x, y) = (-1)4-'ak. (815-12), X, <0.

† MEJO, inf (1,/12, 1-2/p)L.

 8.1 A STRONGLY NON-LINEAR EQUATION 353

Here w denotes the set

W=RUE;US
where

Now the main remark is that the Ak, are still bounded measurable and

that 1,18135 & A4. 15451 =121512

k.1=1
for every feR? and ae. in w. It is also clear that FeL,(w) and that
U€H (w) MH'(w). Consequently we can apply Lemma 8.1.5 to U. This
shows that Wu is Hölder continuous up to the interior of I-
Finally let us consider one of the corners S;. An affine change of
coordinates reduces the general case to the case when S; = 0, I; lies on
the X-axis above zero and I +1 lies on the x-axis on the right of zero.
(Here we have really used the convexity assumption.) Then we perform a
double reflection through {x1 = 0} and {x2 = 0} by setting
X1, *2≥0
-4 (*1-72) *1>0, x2<0
U(X1, X2) = -W(-X1, X2) X1 <0,*2≥0
(41-x10-X2) X, <0,X2<0.
In the same way, we define F from f. Then U is solution of (8,1,22) (there
is no point here in describing the corresponding Ak, in full detail), where
w denotes now the set

Again, we have FELp(w), UE H?(w) MH' (w) and we can define 1 * and
1à such that 0<1, €1% < +00 and such that
171S 2 k.l= 1
for every §€R? and a.e. in w. Applying Lemma 8.1.5 again shows that V u
is Hölder continuous near S;.

Let us again consider u € X, a solution of F(u) = f; thus u is a solution

of (8,1,3). We shall apply Corollary 8.1.6 to u. First we have to find 11
and 12 such that the assumptions hold. We have already shown that we
can set
n, = a(0).
354 MISCELLANEOUS

On the other hand inequality (8,1,19) provides a bound (K fili,,%.) for

IVu. Consequently, we have

This gives a value for 12.

Summing up, we have proved the following result.

Proposition 8.1.7 Let u€ W3(8)M W! (8) be a solution of equation

(8,1,2) where & is a plane open and convex set with a strictly polygonal
boundary. Assume that f€ C'. (R); then there exists u € 10, 1Г (M depends
on fll.o,5) such that Vue C°. " (M), and there exists a constant C (C
depends on fll,.,§) such that (8,1,21) holds.
Now we have all the preliminary material for proving Theorem 8.1.1.
Proof of Theorem 8.1.1 We introduce the subset N of [0, 1] defined by
the following condition (on t€ [0, 11): there exists u, € X which is a
solution of
F(4,) = tf.
We shall prove that N is connected. Since N obviously contains the
value t = 0, it will follow that N also contains the value t = 1, which is the
claim of Theorem 8.1.1 (as far as existence is concerned).
Lemma 8.1.2 implies that N is open. Then we have to show that N is
closed. Let us consider a sequence t; i = 1, 2, ..., of numbers in N which
converges to some limit t. We have to check that t €N. Since t; €N, there
exists a solution u, EX of
F(u,) = t;f.
From Proposition 8.1.7 the sequence U,, j= 1, 2,... is bounded in
C'.M (1) for some u € J0, 1[. By Ascoli's theorem we know that we can
find a subsequence (which we will also denote by u, j= 1, 2,.. for
simplicity) which converges to some limit u in the topology of C'(R). In
particular Vu,, converges uniformly to Vu. Consequently
a(Vu,.?) ->a(Vul?)
uniformly.
Since 4, is a solution of
- V. a (Vu, 3) Vu, +u, = if in 8,
it follows that
- V. a(Vul) Tutu=if in 2
 8.1 A STRONGLY NON-LINEAR EQUATION 355

in the sense of distributions. By uniform convergence we also have

u =0 on T.
On the other hand u € C'. (5) since the sequence u, is bounded in
C'.M(R). Thus the function a(Vul?) is also a Hölder continuous. Conse-
quently Theorem 5.2.1.2 implies that u actually belongs to W3(1)+ since
we have assumed that 2<p <2/(2 - /w). In conclusion we have shown
that tEN (setting U, = u).
Finally the uniqueness of the solution u in X of the equation (8, 1,2) is
easily checked with the help of the usual monotonicity argument. Indeed
let us consider the functional

on X, where '(t) = a(t) for t>0. We have also

where n(t) = 4(t). Thus n is a nondecreasing and convex function.

(n'(6) =2ta(8)=0
In" (1) =2a(1) +417 a' (13) = 2a(0).
This implies that Y is convex on X and that its Frechet derivative is
monotonous. Consequently let u' and u" be two solutions to the problem
(8,1,2); we have

Thus we have u' =u" and this proves the desired uniqueness. •

Remark 8.1.8 As we have already mentioned at the beginning of this

section, one is more likely to find the equation
- V. q(Vu|)Vu=f (8,1,24)
in practical problems. The equation (8,1,1) is obtained by adding a
zero-order term for technical convenience. Actually adding this zero-
order term has been an important simplification only for the proof of
† Here, we take advantage of the uniqueness of the solution w € H'(S) of the equation
- v. a(Vuz)Vw+w=if in 2.
356 MISCELLANEOUS

Lemma 8.1.3. A similar maximum principle for the equation without

zero-order term can be found in Stampacchia (1965) (Remark 4,4, p.
119). This allows one to show that the result of Theorem 8.1.1 is also
valid for equation (8,1,24).

8.2 Some three-dimensional results (an outline)

In Chapters 2 and 3 we proved some smoothness for the solution of

an elliptic boundary value problem in a subset 2 of R" without any
restriction on n. Then, in Chapters 4-7 we assumed that & was a
plane domain. 'There we obtained solutions which split into two parts: one
regular (in the sense of a suitable Sobolev norm) and the other singular
but very explicitly described. Very few similar results are known when 2
is a three-dimensional domain. This subsection is an outline of them. For
the sake of simplicity, we shall restrict our purpose to self-adjoint
problems (in other words we exclude the oblique boundary conditions).
Let us start with a few remarks about the results proved in Chapter 3.
Now we consider M a bounded open subset of 3 whose boundary is a
polyhedron. More precisely we assume that the boundary I' of & is the
union of a finite number of faces I, 1 j&N, each of which is plane. We
assume that & lies on one side of each of the I. We denote by Aik the
edge between I; and Twik being the measure of the corresponding
angle (inside S2). Finally we denote by S the set of all the vertices and by
A the union of all the edges.
Next we split the set {1, 2, ..., N} into two non-overlapping subsets D
and N exactly as we did in Chapter 4. The boundary problem under
consideration is the following. Given f we look for u such that

Au =f in 2
Yiu = 0 on Tis j€D,
(8,2,1)
du= 0
Vi du;
on I; jEN.
If we assume that f belongs to Lp (8), § Sp <0, then the problem has a
unique variational solution u € H'(R). As in the two-dimensional case the
main question is to know what amount of smoothness for u can be
derived from the assumption that f€ W. (S) for some integer m ≥ 0.
Theorem 3.1.1.2 implies the existence of a constant C (which does not
depend on M) such that
2 ID°ullo.2.02€Cllllo.2.0.
lal=2
 8.2 SOME THREE-DIMENSIONAL RESULTS 357

Then if we allow C to depend only on the diameter of 2 and if we

assume for simplicity that D is not empty, we have
(8,2,2)
for u satisfying (8,2,1). Indeed the curvature B is zero on each face I
This is the basic a priori inequality.
Then if N is empty (a Dirichlet problem) or if D is empty (a Neumann
problem) we have a smoothness result when M is convex. Indeed if f
belongs to Lz(8), the corresponding solution u belongs to H3(8).
These are the only results that follow easily from the general state-
ments in Chapter 3. On the other hand, when f is given in W? (S), it
follows from Lemma 2.4.1.4 and Theorem 2.5.1.1 that u € Wm+2(12\V)
for any neighbourhood V of A. We shall now investigate the behaviour of
u near an edge and later a vertex (in this latter case we shall also consider
domains whose boundary has conical points).

8.2.1 Edges
The basic idea (to be stated rigorously later) is that an edge with measure
w leads to a smooth solution iff an angle with the same measure w leads
to a smooth solution (in the two-dimensional problems). Otherwise if ~ is
large enough to produce some singular solution in the two-dimensional
case, then an edge with the same measure will produce infinitely many
singular solutions.
Let us denote by x, y, z the three coordinates and use cylindrical
coordinates around the z-axis (i.e. we use polar coordinates in the x0y
plane: x + iy = re). In addition let us consider a polyhedron & such that
one of its edges is on the z-axis; in other words there exist i and k such
that
Aik = <(0, 0, z) | a <z<b) (8,2, 1,1)
for some real numbers a and b with a <b. We also assume that the face
I; is contained in the xOz plane.
We introduce two other real numbers a' and b' with a <a <b'<b and
a cut-off function depending only on r such that n(r) = 1 for r<p and
n(r) =0 for r > 2p, where p is chosen such that
K={lx, y,z)10<r=2p.0<0<wika'=25b'3C8.
Lemma 8.2.1.1 Define the function u by

u (x, Y, z) =n (r) plat/W, k sin

lTO
- Q (z)
wik
358 MISCELLANEOUS

where a eD(la', b'D. Then we H' (D), Sue W''(R) and ust Wi +2(02) pro-
vided I is an integer such that

max 10.,Wik (m-*4)1<Wik (m -=

2) +2 Wick
p
(8,2,1,2)

The proof of this lemma is very easy. Indeed the support of u is

contained in K and one applies Theorem 1.4.5.3.
This lemma produces infinitely many singular solutions, provided the
condition (8.2.1.2) on I is not empty, since 4 is allowed to span an
infinite-dimensional space.
The above functions u fulfil the Dirichlet boundary condition. Similar
singular solutions for other (self-adjoint) boundary conditions are easily
built: the function

u (X, y, z) = M (r) plat/W, * cOSlaTe @(z)

Wik
fulfils a Neumann boundary condition on I; and I (and vanishes in a
neighbourhood of all the other faces) and the function

q(z)
wink

fulfils a Dirichlet condition on I; and a Neumann condition on Fk (and

vanishes near all the other faces).
In particular a nonconvex edge produces infinitely many variational
solutions u of the Dirichlet problem (or the Neumann problem) such that
Au € La (1), while u$ H3(8).
The above lemma has a negative character; let us now turn to a rather
obvious regularity result.

Theorem 8.2.1.2 Let u€ H'(R) be the solution of the problem (8,2,1)

with f given in C°(R). Then

where V is any neighbourhood of S (the set of all the vertices) provided

(a) m -21p < T/wik When i and kED or when i and KEN.
(b) m -21p <1/2Wik when jED and kEN or when jEN and k € D.

Proof We look at the behaviour of u near A;k. After rotation and

translation we reduce the general case to the particular case when A;k is
on the Oz axis (i.e. (8,2,1,1) holds and we keep the same notation as in
Lemma 8.2.1.1 assuming that I; is in the xOz plane).
 8.2 SOME THREE-DIMENSIONAL RESULTS 359

It is easy to check that u is regular in the z variable. For this purpose,

we replace u by
U(x, y, z) =n(r) Q (z)u (X, Y, z),
with ¢ € D(a', b'D. We already know that u is regular inside O; thus we
have
Av-no"u-2n4, du E C° (5). (8,2,1,3)
dz
Since we start from u € H'(R), we have v € H'(R)
Au E L2(8).
In addition the support of v is contained in K and v fulfils the same
boundary conditions as u.
Let Th be the translation operator in the direction of z, i.e.
ThW (X, Y, Z) = W(x, y, z+ h).
The function (ThU -0)/h is a solution of the same boundary conditions as
u when h is small enough.
In addition, we have
ThU - U
h €H (1).
Integrating by parts the integral

(-1+1) TO-OTAO-" dx dy dz
yields the following inequality

h 541+17 I40-0
11.2,02
.
1-1,2,0

Taking the limit as h SO, we obtain

laz 11.2.0 laz- 1-1+110- 1,2,82

and this shows that do/aze H'(2). Then letting & be any function in
D(a', b'D, we derive that
ди
€H' (8)
for every de D(la', b'D.
360 MISCELLANEOUS

Turning back to (8,2,1,3), we have now

A0єН'(0)
and we apply again the same procedure as above to the function

h U.

