(Pure and Applied Mathematics 65) Robert A. Adams-Sobolev Spaces-Academic Press, Elsevier (1975)

SOBOLEV SPACES
ROBERT A . ADAMS
Department of Mathematics
The University of British Columbia
Vancouver, British Columbia, Canada
A CA DEM I C PR ESS New York San Francisco London 1975

A Subsidiary of Harcourt Brace Jovanovich, Publishers
COPYRIGHT (B 1975, BY ACADEMIC PRESS, INC.
ALL RIQHTS RESERVED.
NO PART OF THIS PUBLICATION MAY BE REPRODUCED OR
TRANSMITTED IN A N Y FORM OR BY ANY MEANS, ELECTRONIC
OR MECHANICAL, INCLUDING PHOTOCOPY, RECORDINQ, OR ANY
INFORMATION STORAGE AND RETRIEVAL SYSTEM, WITHOUT
PERMISSION IN WRITING FROM THE PUBLISHER.
ACADEMIC PRESS, INC.

111 Fifth Avenue, New Yak. New York 10003
United Kingdom Edition published by

ACADEMIC PRESS, INC. (LONDON) LTD.
24/28 Oval Road, London N W l
Library of Congress Cataloging in Publication Data
Adams, Robert A
Sobolev spaces.
(Pure and applied mathematics series; v. 65)

Bibliography: p.
Includes index.
1. Sobolev spaces. I. Title.
11. Series: Pure and applied mathematics; a series
of monographs and textbooks; v. 65
QA3.P8 [QA323] 5 1 W . 8 ~[515'.7] 74-17978
ISBN 0-12-044150-0
AMS (MOS)1970 Subject Classifications: 46E30,46E35
PRINTED IN THE UNITED STATES OF AMERICA

Preface
This monograph is devoted to a study of properties of certain Banach

spaces of weakly differentiable functions of several real variables which arise
in connection with numerous problems in the theory of partial differential
equations and related areas of mathematical analysis, arid which have become
an essential tool in those disciplines. These spaces are now most often asso-
ciated with the name of the Soviet mathematician S . L. Sobolev, though their
origins predate his major contributions to their development in the late 1930s.
Sobolev spaces are very interesting mathematical structures in their own
right, but their principal significance lies in the central role they, and their
numerous generalizations, now play in partial differential equations.
Accordingly, most of this book concentrates on those aspects of the theory of
Sobolev spaces that have proven most useful in applications. Although no
specific applications to problems in partial differential equations are discussed
(these are to be found in almost any modern textbook on partial differential
equations), this monograph is nevertheless intended mainly to serve as a
textbook and reference on Sobolev spaces for graduate students and researchers
in differential equations. Some of the material in Chapters 111-VI has grown
out of lecture notes [181 for a graduate course and seminar given by Professor
Colin Clark at the University of British Columbia in 1967-1968.
The material is organized into eight chapters. Chapter I is a potpourri of
standard topics from real and functional analysis, included, mainly without
Xi
Xii PREFACE
proofs, because they form a necessary background for what follows. Chapter
I1 is also largely “background” but concentrates on a specific topic, the
Lebesgue spaces LP(R),of which Sobolev spaces are special subspaces. For
completeness, proofs are included here. Most of the material in these first
two chapters will be quite familiar to the reader and may be omitted, or simply
given a superficial reading to settle questions of notation and such. (Possible
exceptions are Sections 1.25-1.27, 1.31, and 2.21-2.22 which may be less
familiar.) The inclusion of these elementary chapters makes the book fairly
self-contained. Only a solid undergraduate background in mathematical
analysis is assumed of the reader.
Chapters III-VI may be described as the heart of the book. These develop
all the basic properties of Sobolev spaces of positive integral order and
culminate in the very important Sobolev imbedding theorem (Theorem 5.4)
and the corresponding compact imbedding theorem (Theorem 6.2). Sections
5.33-5.54 and 6.12-6-50consist of refinements and generalizations of these
basic imbedding theorems, and could be omitted from a first reading.
Chapter VII is concerned with generalization of ordinary Sobolev spaces
to allow fractional orders of differentiation. Such spaces are often involved in
research into nonlinear partial differential equations, for instance the Navier-
Stokes equations of fluid mechanics. Several approaches to defining fractional-
order spaces can be taken. We concentrate in Chapter VII on the trace-
interpolation approach of J. L. Lions and E. Magenes, and discuss other
approaches more briefly at the end of the chapter (Sections 7.59-7.74). It is
necessary to develop a reasonable body of abstract functional analysis (the
trace-interpolation theory) before introducing the fractional-order spaces.
Most readers will find that a reading of this material (in Sections 7.2-7.34,
possibly omitting proofs) is essential for an understanding of the discussion
of fractional-order spaces that begins in Section 7.35.
Chapter VIII concerns Orlicz-Sobolev spaces and, for the sake of com-
pleteness, necessarily begins with a self-contained introduction to the theory
of Orlicz spaces. These spaces are finding increasingly important applications
in applied analysis. The main results of Chapter VIII are the theorem of N. S.
Trudinger (Theorem 8.25) establishing a limiting case of the Sobolev imbed-
ding theorem, and the imbedding theorems of Trudinger and T. K. Donaldson
for Orlicz-Sobolev spaces given in Sections 8.29-8.40.
The existing mathematical literature on Sobolev spaces and their general-
izations is vast, and it would be neither easy nor particularly desirable to
include everything that was known about such spaces between the covers of
one book. An attempt has been made in this monograph to present all the core
material in sufficient generality to cover most applications, to give the reader
an overview of the subject that is difficult to obtain by reading research
papers, and finally, as mentioned above, to provide a ready reference for
PREFACE xiii
someone requiring a result about Sobolev spaces for use in some application.
Complete proofs are given for most theorems, but some assertions are left for
the interested reader to verify as exercises. Literature references are given in
square brackets, equation numbers in parentheses, and sections are numbered
in the form m.n with m denoting the chapter.
Acknowledgments
We acknowledge with deep gratitude the considerable assistance we have

received from Professor John Fournier in the preparation of this monograph.
Also much appreciated are the helpful comments received from Professor
Bui An Ton and the encouragement of Professor Colin Clark who originally
suggested that this book be written. Thanks are also due to Mrs. Yit-Sin Choo
for a superb job of typing a difficult manuscript. Finally, of course, we accept
all responsibility for error or obscurity and welcome comments, or
corrections, from readers.
xv
List of Spaces and Norms
The numbers at the right indicate the sections in which the symbols are
introduced. In some cases the notations are not those used in other areas
of analysis.
I * Ile 7.2
I1 * ;B1+B2II 7.11
I ;BSPP(R")II
* 7.67
I *;Bs*p(a)ll 7.72
1.1 1.1
1.25
1.25
1.25
1.25
1.25
1.26
1.27
5.2
8.37
1.51
1.52
7.7, 7.9
xviii LlST OF SPACES AND NORMS
7.29
8.14
3.1
3.12
7.33
8.7
1-40, 1.46
1.53
2.1
3.1
2.5
5.22
7.3
7.3
7.5
7.62
7.66
8.9
1.1
7.14
7.24, 7.26, 7.32
7.32
7.35
7.43
3.1
3.1
3.11
6.54
7.11
7.30
7.36,7.39
7.39
7.48
7.49
8.27
8.27
8.27
1.4, 1.6, 6.51
1.5,1.10
1.22
I
Introductory Topics
Notation
1.1 Throughout this monograph the term domain and the symbol R shall be
reserved for an open set in n-dimensional, real Euclidean space R". We shall
be concerned with differentiability and integrability of functions defined on
Q-these functions are allowed to be complex valued unless the contrary is
stated explicitly. The complex field is denoted by @. For c E @ and two
functions u and u the scalar multiple cu, the sum u + 11, and the product uu are
always taken to be defined pointwise as
(4
( 4 = cu(x),
(u+u)(x) = u ( x ) + u(x),
(uu) ( x ) = u(x),
at all points x where the right sides make sense.
A typical point in R" is denoted by x = ( x l , ..., x,,); its norm 1x1 =
(z=, x ~ ' ) ~ ' ' The
. inner product of x and y is x . y = xjyj.
If 01 = (a1, ...,E,) is an n-tuple of nonnegative integers a j , we call a a
multi-index and denote by xu the monomial x";' ..ex:, which has degree
la1 = C'j= aj. Similarly, if Dj = i?/axj for 1 s j -< n, then
D' = D"1 ...D:
denotes a differential operator of order Jal. D(o*.-*o)u
= u.
1
2 I INTRODUCTORY TOPICS
If a and fl are two multi-indices, we say P I c1 provided /3, I aj for

1 ~j +
s n. In this case a-B is also a multi-index and la-/3I 1/31 = IQI. We
also denote
a! = a , ! ... a"!
and if fl I;a,
The reader may wish to verify the Leibniz formula
valid for functions u and u that are la1 times continuously differentiable near x .
1.2 If G c R", we denote by G the closure of G in R". We shall write G c c R

provided G c R and G is a compact (Lea,closed and bounded) subset of R".
If u is a function defined on G,we define the support of u as
SUPPU = {xEG:u(x)#O}.
We say that u has compact support in R if supp u c c R. We shall denote by
"bdry G" the boundary of G in R", that is, the set G n @where G' = R" G = N
{ x E R" :x 4 G} is the complement of G.

If x E R"and G c R", we denote by "dist(x, G)" the distance from x to G,
that is, the number inf,,,lx-yl. Similarly, if F, G c R",
dist(F,G) = infdist(y,G) = inflx-yl.
Y ~ F XEG
YEF
Topological Vector Spaces
1.3 We assume that the reader is familiar with the concept of a vector space
over the real or complex scalar field, and with the related notions of dimension,
subspace, linear transformation, and convex set. We also assume familiarity
with the basic concepts of general topology, Hausdorff topological spaces,
weaker and stronger topologies, continuous functions, convergent sequences,
topological product spaces, subspaces, and relative topology.
Let it be assumed throughout this monograph that all vector spaces referred
to are taken over the complex field unless the contrary is explicitly stated.
1.4 A topological vector space, hereafter abbreviated TVS, is a Hausdorff

topological space that is also a vector space for which the vector space oper-
NORMED SPACES 3
ations of addition and scalar multiplication are continuous. That is, if Xis a
TVS, then the mappings
( x ,y ) + x +y and (c, x ) + cx
from the topological product spaces X x X and @ x X , respectively, into X

are continuous. Xis a locally conuex TVS if each neighborhood of the origin
in X contains a convex neighborhood.
We shall outline below, mainly omitting proofs and details, those aspects
of the theory of topological and normed vector spaces that play a significant
role in the study of Sobolev spaces. For a more thorough discussion of these
topics the reader is referred to standard textbooks on functional analysis, for
example, those by Yosida [69] or Rudin [59].
1.5 By a functional on a vector space X we mean a scalar-valued function f

defined on X . The functionalfis linear provided
f ( a x + b y ) = af(x) +bfW, x,y E X , a , b E @.
If Xis a TVS, a functional on Xis continuous if it is continuous from X into
@ where @ has its usual topology, induced by the Euclidean metric.
The set of all continuous, linear functionals on X is called the dual o f X
and is denoted by X‘. Under pointwise addition and scalar multiplication
X’ is a vector space :
(f+d =f ( x ) + g(x), (cf 1( x ) = cf(x>, f,9 E X ’ , x E X , c E Q=.
X’ will be a TVS provided a suitable topology is specified for it. One such
topology is the weak-star topology, the weakest topology with respect to which
the functional F, defined on X’ by F’cf) =f ( x ) for each f E X’ is continuous
for each x E X. This topology is used, for instance, in the space of Schwartz
distributions introduced in Section 1.52. The dual of a normed space can be
given a stronger topology with respect to which it is a normed space itself
(Section 1.10).
Normed Spaces
1.6 A norm on a vector space X is a real-valued functional f on X which

satisfies
(i) f ( x ) 2 0 for all x E X with equality if and only if x = 0,
(ii) f ( c x ) = Iclf(x) for every x E X and c E @,
(iii) f ( x + y ) s f ( x ) + f ( y ) for every x , y E X.
4 1 INTRODUCTORY TOPICS
A normed space is a vector space X which is provided with a norm. The norm
will be denoted by 11 .;X 1 except where simpler notations are introduced.
If r > 0, the set
BJx) = { y E X : Ily-x;X(I <r}
is called the open ball of radius r with center x E X.Any subset A of X is called
open if for every x E A there exists r > 0 such that B,(x) c A. The open sets
thus defined constitute a topology for X with respect to which X is a TVS.
This topology is called the norm topology on X . The closure of E,(x) in this
topology is
-
B,(x) = { y E X:(Iy-x;XII 5 r } .
A TVS X is normable if its topology coincides with the topology induced by
some norm on X.Two different norms on a vector space are equivalent if they
induce the same topology on X,that is, if for some constant c > 0
cIIxII1 5 llxllz 5 (~/c)llxll1
for all x E X, )I and 11 . being the two norms.
[I2
If X and Y are two normed spaces and if there exists a one-to-one linear
operator L mapping X onto Y and having the property 1 L(x); Y 11 = Ilx; X 11
for every x E X , then L is called an isometric isomorphisni between X and Y,
and X and Y are said to be isometrically isomorphic; we write X E Y. Such
spaces are often identified since they have identical structures and differ only
in the nature of their elements.
1.7 A sequence {x,} in a normed space X is convergent to the limit xo if and

only if l i m , ~ m ~ ~ x , - x o=
; X0~in~ R. The norm topology of X is completely
determined by the sequences it renders convergent.
A subset S of a normed space X is said to be dense in X if each x E X is
the limit of a sequence of elements of S. The normed space Xis called separable
if it has a countable dense subset.
1.8 A sequence {x,} in a normed space Xis called a Cauchy sequence if and
= 0. If every Cauchy sequence in X converges
only if limffl,n-,m~~xffl-x,;X~~
to a limit in X ,then Xis complete and a Banach space. Every normed space X
is either a Banach space or a dense subset of a Banach space Y whose norm
satisfies
IIx;YII = IIx;XII forevery x E X.
In this latter case Y is called the completion of X .
1.9 If X is a vector space, a functional (., . ) x defined on X x X is called an

THE NORMED DUAL 5
inner product on X provided that for every x, y, z E X and a, b E 62

-
(9 ( x A x = (u, 4 x 9
(ii) (ax + by, z)x = a ( x , Z)X +b (Y, z)x,
(iii) (x, x ) =~ 0 if and only if x = 0,
where C denotes the complex.conjugate of c E @. Given such an inner product,
a norm on X can be defined by
IIx;XII = (x,x)$’2. (1)
If X is a Banach space under this norm, it is called a Hilbert space. The
parallelogram law
Ilx+v;Xll’+ +
Ilx-y;x112 = 2((x;X112 211Y;q12 (2)
holds in any normed space whose norm is obtained from an inner product
via (1).
The Normed Dual
1.10 A norm on the dual X‘ of a normed space X can be defined by setting,

for x’
E X’,
Since @ is complete, X ’ , with the topology induced by this norm, is a Banach

space (whether or not Xis) and is called the norrned dual of X . If X is infinite
dimensional, the norm topology of X’ is stronger (i.e., has more open sets)
than the weak-star topology defined in Section 1.5.
If X is a Hilbert space, it can be identified with its normed dual X‘ in a
natural way.
1.11 THEOREM (Riesz representation theorem) Let X be a Hilbert

space. A linear functional x’ on X belongs to X’ if and only if there exists
x E X such that for every y E X we have
X’(Y) = 0 9 4 x 7
and in this case 1 1 ~ ’ ; X’I(= IIx; XII. Moreover, x is uniquely determined by

x’ E X‘.
A vector subspace M of a normed space Xis itself a normed space under

the norm of X , and so normed is called a subspace of X. A closed subspace of a
Banach space is a Banach space.
1.12 THEOREM (Hahn-Banach extension theorem) Let Mbe a subspace

of the normed space X. If m’ E M’, there exists x’ E X’ such that Ilx’; X’I(=
“m’;M’II and x’(nt) = m’(rn) for every m E M .
1.13 A natural linear injection of a normed space X into its second dual
space X” = (X’)’ is provided by the mapping J, whose value at x E X is
given by
J,x(x’) = xf(x), x’ E X ’ .
Since IJxx(xf)lIIlx’; X’I/ Ilx;XII we have
lIJ,x;X”II I1lx;XIl.
On the other hand, the Hahn-Banach theorem assures us that for any x E X
we can find an x’ E X’ such that 1 1 ~ ’ ; X‘II = 1 and x’(x) = Ilx; X I / . Hence
IlJxx;X”Il = llx;XIl,
and Jx is an isometric isomorphism from X into X”.
If the range of the isomorphism is the entire space X ” , we say that the
normed space X i s reflexive. A reflexive space must be complete and hence a
Banach space.
1.14 THEOREM Let X be a normed space. X is reflexive if and only if

X’ is reflexive. X is separable if X’ is separable. Hence if X is separable and
reflexive, so is X’.
Weak Topology and Weak Convergence
1.15 The weak topology of a normed space X is the weakest topology on X

that still renders continuous each x’ E X’. Unless X is finite dimensional the
weak topology is weaker than the norm topology on X. It is a consequence of
the Hahn-Banach theorem that a closed, convex set in a normed space is also
closed in the weak topology of that space. A sequence convergent with respect
to the weak topology on X is said to converge weakly. Thus x, converges
weakly to x in X provided x’(x,) + x’(x) in C for every x’ E X’. We denote
norm convergence of a sequence {x,} to x in X by x, + x ; weak convergence
by x,-x. Since Ix’(x,,-x)I I;1 1 ~ ’ ; X’II IIx,,-x; XI1 we see that x, + x implies
x,, - x . The converse is generally not true.
Compact Sets
1.16 A subset A of a normed space X i s called compact if every sequence of

points in A has a subsequence converging in X to an element of A. Compact
UNIFORM CONVEXITY 7
sets are closed and bounded, but closed and bounded sets need not be compact
unless X is finite dimensional. A is called precompact if its closure A (in the
norm topology) is compact. A is called weakly sequentially compact if every
sequence in A has a subsequence converging weakly in X to a point in A. The
reflexivity of a Banach space can be characterized in terms of this property.
1.17 - THEOREM A Banach space Xis reflexive if and only if its closed unit
ball B,(O) = { x E X : Ilx; XI1 5 1) is weakly sequentially compact.
1.18 THEOREM A set A is precompact in a Banach space X if and only

if for every positive number there is a finite subset N, of points of X with the
property
A c u
YEN.
Be(y).
A set Ne with this property is called afinife &-netfor A.
Uniform Convexity
1.19 Any normed space is locally convex with respect to its norm topology.
The norm on X is called uniformly conuex if for every number E satisfying
0 < E I2 there exists a number a(&) > 0 such that if x , y E X satisfy
IIx;XII=IIy;XII= 1 and IIx-y;XII> E , then II(x+y)/2;XI)S1-6(~). The
normed space X is itself called “uniformly convex” in this case. It should,
however, be noted that uniform convexity is a property of the norm-X may
possess an equivalent norm that is not uniformly convex. Any normable
space is called “uniformly convex” if it possesses a uniformly convex norm.
The parallelogram law (2) shows that a Hilbert space is uniformly convex.
1.20 THEOREM A uniformly convex Banach space is reflexive.
The following two theorems will be used to establish the separability,

reflexivity, and uniform convexity of the Sobolev spaces introduced in
Chapter 111.
1.21 THEOREM Let X be a Banach space and M a subspace of X closed

with respect to the norm topology on X. Then M is itself a Banach space under
the norm inherited from X.Furthermore,
(i) M is separable if X is separable,
(ii) M is reflexive if X is reflexive,
(iii) M is uniformly convex if X is uniformly convex.
The completeness, separability, and uniform convexity of M follow easily

from the corresponding properties for X. The reflexivity of M is a consequence
of Theorem 1.17 and the fact that M, being closed and convex, is closed in the
weak topology of X.
1.22 THEOREM For j = 1,2, ...,n let X j be a Banach space with norm
11 1 j . The Cartesian product X = n;=
X j ,consisting of points x = (xl, ...,x,,)
with xj E X i , is a vector space under the definitions
x + y = (x1 +y,, ..., x,,+y,,), cx = (cx1, ..., cx,),
and is a Banach space with respect to any of the equivalent norms
Furthermore,
(i) if X j is separable for 1 I j In, then X is separable;
(ii) if Xiis reflexive for 1 I j In, then X is reflexive;
(iii) if X j is uniformly convex for 1 ~j In, then X is uniformly convex.
More precisely, 1 . is a uniformly convex norm on X provided 1 < p c co.
The reader may verify that the functionals 11 I\('), 1 Ip Ico, are in fact
norms on X and that Xis complete with respect to each of them. Equivalance
of the norms follows from the inequalities
IIxll~oo,4 IIxII(P) IIIXII(1) 5 n IIxll(Oo,*
The separability and uniform convexity of X are readily deduced from the
corresponding properties of the spaces X i , The reflexivity of X follows from
that of X i , 1 Ij In, via Theorem 1.17 or via a natural isomorphism between
X' and n;=, Xi'(see, for example, Lemma 3.7).
Operators and Imbeddings
1.23 Since the topology of a normed space Xis determined by its convergent
sequences, an operatorfdefined on Xinto a topological space Y is continuous
if and only iff(x,,) + f ( x ) in Y whenever x,, + x in X. Such is also the case for
any topological space X whose topology is determined by the sequences it
renders convergent (first countable spaces).
Let X , Y be normed spaces andfan operator from X into Y. The operator
f is called compact iff(A) is precompact in Y whenever A is bounded in X . [A
SPACES OF CONTINUOUS FUNCTIONS 9
bounded set in a normed space is one which is contained in the ball BR(0)for
some R.] f is completely continuousif it is continuous and compact. f is bound-
ed if f(A) is bounded in Y whenever A is bounded in X .
Every compact operator is bounded. Every bounded linear operator is
continuous. Hence every compact linear operator is completely continuous.
1.24 We say that the normed space X is imbedded in the normed space Y,
and write X + Y to designate this imbedding, provided
(i) Xis a vector subspace of Y, and
(ii) the identity operator I defined on X into Y by Zx = x for all x E Xis
continuous.
Since I is linear, (ii) is equivalent to the existence of a constant M such that
IlZx;Y1I I MIIx;XII, xE x.
In some circumstances the requirement that X be a subspace of Y and Z be
the identity map is weakened to allow as imbeddings certain canonical linear
transformations of X into Y. (Examples are trace imbeddings of Sobolev
spaces as well as imbeddings of these spaces into spaces of continuous
functions. See Chapter V.)
We say that X is compactly imbedded in Y if the imbedding operator Z
is compact.
Spaces of Continuous Functions
1.25 Let R be a domain in R". For any nonnegative integer m let C"(R) be
the vector space consisting of all functions Q, which, together with all their
partial derivatives D"4 of orders la1 Im, are continuous on R. We abbreviate
Co(Q) s C(Q). Let C"(R) = n;,o C"(R). The subspaces C,(Q) and
COm(R) consist of all those functions in C(Q) and Cm(Q), respectively, which
have compact support in R.
1.26 Since R is open, functions in C"(R) need not be bounded on R. If

Q, E C(Q) is bounded and uniformly continuous on R, then it possesses a
n
unique, bounded, continuous extension to the closure of R. Accordingly,
we define the vector space Cm(n)to consist of all those functions Q, E Cm(Q)
for which DaQ, is bounded and uniformly continuous on R for 0 5 la1 5 m.
[Note that C m ( E )# Cm(Rn).]C"(Q is a Banach space with norm given by
max SUP I
I/ 4; C" (n)I = 0slalr;rn xsn
w(4I.
1.27 If 0 < A S 1, we define CmiA(n) to be the subspace of Cm@)consisting

of those functions 4 for which, for 0 5 la1 s m, D"4 satisfies in R a Htilder
condition of exponent A, that is, there exists a constant K such that
Io"4(x)-o"qi(Y)l s KIx-ylA, X,Y E a.
is a Banach space with norm given by
Cm*A(Q)
It should be noted that for 0 < v < L s 1,

c m * A ( n ) E cm*'(n)5 cyn).
It is also clear that C"*'(n) q! Cm+l(a). In general Cm+l(n)# C"*'(n)
either, but the inclusion is possible for some domains Q, for instance convex
ones as can be seen by appealing to the mean value theorem (see Theorem
1.31).
If R is bounded, the following two well-known thearems provide useful
criteria for denseness and compactness of subsets of C(n).If 4 E C(n),we
may regard 4 as defined on a, that is, we identify 4 with its unique continuous
extension to the closure of GI.
1.28 THEOREM (Stone-Weierstruss theorem) Let R be a bounded

domain in R". A subset d of C(n)is dense in C(n)if it has the following four
properties :
(i) +,
If 4,+ E d and c E @, then 4 + 49, and cqi all belong to d .
(ii) If qi E d , then 6 E d,where 6 is the complex conjugate of 4.
(iii) If x,y E a, x # y, there exists 4 E d such that qi(x) # +(y).
(iv) If x E a, there exists qi E JXIsuch that 4(x) # 0.
1.29 COROLLARY If R is bounded in R", then the set P of all polynomials
in x = (xl, ...,x,,) having rational-complex coefficients is dense in C(Q.
(c E C is rational complex if c = c1 + ic2,where c1 and c2 are rational numbers.)
Hence C(n) is separable.
PROOF The set of all polynomials in x is dense in C(n) by the Stone-
Weierstrass theorem. Any polynomial can be uniformly approximated on the
compact set ll by elements of the countable set P,which is therefore also dense
in C@). I
1.30 THEOREM (Ascoli-Arzelu theorem) Let R be a bounded domain
in R". A subset K of C(n)is precompact in C(n) providing the following two
SPACES OF CONTINUOUS FUNCTIONS 11
conditions hold :
(i) There exists a constant M such that for every 9 E K and x E R,
IdJ(x)lSM.
(ii) For every e > 0 there exists S > 0 such that if 9 E K,x , y E R, and
Ix-ul< 4 then I9(X)-dJ(Y)l< 8.
The following is a straightforward imbedding theorem for the spaces

introduced above.
1.31 THEOREM Let m be a nonnegative integer and let 0 < v < 15 1.

Then the following imbeddings exist :
C"+'(iCz) -P C"(iI), (3)
."ql) -b C"(iCz), (4)
,"*A@) -P C"*V(rI). (5)
If R is bounded, then imbeddings (4) and (5) are compact. If R is convex, we
have the further imbeddings
C"+l(Sz) + c",l(n), (6)
C"+'(iCz) -P C"*V(lI). (7)
If R is convex and bounded, then imbeddings (3) and (7) are compact.
PROOF The existence of imbeddings (3) and (4) follows from the obvious
inequalities
119; C"(QII 5 I19;cm+'(Qll,
119; C"l(W 5 119; C " V ) l l .
To establish ( 5 ) we note that for la1 5 m,
10"9(~)-~"9(Y)l sup P"9(4- p 9 (v)I
SUP
x.Y€n Ix-Yl' x,yen Ix-ylA
o< Ix-yl<l
and
from which we conclude that

119; c"3v(mII
5 3 119;C"W)II.
If R is convex and x , y E R, then by the mean value theorem there is a
point z E R on the line segment joining x and y such that Da+ ( x ) - o"9 ( y ) =
Dzu,...,D,u).Thus
(x-y).VDuq5(z), where Vu = (D,u,
IDa4(x)-Da4(Y)l 5 nIx-YlI14;Cm+'cn~II,
and so
ll4;Crn*'(QIl5 nl14;Crn+'(fi)lI.
Thus (6) is proved and (7) follows from ( 5 ) and (6).
Now suppose that R is bounded. If A is a bounded set in Co"(fi), then
there exists A4 such that I14;Copa((32)111 M for all 4 E A . But then
I+(x)-(P(y)I<Mlx-yla for all ~ E and A all x , y ~ R whence
, A is pre-
compact in C(n) by Theorem 1.30. This proves the compactness of (4) for
m = 0. If m 2 1 and A is bounded in C"*a(n),then A is bounded in Coia((iz)
and there is a sequence {4j}c A such that 4j + 4 in C(n). But {D14j} is
also bounded in Coia(n) so there exists a subsequence of {4j}which we again
denote by { t $ j } such that Dlcjj + $, in C@). Convergence in C(n) being
uniform convergence on R, we have = D14. We may continue to extract
subsequences in this manner until we obtain one for which Da4j+ Da$ in
C ( 0 ) for each 01,0I1011 I m. This proves the compactness of (4). For ( 5 ) we
argue as follows :
I const I Da+ (x) -Da4(JJ)I -v'a (9)

for all 4 in a bounded subset of Crnia(<n). Since (9) shows that any sequence
bounded in C". a (a) and converging in C" (a)also converges in Crnp " (n),
the
compactness of ( 5 ) follows from that of (4).
Finally, if R is convex and bounded, the compactness of (3) and (7) follows
from composing the continuous imbedding (6) with the compact imbeddings
(4) and ( 5 ) for the case I = 1. I
The existence of imbeddings (6) and (7), as well as the compactness of (3)
and (7), can be obtained under less restrictive hypotheses than the convexity
of R. For instance, if every pair of points x , y E R can be joined by a rectifiable
arc in R having length not exceeding some fixed multiple of lx-yl, then we
can obtain an inequality similar to (8) and carry out the proof. We leave it
to the reader to show that (6) is not compact.
The Lebesgue Measure in R"
1.32 Many of the spaces of functions considered in this monograph consist

of functions integrable in the Lebesgue sense over domains in R". While we
THE LEBESGUE MEASURE IN R" 13
assume that most readers are familiar with Lebesgue measure and integration,
we nevertheless include here a brief discussion of that theory, especially those
aspects of it relevant to the study of the LP-spaces and Sobolev spaces con-
sidered hereafter. All proofs are omitted. For a more complete and systematic
discussion of the Lebesgue theory, as well as more general measures and
integrals, the reader is referred to any of the standard works on integration
theory, for example, the book by Munroe [48].
1.33 A collection C of subsets of R" is called a a-algebra if the following

conditions hold :
(i) R" E Z.
(ii) I f A E Z , t h e n # = { x E R " : x $ A } E C .
(iii) If A j E C, j = 1,2, ..., then ur=, A j E C.
It follows from (i)-(iii) that:
(iv) 0E X .
(v) I f A , ~ C , j = 1 , 2,..., then n 7 = J j E C .
(vi) IfA,BEC, then A-B=A n B'EC.
1.34 By a measure p on C we mean a function on Z taking values in either

+
R u { a}(a positive measure) or C (a complex measure) which is countably
additive in the sense that
whenever A j E Z, j = 1,2, ..., and A j n A, = 0 for j # k. (For a complex

measure the series on the right, being convergent for any such sequence { A j } ,
is absolutely convergent.) If p is a positive measure and if A, B E Z and A c B,
then p ( A ) s p(B). Also, if A j E Z, j = 1,2, ... and A, c A, c then
..a,
p ( U ? = , A j ) = limj-,mp(Aj).
1.35 THEOREM There exists a a-algebra C of subsets of R" and a positive

measure ,u on Z having the following properties:
(i) Every open set in R" belongs to C.
(ii) If A c B, B E Z, and p ( B ) = 0, then A E C and p ( A ) = 0.
(iii) If A = { x ~ ( W " : a / < x ~ 1 b ~ , J ,...,
= 1 ,n2} , then A E C and p ( A ) =
n;= 1 (bj - aj).
(iv) p is translation invariant, that is, if x E R" and A E C, then x + A =
{ x + y : y ~ A~}X a n d p ( x + A ) = p ( A ) .
The elements of C are called (Lebesgue) measurable subsets of R"; p is

called the (Lebesgue) measure in R". (We shall normally suppress the word
"Lebesgue" in these terms as it is the only measure on R" which we shall
require for our purposes.) For A E Z we call p ( A ) the measure of A or the
volume of A, since Lebesgue measure is a natural generalization of the concept
of volume in R3. While we make no formal distinction between "measure"
and "volume" we shall often prefer the latter term for sets that are easily
visualized geometrically (balls, cubes, domains) and shall write vol A in
place of p ( A ) in these cases. In R' and Rz the terms length and urea are more
appropriate than volume.
1.36 If B c A c R" and p ( B ) = 0, then any condition that holds at every

point of the set A - B is said to hold almost everywhere (a.e.) in A . It is easily
seen that every countable set in R" has measure zero. The converse is, however,
not true.
A function f defined on a measurable set and having values in R u
{ + 00, - co} is itself called measurable if the set
{x :f(x) > a>
is measurable for every real a. Some of the more important aspects of this
definition are listed in the following theorem.
1.37 THEOREM (a) Iffis measurable, so is I f l .

(b) Iff and g are measurable and real valued, then so aref+g andfg.
(c) If {f.} is a sequence of measurable functions, then supn&, inf,,f.,
lim SUP^,^ f., and lim inf,,, f.are measurable.
(d) Iffis continuous and defined on a measurable set, thenfis measurable.
(e) Iff is continuous on R into R and g is measurable and real valved,
then the composite function f o g defined by f0 g(x) = f ( g ( x ) ) is measurable.
(f) (Lusin's theorem) Iff is measurable and f ( x ) = 0 for x E A' where
p ( A ) < 00, and if E > 0, then there exists a function g E C,(A) such that
suP,,rtnIg(x)I 5 suPx.anlf(x)l and ~1 { x E R" :f(x)f g ( x ) ) < e-
1.38 Let A c W". The function x,, defined by
1 if ~ E A
is called the characteristic function of A. A real-valued function s on R" is

called a simple function if its range is a finite set of real numbers. If for every x ,
zys,
s ( x ) E { a l ,...,a,,,},then clearly s = a, x,,, ,where A, = { x E W" :-s(x) = a,},
and s is measurable if and only if A A?, ..., A,,, are measurable. Because of the
following approximation theorem simple functions are a very useful tool in
integration theory.
THE LEBESGUE INTEGRAL 15
1.39 THEOREM Given a real-valued function f with domain A c [w"

there is a sequence {s,} of simple functions converging pointwise t o f ' o n A .
Iffis bounded, {s,,} may be chosen so that the convergence is uniform. Iffis
measurable, each s, may be chosen measurable. Iffis nonnegative valued, the
sequence {s,} may be chosen to be monotonically increasing at each point.
The Lebesgue Integral
1.40 We are now in a position to define the (Lebesgue) integral of a measur-

able, real-valued function defined on a measurable set A c R". For a simple
function s = Cy= ajxA,, where A j c A , A j is measurable, we define
Iff is measurable and nonnegative valued, we define
J:f ( x ) dx = SUP l s ( x ) dx,
the supremum being taken over measurable simple functions s vanishing

) A . I f f is a nonnegative simple
outside A and satisfying 0 I s(x) ~ f ( x in
function, then the two definitions of SAf(x) dx given by (10) and (1 I ) coincide.
Note that the integral of a nonnegative function may be co. +
Iffis measurable and real valued, we setf =.I.+ -f-,wheref = max (f;0)+
andf - = - minu, 0) are both measurable and nonnegative. We define
J-pdx = J:f'(X)dx-J:f-(X)dx
provided at least one of the integrals on the right is finite. If both integrals are
finite, we say that f is (Lebesgue) integrable on A . The class of integrable
functions on A is denoted L' ( A ) .
1.41 THEOREM Assume all of the functions and sets appearing below
are measurable.
(a) Iffis bounded on A and p ( A ) < co, thenfE L'(A).
(b) If a I f ( x ) 5 b for all x E A and if p ( A ) < co, then
W(A) J:mdx I b(A).
(c) Iff(x) _< g ( x ) for all x E A, and if both integrals exist, then
(d) IfJg E L'(A), thenf+g E L'(A) and
l c f f g ) ( x ) dx = J)X) dx +jp dx.
(e) I f f € L' ( A ) and c E R, then c f ~ L' ( A ) and
(4)(x) dx = c
I f(~)
dx.
(f) I f f € L' ( A ) , then If1 E L' ( A ) and
(g) I f f € L'(A) and B c A , thenfe L'(B); if in additionf(x) 2 0 for all

x E A , then
1f ( 4B dx s p ( x ) dx.
(h) If p ( A ) = 0, then f A f ( x )dx = 0.

(i) If f E L ' ( A ) and if f B f ( x ) d x= 0 for every B c A , then f ( x ) = 0
a.e. on A .
1.42 THEOREM Iff is either an element of L' (Rn) or measurable and

nonnegative on R", then the set function 1 defined by
A(A) =
s1f ( ~ dx
)
is countably additive, and hence a measure, on the a-algebra of Lebesgue

measurable subsets of R".
One consequence of this additivity of the integral is that sets of measure

zero may be ignored for purposes of integration, that is, iffand g are nieasur-
able on A and if f(x) = g ( x ) a.e. on A , then f A f ( x )dx = f A g ( x ) dx.
Accordingly, two elements of L' ( A ) are considered identical if they are equal
almost everywhere.
The following three theorems are concerned with the interchange of
integration and limit processes.
1.43 THEOREM (Monotone convergence theorem) Let A t R" be

measurable, and {f,} a sequence of measurable functions satisfying
0 s f 1 ( x )s f 2 ( x )I ... for every x E A . Then
lim
n-+m
I f , ( x ) dx = /A
(lim f . ( x ) ) dx.
n+w
THE LEBESGUE INTEGRAL 17
1.44 THEOREM (Fatou's lemma) Let A c R" be measurable and let

{ f , }be a sequence of nonnegative, measurable functions. Then
1.45 THEOREM (Dominated convergence theorem) Let A c R" be

measurable and let {f,} be a sequence of measurable functions converging to
a limit pointwise on A . If there is a function g E L'(A) such that If,(x)I I g(x)
for every n and all x E A , then
lirn JA f,(x) dx = (lim S, (x)) dx.

n- m A n-rm
1.46 The integral of a complex-valued function over a measurable set

+
A c R" is defined as follows. Set f = u iu, where u and v are real valued and
callfmeasurable if and only if u and u are measurable. We shall say that f is
integrable over A , and writefe L' ( A ) , provided I f 1 = (u2 +u')'/~ belongs to
L' ( A ) in the sense described in Section 1.40. Forf'E L' ( A ) ,and only for suchf,
the integral is defined by
/A
f ( x ) dx = I.(.) + dx i L v ( x ) dx.
It is easily checked that , f ~ L'(A) if and only if u,v E L'(A). Theorem

1.37(a,b,d-f), Theorem 1.41(a,d-i), Theorem 1.42 [assuming f~ L'(R")],
and Theorem 1.45 all extend to cover the case of complexf.
The following theorem enables us to express certain complex measures in

terms of Lebesgue measure p . It is the converse of Theorem 1.42.
1.47 THEOREM (The Radon-Nikodym theorem) Let A be a complex

measure defined on the a-algebra C of Lebesgue measurable subsets of R".
Suppose that ] . ( A ) = 0 for every A E C for which p ( A ) = 0. Then there exists
f E L' (R") such that for every A E C
A ( A )=
I f ( x ) dx.
The functionfis uniquely determined by I. up to sets of measure zero.
1.48 Iff is a function defined on a subset A of R"'"', we may regard f as

depending on the pair of variables ( x , y ) with x E R" and y E R". The integral
off over A is then commonly denoted by
Lf(X,Y)dx dY
or, if it is desired to have the integral extend over all of Rn+m,
where zA is the characteristic function of A . In particular, if A c R", we may

write
I f ( x ) dx = lf(xl, ...,x,) dxl ... dx,.
1.49 THEOREM (Fubini's theorem) Let f be a measurable function on

R"+mand suppose that at least one of the integrals
Distributions and Weak Derivatives
1.50 We shall require in subsequent chapters some of the basic concepts and
techniques of the Schwartz theory of distributions [60], and we present here
a brief description of those aspects of the theory that are relevant for our
purposes. Of special importance is the notion of weak or distributional
derivative of an integrable function. One of the standard definitions of Sobolev
spaces is phrased in terms of such derivatives (Section 3.1). In addition to
Ref. [60] the reader is referred to Rudin [59] and Yosida [69] for more
complete treatments of the spaces 9(n)and W(R) introduced below, as well
as useful generalizations of these spaces.
DISTRIBUTIONS AND WEAK DERIVATIVES 19
1.51 Let R be a domain in R". A sequence {+"} of functions belonging to

Corn@)is said to converge in the sense of the space 9 ( R ) to the function
4 E Corn(R)provided the following conditions are satisfied:
(i) there exists K c c R such that supp(4, - 4) c K for every n, and
(ii) limn+rnD " ~ , ( x=) D " ~ ( xuniformly
) on K for each multi-index a.
There exists a locally convex topology on the vector space Corn(R)with
respect to which a linear functional Tis continuous if and only if T(4,) -+ T ( 4 )
in C whenever 4" -+ 4 in the sense of the space 9 ( R ) . This TVS is called 9 ( R )
and its elements testing functions. 9 ( R ) is not a normable space. (We ignore
the question of uniqueness of the topology asserted above. It uniquely
determines the dual of 9 ( R ) which is sufficient for our purposes.)
1.52 The dual space 9'(R) of 9 ( R ) is called the space of (Schwartz) distri-
butions. 9'(R) is given the weak-star topology as dual of 9(R), and is a locally
convex TVS with that topology. We summarize the vector space and con-
vergence operations in 9'(R) as follows: if S, T, T. E W(R) and c E C, then
(S+ n(4)= S(4)+ T(4h 4 E 9(Q),
(cT)(4)= cT(4), 4 E 2(R),
T, -+ Tin 2'@)if and only if Tn(4) -+ T ( 4 ) in C for every 4 E 9(R).
1.53 A function u defined almost everywhere on R is said to be locally

integrable on R provided u E L' ( A ) for every measurable A cc R. In this
case we write u E L;oc(S2). Corresponding to every u E L:oc(a) there is a
distribution TuE 9 ( R ) defined by
It is clear that Tu,thus defined, is a linear functional on 9(R). To see that it is

continuous suppose that +,, -+ 4 in 9 ( R ) . Then there exists K cc R such
that supp(4,,- 4) c K for n = 1,2,3, ... . Thus
lL(4n)-T,(+)/ I supI$n(X)-4(x)I
XEK sh
The right side of the inequality above tends to zero as n
lu(x)l dx.
-+ co since +,, + +
uniformly on K .
1.54 Not every distribution T E9'(R) is of the form T = T, [defined by (13)]

for some u E L,'oc(Q).Indeed, assuming 0 E R,the reader may wish to convince
himself that there can be no locally integrable function 6 on Q such that for
every 4 E 9 ( Q )
The linear functional 6 defined on 9 (0)

by
6(+) = 4(0)
is, however, easily seen to be continuous, and hence a distribution on Q.
1.55 Let U E C'(Q) and 4 EQ((R). Since 4 vanishes identically outside

some compact subset of 0, we obtain by integration by parts in the variable x j
Similarly, integration by parts IQI times leads to
if u E Cl"l(Q).This motivates the following definition of the derivative D"T

of a distribution T E gf(0):
(D"T)(4) = (- l)'"'T(D"+). (1 5 )
Since 0'4 E 9 (0) whenever 4 E 9(a),D"T is a functional on 9 (Q). Clearly
D"Tis linear on 9(Q). We show that D"Tis continuous, and hence a distribu-
tion on R. To this end suppose 4, + 4 in 9(Q). Then
SUPP(D"(4,-4)) = SUPP(4,,-4) = K
for some K c c R. Moreover,
DPCD" (4%- 4)l = DP+ r - 4)
converges to zero uniformly on K as n -+ co for each multi-index p. Hence
D"4. -i0'4 in 9 (Q). Since T E 9'(Q)
it follows that
D"T(4,) = (- I > ' " ' T ( D " ~ ,+
) ( - l ) 1 a ' T ( D a 4=) D " T ( 4 )
in @. Thus D " T Egf(f2).
We have shown that every distribution in W ( Q )possesses derivatives of
arbitrary orders in W ( Q )in the sense of the definition ( 1 5). Furthermore, the
mapping D" from 9'(Q) into W ( Q )is continuous. If + T in 9'(Q) and if
4 E 9 (Q), then
D"T,(f$) = ( - 1)'"'T,(D"4)+ (- l)'"IT(D"f#))= D " T ( 4 ) .
DISTRIBUTIONS AND WEAK DERIVATIVES 21
1.56 EXAMPLES (1) If 0 E R and 6 E W(R) is defined by (14), then D"6

is given by
D"6 (4) = (- 1)'"'Da4(0).
(2) If R = R (;.eat n = 1) and H E L;,,([w) is defined by
1 if x 2 0
H(x) =
0 if x < 0,
then the derivative (TI,)'is 6, for if 4 ~ 9 ( R has
) support in the interval
[ - a , a ] , then
(TH)'(4)= -TH(4') = -S'@(x) dx = $(O) = S(4).

0
1.57 We now define the concept of weak derivative of a locally integrable

function. Let u E L;,,(R). There may or may not exist a function L), E L~,,(Q)
such that T,, = D"(T,) in 9'(!2).If such a v, exists, it is unique up to sets of
measure zero and it is called the weak or distributional partial derivative of u
and is denoted by D"u. Thus D"u = u, in the weak (distributional) sense
provided v, E Lkc (R) satisfies
for every
J)(x)D"$(x)
4 E 9(R).
dx = (- 1)1"1
I*
L),(x)@(x)dx
If u is sufficiently smooth to have a continuous partial derivative D"u in the

usual (classical) sense, then D"u is also a distributional partial derivative of u.
Of course D"u may exist in the distributional sense without existing in the
classical sense. For example a function u, continuous on R, which has a
bounded derivative u' except at finitely many points, has a derivative in the
distributional sense. We shall show in Theorem 3.16 that functions having
weak derivatives can be suitably approximated by smooth functions.
1.58 Let US note in conclusion that distributions on R can be multiplied by

smooth functions. If T E 9'(Q) and o E Ca'(Q),the product W T Eg'(Q) is
defined by
(wT)(+)= T ( w 4 ) , 4 E gm.
If T = 7;, for some u E L;,,(Q), then W T= T,,,,.The Leibniz rule (see Section
1.1) is easily checked to hold for D'(wT).
The Spaces LP(a)
Definition and Basic Properties
2.1 Let R be a domain in R” and let p be a positive real number. We denote

by Lp(R) the class of all measurable functions u, defined on R, for which
We identify in LP(R) functions that are equal almost everywhere on R. The

elements of LP(R)are thus actually equivalence classes of measurable functions
satisfying (l), two functions being equivalent if they are equal a.e. in R. For
convenience, however, we ignore this distinction and write u E Lp(R) if u
satisfies ( l ) , and u = 0 in Lp(R)if u ( x ) = 0 a.e. in R. It is clear that if u E LP(R)
and c E @, then cu E LP(R). Moreover, if u, u E Lp(R), then since
+ WIPI(I.(x>l+ Iu(x>I)~I2P(Iu(x)IP+ lo(x>l”9
u+v E Lp(R), so LP(R) is a vector space.
2.2 We shall verify presently that the functional 11 . \ I p defined by
is a norm on LP(R) provided 1 Ip < co. (It is not a norm if 0 < p < I.) In
22
DEFINITION AND BASIC PROPERTIES 23
arguments where confusion of domains might occur we shall use I(. in

place of 11. IJp. It is clear that lluilp 2 0 and equality occurs if and only if u = 0
in LP(R). Moreover,
Ilcullp = IClIl~llP~ cE @.
It remains to be shown, then, that if 1 I p c co,
IIu+ullp llullp + I I u l l P ? (2)

which is known as Minkowski’s inequality. Condition ( 2 ) certainly holds for
p = 1 since
If 1 < p < co,we denote byp’ the number p / ( p - 1) so that 1 c pr c co and

(1/P) + ( l / P ’ ) = 1.
pr is called the exponent conjugate t o p ,
2.3 THEOREM (Hiilder’s inequality) If 1 < p c co and u E LP(R),

2) E LP’(R), then uu E L’ (R) and
+
PROOF The function f ( r ) = ( t p / p ) (l/p’)- t has, for r 2 0, the minimum
value zero, and this minimum is attained only at t = 1. Setting t = ab-p”p,
we conclude, for nonnegative numbers a and b, that
+
ab I (aP/p) (bP‘/pr) (4)
with equality occurring if and only if up = bP’. If either llullp = 0 or
(lullp, = 0, then u ( x ) u ( x )= 0 a.e. in R so (3) is satisfied. Otherwise we obtain
(3) by setting a = l u ( x ) l / l ~ u ~and
l , b = lo(x)I/IIu/1,,,in (4)and integrating over
R. Equality occurs in (3) if and only if lu(x)lp and lu(x)lp’are proportional
a.e. in R. I
We remark that a form of Holder’s inequality for finite or infinite sums,

c lakbkl I (C IaklPll’p(CI b k l P ’ Y P ’ 9
can be proved in the same manner.
2.4 THEOREM (Minkowski’s inequality) If 1 5 p < co,then

IIIJ+UllP l l 4 P + Ilullp. (5)
24 11 THE SPACES LP(n)
PROOF We have already done the case in which p = 1 so we assume

1 < p < co. We may also assume that u,u E Lp(R), for otherwise the right
side of ( 5 ) is infinite. Now
L n
by separate applications of Holder's inequality. Inequality ( 5 ) follows by

cancellation, which is valid since IIu+ullp < 03. I
2.5 A function u, measurable on R, is said to be essentially bounded on $2

provided there exists a constant K for which Iu(x))< K a.e. on R. The
greatest lower bound of such constants K is called the essential supremum of
IuI on R and is denoted by ess ~ u p , . , ~ l u ( x )We
l . denote by L"(R) the vector
space consisting of all functions u that are essentially bounded on R, functions
being once again identified if they are equal a.e. on R. It is easily verified that
the functional 1 Ilm defined by
IlUll" = ess SUP lu(x)l

X€R
is a norm on Lm(R).Moreover, Holder's inequality (3) clearly extends to cover

the two cases p = 1, p' = 00, and p = co,p' = 1.
The following pair of theorems establishes reverse forms of Holder's and

Minkowski's inequalities for the case 0 < p < 1. The latter inequality will be
used later in establishing the uniform convexity of certain LP-spaces.
2.6 THEOREM Let 0 < p < 1 so that p' = p / ( p - 1) < 0. Suppose

f E Lp(R) and
0 < /nlg(x)lp'dx < co.
Then
PROOF We may assumefg E L' (R); otherwise the left side of (6) is infinite.
Set 4 = l g l - p and $=IfgIp so that 4$=IflP. Then $ E L ~ ( R ) where
q = I/p > 1, and since p' = -pq' where q' = q/(q- 1) we have 4 E Lq'(R). By
DEFINITION AND BASIC PROPERTIES 25
the direct version of Holder’s inequality (3) we have
j*If(x)lPdx = J - * + ( x ) W d x 5 ll~llqll+llq~
Taking pth roots and dividing by the last factor on the right side we obtain
(6). I
2.7 THEOREM Let 0 < p < 1. If u, u E LP(Q),then
PROOF If u = u = 0 in LP(R), then (7) is trivial. Otherwise the left-hand side

is greater than zero. Applying the reverse Holder’s inequality ( 6 ) , we obtain
and (7) follows by cancellation. I

The following theorem gives a useful imbedding result for LP-spaces over
domains with finite volume, and some consequences of this imbedding.
2.8 THEOREM Suppose volR =Sol dx < co and 1 5 p I y I co. If

u E Lq(n), then u E LP(R) and
/lullp I (voln)(”P)-(”q) l l ~ l l q . (8)
Hence
Lq(R) + LP(Q). (9)
If u E L“(R), then
lim IIullp = II4Im. (10)
P- m
Finally, if u E LP(Q)for 1 I p c co and if there is a constant K such that for

all such p
llullp 5 K? (1 1)
26 11 THE SPACES Lp(R)
then u E Lm(R) and

llullm K.
PROOF If p = q, (8) and (9) are trivial. If 1 I p -= q I co and u E Lq(R),

Holder's inequality gives
from which (8) and (9) follow immediately. If u E Lm(R),we obtain from (8)
On the other hand, for any E > 0 there exists a set A c R having positive
measure p (A) such that
lu(x)l 2 llullm- E if x E A.
Hence
Equation (10) now follows from (13) and (14).

Now suppose (1 1) holds for 1 I p < co. If u $ La(R)or else if (12) does
not hold, then we can find a constant K, > K and a set A c R with p ( A ) > 0
such that for x E A, l u ( x ) l 2 K,. The same argument used to obtain (14) now
shows that
lim inf llullp 2 K , ,
p-m
which contradicts (1 1). I

2.9 COROLLARY LP(R) c Li,,(R) for 1 I p I co and any domain R.
Completeness of LP(R)
2.10 THEOREM LP(R) is a Banach space if 1 I p I co.
PROOF First assume 1 I p < co and let {u,} be a Cauchy sequence in LP(Q).
COMPLETENESS OF Lp(a) 27
There is a subsequence {u,,} of {u,} such that

I1/2j,
J~u,,+,-u,,~~p j = 1,2,....
Let u,(x) = xj!l Then
Junj+,(x)-uu,,(x)~.
5
JI~,IJ~5 j = 1 (1/2j) < I , m = 1,2,....
Putting u ( x ) = limm,,um(x), which may be infinite for some x, we obtain by

Fatou's lemma 1.44
Jnlv(x)lpdx 5 lim inf

m+m
Hence u ( x ) < co a.e. in f2 and the series
converges to a limit u ( x ) a.e. in R. Let u ( x ) = 0 whenever it is undefined as the

limit of (1 5). Since (1 5 ) telescopes we have
lim unm(x)= u ( x ) a.e. in R.
m-tm
For any E > 0 there exists Nsuch that ifm,n 2 N, then IIu,,,-unllp < E . Hence,
by Fatou's lemma again
5 lim infJ~Iu~,(x)-u,(x)lpdx
j + 00 EP
if n > M . Thus u = ( u - u , ) + u , ~ L ~ ( R ) and JJu-uu,JJP-'Oas n - t c o .

Therefore Lp(R) is complete.
Finally, if {u,} is a Cauchy sequence in Lm(R),then there exists a set A c R
having measure zero such that if x 4 A , then for every n, m = 1,2, ...
-
lun(x)l 5 ll~nllm~ lun(x>--u,(x)I II~n-umllm*
Since { ~ ~ uis bounded

, ~ ~ ~in R, ) u, converges uniformly on R A to a bounded
function u. Setting u ( x ) = 0 for x E A , we have u E Lm(R)and JJu,-uJJ, + 0
as n + 00. Thus Lm(R)is complete. I
2.11 COROLLARY If 1 5 p I co, a Cauchy sequence in Lp(R) has a

subsequence converging pointwise almost everywhere on R.
2.12 COROLLARY L2(R) is a Hilbert space with respect to the inner

product
(u,u) = J~R u(x)v(x)dx.
Holder’s inequality for L2 (a) is actually just the well-known Schwarz
inequality
I(u,u)l 5 llullz IIuIIz.
Approximation by Continuous Functions, Separability
2.13 THEOREM C,(R) is dense in LP(R) if 1 I p < co.
PROOF Let u E LP(R) and let E > 0. We show that there exists a function
4 E C,(R) such that < E . Setting u = u,-u,+i(u,-u,) where each
u j , 1 I j I 4, is real valued and nonnegative, we find 4j E Co(R) such that
l/4j-ujllp < ~ / 4 ,1 < j 1 4 . Then ilu-41+42-i(43-44)llp<E . We assume
without loss of generality, therefore, that u is real valued and nonnegative. By
Theorem 1.39 there exists a monotonically increasing sequence {s,} of non-
negative simple functions converging pointwise to u on R. Since 0 I s,(x) I
) ~( ~ ( x ) ) ”we have s, + u in
u ( x ) we have s, E LP(R). Since ( U ( X ) - ~ , ( X )I
LP(R) by the dominated convergence theorem 1.45. We may thus pick
s E {s,} such that IIu-sllp < 42. Since s is simple and p < 00 the support of s
must have finite volume. We may also assume that s(x) = 0 for all x E Rc.
Applying Lush’s theorem 1.37(f) we obtain a function 4 E C,(R) such that
I ~ ~ s ~for~allm x E R,
I ~ ( X )I
and
vol{x E n : s(x) # 4(x)} < (44 IIsII,)P.
Hence by Theorem 2.8 we have
lls-411p I Ils-411a,(vol{x E Q : s(x) # 4(x)})1’p
< 2 llslla3(E/4 Ilsl,) = 42.
It follows that ~ ~ < E.u - ~ ~ ~ ~
2.14 The above proof shows that in fact the set of simple functions in LP(Q)
is dense in LP(R) for 1 I p < co. That this is also true for Lm(Q)is a direct
consequence of Theorem 1.39.
2.15 THEOREM Lp(R) is separable if 1 I y < co.
PROOF For m = 1,2, ... let

MOLLIFIERS, APPROXIMATION BY SMOOTH FUNCTIONS 29
Thus Ti, is a compact subset of 0. Let P be the set of all polynomials on R"
having rational-complex coefficients. Let P, = {xn,f:J'~ P} where xa, is the
characteristic function of a,. By Corollary 1.29, P, is dense in C(Ti,). More-
over, u;=, P, is countable.
If u E LP(0)and E > 0, there exists 4 E Co(Q)such that ~ I U - < ~4 2~ . ~ ~
If l/m < dist(supp4, bdryn)), there exists J'E P, such that ~ ~ ~ < - j ' ~ ~ m
(c/2)(vola,)- ' I p . It follows that
II4-fIIp 5 II~-fIIw(vol~m)"p< &/2

and so ~ ~ u - <f E~. Thus
~ p the countable set uz,l P, is dense in LP(R)and
Lp(Q)is separable.
2.16 C(Q), being a proper closed subspace of La@), is not dense in that
space. Thus neither is Co(Q)nor Com(Q),and L"(Q) is not separable.
Mollifiers, Approximation by Smooth Functions
2.17 Let J be a nonnegative, real-valued function belonging to Com(R")

and having the properties
(i) J ( x ) = 0 if 1x1 2 1, and
(ii) J a n J ( x )dx = 1.
For example, we may take
k exp[-1/(1- 1xI2)] if 1x1 < 1
J(x) =
if 1x1 2 1,
where k > 0 is so chosen that condition (ii) is satisfied. If E > 0, the function
J , ( x ) = E - " J ( x / Eis) nonnegative, belongs to Com(R"),and satisfies
(i) J,(x) = 0 if 1x1 2 8, and
(ii) Joan J, ( x ) dx = 1.
J, is called a mollijier, and the convolution
defined for functions u for which the right side of (16) makes sense, is called a
mollijication or regularization o f u. We summarize some properties of
mollification in the following lemma.
2.18 LEMMA Let u be a function which is defined on R" and vanishes

identically outside the domain Q.
30 11 THE SPACES LP(R)
(a) If u E ,!,iOc(Q), then J, * u E C"(R").

(b) If also supp u c c R, then J, * u E C,"(Q) provided
E < dist (supp u, bdry 0).
(c) If U E L P ( Q ) where 1 ~p < 00, then J , * u E LP(Q). Moreover,
(d) If u E C(R) and G c c Q, then J, u ( x ) = u ( x ) uniformly on G .

(e) If u E C@), then lim,,,, J,u(x) = u ( x ) uniformly on R.
PROOF Since J , ( x - y ) is an infinitely differentiable function of x and

vanishes if Iy-xI 2 E , and since for every multi-index CI and every function u
that is integrable on compact sets in R" we have
r
it follows that conclusions (a) and (b) are valid.

Suppose u E LP(R). If 1 0. By Theorem 2.13 there exists t$ E C,(Q)such that ~ ~ u <-q/3.

~ ~ ~ p
Thus by (18), llJ,*~-J,*$11,, < q/3. NOW
5 SUP I$(Y)-4(X)l. (19)

ly-XI < e
Since 4 is uniformly continuous on R the right side of (19) tends to zero as
PRECOMPACT SETS IN LP(n) 31
E + O + . Since supp4 is compact we may therefore arrange to have

liJ, * 4-41/,, < r]/3 by choosing E sufficiently small. For such E we therefore
have IIJ,*u-uIIp < r] and (c) follows. l f p = 1, (18) follows directly from (16)
without use of Holder's inequality, and the rest of the proof of (c) is the same
as above. The proofs of (d) and (e) may be obtained by replacing 4 by u in
(19). I
2.19 THEOREM C,"(R) is dense in LP(R) if 1 I p < cx).
The proof is an immediate consequence of Theorem 2.13 and Lemma

2.18 (b, e).
Precompact Sets in E ( R )
2.20 The following theorem plays a role in the study of LP-spaces similar
to that played by the Ascoli-Arzela theorem 1.30 in the study of spaces of
continuous functions. If u is a function defined (a.e.) on a domain R c R", we
denote by ii the zero extension of u outside R, that is,
G(x) =
if x E R" N R.
2.21 THEOREM Let 1 I p < co. A bounded subset K c LP(R) is pre-

compact in Lp(R) if and only if for every number E > 0 there exists a number
6 > 0 and a subset G c c R such that for every u E K and every h E R" with
Ih( c S
and
PROOF I t is sufficient to prove the theorem for the special case R = R", as
the theorem follows for general R from its application in this special case to
the set I? = {ii :u E K } .
Let us assume first that K is precompact in Lp(R").Let E > 0 be given.
Since K has a finite (e/6)-net (Theorem 1.18), and since C,(R") is dense in
Lp(R") (Theorem 2.13), there exists a finite set S of continuous functions
having compact support, such that for each u E K there exists 4 E S satisfying
~ ~ U - C< #~ ~/ 3~Since
.~ S is finite there exists Y > 0 such that supp4 c B, for
every 4 E S, where B, is the ball {x E R" : 1x1 < r } . Setting G = B,, we obtain
(21). Also, 4 ( x + h ) - 4 ( x ) is uniformly continuous for all x and vanishes
identically outside B,,, provided 1/1 < 1. Hence
Since S is finite, (22) is uniform for 4 E S. For u E K let Thu be the translate
of u by h :
Thu(x) = # ( X + h ) . (23)
If 4 E S satisfies 11u-411, < ~ / 3 then
, also IjThu-Thfjllp < ~ / 3 Hence
. by (22)
we have for Ihl sufficiently small (independent of u E K ) ,
~ ~ T h u - u ~ ~5p I I T h u - T T h 4 1 1 p + l l T h 4 - 4 1 1 p + ~ ~ ~ - u ~ ~ p
< (2e/3) + llTh4-411p < 6,

and (20) follows. [This argument shows translation is continuous in LP(R).]
To prove the converse let > 0 be given and choose G cc R" such that
for all u E K
j W-G
lu(x)lPdx < &/3. (24)
For any q > 0 the function J,, * u defined as in (16) belongs to Cm(Rn)and in
particular to C(G). If E C,(R"), then by Holder's inequality
where Th4 is given as in (23). Hence
If u E Lp(R"), let {&} be a sequence in Co(R") converging to u in Lp-norm.

By Lemma 2.18(c), { J , * 4,,} is a Cauchy sequence converging to J,, * u in
L p ( R " ) . Since also ThI$,, + Thu in Lp(R"), (25) extends to all u E Lp(R") :
llJq * u-ullp sup llTh~~-ullp.

heB,
Now (20) implies that l i m l h l . + o ~ ~ T h u=- u0~uniformly

~p for u E K . Hence
PRECOMPACT SETS IN L'(Q) 33
~ i m q + o ~ ~ J q * u=- 0u ~uniformly
~p for UE K . We now fix 4 > 0 so that
E
3.2'
for all u E K .
We show that {J,,* u : u E K } satisfies the conditions of the Ascoli-Arzela
theorem 1.30 on G and hence is precompact in C(G). By (19) we have
IJ, * U W l 5 (~~~"J,,(x~)l"llullp
which is bounded uniformly for x E R" and u E K since K is a bounded set in
LP(Q)and q is fixed. Similarly
and so lim,,l,,J,*u(x+h) = J , * u ( x ) uniformly for X E R" and U E K .

Thus {J,, * u : u E K } is precompact in C(G) and by Theorem 1.18 there
exists a finite set ..., $,,,} of functions in C(G) such that if u E K , then for
somej, 1 IjI m, and all x E G we have
Denoting by Gj the zero extension of $ j outside G, we obtain from (24), (26),

(27), and the inequality (la1 + lbl)' I 2P(lalP+ /blp)
Hence K has a finite &-netin Lp(R"), namely { q j :1 < j 5 m},and so is pre-

compact by Theorem 1.18. I
2.22 THEOREM Let 1 I p < co and let K c LP(R).Suppose there exists

a sequence {Rj} of subdomains of R having the following properties:
(a) for each j , S2, c Q,, ; ,
(b) for each j the set of restrictions to Qj of the functions in K is pre-
compact in Lp(Qj);
34 11 THE SPACES Lp(Q)
(c) for every E > 0, there exists j such that
,.I lu(x)lpdx
Then K is precompact in Lp(a).