By iteration we eventually prove that

for every integer k>1 and every de D(Ja', b'D). Thus nu is infinitely
differentiable function of z (in Ja', b'D with values in H'(G), where we
define G as follows:
G= {(x, y) |0<r <2p, 0<0 <wik}.
In particular nu is an infinitely differentiable function with values in
Lp (G), for every p € ]1, ∞. Now let us denote by A' the Laplace operator
in the variables x and y. We have
A'(nu) = A(nu) - nDZu. (8,2, 1,4)
We already know that u is regular in & \A; thus we have
A(nu)E C°(K)
and consequently A'(nu) is infinitely differentiable with values in Lp (G).
Applying Theorem 4.4.3.7 shows that nu is infinitely differentiable with
values in W(G) provided p is such that 2-2/p <m/w; when j and k
belong to the same set (D or N) and 2-2/p < m/ 2w;k otherwise.
Going back to (8,2,1,4), we see now that A'(nu) is infinitely differenti-
able with values in W(G). Then Theorem 5.1.3.5 shows that nu is
infinitely differentiable with values in W(G) provided p is such that
4-21p <m/wik when i and k belong to the same set (D or N) and
4-21p <1/2w;k otherwise.
Iterating the above procedure yields the desired result.

Remark 8.2.1.3 When the condition (a) (respectively (b)) is violated, the
above proof shows that there exist functions k, € C°(Ja', b'D such that
u (X, y, z)-2+2/p<1<0
- I k(z)E, (X, y)
is an infinitely differentiable function of z with values in W3(G). (We
recall here that a' and b' are any real numbers such that a <a' <b'<b
and consequently k, belongs to C° Ja, bD). In other words k, € C° (A; k).)
 8.2 SOME THREE-DIMENSIONAL RESULTS 361

Here the notation is the following:

1, =
It when i and k both belong to D or N
Wik
1, = (173) I otherwise.
Wik
In addition
G, (x, y) =r-' sin (1,0), when i and k belongs to D
G, (x, y) =5-¼, cos (1,0), when i and k belong to N
G, (x, y) =r" sin (1,0), when jED and KEN
when 1, is not an integer (go back to Definition 5.1.3.4 for the modified
definition of G, when 1, is an integer).

Remark 8.2.1.4 The assumption that fe C() in Theorem 8.2.1.2 is

not necessary. Assuming only that f belongs to Wm-2(82) (with m >2)
leads to the same conclusion; however the proof is much more compli-
cated. It can be found in Kondratiev (1970) when p = 2 and in Grisvard
(1975a) in the general case. Unfortunately we do not know the amount of
regularity of the functions k (cf. the previous remark) that follows from
the assumption that f€ Wm-2 (8).+
Additional results on the behaviour of k, when p = 2 are derived in
Grisvard (1982).

Remark 8.2.1.5 As in Section 5.3, one can extend the result of

Theorem 8.2.1.2 to the case of an operator with variable coefficients in a
domain with a curvilinear edge (a precise definition is left to the taste of
the reader). The basic idea is still that the solution belongs to a given
Sobolev space iff the two dimensional angles with the same measure as
the edge (at any point of the edge) yield regularity (in the corresponding
Sobolev space in two variables of course.

8.2.2 Conical points and vertices

We proceed with our investigation of the behaviour of u, the solution of
problem (8,2,1), by considering u near one of the vertices. For conveni-
ence, we translate this vertex to zero. Thus, in a neighbourhood of 0, S

† The particular case when jeD, k €N and wik = i (mixed boundary condition along a
"flat' edge) has been investigated in Eskin (1973). This author has given an explicit formula
for the functions k, involving the data f. Inspection of this formula easily shows that the
assumption that f belongs to H' (8) implies that k, belongs to H$+13(A;k). Related results
can be found in Nikishkin (1979) and Maz'ya and Plamenevski (1978).
362 MISCELLANEOUS

coincides with a cone C whose intersection with the unit sphere S? is

denoted by G. Thus G is an open subset of the unit sphere whose
boundary is the union of a finite number of arcs of great circles.
Here, in order to include cones which have a regular basis, we shall
make a more general assumption on G. We shall only assume that AG is a
curvilinear polygon (a definition similar to Definition 1.4.5.1 can be made
here; roughly speaking dG is the union of a finite number of curves which
cut at angles and G lies on one side of each of these curves only).
As it is natural we introduce the spherical coordinates (r, A, $). For the
sake of definiteness & denotes the angle between OM (M the point
whose coordinates are x, y and z) and the z-axis; o denotes the angle
between Om (m the projection of M on the xOy plane) and the x-axis.
The basic idea here is that such a cone can produce two kinds of
singular solution:
(a) Solutions of the form
(8,2,2,1)
where I is an eigenfunction of the Laplace-Beltrami operator 4' in G
(with the suitable boundary conditions) and ^ is related to the eigenvalue.
These are singular functions similar to the singular functions G of the
two-dimensional case (cf. Chapter 4). Such singular functions will arise
even when the cone C has a regular basis (i.e. G is smooth). The amount
of singularity is related to 1 and only a finite number of such functions
will be generated outside a given Sobolev space.
(b) Solutions of the form
u (x, y, z) =4(8) G(0, 4) (8,2,2,2)
where I is a regular function of r, at least away from zero, while G is a
singular solution of 4' on G (again fulfilling suitable boundary conditions
which will be made precise later). These are singular functions similar to
those produced by an edge (cf. Remark 8.2.1.3). Such singular functions
arise only when G has corners (or there is a mixed boundary condition).
In addition, & spans an infinite-dimensional space and accordingly there
are infinitely many such singular solutions.
In order to make the above outline more precise we shall first state a
result which shows how the singular solutions (8,2,2,1) arise. For this
purpose it is more convenient to consider a domain 1 C3 having only
conical points corresponding to a regular basis GeS?. Exactly as in
Chapter 4 we shall start from an a priori inequality in H7(8), which has
been derived in Hanna and Smith (1967).
We consider here a bounded open subset & of l3 such that O belongs
to its boundary I. We assume that I'\fO} is of class C2, we assume in
 8.2 SOME THREE-DIMENSIONAL RESULTS 363

Figure 8.1

addition that there exists a neighbourhood V of 0 such that, in V, S

coincides with the infinite cone C whose intersection with the unit sphere
S? is a subset G of S? whose boundary aG is of class C2
Given fe L2(8) we look for a solution u € H2(1) of
Au =f in 2 (8,2,2,3)
with either a Dirichlet boundary condition
vu = 0 on Г (8,2,2,4)
or a Neumann boundary condition
du= 0 (8,2,2,5)
Y on Г1<O}.
dy
We denote by A' the Laplace-Beltrami operator on G with the
corresponding boundary condition, i.e. its domain is either H?(G) n
H'(G) for a Dirichlet problem, or

429 = 0 on 268
for a Neumann problem. In both cases this is a self-adjoint operator in
H= L(G) whose spectrum is an infinite sequence of real numbers
364 MISCELLANEOUS

-1, 1 = 1, 2, ... where 1, ≥0, with no limit point. We denote by 44, 1=

1, 2, . .. the orthonormalized sequence of the related eigenfunctions. Thus
we have
-4'44 = 1,04 in G (8,2,2,6)
where de H (G) and either 4, € H&(G) for a Dirichlet problem or
y dad/av = 0 on AG for a Neuman problem.
The basic a priori inequality for the equation (8,2,2,3) with one of the
boundary conditions (8,2,2,4) or (8,2,2,5), does not follow from (8,2,2)
since M is not a polyhedron. We shall prove here the following statement.

Theorem 8.2.2.1 Assume that At 3 for every I; then there exists a

constant C such that
lul2.20sClAullo.2.n+|lullo.2.nl (8,2,2,7)
for every u € H? (1) which fulfils either the boundary condition (8,2,2,4) or
the boundary condition (8,2,2,5).
(Of course one can drop the term lullo. in the case of a Dirichlet
problem.)
It is clearly seen with the help of a partition of unity that inequality
(8,2,2,7) follows from (3,1,1) and a similar inequality in C for functions
with bounded support. (In addition, (8,2,2,7) is just (3,1,1) when the cone
C is convex.)
Next the proof of the inequality in C relies on the use of weighted
spaces similar to those introduced in Subsection 4.3.2. Indeed we denote
by PI(C) the space of all the functions u defined in C such that
plat-mD°u€Iz(C)
for all al m.† It is obvious that a function u € P(C) which has bounded
support also belongs to H' (C). The converse statement is true up to the
addition of a finite-dimensional space. This will be stated in a precise
fashion below; however, we must observe at once that a similar statement
for the two-dimensional case does not hold (see Kondratiev (1967),
Theorem 4.3.2.2, which excludes the case when p =2 and the weaker
statement in Theorem 7.2.1.1 when p = 2).
Theorem 8.2.2.2 Let u € H? (C); then u € P'(C) iff u (0) = 0.

The condition u (0) = 0 is meaningful since every u € H? (C) is continu-

ous in C (see inclusion (1,4,4,6)). Conversely every u € P2(C) is locally in
H? and therefore also continuous in C for the same reason.
• P™(C) is equipped with the obvious norm (see Subsection 4.3.2).
 8.2 SOME THREE-DIMENSIONAL RESULTS 365

Next the condition u(0) = 0 is necessary for u to belong to P¿(C), since

this requires u/r? to be square integrable near zero and u is continuous
there. We just have to prove that the above condition is sufficient. This
will be done in two steps.

Lemma 8.2.2.3 The inequality

(8,2,2,8)
1x1?
holds for every u € C'(C).

Proof For u € C'(C), we have

du (tor) de

for every o € G.
It follows that
u(ro)5-
r r 1
I. INu(co) de
Integrating, we obtain (8,2,2,8) by applying the second Hardy inequality
(see Subsection 1.4.4) with a = 1.
Lemma 8.2.2.4 The inequality

1x14
Fart's (8,2,2,9)
C
holds for every u€ C'(C) such that u (0) = 0.
Proof Here we have

u(ro) =Idt
5k=1on du
ate(to) de
for every a € G.
It follows that
u(ro)
p?
Integrating, we obtain (8,2,2,9) by applying the first Hardy inequality (see
Subsection 1.4.4) with a = 0. -
366 MISCELLANEOUS

Proof of Theorem 8.2.2.2 We consider u € C'(C) a dense subspace in

H?(C) (by Theorem 1.4.2.1). We fix n € D(C) a cut-off function, such that
n(0) =0, and we apply the previous lemmas to
и-u(0)n
and its first derivatives respectively. We obtain
[u(x) -u(0)n(x)12
x'A C Ix/?
C ii=1

for some constant K. By density (and Sobolev's imbedding theorem) the

same inequality holds for every u € H?(C). The conclusion follows when
u (0) =0. 1
Now a first step toward the proof of Theorem 8.2.2.1 is the following
preliminary result.

Lemma 8.2.2.5 There exists a constant K such that

Tulip3c-K|Aullo.2.c (8,2,2,10)
for every u eP¿(C) such that either yu =0 on aC or y dulav =0 on 2C,
provided A, # ⅔ for every I.

Proof In spherical coordinates, the equation

Au =f
means
22и 2 an
r Ar 4'u =f.
As in Subsection 4.3.2, we perform the change of variable r=e', setting
u(t.o) =ule'o)
g(t, o) =e21f(e'0)
for every teR and o € G. We obtain the equation
220 20+4'0 = 8
dt (8,2,2,11)
in the infinite cylinder B = RX G.
If we assume that yu = 0 on dC, then we have yo = 0 on AB; otherwise
 8.2 SOME THREE-DIMENSIONAL RESULTS 367

we assume that y du/av = 0 on dC and we have y do/av = 0 on AB. Finally

the assumption that u belongs to P(C) implies that
é12D° € L2(B) = Lz(R; H) (8,2,2,12)
for every a =2.
Expanding both sides of equation (8,2,2,11) on the eigenfunctions 4
one obtains
d'(t) +01(8) -110, (0) = 81 (1), (8,2,2,13)
where
0; (1) = 064,0741(6) do

84(1) = 8(1,0)141(6) do.

G
In addition the condition (8,2,2,12) reads as follows:

1=1
The function g, is given such that e2g, € La(R); consequently the
equation (8,2,2,13) has a unique solution u such that

e'v'(t)12+w(t)12]dt<+00
('*.
iff > is not a root of the characteristic equation (of the differential
equation)
a?+a-1, =0.
This requirement means 1, + 3.
In addition it is easy to check that there exists a constant K such that
° +00

when I-> +0. Summing up, we have

§1-1*o 1=1 é'|gi(t)|? dt

1-00

for some K' and this implies

2 lle-U2D°ollo.u.B < K"llegllo.o,B
la|=2
by the Bessel identity.
368 MISCELLANEOUS

Performing the inverse change of variable = In r, one obtains the

desired inequality (8,2,2,10).