<E for every u E K.
PROOF Let {u,} be a sequence in K. Then by (b) there exists a subsequence

{ u i ' ) } such that the restrictions { u i l )In,} converge in Lp(Q,). Having selected
{ u ! ' ) } , ..., { u y ) } , we may select a subsequence { u ~ ~ " )of
} { u p ) } such that
{ u i k + l )Ink+,} converges in LP(Qk+ '). Hence also {u:~+')(~~} converges in
Lp(Qj) for 1 < j 5 k + 1 by (a).
Let u, = u r ) for n = 1,2, .... Clearly {u,} is a subsequence of {u,}. Given
E > 0, there existsj [by (c)] such that
.. .
64, ~U,(X)-Uu,(X)~PdX < E/2
for all n,m = 1,2, . Except for the first j - 1 terms, {u,} is a subsequence of
{u,?} and so {u, In,} is a Cauchy sequence in Lp(Qj). Thus for n,m sufficiently
large we have
n
Combining (28) and (29) we see that {u,} is a Cauchy sequence in LP(Q)and
so converges there. Hence K is precompact in LP(Q). I
We remark that Theorem 2.22 is just a setting, suitable for our purposes,
of a well-known theorem stating that the operator-norm limit of a sequence of
compact operators is compact.
The Uniform Convexity of Lp(Q)
-=
2.23 For 1 < p co the space LP(Q)is uniformly convex, its norm 11. / I p
satisfying the condition prescribed in Section 1.19. This result, due to
Clarkson [19], is obtained via a set of inequalities for LP(R)that generalizes
the parallelogram law in L2(Q). These inequalities are given in Theorem 2.28,
for the proof of which we prepare the following lemmas.
2.24 LEMMA If 1 I p < co and a 2 0, b 2 0, then

(a+b)P 5 2P-1(ap+bp).
THE UNIFORM CONVEXITY OF LP(n) 35
PROOF If a = 0, (30) clearly holds. If a > 0, (30) may be rewritten in the form
(1+x)P 5 2P-l(l+xp) (31)
where 0 I x = b/a. The function f ( x ) = (1 +x)”(l + x p ) satisfies f(0)= 1 =
limx-tcof ( x ) andf(x) > 1 if 0 < x < co.Hencefhas its maximum for x 2 0 at
its only critical point, x = 1. Sincef(1) = 2p-1,(31) follows. I
2.25 LEMMA If 0 < s < 1, the function f ( x ) = (1 -sx)/x is a decreasing

function of x > 0.
PROOFf ’ ( x ) = (1/x2)(g(sx)- 1) where g ( t ) = t - t In c. Since 0 < sx < 1 and

since g’(t) = -1nt 2 0 for 0 < t s 1, it follows that g(sx) < g(1) = 1 whence
f ’ ( x ) < 0. I
2.26 LEMMA If 1 < p 5 2 and 0 5 t I1, then
where p’ = p / ( p - I) is the exponent conjugate to p .
PROOF Since equality clearly holds in (32) if either p = 2 or t = 0 or t = 1,

we may assume that 1 s > 0, (32) reduces to the equivalent form
+[(l+s)P+(l-s)P] -(l+sp’)p-1 2 0. (33)
If we denote
the power series expansion of the left side of (33) takes the form
The latter series is convergent for 0 < s < I . We prove (33) by showing that
-=
each term in the series is positive for 0 < s 1. The kth term can be written
36 11 THE SPACES Lp(a)
in the form
The first factor above is positive since p c 2; the factor in brackets is positive
by Lemma 2.25 since 0 c ( 2 k - p ) / ( p - 1) < 2 k / ( p - 1). Thus (33) and hence
(32) is established. I
2.27 LEMMA Let z, w E C. If 1 0. In this case (34) can be rewritten in
the form
where MJ/Z = rei*, r 2 0, 0 I 8 c 2x. If 0 = 0, (36) is already proved in

Lemma 2.26. We complete the proof of (36) by showing that for fixed r the
function
f ( 8 ) = I1+reieIP’+I1-rei*IP‘
has a maximum value for 0 s 8 < 2x at 8 = 0. Since
j ( e ) = (1+r2+2rcose)p’/2 + (l+r2-2rcos8)p’/2,
it is clear thatf(2n-8) = f ( x - 8 ) = f ( 8 ) , so that we need considerfonly on
the interval 0 I8 I 7~12.Since pr 2 2, we have on that interval
f y e ) = -p‘r sin8[(1 + r 2 + 2 r cose)(p’/2)-1- (1 + r 2 - 2 r C O S B ) ( P ’ / ~I) - 0.
~]
Thus the maximum value off occurs at 8 = 0 and (36) is proved.
THE UNIFORM CONVEXITY OF LP(Q) 37
If 2 Ip < 00, then 1 < p’ I2 and we have, interchanging p and p’ in (34)

and using Lemma 2.24,
so that (35) is also proved. I

2.28 THEOREM (Clarkson’s inequalities) Let u,v E L P ( 0 ) . For 1 <
p < co letp’ = p / ( p - 1). If 2 ~p < co, then
If 1 < p s 2, then
PROOF For 2 I p < co, (37) is obtained by taking z = u ( x ) and w = v ( x ) in

(35) and integrating over 0. To prove (39) for 1 < p I 2 we first note that
1 l U I p ’ i l p - ~= 11~118’for any u E LP(Q).Using the reverse Minkowski inequality
(7) corresponding to the exponent p - 1 < 1, and (34) with z = u ( x ) ,
M’ = v ( x ) , we obtain
equation continues
38 11 THE SPACES J!!!(R)
(1(!
I n 2 lu(x>l” + 2 lL.(X)lP) d x )
p’- 1
p’- 1
1 1
= ( j llull; + 5 lul;)
which is (39).
Inequality (38) can be proved for 2 I p < co by the same method used to
prove (39), except that the direct Minkowski inequality (5), corresponding to
p - 1 2 1, is used in place of the reverse inequality, and in place of (34) is used
the inequality
which is obtained from (34) by replacing p by p’, z by ( q, and w by ( - q. +

Finally, (40) can be obtained from a similar revision of (33). We remark that
all four of Clarkson’s inequalities reduce to the parallelogram law
IIU+uII: + llu-ull; = 2IlUll: + 2IlulI;
in the case p = 2. I
2.29 COROLLARY If 1 0. If
2 5 p < co, we have from (37)
1 p q ; s l-2.p.&P
If 1 0 such that II(u+u)/2jlP I 1-8. I
Being uniformly convex, Lp(R) is reflexive for 1 c p < co by Theorem
1.20. We shall give a direct proof of this reflexivity after computing the dual
of L‘(R).
The Normed Dual of Lp(R)

2.30 Let 1 I p I 00 and let p’ denote the exponent conjugate to p . For each
element u E LP’(R)we can define a linear functional L, on LP(R)via
~ ( X ) U ( X dx,
) u E LP(R).
THE NORMED DUAL OF LP(n) 39
By Holder’s inequality JL,(u)lI \ l ~ l / ~ I I t ’so

l l ~that
. L, E [Lp(Q)]’and
I I L u ; CLP(Q>l’ll lI~llp*. (41)
We show that equality must hold in (41). If 1 < p Ico, let u ( x ) =
~ o ( ~ ) ~ ” ’ - ~ u (ifX )o(x) # 0 and u ( x ) = 0 otherwise. Then u E Lp(Q) and
L,(u) = JIu/lp/IuIIpt. Now suppose p = 1 so p‘ = co. If IIull,,. = 0, let u ( x ) = 0.
Otherwise let 0 < E < 1 1 0 1 1 ~ and let A be a measurable subset of R such that
0 < p ( A ) < co and I u ( x ) l 2 I I U I I ~ - E on A . Let u ( x ) = It’(x)l-’v(x) for x E A ;
u ( x ) = 0 otherwise. Then u E Ll(R) and L,(u) 2 ~~u~~ l ( ~ ~ u ~ ~ mThus
- ~ ) we
. have
shown that
I I L u ; CLP(Q)1’ll = Ilt’llp, (42)
so that the operator L mapping t’ to L, is an isometric isomorphism of LP’(Q)
onto a subspace of [LP(Q)]’.
2.31 It is natural to ask whether the range of the isomorphism L is all of

[Lp(Q)]’, that is, whether every continuous linear functional on LP(R) is of
the form L, for some o E LP’(Q).We shall show that such is the case provided
1 I p < co. For p = 2 this is an immediate consequence of the Riesz repre-
sentation theorem for Hilbert spaces. For general p a direct proof can be
given using the Radon-Nikodym theorem (see Rudin [ S S ] or Theorem
8.18). We shall give a more elementary proof based on the uniform convexity
of Lp(R)and a variational argument. This method of proof is also used by
Hewitt and Stromberg [32]. Finally we shall use a limiting argument to obtain
the case p = 1 from the case p > 1.
2.32 LEMMA Let 1 < p < 03. If L E [LP(R)]‘ and 1 L; [Lp(Q)]’Ij = 1, then
there exists unique w E LP(R) such that IIw’lI,, = L ( w ) = 1. Dually, if w E Lp(R)
is given and llwllp = 1, then there exists unique L E [LP(R)]’ such that
JJL;[LP(Q)]’)I= L(w) = 1.
PROOF First assume that L E [Lp(R)]’ is given and llLll= I . There exists a
sequence {w,,}E LP(R) such that ~ ~ w=, ,1 ~and ~ limn+m~L(wn)~ = 1. We may
+
assume that IL(w,,)l> for each n, and, replacing w,, by a suitable multiple
of w,, by a complex number of unit modulus, that L(wJ > 0. Suppose the
sequence {w,,} is not a Cauchy sequence in Lp(Q).Then there exists E > 0 such
that ~ ~ w , ,wmIlp
-- 2 E for some arbitrarily large values of m and n, so that by
uniform convexity we have ~ \ + ( W , , + W , , , ) ~ ~ ,5, 1-6, where 6 is a fixed positive
number. Thus
1 1
1-6 --[L(w,)+L(w,)]
2 - (43)
2
40 11 THE SPACES u(n)
Since the last expression approaches 1/(1-6) as n, m + 00, we have a contra-

diction. Thus {w,} is a Cauchy sequence in Lp(n)and so converges to an
element w of that space. Clearly 1 wllp = 1 and L(w) = L(w,) = 1. The
uniqueness of w follows from (43) applied to two distinct candidates.
Now suppose w E LP(n)is given and /I wJlp= 1. As noted in Section 2.30
the functional L, defined by
where
satisfies L,(w) = 11 wll; = 1 and 1 L,; [LP(n)]’II= IIullp, = )I wl/;’P’ = 1. It remains

to be shown, therefore, that if L,, L, E [LP(Q)]‘satisfy 1 L , 11 = IIL,II =
L,(w) = L,(w) = I , then L , = L,. Suppose not. Then there exists u E LP(Q)
such that L,(u) # L2(u). Replacing u by a suitable multiple of u, we may
assume that L,(u)-L,(u) = 2. Then replacing u by its sum with a suitable
multiple of w, we can arrange that L,(u) = 1 and L,(u) = - 1. If t > 0, then
L , ( w + t u ) = l + t ; since IILIII= 1, therefore IIw+tujlp2 1 + t . Similarly,
L,(w-tu)=l+t so I I w - r u l l p 2 1 + f . If 1 < p I 2 , Clarkson’s inequality
(40) gives
1 + PIIUII; = (w+tu)+(w-tu)
(46)
If 2 Ip < co,Clarkson’s inequality (38) gives
1 + tP’IIuII; = (w+tu) + (w-tu)
(47)
Equations (46) and (47) are not possible for all t > 0 unless llullp = 0 which is
impossible. Thus L , = L,.
2.33 THEOREM (The Riesz representation theorem for LP(R)) Let

1 < p < 00 and let L E [LP(Q)]‘.Then there exists u E LP’(R)such that for all
THE NORMED DUAL OF LP(Q) 41
u E LP(Q)
L(u) =
Sn u(x)u(x)dx.
Moreover, lIullp. = 11 L; [LP(Q)]'il.Thus [LP(Q)]'z LP'(R).
PROOF Lf L = 0, we may take u = 0. Accordingly, we assume L # 0 and,

without loss of generality, that IIL; [Lp(R)]'/I = 1. By Lemma 2.32 there
exists w E LP(Q)with 11 wllp = 1 such that L ( w )= 1. Let u be given by (45).
Then L,, defined by (44), satisfies IIL"; [Lp(Q)]'ll= 1 and L,(w) = 1. By
Lemma 2.32, again we have L = L , . Since IIullp. = 1, the proof is complete. I
2.34 THEOREM (Riesz representation theorem for L'(Q)) Let L E

[L' (a)]'.Then there exists u E Lm(R) such that for all u E L' (0)
L(u)=
Sn u ( x ) u ( x ) dx
and l\ulla = IIL; [L'(Q)]'ll. Thus [L'(R)]' z L"(Q).
PROOF Once again we may assume that L # 0 and I1L; [L'(R)]'I/ = 1. Let
us suppose, for the moment, that R has finite volume. Then by Theorem 2.8
if 1 < p < co, we have Lp(Q)c L' (Q) and
IL(u)I IIIui(, I(volQ)'-(''P) llullp

for any u E LP(Q). Hence L E [LP(R)]' and by Theorem 2.33 there exists
up E Lp'(Q)such that
/IUPllP' I(volQ)'-(1'p) (48)
and for every u E LP(R)
(49)
Since Coa(Q)is dense in LP(Q)for 1 < p < co,and since for anyp, q satisfying
1 < p , q < co and any 4 E Co"(Q) we have
it follows that up = uq a.e. on R. Hence we may replace up in (49) bv a function

u belonging to LP(Q)for each p. 1 c p < co, and satisfying, following (48),
IIullpt I (volR)'-~''~).
It follows by Theorem 2.8 again that v E La(Q) and
42 11 THE SPACES LP(R)
The argument at the beginning of Section 2.30 shows that equality must
prevail in (50).
If R does not have finite volume, we may nevertheless write R = u?'
I Gj,
where Gj = { x E R : j - 1 I 1x1 < j } has finite volume. The sets G j are mutually
disjoint. Let x j ( x ) be the characteristic function of G j . If u j € L'(Gj),
let lij denote the zero extension of uj outside G j , that is, i i j ( x ) = u j ( x ) if
x E G j , iij(x) = 0 otherwise. Let Lj(uj)= L(fij). Then Lj E [L' (Gj)]' and
l[Lj;[L1(Gj)]'ll 5 1. By the finite volume case considered above there exists
uj E L"(G,) such that JIujJlm,GjI1 and
where u ( x ) = u j ( x ) for X E Gj ( j =1,2, ...), so that llulloo 1. If u ~ L ' ( n ) ,

we put u = C?==, x j u ; the series converging in norm in L'(R) by dominated
convergence. Since
we obtain, passing to the limit by dominated convergence,
L(u) = U ( X ) U ( Xdx.
)
SR
It then follows, as in the finite volume case, that IIulJrn= 1. I
2.35 THEOREM Lp(R) is reflexive if and only if 1 c p < 00.
PROOF Let X = LP(R), where 1 < p c 03. Since X' r LP'(R), to any w E X"
there corresponds fi E [Lp'(R)]' such that w ( u ) = @(a), where
u(u) = v"(x)u(x)dx, u E X.
SR
Similarly, corresponding to fi E [LP'(R)]' there exists u E X such that
G(E) = ii(x)u(x) dx, v" E LP'(i-2).

sn
It follows that
w ( u ) = fi(5) = i i ( x ) u ( x ) d x = u(u) = J,u(u)

Sn
for all u E X', Jx being the natural isometric isomorphism (see Section 1.13)
of Xinto X " . This shows that Jx maps Xonto X" so that X = Lp(R) is reflexive.
Since L' (0) is separable while its dual, which is isometrically isomorphic
to Lrn(R),is not separable, neither L' (R) nor L"(R) can be reflexive. I
THE NORMED DUAL OF Lp(n) 43
2.36 The Riesz representation theorem cannot hold for the space Lm(Q)in
a form analogous to Theorem 2.33, for if so, then the argument of Theorem
2.35 would show that L'(R) was reflexive. The dual of L"(R) is larger than
L'(Q). It may be identified with a space of absolutely continuous, finitely
additive set functions of bounded total variation on R. The reader is referred
to Yosida [69, p. 1 IS] for details.
I11
The Spaces WmJ’(&2)
Definitions and Basic Properties

In this chapter we introduce Sobolev spaces of integer order and establish
some of their basic properties. These spaces are defined over an arbitrary
domain R c R” and are vector subspaces of various spaces LP(R).
3.1 We define a functional 11 where m is a nonnegative integer and

1 Ip s 03, as follows:
IIUllm,m = max IIDauIIco (2)

Oslolsm
for any function u for which the right side makes sense, 11 I l p being, of course,
3
the LP(R)-norm. (In situations where confusion of domains may occur we

shall write / I ~ l l , , ~ , ~ in place of ~ ~ u ~ ~Itmis, clear
p . ) that ( I ) or (2) defines a norm
on any vector space of functions on which the right side takes finite values
provided functions are identified in the space if they are equal almost every-
where in R. We consider three such spaces corresponding to any given values
of m and p :
H m S P ( R=) the completion of { u E Cm(Q): < 03} with respect to the
norm I1 * II m, p
9
44
DEFINITIONS AND BASIC PROPERTIES 45
W"'9p(R)= {u E Lp(R) : Dau E LP(R) for 0 5 ( a (5 m,where Dau is the weak

(or distributional) partial derivative of Section 1.57}, and
Wt,p(R) = the closure of Com(R)in the space W",P(R).
Equipped with the appropriate norm (1) or (2), these are called Sobolev spaces
-=
over R. Clearly Wo3p(R)= Lp(R), and if 1 s p co, W$p(R) = Lp(Q) by
Theorem 2.19. For any m the chain of imbeddings
wo" * p (n1 -+ W"*P(Q)+ Lp(n)
is also clear. We shall show in Theorem 3.16 that H"9P(R)= Wm,p(R) for
every domain R. This result, published in 1964 by Meyers and Serrin [46]
dispelled considerable confusion about the relationship of these spaces that
had existed in the literature before that time. It is surprising that this elementary
result remained undiscovered for so long.
The spaces Wm*p(R)were introduced by Sobolev [62,63] with many
related spaces being studied by other writers, in particular Morrey [47] and
Deny and Lions [21]. Many different symbols ( W m g P , etc.)
, P m g PLmP,
Hmsp,
have been (and are being) used to denote these spaces and their variants, and
before they became generally associated with the name of Sobolev they were
sometimes referred to under other names, for example, as "Beppo Levi
spaces."
Numerous generalizations and extensions of the basic spaces W m . P (R)
have been made in recent times, the great bulk of the literature originating in
the Soviet Union. In particular we mention extensions that allow arbitrary
real values of m (see Chapter VII) and are interpreted as corresponding to
fractional orders of differentiation, weighted spaces that introduce weight
functions into the LP-norms, spaces W'sP that involve different orders of
differentiation and different LP-norms in the various coordinate directions
(anisotropic spaces), and Orlicz-Sobolev spaces (Chapter VIII) modeled on
the generalizations of LP-spacesknown as "Orlicz spaces."
It will not be possible to investigate the complete spectrum of possible
generalizations in this monograph.
3.2 THEOREM W"*P(R)is a Banach space.
PROOF Let {u,} be a Cauchy sequence in W " ~ p ( QThen

). {D"u,} is a Cauchy
sequence in LP(R)for 0 I la1 Im.Since LP(R)is complete there exist functions
u and u,, 0 < JaIIm, in LP(R) such that u,, u and Dau, -+ u, in LP(R) as
-+
n --t co. Now LP(R)c L:,,,(R) and so un determines a distribution Tun E 9'(R)
as in Section 1.53. For any 4 E 9 ( R ) we have
1
I K ~ ( + ) - T ~ ( ~ ) I lun(x)-u(x)I I + ( ~ )dx
R
I 5 II+IIp'IIun-ullp
46 111 THE SPACES w"'p(a)
by Holder's inequality, where p' = p / ( p - 1) (or p' = co if p = 1, p' = 1 if
p = a).Hence Tun($) +Tu(b,) for every b, E Q((Sz) as n + co. Similarly,
for every b, E 9 (a).It follows that
TD.,,(b,)+ Tue(b,)
= lim T',,,,(b,) = lim (- l)lalT,n(Dab,)
Tu=(b,) = (- l ) W u ( ~ a b , )
n-rm n-r m
for every b , ~ . 9 ( R ) .Thus u, = D"u in the ditributional sense on R for

011ct1<m, whence U E W'"*p(a).Since l i m f l + m ~ ~ u n - u=~ 0,
~ mW"*P(sZ)
,p is
complete. 1
Distributional and classical partial derivatives coincide when the latter

exist and are continuous; thus it is clear that the set
S = {b, E Cm(Q): Ilb,llm,p < m}
is contained in Wa*p(Q). Since W m V P ( is
Q )complete, the identity operator in
S extends to an isometric isomorphism between HmVP(Q), the completion of S,
and the closure of S in W"*p(Q).It is thus natural to identify H m * p ( Q )with
this closure and so obtain the following corollary.
3.3 COROLLARY Hm*p(S2)c

W"*p(LR).
3.4 Several important properties of the spaces W"*p(Q)are most easily

obtained by regarding Wm*P(i2) as a closed subspace of a Cartesian product
of spaces LP(a).Let N E N ( n , m ) = COsllolsm 1 be the number of multi-
indices u satisfying 0 I lul I m. For 1 I p Ico let LNp = ny=,
LP(a),the
norm of u = ( u l , ...,uN)in LNpbeing given by
By Theorems 1.22, 2.10, 2.17, 2.25, and 2.31, LNp s a Banach space that is
separable if 1 I p c co and reflexive and uniformly convex if 1 < p c co.
Let us suppose that the N multi-indices u satisfying 0 I1011
I m are linearly
ordered in some convenient fashion so that to each U E W"sP(R) we may
associate the well-defined vector Pu in LNPgiven by
PU = ( D a u ) O < l a l < m * (3)
Since IIPu; LNp)I= I I U ~ ~ , , , ~ , P is an isometric isomorphism of W m . p ( nonto
) a
subspace W c LNp.Since Wm*p(i2) is complete, W is a closed subspace of LNP.
By Theorem 1.21, W is separable if 1 Ip c co and reflexive and uniformly
convex if 1 < p c co. The same conclusions must therefore hold for
W"*P(Q)= P-'(W).
DUALITY, THE SPACES W-””‘(a) 47
3.5 THEOREM W m * p ( Qis) separable if 1 s p < co, and is reflexive and

-=
uniformly convex if 1 < p co.In particular, therefore, Wm*(Q) is a separable
Hilbert space with inner product
where (u, u) = fn u(x)v(xr)dx is the inner product in L2(Q).
Duality, the Spaces W-mpp’(Q)
3.6 In the following sections we shall take, for fixed Q, m,andp, the number
N, the spaces L N pand W, and the operator P to be as specified in Section 3.4.
We also define
<u,u> = j-nu(x)u(x)dx
for any functions u, u for which the right side makes sense. For given p, p‘
shall always designate the conjugate exponent :
if p = l
if 1 < p c co
if p = 00.
3.7 LEMMA Let 1 s p < co. To every L E (LNp)’there corresponds

unique u E L$ such that for every u E LNp
L(u) = 2 (Uj,Uj).
N
j= 1
PROOF If 1 < j < N and W E LP(a), let w ( ~=) (0,..., 0,w,O, ...,0) be that
element of LNp whosejth component is w, all other components being zero.
Setting Lj(w) = L ( W ( ~ )we
) , see that Lj E (LP(Q))’ and so by Theorems 2.33
and 2.34 there exists (unique) uj E LP’(Q)such that for every w E LP(Q)
L(w,j,) = L j ( W ) = {W,Uj).
If u E LNp,then
48 111 THE SPACES WrnSp(R)
By Holder's inequality (for functions and for finite sums), we have

N
IL(u)l j=l
llujllpIIujllp' 5 ~ ~ u ~ L N p ~ ~ ~ ~ u ; L $ / ~
so that IIL;(LNP)'IIs IIu; L$ 11. We show that these norms are in fact equal as
follows: if 1 0 there is a measur-
able subset A c R having finite, nonzero volume, such that Iuk(x)I 2 IIukIla, - E
{r
for x E A. Set
uk(x)/uk(x) if x E A and uk(x) # 0
u(x) =
otherwise.
Then
L(u(k)) = (%uk) =
s* l'k(x)ldx 2 (\\Uk\\a,-E)lIu\ll
= (II~;LNDDII-~)II~~~);LNIII.
Since E is arbitrary, (4) must follow in this case also. I
3.8 THEOREM Let 1 5 p < co. For every L E ( Wrn*p(R))' there exists an
element u E L$ such that, writing the vector u in the form (u,)oslulsrn, we
have for all u E W",p(Q)
L(u) = c
Osllalsrn
(DPIU,UU).
Moreover,
11 L ; (W"*P(R))' )I = inf 1 u; L$II = min 11 o; L$ I), (6)
the infimum being taken over, and attained on the set of all u E L$ for which
( 5 ) holds for every u E Wrn*p(.R).
PROOF A linear functional L* is defined as follows on the range W of the

operator P defined by (3) :
L*(PU) = L(u), u E W".P(R).
Since P is an isometric isomorphism, L* E W' and
JIL*;W'II = IIL;(W"*P(Q))q.
DUALITY, THE SPACES W-rn'P'(a) 49
By the Hahn-Banach theorem there exists a norm preserving extension L of

L* to all of LNp, and by Lemma 3.7 there exists U E Lj;' such that if
u = (uah 4 la1 s rn E LN', then
z(u) = 1
0 < ( a ( srn
<ua,ua)*
Thus for u E Wrn*"(Q)

we obtain
L(u) = L*(Pu) = Z(h)= 1
Os(a(<rn
(Dau,ua).
Moreover,
Now any element u E Lj;' for which ( 5 ) holds for every u E W"'*p(Q)corre-
ponds to an extension L of L* and so will have norm Ilu; Lj;'l/ not less than
Combining this with (7), we obtain (6). I
IIL;(W'"*P(Q))'II.
We remark that, at least if 1 < p < co, the element u E Lj;' satisfying (5)
and (6) is unique. Since LN' and Lj;' are uniformly convex it follows by an
argument similar to that of Lemma 2.32 that linear functionals defined on
closed subspaces of LNp have unique norm preserving extensions to LNp.
3.9 For I I p < co every element L of the space ( Wrn.P(Q))' is an extension

to W'"n,p(Q)of a distribution T E g'(C2).To see this suppose L is given by ( 5 )
for some v E L$ and define Tvm,T E g'(Q), by
Ke(4)= (4,va), 4 E g(W, 0 I 1x1 5 m,
T= C (- l)l"lD"T,a.
O s l a l srn
For every 4 E 9(Q) we) have
c WmSp(Q
T(4) =
0s
c
Jalsrn
T,.(Da4) = L ( 4 )
so that Id is clearly an extension of T. Moreover, we have, following (6),

= min{llv; LCII : L extends T given by (8)).
\lL;(Wrn*p(Q))'))
The above remarks also hold for L E (W;*p(Q))'since any such functional
possesses a norm-preserving extension to W m *(a).
Now suppose T is any element of 9'(C2)having the form (8) for some
u E L$, where 1 I p' 5 co. Then T possesses possibly nonunique continuous
extensions to Wrn*p(Q).We show, however, that Tdoes possess a unique such
extension to W$*P(Q).If u E W;$p(Q),let {4,} be a sequence in Co"(i2) =
such that I l q 5 , , - ~ l l ~ , ~ + O as n + 00. Then
9(Q)
5 c
O i l a l srn
II'"(4k-4n)IIp IIuallp'
~ ~ ~ k - ~ n ~ ~ 0,
m p as
~ ~ k,n
~ ;--t~ 00-
$ ~ ~
Therefore {T($,)} is a Cauchy sequence in C and so converges to a limit that

we may denote by L(u) since it is clear that if also {$"} c 9(Q)and
~~i,bn-u~~m,p-i0, then T(4")- T($,) -+ 0 as n + 00. The functional L thus
defined is linear and belongs to (W$*P(n))', for if u = limn+m4, as above, then
IL(u)l = limIT(4n)l 5 ~imll4nllm,pll~;G'II
= II~IIrn,plI~;~$II-
n+ 03 n+cc
We have therefore proved the following theorem.
3.10 THEOREM Let 1 5 p < co. The dual space (Wtvp(n))' is iso-
metrically isomorphic to the Banach space consisting of those distributions
T EW(Q)satisfying (8) for some v E Lf;',normed by
(1 T 11 = inf { 11 u; Lf;'11 : u satisfies (8)).
In general one cannot expect any such simple characterization of (Wrn*p(Q))r
if W t * p ( Q )is a proper subspace of Wmsp(Q).
3.11 If m = 1,2,...and 1 5 p < 00, let p' denote the exponent conjugate to
p and denote by W-m,p'(Q)the Banach space of distributions on Q referred to
in the above theorem. (The completeness of this space is a consequence of the
isometric isomorphism asserted there.) Evidently W - r n , p'(Q) is separable and
reflexive if 1 l5 I I ~ l l p ~ I I ~5l l pI I ~ l l p ~ I I ~ l l r n , p ~
We define the (-m,p')-norm of v E LP'(R)to be the norm of L,, that is,
1141-m,p' = l l ~ " ; ( w t ~ p ( Q ) )=~ l l uSUP
ew
I(% U>l,
IIUllm, P 41
where we have written W to represent WFSP(Q)on the right side, a practice we

continue below, for simplicity. Clearly, for any u E Wand v E LP'(R)we have
The latter formula is a generalized Holder's inequality.

APPROXIMATION BY SMOOTH FUNCTIONS ON 51
Let V = { L , : u E Lp'(Q)}.Thus V is a vector subspace of W' = (W;*P(Q))'.

We show that V is in fact dense in W ' . This is easily seen to be equivalent to
showing that if F E W" satisfies F(L,) = 0 for every L, E V, then F = 0 in
W". Since W is reflexive, there existsf E W corresponding to F E W" such that
(f,u> = L,( f ) = F(L,) = 0 for every u E Lp'(Q).Since f E Lp(Q),it follows
that f ( x ) = 0 a.e. in Q. Hence f = 0 in W and F = 0 in W".
Let H-m*p'(Q)denote the completion of LP'(R) with respect to the norm
1 - m , p , . Then we have
H -"*"(n)E (W;,p(Q))' z w - m * p ' ( Q ) .
In particular, corresponding to each u E H-m*P'(Q),there exists T, E
W-m*p'(Q) such that T,(R) = l i m n + m ( ~u,), for every 4 E 9(Q) and every
sequence {u,} c LP'(R)such that limn-tmllu,,-uII- m , p , = 0 and conversely any
T E H-m*P'(Q)satisfies T = T, for some such u. Moreover, by (9), IT,,(4)II
I l 4llm,p llull - m , p ' *
3.13 By an argument similar to that in Section 3.12 the dual space ( Wm*p(Q))'
can be characterized for 1 < p < co as the completion of Lp'(Q)with respect
to the norm
l l ~ l l ~ m , p=,
UE
SUP
Wm.qn)
I(wu)I*
llullm, P 5 1
Approximation by Smooth Functions on Q
We wish to prove that {d, E C"(Q):IIC#I~~,~

c co} is dense in W m * p ( QTo
).
this end we require the following standard existence theorem for infinitely
differentiable partitions of unity.
3.14 THEOREM Let A be an arbitrary subset of R" and let 8 be a

collection of open sets in R" which cover A , that is, such that A c U. uUEB
Then there exists a collection Y of functions I) E Com(Rn)having the following
properties :
(i) For every $ E Y and every x E R", 0 I $(x) I1.
(ii) If K cc A , all but possibly finitely many I)E Y vanish identically
on K .
(iii) For every I,$ E Y there exists U E 0 such that supp I) c U .
(iv) For every x E A , &e.uI,$(x) = 1.
Such a collection Y is called a C"-partition oj'unity for A subordinate to 8.
PROOF Since the proof can be found in many texts we give only an outline
of it here, leaving the details to the reader. Suppose first that A is compact so
52 111 THE SPACES W"*'(R)
that A c (JYElU j , where U , , ..., UNE 0. Compact sets Kl c U , , ..., KN c UN

can be constructed so that A c uy= K j . For 1 Ij I N there exists a non-
negative-valued function 4j E Com(Uj)such that ~ $ ~ ( > x )0 for x E K j . A
function 4 can then be constructed so as to be infinitely differentiable and
positive on R" and to satisfy 4 ( x ) = z:=, $,(x) for x E A. Now Y =
{I)~: t,bj(x) = + j ( x ) / 4 ( x ) , 1 5 j I N} has the desired properties. Now suppose
A is open. Then A = (JYxl A j , where
A, = { x E A : 1x1 ~j and dist ( x , bdry A) 2 lo}
is compact. For eachj the collection
0, = { U n (interiorAj+, n AS-,): U E U }
covers A j and so there exists a finite C"-partition of unity Y j for A, sub-
ordinate to U j . The sum a(x) = cT=l
&.,,,.,+(x) involves only finitely many
nonzero terms at each point, and is positive at each x E A. The collection
Y = { $ : $ ( x ) = +(x)/a(x> for some + in some Y, if X E A, $ ( x ) = 0 if
x # A} has the prescribed properties. Finally, if A is arbitrary, then A c B =
uUEd U,where B is open. Any partition of unity for B will do for A as well. I
3.15 LEMMA Let J, be defined as in Section 2.17 and let 1 I p < co and
u E W"*P(R). If R' c c R, then lime,,+ J, * u = u in W"*p(sZ').
PROOF Let E c dist(R', bdry R). For any 4 E 9(R') we have
In.
* J, u ( x ) D"4 ( x ) dx = k" ) dr
~ " a b - Y ) J . ( Y ) D a 4 ( xdx
= (-1)I"lI
W"
1
R"
D,"U"(x-y)J,(y)4(x)dxdy
= (- 1)la1j- J, * D a u ( x ) 4 ( x )dx,
n.
where u" is the zero extension of u outside R. Thus D"J,*u = J,* D"u in the
distributional sense in R'. Since D"u E Lp(R)for 0 < la1 Im we have by Lemma
2.18 (c)
Thus lim,+O+~~J,*u-u~~,,p,R'
= 0.
3.16 THEOREM (Meyers and Serrin [46]) If 1 I p c 03, then

H"*P(R) = W"*P(R).
PROOF By virtue of Corollary 3.3 it is sufficient to show that W"*p(R)c
that is, that { 4 E Cm(R):~~q5~[,n,p
HmPP(R), c 03} is dense in W",P(R). If
APPROXIMATION BY SMOOTH FUNCTlONS ON R" 53
u E W"*p(R) and E > 0, we in fact show that there exists 4 E Cm(R) such
, , pk = 1,2, ... let
that ~ ~ u - ~ ~< ~E . ,For
Rk = {x E Q : 1x1 .c k and dist(x, bdryR) > l / k } ,
and let R, = R - , = 0,the empty set. Then
-
0= { u k : uk = Qk+r n ( R k - i y , k = 1,2, ...}
is a collection of open subsets of R that covers R. Let Y be a C"-partition
of unity for C? subordinate to 0. Let $k denote the sum of the finitely many
functions 9 E Y whose supports are contained in u k . Then $k E Corn( U,)and
Ckm,l$&) = 1 on R.
I f 0 < E < l / ( k + l ) ( k + 2 ) ,thenJe*(t)ku)hassupportinR,+zn (&2)'=
c c R. Since $ku E W".p(R) we may choose & k , 0 c &k < l / ( k + l ) ( k + 2 ) ,
such that
llJek * (#ku) - $k'IIm,P,f2 = IIJEk * ($ku) - $ k u J I m , p . Y k < ' / Z k .

Let 4 = Ckm,lJek* ( ~ ) ~ u On
) . any 0'c c R only finitely many terms in the
sum can fail to vanish. Thus 4 E Cm(R). For x E Rk we have
k+2 k+Z
~ ( x =) C $j(x)u<x>,
j=l
4(x) = C
j=l
Jej * <$ju><x)*
Thus
k+Z
IIu- $Ilmn,p,nk
j=1
C IIJej * (+ju)-+jUllm,p,n < 8.
By the monotone convergence theorem 1.43, ~ J U - + I ~ , , , ~ , ~ .c E. I

We remark that the theorem does not extend to the case p = co. For
instance, if R = { x E R : - 1 < x < 1 ) and u ( x ) = 1x1, then u E W1*m(R)but
u .$ H""(R); in fact, if E < +, there exists no function 4 E C'(n) such that
II4'-U'Ilm < E.
Approximation by Smooth Functions on R"
3.17 Having shown that an element of W"*P(R)can always be approximated

by functions smooth on R we now ask whether or not the approximation can
in fact be done with bounded functions having bounded derivatives of all
orders, or, say, of all orders up to m. That is, we are asking whether for any
values of k 2 m the space C"(n)is dense in W"qP(R). The answer may be
negative as the following example shows :
Let R = { ( x , y ) E Rz :0 < 1x1 < 1 , O < y < 1). Then the function u specified
54 111 THE SPACES Wm*'(R)
by
evidently belongs to W'*p(R). The reader may verify, however, that for
sufficiently small E > o no function 4 E C1(Ti) can satisfy ))I(- 4 )I < E.
The difficultywith this domain is that it lies on both sides of part ofits boundary
(the segment x = 0,O < y < 1).
We shall say that a domain R has the segment property if for every
x E bdryR there exists an open set U, and a nonzero vector y, such that
x E U, and if z E n V,, then z + t y , E R for 0 < t < 1. A domain having
this property must have (n - 1)-dimensional boundary and cannot simul-
taneously lie on both sides of any given part of its boundary.
The following theorem shows that this property is sufficient to guarantee

that Co"(R") is dense in Wm.p(R),and hence in particular that Ck(Ti)is dense
in Wm.p(R)for any m.
3.18 THEOREM If R has the segment property, then the set of restrictions
to R of functions in Com(Rn)is dense in WmSp(R)for 1 I;p < 00.
PROOF Letfbe a fixed function in Coco(R")satisfying

(i) f ( x ) = 1 if 1x1 I; 1,
(ii) f ( x ) = 0 if 1x1 2 2,
(iii) ID"f(x)l I M (constant) for all x and 0 I1.1 5 rn.
Let j i ( x ) EX) for E > 0. Then f , ( x ) = 1 if 1x1I I/& and ID"fe(x)lI
M~l'l 5 M if E 5 1. If u E WmSp(R),then u, = f , . u belongs to Wm.p(R)and
has bounded support. Since, for 0 < E I 1 and1.1 I m,
we have, setting Re = { x E R:lxl> l / ~ } ,

Il~--.zllm,p,n = IIu- uellm,p,n.
5 llullm,p,n8 + IIUellm,p,n. I c~n~tII~llm,p,n..

The right side tends to zero as E tends to 0. Thus any u E Wmn,p(R) can be
approximated in that space by functions with bounded supports.
-
We may now, therefore, assume K = { x E R : u(x) # 0} is bounded. The
set F = K (Uxebdryn Ux)is thus compact and contained in R, {U,} being
the collection of open sets referred to in the definition of the segment property.
APPROXIMATION BY SMOOTH FUNCTIONS ON R" 55
There exists an open set Uosuch that F c c Uo c c a. Since R is compact,

there exist finitely many of the sets U,;k t us rename them U,,...)Uky such
that R c Uo v U1 u ... u uk. Moreover, we may find other open sets
80,0,, ..., o k SUChthat8j C C U j ,0 l j l k, but stillx C 8,v 8,V U 0,. a * *
Let Y be a C"-partition of unity subordinate to { o j : O S j s k}, and let

$ j be the sum of the finitely many functions $ E Y whose supports lie in oj.
Let u, = 1,9~u.Suppose that for each j we can find (pj E Com(R")such that
lluj-+jllm,p,n < d(k+ 1). (10)
Then putting (p = C;=,
+ j , we would obtain
k
II+-uIIm,p.R 5 C I I ~ j - ~ j I I m , p , n< E *
j=o
A function r$o E Co"(R")satisfying (1 0) f o r j = 0 can be found via Lemma 3.15

since suppuo c 0, c c R. It remains, therefore, to find 4j satisfying (10) for
-
1 s j Ik. For fixed suchj we exend uj to be identically zero outside R. Thus
uj E W m * P ( R " oj
r), where r = n bdryR. Let y be the nonzero vector
associated with the set Uj in the definition of the segment property (Fig. I).
Let r, = T - t y , where t is so chosen that
0 < t < min(l,dist(oj, R" - Uj)/lyl).
FIG.1
56 111 THE SPACES Wm'p(R)
r,c Uj and r, n is empty by the segment property. Let Vj,,(x)=

Then
-
uj(x+ty). Then uj,, E W"*P(R" lJ. Translation is continuous in LP(R)so
Duuj,,+Duuj in Lp(R) as t + O + , lalsrn. Thus uj,,+uj in WmSP(R)as
t +0 +
- ,
and so it is sufficient to find 4, E COm(R")such that 11 uj, - 4j I( ,, is
sufficiently small. However, R n Ujt c R" & and so by Lemma 3.15 we
may take 4j = Ja* uj,, for suitably small 6 > 0. This completes the proof. 1
3.19 COROLLARY W?p(Rn) = W"3p(Rn).
Approximation by Functions in Com(12);(m,p')-Polar Sets
Corollary 3.19 suggests the question: For what domains R is it true that
W"*p(l2)= Wz,P(l2), that is, when is CoOo(R)dense in Wm3P(R)? A partial
answer to this problem can be formulated in terms of the nature of the distri-
butions belonging to W-"*p'(Rn).The approach given below is due to
Lions [39].
3.20 Throughout the following discussion we assume that 1 < p < 00 and
p' is conjugate to p. Let F be a closed subset of R".A distribution T E 9'(Rn)
has support in F (supp T c F) provided that T ( 4 ) = 0 for every 4 E 9(Rn)
vanishing identically on F. We say that the closed set F is (m,p')-polar if the
only distribution Tin W-", p'(R") having support in F is the zero distribution
T=O.
If F has positive measure, it cannot be (m, p')-polar for the characteristic
function of any compact subset of F having positive measure belongs to
LP'(R") and hence to W-'"*p'(R").
We shall show later that if mp > n, then W"*P(R") + C(W) (see Theorem
5.4) in the sense that if u E Wm*p(Rn), then there exists uo E C(Rn) such that
u(x) = uo(x) a.e. and
luo(x)l 5 constllullfn,,,
the constant being independent of x and u. It follows that the Dirac distribution
6, given by 6,(4) = 4 ( x ) belongs to (Wm*p(Rn))' = ( W;,p(IWn))' s W-"-p'(Rn).
Hence, if mp > n, a set F cannot be (m,p')-polar unless it is empty.
3.21 Since Wm+l*p(R") + W"*P(R")any bounded linear functional on the

latter space is bounded on the former as well, that is, W-m*p'(sL) c
W-"-'.P'(l2). Hence any (m+ 1,p')-polar set is also (m,p')-polar. The con-
verse is, of course, generally not true.
Let the mapping u + G denote zero extension of u outside a domain

APPROXIMATION BY FUNCTIONS IN ; POLAR
COm(R) SETS 57
R c R":
u"(x) =
if x E R',
The following lemma shows that this mapping takes W;sp(R) (isometrically)
into W"*p(Rn).
3.22 LEMMA Let u E W$*p(R). If JCIJ5 m, then D"u" = (D'u)" in the

distributional sense in R". Hence u" E W"9P(Rn).
PROOF Let {+,,} be a sequence in C,"(R) converging to u in the space

W$*P(R).If i,b E !B(R"), we have for la1 5 m
Thus D"u" = (D'u)" in the distributional sense on R" and hence these locally
integrable functions are equal a.e. in R". It follows that IIu"llm,p,Rn =
IIullm,p,R* I
We now give a necessary and sufficient condition on R that mapping (1 1)
carry W$sP(R)isometrically onto W"*p(R").
3.23 THEOREM Com(R)is dense in W"*p(R")

if and only if the comple-
ment R' of R is (m,p')-polar.
PROOF First suppose COm(R) is dense in W"-P(R").Let T E W - m * p ' ( R "have

)
support contained in R'. If u E W m . p ( R nthen
) , there exists a sequence {+,,} c
COm(Q) converging in Wm,p(Rn) to u. Hence T ( u )= limn+" T(4,) = 0 and so
T = 0 and Qc is (m,p')-polar.
Conversely, if COm(R)is not dense in W m 7 p ( R nthen ), there exists
U E W m * p ( R nsuch
) that 2 k > 0 for every 4 E COm(R), k being
independent of 4. By the Hahn-Banach extension theorem there exists
T E W-"*p'(R")such that T ( 4 )= 0 for all 4 E Com(R)but T(u) # 0. Since
supp T c Rc but T # 0, R' cannot be (m,p')-polar. I
58 111 THE SPACES W""(a)
As a final preparation for our investigation of the possible identity of

W?p(Q) and W"'*p(Q)we establish a distributional analog of the fact, obvious
for differentiable functions, that identical vanishing of first derivatives over a
rectangle implies constancy on that rectangle. We extend this first to distri-
butions and then to locally integrable functions.
3.24 LEMMA Let B = (u,, 6 , ) x (a2,b,) x ... x (u,, b,) be an open rectan-
gular box in Iw". Let 4 E 9 ( B ) . If fe$(x) dx = 0, then 4(x) = 4j(x), x'&,
where 4j E 9 ( B ) and
p j ( x l , ..., x j , ...,x,) dXj = 0 (12)
for every fixed (xl, ..., xi- ,,xi+ ,,...,x,) E W""
L L
= [b14(xl,...,x,) dx, - t,h2(x2,...,x,) [bl~l(~l) dx, = 0,
= u,(x,)***uj-,(xj-,)
=0, 2 ~ j ~ n - 1 ,
APPROXIMATION BY FUNCTIONS IN Com(R); (I)I,P')-POLAR SETS 59
= ul(xl) un- l(xn-1) 1

B
+(x) dx = 0. I
3.25 COROLLARY If T E 9 ' ( B ) and Dj T = 0 for 1 I j I n, then there

exists a constant k such that for all 4 E 9 ( B )
PROOF First note that if J B + ( x )dx = 0, then T ( 4 ) = 0, for, by the above

lemma we may write 4 = Z!=l 4j, where 4j E 9 ( B )satisfies (12), and hence
4j = Dj O,, where 0, E 9(B) is defined by
xYGl
Thus T ( 4 ) = T(DjOj)=-x;=, DjT(Oj)= 0.
Now suppose T # 0. Then there exists 4, E 9 ( B ) such that T(4,) =
kl # 0. Hence J B 4o(x) dx = k2 # 0 and T ( 4 , ) = k f B 4, (x) dx, where
. 4 E 9 ( B ) is arbitrary, let K ( 4 ) = JB4(x)dx. Then
k = k 1 / k 2 If
and so T ( 4 - [K($)/k2]4,) = 0. It follows that
It should be remarked that this corollary can be extended to any open,

connected set B in Rnvia a partition of unity for R subordinate to some open
cover of R by open rectangular boxes that are contained in R. We shall not,
however, require this extension.
The following lemma shows that different locally integrable functions on
an open set R determine different distributions on R.
3.26 LEMMA Let u E L:,,(R) satisfy Jnu(x) 4(x) dx = 0 for every

4 E 9(R). Then u(x) = 0 a.e. in R.
PROOF If $ E C,(R), then for sufficiently small positive E, the mollifier
J, * $ belongs to 9 (a).By Lemma 2.18, J , * $ + $ uniformly on R as E + 0 .+
Hence Jnu(x) $(x) dx = 0 for every $ E C,(R).
Let K c c R and let E > 0. Let xK be the characteristic function of K.
SK
Then lu(x)l dx < co. There exists 6 > 0 such that for any measurable set
A c Kwith p ( A ) < 6 we have f A lu(x)l dx < 4 2 (see, e.g., the book by Munroe
[48, p. 1361).
By Lusin's theorem 1.37(f) there exists $ E C,(R) with suppt) c K and
II) ( x )I I 1 for all x , such that
P({X E Q : Il/(x) z X K M s g n u o l ) < 6.
Here
Hence
3.27 COROLLARY If B is a rectangular box as in Lemma 3.24 and

u E L:b,(B) possesses weak derivatives Dju = 0 for 1 5 j In, then for some
constant k, u ( x ) = k a.e. in B.
PROOF By Corollary 3.25 we have, since DjTu= 0, 1 I j I n,
~ u ( x ) 9 ( x ) d x= TU(9) = k
Hence u ( x ) - k = 0 a.e. in B. I
SB 9(x,dx.
3.28 THEOREM (1) If Wm*p(R)= W;sp(R), then Qc is (m,p')-polar.

(2) If R" is both (1,p)-polar and (m,p')-polar, then W"*p(R)= W;pp(R).
PROOF (1) Assume W".p(R) = W;*P(R). We deduce first that R" must
have measure zero. If not, there would exist some open rectangle B c Iw"
which intersects both R and Rc in sets of positive measure. Let u be the restric-
tion to R of a function in COm(Iw") which is identically one on B n R. Then
U E WmSP(R)and so U E W;,p(R). By Lemma 3.22, u" E WmiP(Iw")and
Dju"= ( D j u ) - in the distributional sense in R", for 1 l j 5 n. Now ( D j u ) -
vanishes identically on B whence so does Dju" as a distribution on B. By
Corollary 3.27, u" must have a constant value a.e. in B. Since u"(x) = 1 on
APPROXIMATION BY FUNCTIONS IN Com(R); POLAR SETS 61
B n R and u" (x) = 0 on B n a', we have a contradiction. Thus Qc has measure

zero.
Now if u E W"9p(R")and u is the restriction of u to R, then u belongs to
WmSP(R)and hence, by assumption, to WTiP(R). By Lemma 3.22, u" E
W".P(Rn)and can be approximated by elements of Com(R). But u ( x ) = u"(x)
on R, that is, a.e. in R". Hence u and u" have the same distributional derivatives
and so coincide in W"*p(R"). Therefore Coa(R)is dense in W'"*p(R") and R"
is (rn,p')-polar by Theorem 3.23. '
(2) Now assume R" is (1,p)-polar and (m,p')-polar. Let u E Wm,p(R),

We show that u E Wg*J'(R). Since ti E LP(Rn), the distribution T,,,,
corresponding to Dju",belongs to W-'.p(R"). Since (Dju)* E LP(Rn) c
H - l D P ( R ntherefore
), qDj")- E W - ' v p ( R " ) .Hence TD,I-(D,u)-
E W - l S @").
p
But D,u"-(Dju)- vanishes on R so suppTDj,-(D,,)- c R". Since R, is (1,p)-
polar Dju"= (0,~)- in the distributional sense on R". By induction on Jal
we can show similarly that D"u" = (D"u)" in the distributional sense, for
la1 I nz. Therefore u" E W"*p(Ra)whence, by Theorem 3.23, u, the restriction
of u" to R, belongs to Wg*p(R),Rc being (m,p')-polar.
If (m, p')-polarity implies (1 ,p)-polarity, then Theorem 3.28 amounts to
the assertion that (m,p')-polarity of Rc is necessary and sufficient for the
equality of Wmvp(R)and W2P(R). We now examine this possibility, estab-
lishing first two lemmas containing important properties of polarity. The
first of these shows that (rn,p')-polarity is a local property.
3.29 LEMMA F c R" is (m,p')-polar if and only if F n K is (rn,p')-polar
for every compact set K c R".
PROOF Clearly the (rn,p')-polarity of F implies that of F n K for every
compact K. We prove the converse. Let T E W-"*p'(R") be given by (8) and
have support in F. We must show that T = 0. Letf'E Com(Rn)satisfyf(x) = 1
if 1x1 I1 andf(x) = 0 if 1x12 2. For E > 0 letjL(x) EX) so that D"f,(x)=
EI'IDY(EX)0 uniformly in x as E --f O + . Thenfi T E W-"~p'(R")and for any
$ E 9(Rn) we have
I T($) -f, T ( $ )I = I WJ) $11
- Vf,
62
where
Since S,(x) = 1 for 1x1 II/&, we have lim,,,, llwallp = 0. Thus S,T+ T in
W-m.p'(R")as -+ O+ . Since S, T has compact support in K, it vanishes by
assumption. Thus T = 0. [
3.30 LEMMA If p' Iq' and F c R" is (m,p')-polar, then F is also

(rn,q')-polar.
PROOF Let K c R" be compact. By Lemma 3.29 it is sufficient to show that

F n K is (rn, 9')-polar. Let G be an open, bounded set in R" containing K. By
Lemma 2.8, W t s P ( G ) + WC*q(G)so that W-mn,q'(G) c W-"lSp'(G).Any
distribution T E W-m.g'([w")having support in K n F also belongs to
W-"*q'(G)and hence to W-m*p'(G). Since K n F i s (m,p')-polar, T = 0. Thus
K n F i s (m,q')-polar.
3.31 THEOREM Let p 2 2. Then W m q P ( R

=)W t * p ( l lif) and only if 0
'
is (m,p')-polar.
PROOF Since p' 5 p, f2 is (m,p)-polar, and hence (1,p)-polar, if it is (m,p')-

polar. The result now follows by Theorem 3.28. [
3.32 The Sobolev imbedding theorem (Theorem 5.4) can be used to extend
Theorem 3.31 to cover certain values o f p c 2. If (m - I ) p < n, the imbedding
theorem gives
wm*p(R")-+ W'*q(R"), q = n p / [ n - ( m - l)p],
whichin turnimplies that W-'*q'(Rn) c W-'.p'(Rn).Ifalsop I2n/(n+m- l),
then q' Ip, and so by Lemma 3.30, R' is (1,p)-polar if it is (m,p')-polar. Note
that 2n/(n+m- 1) < 2 provided m > 1. If, on the other hand, (m- 1)p 2 n,
=-
then mp n and, as pointed out in Section 3.20, R" cannot be (rn,p')-polar
unless it is empty, in which case it is (1,p)-polar trivially.
Thus, the only values o f p for which the (m,p')-polarity of R"is not known
to imply (1,p)-polarity and hence be equivalent to the identity of W m * p ( R )
and W:*p(R) are given by 1 Ip Imin(n/(m - l), 2n/(n+ m - 1)).
TRANSFORMATION OF COORDINATES 63
3.33 Whenever W r s p ( Q # ) Wm*p(C2), the former space is a closed subspace

of the latter. In the Hilbert space case, p = 2, we may consider the space Wol
consisting of all u E W'"*p(f2) such that (u, 4)" = 0 for every 4 E Com(Cl).
Every u E W'"*P(Q)can be uniquely decomposed in the form u = u,+v,
where uo E W;.p(Q) and v E Wol. Integration by parts shows that any v E WoL
must satisfy
c
OsJalsm
(- l ) ' " ' P u ( x )= 0
in the weak sense, and hence a.e. in C2.
Transformation of Coordinates
3.34 Let @ be a one-to-one transformation of a domain $2 c R" onto a

domain G c R", having inverse Y = @ - I . We call @ m-smooth if, writing
y = @ ( x ) and
~1 = 41(x1, **.,xn), = $ ~ ( Y -I. ,. , . ~ n )
Y Z = 4Z(x,, **.,xn), XZ = $ ~ ( Y I ,..*,Yn)
Yn = 4n(X1, ...,xn), xn = $n(Yl, .-*,Yn),

the functions 41,..., 4" belong to Cm(Q)and the functions $ 1 , ..., $,, belong
to CyG).
If u is a measurable function defined on R, we can define a measurable
function on G by
A4.Y) = .(Y((Y)). (13)
Suppose that @ is 1-smooth so that for all x E Cl
c I (det@'(x)l I C (14)
for certain constants c, C,O < c I C. [Here, of course, @'(x) denotes the
Jacobian matrix d(yl, ...,y,,)/ d ( x l , ...,x,,).] It is readily seen that the operator
A defined by (13) transforms LP(R) boundedly onto Lp(G)and has a bounded
inverse; in fact (for 1 Ip < a),
IIullp,n 5 IIA~ll,,. I C1'PIIuIIp,R.
We establish a similar result for Sobolev spaces.
3.35 THEOREM Let @ be m-smooth, where m 2 I . Then A transforms

W"*p(Cl)boundedly onto W"*p(G)and has a bounded inverse.
64 111 THE SPACES w"'p(sz>
PROOF We show that the inequality I I A U ~ 5~ constllullm,p,n

~,~,~ holds for any
u E W"3P(Q),the constant depending only on the transformation @. The
reverse inequality / I A U ~ ~2, ,constllullm,p,n
~,~ can be established in a similar
manner using the operator A-' taking functions defined on G into functions
defined on C2.
By Theorem 3.16 there exists for any u E Wmip(Q) a sequence {u,} of
functions in Cm(C2)converging to u in WmsP(Q)-norm.For such smooth u, it
is easily checked by induction that
where Mapis a polynomial of degree not exceeding [PI in derivatives, of orders

not exceeding [MI, of the various components of Y. If C#I E 9 ( G ) , we obtain
from (1 5 ) and integration by parts
or, replacing y by @ ( x ) and expressing the integrals over a,

P
dx
(- l)lalJ u,(x>(DaC#I)(@(x))(detQ)'(x)l
n
= 2 /nDpun(x) MaS(@(x))ldet
181s 14
@'(x)l dx.
Since D h , -+ u in LP(Q)for IPI < m, we may take the limit through (17) as
n -+ 00 and hence obtain (16) with u replacing un. Thus (15) holds in the weak
sense for any u E W m P P ( QWe
) . now obtain from (15) and (14)
Of special importance in later chapters is the case of the above theorem

corresponding to nonsingular linear transformations CP or, more generally,
affine transformations (compositions of nonsingular linear transformations
and translations). For such transformations det @'(x) is a nonzero constant.
IV
Interpolation and Extension Theorems
Geometrical Properties of Domains

4.1 Many properties of Sobolev spaces defined on a domain R, and in
particular the imbedding properties of these spaces, depend on regularity
properties of R. Such regularity is normally expressed in terms of geometrical
conditions that may or may not be satisfied by any given domain. We specify
below five such geometric conditions, including the segment property already
encountered in Section 3.17, and consider their interrelationships. First,
however, we standardize some geometrical concepts and notations that will
prove useful.
Given a point x E R", an open ball B , with center x, and an open ball B ,
not containing x, the set C, = B , n { x + i ( y - x ) : y E B,, I > 0} is called a
+
finite cone in R" having vertex at x. We also denote by x C, = {x + y :y E C,}
the finite cone with vertex at x obtained by parallel translation of a finite cone
Cowith vertex at 0.
Given linearly independent vectors y,,y,, ...,y, E R", the set P =
{zy=l I j y j : O < I j < 1, 1 <j < n} is a parallelepiped with one vertex at the
origin. Similarly, x + P is a parallel translate of P having one vertex at x . By
+ +
the center of x P we mean, of course, the point c ( x + P) = x + + ( y , ... +y,).
Every parallelepiped with one vertex at x contains a finite cone with vertex at
x, and conversely is also contained in such a cone.
An open cover 0 of a set S c R" is said to be locallyfinite if any compact
65
66 Iv INTERPOLATION AND EXTENSION THEOREMS
set in R" can intersect at most finitely many elements of 0.Such locally finite
collections of sets must be countable, so their elements can be listed in sequence.
If S is closed, then any open cover of S possesses a locally finite subcover.
We now define five regularity properties which an open domain R c R"
may possess.
4.2 R has the segment property if there exists a locally finite open cover { U j }
of bdry R and a corresponding sequence { y j } of nonzero vectors such that if
x E Ti n U j for some j , then x + tyj E R for 0 c t < 1.
4.3 R has the cone property if there exists a finite cone C such that each point
x E is the vertex of a finite cone C, contained in R and congruent to C.
(Note that C, need not be obtained from C by parallel translation, just by
rigid motion.)
4.4 R has the unijorm cone property if there exists a locally finite open cover
{ U j } of bdry R, and a corresponding sequence { C j } of finite cones, each con-
gruent to some fixed finite cone C, such that:
(i) For some finite M , every Uj has diameter less than M.
(ii) For some 6 > 0, uj"=
U j =I RbE {x E R : dist (x, bdry R) c S}.
(iii) For everyj, uxEnnu,(x+Cj)= Qj c R.
(iv) For some finite R, every collection of R + 1 of the sets Q j has empty
intersection.
4.5 R has the strong local Lipschitz property provided there exist positive
numbers 6 and M, a locally finite open cover { U j } of bdryn, and for each Uj
a real-valued function & of n - 1 real variables, such that the following con-
ditions hold :
(i) For some finite R, every collection of R + 1 of the sets Uj has empty
intersection.
(ii) For every pair of points x, y E Rb = {x E R : dist (x, bdry R) < S}
such that Ix-yl< 6 there existsj such that
x , y E Yj = { x E Uj : dist(x, bdry V j )> S}.
(iii) Each function& satisfies a Lipschitz condition with constant M:
lf(t1,...,gn-,)-f(rll,...,rtn-l)l 5 ...,tn-l-vn-1>l*
MI(t1-~1,
(iv) For some Cartesian coordinate system (tj,1, ..., t j , Jin Uj the set
R n Ujis represented by the inequality
4;,n C fi<tj.1 9 *.*) C j , n - 1).
GEOMETRICAL PROPERTIES OF DOMAINS 67
We remark that if R is bounded, the rather complicated conditions above

reduce to the simple condition that R have a locally Lipschitz boundary,
that is, that each point x on the boundary of R should have a neighborhood
U, such that bdryR n U, is the graph of a Lipschitz continuous function.
4.6 R has the uniform C"-regularity property if there exists a locally finite
open cover { U j } of bdry R, and a corresponding sequence {mj} of m-smooth
one-to-one transformations (see Section 3.34) with Qj taking Uj onto
B = { y ~ R " : l y l < l }suchthat:
,
(i) For some 6 > O, ( J T = , Y j ( { y R":(yl<+})

~ =aRd, where Y j = @ ; ' .
(ii) For some finite R, every collection of R + 1 of the sets U j has empty
intersection.
(iii) For eachj, @,(Uj n R) = { y E B : y , > 0).
(iv) If ( $ j , ..., 4j,n)and ( $ j , ..., $,,J denote the components of Oj
and Y j ,respectively, then there exists a finite M such that for all a, la1 Im,
for every i, 1 Ii In, and for every j , we have
lDa#j,i(x)1 I M , xE Uj
IM ,
lDVj,i(.~)l Y E B.
4.7 With the exception of the cone property, all the other properties above
require R to lie on only one side of its boundary. The two-dimensional domain
R mentioned in Section 3.17, that is,
R = {(x,y)ER2:O<JX~<1,0<y<1}
has the cone property but none of the other four. The reader may wish to
convince himself that
uniform C"-regularity property (m 2 1)
strong local Lipschitz property
==uniform cone property
segment property
for any domain R.
Most of the important imbedding results of Chapter V require only the
cone property though one requires the strong local Lipschitz property.
Although the cone property implies none of the other above properties it
"almost" implies the strong local Lipschitz property for bounded domains,
in a sense made precise in the following theorem of Gagliardo [24].
4.8 THEOREM (Gagliardo [24]) Let R be a bounded domain in 08”

having the cone property. For each p > 0 there exists a finite collection
{R,, R,, . .., Rnt} of open subsets of R such that R = Uj”=,Q j , and such that
aj
to each R, there corresponds a subset A j of having diameter not exceeding p,
and an open parallelepiped Pj with one vertex at 0, such that Qj =
UxEA,(x+Pj). Moreover, if p is sufficiently small, then each Rj has the strong
local Lipschitz property.
PROOF Let Co be a finite cone with vertex at 0 such that any x E R is the
vertex of a finite cone C, c R congruent to C,,. It is clearly possible to select
a finite number of finite cones C , , ..., C, each having vertex at 0 (and each
having aperture angle smaller than that of C,) such that any finite cone
congruent to C, and having vertex at 0 must contain one of the cones C j ,
1 I j I k . For each Cj let Pj be an open parallelepiped with one vertex at the
origin and such that Pj c C j . Then for each x E R there exists j , 1 ~j Ik,
such that
x + Pj c x + cj c c, c R.
-
Since R is open and x + Pj is compact, it follows that y + Pj c R for all y
sufficiently close to x. Hence for every x E R we can find y E R such that
x E y + Pj c R for somej, 1 I j I k. (Any domain with the cone property can
therefore be expressed as a union of translates of finitely many parallelepipeds.)
Let -
zj= { x ~ n : x + cP R}.~ If d i a m x j s p for eachj, we take m = k,
set A j = A j and R j = U X E A , ( xPj),
+ and note that the first part of the theorem
xj
is proved. Otherwise we decompose into a finite union of sets A, such that
diam A j i Ip, set corresponding Pji = Pi, rearrange the totality of sets Aji
into a finite sequence A,, ..., A,, rename the corresponding Pj;s as P,, ..., P,,
and finally set R j = U x s A i ( x +Pj) to achieve the same end. (Figure 2 attempts
to illustrate these notions for the case
R = { ( x , y ) E 08, : O < l X l < 1, 0 < y < l},
c, = {(x,y) E RZ : x > 0,y > o , x , + y 2 < )},

p = 13/16.
In this case R can be covered by as few as four open subsets R, corresponding

to only two distinct parallelepipeds.)
It remains to be shown that if p is sufficiently small, then each R, has the
strong local Lipschitz property. For simplicity of notation we suppose,
therefore, that R = U x E A ( x P+ ) , where diam A I p and Pis a fixed parallele-
piped, and we show that R has the strong local Lipschitz property.
GEOMETRICAL PROPERTIES OF DOMAINS 69
t
@
P, aiid P,
f'
FIG.2
For each vertex uj of P let Q j = ( y = uj+ h ( x - v j ) :x E P, 2 > 0 } be the

infinite pyramid with vertex uj generated by P . Then P = Q j , the inter-
section being taken over all 2" vertices of P . Let R,j, = U X E A ( x + Q j )Let .
6 = dist (center of P, bdry P ) and let B be an arbitrary ball of radius n = 6/2.
For any fixed x E R, B cannot intersect opposite faces of x + P so we may pick
a vertex uj of P with the property that x+uj is common to all faces of x + P
that meet B, if any such faces exist. Then B n ( x + P ) = B n ( x + Q j ) . Now
+
let x, y E A and suppose B could intersect relatively opposite faces of x P and
y + P , that is, there exist points a and b on opposite faces of P such that
x + a E B and y + b E B. Then
p 2 dist(x,y) = dist(x+b, y+b)
2 dist(x+b, x+a) - dist(x+a, y+b)
2 26 - 20 = 6.
It follows that if p < 6, then B cannot meet relatively opposite faces of x + P
and y + P for any x,y E A. Thus B n ( x + P ) = B n ( x + Q j ) for some fixedj
independent of x E A, whence B n R = B n a,, .
Choose coordinates 5 = (t‘,5,) = (tl, ...,t,,- g,,) in B so that the 5, axis
lies in the direction of the vector from the center of P to the point v j . Then
( x + Q,) n B is specified in B by an inequality of the form c,, <fx(t’),where
fx satisfies a Lipschitz condition with constant independent of x . Thus
R,,, n B, and hence R n B, is specified by t,, <where I( <
f(t’) = ’)
sup,,,f,(t’) is itself a Lipschitz continuous function. Since this can be done
for a neighborhood B of any point on bdryR it follows that R has the strong
local Lipschitz property. I
Interpolation Inequalities for Intermediate Derivatives

4.9 We consider the problem of determining upper bounds for Lp-norms
of derivatives DBu,IbISm, of functions U E Wmn,p(R), in terms of the LP-
norms of u and its derivatives D“u of order lal=m. Such interpolation in-
equalities have been obtained by many writers including Ehrling [23],
Nirenberg [53, 541, Browder [ l l , 123, and Gagliardo [24,25], and are
amenable to numerous generalizations. Extensions of the definition of
WmmP(R) to cover the case of nonintegral values of m can be carried out (see
Chapter VII) via suitable interpolation arguments.
It is convenient to begin with a straightforward one-dimensional inter-
polation inequality which nevertheless typifies and provides a basis for the
proof of the more general theorems which follow.
4.10 LEMMA Let - 0 0 < a < b < m,let 1 < p < 00, and let 0 < E~ < 00.
There exists a finite constant K = K(&,,p,b- a), depending continuously on
b-a for 0 < 6-a I 00, such that for every E satisfying 0 < E I e0, and for
every functionftwice continuously differentiable on the open interval (a,b)
~ b l Y ( t ) l p d<
t KE l b l Y ( r ) l ’ d t + KE-’ l bI . f ( t ) l Pdt . (1)
Moreover, if b-a = 00, then K = K(p) can be found so that (1) holds for
every positive number e.
INTERPOLATION INEQUALITIES FOR INTERMEDIATE DERIVATIVES 71
PROOF It is sufficient to prove (1) for real-valued functionsx for, assuming

this done and writing arbitraryfin the form f = u+iu with u, u real valued,
we obtain
+
Im a ~ ( l , 2 ( ~ -a ~ ) ~ ~ ) ~ lu’(t)lP]
~ [ ~ u ’dt( r ) ~ ~
We may also assume, without loss of generality, that .so = 1, for, assuming
the lemma proved in this case, we obtain from (l), since 0 < c/eo I1,
lblf’ct)l’dt 5 K.(E/E~)
II I.p(r)lpddt + ~.(q,/e) I’ If(t)lpdt.
This, in turn, implies (1) with K = K ( e o , p , b - a ) = K ( I , p , b - a ) max(cO,e;l).