Proof of Theorem 8.2.2.1 Let us assume that u€ H7(2) and fulfils a

Dirichlet boundary condition, then we have u(0) = 0 and consequently
nue P2(C)
by Theorem 8.2.2.2, where n is a cut-off function such that ME D(R),
n (x) = 1 near zero and the support of n is contained in V (the neighbour-
hood of zero in which & coincides with C). Therefore the inequality
(8,2,2,7) follows from (8,2,2,10) and from (3,1,1).
When u€ H?(M) and fulfils a Neumann boundary condition the same
procedure leads to the inequality (8,2,2,7) only for those functions
u€ H?(8) such that
ди
Y =0 on T'10!
du
and such that in addition
u (0) =0.
This last condition defines a subspace of codimension one in the space

"-0!
in which we want to prove the inequality.
The same technique as in the proof of Theorem 4.3.2.4 allows one to
derive the weaker inequality
ul22n-ClAullo.2.o+|/ullo.2.at
This completes the proof. S

From the inequality (8,2,2,7), we shall derive a Fredholm alternative

quite similar to the one in Subsection 4.4.1. Let us derive it briefly now.
We apply Lemma 4.4.1.1 to the operator A = 4 considered as an
operator from
du
(respectively y =0 on F107)!

into Ez= L2(8). Thus A has a finite-dimensional kernel and a closed

range.
The kernel of A is obviously {0} in the case of a Dirichlet problem and
the space of the constant functions in the case of a Neumann problem.
 8.2 SOME THREE-DIMENSIONAL RESULTS 369

The closed range property is far more important. Let us denote by N the
orthogonal of the range of A in L2(8). As in Subsection 4.4.1, it is easily
seen that every eN is harmonic in M and such that y =0 on
F\0} (respectively do/av = 0 on F\{0}). Since I'\10} is of class C2 the
results in Chapter 2 imply that

for every p <∞ and every neighbourhood W of 0.

Then we look at the behaviour of v in V. In spherical coordinates, we
have
220 220
+ 4'0=0.
Expanding o on the orthonormal system 4, I = 1, 2, . .., we obtain

for r small enough (say <ro), where

(8,2,2,14)

In addition, we have
2
0°(r) +=01(r) -
1, O<r<ro,
r
since v is harmonic. Consequently
v. (r) =a,p°, +b,pB,
where a, and b, are arbitrary real numbers and
-1+1(1+41,)
04, =
2 B, = 2
-1-V(1 + 41,)

The condition (8,2,2,14) readily implies that b, = 0 unless 1, <4. Thus

we have

1, <3/4 1=1
It is clearly checked that the first sum (which is finite) does not belong to
H'(OMir <ror) while each term of the second sum belongs to this space.
More precisely, applying the method of Proposition 4.4.2.2, one shows
that this series actually converges in H'(On{r <rl) for every ró € 10, rol.
Finally this implies that the dimension of N is the number u of the
370 MISCELLANEOUS

eigenvalues 1, which are less than & in the case of a Dirichlet problem and
u +1 in the case of a Neumann problem (see the proof of Theorem
4.4.3.3).
Summing up, we have sketched the proof of the following result of
Kondratiev (1967a).

Theorem 8.2.2.6 Assume that M, + for every I and denote by u the

number of eigenvalues 1, such that
1,33
then the space A(H? (1) MH' (8)) (respectively A{u eH? (1); y aulav =0
on F\{0}} has codimension u (respectively u + 1) in L, (12).

Consequently we must add u singular functions to the space H7(8) in

order to describe all the solutions of the equation (8,2,2,3) with f given in
L2(8) under the boundary condition (8,2,2,4) (respectively (8,2,2,5)).
Applying the technique of Theorem 4.4.3.7, we set
si(ro) =p° (0)n(ro)
for 1, <4. These functions are u functions belonging to
H?(0)\H'(0)
and such that
As € I?(1)
and they fulfil the boundary condition (8,2,2,4) (respectively (8,2,2,5)).
Accordingly, we have the following statement.

Corollary 8.2.2.7 Assume that 1,7à for every I; then for every f€ L, (12)
there exist constants c and a function u such that

1, <3/4
u is solution of the equation (8,2,2,3) and fulfils the boundary condition
(8,2,2,4) (respectively (8,2,2,5) provided In fdx = 0); u is unique (respec-
tively unique up to the addition of a constant).
This shows that a conical point produces singular solutions of the form
(8,2,2, 1).
Let us now discuss the assumption A, + a, which has been useful in
deriving the a priori inequality (8,2,2,7). It is easy to check that when G is
contained in a hemisphere then the first eigenvalue 1, corresponding to
the Dirichlet problem is greater than or equal to 1 (cf. for instance
Grisvard (1975b)). Accordingly we have shown that u belongs to H? (8),
 8.2 SOME THREE-DIMENSIONAL RESULTS 371

when the cone C is convex; this was proved in Chapter 3. On the other
hand let us consider the particular case when G is a circular cone of angle
B. It is shown in Hanna and Smith (1967) that when B increases from
to 2m, the first eigenvalue 1 decreases from 1 to 0. Consequently there is
one value (B = 1.45) for which the above method of proof is inconclu-
sive (actually, the a priori inequality (8,2,2,7) does not hold either).
Curiously the inequality (8,2,2,7) always holds for polyhedral cones, i.e.
cones for which aG is the union of a finite number of arcs of great circles
(cf. inequality (8,2,2)). When the domain of A' is contained in H?(G),
the method of proof of Theorem 8.2.2.6 still works. This shows again the
regularity of u in H7(8) when M is a convex polyhedron.
So far, we have shown precisely how the singular functions of the form
(8,2,2,1) arise. When & is a polyhedron several edges meet at the same
vertex. Each edge is likely to produce the kind of singular solutions that
we have described in Subsection 8.2.1: the functions in Lemma 8.2.1.1
are of the form (8,2,2,2) near each vertex. Unfortunately the precise
behaviour of the function & in (8,2,2,2) as r->0 is not yet known.
To conclude this subsection, we give a statement which summarizes the
results in Theorem 8.2.1.2 when m =2 and an extension to p*2 of
Corollary 8.2.2.7 when M is a polyhedron. The basic assumption is that
the angle of the edges are small enough not to produce edge singularities.
The proof is very technical and may be found in Grisvard (1975a).
We need some auxiliary notation. We denote by S; the vertices of &2,
1 sisI, and by G; the intersection of the unit sphere centred at S; with
the cone C; corresponding to S; We denote by Mi, 1 = 1, 2, ... the
sequence of the eigenvalues of -A' in G; with the corresponding bound-
ary conditions.

Theorem 8.2.2.8 We assume that

(a) 2-21p < IT/wik when i and kED or when j and kEN.
(b) 2-21 p < 1/2w; k when j€ D and KEN or when jEN and kED.
We assume in addition that 1; + (2-3/p)(3-3/p) for every i and every I.
Then the space

has codimension (we assume that D is not empty for simplicity)

14 = 2 card fe 1. 5(2-3) (3-5)3

in L. (8).
372 MISCELLANEOUS

Consequently the solution u of the problem (8,2,1) with f given in

Lp (S) is such that

u-S i =1 1, < (2-3/p) (3-3/p)

where Ci are real constants and

Si, (r,O;) = r; 2
-1+411+41,1) 441(03) no (5,01)
in an obvious notation; namely: r; denotes the distance to Si, O; a point of
Gi, Mi, is the normalized eigenfunction, corresponding to - Mi,, of 4' with
suitable boundary conditions; finally n is a cut-off function depending
only on r; such that Ni € D(R), Ni = 1 near S; and Mi = 0 outside some
neighbourhood of S;

Remark 8.2.2.9 An asymptotic expansion near the vertices of the

singular part of the solution corresponding to an edge is derived in
Grisvard (1982) in the particular case p = 2 when the assumption (a) in
Theorem 8.2.2.8 is not fulfilled.

8.3 The heat equation

It is well known that one can derive several properties of the heat
equation by applying semigroup theory, provided one has a good know-
ledge of the properties of the Laplace operator (and its resolvent
operator). This method has been applied successfully for solving the heat
equation with various mixed boundary conditions in regular cylinders. We
mean here a cylinder Q of the form
Q=10, TI×D
where & is a domain with a smooth boundary in R" and 10, TI is an
interval in time. Possible references are Lions (1956), Lions and Magenes
(1960-63), Krein (1967). One can apply the same kind of method when
M is a plane polygon and consequently & is a cylinder with edges.
However, we shall consider here a different problem. We shall solve a
mixed boundary value problem for the heat equation in a domain Q
which may not be the Cartesian product of an interval in time by a
domain in space. Here we follow work by Sadallah (1976, 1977, 1978).
For simplicity, we consider a problem with only one space variable. To be
precise, we assume that
Q=<(t, *) |0<1<T, 41(1) <x <42 (1)),
 8.3 THE HEAT EQUATION 373

X=92(5)

(t!
X=Ф,

Figure 8.2

where T is a finite positive number, while 41 and 42 are continuous

real-valued functions defined in [O, T] Lipschitz continuous in JO, TL, such
that
41(t) <42(t)
for t € J0, TL. Given f€ L2(Q) we look for a solution u (as regular as
possible) of
Diu -D<u=f in O
u (0, x) =0, 41(0) <* <42(0) (8,3,1)
ult, 4: (1)) =4 (1, 42(1)) =0, O<I<T.
We emphasize that we shall allow 4, to coincide with 42 for t=0 and
for t = T. Actually domains of the same kind under a weaker assumption
on 41 and ¢ are considered in the works by Baderko (1973, 1975, 1976).
This author assumes that 1 and $2 are only Hölder continuous (with
exponent larger than or equal to =) and solves the heat equation in Q by
the potential method introduced by Gevrey (1913). However, in order to
apply this method, one must assume that 41(0) < 42(0) and that 41(T) <
Ф2(T).
Actually the method of Sadallah, which we shall outline here, is a
straightforward extension of Chapter 3. We first prove an a priori esti-
mate when Q is nice (in a sense to be defined later), and then we take
limits in Q in order to reach the kind of domains described above. The a
priori inequality is proved simply by integration by parts, and again we
have very accurate control of the constants involved, as functions of ;
this is why we are able to take limits in Q.
374 MISCELLANEOUS

The result we are going to prove here is the following, for which we
need the technical assumption
%: (1) [42(1) -41(8)]->0 as +->0, i = 1,2 (8,3,2)
if Q1(0) = 42(0) and
%: (1) (42(8) -41(8)]->0 as t-T, i = 1,2 (8,3,3)
if 41(T) = 42(T).

Theorem 8.3.1 Assume that (8,3,2) and (8,3,3) hold and that f is given in
La(Q); then there exists a unique function u which is a solution of
Diu -D<u =f in Q (8,3,4)
such that u, D,u, Du and D<u belong to L2(Q) and
0<+<T, i=1, 2
(8,3,5)
(ya (0, x) =0, Ф1(0) <X <42(0).

We observe that u belongs to H'(Q) and consequently the trace yu is

well defined on 2Q away from the points (0, 4; (0)) and (T, Q; (T)), i = 1, 2.
Indeed these are the only possible points in a neighbourhood of which the
boundary 2Q may not be Lipschitz.
We emphasize that this is an existence result for a strong solution. The
existence and uniqueness of a weak solution with, say,
u and DuEL, (Q),
are easy to derive (see Oleinik and Radkievitz (1971) for instance). Thus
Theorem 8.3.1 is mainly a smoothness result.
Let us now carry out some preliminaries. For the time being we
consider the simpler case which follows. We replace Q by
Qa = {(I, X) | a <+<T-a, 41 (1) <x <42(1)}
with a >0. Thus we have
(41(a) <42(a)

and 41 and 42 are uniformly Lipschitz continuous in [a, t - a]. In this case
it is very easy to prove the result corresponding to Theorem 8.3.1.
Indeed, we can easily find a change of variable & mapping a onto the
rectangle
Ra = Ja, T-a[X]0,11,
 8.3 THE HEAT EQUATION 375

which leaves the t variable unchanged. We define & as follows:

X - Q1(1)
42(8) -4,15))
Then we define the functions v and g in Ra by
0=40151 and
The equation
Diu -D<u =f
in a is equivalent to the following:
Did + a (t, x)D+U-b(1) D70=8 (8,3,6)
in Ra where a and b are defined by

all. I) = D. (645-914040)

b (t)
1
The mapping & is bi-Lipschitz and therefore it preserves the space H'
In other words u belongs to H'(Qa) iff we H'(Ra). The boundary
conditions on v which correspond to the boundary conditions on u are
the following:
(ru(a, x) =0, 0<x<1
I yo (t, 0) = 0(4, 1) =0, a<t<T-a. (8,3,7)

In a first step we consider the simplified equation

D,0 -b(t)D=v =g in Ra (8,3,8)
with the same boundary conditions on U.

Lemma 8.3.2 For every g € La (Ra) there exists a unique ve H'(Ra), with
DIvELz(Ra), which is a solution of (8,3,8) and (8,3,7).

Proof A simple change of variable (in t) reduces equation (8,3,6) to the

heat equation and we can apply some classical results.
Indeed we define the function B as follows:

B(= b(5) ds.

This is an invertible C'. mapping from [a, T-a] onto [0, B(T-a)]. It
376 MISCELLANEOUS

follows from Lions and Magenes (1968), for instance, that there exists a
solution w€H'(R) of
D,w-D<w=g/b
in R' = J0, B(T- o)[× J0, 1[, with
(yw(O, x) =0, 0<x<1
(yw(t, 0) =yw(4, 1) =0, 0<+<B(T-a)
and such that in addition D{w € Lz(R&).
We obtain the desired function v by setting
u(t, x) =w(B (t),x). •
Lemma 8.3.3 For every g € La (Ra) there exists a unique solution U €
H'(Ra), with DIve La(Ra), of (8,3,6) and (8,3,7).

Proof We denote by A the operator D, -b(t)D} defined from

V=fUEH' (Ra) | DEV EL, (Ra), u fulfils (8,3,7))
into L2(R). We have shown in Lemma 8.3.2 that A is an isomorphism.
Then it is known (cf. for instance Besov [1969]) that Dx is a compact
operator from V into La(R).t Since a is a bounded function, the
operator aD is also compact from V into L2(Ra). Consequently A + aD
is a Fredholm operator (with index zero) from A into La (Ra). Thus the
invertibility of A + aDx will follow from its injectivity.
Accordingly let us consider v € V, a solution of
Diu+aD.0-bD<v=0,

in Ra. We perform the inverse change of variable of 4. Thus we set

U=vow.
It turns out that u € H'(Qa), Due La (Qa) and
Diu -D<u = 0 in Qa
In addition u fulfils the homogeneous boundary condition (8,3,5). As
† Actually since Ra is a rectangle, it can easily be shown that there exists a continuous
extension operator from H12(R) into H1.2(13); here for a general plane domain 2,
H1.2(0) is defined by

H'2(0) = fu | u, D,u, Dzu, DzuEL, (1)}.

Then the compactness of u-»qDyu from H12(R3) into L (3) is easily checked by Fourier
transform, provided 4 € 9 (3).
 8.3 THE HEAT EQUATION 377

usual we calculate the integral:

This yields

yDaurav,.dst|D.uf-didx=0.
All the boundary integrals vanish but

Ivu(T-a, x) dx,
10, (7-a)
which is nonnegative. This yields the inequality

Qa

which implies that u vanishes; this is the desired injectivity. W

So far we have proved the desired result in the better domains Qa. Now
we shall prove an a priori estimate which will allow us to take limits in a.

Lemma 8.3.4 Let ]a, b[ be a nite real interval. There exists a constant C
(independent of a and b) such that

for every u € H'(la, bD such that Saw(x) dx = 0.

The proof of this inequality may be found in Necas (1967) for instance
(it is elementary: actually the general case follows from the particular case
Ja, bl = 10, 11 by an affine change of variable).
We shall apply this inequality later to the function Du, where u fulfils
the assumptions of Theorem 8.3.1. This yields the inequality

(D}u(t, x)17 dx.(8,3,9)

Ф) (t)
Since we have
(Ф, (1)
Du (t, x) d* = yu (t, 42(1)) - yu (t, 41 (1)) = 0.

Lemma 8.3.5 We assume that , and 42 ful l the conditions (8,3,2) and
fi
fi
 378 MISCELLANEOUS

(8,3,3). We assume in addition that u € C(@a) and

(u (a, x) =0, 41(a) <*=42(01)
Tult, c, (1)) =4 (4, 42(4)) =0, astST-a.
There exists a constant K which does not depend on a and u such that
[lull+||D.u|? +||D.u|?+||D<u|REK||D,u- D<ul? (8,3,10)
in the norm of La(a).

Proof There are two main steps. First we derive a bound for Du by
calculating

( (Du - Diu)u dr dx.

la
Second we derive a bound for Du and Du by calculating
(Du -D<u)? dt dx
as we did in Chapter 3.
Setting f = Du - Dzu, we have

= Dul- didx.
Oa
It follows that

in the norm of L2(a).

On the other hand, Poincaré's inequality implies that

where L = maxicO,TI [42(t) - 41(t)]. It follows that

(8,3,11)
ID.uSLI .
Then we have

Pardx=|IDuPdrdx+ID,uf-dids
JOd
(8,3,12)
Jada
fl
 8.3 THE HEAT EQUATION 379

We shall rewrite the boundary integral making use of the boundary

conditions (8,3,5).
On the part of the boundary where t = a, we have v2=0, u =0 and
consequently Du = 0. The corresponding boundary integral vanishes.
Then on the part of the boundary where t = T- a, we have again v2 = 0
and v1 = 1. Accordingly the corresponding boundary integral is nonnega-
tive and we can forget it.
On the part of the boundary where * = Q; (t), i = 1, 2, we have
ult, Qi (t)) =0.
Differentiating with respect to t we obtain
Diu =-q'(t) Dxu.
Consequently the corresponding boundary integral is

g==T-
" Duce a son eric de + S. Died, a cast a sin(8,3,13)
de
We convert this boundary integral into a surface integral by setting
42(t) -X
'X = p. (t)

=- (9, 18) a 1 42(1) -X

=- (Ф2(t)b,(t) -x

+ for
1
and consequently

Dxult, c. (t)) 12=2

t
Sp, (1)
• Ф, (t) 1
A similar inequality holds for Du(t, 42(t)) and this yields

Our

10. 42(1) -41(1)

380 MISCELLANEOUS

It follows that

With inequality (8,3,9) this yields

where

92(1) - 41 (1) S
and
Ne=red.1.u.b.
iu.b: e.T. 12114 1 (1) +14 5(0)137483 (6) - 41617

+44(1) +42(1) 1142(1) -41(4)13

for every & > 0.
Going back to (8,3,12) we have

with (8,3,11), this yields

Finally we take advantage of the assumptions (8,3,2) and (8,3,3). Thus

choosing & small enough, we have N. ≤‡ and consequently

Summing up, we have

provided N. ≤4.
We shall need an extension of inequality (8,3,10) to functions with less
regularity. This requires a density lemma:

Lemma 8.3.6 Every u € H (Qo) such that

(DIuEL2(Qa)

(yu (a, x) = 0, 41(a) <*<42(01)

can be approximated by a sequence Un n = 1, 2, ... of functions belonging
 8.3 THE HEAT EQUATION 381

to C2(O) such that

(Un. (t. 41 (1)) =Un (4 42(1)) = 0, a<<T-a
Tun (as, x) =0, 41(a) <* <42(01).
The convergence is such that

as n=> +2.

Proof One easily replaces a by Ra with the help of the change of

variable a defined above. Then the proof in the case of Ra is just an
exercise. •
This implies clearly that the inequality (8,3,10) holds for every u€
H'(Q.) which fulfils the assumptions in Lemma 8.3.6. We are now able
to prove Theorem 8.3.1.

Proof of Theorem 8.3.1 By Lemma 8.3.3, there exists for each a >0
(small enough), a unique

such that Dila € Lz(a) and

(Dux -DIUa=f in la
Ma (a, x) = 0, 41 (a) <x<42(0)
(Walt, 41 (1)) = Ma (4, 42(1)) = 0, a<1<T -a.
In addition the inequality (8,3,10) holds for u. and accordingly we have
wali.2.o.+|/Dsuello.2.0.<K|Al6.2.0.
We consider a sequence an ->0, as n-> +0. The related sequences of
functions
U.., DiMan Dylan and DElays
n =1,2,..., are bounded in L¿(0). By replacing an by a suitable
subsequence (that we denote again by On n = 1, 2, ... for simplicity).
there exist functions
u, Up, Us and w
in L2(O) such that

DeMon 702 Dik.,-> W

weakly in L2(Q) as n=> +00.
382 MISCELLANEOUS

The remainder of the proof is similar to the proof of Theorem 3.2.1.2.

We have clearly
U, = Diu, U2= Dxu and w = D<u
in the sense of distributions in Q and consequently we have
Diu -D<u =f in Q.
Finally let 4 € D(R) be such that
(Ф(t)=1 for t<T-s
(4 (1) =0 for t≥ T-El2.
with € >0. We have clearly
фй, єН' (0)
and Qi. remains bounded in I'(Q). Thus, taking the limit, we have
quE H'(Q).
This implies that
(yu) (O,x) =0, 41(0) <X <42(0)
(yU) (1, 41 (1)) = (yU) (1, 42(1)) = 0, 0<1<T -E.
Since the above boundary conditions hold for every & >0 we have
proved the existence of a function u having the properties listed in
Theorem 8.3.1. As we have already observed the uniqueness of u is
classical. _
Remark 8.3.7 Inspection of the identity (8,3,12) shows that the condi-
tion (8,3,2) is useless when 42 is nondecreasing and 4, is nonincreasing
near zero. In the same way, the condition (8,3,3) is useless when 42 is
nondecreasing and ¢1 is nonincreasing near T.

Remark 8.3.8 The works by Sadallah mentioned above include similar

results for the heat equation in more space variables and for the equation
Diu + (-1)MD?Mu =f,
m an arbitrary positive integer (such an equation is also studied in the
works by Baderko mentioned above).
The domains considered in Theorem 8.3.1 include all the convex
polygons but not all the polygons. However, it is very easy to derive a
similar result for any polygon.

Corollary 8.3.9 Let Q be a plane open domain with a polygonal bound-

 8.3 THE HEAT EQUATION 383

s. k

aQ\r'

Figure 8.3

ary. Then the operator D, - D' is a one-to-one operator from

V ={u€H'(O); DueL2(Q), vu =0 on r'}
into L¿(Q), where I' is the part of the boundary aQ where v, <1. In
addition the image of D, -D≥ is closed and has finite codimension M,
the number of corners S; = (t; X;), with the following property: There exists
a neighbourhood U; of S; such that
t>t
for every point (t, x) € U,ndQ.

The proof consists in applying Theorem 8.3.1 in each polygon Qi of a

covering of Q such that
(a) Qi is a convex open subset of O with a polygonal boundary
(b) 0.00. = O for i‡ j

The details can be found in Sadallah (1976).