We assume, therefore, thatfis real valued and e0 = 1. For the moment we
suppose also that a = 0 and b = 1. If 0 < 5 < 3 and 5 < q < 1, then there
exists A E (5, q) such that
It follows that for any x E (0,l)
I3 V(OI 1’
+ 3 I ~ W+I 0 If”(t)l dt.
Integrating the above inequality with respect to over (O,+) and with respect
to q over (3, I), we obtain
whence, by Holder’s inequality,

Hence,
where Kp = 2p-1.9p.It follows by a change of variable that for any finite

interval (a, b)
+ Kp(b-a)-pllf(r)lpdr.
b
lblf’(r)lpdrI K p ( b - a ) P b If”(t)lPdt (2)
Since 0 < E I 1 there exists a positive integer n such that

+F?JP I l/n I ~ “ p .
Setting aj = a+(b-a)j/n for j = 0,1,...,n, we obtain from (2), noting that
aj-aj-l = (b-a)/n,
where R ( p , b - a ) = Kpmax[(b-a)P,2P(b-a)-P].
Now let
i
max R ( p , s ) if b - a 2 1
1 sss2
K(l,p, b-a) =
max R ( p , s ) if 0 c b - a < 1.
b-ass52
Then K ( l , p ,b-a) is finite for 0 c b-a I GO and depends continuously on

b-a. For b-a c 1, (1) follows directly from (3). For 1 I b-a I co, the
interval (a, b) may be partitioned into (possibly infinitely many) subintervals
each of length between 1 and 2, whence (1) follows upon summing (3) applied
to each of these subintervals.
Finally, suppose that b-a = 00. T o be specific we assume a is finite and
b = co, the other possibilities being similar. For given E > 0 let aj = a+jE1/P,
j = 0, 1,2, .. . . Then aj - aj- = d i pand we have, using (2),
which is (1) with K = Kpdepending only on p . I

4.11 For 1 I p < co and for integers j , 0 5 j s m,we introduce functionals

1.1 j , p on W"*P(R) as follows:
Clearly, lu(o,p= l)ullo,p= llullp is the norm of u in Lp(R) and
If j 2 1, 1.1 j , p is a seminorm-it has all the properties of a norm except that

( U J ~=, 0~ does not imply that u vanishes in Wrn.p(R);for instance, u may be a
nonzero constant on a domain R having finite volume. Under certain circum-
stances which we investigate later, 1 is an equivalent norm for the space
WFP(R).In particular this is so if R is bounded.
At the moment we are concerned with establishing interpolation in-
equalities of the form
+ K & - j / ( ' " - j140,
5 K~(ul,,~ ) P' (4)
where 0 s j Im - 1. The following lemma shows that in general we need only
establish (4) for the special casej = 1, m = 2, a reduction that will be used in
the three interpolation theorems that follow.
4.12 LEMMA Let 0 < do < co, let m 2 2, and let

E~ = min(bo, do2, ...,6z-l).
Suppose that for given p , 1 5 p < co,and given R c R" there exists a constant
K = K(d0,p,R) such that for every finite 6, 0 < 6 s do, and for every
u E W2*p(R),we have
IK61u12,p + K~-'l40,p. (5)
Then there exists a constant K = K ( ~ ~ , m , p ,such
n ) that for every finite E,
0 < E I e0, every integer j, 0 I j I m- 1, and every u E W".P(R),we have
I K E J U ~+
, , K, E~- j / ' " - j ) I40,P ' (6)
PROOF Since (6) is obvious f o r j = 0 we consider only the case 1 ~j I m - 1.

The proof is accomplished by double induction on WI and j . The constants
K , , K2,.. . appearing in the argument may depend on So (or E ~ ) m,
, p , and R.
We first prove (6) for j = m- 1 by induction on m, so that ( 5 ) is the special
case m = 2. Assume, therefore, that for some k, 2 Ik Im- 1,
(ulk-l,p 5 K,dIU(k,p+ K,d-(k-l) l 4 O . P (7)
holds for all 6, 0 < 6 5 do, and all u E Wk.p(Cl).If u E Wk+l,p(R),
we prove
that (7) also holds with k+1 replacing k (and a different constant K l ) . If
ILY~ = k- 1, we obtain from ( 5 )
+
IK,61D"ul*,p K , 6 - 1 1 D " u l o , p .
Combining this inequality with (7), we obtain, for 0 < q I6,,
1'lk.p 5 K3 c
Ial=k-l
IDaull,p
IK 4 6 1 u l k + l , p + K48-11ulk-l,p
5 K461ulk+l,p + K4K16-1?Iulk,p + K4K16-'q1-kIU10,p.
We may assume without prejudice that 2K1K4 2 1. Hence we may take
q = 6 / 2 K , K4 and so obtain
Iulk,p 5 2K461ulk+l,p + (6/2KlK4)-klu10,p
I K561ulk+l,p + K56-kIuI0,p*
This completes the induction establishing (7) for 0 < 6 5 6 , and hence (6)
for j = m-1 and 0 < E I d , .
We now prove by downward induction on j that
I'/j.p 5 K66m-jlulm,p -k K68-jluI0,p (8)
holds for 1 I . j Im - 1 and 0 < 6 I So. Note that (7) with k = m is the
special case j = m - 1 of (8). Assume, therefore, that (8) holds for some j ,
2 Ij I m - 1. We prove that it also holds with j replaced by j - I (and a
different constant &). From (7) and (8) we obtain
Iulj-1,p K7Jlulj,p +K781-il~lo,p
< K 7 6{ K 6 a m - j
- lu(,,p+&d-j lulO,p} + K 7 6 l-j IuIO,p
I ~ , ~ m - ( j - l ) +~ , 6 - ( j - l )
Iulm,p Iul0,p.
Thus (8) holds, and (6) follows by setting S = ~ ' ' ( " ' - j ) in (8) and noting that
E I E, if 6 I6,. 1
4.13 THEOREM There exists a constant K = K ( m , p ,n) such that for any
R any E > 0, any integer j , 0 Ij 5 m - 1, and any u E Wz*p(R),
c R",
+
IK E I U J , , , ~K e - j ' ( m - j ) I40 , P' (9)
PROOF By Lemma 3.22 the operator of zero extension outside R is an iso-

metric isomorphism of W$.p(R) into Wm-P(R"). Thus it is sufficient to establish
(9) for R = R". Also, by Lemma 4.12 we need consider only the case j = 1,
m = 2. (The case j = 0 m = 1 is trivial.) For any E > 0 and any E COw(R")
empty intersection, (1 1) and (12) imply that

m
lul!,p,~ II ~ l f , p , n o+ C
j = 1 IuI!,p,ujnn
m ’I
5 ( R+ 1) m a w , 3 Kz)bPlul%,r.I, n+ E - p I48,p . n>

and this inequality yields (10) (case j = 1, m = 2) on taking pth roots. It
remains therefore to verify the validity of (1 1) and (12).
If u E Cm(R)n W2*P(R),we apply Lemma 4.10 to u considered as a
function of xk on the interval from ~,6/2J;; to (,I,1)6/2fi,+ and then
integrate the remaining variables over similar intervals to obtain, for any
HA c 0,
where K3 depends only on p and the length of a side of H A(which in turn

depends on R via 6 and n). Summing (13) for 1 Ik In, we obtain
IUl!,p,H* + nK3&-Plu16,p,Ha.
5 K3&P144,p,Ha (14)
Since the cubes H do not overlap, we sum (14) for all cubes H Ac R and
obtain (1 1) with K, = nK3. Since Cm(R)n W2’P(R)is dense in W 2 g P ( R(1 ) ,1)
holds for all u E W2*p(R).
The constant K , in (12) will turn out to depend only on p, M and the
dimensions of the cone Cj (see Section 4.4). Anticipating this, and noting that
these dimensions are specified by the single cone C to which all cones Cj are
congruent, we drop, for simplicity, all subscripts j in considering (12). Let 5
be a unit vector in a direction in C, and let R, = { y + tt :y E R n U,0 5 t 5 h},
where h is the height of the cone C (Fig. 3). Thus (R n U ) c R, c Q by con-
dition (iii) of the uniform cone property. Any line L parallel to 5 either has
empty intersection with R, or else intersects R, in an interval of length p,
where h I p I h+ diam U 5 h+ M by condition (i) of the uniform cone
property. By Lemma 4.10, if u E Cm(R)n W2,P(R),
Ln*c
ID,ulpds I K4cP
Lnnc
+
IDt2uIPds K 4 ~ - P
L n c
IuIpds, (15)
where D, denotes differentiation in the direction of 5 and where K4 can be

chosen to depend only on p, h, and M, that is, on p and R. We now integrate
we obtain from Lemma 4.10
Integrating the remaining components of x, we are led to

IIDj4II; IKEpIIDj2411;+ K&-pl1411;,
whence
The case j = 1 , m = 2 of (9) now follows by takingpth roots and noting that
COw(Iwn)is dense in Wm*p(Iw").I
4.14 THEOREM (Ehrling [23], Nirenberg [53], Gagliardo [24]) Let

R c Iw" have the uniform cone property (Section 4.4), and let E, be a finite,
positive number. Then there exists a constant K = K(E,,m,p, R) such that
for any E , 0 < E IE,, any integer j , 0 Ij Im- 1 , and any u E WmVp(R)
+
Iulj,p IK E I U ~ , , ~Ke-j'(m-j)lulO,p. (10)
PROOF The case m = 1 is trivial; again Lemma 4.12 shows that it is sufficient
to establish (10) f o r j = 1, m = 2. In addition, the argument used in the second
paragraph of the proof of Lemma 4.10 shows that we may assume c0 = 1 .
In this proof we make constant use of the notations of Section 4.4 describing
the uniform cone property possessed by R. If 6 is the constant of condition (ii)
of that section and if A = (Al, ..., A,,) is an n-tuple of integers, we consider the
cube
H A = {x E R" : nk6/2& 5 xk 5 (Ak+1)6/2&}.
Then Iwn = (J1 H Aand diam H A= 612. Let Ro = (JHlcnH A .Thus R R, c
Ro c R. If the sets U , , U 2 ,... and Q l , Q2, ... are as in Section 4.4, then
-
R =
W
U (Uj n R) u R,
j=1
= uW
j =1
Qj u R,.
We shall prove that for any u E W2.p(R)

I47,p,n0 5 K , E P 1 4 R p , R o + Kl E - p 14Pg,p,no (1 1)
a n d f o r j = 1,2,3, ...
IUIf,p,UjnC2 5 KZEpIuI$,p,Qj + KZE-PIUIPg,p,Q,, (12)
where K2 is independent of j . Since any R + 2 of the sets R,, Q,, Qz, ... have
FIG.3
(1 5) over the projection of R, on a hyperplane perpendicular to ( and so obtain

Now let tly..., 5. be a basis of unit vectors in R" each lying in a direction
contained in the cone C. For 1 Ik In, Dku(x)= ~.;=lujDt,u(x) where
the constants ai satisfy lujl I1/V, 1 ~j In, V being the volume of the
,,
parallelepiped spanned by t ..., t". (The reader might verify this assertion.
It is a simple exercise in linear algebra.) A lower bound for V can be specified
in terms of the aperture angle of the cone C-that is to say, the basis tl, ..., 5, ,
which varies with the covering patch U, may always be chosen in such a way
that Vis independent of U . It now follows from (16) that
IK6eP14Z,p,Q + K6e-Plul~,p,*.
The desired inequality (12) now follows by summing on k and using the
density of Cm(R)n W2*P(R)in W2*p(R). I
IfRis bounded, the above theorem can be proved under weaker hypotheses.
4.15 THEOREM Let R be a bounded domain in R" having the cone

property. Then the conclusion of Theorem 4.14 holds for Q.
PROOF By Theorem 4.8 there exists a finite collection {Q1, ...,R,} of open
subsets of R such that R = u,"= Qj and such that each Rj is a union of trans-
lates of some open parallelepiped. It is clearly sufficient to prove an inequality
analogous to (11) for each Rj. We may therefore assume without loss of
generality that R = U x s A ( x +P),where A is a bounded set in R" and P is an
open parallelepiped with one vertex at the origin. Let tl, . .., 5. be unit vectors
in the directions of the n edges of P that concur in the origin, and let 1 be the
minimum length of these edges. Then the intersection with R of any line L
parallel to one of the vectors t j is either empty or a finite collection of segments
each having length between 1 and diamf2. It follows as in (16) that
for smooth functions u. Since {tl,t2,...,5,) is a basis for R" one can now
show, by an argument similar to that following (16), that
lull.p,n 5 ~ 2 4 u I Z , p ,+nK2&-PIu19,p.n
holds for all u E W2,p(R). I
INTERPOLATION INEQUALITIES INVOLVING COMPACT SUBDOMAINS 79
4.16 COROLLARY The functional ((-)),, p,n defined by
( ( u ) ) m , p , n = {IuIg,p,n + IuIE,p,cJ"p
is a norm, equivalent to the usual norm 11. (Im, p , n , on each of the following
spaces :
(i) W ~ ~ p ( f for
2 ) any domain R,
(ii) Wmsp(R)for any domain R having the uniform cone property,
(iii) W m * p ( Rfor
) any bounded domain R having the cone property.
4.17 THEOREM (Ehrling [23], Browder [12]) If c R"has the uniform

cone property or if it is bounded and has the cone property, and if 1 Ip < 03,
then there exists a constant K = K ( m , p , R ) such that for 0 Ij I m and any
u E W"l*P(R),
I1ullj.r 5 Kll~lli&l~lL:;i)/m* (17)

In addition, (17) is valid for all u E Wg*p(R)with a constant K = K ( m , p ,n)
independent of R.
PROOF Inequality (17) is obvious if eitherj = 0 o r j = m. For 0 c j c m we

can obtain from successive applications of (10) that
IIullj.p I K1EIIullm.p + K ~ Ej / (-m- i)Ilullo.p (18)
holds for all E, 0 < E 5 1, and all u E W".p(R) with K , depending only on m,
p , and R. (By Theorem 4.13 the same inequality holds for all u E Wt*p(R)
with K , depending only on m, p , and n.) Inequality (17) now follows for u # 0
if we set E = ( ~ ~ ~ ~ ~ o , p / ~ ~ in ~ m ,1
u ~(18). p)(m-~)~m
We remark that (17) also implies (I 8) algebraically : specifically, set

p = m/j, p' = m / ( m - j ) , a = ( E ~ ~ u ~ ~ , , J iand
~ " ' ,b = ~ - ~ / ~ ~ ~ u l l &in~ ;in-
~)/"'
equality (4) of Chapter 11,
ab I (aP/p)+ (bP'/p'), (UP)+ (UP') = 1,
to show that the right side of (17) does not exceed the right side of (18).
Interpolation Inequalities Involving Compact Subdomains
4.18 Upper bounds for the Lp(R)-norms of intermediate derivatives DBu,

Im- 1, of a function u E Wm.P(R)can be expressed in terms of the semi-
norm I U ~ , , ~ , and
~ the LP-norm of u over a suitable subdomain whose closure
is a compact subset of the bounded domain R. We establish some hybrid
interpolation inequalities of this sort by methods that follow somewhat the

same lines as those used for the interpolation inequalities derived above.
(See, for instance, the work of Agmon [S].)
4.19 LEMMA Let (a, b) be a finite open interval in R and let 1 Ip < co.
There exists a finite constant K = K ( p , b-a) and, for every positive number E ,
a number 6 = 6( E , b - a) satisfying 0 < 26 < b -a such that every continuously
differentiable functionfon (a,b) satisfies
p-(r)lPdt 5 K E p ( r ) l ’ d t + .S,:,9.l)lPdt. (19)
Moreover, fixed values of K and 6, independent of b -a, can be chosen so that

(19) holds for all intervals (a, b) whose lengths lie between fixed positive
bounds: 0 < Il 5 b - a 5 l2 00.-=
PROOF The proof is similar to that of Lemma 4.10. Suppose for the moment
that a = 0 and b = 1, and let 3 < q < 3. If 0 < x < 1, then
Integrating q over (3, +), we are led to
If(x)l I 3 I;rlf(v)l + l1
0 If’(Oldt,
so that by Holder’s inequality if p > I ,
Integrating x over (0, l), we obtain
where Kp = 3.2p-1. The change of variable a + t ( b - a ) + t now yields, for

any finite interval (a, b),
b-(b-a)/3
lblf(r)lpdt +
5 K ~ ( ~ - ~ ) P ~ I Y ( o I wa + ( b - a ) / 3 I ~ ( o I P ~ ~ .
For given E > 0 pick a positive integer n such that n - p I E . Let uj =

+
a (b -u)j/n for j = 0, 1, ...,n and pick 6 so that 0 < 6 I (b -a)/3n. Then
equation continues
INTERPOLATION INEQUALITIES INVOLVING COMPACT SUBDOMAINS 81
which is the desired inequality. a

The reader may convince himself that Lemma 4.19, unlike Lemma 4.10,
does not extend to infinite intervals (a,b) if the intention is that the second
integral on the right side of (19) should be taken over a compact subinterval.
4.20 THEOREM Let R be a bounded domain in 08” that has the segment
property. Then there exists a constant K = K ( p ,0)such that to any positive
number E there corresponds a domain Re c c 32 such that
holds for every u E W1ip(R).
PROOF The proof is similar to that of Theorem 4.14. Since R is bounded the
locally finite open cover { Uj}of bdryR and the corresponding set { y j } of
nonzero vectors referred to in the description of the segment property (Section
4.2) are both finite sets. Therefore open sets 5.c c Ujcan be found so that
bdry R c uj 5..(See the first part of the proof of Theorem 3.14.) Moreover,
for some 6 > 0, R6= {x E R :dist(x, bdry R) < 6) c U j“v; so that we may
write R = u,
that for each j
(5n Q) u fi, where fi c c R. It is thus sufficient to prove
lulPo,p,Y,nn K ~ & P I ~ l LK1IUlPo,p,n.,,

n+
for some Re,j c c R. For simplicity we drop all subscripts j .
Consider the sets Q, Q,, 0 I q c 1, defined by
Q = { x + t y : x ~U n Q , O < t < l } ,
Q, = { x + t y : x E V n R , q < t < l } .
If q > 0, Q, c c Q. By the segment property, Q c R and any line L parallel

to y and passing through a point of Y n R intersects Qo in one or more
intervals each having length between lyl and diamR. By Lemma 4.19 there
exists q > 0 and a constant K such that for any u E Cm(R)and any such line L
lu(x)lpds K1eP +
L n Q o I D , u ( ~ ) ( ~ d s K1 LnQql~(x)lpds,
LnQo
Dydenoting differentiation in the direction of y. We integrate this inequality

over the projection of Q, on a hyperplane perpendicular to y and so obtain
lul&p;nnV 5 lul!,p,Qo 5 + Kllul&p,Q,
KlePlulf,p,Qo
5 K1&plulf.p,n+ K1I4!,p,n,,
where Re = R, c c R. By density this inequality holds for any u E W1*p(R). I
4.21 COROLLARY The conclusion of Theorem 4.20 is also valid if R is
bounded and has the cone property.
PROOF As remarked earlier, a domain R with the cone property need not
have the segment'property. By Theorem 4.8, however, i2 is a finite union of
domains having the strong local Lipschitz property. We leave it to the reader
to show that a bounded domain with the strong local Lipschitz property has
the segment property, and thus complete the proof. I
4.22 LEMMA Let Ro,R be domains in R" with R, c c R. Then there
exists a domain R' having the cone property such that R, c R' c c R.
PROOF Since Q, is a compact subset of R there exists 6 > 0 such that

>
dist (a,, bdry R) 6. The domain R' = { y E R" : 1 y - X I < 6 for some x E R,}
clearly has the desired properties.
4.23 THEROEM Let R be a bounded domain in R" having either the

segment property or the cone property. Let 0 < E, < 03, let 1 < p < 03, and
let j and m be integers with 0 5j 5 m- 1. Then there exists a constant
K = K(E,,m,p, R), and for each E, 0 < E I E,, a domain Re such that Re c c R
and such that for every u E Wmip(R)
PROOF We apply Theorem 4.20 or its corollary to derivatives Db,

= m- 1, to obtain
lulm-1,p.n I KlEIuJm,p,n + K1IUlm-l,p,ne, (22)

where Re c c R. By Lemma 4.22 we may assume that Re has the cone
property. For 0 < E IE,, we have by Theorem 4.15
Iulm-l,p,nc 5 KzEIulm,p,nc + K ~ E
-(m-l)
I40 , p , n. * (23)
Combining (22) and (231, we get the case j = m- 1 of (21). We complete the
proof by downward induction on j . Assuming (21) holds for some j 2 1 and
replacing E by ~ ' " - j (with consequent alteration of K and Re), we obtain
IuIj,p,n K~p-'IuIm,p,n+ K ~ E - ' I U I O , ~ , ~ ; * (24)
EXTENSION THEOREMS 83
Also by (21) withj and m replaced byj- 1 andj, respectively (the case already
proved), we have
IUIj - 1, p . n &E IuIj , p . n + K4E-(j- ')I uI0 ,p , n,-

* (25)
Combining (24) and (25), we get
Iulj-l.p,n <
- K5 p - U - 1 ) IUlm,p,n + K ~- UE- U Iulo,p,n.,
where K5 = K4(K3+ 1) and Re = Re' u R:. Replacing E by ~'/("-j+')9 we
complete the induction. 1
Extension Theorems
4.24 Let R be a domain in R". For given m a n d p a linear operator Emapping

W"'*P(R) into W"*P(R")is called a simple (m,p)-extension operator for R
provided there exists a constant K = K(m,p) such that for every u E W"'*p(R)
the following conditions hold :
(i) Eu(x) = u ( x ) a.e. in 0,
(ii) IIEUlln,p,pn s K I I ~ l l m , p , n -
E is called a strong m-extension operator for R if E is a linear operator mapping
functions defined a.e. in R into functions defined a.e. in R" and if for every p,
-=
I 5 p 00, and every k, 0 5 k s m, the restriction of E to Wk*p(R)is a
simple (k,p)-extension operator for R. Finally, E is called a total extension
operutor for R provided E is a strong m-extension operator for R, for every m.
4.25 The existence of even a simple (m,p)-extension operator for a domain

R guarantees that W"*p(f2)inherits many properties possessed by W"*p(R").
For instance, if the imbedding W m V p ( W
+)L4(R") is known to hold, then the
imbedding W"*p(R) Lq(R) follows via the chain of inequalities
llUllo,q,n lIEu110,q,~ KI IIEUllm,p,wn 5 K1KIIuIIm,p,n*

We shall not, however, use this technique to prove the Sobolev imbedding
theorem in Chapter V as we shall prove that theorem under rather weaker
hypothesis on R than are needed to guarantee the existence of an (m,p)-
extension operator.
We shall construct extension operators of each of the three types defined
above. The method used in two of these constructions is based on successive
reflections in smooth boundaries. It is attributed to Lichenstein [35] and later
Hestenes [31] and Seeley [61]. The third construction, due to Calder6n [14]
involves the use of the Calderbn-Zygmund theory of singular integrals. It is

less transparent than the reflections method and yields a weaker result, but
requires much less regularity of the domain R. Except for very simple domains
all of the constructions require the use of partitions of unity subordinate to
open covers of bdryR chosen in such a way that the functions in the partition
have uniformly bounded derivatives. Because of this, domains with bounded
boundaries (both exterior domains and bounded domains) are more likely to
be easily seen to satisfy the conditions of our extension theorems. Exceptions
are half-spaces, quadrants, etc., and smooth images of these.
4.26 THEOREM Let R be either

(i) a half-space in R", or
(ii) a domain in R" having the uniform C"-regularity property, and also
having a bounded boundary.
For any positive integer m there exists a strong m-extension operator E for R.
Moreover, if a and y are multi-indices with IyI I IaJI m,there exists a linear
operator Eay continuous from W j S P ( R into) W j ? P ( R " )for 1 s j 5 m - ItlJ
such that if u E Wlal*p(f2),
then
D"Eu(x) = 1 Ea,Dyu(x) a.e. in R".
IYI5 I4
r)
PROOF First let R be the half-space R+" = { x E R" :x,, > O } . For functions u
defined a.e. on R," we define extensions Eu and Eau, la1 I m, a.e. on R" via
x, > 0
Eu(x) = m+l
C I j u ( x l ,..., x,-
[j= 1
1, -jx,,) if x,, I 0,
(27)
if x, > 0
C (-j>..Aju(xl, ...,x,,-
j=l
-jx,,) if x, s 0,
where the coefficients I , , are the unique solutions of the
(m+ 1) x (m+ 1) system of linear equations
m+l
C (-j)'Aj
j=1
= 1, k = O , l , ..., m.
If u E C m ( c ) then
, it is readily checked that Eu E Cm(R")and
P E u ( x ) = EaDau(x), la1 s m.
Thus
I K(m,p,u)~+"lD'u(x)lpdx.
The above inequality extends, by virtue of Theorem 3.18, to functions
u E W k * p ( R + "m) ,2 k 2 Iu1. Hence Eis a strongm-extension operator for R+".
Since DBE,u(x)= E a + a ~ ( a~ similar
), calculation shows that E, is a strong
(m - lul)-extension. Thus the theorem is proved for half-spaces (with
E,, = E,, Euy= 0 otherwise).
Now suppose R is uniformly C"-regular and has bounded boundary. Then
the open cover { U j } of bdryR, and the corresponding m-smooth maps Qj
from Uj onto B, referred to in Section 4.6, are finite collections, say 1 < j < N .
Let Q = {y E R" : ly'( = (x'jiiyj2)1/2< +,ly,,l< 4 / 2 1 . Then
{ yE R" : IyI < +} c Q t B = { y E R" : IyI < l},
By condition (i) of Section 4.6 the open sets 9'j = Y j ( Q ) ,1 < j I N , form an
open cover of Rd = { x E R : dist(x, bdry 0)< 6 ) for some S > 0. There exists
an open set Yoof R, bounded away from bdryR, such that R c (JiN_, 5..
By Theorem 3.14 we may find infinitely differentiable functions w,, w l , ..,,wN
such that suppwj c 5. x:j"=,
and w j ( x )= 1 for all x E R. (Note that suppw,
need not be compact if R is unbounded.)
Since R is uniformly C"-regular it has the segment property and so
restrictions to R of functions in Com(R")are dense in W k V P ( RIf) .4 E COm(Rn),
then #J agrees on R with the function xi"=, 4 j , where r$j = w j . 4 E Co"(q).
) ) . Il/i E COm(Q).We extend
F o r j 2 1 and y E B let ILj(y) = ~ $ ~ ( Y l ~ ( yThen
ijj to be identically zero outside Q. With E (and E,) defined as in (27), we have
E J / / E c ~ " < Q > , E $ j = $ j oQ+
n ={~~Q:yn>O},and
~ ~ E ~ j ~ ~ 5
k , Kp l, ~Q~ $ j ~ l k , p , Q + , 5 5 m?
where K, depends on k, m, and p. If Oj(x)= Ei,hj(Qj(x)),then 0, E C,"(Vj)
and Oj(x)= 4 j ( x )if x E R. It may be checked by induction that if IuI I m, then
where aj;,BE Cm-lal(T7j)

and b j i g yE C" -IB'(B)depend on the transformations
Qjand Y, = 0,:' and satisfy
By Theorem 3.35 we have for k I m,

llejllk,p,Rn,s K 2 IIE$jllk,p,Q K l KZ Il#jllk,p,Q+ I K3 ~ ~ # j ~ ~ k , p , t l ~
where K3 may be chosen to be independent of j. The operator ,!? defined by
&<x> = 4 O W +j = 1 cN
ej<x>
clearly satisfies ,!?4( x ) = 4 ( x ) if x E 0, and
II’%llk,p,R’’ 5 II40llk,p,n + K3 j = l c
N
~ ~ ~ j ~ ~ k , p , f l
5 K4(1 + N K 3 ) 114Ilk,p,tl,
where
K4 = max max sup lo”wj(x)l < ao.
O S j S N lalSmxeOP”
Thus ,!? is a strong m-extension operator for R. Also
if ci # y, and
4au(x) = (V‘WO)(X)+ c F
N
I=1 IS Sl.1
a j ;a j ( x ) [EB( b j ; pa * (v * w j ) 0 ‘ Y j > I (Qj (XI).
We note that if x E R, E,,u(x) = 0 for ci # y and Eaau(x)= v(x). Clearly Eay
is a linear operator. By the differentiability properties of uj;aband b j i S YEay , is
continuous on W i P p ( Rinto
) W j p p ( R n ) for 1 < j S m - I c i l . This completes
the proof.
The representation (26) for derivatives of extended functions was included

in the above theorem because it will be needed when we study fractional-order
spaces in Chapter VII.
The reflection technique used in the above proof can be modified to yield a
total extension operator for smoothly bounded domains. The proof, due to
Seeley [61], is based on the following lemma.
4.27 LEMMA There exists a real sequence {ak}p=o such that for every
nonnegative integer n we have
c
k=O
m
2nkak = (- I)”
and
PROOF For fixed N , let a k , N , k = 0,1,2, . .., N, be the solution of the system
of linear equations
N
In terms of the Vandermonde determinant
xo x 1 *-. XN
v(xO,x1, ..., X N ) = Xo2 XI2 ... XNZ
XoN XIN XNN
V(1,2,..., 2 k - 1 , - 1 , 2 k + l , ...)2 N )
=
ak,N
V(1,2,...,2")
= AkBk,N
where
it being understood that ny=.=,

Pi = 1 if 1 > m. Now
Also
where we have used the inequality log(1 + x ) < x valid for x > 0. I t follows
that the increasing sequence {Bk,N};=O converges to a limit Bk Ie4. Let
88 1v INTERPOLATION AND EXTENSION THEOREMS
Then for any n
Letting N tend to infinity in (31), we complete the proof. I

4.28 THEOREM Let R be either
(i) a half-space in R", or
(ii) a domain in R" having the uniform C"-regularity property for every m,
and also having a bounded boundary.
Then there exists a total extension operator for R.
PROOF It is sufficient to prove the theorem for the half-space R+"; the
proof for R satisfying (ii) then follows just as in Theorem 4.26.
The restrictions to R+" of functions 4 E Com(R")being dense in W"9p(R+")
for any m and p , we define the extension operator only on such functions. Let
f be a real-valued function, infinitely differentiable on [O,oo) and satisfying
where {ak} is the sequence constructed in the above lemma, and x' =
( x l , .. ., x,- ,). Then clearly €4 is well defined on R" since the sum in (32) has
only finitely many nonvanishing terms for any particular x E R-" =
{ X E Iw":x,<O}. Moreover, Ed has compact support and belongs to
Cm(lW,") n C m ( p ) . If x E R-", we have
Since JIk(x) = 0 when -x, > 1/2k-' it follows from (30) that the above series
converges absolutely and uniformly as x, tends to zero from below. Hence by
(29)
Thus E4 E Com(R"). Moreover, if 1.1 5 m,
where K1 depends on m,p , n, and.f. Thus
= KllUk12km(1/2k)
{ 1pIsm L + n
5 Kllak12k" Il4llrn,p,W+"*
It follows by (30) that

m
IIDaE4110,p,W-" 5 Kll1411m,p,R+n 1
k=O
2kmlakl Kt I l $ l l m , p , R + " .
Combining this with a similar (trivial) inequality for IIDaE41]0,p,R+,,

, we
obtain
I E4 Ilm, p, R" 5 K3 I14Ilm, p. R + "
with K3 = K3 (m,p,n). This completes the proof. I
4.29 The restriction that bdryn be bounded was imposed in Theorem 4.26
(and similarly in Theorem 4.28) so that the cover {T.} would be finite. This
finiteness was used in two places in the proof, first in asserting the existence of
the constant K4, and secondly in obtaining the last inequality in (28). This
latter use is, however, not essential for the proof, for, were the cover {q.} not
finite, (28) could still be obtained via the finite intersection property [Section
4.6, conidition ($1. Theorems 4.26 and 4.28 extend to any suitably regular
domains for which there exists a partition of unity { a j }subordinate to the
cover {q.} with D'wj bounded on R" uniformly in,jfor any given o(. The reader
may find it interesting to construct, by the above techniques, extension
operators for domains not covered by the above theorems, for example,
quadrants, strips, rectangular boxes, and smooth images of these.
It might be noted here that although the Calderbn extension theorem
(Theorem 4.32) is proved by methods quite different from the reflection
approach used above, nevertheless the proof does make use of a partition of
unity in the same way as does that of Theorem 4.26. Accordingly, the above
considerations also apply to it. The theorem is proved under a strengthened
form of uniform cone condition that reduces to the uniform cone condition of
Section 4.4 if R has bounded boundary.
Before proceeding to Calder6n's theorem we present two well-known
results on convolution operators that will be needed for the proof. The first
is a special case of a theorem of W. H. Young.
90 Iv INTERPOLATION A N D EXTENSION THEOREMS
4.30 THEOREM (Young) Let 1 Ip c co and suppose that u E L'(R")

and u E Lp(R").Then the convolution products
u * u(x) =
s.. u ( x - y ) u ( y ) dy,
are well defined and equal for almost all x

u * u(x)
E
=
s..
u ( x - y ) u ( y ) dy
R". Moreover, u * u E L p ( R n ) and

IIu*ullp llull1 I I 4 l P . (33)
PROOF The proof is a simple consequence of Fubini's theorem if p = 1, so

we assume 1 < p < co. Let w E LP'(n).Then
= IluII1 11~1lpIIwllp'.
Since w may be chosen so as to vanish nowhere, it follows that u * u ( x ) and
u * u ( x ) must be finite a.e. Moreover, the functional
F,,,(w) =
s.. u * U(X) w ( x ) dx
belongs to [Lp'(R")]'and so by Theorem 2.33 there exists A E Lp(R")with

l l 4 p llull1 llullp such that
s.. 1(x) w ( x ) dx =
s.. u * u(x) w ( x ) dx
for every w E Lp'(R").Hence 1 = u * u E LP(R")and (33) is proved. The

equality of u * u and u * u is elementary. I
The following theorem is a special case, suitable for our purposes, of a

well-known inequality of Calderbn and Zygmund [161 for convolutions
involving kernels with nonintegrable singularities. The proof, which is rather
lengthy and may be found in many sources (e.g., Stein and Weiss [SS]), is
omitted here. Neither the inequality nor the extension theorem based on it
will be required hereafter in this monograph.
Let BR = { x E R" : 1x1 5 R } , S , = { x E R" : 1x1 = R}, and let doR be the
area element [Lebesgue (n - I)-measure] on S, . A function g is said to be
-
homogeneous of degree p on B, ( 0 ) if g ( t x ) = t " g ( x ) for all X E B R {0}
and 0 < t 5 1.
N
4.31 THEOREM (The Calderdn-Zygmund inequality) Let
A x ) = G ( 4 1x1-",
where
(i) G is bounded on R" {0} and has compact support,
N
-
(ii) G is homogeneous of degree 0 on BR (0)for some R > 0, and
(iii) Jsn G(x) du, = 0.
If 1 c p c co and u E L p ( R " ) , then the principal value convolution integral
u * g ( x ) = lim
e+O+
/
W-B.
u ( x - . Y ) s ( r ) d.Y
exists for almost all x E R", and there exists a constant K = K ( G , p ) such that
for all such u
II~*SllP KII~llP~
Conversely, if G satisfies (i) and (ii) and if u * g exists for all u E Com(R"),then
G satisfies (iii).
4.32 THEOREM (The Calderdn extension theorem) Let R be a domain

in R" having the uniform cone property (Section 4.4) modified as follows:
(i) the open cover { V j }of bdryR is required to be finite ,and
(ii) the sets Uj are not required to be bounded.
Then for any m E {1,2, ...} and any p , I < p c a,there exists a simple (m,p)-
extension operator E = E(m,p) for R.
PROOF Let { U , , U,, ..., U,} be the open cover of bdry R given by the uniform
cone property, and let Uobe an open subset of R bounded away from bdryR
such that R c u;=,, U j . [Such Uo exists by condition (ii), Section 4.4.) Let
oo,ol, ...,w N be a C"-partition of unity for R with suppoi c U j . For
1 5 j < N we shall define operators Ej so that if U E W"*P(R), then E j u €
Wm*P(Rn) and satisfies
Eju = u in U j n R,
The desired extension operator is then clearly given by

u
We shall write x E R" in the polar coordinate form x = pa where p 2 0 and a

is a unit vector. Let C,, the cone associated with Uj in the description of the
-
uniform cone property, have vertex at 0. Let 4j be a function defined in
R" (0)and satisfying
(i) 4 j ( x ) 2 0 for all x # 0,
-
(ii) supp 4j c - cj u {O},
(iii) +j E Cm(R" {0}),
(iv) for some E > 0, (bj is homogeneous of degree m - n in Be - (0).
function $ j ( x ) = (a/8p)m[p"-1~j(~)] --
Now p " - 1 4 j is homogeneous of degree m- 1 2 0 on BE (0) and so the
vanishes on Be (0). Hence $ j ,
extended to be zero at x = 0, belongs to Cow(- Cj).Define
-L [$j(pa)u(y-pg) dp do (35)
where Js. da denotes integration over the unit sphere, and the constant Kj
will be determined shortly. If y E Uj n 0, then, assuming for the moment that
u E Cm(0),we have by condition (iii) of Section 4.4 that u ( y - p a ) is infinitely
differentiable for pa E supp 4 j . Now integration by parts rn times yields
Hence
Since (a/ap)"'-' [p"-'~$~(pa)] is homogeneous of degree zero near 0, the

above integral does not vanish if C$j is not identically zero. Hence Kjcan be
chosen so that E j u ( y ) = u ( y ) for y E Uj n R and all u E Cm(R).Since P ( R )
is dense in Wmn,p(R)we have E j u ( y ) = u ( y ) a.e. in U j n R for every
u E Wm*P(R).It remains, therefore, to show that (34) holds, that is, that
I/D"Eju It 0,p , 08" 5 JG I u I m, p . n
for any a, la( I;m.
The last integral in (35) is of the form 0, * u(y), where Oj(x) = $j(x)~x~l-".
Since Oj E L'(R") and has compact support we obtain via Young's theorem
4.30 and a suitable approximation of u by smooth functions,
* u)llo, p , 08" =
IIDa<ej llej * (Dau)I o , p , 08" I llejllo, 1, 08" IIDaullo,p,n.
It now remains to be shown that the first integral in (35) defines a bounded map
from W"'*p(R)into Wm*P(Rn). Since (a/dp)"' = &I= ,(m!/a!)oaDawe obtain
= C 5, * D'u,
-
where & = (- l)lal(m!/a!)aaC$j
1.1 =m
is homogeneous of degree m-n in Be {0}

and belongs to Cm(R" {O}). It is now clearly sufficient to show that for any
-
B, I B l l m
It D p ( t a * 0)110, p , 08" 5 Ka,p IIu It 0 , p . R . (36)
-
Tf IPlsm-1, then O p t ais homogeneous of degree not exceeding 1-n in
Be {0} and so belong to L' (W). Inequality (36) now follows by a Young's
theorem argument. Thus we need consider only the case 181= m, in which we
write Dp= (a/dxi)DYfor some y , IyI = m - 1, and some i, 1 I i I n. Suppose,
for the moment, that u E C,"(R). Then we may write
= lim
d+o+
J w"-Bd
oiU(x-y)Dyta(y) du.
We now integrate by parts in the last integral to free u and obtain Optaunder
the integral. The integrated term is a surface integral over the sphere S, of the
product of U ( X - .) and a function homogeneous of degree 1 - n near zero.
This surface integral must therefore tend to Ku ( x ) as 6 + 0 + ,for some constant

K. Noting that D,u(x-y) = -(a/8yr)u(x-y), we now have
n
Now 0 8 5 , is homogeneous of degree - n near the origin and so, by the last
assertion of Theorem 4.31, D8ta satisfies all the conditions for the singular
kernel g of that theorem. Since p > 1 we have for any v E LP(R)(regarded as
being identically zero outside a)
IIDB~a*~llo,p,~~
Ka,p II~IIo,p.n*
This completes the proof. I
V
Imbeddings of Wm9
p(Q)
The Sobolev Imbedding Theorem
5.1 It is primarily the imbedding characteristics of Sobolev spaces that render

these spaces so useful in analysis, especially in the study of differential and
integral operators. The most important of the imbedding properties of the
spaces W m * p ( Qare
) usually lumped together in a single theorem referred to
as the Sobolev imbedding theorem. The core results are due to Sobolev [63]
but our statement (Theorem 5.4) includes refinements due to others, in
particular to Morrey [47] and Gagliardo [24].
Most of the imbedding results hold for domains R in R" having the cone
property but otherwise unrestricted ; some imbeddings however require the
strong local Lipschitz property. Specifically no imbedding of W m . (0)P into
a space of uniformly continuous functions on R is possible under only the cone
property, as can be seen by considering the example given in the second
paragraph of Section 3.17.
5.2The Sobolev imbedding theorem asserts the existence of imbeddings of

into spaces of the following types:
WmqP(R)
(i) Wi*q(R),j I m, and in particular Lq(R).
(ii) CBi(R) = {u E Ci(R): Dau is bounded on R for lctl ~ j }This . space
is larger than C j ( Q in that its elements need not be uniformly continuous on
95
96 v IMBEDDINGS OF Wm"(n)
0,However, C,j(R) is a Banach space under the norm

1 u; C,J (a)11 = max sup IDau(x) I.
Oslalsj xen
(iii) Cj*'(fi)(see Section 1.27) and in particular Cj(n).

(iv) Wisq(Rk), and in particular Lq(Rk).Here Rkdenotes the intersection
of R with a k-dimensional plane in R", considered as a domain in Rk.
Since elements of W". P(R)are, strictly speaking, not functions defined
everywhere on R but rather equivalence classes of such functions defined and
equal up to sets of measure zero, we must clarify what is meant by an imbedding
of Wmpp(f2)into a space of type (ii)-(iv). In the case of (ii) or (iii) what is in-
tended is that the "equivalence class" u E Wm*p(R)should contain an element
belonging to the continuous function space that is target of the imbedding,
and bounded in that space by a constant times I I U ~ ~ ~ , Hence,
~ , ~ . for example,
Wm*p(R)-+ Cj(5i) means that each u E Wmn,p(R) can, when considered as a
function, be redefined on a set of zero measure in R in such a way that the
modified function 12 [which equals u in Wmn,p(!2)] belongs to Cj(n) and
satisfies 1 0; Ci@) 11 IK 1 u 1 ,, p , with K independent of u.
Even more care is necessary in interpreting the imbedding W"*p(R) -+
Wjiq(Rk) where k c n. Each element u E Wmn,p(R)is, by Theorem 3.16, a
limit in that space of a sequence {u,} of functions in Cm(R).The functions u,
have traces on Rkbelonging to Cm(Rk).The above imbedding signifies that
these traces converge in WjPq(Rk)to a function # satisfying IIU"llj,q,nk I
K I\u llm, p , with K independent of u.
Let us note as a point of interest (though of no use to us later) that the
imbedding Wm.p(R)+ WjVq(R)is equivalent to the simple containment
W"*P(R) c Wi.q(R). Certainly the former implies the latter. To see the con-
verse suppose W"*p(R) c Wj*q(R) and let I be the linear operator defined on
Wm*J'(R) into Wj*Q(R)by Zu = u. If u, -+ u in WmSP(R)[and hence in Lp(R)]
and Zu, + u in Wi*q(R)[and hence in Lq(R)], then, passing to a subsequence
if necessary, we have by Corollary 2.1 1 that u, (x) -+ u (x) a.e. in R and u, (x) =
Zu,(x) -+ u ( x ) a.e. in $2.Thus u ( x ) = u(x) a.e. in R, that is, lu = u, and Z is
continuous by the closed graph theorem of functional analysis.
5.3 Let R be a domain in R" having the cone property specified by a certain
finite cone C (see Section 4.3). C may be regarded as the intersection of an
infinite cone C* having the same vertex as C with a ball B centered at that
vertex. By the height of C we mean the radius of B. By the opening of C we mean
the surface area [(n- 1)-measure] of the intersection of C* with the sphere of
unit radius having center at the vertex of C. These geometric parameters are
clearly invariant under rigid transformations of C.
THE SOBOLEV IMBEDDING THEOREM
In asserting that an imbedding of the form

W"".(R) + x (1)
(where Xis a Banach space of functions defined over R) holds for R having the
cone property, it is intended that an imbedding constant for (l), that is, a
constant K for which the inequality
11% XI1 4 ~ I I ~ I l m * p . R
is satisfied for all u E Wmn,p(R),
can be chosen to depend on R only through the
dimension n and various such parameters of the cone C which are invariant
under rigid motions of C.
5.4 THEOREM (The Sobolev imbedding theorem) Let R be a domain in

R" and let Qk be the k-dimensional domain obtained by intersecting R with a
k-dimensional plane in R", 1 I k In. (Thus R" = n.) Let j and m be non-
negative integers and let p satisfy 1 I p < 00.
PART I If R has the cone property, then there exist the following imbeddings :
CASEA Suppose mp < n and n - m p < k I n. Then

wi+m.P(n) + wj,q(ak>, P 5 q I kp/(n-mp), (2)
and in particular,
wj+"l.p(R)-, Wj*q(R)), p I q I np/(n-mp), (3)
or
W"*P(Q) -b J!?(R), p I q I np/(n-mp). (4)
Moreover, if p = 1, so that m < n, imbedding ( 2 ) also exists for k = n - m .
CASEB Suppose mp = n. Then for each k, 1 Ik 5 n,

wj+m*P(n)+ Wjqzk), p 5q < 00, (5)
so that in particular
W"*P(R) -+ JY(R), p 5q < 00. (6)
Moreover, if p = 1 so that m = n, imbeddings ( 5 ) and (6) exist with q = 03 as
well; in fact,
w j + " * i ( n-
) , C*j(R). (7)
CASEC Suppose mp > n. Then

Wj+"~P(R)+ CB'(R).
98 v IMBEDDINGS OF W""(R)
PART II If R has the strong local Lipschitz property, then Case C of Part I
can be refined as follows :
CASEC' Suppose mp > n > (m- 1)p. Then

Wj+mp(o) -, cj.A(a), 0 < 1 I m - (n/p). (9)
CASEC" Suppose n = (in- 1)p. Then
Wj+m,p(R) -, cj.A(sz),
0 < A < 1. (10)
Also, if n = m- 1 and p = 1, then (10) holds for 1 = 1 as well.
PART I11 All the conclusions of Parts 1 and 11are valid for arbitrary domains
provided the W-spaces undergoing imbedding are replaced with the corre-
sponding Wo-spaces.
5.5 REMARKS (1) Imbeddings (2)-(8) are essentially due to Sobolev

[62,63] whose original proof did not, however, cover the cases q = kp/(n- mp)
in (2), or q = np/(n-mp) in (3) and (4). Imbeddings (9) and (10) find their
origins in the work of Morrey [47].
(2) Imbeddings of type (2) and (5) involving traces of functions on planes
of lower dimension can be extended in a reasonable manner to apply to traces
on more general smooth manifolds. For example, see Theorem 5.22.
(3) Part 111 of the theorem is an immediate consequence of Parts I and 11
applied to Iw" because, by Lemma 3.22, the operator of zero extension of
functions outside R maps Worn* isometrically into W'", p(R").
(4) Suppose that all the conclusions of the imbedding theorem have been
proven for R = R". It then follows that they must also hold for any domain R
satisfying the requirements of the Calder6n extension theorem 4.32. For
example, if Wm*P(Iw") + Lq(Iwn),and if E is an (m,p)-extension operator for R,
then for any u E Wm*p(SZ) we have
Ilullo,q,n 5 K1K2 II4rn.p.n
I I ~ u l I o , q ,5;~ ~K~IIEullrn,p,~n
with K , and K2independent of u. We shall not, however, prove the imbedding
theorem by such extension arguments.
(5) It is sufficient to establish each of the imbeddings (2), (3), (5), (7)-
(10) for the special case j = 0. For example, if Wm*p(R)+ Lq(R) has been
established, then for any u E Wj+m*p(R) we have DauE W"+P(Q) for JctJIj,
whence P u E Lq(R); thus u E Wj*q(R);and
PROOF OF THE IMBEDDING THEOREM 99
Accordingly, we will always specializej = 0 in the proofs.

(6) If Rk (or 0) has finite volume, it follows by Theorem 2.8 that im-
beddings (2)-(6) hold for 1 Iq < p in addition to the values of q asserted in
the theorem. It will be shown later (Section 6.38) that no imbedding ofthe form
Wm*P(R)+ Lq(R)where q < p is possible unless R has finite volume.
Proof of the Imbedding Theorem
5.6 The proof given here is due to Gagliardo [24]. Though it is rather lengthy,
the techniques involved are quite elementary, being based on little more than
simple calculus combined with astute applications of Holder's inequality.
Moreover, Gagliardo's proof establishes the imbedding theorem in the
greatest possible generality and is capable of generalization to produce
imbedding results for some domains not having the cane property (see
Theorems 5.35-5.37).
The proof is carried out in a chain of auxiliary lemmas. In each such lemma
constants K , , K 2 , ... appearing in the proof are allowed to depend on the same
parameters as the constant K referred to in the statement of the lemma.
5.7 LEMMA Let

R = ( x ~ R " : a , < x ~ <1 Ib i~S; n }
and
R' = {x' = ( x l , ...,x , , - ~ )E R"-' : ai < xi < b,; 15 i S n - 1 )
be bounded open rectangles in R" and Rn-', respectively. If a, < 5 < b, and
p 2 1, then for every u E Cm(R)n W l S p ( Rwe
) have
where K = K(p,b,-an). Thus the trace mapping u + u ( . , [ ) extends to an

imbedding of W1*P(R)into LP(R;-l),where Rr-' = R n { x E R":x,, = 4).
PROOF By Theorem 3.18, C"(ir> is dense in W1*I(R)so we may assume

u E C=(K). Thus jR'Iu(x', .)Ipdx' belongs to C"([a,,b,,]) and by the mean
value theorem for integrals we have
for some Q E [a,, b,]. Now
I s .+ '
Iu(x', ()IP = u(x', 0 ) D,,u(x', t ) dt
100 v IMBEDDINGS OF w""(R)
by Holder's inequality. Integration over R' leads to

~ ~ u ( ' , ~ ) ~ ~ 6 ,5
p , 2Rp '- 1 [ I I u ( ' , o ) l l g , p , R ' + ( b n - a n ) P - l llDnull$,p,R1
2 p - 1 [(bn-an)-' ~ ~ u ~ ~+ ~ (bn-an)P-'
, p , R llDnul16,p,Rl
which yields (1 1) with K = [2p-1 max((b,-a,)-l,(bn-an)P-')]l'P.We note

that K depends continuously on bn- a, but may tend to infinity if bn- a, tends
to zero or infinity. I
5.8 LEMMA Let R be as in the previous lemma. Then

W"*'(R)+ C(R).
The imbedding constant depends only on n and the dimensions of R.
PROOF Let x be any point of R, and let R' be as in the previous lemma. If
u E C"(K) and lac(5 n - 1 , we have by that lemma that
IIDau(',xn)IIO.lR' IK l ~ ~ D a u ~ ~ l , l , R -
Thus
~ ! ' ( * , x n ) ~ ~ n - l , l , R 'IK 2 ~ ~ u ~ ~ n , l , R
with K2 depending on b, - an. Iteration of this argument over successively

lower-diemnsional rectangles leads to
IIu( '3 X 2 , X39 **., 11 1,1, ((11, b1) 5 K 3 IIulln, 1, R
with K , depending on b j - a j , 2 I j5 n. By the mean value theorem for

integrals there exists Q E [ a l ,b,] such that
/Iu( '3 x21 1 0,1, (a1, b l ) = ( b l - al)l (a,X 2 r *. * ) xn)l *
Hence
5 [l / ( b l - all1 1 u ( * x29 *. *, xn) 11 0 , 1,

(al, b l )
+ llDl u ( * , x 2 , . . . , X n ) 1 1 0 , 1 , ( a 1 , b l )
KII4ln,,,R. (12)
Now suppose u E W"*'(R).By Theorem 3.18, u is the limit in W n * ' ( R )
of a sequence of functions belonging to C"(K). It follows from (12) that this
sequence converges uniformly on R to a function u" E C(K).Since u"(x)= u(x)
a.e. in R, the lemma is proved. I
We now turn our attention to more general domains. The following lemma
of Gagliardo, which is essentially combinatorial in nature, is the foundation
on which his proof of the imbedding theorem rests.
5.9 LEMMA Let R be a domain in Iw" where n 2 2. Let k be an integer

satisfying 1 Ik 5 n, and let K = ( K ~ , K ...,
~ , K,,) denote a k-tuple of integers
satisfying 1 Ix1 c K~ < ..+c K~ I n. Let's be the set of all (E) such k-tuples.
Also, given x E R", let x, denote the point (x,,, . . ., x,J E Rk;dx, = dxKl dx,, .
For given K E S let E, be the k-dimensional plane in R" spanned by the
coordinate axes corresponding to the components of x, :
E, = { x E R " : x i = O i f i $ K } ;
and for any set G c R"let G, be the projection of G onto E,; in particular
R, = { x E E, : 3y E R such that y , = x,}.
Let F, be a function depending on the k components of x, and belonging
to LA(R,),where ,I= (:I:). Then the function F defined on R by
44 = n
KES
F,(x,)
belongs to L1(Q),and llFlll,n I nreS

llF,llA,ns,that is,
PROOF For K E S and <,

E Rk let Q(<,) denote the k-dimensional plane
section of R by the plane x, = <,:
a(<,)= {x E R :xu = t,}.
We establish (13) by induction on n, and so consider first the case n = 2. We
may also suppose that k = 1 since, for any n, the subcase k = n of (13) is
trivial. For n = 2, k = 1, we have ,I= 1 and S has only two elements, K = 1
and K = 2. Hence
since clearly ( Q ( X ~ )c) ~Q2 for any xl. This is (13) for the case being con-
sidered. (A similar calculation will yield (1 3) for arbitrary n and k = 1.)
102 v IMBEDDINGS OF wrnsP(a)
Now we assume that (13) has been established for n = N - 1. We consider

the case n = N and, as noted above, may assume 2 I k I N- I . Thus
I = (;I:). Let p = (fl::) and v = (;I,”). The integrand on the left side of (13)
is a product of (f) factors IFKI each belonging to the corresponding space
LA(nK). Exactly of these factors, say those corresponding to K E A c S,
are independent of x, . It follows from applying the induction hypothesis over
the (N- I)-dimensional domain O(X,) and noting that ( R ( x , ) ) ~c R, that
The remaining (f)-(”;’) = A factors IF,,] depend on x N , and so when

restricted to n ( x N ) depend on only k- 1 variables. Applying the induction
hypothesis over Q(x,) again, but this time with k- 1 in place of k, we obtain
sn(XN)
n
KCS-A
..*dxN-l
IFK(XK)I%iX1
I KES-A [s *-kk-*
IFK(XK)lAdXK,
(n(xN)), I”’-
Now p+ v = I and so by Holder’s inequality, and (14) and (15),
(15)
-
Since S A contains I elements we obtain by (the several function form of)
Halder’s inequality that
It follows by insertion of (1 7) into (16) that
which completes the induction and the proof of (13). I

5.10 LEMMA Let R be a bounded domain in R" having the cone property.
-=
If 1 I p n, then W'*P(R) -,Lq(R), where q = np/(n-p). The imbedding
constant may be chosen to depend only on m, p , n, and the cone C determining
the cone property for R.
PROOF We must show that for any u E W1.P(R),
with K = K(m,p,n, C). By Theorem 4.8, R may be expressed as a union of

finitely many subdomains each of which has the strong local Lipschitz
property (and therefore the segment property), and each of which is itself a
union of parallel translates of a corresponding parallelepiped. A review of the
proof of that theorem shows that the number of subdomains and the dimen-
sions of the corresponding parallelepipeds depend on n and C. It is therefore
sufficient to establish (18) for one of these subdomains.
By Theorem 3.35 and a suitable nonsingular linear transformation we may
assume that the parallelepiped involved is, in fact, a cube Q having edge
length 2 units, and having edges parallel to the coordinate axes. Accordingly
we assume hereafter that R = U x E A ( x + Qwith ) A c R, and that R has the
segment property. By Theorem 3.18 it is sufficient to establish (18) for
uE cyn).
For X E O let wI(x) denote the intersection of R with the straight line
through x parallel to the xi coordinate axis. Clearly, wI(x) contains a segment
of unit length with one endpoint at x, say the segment x + t e i , 0 It < 1,
where e, is a unit vector along the xi-axis.
Let y = (np-p)/(n-p) so that y 2 1. Integration by parts gives, for
u E cyn),
61 lu(x + (1 -t)e,)lYdt
Let 2t = (xl, ..., xi- 1, xi+ , ..., x,) and set

F,(Q = sup Ju(y)JP'("--P).
yewr(x)
104 v IMBEDDINGS OF Wm'p(R)
Then (19) gives
If p > 1, then y > 1 and an application of Holder's inequality gives
since (y- 1)p' = q.

We now apply Lemma 5.9 to the functions F,,1 I i < n, noting that
k = n - 1 so that the exponent ;Z of that lemma is itself n - 1 :
< (2(p-1)/pYllull1,p,n llu118!$,*n)n/(n-1)*

-
-
Since (n 1)q/n- q/p' = 1, (18) follows by cancellation. The cancellation is
justified, for since u E C"(0) and R is bounded, ~ ~ u / ~is finite.
~ , , , Since
, ~ Cm(0)
is dense in Wi*p(R),(18) extends by continuity to all of WiSp(R). I
5.11 REMARK Let u E Co(Rn)and let q, r be as in the above proof.

From the identity
we obtain
where w i ( x ) is the line through x parallel to the xi-axis. Comparing this with
(20), we see that the computations of thc above proof can be reproduced to
yield in this case
I I ~ I I o , q , w " 5 J+Il,P,RV (21)
where the seminorm 1. is defined in Section 4.1 1. Inequality (21) is known
as Sobolev's inequality.
5.12 LEMMA Let R be a bounded domain in R" having the cone property.
If mp < n, then WmVp(R) -+ Lq(R) for p I; q Inp/(n-mp). The imbedding
constant may be chosen to depend only on m, p, n, q, and the cone C deter-
mining the cone property for R.
PROOF Let qo = np/(n-mp). We first prove by induction on m that

W"$p(R)-+ LqO(R).Note that Lemma 5.10 establishes the case m = 1.
Assume, therefore, that W"- l i p(Q) + Lr(R) for r = np/(n- mp + p ) when-
ever n > (m - 1)p. If u E Wm*p(R),where n > mp, then u and D p (1 < j < n)
belong to W"-l.p(R). It follows that u E W'.r(R) and
I I ~ I I I , ~<, ~K,IIUllm,p,n*
P (a)-+ Lqo(R)
Since mp < n, we have r < n and so by Lemma 5.10 we have W 1
where qo = nr/(n- r) = np/(n- mp) and
IIullo,qo,n 5 K2 ll4 l,r,n KJ IIUllm,p,n*

This completes the induction.
Now suppose p < q s qo. We set
s = (Yo -q) P/ ( qo- P) and r = P / S = (qo-p)/(qo-q)
and obtain by HBlder's inequality
5.13 COROLLARY If mp = n, then W"*p(R) -+Lq(R) for p Iq < 03.

The imbedding constant here may also depend on volR.
PROOF If q 2 p' = p / ( p- l), then q = ns/(n- ms), where s = p q / ( p+ q)
satisfies 1 Is < p . By Theorem 2.8, W"*P(R) W m i s ( Q with
) -+ imbedding
constant dependent on volR. Since ms < n, W'",'(R) + Lq(R)by Lemma 5.12.
If p I q I p' the desired imbedding follows by interpolation between
W".P(R) -+ LP(R) and W"*p(R) -+ LP'(n) as in (23). I
For mp = n and q 2 p the dependence of the imbedding constant on

volR may be removed as we show in the following lemma which removes the
restriction of boundedness of R from Lemma 5.12 and Corollary 5.13.
106 v IMBEDDINGS OF Wmn,p(R)
5.14 LEMMA Let R be ararbitrary domain in R"having the cone property.