Remark 8.3.10 Similar results for the operator D, + (-1)' Dam are de-
rived in Sadallah (1983).
384 MISCELLANEOUS

8.4 The numerical solution of elliptic problems with

singularities
In this section, we take for granted that the reader is familiar with the
finite element method for solving elliptic boundary problems in domains
with smooth boundaries (cf. for instance Carlet (1978)). The analysis of
the finite element method usually relies on the assumption that the
solution of the given problem is regular enough. However, the implemen-
tation of this method is very often done on problems in polygonal
domains which prevent the solution from being smooth everywhere.
As we saw in previous chapters, the presence of corners leads to
singular behaviour of the solution only near the corners. This singular be-
haviour occurs even when the data of the problem are very smooth. It
strongly affects the accuracy of the finite element method throughout the
whole domain. We shall outline here the two main procedures which have
been proposed to overcome this difficulty. The first is based on mesh
refinements and has been analysed by several authors; see Babuska and
Kellogg (1972), Babuska et al. (1979), Raugel (1978), Schatz and
Wahlbin (1978-79), Thatcher (1975) for instance. This method may be
applied to most of the practical problems since it requires only a qualita-
tive knowledge of the behaviour of the solution near the corners (see
details in Subsection 8.4.1). The second consists in augmenting the space
of trial functions in which one looks for the approximate solution. This is
done by adding some of the singular solutions of the problem to the usual
spaces of piecewise polynomial functions (cf., for instance, Fix et al.
(1976), Babuska and Kellogg (1972), Lelievre (1976b), Diaoua (1977)
and Ladeveze and Peyret (1974)). This procedure requires a very accurate
knowledge of the singular solutions and consequently it can be applied
only to special problems (see details in Subsection 8.4.2).
Since the purpose of this section is only to illustrate the procedures
mentioned above, we shall consider only the simplest model problem;
namely we shall consider the Dirichlet problem for the Laplace equation
in a plane polygon with only one nonconvex corner. We shall approxi-
mate its solution by means of a Galerkin method using trial functions
which are piecewise first-order polynomials for simplicity.
Some slightly different approaches to the singularity problems, using
integral equations, may be found in Wendland et al. (1979).
8.4.1 Weighted spaces and mesh refinements
Let us again fix some notation which we keep consistent with that of
Chapter 4 (see Section 4.1). Accordingly & is a plane domain with a
polygonal boundary I, the union of a finite number N of linear segments
; numbered according to the positive orientation. We denote by w; the
 8.4 PROBLEMS WITH SINGULARITIES 385

my

I'з

W=WN

0=5 N

Figure 8.4

angle between I; and I+1 and we assume that c; <m for every i but
¡ = N. For simplicity we assume that S~, the corner point between IN and
Th, has been translated to the origin. In addition we assume that I is
included in the positive abscissa axis (Ox) while In is supported by the
half line whose angle with Ox is ww (counted counterclockwise). For
further simplicity we set w = wn.
Given f€ La(8) we look for a solution u€ I' (8) of
-Au =f in l. (8,4,1,1)
We know (Chapter 4) that there exists a unique solution u and in addition
there exists a unique number ^ such that

u-. Art/ sin - EH? (8). (8,4,1,2)

In theory this solution u is obtained by applying the variational method

of Lemma 2.2.1.1. We set
V =H'(8)
386 MISCELLANEOUS

and
U, U€V.
Then u is the unique element of V such that

alu, o)= fodx (8,4,1,3)

for every v€ V.
The Galerkin method for approximating u consists in replacing the
space V by a finite-dimensional subspace V of V in the above setting.
Thus we consider un the unique element of V such that

(8,4,1,4)
for every ve Vh. This new problem is equivalent to a set of n linear
algebraic equations with n unknowns, n being the dimension of V. On
the one hand one expects to solve explicitly this set of equations with
the minimal amount of calculations. On the other hand one expects that un
significantly approximates u; in other words one wants the error

to be small for some suitable norm. Satisfying both requirements depends

strongly on the choice of Wh
The basic tool for estimating the error is Céa's lemma (see Theorem
2.4.1 in Carlet (1978)):
Minfu-uliv,
u-unliv s= (8,4, 1,5)

where a denotes the coerciveness constant (cf. Lemma 2.2.1.1) while M

is the constant such that

for every u and o in V.

In the finite element method h is built with the help of a triangulation
In over M; In is a set of closed triangles (we assume that the triangles
are not "degenerate i.e. their interiors are not empty) such that

(a) p= KETh
U K;
(b) for each distinct K1, K2€ Th one has
 8.4 PROBLEMS WITH SINGULARITIES 387

(c) any edge of any triangle K, is either a subset of the boundary I' or
an edge of another triangle Kz in the triangulation.

The number h related to the triangulation In is defined by

h= max
KETh
hk
where hk is the diameter of K. This number h is supposed to vary and
approach zero. While h varies we assume that the corresponding family
of triangulations is regular, i.e. that there exists a constant o such that

AK€O
Рк
for every K in Th where Pr is the interior diameter of K. In other words,
Pr is the diameter of the biggest disc included in K.
Once a family of such triangulations has been chosen the simplest
choice of a related family of spaces n is as follows: the functions
belonging to Vh are the continuous functions on O which vanish on I and
whose restrictions to each K€ Th are 'linear' (i.e. affine).
In order to take advantage of Céa's lemma we need an estimate for
inf lu-vlv = dy (u; Vh).
vEVn

The classical result is that there exists a constant C such that

dylu;Vh)sChkHlull+1.2.02 (8,4,1,6)
provided u € Hk+ (1), k = 1, 2 (see Section 3.2 in Carlet (1978)). Ex-
trapolating this inequality to non-integral values of k, 1 <k <2, and taking
in account (8,4,1,2), one expects here the estimate
dv (u; Vh) =O(h'/w==)
for every & >0. Indeed, by Theorem 1.4.5.3, we have
UE H'+I/) = (D)\ HI+(T/w) (12)
for every & >0, if 1 does not vanish. Even choosing higher-order finite
element spaces leads to the same limitation of the asymptotic rate of
convergence of the error as h->0. However, as we shall show now, the
above inequality (8,4,1,6) does not yield the best estimate of the asymp-
totic rate of convergence provided some additional assumptions are made
on Th
Indeed, the property (8,4,1,2) prevents u from belonging to H7(2)
when 1 does not vanish. Nevertheless it allows u to belong to a weighted
space corresponding to the second order of differentiation. For this
purpose let us set a new definition.
388 MISCELLANEOUS

Definition 8.4.1.1 For a a nonnegative real number, we denote by

H2. (R) the space of all functions u € H' (R) such that in addition
r°DBuE L2 (12)
for every B such that B = 2.

We observe that for w € ]m, 2m [ we have u € H2.9(2) for every a such

that
(8,4,1,7)
Some preliminary properties of those spaces will be useful.

Lemma 8.4.1.2 We equip H2ª(D) with the norm

131=2

Then the natural imbedding of H2. (R) into H'(1) is compact for a < 1. In
addition H2.9(1) is continuously imbedded in C° (R).

Proof A mere application of Hölder's inequality shows that

H2.°(1) c W=(S)
for every p such that 1 <p <2/(a + 1). Furthermore the corresponding
imbedding is continuous.
It follows that H29(1) is continuously imbedded in C°(R) by Theorem
1.4.5.2, provided & <1. The compactness of the imbedding of H2,a (2)
into H'(8) is a consequence of Theorem 1.4.3.2.

Lemma 8.4.1.3 Let P.(R) be the space of the first-order polynomials

restricted to S. Then there exists a constant C such that

pEP (52)
(8,4, 1,8)
3 =2

for every u € H2.a (52).

It is worth observing the similarity of this lemma with Theorem 3.1.1 in

Carlet (1978).
Proof A first step is the proof of the following inequality
(8,4,1,9)
3\=2
for every ve P.(M)' the orthogonal of P(1) in H29(12).
 8.4 PROBLEMS WITH SINGULARITIES 389

Indeed if (8,4,1,9) does not hold, there exists a sequence Uns h =

1, 2, ... of functions in P1(8)- such that
(8,4,1,10)
for every n, while
F°DEUn ->0, 1B|=2 (8,4,1,11)
in Lz(12) as n -> +00.
The compactness of the imbedding of H2, (1) into H'(1) (Lemma
8.4.1.2) implies that there exists a subsequence which is strongly con-
vergent in H'(1). Again we denote this subsequence by Uns n = 1, 2, ...
and thus there exists v € H'(Q) such that
Un -> U
in H'(2) as n-> +00.
Next, by the very definition of the norm in H2.4(8), Uns n = 1, 2, ... is a
Cauchy sequence in H?.«(1). Indeed we have

161=2

and both terms on the right-hand side converge to zero as n and m -> +00.
Accordingly we have

and
Un >U
in the norm of H29(8). It follows that U€ P, (5)" since U, € P, (9)+ for
every n, and furthermore (8,4,1,11) implies that
B=2.
It follows that v € P. (2)n P. (2)', i.e. v = 0. This contradicts (8,4,1,10),
which implies that
UH?"(p) = 1.
Now we complete the proof by observing that (8,4,1,8) follows from
(8,4,1,9) by setting
v = u - p,
where p is the orthogonal projection of u onto P(9). •
From now on we denote by R the model triangle whose vertices are
390 MISCELLANEOUS

(0, 0), (0, 1) and (1, 0). For any function u € H2.° (K), we denote by [lu
the first order interpolating polynomial i.e.
ПиєР.(K)
and
[Tu = u at (0, 0), (0, 1) and (1, 0).
This makes sense since u is continuous by Lemma 8.4.1.2. Then for every
peP. (K) we have
и-Ilu =(1-I)(u-p).
Both the identity operator and I are continuous from Ha(K) into
H'(K). Consequently there exists a constant @ such that

and thus we have

Taking the infimum in p it follows from (8,4,1,8) that

(8,4,1,12)
161=2

for every u € H2.9(K).

The above inequality is fundamental in the sequel. We shall need a
similar inequality on an arbitrary triangle. For this purpose let us consider
a triangle K whose vertices are a, b, c with
a = (a1, az), b=(a, +by, az+b2), c= (a, +c, az+ (2).
The triangle K is the image of the model triangle  under the affine
mapping
Pk: *-> a + TkX,
where the matrix of Tk is
b, cl
Los c2.
We have already introduced above the numbers h (diameter of K)
and Pr (radius of the biggest circle contained in K). We can estimate Tk
with the help of these numbers: obviously we have

and V2 (8,4,1,13)
MIkll<v2 hk
Рк
For u € H?. (K) we denote by Iku the first-order interpolating polyno-
 8.4 PROBLEMS WITH SINGULARITIES 391

mial, i.e.
Пки€ Р.(K)
and
Пки = и at the vertices of K.
We have the following estimate:

Lemma 8.4.1.4 There exists a constant C independent of the triangle K

such that
IW(u -Пки)6.2k

for every u€ H?.« (K).

Proof We set ü = u o k. Obviously we have § € H2.°(K), and in addi-

tion
(Пки) о Фк = Пй.
Then we can apply inequality (8,4,1,12) to : this yields

or equivalently

Next applying the chain rule for differentiation we get

Finally we perform the obvious change of variable, setting x = k(x).

Thus we obtain

131=2

The desired inequality follows since we have

r(Dk' (x)) = d(Dk' (x); Ok'(a)) S||Tk'll d(x, a). •
392 MISCELLANEOUS

Remark 8.4. 1.5 We shall use inequality (8,4, 1,14) only in two particular
cases. First, when a =0, we get the inequality (already proved in Carlet
(1978)):
(8,4,1,15)
|B|=2
Second when a <1 and a = 0 (i.e. one of the vertices of K is the origin)
one gets the weighted inequality
(8,4,1,16)
16\=2

The next statement is an easy consequence of these preliminaries. We

consider a triangulation over 0 as above and Th the interpolation
operator, defined as follows for every u€ H2.9(1):
(a) Inuk EP (K) for every K€ Th
(b) Inu = u at any vertex of any K€ Th
Theorem 8.4.1.6 We assume that the family of triangulations In satisfies
the following conditions as h->0; there exists o such that
(a) maxes. he/Pk So for every hi
(b) hk Soh'''-a) for every K€ In such that one of the corners of K is at
O;
(c) he Sch infk y* for every KEIn with no corner at O.
Then there exists a constant C such that
(8,4, 1,17)
for every h >0 and every u€ H3.9(1), provided a <1.

Proof We observe that for every k € In the restriction Inulk of Ink to

K is just Ikíuk), where uk is the restriction of u to K. Thus we can
apply one of the inequalities (8,4,1,15), (8,4,1,16) to UK.
If one of the vertices of K is 0 we make use of (8,4,1,16): this yields

3=2
On the other hand if no vertex of K is O we make use of (8,4,1,15). Thus
we get

IB|=2
 8.4 PROBLEMS WITH SINGULARITIES 393

In both cases the inequalities (8,4,1,13) and the assumptions (a)-(c) in

the statement of Theorem 8.4.1.6 imply the following inequality:

131=2

with possibly another value for the constant C (yet independent of K and
u). Inequality (8,4,1,17) follows by addition (over K€ Th) and with the
help of Poincaré's inequality (cf. Theorem 1.4.3.4). I

Corollary 8.4.1.7 If the triangulation Th fulfils the conditions in Theorem

8.4.1.6, then there exists a constant C which depends on neither u nor h
such that
(8,4,1,18)
(we recall that u and un are defined by (8,4,1,1) and (8,4,1,4) respectively).

This follows obviously from inequality (8,4,1,17) and Céa's lemma

(inequality (8,4,1,5)).
This result shows that one can expect the same asymptotic rate of
convergence (as h->0) of the error (in the norm of H'(R)) as in the
regular case provided the mesh is refined in a suitable way near 0. In
addition it is also shown in Babuska et al. (1979) that this is the best
asymptotic rate of convergence that one can expect for spaces In such
that the dimension does not grow faster than O(h 2) as h->0. This can
also be derived from the asymptotic estimates of diameters in El-Kolli
(1971).
In practice one has to make sure that meshes refined in the above way
do exist. In Raugel (1978), the following method is proposed:
First step: divide O into 'big' triangles;
Second step: divide each of the big triangles which have no vertex at
zero in the usual uniform way (i.e. divide each side into n
subsegments of the same length and proceed).
Third step: divide each of the big triangles which have a vertex at zero,
according to the ratios
1/(1-0)

along the sides which end at zero. Divide the third side in
the usual way (see figure) and proceed as usual.
With such a procedure the dimension of Wh is equivalent to n? (as
n=h- goes to infinity).
394 MISCELLANEOUS

131
(4)
Figure 8.5

Remark 8.4.1.8 It is important to emphasize that, once a refined mesh

has been chosen, the algebra of the approximate problem (8,4, 1,4) is
unchanged with respect to the regular case. In other words the stiffness
matrix has the same structure as in the case when all the triangles have
the same order of magnitude.

Remark 8.4.1.9 We observe that all the previous analysis has been done
starting from the knowledge that u belongs to some weighted space
H29(8). We never used the explicit form of the singularity given by
(8,4,1,2). This is the reason why such an analysis may be carried out for
problems which are much more general than the model problem (8,4,1,1).

Remark 8.4.1.10 No attempt is made in the current literature about the

numerical treatment of singular problems to estimate the constant C in
inequality (8,4,1,18). In practice, the implementation is done with a
chosen h>0 and the order of magnitude of C is as important as the
asvmptotic rate of convergence in h for estimating the error.

8.4.2 Augmenting the space of trial functions

Another method for overcoming the polluting effect of the corners on the
finite element method has been proposed. Instead of refining the mesh
one keeps a regular mesh all over S, i.e. we consider triangulations In
such that there exists o, and oz such that
ohspkshk50zh
 8.4 PROBLEMS WITH SINGULARITIES 395

for every K € Th. However, we shall consider a bigger space n than before.
To be precise let us denote by En the space of the continuous functions
u on § whose restriction to each K€ Th belongs to P(K) and which
vanish on I. We also introduce a cut-off function n, of r, which is
identically 1 near zero and which vanishes near all the I; except I and
I~ (notation of Subsection 8.4. 1). Then we define Vh as the direct sum of
En and the one-dimensional space generated by the function

TO =n (r) Us (r, 0).

In other words we set

Vh=En DRinus'.
With this new space h we consider the approximate problem (8,4, 1,4).
Again we can apply Cea's lemma and we get

M infu-vll.2.o.
a UEVn

From (8,4,1,2) we know that

u=w+AnUs,
where we H?(R). Thus we have
influ-olla.2,o
vEVn = inf W -Q|11.2,0
PEEn

and from the results in Section 3.2 of Carlet (1978), we conclude that
there exists a constant C such that
infw-Al1.2.o
SEEn SChW|2.2.p.
Summing up we have
(8,4,2,1)
with, possibly, another value for C. With such a choice of Vhs we have
obtained the same asymptotic rate of convergence as in the regular case
and again here the dimension of Wh is of the order h 2 as in Subsection
8.4.1).
Adding the one-dimensional space generated by us to Eh the usual
space of trial functions, disturbs the sparseness of the matrix of problem
(8,4,1,4). Therefore it is advisable to choose the support of n small
enough to be contained in those K€ Th which have one corner at 0.
Accordingly an accurate study of the asymptotic error estimate might be
done with a cut-off function n depending also on h. Some details can be
found in Destuynder and Diaoua (1979) and Lelievre (1976b).
396 MISCELLANEOUS

8.4.3 Calculating the stress intensity factor

The coefficient ^ of the singular part of the solution in (8,4,1,2) is often
called the 'stress intensity factor'. Indeed in mechanical problems the
stresses are given by the gradient of the solution u. Here we have

u = Art/o sin - + W,

where we H?(M). Actually it was shown in Chapter 4 that there exists

Po > 2 such that we W$(8) for every feLp (1) provided p <Po.† Conse-
quently the gradient of w is bounded and the unbounded part of Wu is
Art/w sin
In most practical problems one is mainly interested in calculating (ap-
proximately) ^ rather than the whole solution u. When one works with the
method outlined in Subsection 8.4.2, one can show an error estimate of
the form

where An is defined by
Un-Annus€Eh.
This is proved in Djaoua (1983) in the particular case of a crack. This
shows poor convergence of in toward 1.
In addition it is not clear how to compute ^ when one uses only mesh
refinements as outlined in Subsection 8.4.1.
Several approaches have been proposed in Schatz and Wahlin (1978-
79), Destuynder and Djaoua (1979) and finally in Destuynder et al.
(1981, 1983) for the particular case of a crack. Here we shall rely on a
very simple method devised in Bellout and Moussaoui (1981). We rely on
the results in Section 4.4. There we introduced (Subsection 4.4.1) the
space Nz of all functions u in L2(8) which are orthogonal to the image of
H?(8)M H' (S) by 4. We saw that every o belonging to N- is harmonic in
1 and vanishes on the boundary I in the weak sense. In addition we have
proven (in Subsection 4.4.2) that N is one-dimensional and generated by
one function o such that

O-p-To sin є H' (8).

† To be precise we have

Po= min 3(1-. min

1=1,2,
UN
 8.4 PROBLEMS WITH SINGULARITIES 397

This suggests that 1 is proportional to the scalar product

La for ax.
Unfortunately the function o is not known explicitly because & is too
general a domain. Consequently we are unable to calculate the propor-
tionality ratio above. To make this calculation feasible we replace 2 by
the simpler domain

where p is chosen small enough such that

minN-1d(0,
p< i=2. I;).
In addition we assume that n vanishes for r >p/2. We can now state the
basic preliminary result of this subsection.

Lemma 8.4.3.1 Assume that u€ H'(D) and that

U-Apt/wsin - € H2(0);

then

^= dx. (8,4,3,1)
w

Proof We use the same method as in Lemma 4.4.4.10.

We denote by w the difference

И- Apt/w sinTA= U - NUs.

First we shall calculate the integral

TO• dx
by applying Theorem 1.5.3.6. For this purpose we introduce a second
cut-off function n, which is equal to one on the support of n and which
vanishes for r > p. Thus we have

1,=[ A(nw)n,p the site. TO- dx.

398 MISCELLANEOUS

It is clear that we W4(8) for q s2 and that the function

r,O->n rT/w sin = (r, e)

belongs to D(A; Lp (R)) for p < 2w/m. In addition we have
nw(S;)=0
for every j. Thus (1,5,3,6) yields

" I) dx =0.
since the traces of all the involved functions are zero on the boundary and
the function r*/w sin 0/ is harmonic on the support of n.
In order to calculate

12= Anywho sinTO'p-T/w sin dx,

we apply the Green formula on the subdomain

and we take the limit as &->0. This yields

I=lim A(herton sin - dx

= lim & - ET/w-1 sin g-Tw sin

do la
+ a Tw sin
TO T Te
- E-I/w--1 sin

2T
sin_7072de = T.
Summing up, we have proved that

A(nu)r to sin
Let us set
W=r-Twsin
 8.4 PROBLEMS WITH SINGULARITIES 399

Integrating by parts we derive from identity (8,4,3,1) that

^=

=- (8,4,3,2)
due to the fact that Wn vanishes for small r and for r > p/2.
This suggests that we define an approximate value for ^ by setting

Ah == (8,4,3,3)
"I wear =
where un is the solution of (8,4,1,4). It is easy to estimate 1 - 1.

Theorem 8.4.3.2 Let u be the solution of the problem (8,4,1,1), Un the

solution of the approximate problem (8,4,1,4) and ^ such that (8,4,1,2)
holds. Assume In is defined by (8,4,3,3) and that the triangulations In are
regular. Then there exists a constant C (which does not depend on h or f)
such that
1-141-Ch flo.2.n.
Proof Clearly it follows from (8,4,3,2) and (8,4,3,3) that
(8,4,3,4)
for some constant K. Then as we have shown in Section 8.4.1 it follows
from (8,4,1,6) that
Iu-unlli.2.5=O(vh)
at least. In addition one has u€ H3/2(8) at least (assuming a <2).
Therefore the Aubin-Nitsche trick (cf. Carlet (1978)) implies that
u-unlo.2.n=O(h). (8,4,3,5)
The result obviously follows from (8,4,3,4) and (8,4,3,5).

Remark 8.4.3.3 Let us emphasize that no mesh refinement is needed in

Theorem 8.4.3.2.
Bibliography

Adams, R. (1975) Sobolev Spaces, Academic Press.

Agmon, S. (1965) Lectures on Elliptic Boundary Value Problems, Van Nostrand
Mathematical Studies No. 2.
Agmon, S., Douglis, A. and Nirenberg, L. (1959) Estimates near the boundary for
solutions of elliptic partial differential equations satisfying general boundary
conditions, 1, Commun. Pure Appl. Math., 12, 623-727.
Agranovitch, M. S. and Vichik, M. I. (1964) Elliptic problems with parameters
and parabolic problems of general type, Uspeki Mat. Nauk, 19 (3), 53--161 (and
Russian Mathematical Survey, 19, 53-157).
Akin, J. E. and Whiteman, J. R. (1978) Finite Elements, Singularities and
Fracture, Technical report, Brunel University.
Ambrosetti, A. and Prodi, G. (1973) Analisi Non Lineare, Scuola Normale
Superiore, Pisa, Quaderno I.
Aronszajn, N. (1951) Green's functions and reproducing kernels, Proc. Sym-
posium on Spectral Theory and Differential Problems, Stillwater, Oklahoma,
355-412.
Aronszajn, N. and Smith, K. T. (1961) Theory of Bessel potentials, 1, Ann. Inst.
Fourier, Grenoble, 11, 385-475.
Avantaggiati, A. (1975) Due teoremi di isomorfismo per operatori associati a
sis mi ellittici di equazioni a derivate parziali a coefficienti costanti in spazi di
Sobolev con peso su R? * Rendiconti di Matematica, (2) 8, Series 6, 509-517.
Avantaggiati, A. (1976) Spazi di Sobolev con peso ed alcune applicazioni,
Bolletino U.M.I., (5) 13-A, 1-52.
Avantaggiati, A. and Troisi, M. (1973-74) Spazi di Sobolev con peso e problemi
ellittici in un angolo, Annali di Matematica pura ed applicata, Series. 95,
361-408 (1973); 97, 207-252 (1973); 99, 1-64 (1974).
Avantaggiati, A. and Troisi, M. (1974a) Ulteriori contributi allo studio dei
problemi ellittici in un angolo, Annali di Matematica pura ed applicata, Serie
4, 100, 153-168.