If mp < n, then Wm9p(R) + Lq(R) for p 5 q I;np/(n-mp). If mp = n, then
Wm*P(R)--t LQ(R) for p I;q < 00. If p = 1 and m = n, then Wm*P(R)--t
Ceo(a).The constants for these imbeddings may depend on m, p , n, q, and the
cone C determining the cone property for $2.
PROOF We tesselate Rnby cubes of unit side. If ,I= (Al, ..., I,)is an n-tuple of
integers, let H = { x E R" : I , I;xi 5 I.,+ 1; 1 I;i I;n}. Then 08" = HA. u,
As remarked in the first paragraph of the proof of Theorem 4.8, even an
unbounded domain R with the cone property can be expressed as a union of
finitely many subdomains, say R = uYzl
R,, such that Rj = UxEA,(x+Pj),
where A, c R and Pi is a parallelepiped with one vertex at the origin. The
number N and the dimensions of the parallelepipeds P j depend on n and the
cone C determining the cone property for R. For each I and for 1 I;] 5 N let
Q,,j = u
X EA j d f +
(x+pj)*
The domains R,, evidently possess the following properties :
(0
,= uA,jfiA,,;
(ii) RA, is bounded;
(iii) there exists a finite cone C' depending only on P,, ..., PN (and hence
,
only on n and C) such that each R,, has the cone property determined by C' ;
(iv) there exists a positive integer R depending on n and C such that any
,
R + 1 of the domains an,have empty intersection;
(v) there exist constants K' and K" depending on n and C such that for
each R,, j ,
K' 5 vol R,, I; K". ,
Suppose mp < n and let u E WmPp(R).If p I;q I; np/(n-mp), then by (ii),
(iii), and Lemma 5.12, we have
I I u l l o , S , *A,, I; II'llm, (24)
P, nl,,,
where K = K(m,p, n, q, C ) is independent of I and]. Hence by (i) and (iv) and
since q 2 p
Thus Wmip(R)+ Lq(R) with imbedding constant KR1IP.

If mp = n, (24) holds for any q such that p I,q < co by virtue of Corollary
5.13, and the constant K can be chosen independent of A and j thanks to (v).
The rest of the above proof then carries over to this case.
Finally, if p = 1 and m = n, we have by Lemma 5.8 and a nonsingular

linear transformation that W"*'(P)+ Co((P)for any parallelepiped P c R,
the imbedding constant depending only on n and the dimensions of P.Hence
W", (R) + CBo(a) by virtue of the decomposition R = R , , . u
We have now proved Part I, Cases A and B of Theorem 5.1 for the case
k = n. Before completing these cases by considering the trace imbeddinge
(k < n), we establish the continuous function space imbeddings, Part I, Case
C, and Part 11.
5.15 LEMMA Let R be a domain in R" having the cone property. If

mp > n, then Wmip(R)+ CBo(f2),the imbedding constant depending only on
m,p, n, and the cone C determining the cone property for n.
PROOF Suppose that we can prove that for any 4 E Cm(n),

~ U P I ~ ( ~Kll4llm,p,n,
)I (25)
Jon
where K = K(m,p,n, C).If u E Wm*p(R),then by Theorem 3.16 there exists
a sequence {$,,} in Cw(R)converging to u in norm in Wmsp(R).Since (4,) is
a Cauchy sequence in Wm* p(R),(25) implies that { 4,,} converges to a continuous
function on R. Thus u must coincide a.e. with an element of CBo(R).It is
therefore sufficient to establish (25).
=-
First suppose m = 1 so that p n. Let x E R and let C, c R be a finite
cone congruent to C and having vertex at x. Let h be the height of C. Let
(r,8) denote spherical polar coordinates in R" with origin at x so that C, is
specified by 0 < r < h, 8 E A. The volume element in this system is denoted
by r"- 'o(8) dr do. We have
from which we conclude, for 0 < r < h,
I4(x)l sl4(r,@l+
sa lgrad4(t9e)ldt.
Multiplying this inequality by r"- 'o(8) and integrating r over (0,h) and 8
over A, we obtain
108 v IMBBDDINGS OF Wm”(Q)
the last inequality following from two applications of Holder’s inequality.

Since p > n we have (n- 1)(1-p‘) > - 1 and so
Ix-,,l-(n-1)~’dy = r(m-l)(l-~’)dr <
Hence
I4(x)l I KI1411i,p,cx5 Kll4lli.p.n
with K = K(m,p,n, Cx)= K(m,p,n, C ) . Thus (25) is proved for rn = 1.
If m > 1 but p > n, we still have
If p I n < mp, there exists an integer j satisfying 1 s j Im - 1 such that

j p I n < ( j + 1)p. Ifjp < n, set r = np/(n-jp); ifjp = n, choose r > max(n,p).
In either case we have by the result proved above and by Lemma 5.14 that
I4(x)l I K1II4II1.r.Cx I ~ ~ l 1 4 1 1 m - j , r , c xI ~ l 1 4 1 1 m , p , c x5 KII4IIm,p,n,
the constants depending only on m, p , n, and C. This completes the proof. I
5.16 COROLLARY If mp > n, then Wm7p(Q) -+ Lq(Q)for p I q s 00.
The imbedding constants depend only on m, p , n, q, and the cone C.
PROOF We have already established that

IIuIIo,~,R = esssu~lu(x)l5 Kllullm,p,n
X O R
If p I q < co, we have

for all u E Wm*p(Q).
5.17 LEMMA Let Q be a domain in Rn having the strohg local Lipschitz

property, and suppose that mp > n 2 (m- 1)p. Then Wm*P(Q)-+ CoiA((rz)
for:
(i) 0 c I I m-n/p if n > (m- l ) p , or
(ii) 0 < I < 1 if n = ( m - l ) p , or
(iii) 0 < 1 1 1 i f p = l , n = m - 1 .
In particular W m * P ( R
-+)C0(a).The imbedding constants depend on m, p , n
and the parameters S,M specified in the description of the strong local
Lipschitz property for Q (see Section 4.5).
PROOF Let u E Wm.p(R).The strong local Lipschitz property implies the cone
property so by Lemma 5.15 we may assume that u is continuous on a and
satisfies
It is therefore sufficient to establish further that for suitable d,
Since mp > n 2 (m- 1)p we have by Lemma 5.14 that Wm*p(R)+

W1.'(R) where:
(i) r = np/(n- mp + p ) and 1 - (n/r) = m - (n/p) if n > (m - l)p, or
(ii) r is arbitrary, p c r c 00 and 0 c 1 -(n/r) c 1 if n = (m- l)p, or
(iii) r = 00, 1 - (n/r) = m - (n/p) = 1 if p = 1 and n = m- 1.
It is therefore sufficient to eastablish (27) for m = 1 ;that is, we wish to prove
that if n c p I co and 0 < d 5 l - ( n / p ) , then
Suppose, for the moment, that R is a cube, which we may also assume
without loss of generality to have unit edge. For 0 < t < 1, R, will denote a
cube of edge r with faces parallel to those of R and such that B,c R. Let
u E Crn(R).
Let x , y E R, I x - y l = 0 < 1. Then there exists a fixed cube R, with x , y E
3z, c R. If ZER,, then
u ( x ) = U(Z) -
so:l -u(x + t(z-x)) dr,
so that
Hence
lu(x)-u(z)l I&a
I' lgradu(x + z(z-x))l dz.
equation continues
110 v IMBEDDINGS OF Wm*'(R)
K40
-
1 (n/p)
Ilgradullo,p,n,
where K4= K4(n,p)= Jnj; t-"lPdt < m. A similar inequality holds with y
in place of x and so
lu(x)-U(Y)l 5 2K4 1x-J" 1 -(nip)
IIgrad 410 , p . R *
It follows for 0 < I 5 1- (n/p) that (28) holds for R a cube, and so, via a
nonsingular linear transformation, for R a parallelepiped.
Now suppose that R has the strong local Lipschitz property. Let 6, M , R,,
Uj,and 5 be as specified in Section 4.5. There exists a parallelepiped P of
diameter 6 whose dimensions depend only on 6 and M such that to each j
there corresponds a parallelepiped Pj congruent to P and having one vertex
at the origin, such that for every x E 5 n R we have x + Pi c Q. Further-
more there exist constants So and d1 depending only on 6 and P,with 6,s 6,
such that if x , y E Vj n R and I x - y l < do, then there exists z E (x+ Pj) n
( y + Pj) with Ix-zl+ l y - z l S 6 , ( x - y l . It follows from application of (28)
to x + Pr and y+ Pj that if u E Cm(R),then
lu(x)-u(Y)l 5 lu(x)-u(z)l+ lU(Y)-U(Z)l
5 ~s lx-zI'Il~lli,p,n + Ks IY-ZI' IIUlli,p,n
5 2 1 - ' & ~ i ' I ~ - ~ I ' IIUII1,p.n. (30)

Now let x , y E R be arbitrary. If I x - y l < 6 , 56 and x , y E R,, then x , y E 5
for some j and estimate (30) holds. If I x - y l < do, x E Rd, y E R ad,then
x E 5 for some j and (30)follows by application of (28) to x+ Pj and y + Pj
-
-
again. If Ix-yI < So and x , y E R ad,then (30) follows from application of
(28) to x + P',y +P',where P' is any parallelepiped congruent to P and having
one vertex at the origin. Finally, if I x - y l r a0, then we have
Iu(x)-u(Y)I 5; l ~ ( x )+I lu(~)I5 Kci IIUII1,p.n 5 ~~~~'l~-~l'll~lli,~,n
This completes the proof of (28) for u E Cm(R),and so by Theorem 3.16, for
all continuous u. I
We have now completed the proof of all parts of the imbedding theorem
5.4 except the trace imbeddings of Cases A and B (corresponding to k < n).
For the proof of these we will need the following interpolation result.
5.18 LEMMA Let Q be a cube of edge length 1, having edges parallel to

=-
the coordinate axes in R". If p 1, q 2 1 and mp-p c n < mp, then there
exists a constant K = K ( p , q, m, n, 1 ) such that for every u E W m * p ( Qwe
) have
(a.e. in Q )
lu(x)l
5 KIIuII$,q,Q IlullA:LQ, (3 1)
where s = (mp-n)q/[np+(mp-n)q].
PROOF It is sufficient to establish (31) for u E Cm(Q).Since each point of Q

is a corner point of a cube contained in Q, having edges parallel to those of Q ,
and having edge length 1/2, we may assume without loss of generality that x is
itself a corner point of Q , say Q = { y E R" :xi < yi < xi + 1; 1 5 i 5 n}.
By Lemma 5.17 we have for y E Q ,
lu(x)l- lu(y)I 5 lu(x)-u(y)I IK l l X - y I " - ( " I p ) llullrn,p,~* (32)
Let U = which we may assume to be positive; let p = Ix-yl and
( = [lu(x)I/K1U ] p / ( m p - nSuppose
). for the moment that ( I1. We have for
P 5 5,
lu(y)I 2 lu(x)l- K1 u p m - ( " / p ) >
- 0.
Raising the above inequality to the power q and integrating y over Q , we

obtain
= K~I ( x )1q+ ( n p / ( m p - n ) ) -np/(mp-n)
from which (31) follows at once.

If, on the other hand, > 1, then from (32) we obtain
l ~ ( y )2
I Iu(x)~ - Iu ( x )I - I u ( x )I ( p / l y- ( " / P )
- K1 Up'"-("'p)>
20 ifpsl.
If t > 0, then
splu(y)l'dy 2 K 2 j i l u ( x ) l ' ( l -(p/l)m-'"/p')'p"-ldp = K'$IU(X)l'.
0
Set t = [(mp-n)q+np]/mp. Then

112 v IMBEDDINGS OF Wm"(R)
by an application of Holder's inequality. Since I I U ~ ] ~I, ~~ , ~~ u ~(31)~ ~ , ~ ,

follows at once. I
We remark that the above lemma also holds for the case p = 1, m = n.
In this case we have from Lemma 5.14 that W"*'(R) --i Lm(R)so that lu(x)l I
K I I U ~ ~ , , ~a.e.
, ~ in Q, which is (31) in this case.
5.19 LEMMA Let R be a domain in R" having the cone property, and let
Rk denote the intersection of R with some k-dimensional plane, where
1 I k I n (n"= 0). If n 2 mp and n-mp < k I n, then
W"*P(R) --i L*(Rk) (33)
for p I q I kp/(n-mp) if n > mp, or for p I q < 00 if n = mp. If p = 1,
n > m and n - m I;k I n, then (33) holds for 1 I q Ik/(n-m).
The imbedding constants depend only on m, p , k, n, q, and the cone C
determining the cone property for 0.
PROOF It is sufficient to establish the above conclusions for R bounded,

n > mp, and q = kp/(n- mp), as extension to the other cases can be carried out
in the same manner as was described for the case k = n in Corollary 5.13 and
Lemma 5.14. We may also assume, as in Lemma 5.10, that R is a union of
coordinate cubes of edge 2 units.
Let ROkbe a k-dimensional coordinate subspace of R" on which Qk has a
one-to-one projection R,k. Suppose, for the moment, that p > 1. Let v be
the largest integer less than mp. Then mp-p c v < mp and since n - m p < k
we have n - v Ik. (Note that if p = 1, the same conclusion holds with
k = n - m , v = m.) Let p = (,,!J and let Ei (1 I i I p ) denote the various
coordinate subspaces of R,k having dimension n - v. Let R, denote the projec-
tion of nok(and hence of Rk)onto E,. Also, for each x E R, let denote the
intersection of R with the v-dimensional plane through x perpendicular to
E,. Then R,, contains a v-dimensional coordinate cube of unit edge with one
vertex at x. By Lemma 5.18, with q = qo = np/(n- nip), we have for u E Cm(R)
sup 1 (v)1 (n - v)p/(n - mp) I K1 llull(mnp- v)qo/mp
0 , qo, n,,, I u l l mV ,mp , * (34)
YSnl,,
Let dx' and dx,' denote the volume elements in Ei and the orthogonal comple-
ment of E,, respectively. Integration of (34) over Ri leads to
equation continues
TRACES OF FUNCTIONS IN wm*p(R)
ON THE BOUNDARY OF 113
IKl[jnlu(x)lqodx] (mp- v ) l w [I C
Rlalsm
~Dau(x)lpdx~
= Kl la!:, R
I141~2iv)/mp (35)
by Holder’s inequality.
Finally, we apply Lemma 5.9 to the subspaces Eiof Rok. Note that the constant
2 of that lemma is here equal to (&Il). Letting d d k ) denote the volume
element in R,k and setting q = kp/(n- mp),we obtain
sup 1 u ( y ) pfld d k )
Since @/,u = (n- v)p/(n-mp), it follows from (35) and (36) and from Lemma
5.14 that
This establishes the desired imbedding. 1

We have now completed the proof of Theorem 5.4.
Traces of Functions in WmvP(R) on the Boundary of R
5.20 Of importance in the study of boundary value problems for differential

operators defined on a domain R is the determination of spaces of functions
defined on the boundary of R containing the traces uIbdrynof functions u in
W”.P(R). For example, if Wm!P(R)-, C(Ti), then clearly UIbdryn belongs to
C(bdryR). We outline below an Lq-imbeddingtheorem for such traces which
can be obtained as a corollary of Theorem 5.4.
The problem of characterizing the image of Wm*p(R)under the operator
u + U \ b d r y n has been extensively studied by many authors. The solution,
which involves Sobolev spaces of fractional order m will be given in Chapter
VII (see in particular Theorem 7.53). The approach used in that chapter is
due to Lions [37, 381.
114 v IMBEDDINGS OF WmSp(R)
5.21 Let $2be a domain in R" having the uniform C"-regularity property.
Thus there exists a locally finite open cover { U j }of bdry R, and corresponding
m-smooth transformations Yj mapping B = { yE R" : IyI < I } onto Uj such
that Uj n bdryn = Y j ( B o ) Bo
; = { yE B : yn = 0). Iff is a function having
support in Uj, we may define the integral off over bdry R via
sbdry Cl
f(x) da =
sUjnbdry n
AX)
where y' = (yl, .. .,Y,,-~)and if x = Y j ( y ) , then

da = lo f i y j (y', 0) Jj (y'>W ,
Iffis an arbitrary function defined on bdryR, we may set
1bdry n
~ ( xdo
) = c1 j bdryn
f ( x ) uj ( x ) do,
where {uj} is a partition of unity for bdryn subordinate to { U j } .
5.22 THEOREM Let R be a domain in R" having the uniform C"-

regularity property, and suppose there exists a simple (m,p)-extension operator
E for R. If mp < n and p S q I (n- I)p/(n - mp), then
Wmpp(R) + Lq(bdryR). (37)
If mp = n, then (37) holds for p 5 q < co.
PROOF Imbedding (37) should be interpreted in the following sense: If

u E Wm3p(R), then EMhas a trace on bdryR in the sense described in the final
paragraph of Section 5.2, and (JEul(o,q,bdryR 5 K I I U ~ ~ , , , , ~ , ~with K independent
of u. [Note that since Co(Rn) is dense in WmVP(R), 11 E U ( J ~ , ~ , independent
~~,,,~~S
of the particular extension operator E used.]
It is sufficient to prove the theorem for mp c n and q = (n- l)p/(n-mp).
There exists a constant K , such that for every u E W"*p(R)
I I E ~ l l m , p , ~5
n K1IIullm,p,n*
By the conditions of the uniform C"-regularity property (Section 4.6), there

exists a constant K2 such that for e a c h j and every y E B, x = Y j ( y ) E Uj
Noting that 0 d; uJ(x) 5 1 on R", and using imbedding (2) of Theorem 5.4
W"*'(R) AS A BANACH ALGEBRA 115
applied over B, we have for u E Wm*P(R),
The second last inequality above makes use of the finite intersection property
possessed by the cover {Uj}. The constant K4 is independent of j since
IDaYj,i(y)I ,,
I const for all i,j , where Y j = ( Y j , ..., Yj,,,).This completes
the proof. I
Wm*p(R)as a Banach Algebra
Given functions u and u in WmaP(R),where R is a domain in R", one

cannot in general expect their pointwise product uu to belong to Wm*p(R).
[(uu)(x) = u ( x ) v ( x ) a.e. in 0.1 It is, however, a straightforward application
of the Sobolev imbedding theorem to show that this is the case provided
mp > n and R has the cone property.
5.23 THEOREM Let R be a domain in R" having the cone property. If

mp > n, then there exists a constant K * depending on m,p, n, and the finite
cone C determining the cone property for R, such that for all u, u E Wm9P(R)
the product uu, defined pointwise a.e. in R, belongs to Wm*p(R)and satisfies
K* 1I UIIm, p . R I ullm, p . n
I1uullm, p . R (38)
In particular, W'"*P(R) is a commutative Banach algebra with respect to
pointwise multiplication and the equivalent norm
llull~,p,*= K*IIullm,p,n*
PROOF In order to establish (38) it is sufficient to show that if la1 Im,then

n
116 v IMBEDDINGS OF Wm*P(R)
where K = K(m,p,n, C ) . Let us assume for the moment that u E Cm(R). By

Leibniz's rule for distributional derivatives, that is,
it is sufficient to show that for any P S a, IaJI m, we have
where Ka,a= K,,,,(m,p,n, C ) . By the imbedding theorem there exists, for

each j with m, a constant K(P) = K(P,m,p,n,C ) such that for any
W E W"q2)
provided (m - IPI)p In and p I r 5 np/(n - (m - IPI)p) [or p I r < 00 if

(m - IPI)P = nl, or
ID'w(x)I 5 K(P) IIwIIm,p,n a.e. in R
provided (m - IPI)p > n.
Let k be the largest integer such that (m- k)p > n. Since mp > n we have
that k 2 0. If I k, then (m - IbI)p > n, so
Similarly, if la - PI I k, then
Nowif IPl>kand Ia-PI>k,then,infact, I P I 2 k + l and I a - P 1 2 k + l

so that n 2 (rn - IjI)p and n 2 (m - Ia-flI)p. Moreover,
n - (m - IPI)P + n - (m - I ~ - P I ) P = 2 - (2m - lal)P < 2 - -
mP < 1.
n n n n
Hence there exist positive numbers r,r' with ( l / r ) + ( l / r ' )= 1 such that
COUNTEREXAMPLES AND NONIMBEDDING THEOREMS 117
Thus by Holder's inequality and (39) we have
This completes the proof of (38) for u E Cm(R), u E WmSp(R).

If u E Wmn,p(R),then by Theorem 3.16 there exists a sequence {u,} of
CoD(R)functions converging to u in W'".p(R). Then by the above argument
{u,,u} is a Cauchy sequence in Wm.P(R)so converges to an element w of that
space. Since mp > n, u and u may be assumed continuous and bounded on R.
Thus
II w--u~II 0,p . n 5 II w - un 0110,p . n + I (un-u)ulI 0,p , ~
5 lIW-unUllO,p,n +ll~llO,oo,~ll~n-~llO,p,~
-0 as n-co.
Hence w = uu in Lp(R) and so w = uu in the sense of distributions. Therefore
w = uu in W m ~ p ( and
0)
I I ~ ~ I l m , p ,=
n IIwllm,p,n 5 lim SUP = Il~llm,p,~ll~Ilm,p,~~
lI~n~llm,p,R
n+m
This completes the proof of the theorem.
We remark that Banach algebra Wm*P(R)has an identity if and only if R

is bounded, that is, the function e ( x ) = 1 belongs to Wmrp(R)if and only if
-=
volR co, but there are no unbounded domains with finite volume having
the cone property.
Counterexamples and Nonimbedding Theorems
5.24 Consideration of the statement of the Sobolev imbedding Theorem 5.4

may lead the reader to speculate on several directions of possible generaliza-
tion. Before exploring the possibility of proving imbedding theorems for
domains not satisfying the conditions of Theorem 5.4, we first construct
examples showing that in certain respects that theorem gives "best possible"
imbedding results for the domains considered, and indeed for any domain.
Let R be an arbitrary domain in R" and assume, without loss of generality,
=-
that the origin belongs to 0.Let R 0 be such that the closed ball E T is
examples we construct a function u E C"(B,, -

contained in R. (Here E , = { X E R " : I x I c R}.) In each of the following
(0)) depending only on
p = 1x1. Iff€ Cm(O,co) satisfiesf(t) = 1 if t 5 R whilef(t) = 0 if t 2 2R, then
118 v IMBEDDINGS OF W"*'(fi)
the function w defined by
has compact support in R and belongs to W"*p(R)if and only if u E W"*p(BR).
5.25 EXAMPLE Let k be an integer such that 1 I k I n and suppose that

mp < n and q > kp/(n-mp). We construct u so that u E W"*P(BR)but
u $ Lq(BRk),where BRk= { x E BR:xk+ = ... = x,, = 0). Hence no imbedding
of the type WmiP(R) + Lq(Rk)is possible if q > kp/(n- mp).
Let u(x) = p', where p = 1x1 and the exponent I will be specified below.
It is readily checked by induction on la1 that
D'u ( x ) = 'P (x)p' - z l a J (40)
where Pa is a polynomial homogeneous of degree la1 in the components of x .
Hence ID"u(x)JI Kap"l"l and
Therefore u belongs to W m S p ( Bprovided

R)
(A-m)p + n > 0.
On the other hand, if u = (x12 + ...+ x ~ ~ ) then
"~,
SO that u $ Lq(BRk)if
I q + k < 0. (42)
Since q > kp/(n-mp), it is possible to select I to satisfy both (41) and (42)
as required. 1
Since the function u constructed above is unbounded near 0, no imbedding

of the form W m * p ( R-+ ) CBo(R)is possible if mp < n.
5.26 EXAMPLE Suppose p > 1 and mp = n. We construct u E W'"*p(BR)

so that u 4 L"O(BR).Hence the imbeddings WmiP(R)-+ Lq(R), valid for
p I q < co if mp = n and R has the cone property, cannot be extended to
yield Wm*p(f2) + Lm(R) or W"*P(R)+ CBo(R) unless p = 1 and n = m (see,
however, Theorem 8.25).
Let u(x) = log(log4R/p), where p = 1x1. Clearly u $ L"O(BR).Again it is
easily checked by induction that
D"u ( x ) = y
j= 1
Pa, ( x ) p - ("I [log(4R/p)] -j , (43)
where Pa,jis a polynomial homogeneous of degree 1.1 in the components of

x. Since p = n/m we have
so that
The right side of the above inequality is certainly finite if lcll< m. If = m,

we have, setting Q = log(4R/p),
which is finite since p > 1 . Thus u E WmSp(BR).I
It is interesting that the function u above is independent of the choice of

m and p with mp = n.
5.27 EXAMPLE Suppose mp > n > ( m- l ) p , and let 1 > m - (n/p). We

construct u E W m P p ( B such
R ) that u $ Co,'((B,).Hence no imbedding of the
form WmgP(R) -+ Co*'(a) is possible if mp > n > ( m - 1)p and 1 > m - (nlp).
As in Example 5.25 we take u ( x ) = p", p = 1x1. From (41) we have
U E Wm9P(BR) provided p > rn-(n/p). NOW ~ u ( x ) - u ( O ) ~ / ~ X - O ~ ~ SO
that u $ CoiA(&)when p < A , Thus u has the required properties if we choose
p to satisfy m - (n/p) 1 . We construct u E

W".p(BR)such that u 4 Ca7'(&). Hence the imbedding Wm*p(R)+ Co**(Q),
valid for 0 < 1 < 1 whenever R has the strong local Lipschitz property,
cannot be extended to I = 1 unless p = 1 , m - 1 = n.
Let u(x) = p log(1og 4R/p) where p = 1x1. Since
lu(x)-u(O)l/lx-Ol = log(log4R/p) + co as x + 0
it is clear that u $ Cop' (K).
Following (40) and (43) we have
D"u(x) = y Pa,j(x)pl-2~"~[log(4R/p)]-j,
j=l
120 v IMBEDDINGS OF W*"(R)
where P,, is a polynomial homogeneous of degree la], Hence
ID"u(x)lP I f K,,jpP"-'~')[log(4R/p)]-jP.
j=l
It then follows as in Example 5.26 that u E W".P(BR). I
5.29 The above examples show that even for very regular domains there can
exist no imbeddings of the types considered in Theorem 5.4, except those
explicitly stated there. It remains to be seen whether any imbeddings of these
types can exist for irregular domains not having the cone property. We shall
show that Theorem 5.4 can be extended, with weakened conclusions, to certain
types of such irregular domains, but we first show that no extension is possible
if the domain is "too irregular."
An unbounded domain R in R" may have a smooth boundary and still
fail to have the cone property if it becomes narrow at infinity, that is, if
lim dist(x, bdry R) = 0.
Iel;+nm
The following theorem shows that Parts I and I1 of Theorem 5.4 fail completely
for any such unbounded R which has finite volume.
5.30 THEOREM Let R be an unbounded domain in R" having finite

volume, and let q > p. Then W"sP(R) is not imbedded in Lq(fZ).
PROOF We construct a function u(x), depending only on the distance

p = 1x1 of x from the origin, whose growth as p increases is rapid enough to
preclude membership of u in Lq(f2)but not so rapid as to prevent u E W'"*P(f2).
Without loss of generality we assume volR = 1 . Let A ( p ) denote the
surface area [(n- 1)-measure] of the intersection of R with the spherical
surface of radius r centered at the origin. Then
p ( p ) d p = 1.
Let ro = 0 and definer,, for n = 1,2, ... by
Clearly r,, increases to infinity with n. Let Ar,, = r,,+ -r,, and fix E such that
0 < E < [l/(mp)]- [l/(mq)]. There must exist an increasing sequence {nj}T=,
such that Ar,,, 2 2-&"I, for otherwise Ar,, < for all but possibly finitely
many values of n whence we would have C,"=,, Ar,, < 00, a contradiction. For
convenience we assume n, 2 1 so that nj 2 j for all j . Let a. = 0 aj = r,,,, 1,
and bj = r,,, . Note that aj- I bj c aj and aj - bj = Ar,,, 2 2-'"*.

Let f be a nonnegative, infinitely differentiable function on R having the
properties :
(i) 0 I f ( t )I 1 for all t,
(ii) f ( t ) = 0 if t 5 O , f ( t ) = 1 if t 2 1,
(iii) I ( d / d t ) k f ( t ) l I M for all t if 1 I k I m.
If x E R and p = 1x1, set
c 2"' - for a j - l I p Ibj
Clearly u E Cm(R).Denoting Qj = { x E i2 :aj- I p I aj}, we have
< 2n'-1p'q [ A (p) dp + 2n@1qr A ( p ) dp

00
-
Since p < q the above inequality forces
Also, if 1 I k Im, we have
I*,I=I,
@ IJI l
dku P
dx
dkU P
A ( p ) dP
I M p 2 " J p i q [ a j - b j bi] - k p ~ dp
(p)
= .$~~2-nAl-~/q-ekp) -
< 3MP2-'j,
where C = 1 - p / q - e k p > 0 since E < [ l / ( m p ) ] - [l/(rnq)].Hence D4u E Lp(R)
for IQI Im, that is, u E W",p(i2).However, u $ Lq(i2),for we have for each j,
= 2"1-1[2-fl1-1-1-2-nj-l] 2 a.
Therefore WmVp(R)
cannot be imbedded in Lq(R). I
122 v IMBEDDINGS OF W""(R)
The conclusions of the above theorem can be extended (see Section 6.35)
to unbounded domains R having infinite volume but satisfying
lim sup vol { x E R : N 5; 1x1 5; N + 1) = 0.
N-rm
5.31 Parts I and I1 of Theorem 5.4 also fail completely for domains with
sufficiently sharp boundary cusps. If R is a domain in R" and xo is a point on
the boundary of R, let B, = B,(xo) denote the open ball of radius r about xo,
let R, = B, n R, let S, = (bdry B,) n R, and let A (r, R) be the surface area
[(n-1)-measure] of S,. We shall say that R has an exponential cusp at
xo E bdryR if for every real number k, we have
5.32 THEOREM If R is a domain in R" having an exponential cusp at the

point xo on bdry R, and if q > p, then Wm9P(R)is not imbedded in Lq(R).
PROOF We construct u E Wm*p(R)which fails to belong to Lq(R) because it

becomes unbounded too rapidly near xo. Without loss of generality we may
assume xo = 0 so that r = 1x1. Let R* = {y = x/lxlz: x E R, 1x1 < 1). It is
easily seen that R* is unbounded and has finite volume, and that
= )r2("-')~(1/r,R).
~(r,n*
Let t satisfyp < t < q. By Theorem 5.30there exists a function ii E C"(0,co)
such that
(i) ii(r) = 0 if 0 < r 5; 1,
~ *<
(ii) / m l u ( j ) ( r ) l r ~ ( r , dr ) co if o < j 5 m,
1
(iii) ~ m l i i ( r ) l q A ( r , R *dr) = co.
[If r = Iyl, then u ( y ) = ii(r) defines u E WmVp(R*)but u 4 Lq(R*).] Let

x =y/JyI2 so that p = I x J = l/lyl= l/r. Set 1 = 2n/q and define u ( x ) =
Z ( p ) = r'ii(r) = lyl'u(y). It follows for la1 = j I: m that
where the coefficients cli depend only on 1. Now u ( x ) vanishes for 1x12 1
and so
IMBEDDING THEOREMS FOR DOMAINS WITH CUSPS 123
On the other hand, if 0 51.1 = j I m,we have
If it happens that (J.+2m)p I 2n, then, since p < i and vol!2* < co, all the
integrals in (45) are finite by Holder's inequality and so u E Wm*p(R).Otherwise
let
k = [(A+2m)p- h][t/(t-p)] + 2n.
By (44)there exists a 5 1 such that if p s a, then A@,R) I pk. It follows
that if r 2 l/a, then
rk-'"A(r,Jz*) 5 rk-2pk = r - 2 .
Thus
which is finite. Hence u E Wm*p(R)and the proof is complete. I
Imbedding Theorems for Domains with Cusps
5.33 Having proved that Theorem 5.4 fails completely for domains that are
sufficiently irregular we now show that certain imbeddings of the types con-
sidered in that theorem do hold for less irregular domains that however fail
to have the cone property. Questions of this sort have been considered by
several writers-see, for instance, the work of Globenko [26,27] and Maz'ja
[44,45]. The treatment given below follows that given in one of the author's
papers c 11.
We consider domains Jz c R" whose boundaries consist only of (n- 1)-
dimensional surfaces, and it is assumed that R lies on only one side of its
boundary. R is said to have a cusp at the point xo E bdryR if no finite open
cone of positive volume contained in R can have vertex at xo. The failure of a
domain R to have any cusps does not, of course, imply that R has the cone
124 v IMBEDDINGS OF Wm”(a)
property. We consider, for the moment, a family of special domains that we

call standard cusps and that have cusps of power sharpness (less sharp than
exponential cusps).
5.34 If 1 5 k I n- 1 and A > 1, let QkS1denote the standard cusp in En

specified by the inequalities
xi’ + * ’ *
+ xk’ < xi:,, > o,...,x, > 0,
xk+l
(x,’ + + Xk’)’’’ + xi+, + ... + x,’ < a’, (46)

where a is the radius of the ball of unit volume in 08”. We note that a < 1, Qk,A
,,
has axial plane spanned by the xk+ ...,x, axes, and verticial plane (cusp
plane) spanned by x k + Z , ...,x, . If k = n - 1, the origin is the only vertex point
of Qk, The outer boundary surface of Qk, is taken to be of the form (46)in
order to simplify calculations. A sphere, or other suitable surface bounded
away from the origin, could be used instead.
Corresponding to the standard cusp Qk,A we consider the associated
standard cone Qk = Q,, specified in terms of Cartesian coordinates y,, ...,y,
by
Y1’ + + yk’
*” < Yk2+1, Yk+l > 0,***,Ye
> 0,
y,’ + + y,’
*.* < a’.
Figure 4 illustrates standard cusps and their associated standard cones in R2
and R3. In R3 the cusp Q2,2 has a single cusp point at the origin, while
has a cusp line along the +axis.
It is convenient to adopt generalized “cylindrical” coordinates
(rk,+l, ..., 4k-1,YktI,...,Yn)in E n s o t h a t r , 2 0 , - n s 4 , < n , 0 ~,...,4 ~
4,- s I[, and
yl = r, sin4, sin& ... sin+,-,,
y2 = r,cos$,sin$2 +..sine$,-,,
Y3 = rkcos42-..sinq5k-,, (47)
Yk = rk cos4k-1.
In terms of these coordinates Qkis represented by
5 rk < Yk+l, Yk+l > O, .**,Yn> O,
rk2 yit1+ + .-.+y,’ < a’.
The standard cusp may be transformed into the associated cone Qk by
t y2
1”
(b)
FIG.4 Standard cusps and cones in (a) R2 and (b) R3.