400
 BIBLIOGRAPHY 401

Avantaggiati, A. and Troisi, M. (1974b) Problemi al contorno di tipo ellittico in

un dominio piano limitato e dotato di punti angolosi, Atti della Accademia
Nazionale dei Lincei (Memoria), Series 8, 12 (4), 271-308.
Azzam, A. (1979) On the first boundary value problem for elliptic equations in
regions with corners, Arabian J. Science and Engineering, 4, 2.
Azzam, A. (1980) Smoothness properties of bounded solutions of Dirichlet's
problem for elliptic equations in regions with corners on the boundary, Canadian
Math. Bull., 23 (2), 213-226.
Azzam, A. and Kreyszig, E. (1980) On parabolic equations in n space variables
and their solutions in regions with edges, Hokkaido Math. J., 9, 140-154.
Babuska, I. and Kellogg, R. B. (1972) Numerical solution of the neutron diffusion
equation in the presence of corners and interfaces, Numerical Reactor Calcula-
tions, I.A.E.A., 473-486.
Babuska, I., Kellogg, R. B. and Pitkäranta, J. (1979) Direct and inverse error
estimates for finite elements with mesh refinements, Numerical Math., 33,
447-471.
Baderko, E. A. (1973) On smoothness in Hölder spaces of potentials for 2p
parabolic equations, Dokl. Akad Nauk S.S.S.R., 211 (5) 1017-1020 (and Soviet
Math. Doklady, 14 (4), 1128-1132).
Baderko, E. A (1975) On the solution of boundary value problems for 2p-
parabolic equations in domains with curvilinear lateral boundaries, Dokl. Akad
Nauk S.S.S.R., 225 (1), 21-23 (and Soviet Math. Doklady, 16 (6), 1425-1427).
Baderko, E. A. (1976) On the solution of boundary value problems for parabolic
equations of high order in domains with curvilinear lateral boundaries, Diff.
Urav., 12 (10), 1781-1792.
Bailet-Intissar, J. (1978) Problèmes aux limites non linéaires avec conditions
mélées, Note C.R.A.S. Paris, 287, 1061-1063.
Bellout, H. and Moussaoui, M. (1981) Quelques remarques sur l'approximation
numérique du problème de Dirichlet pour le Laplacien dans un ouvert avec coin,
Prépublications, Alger.
Besov, O. V. (1967) The continuation of functions in L, and W", Proc. Steklou
Inst. Math., 88 (and Translations A.M.S. (1969)).
Boutet de Monvel, L. (1971) Boundary problems for pseudo-differential
operators, Acta Mathematica, 126, 11-51.
Brezis, H. (1971) Monotonicity Methods in Hilbert Spaces, Academic Press.
Caccioppoli, R. (1950-51) Limitazioni integrali per le soluzioni di un'equazione
lineare ellittica a derivate parziali, Giornale di Battaglini, Series 4, 5, 80,
186-212.
Calderon, A. (1961) Lebesgue spaces of differentiable functions and distributions,
Proc. Symposium in Pure Mathematics, A.M.S., 4, 33-49.
Chenais, D. (1973) Un résultat de compacité d'un ensemble de parties de R",
C.R.A.S. Paris, 277, 905-907.
402 BIBLIOGRAPHY

Ciarlet, P. G. (1978) The Finite Element Method for Elliptic Problems, North
Holland, Amsterdam.
Courant, R. and Hilbert, D. (1953/1962) Methods of Mathematical Physics, Vols. 1
and 2, Interscience, New York.
Destuynder, P. and Djaoua, M. (1979) Estimation de l'Erreur de la Singularité de
la Solution d'un Problème Elliptique sur un Ouvert avec Coin, Centre de
Mathématiques Appliquées, Ecole Polytechnique, Palaiseau.
Destuynder, P., Diaoua, M. and Lescure, S. (1981) Une nouvelle méthode de calcul
de la singularité pour un problème elliptique pose sur un ouvert avec fente,
C.R.A.S. Paris, 293, 343-346.
Destuynder, P., Diaoua, M. and Lescure, S. (1983) Quelques remarques sur la
mécanique de la rupture élastique, J. Mécanique Théorique et Appliquée, 2 (1),
113-135.
Djaoua, M. (1977) Equations Intégrales pour un Problème Singulier dans le Plan,
Thèse de Troisième Cycle, Université Pierre et Marie Curie, Paris.
Djavua, M. (1978) A Method of Calculation of Lifting Flows around 2-
Dimensional Corner Shaped Bodies, Centre de Mathématiques Appliquées.
Ecole Polytechnique, Palaiseau.
Djaoua, M. (1983) Méthodes Numériques en Mécanique de la Rupture Élastique et
en Viscoplasticité, Thèse, Ecole Polytechnique de Palaiseau.
Eggleston, H. G. (1958) Convexity, Cambridge Tracts in Maths and Mathematical
Physics No. 47, Cambridge University Press.
El-Kolli, A. (1971) n-ième épaisseur dans les espaces de Sobolev avec poids, Note
C.R.A.S. Paris, 273, 450-453.
Eskin, G. I. (1973) Boundary Value Problems for Elliptic Pseudo-differential
Equations, Nauka, Moscow.
Fabes, E. B., Jodeit, M. and Lewis, J. E. (1977) Double layer potentials for
domains with corners and edges, Indiana University Math. J., 26 (1), 95-114.
Fabes, E. B., Jodeit, M. and Riviere, N. M. (1978) Potential techniques for
boundary value problems on C'-domains, Acta Mathematica, 141, 165-186.
Fix, G. J., Gulati, S. and Wakoff, G. I. (1976) On the use of singular functions
with finite elements approximations, J. Computational Physics, 13, 209-228.
Freeman, R. S. and Schechter, M. (1974) On the existence and regularity of
solutions to general elliptic boundary value problems, J. Differential Equations,
1(15). 213-246.
Gagliardo, E. (1957) Caratterizzazioni delle trace sulla frontiera relative ad
alcune classi di funzioni in n variabili, Ren. Sem. Mat. Univ. Padova, 27, 284-305.
Gagliardo, E. (1958) Proprietà di alcune classi di funzioni in più variabili, Ric. di
Mat., 1, 102-137.
Gevrey, M. (1913) Sur les équations aux dérivées partielles du type parabolique,
J. Math. Pures Appliquées, (6) 9, 305-471.
Goulaouic, Ch. and Grisvard, P. (1970) Existence de traces pour les éléments
 BIBLIOGRAPHY 403

d'espaces de distributions définis comme domaines d'opérateurs différentiels

maximaux, Inventiones Math., 9, 308-317.
Grisvard, P. (1963) Espaces intermédiaires entre espaces de Sobolev avec poids,
Annali della Scuola Normale Superiore di Pisa, Series 3, 17 (3), 255-296.
Grisvard, P. (1966) Commutativité de deux foncteurs d'interpolation et applica-
tions, J. Math. Pures Appliquées, 45, 143-290.
Grisvard, P. (1972) Alternative de Fredholm relative au problème de Dirichlet
dans un polygone ou un polyèdre, Bollettino U.M.I., 132-164.
Grisvard, P. (1975a) Behavior of the solutions of an elliptic boundary value
problem in a polygonal or polyhedral domain, Symposium on Numerical Solu-
tions of Partial Differential Equations, III, B. Hubbard, Ed., Academic Press,
207-274.
Grisvard, P. (1975b) Alternative de Fredholm relative au problème de Dirichlet
dans un polyèdre, Annali S.N.S. Pisa, Series 4, 2 (3), 359-388.
Grisvard, P. (1984) Singular solutions of elliptic boundary value problems in
polyhedra, Portugaliae Mathematica, 41, 4.
Grisvard, P. and looss, G. (1975) Problèmes aux limites unilatéraux dans des
domaines non réguliers, Publications des Séminaires de Mathématiques,
Université de Rennes, exposé 9, 1-26.
Hanna, M. and Smith, T. (1967) Some remarks on the Dirichlet problem in
piecewise smooth domains, Comm. Pure and Applied Math., 20, 575-593.
Hardy, G., Littlewood, D. and Polya, G. (1952) Inequalities, Cambridge Univer-
sity Press.
Hopf, E. (1927) Elementare Bemerkungen über die Lösungen partieller Differen-
tialgleichungen zweiter Ordnung vom elliptischen Typus, S.-B Preussischen
Akad. Wiss, 19, 147-152.
Hörmander, L. (1960) Estimates for translation invariant operators in L° spaces,
Acta Mathematica, 104, 1-140.
Hörmander, L. (1963) Linear Partial Differential Operators, Springer-Verlag.
Ibuki, K. (1974) Dirichlet problem for elliptic equations of the second order in a
singular domain of R?, J. Math. Kyoto Univ., 14 (1), 54-71.
Il'in, A. M. (1976) A boundary value problem for the elliptic equation of second
order in a domain with a narrow slit. 1: The two-dimensional case, Mat.
Sbornik, 99 (141), 4 (and Math. U.S.S. R. Sbornik, 28 (4), 459).
Kadlec, J. (1964) On the regularity of the solution of the Poisson problem on a
domain with boundary locally similar to the boundary of a convex open set,
Czechoslovak Math. J., 14 (89), 386-393.
Kato, T. (1966) Perturbation Theory for Linear Operators, Springer Verlag, New
York.
Kawohl, B. (1980-81) On nonlinear mixed boundary value problems for second
order elliptic differential equations on domains with corners, Proc. Royal
Society of Edinburgh, Sect. A, 87 (1-2), 35-51.
404 BIBLIOGRAPHY

Kellogg, R. B. (1971) Singularities in interface problems, Symposium on Numeri-

cal Solutions of Partial Differential Equations, II, Academic Press, 351-400.
Kellogg, R. B. and Osborn, J. E. (1976) A regularity result for the Stokes
problem in a convex polygon, J. Functional Analysis, 21 (4), 397-431.
Khelif, A. (1978) Problèmes aux limites pour le Laplacien dans un domaine à
points cuspides, C.R.A.S. Paris, 287, 1113-1116.
Kondrasov (1945) Certain properties of functions in the spaces L', Dokl. Akad
Nauk, S.S.s.R., 48, 535-538.
Kondratiev, V. A. (1967a) Boundary value problems for elliptic equations in
domains with conical or angular points, Trudy Moskoukogo Mat. Obschetsva,
16, 209-292 (and Transactions Moscow Mat. Soc., 227-313).
Kondratiev, V. A. (1967b) Asymptotic of the solution of the Navier-Stokes
equation near the angular part of the boundary, Prikl. Mat. Mech. 31, 119-123;
(and J. Appl. Math. Mech., 31, 125-129).
Kondratiev, V. A. (1970) The smoothness of a solution of Dirichlet's problem for
2nd order elliptic equations in a region with a piecewise smooth boundary,
Differential' ye Uravneniya, 6 (10), 1831-1843 (and Differential Equations, 6,
1392-1401).
Krein, S. G. (1967) Linear Differential Equations in Banach Spaces, Nauka,
Moscow (and (1971) Providence, R.I., A.M.S.).
Ladeveze, J. and Peyret, R. (1974) Calcul numérique d'une solution ave
singularité des équations de Navier-Stokes: écoulement dans un canal ave
variation brusque de section, Journal de Mécanique, 13 (3), 367-396.
Ladyzhenskaia, O. A. and Ural'ceva, N. N. (1968) Equations à Dérivées Partielles
de Type Elliptique, Dunod, Paris.
Lehmann, R. S. (1959) Developments at an analytic corner of solutions of elliptic
partial differential equations, J. Mathematics and Mechanics, 8 (5), 727-760.
Lelievre. J. (1976a) Sur l'utilisation de fonctions singulières dans la méthode des
éléments finis, C.R.A.S. Paris, 283, 863-865.
Lelievre, J. (1976b) Sur les éléments finis singuliers, C.R.A.S. Paris, 283, 1029-
1032.
Lemrabet, K. (1977) Régularité de la solution d'un problème de transmission, J.
Math. Pures et Appliquées, 56, 1-38.
Lemrabet, K. (1978) An interface problem in a domain of 13, J. Mathematical
Analysis and its Applications, 63, 549-562.
Lewis, J. E. (1978) Estimates for Poison's equations when the curvature of the
boundary is in an Orlicz class, Boll. U.M.I., (5) 15-A, 87-93.
Lions, J. L. (1956) Sur les problèmes aux limites du type dérivée oblique, Annals
of Mathematics, 64 (2).
Lions, J. L. (1957) Lectures on Elliptic Partial Differential Equations, Tata
Institute of Fundamental Research, Bombay.
 BIBLIOGRAPHY 405

Lions, J. L. (1961a) Equations Différentielles Opérationnelles et Problèmes aux

Limites, Springer-Verlag.
Lions, J. L. (1961b) Sur les espaces 'interpolation, dualité, Math. Scand., 9,
147-177.
Lions, J. L. (1962) Problèmes aux Limites dans les Équations aux Dérivées
Partielles, Université de Montréal.
Lions, J. L. and Magenes, E. (1960-63) Problèmes aux limites non homogènes
(Problemi ai limiti non omogenei), 1, Annali S.N.S. Pisa, 14 (3), 269-308
(1960); 2, Annales de l'institut Fourier, 11, 137-178 (1961); 3, Annali S. N.S.
Pisa, 15, 39-101 (1961); 4, Annali S.N.S. Pisa, 15, 311-326 (1961); 5, Annali
S. N.S. Pisa, 16, 1-44 (1962); 6, J. d'Analyse Mathématique, Jérusalem, 11,
165-188 (1963).
Lions, J. L. and Magenes, E. (1968) Problèmes aux Limites Non Homogènes et
Applications, Dunod, Paris.
Lions, J. L. and Peetre, J. (1964) Sur une classe d'espaces 'interpolation,
Publications I.H.E.S., 19, 5-68.
Lorenzi, A. (1975-76) A mixed boundary value problem for the Laplace equation
in an angle, 1, 2, Ren. Sem. Mat. Padova, 54, 147-183 (1975); 55 (1976).
Lorenzi, A. (1978) Mixed boundary value problems for non-variational elliptic
equations on a polygon, Boll. U.M.I., (5) 15-B, 401-435.
Lozi, R. (1978) Résultats numériques de régularité du problème de Stokes et du
Laplacien itéré dans un polygone, RAIRO, Analyse Numérique, 12 (3), 267-
282.
Magenes, E. and Stampacchia, G. (1958) I problemi al contorno per le equazioni
differenziali di tipo ellittico, Annali S. N. S. Pisa, Serie III, 12 (II1), 247-357.
Maz'ya, V. G. and Plamenevskiì, V. A. (1978) L, estimates of solutions of elliptic
boundary value problems in domains with edges, Trudy Moskov. Mat. Ob. 37
(1978) (and Transactions of the Moscow Mathematical Society (1980), Issue no.
1, 49-97).
Merigot, M. (1971) Régularité des dérivées de la solution du problème de
Dirichlet dans un secteur plan, C.R.A.S. Paris, 273, 356-359.
Merigot, M. (1972) Regularité des dérivées de la solution du problème de
Dirichlet dans un secteur plan, Le Matematiche, 28 (2), 1-36.
Merigot, M. (1974a) Etude de problème Au = f dans un polygone plan. Inégalités à
priori, Boll. U.M.I., (4) 10, 577-597.
Merigot, M. (1974b) Potentiels et régularité L pour les problèmes elliptiques
dans un secteur pian, C.R.A.S. Paris, 278, 1487-1489.
Meyer, I. (1978) Nouvelles Estimations pour les Opérateurs Pseudo-différentiels,
Séminaire Goulaouic-Schwartz, exposé No. 10.
Mghazli, Z. (1983) Une méthode intégrale pour un problème mixte de Dirichlet-
Robin dans un polygone, Thèse de Troisième Cycle, Marseille.
406 BIBLIOGRAPHY

Mih'lin, S. (1956) Sur les multiplicateurs des intégrales de Fourier, Dokl. Akad.
Nauk. S.S.S.R., (109) 4, 701-703.
Minty, G. (1962) Monotone (non linear) operators in Hilbert spaces, Duke Math.
J., 29, 341-346.
Minty, G. (1964) On the monotonicity of the gradient of a convex function, Pac.
J. Math., 14, 341-346.
Miranda, C. (1970) Partial Differential Equations of Elliptic Type, Springer-
Verlag.
Moreau, J. (1965) Proximité et dualité dans un espace hilbertien, Bull. Soc. Math.
France, 93, 273-299.
Moussaoui, M. (1974) Régularité de la solution d'un problème à dérivée oblique,
C.R.A.S. Paris, 279, 869-872.
Moussaoui, M. (1977) Etude, dans les Espaces de Hölder, de Problèmes Elliptiques
dans un Secteur Plan et, dans les Espaces de Sobolev, d'un Problème á Dérivée
Oblique dans un Polygone Plan, Thèse à 1'Université de Nice.
Najmi, M. (1978) Résolution d'un problème de Dirichlet fortement non linéaire
dans un domaine non régulier, Note C.R.A.S. Paris, 287, 1109-1112.
Necas, J. (1967) Les Méthodes Directes en Théorie des Équations Elliptiques,
Masson, Paris.
Nikishkin, A. (1979) Singularities of the solution to the Dirichlet problem for a
2nd order equation in a neighborhood of an edge, Vestnik Moskov. Univ. Mat.,
34 (2), 51-62 (and Moscow University Mathematics Bulletin, 34 (2) 53-64).
Nikol'skii, S. M. (1956a) On Dirichlet's problem in a domain with corners, Dokl.
Akad. Nauk S.S.S.R., 109 (1), 33:35.
Nikol'skii, S. M. (1956b) Boundary properties of functions in domains with
corners, Dokl. Akad. Nauk. S.S.S.R., 3 (1), 26-28.
Nikol'skii, S. M. (1956-58) Boundary properties of functions defined in domains
with corners, 1-3, Mat. Sbornik., 40 (82) (3), 303-318 (1956); 44 (86) (1),
127-144 (1957); 45 (87) (2), 181-194 (1958).
Nikol'skii, S. M. (1961) Imbedding continuation and approximation theorems for
differentiable functions of several variables, Uspeki Mat. Nauk, 16 (5) (101),
63-114.
Nikol'skii, S. M. (1975) Approximations of Functions of Several Variables and
Imbedding Theorems, Springer Verlag.
Nirenberg, L. (1959) On elliptic partial differential equations, Annali S.N.S. Pisa,
13, 115-162.
Oleinik, O. A. and Radkievitz, E. V. (1971) Second Order Equations with
Non -negative Characteristic Form, Moscow.
Osborn, J. (1975) Regularity of solutions of the Stokes problem in a polygonal
domain, Symposium on Numerical Solutions of Partial Differential Equations,
III, B. Hubbard, Ed., Academic Press, 393-411.
 BIBLIOGRAPHY 407

Protter, M. H. and Weinberger, H. F. (1967) Maximum Principles in Differential

Equations, Prentice-Hall.
Rademacher (1919) Über partielle und totale Differenzierbarkeit, Math. Ann.,
79, 340-359 (1919); Math. Ann., 81, 52-63 (1920).
Raugel, G. (1978) Résolution numérique par une méthode d'éléments finis du
problème de Dirichlet pour le Laplacien dans un polygone, C.R.A.S. Paris, 286,
791-794.
Sadallah, B. K. (1976) Régularité de la solution de l'équation de la chaleur dans
un domaine plan non rectangulaire, Bollettino U.M.I., (5) 13-B, 32-54.
Sadallah, B.K. (1977) Régularité de la solution de l'équation de la chaleur dans
un domaine plan sans condition de corne, Note C.R.A.S. Paris, 284, 599-602.
Sadallah, B. K. (1978) Une équation 2m-parabolique dans un domaine plan
singulier, Note C.R.A.S. Paris, 287, 1025-1027.
Sadallah, B. K. (1983) Etude d'un problème 2m-parabolique dans des domaines
plans non rectangulaires, Bollettino U.M.I., (6) 2-B, 51-112.
Schatz, A. and Wahlbin, L. (1978-79) Maximum norm estimates in the finite
element method on plane polygonal domains, Part 1, Mathematics of Computa-
tion, 32 (141), 73-109 (1978); Part 2 (refinements), Mathematics of Computa-
tion, 33 (146), 465-492 (1979).
Schwarz, J. T. (1965) Nonlinear Functional Analysis Lecture Notes, Courant
Institute, New York.
Seeley, S. (1964) Extensions of C° functions defined in a half-space, Proc. Amer.
Math. Soc., 15, 625-626.
Seeley, S. (1966) Singular integrals and boundary value problems, American J.
Maths, 80, 781-809.
Seif, J. B. (1973) On the Green's function for the biharmonic equation in an
infinite wedge, Transactions A.M.S., 182, 241-260.
Stampacchia, G. (1965) Le problème de Dirichlet pour les équations elliptiques
du second ordre à coefficients discontinus, Ann. Inst. Fourier, 15, 189-258.
Stein, E. M. (1970) Singular Integrals and Differentiability Properties of Functions,
Princeton University Press.
Stephan, E. (1979) Conform and mixed finite element schemes for the Dirichlet
problem for the BiLaplacian in plane domains with corners, Math. Methods in
Applied Sci., 1, 354-382.
Taibleson, M. (1964) On the theory of Lipschitz spaces of distributions on
Euclidean n space, J. Mathematics and Mechanics, 13 (3), 407-479.
Talenti, G. (1966) Equazioni ellittiche in due variabili, Le Matematiche, Catania,
21, 1-38.
Temam, R. (1977) Navier- Stokes Equations, North Holland, Amsterdam.
Thatcher, R. W. (1975) Singularities in the solution of Laplace's equation in two
dimensions, J. Inst. Maths. Applics., 16, 303-319.
408 BIBLIOGRAPHY

Triebel, H. (1978) Spaces of Besov-Hardy-Sobolev Type, Teubner Texte zur

Mathematik, Leipzig.
Uspenskii, S. V. (1962) Théorèmes d'immersion pour les classes W* généralisées
de S. L. Sobolev, Sibirskii Mat. J., 3, 418-445.
Volkov (1965a) On the differential properties of solutions of boundary value
problems for the Laplace and Poison equations on a rectangle, Trudy Mat.
Instituta Steklova, 77, 89-112 (1965) (and Proc. Steklou Institute of Maths,
101-126 (1967)).
Volkov (1965b) On the differential properties of solutions of boundary value
problems for the Laplace equations on a polygon, Trudy Mat. Instituta Steklova,
77, 113-142 (1965) (and Proc. Steklou Institute of Maths, 127-159 (1967)).
Wendland, W. L., Stephan, E. and Hsiao, G. C. (1979) On the integral equation
method for the plane mixed boundary value problem of the Laplacian, Math.
Methods in Applied Sci., 1, 265-321.
Widom, H. (1969) Lectures on Integral Equations, Van Nostrand Mathematical
Studies No. 17.
Yakovlev, G. N. (1965) The Dirichlet problem for domains with non-Lipschitzian
boundaries, Differentisial'nye Uravneniya, 1 (8) 1085-1098.
Zolesio, J. L. (1977) Multiplication dans les espaces de Besov, Proc. Royal Soc.
Edinburgh, 78A, 113-117.
Index

A priori estimates 92-111, 138-146, 184- Hardy inequality 28

208, 328-335, 364-368 Heat equation 372-383
Adjoint operator 53 Hölder spaces 21, 275-282
Algebra (Sobolev) 28
Imbedding 27-31
Barrier function 349-351 Index (of an operator) 111
Bessel potential 15
Biharmonic equation 301-340
Bi-Lipschitz homeomorphism 7 Korn procedure 107

Characteristic equation 308-322 Laplace-Beltrami operator 363

Cone property 10 Laplace operator 83
Conical points 361-372 Lax-Milgram lemma 85
Conormal derivatives 53 Lipschitz boundary 5
Continuation in Hölder spaces 276-278 Lipschitz epigraph 10
Continuation in Sobolev spaces 25 Lipschitz submanifold 7
Continuous boundary 5 Locally invertible mapping 346-347
Continuous submanifold 7
Continuously differentiable boundary 5 Maximal extension (of an operator) 54-59
Continuously differentiable submanifold 7 Maximum principle 121-124, 348-349
Curvilinear polygon 34, 42 Mesh refinements 384-394
Cuts 13, 74-80, 248, 320 Monotone operator/Maximal monotone
operator 151
Differential quotients 87-90 Multipliers 98, 276
Differentiation in Sobolev spaces 21-24,
31-32 Navier-Stokes problem 343-344
Dirichlet
131, 139,problem 83,147-149,
142-144, 85-86. 91, 124-125, Neumann problem 83, 86, 91-92, 131,
174-181
Duality mapping 92 140-142, 144-146, 149-151, 181
Noncharacteristic operators 63
Normal system of boundary conditions 62
Edges 357-361
Elementary solution 98
Epigraph 10 Oblique boundary condition 167-174

Finite element method 384-399 Partition of the unity 91

Fourier transform 15, 99, 184, 307 Poincaré inequality 26
Friedrichs method 114 Polygons 42
Proper mapping 346
Galerkin method 386
Green formula 52-62 Regular problems 81

409
 410 INDEX

Schauder inequality 282 Weighted spaces 28-29, 200-206, 305-306,

Segment property 10 364-368, 388
Semi-Fredholm operator 111-208
Sobolev imbedding theorem 27, 34 Yosida approximation 152
Sobolev space 1, 14-20
Stokes problem 340-344
Stress intensity factor 396
Strongly elliptic operator 82
Subdifferential 151 Index of non-customary symbols

. D 182
Third boundary value problem 125
Traces of Hölder spaces 278-281
Traces of Sobolev spaces 36-62
Transposition (à la Lions-Magenes) 129-
8, 183
131
Triangulation 386 N 182
Turning points 174-181
Na 211
Uniform cone property 10 200
Uniform segment property 10 pm.« 295
Variational solution 84-86 Th 359
Vertices 371-372 Th.h