126 v IMBEDDINGS OF Wrn”(R)
means of the one-to-one transformation

x1 = r,”sin+, sin& ... sin+,-,,
x2 = r,lcos4,sin4, -..sin+,-,,
xs = r,l cos +2 sin 4,-
xn = Yn,
which has Jacobian determinant
We now state three theorems extending imbeddings of the types considered

in Theorem 5.4 (except trace imbeddings)to domains with boundary irregular-
ities comparable to standard cusps. The proof of these theorems will be given
later in this chapter.
5.35 THEOREM Let R be a domain in R” having the following property:

There exists a family r of open subsets of R such that
(0 a= U G E &
(ii) r has the finite intersection property, that is, there exists a positive
integer N such that any N + 1 distinct sets in r have empty intersection;
(iii) at most one set G E r has the cone property;
(iv) there exist positive constants v > mp-n and A such that for any
G E r not having the cone property there exists a one to one function $
mapping G onto a standard cusp Q k , l , where (A- 1)k 5 v and such that for
all i,j (1 I i, j In), all x E G,and all y E Q k , A ,
Then
[If v = mp-n, (50) holds forp 1s q < 00 (and q = rn i f p = 1). If v < mp-n,
(50) holds for p 5 q I 03.1
5.36 THEOREM Let f2be a domain in R” having the following property:

There exist positive constants v < mp-n and A such that for each x E R there
exists an open set G with x E G c R and a one-to-one mapping J/ of G onto a
standard cusp Qk,Iwith (12- I)k I v and such that for all i,j (1 Ii, j S n), all
x E G, and all y E
Then
More generally, if v < ( m - j ) p - n , where 0 s j 5 m- 1, then

W”.P(R) + CJ(R).
5.37 THEOREM Let R be a domain in R” with the following property:

There exist positive constants v, 6, and A such that for each pair of points
x, y E R with I x - y l s 6 there exists an open set G with x, y E G c R, and a
one to one mapping J/ of G onto some standard cusp with (12- 1)k I v ,
and such that for all i,j (1 I i, j I n), all x E G, and all y E Qk,
Suppose that for some j with 0 I j I m - 1 we have ( m - j - l ) p < v + n c

( m - j ) p . Then
wmpp(R) + Cj.p(si), 0 c p I m - j - [(n+v)/p]. (52)
If ( m - j - 1)p = v + n , then (52) holds with 0 < p c 1 . In either event we have
Wrn5P(R)+ @(?I).
5.38 REMARKS (1) The reader may wish to construct examples similar
to those of Sections 5.25-5.28 to show that the three theorems above give the
best possible imbeddings for the domains considered.
(2) The following example may help to illustrate Theorem 5.35: Let
R = {(xI,x2,x3) E R3 : x 2 > 0, x22< x, < 3 ~ ~ Setting
~ ) . a = (3/471)’’~, the
radius of the ball of unit volume in R’, we may readily verify that the trans-
formation
y , = XI+ 2x22 , ~2 = ~ 2 , ~3 = ~3 - (k/~), k = 0, +1, + 2 ,... ,
transforms a subdomain G, of R onto the standard cusp in the manner

required of the functions $ in the statement of Theorem 5.35. Moreover,
{Gk}km,--oo has the finite intersection property and covers R up to a set with the
128 v IMBEDDINGS OF Wm"(n)
cone property. Hence W r n y p (+ n )Lq(n) for p s q I4p/(4-mp) if mp < 4 ,

or f o r p Iq < co if mp = 4, o r p 5 q 5 00 if mp > 4.
Imbedding Inequalities Involving Weighted Norms
5.39 The technique of mapping a standard cusp onto its associated standard
cone via (47) and (48) is central to the proof of Theorem 5.35. Such a trans-
formation introduces into any integrals involved a weight factor in the form
of the Jacobian determinant (49). Accordingly, we must obtain imbedding
inequalities for such standard cones, corresponding to LP-norms weighted by
powers of the distance from the axial plane of the cone. Such inequalities are
also useful in extending the imbedding theorem 5.4 to more general Sobolev
spaces involving weighted norms.
We begin with some one-dimensional inequalities for functions contin-
uously differentiable on a fixed open interval (0, T) in R.
5.40 LEMMA If v > 0 and u E C'(0, T), and if ~ ~ l u ' ( t ) l t ' d

<tco, then
limf+o+~u(t)l
t' = 0.
PROOF Let E > 0 be given and fix s, 0 < s < T/2, small enough so that for
any t, 0 < t < s, we have
Jt
Now there exists 6, 0 < 6 < s, such that
6'lu'(T/2)1 < 4 3 and (a/$)' lTi2lu'(z)lz'dz < ~ / 3 .

If 0 < t I6, we have
so that
Hence limf,o+tv~u(t)~=O.I
5.41 LEMMA If v > 0, p 2 1, and u E C1(0,T), then

IMBEDDING INEQUALITIES INVOLVING WEIGHTED NORMS 129
PROOF We may assume without loss of generality that the right side of (53)
is finite, and that p = 1. Integration by parts gives
Lemma 5.40 assures the vanishing of the integrated term at zero. Transposition
and estimation of the term on the right now yields
which is (53) for p = 1. I

5.42 LEMMA If v > 0, p 2 I, and u E C1(0, T ) , we have the following
pair of inequalities :
7 PT PT
PROOF Again the inequalities need only be proved for p = 1. If 0 < t I T/2,
we obtain by integration by parts
whence
lu(t)l 5 Tl +l
2 T
lu(4l da
T
lu’(4l do.
For T / 2 I t < T the same inequality results from partial integration of

J,T’21#(t+~-T/2)1dz. This proves (54) for p = 1. Replacing u ( t ) by u ( t ) t v
in this inequality, we obtain
where (53) has been used to obtain the last inequality. This is the desired
result (55) for p = 1. I
130 v IMBEDDINGS OF wrngp(Q)
5.43 Now we turn to R", n 2 2. If x E R", we shall make use of the spherical
polar coordinate representation
x = (P, 4) = (P, $1. ..*,4 n - 1 > ,
where p 2 0, - n 5 +1 I n, 0 I iP2, ...,4n-lI ?I,and
x 1 = p sin& sin$,
x2 = pcos&,sin4,
x3 = p cosc#~~...sin4,, -,,
x, = P cos 4 n - 1.
The volume element is
dx = dxl dx2 dx, = p"-' n sin'-

n-1
j=l
+j dp d4,
where d4 = d41.-.di$n-l.
We define functions rk = rk(x)for 1 5 k I n as follows:
n-1
rn(x) = P*
For 1 Ik I n - 1, rk(x)is the distance of x from the coordinate plane spanned
by the axes X k + l , ..., x,,; r,(x) being just the distance of x from the origin. In
connection with the use of product symbols of the form P = Pi, be it
agreed hereafter that P = 1 if m < k.
Let Q be an open, conical domain in R" specified by the inequalities
O<p<a, -fi1<41<B1, 044j<<j, j = 2 , 3,...,n - 1 ,
(56)
where 0 < pi 5 K. [Inequalities "<" in (56), corresponding to any pi = n,
are replaced by "I ." If all pi = K, the first inequality is replaced by 0 I p < a.]
It should be noted that any standard cone Qk (introduced in Section 5.34) is
of the form (56) for some choice of the parameters pi, 1 Ii I n- 1.
The following lemma generalizes Lemma 5.41 in a manner suitable for
our purposes.
5.44 LEMMA Let Q be specified by (56) and let p 2 1. Suppose that either
m = k = 1, or 2 5 m 5 n and 1 I; k 5 n. Suppose also that 1 - k < v1 I v I
v2 < co. Then there exists a constant K = K(m,k , n,p, v l , v2,P1, ..., Pn- 1)
independent of v and a, such that for every function u E C' (Q) we have
-
S,I (x) I Crk Crm dx
PROOF Once again it is sufficient to establish (57) for p = 1. Let Q+ =

{x = ( p , 4) E Q :41 2 0}, Q- = {x E Q :q51 s O}. Then Q = Q + u Q-. We
shall prove (57) only for Q+ (which, however, we continue to call Q); a similar
proof holds for Q- so that (57) holds for the given Q. Accordingly, assume
Q = Q+-
For k I m we may write (57) in the form (taking p = 1)
L lul
j=2
sinj-14j n sinv+j-l 4 j " isinv+i-2 4j dp
m- 1
j=k j=m
pv+n-2 d+
K S p [ ( ~ / a ) I u l + / g r a d u l ] ~ ~ i n j - 'n
n-1
5 j=2 j=k 4 ~sinV+j-'
4jpv+n-1dp d4,
(58)
For k > m 2 2 we may write (57) in the form
By virtue of the restrictions placed on v, m, and k in the statement of the

lemma, (58) and (59) are both special cases of
I C ( l / a ) [ u l + I g rj =a1d ~ l ] ~ ~ s i n ~ ~ ~ ~ p ~ ~ ~(60)
-'dpd~,
Q
where 1 5 i I n, pj 2 0, and 0 < pj* I pj i f j 2 i. We prove (60) by backwards
induction on i. For i = n, (60) is obtained by applying Lemma 5.41 to u con-
sidered as a function of p on (0,a ) and then integrating the remaining variables
with the appropriate weights. Assume, therefore, that (60) has been proved for
i = k+ 1 where 1 I k s n - 1. We prove it now holds for i = k.
If B k < n, we have
sin4, I 4 k I K,sin4,, 0s 4 k I Pk, (61)
132 v IMBEDDINGS OF w"'p(a)
Note that K,, and hence K,, depend on Pk but may be chosen independent of
pk, and hence of v, under the conditions of the lemma. If P k = R, we obtain
(62)by writing J; = J ~ " + J ~ , ,and using, in place of (61),the inequalities
sin 4 k I 4 k I(7r/2)sin $k if 0 s4k I4 2
(63)
sin $k IR - +k I(7c/2)sin4, if n/2 I4 k IR.
We now have, using (62)and the induction hypothesis,
This completes the induction establishing (60)and hence the lemma. I

In the following lemma we obtain an imbedding inequality similar to that
of Lemma 5.10 for the domain Q and appropriately weighted LP-norms.
5.45 LEMMA Let Q be as specified by (56) and let p 2 I and 1 Ik 5 n.

-=
Suppose that max(1-k,p-n) c v 1 v2 c co. Then there exists a constant
K = K(k, n,p, vl, v2, B1, ...,fin- '), independent of a, such that for every v
satisfying v 1 5 v Iv2, and every function u E C1(Q) n C @ ) we have
+
where q = (v n)p/(v n -p).+
PROOF Let 6=(v+n-l)p/(v+n-p), s=(v+n-l)/v, s'=(v+n-l)/(n-l).
We have by Holder's inequality and Lemma 5.44 (the case rn = k )
In order to estimate the last integral above we adopt the notation

p* = (41'*..74n-1I, +j* = (~,41,...,~j,...,4n-1), j = 1 , 2 9 . . . y ~I,-
where the caret denotes omission of a component. Let
Qo* = {p* : ( p , p*) Q for 0 < p c a }
E
Qj* = {$j* : ( p , 4 ) E Q for 0 < 4j c pj}.

Qo* and Qj* (1 l j In - 1) are domains in R"' ' . We define functions Fo and
Fj on Qo* and Qj*, respectively, as follows :
L-Fo(P*)ln-' = CFO(41,* . .9 4 n - 1)T-
=
O<p<a
n-1
I7
sup [ I ~ l ' p ~ + ~ - l ]
i=k
n sini-14i,
n-1
i=2
[Fj(4j*)]"-' = [Fj(p , 4 1 , . . . , ~ j , . . . , 4 n - ~ ) I n - '
= sup [Iul'sin'+j- '4j]pv+n-

0 < 61 <81
fl sinv4in sin'-' n sini-24i.

n-1 j- I n- I
x 4i
i=k i=2 i=j+l
134 v IMBEDDINGS OF wm'p(a)
Applying the combinatorial Lemma 5.9, we obtain
$, lulna/(n - 1)rkn v / ( n - 1) dx
Now by Lemma 5.42, and since l a u / d p l I lgradul,
SUP Iuldpv+"-' 5 K2 L l u l d - ' [(l/u)lul+ Igradul]pV+"-l dp,

O<p<a
where K2 is independent of v for 1 - n < v1 I v I v2 .c 00. It follows that
Similarly, by making use of inequality (61) or (63) as in Lemma 5.44, we obtain

from Lemma 5.42
sup [uldsinv+j-'$,
o<+J <bl
since Iau/a$jjl Ip n;;f+ sin $r.Hence

where we have used Lemma 5.44 again to obtain the last inequality. Note that
the constants K 2 , j and K 3 , jcan be chosen independent of u for the values of
v allowed. Substitution of (67) and (68) into (66) and then into (65) leads to
[(l/aP)lulP+ (gradulP]rk'dx
Since (v +n - l)/(v +n) - ( p - 1)/p = l / q , inequality (64) follows by cancel-

lation for, since u is bounded on Q and v > 1 -n, ~QIU1'rk'dX is finite. I
5.46 REMARKS (1) The assumption that u E C(Q) was made only to
ensure that the above cancellation was justified. In fact the lemma holds for
any u E C'(Q).
(2) If 1 -k < v1 < v2 < 00 and v 1 I; v 5 v2, where v 5 p - n , then (64)
holds for any q satisfying 1 I q < 00. It is sufficient to prove this for large q.
If q 2 (v+n)/(v+n-1), then q = (v+n)s/(v+n-s) for some s satisfying
1 5 s < p . Thus
I; K
L [( l/aJ)Iu Is + 1 grad uls] r,Y dx
5K kp-')/'
s, [( l/ap)Iu I + Igrad uIp]r,Y dx {s, r,' dxj(p-s)ip,
which yields (64) since the last factor on the right is finite.
(3) If v = m,a positive integer, then (64) can be obtained very simply as
follows. Let y = ( x , z ) = (xl, ...,x,, zl, ...,z,,,) denote a point in W"'"'and
define u*(y) = u(x) for x E Q. If
Q* = { y E R"'"' : y = (x,z), x E Q,0 < zj < rk(x), 1 I;j 5 m},
then Q* has the cone property in UP+"',
whence by Theorem 5.4 we have,
136 v IMBEDDINGS OF wrnSp(R)
+ +
putting q = (n m)p/(n m -p),
=
[( l/ap)Iu I + Igrad u Ip] rkmdx
since Igradu*(y)l= lgradu(x)l, u* being independent of z.

(4) Suppose that u E COm([wn),or, more generally, that
~ R ~ ~ u ( x ) ~ p [ r k ( x<
) l vad x
with v as in the above lemma. If we take pi = II, 1 5 i I;n - 1, and let a + 00

in (64), we obtain
This generalizes Sobolev's inequality as given in Section 5.1 1.
We now generalize Lemma 5.15 to allow for weighted norms. It is con-

venient to deal here with arbitrary domains having the cone property, rather
than the special case Q considered above. The following elementary result
will be needed.
5.47 LEMMA Let z E Rk and let R be a domain of finite volume in Rk. If

0 I v < k, then
Ix-z1-"dx I; K(volR)l-vlk,
where the constant K depends on v and k but not on z or R.
PROOF Let B be the ball in [wk having center z and the same volume as R. It
is readily verified that the left side of (69) does not exceed J,Ix-zI-'dx.
But (69) clearly holds for R = B.
5.48 LEMMA Let 0 be a domain with the cone property in [w". Let
1 I k I n and let P be an (n- k)-dimensional plane in R". Denote by r(x) the
distance from x to P. If 0 I v < p - n , then for all u E C'(0) we have
where the constant K may depend on v, n,p, k and the cone C determining the
cone property for n, but not on u.
PROOF Throughout this proof Ai and Ki will denote various constants

depending on one or more of the parameters on which K is allowed to depend
in (70). It is sufficient to prove that if C is a finite cone contained in R having
vertex at, say, the origin, then
For 0 I j I n let A j denote the supremum of the Lebesgue j-dimensional

measure of the projection of C onto Rj, taken over allj-dimensional subspaces
!@of R”. Writingx = (x’, x”) where x’ = ( x l , .. ., x,,-k) and xff= (x,,-k+ l , ..., x,,),
we may assume, without loss of generality, that P is orthogonal to the co-
ordinate axes corresponding to the components of x”. Define
For 0 I t I 1 we denote by C, the cone { t x :x E C } so that C, c C and

C, = C if t = 1. For C, we define the quantities A ,j , S,,and R,(x’) analogously
to the similar quantities defined above for C. Clearly = t j A j . If x E C,
we have
so that
Setting V = vol C and a = supxEc1x1, and integrating the above inequality

over C, we obtain
(72)
Let i denote the orthogonal projection of x onto P. Then v(x) = lx” -z”I.
=-
Since 0 I v < p - n we have p 1, and so by Lemma 5.47
138 v IMBEDDINGS OF wmrP(a)
It follows that
Hence, since v < p - n,

c
Similarly,
Inequality (71) now follows from (72), (74), and (75). 1

5.49 LEMMA Suppose all the conditions of Lemma 5.48 are satisfied and,
in addition, R has the strong local Lipschitz property. Then for all u E C'(Q)
we have
where p = 1 - ( v + n ) / p satisfies 0 < fi < 1 and K is independent of u.
PROOF The proof is the same as that given for inequality (28) in Lemma 5.17,
except that the inequality
Jgradu(z)ldz I K1t"-(V+")/P Igradu(~)1~[r(.z)]'dz} ' I p (77)
is used in (29) in place of the special case v = 0 actually used there. Inequality
(77) is obtained in the same way as (73) above. 1
Proofs of Theorems 5.35-5.37
5.50 LEMMA Let $20. If V > p - n , let 1 I qs ( V + n ) p / ( $ + n - p ) ;

otherwise let 1 I q < 00. There exists a constant K = K(n,p,V) such that for
PROOFS OF THEOREMS 5.35-5.37 139
every standard cusp domain Q k , A (see Section 5.34) for which (A- l)k =
v I V, and every u E C' (Qk,A), we have
IIullo.4, Q k , A ' llull 1,P.Qk.A ' (78)
PROOF Since each Qk, A has the segment property, it suffices to prove (78) for
u E C' (G).We first do so for given k and 1and then show that K may be
chosen so as to be independent of these parameters.
First suppose V > p - n . It is sufficient to prove (78) for
q = (ij+n)/(V+n-p).
For u E C' (KA) define C ( y ) = u(x), where y is related to x by (47) and (48).
Thus 6 E C' (Qk) n C(Qk),where Qk is the standard cone associated with
Qk, A.By Lemma 5.45, and since q I (v+n)p/(v+n-p) we have
Now xi = r:-lyj if 1 Ijl k ; xj = y j if k + l I j S n. Since rkz =

yl2+".+ykz we have
axj Syri-' + (L-l)r;-'yiyj if 1 Ii, j Ik
-ayi
={s, otherwise.
Since rk(Y) I 1 on Qk it follows that
lgradW)l 5 KZ h a d Wl.
Hence (78) follows from (79) in this case. For 5 I p - n and arbitrary q the
proof is similar, being based on Remark 5.46(2).
In order to show that the constant K i n (78) can be chosen independent of
k and A provided v = (A- 1)k I 5, we note that it is sufficient to prove that
there is a constant R such that for any such k,A and all 1) E c'(Qk)n C(0k)
we have
In fact, it is sufficient to establish (80) with Z? depending on k as we can then

use the maximum of R ( k ) over the finitely many values of k allowed. We
distinguish three cases.
140 v IMBEDDINGS OF Wmvp(Q)
CASEI V < p - n , 1 s q c m . ByLemma5.48wehaveforO1~1V,
sup
XEQk Iv(x)l + lgradv(y)lpl[rk(y)lvd~}l'p. (gl)
5 K(V){ik[lv(Y)p
Since the integral on the right decreases as v increases we have K ( v ) I K(V)

and (80) now follows from (81) and the boundedness of Q,.
CASEI1 1 > p-n Again it is enough to deal with q = (V+n)p/(V+n-p).

From Lemma 5.45 we obtain
where s = (v+n)p/(v+n-p) 2 q and K , is independent of v for p - n c

vo Iv IV. By Holder's inequality, and since r k ( y )I1 on Q,, we have
so that if vo Iv IV, then (80) follows from (82).

If p-n c 0, we can take vo = 0 and be done. Otherwisep 2 n 2 2. Fixing
vo = (V-n+p)/2, we can find v1 such that 0 I v1 S p - n (or v1 = 0 i f p = n)
such that for v1 s v Ivo we have
+ +
(V n) (V n)p P
l I t =
+
(v+n)(V+n) (V-v)p
I-,
1+EO
where e0 > 0 and depends only on V, n, andp. Because of the latter inequality
we may also assume t-n c v l . Since (v+n)r/(v+n - r ) = q we have, again by
Lemma 5.45 and Holder's inequality,
where K 2 is independent o f v for v1 I v I yo.

In the case v1 > 0 we can obtain a similar (uniform) estimate for
0 I v Iv1 by the method of Case I. Combining this with (82) and (83), we
prove (80) for this case.
CASE111 V = p - n , 1 Iq c 00 Fix s 2 max(q,n/(n- 1)) and let t =

(v+n)s/(v+n+s) so that s = (v+n)t/(v+n-t). Then 1 I t s p s / ( p + s ) c p
PROOFS OF THEOREMS 5.35-5.37 141
for 0 I v I 7. Hence we can select v l 2 0 such that t - n < v1 < p - n . The rest
of the proof is similar to Case 11. This completes the proof. I
5.51 PROOFOF THEOREM 5.35 By the same argument used in the proof of
Lemma 5.12 it is sufficient to consider here only the special case m = 1. Let q
+ +
satisfy p I q I (v n)p/(v n - p ) if v + n > p , or p I q < co otherwise.
Clearly q < np/(n- p ) if n > p so in either Case we have by Theorem 5.4
I1410,q.G 5 K l I I ~ I I l , P , G
for every u E C' (0)and that element G of r which has the cone property (if
such G exists.) If G E r does not have the cone property, and if $: G + Q k , l ,
where (A- 1)k 5 v, is the 1-smooth mapping specified in the statement of the
theorem, then by Theorem 3.35 and Lemma 5.50
II'II0,q.G K211uo$-'l10,q,Qk,rl K311uo$-1111,p,Qr,rl K411uI11,p,G,
where K4 is independent of G. We have, therefore, noting that q / p 2 1,
where we have used the finite intersection property of r to obtain the final
inequality. Imbedding (50) now follows by completion. [If v < mp-n, we
require that (50) hold for q = CO. This is a consequence of Theorem 5.36
proved below.] I
5.52 LEMMA Let 0 I 7 < mp - n. There exists a constant K = K ( m , p ,n, V)

such that if Qk,Ais any standard cusp domain for which (A- 1)k = v I V and
if u E Cm(Qk,,),then
lu(x)l KII'llm,P,Qk,rl* (84)
xEQk.r
PROOF It is sufficient to prove the lemma for the case m = 1 ; the proof for
general m then follows by the same type of argument used in the last paragraph
of the proof of Lemma 5.15.
If u E C' ( Q k , J , (A-1)k = v I V, we have by Lemma 5.48 and via the
method of the second paragraph of the proof of Lemma 5.50,
SUP lu(x)l = SUP lfi(Y)I
xEQk,i YEQk
K z { 1 Q k . A [lu(x)lp + lgradu(x)lP] dx}'". (85)

142 v IMBEDDINGS OF wmn,p(R)
Since rk(y)I 1 for y E Qk it is evident that K,, and hence K2,can be chosen
independent of k, 1provided 0 5 v = ( A- 1)k S 5. I
5.53 PROOFOF THEOREM 5.36 It is sufficient to prove (51). Let u E Cm(sl).

If x E R, then x E G c R for some domain G for which there exists a l-smooth
transformation $: G ( A - 1)k I v, as specified in the statement of the
theorem. Thus
(86)
where K , and K2 are independent of G. The rest of the proof is similar to the
first paragraph of the proof of Lemma 5.15. I
5.54 PROOF OF THEOREM 5.37 As in Lemma 5.17 it is sufficient to prove that

(52) holds when j = 0 and m = 1, that is, that
holds when v + n c p and 0 c p 5 1 -(v+n)/p. For x , y E R, Ix-yl> 6, (87)

holds by virtue of (86). If Ix-yl c 6, there exists G with x , y E G c !2, and a
l-smooth transformation $ from G onto a standard cusp Q k , I with
(1- 1)k 5 v, satisfying the conditions of the theorem. Inequality (87) can then
be derived from Lemma 5.49 by the same method used in the proof of Lemma
5.52. The details are left to the reader.
VI
Compact Imbeddings of Wm9(a)
The Rellich-Kondrachov Theorem
6.1 Let R be a domain in R" and let R, be a subdomain of R. Let X ( Q )

denote any of the possible target spaces for imbeddings of Wm*P(R),that is,
X ( n ) is a space of the form C,j(R), Cj*'(Csi>,Lq(Rk),or w j ~ ~ ( Rwhere
~ ) , Rk,
1 I k S n, is the intersection of R with a k-dimensional plane in R". Since the
linear restriction operator ino:u -,#Ino is bounded from X(R) into X(R,)
[in fact [linou;X(R,)II I Ilu; X(Q) 1
11 any imbedding of the form
Wrn*P(Q)-+ X(Q) (1)
can be composed with this restriction to yield the imbedding
Wrn*P(R)-+ X(S-2,) (2)
and (2) has imbedding constant no larger than (1).
If R satisfies the hypotheses of the Sobolov imbedding Theorem 5.4 and
if R, is bounded, then, with the exception of certain extreme cases, all imbed-
dings (2) (corresponding to imbeddings asserted in Theorem 5.4) are compact.
The most important of these compact imbedding results originated in a
lemma of Rellich [57] and was proved specifically for Sobolev spaces by
Kondrachov [33]. Such compact imbeddings have many important applica-
tions in analysis, especially to showing the discreteness of the spectra of linear
elliptic partial differential operators defined over bounded domains.
143
144 VI COMPACT IMBEDDINGS OF W""(LR>
We summarize the various compact imbeddings of W"pp(Q)in the following

theorem.
6.2 THEOREM (The Rellich-Kondrachou theorem) Let R be a domain

in R",R, a bounded subdomain of R, and $2; the intersection of Qo with a
k-dimensional plane in R". Letj, m be integers,j 2 0, m 2 1, and let 1 Ip < 03.
PART I If R has the cone property and mp I n, then the following imbed-
dings are compact :
Wj+'"*P(R)
-t Wiiq(Rgk) if 0 < n - mp < k <n and
1I 4 < kP/(n-mP), (3)
Wj+'".P(R)-+ Wj*q(Rok) if n = mp, 1 I X I n and
l<q<co. (4)
PART 11 If R has the cone property and mp > n, then the following
imbeddings are compact:
Wj+m.p(R)-t C / ( R ) , (5)
~ j + ' " * p (-t
Q )Wj.q(Rok) if 1 I q I 03. (6)
PART III If R has the strong local Lipschitz property, then the following
imbeddings are compact :
wj+"*p(R)+ Cj(Pi,) if mp > n, (7)
wi+m.p(n) --+ (no)
cj.i if mp > n 2 ( m - 1 ) p and
0 < A < m-(n/p). (8)
PART IV If R is an arbitrary domain in R", all imbeddings (3)-(8) are

compact provided Wj+'"*P(R)
is replaced by Wi+'"*P(R).
6.3 REMARKS ( 1 ) If X , Y, and 2 are spaces for which we have the im-
beddings X + Y and Y -t 2 and if one of these imbeddings is compact, then
the composite imbedding X - t 2 is compact. Thus, for example, if Y + 2 is
compact, then any sequence {ui} bounded in X will be bounded in Y and
therefore have a subsequence {u,'} convergent in 2.
Since the extension operator u --* u" where u"(x) = u ( x ) if x E R, u"(x) = 0
if ~ $ defines
0 an imbedding Wi+mvp(R)-t Wj+m.P(R") by Lemma 3.22,
Part IV of Theorem 6.2 follows from the application of Parts 1-111 to W".
THE RELLICH-KONDRACHOV THEOREM 145
(2) In proving the compactness of any of the imbeddings (3)-(8) it is

sufficient to consider only the case j = 0. Suppose, for example, that (3) has
been proven compact if j = 0. For j 2 1 and {ui} a bounded sequence in
Wj+'".P(R)it is clear that {D'u,} is bounded in Wm3p(R)for each M: such that
la15 j . Hence {D"ui} is precompact in Lq(Rgk)with y specified as in (3). It is
possible, therefore, to select (by finite induction) a subsequence {u,'} of {u,} for
which {D"u,l} converges in Lq(Rgk) for each M such that IQI I j . Thus {u,'}
converges in Wjsq(Rok)and (3) is compact.
(3) Since Ro is bounded, CBo(R,k)--t Lq(Rt) for 1 I q I co;in fact,
I I # I I O , ~ , ~ O L 1124; cBo(f20k)II[VolkROk]1'q. Thus the COmpaCtIleSS Of (6) (for
j = 0) follows from that of (5).
(4) For the purpose of proving Theorem 6.2 the bounded subdomain R,
of R may always be assumed to have the cone property if R does. If C is a
finite cone determining the cone property for R, let fi be the union of all
finite cones congruent to C, contained in R and having nonempty intersection
with RoeThen Ro c c R and fi is bounded and has the cone property. If
W'",p(R) -b X(fi) is compact, then so is WmpP(R) -b X(R,) by restriction.
Note that if R is bounded, we may have f2, = R in the statement of the

theorem.
6.4 PROOFOF THEOREM 6.2, PART111 If mp > n 2 ( m - 1)p and 0 < Iz <
(rn- n)/p, then there exists ,u such that 2 < p < m - (n/p).Since Ro is bounded,
the imbedding Co.p(320) --t C0.'((35,) is compact by Theorem 1.31. Since
W'"*p(R)-b Coi"((n)+ Co*"(O,)by Theorem 5.4 and restriction, imbedding

(8) is compact for j = 0 by Remark 6.3(1).
If mp > n, let j * be the nonnegative integer satisfying ( m - j * ) p > n 2
(rn -j * - 1)p. Then we have the imbedding chain
W'"*P(R) -b Wm-'**p(R) + Coqn,) -b C(ll,) (9)
where 0 < p < m -j * - (n/p).The last imbedding in (9) is compact by Theorem
1.31. Thus (7) is compact for j = 0. I
6.5 PROOFOF THEOREM 6.2, PART 11 As noted in Remark 6.3(4), Romay be

assumed to have the cone property. Since R, is also bounded it can, by
Theorem 4.8, be written as a finite union, R, = ur=,
Qk, where each Rk has
the strong local Lipschitz property. If mp > n, then W'"*p(n) -b W m S P ( R -+ k )
C(n,), the latter imbedding being compact as proved above. If {ui} is a
sequence bounded in WmsP(R), we may select (by finite induction on k ) a
subsequence {u,'} whose restriction to Rk converges in C(0k) for each k,
1 k 2 M. But then {u,'} converges in CBo(Ro) proving that ( 5 ) is compact
f o r j = 0. Therefore (6) is also compact by Remark 6.3(3). I
146 VI COMPACT IMBEDDINGS OF Wrn"(Q)
6.6 LEMMA Let R be a domain in R", 62, a subdomain of R, and 0: the

intersection of Ro with a k-dimensional plane in 03" (1 Ik In). Let
1 Iq1 < qo and suppose
W"JyR) + LqO(Rok)) (10)
W",P(R) + Lq'(Rok). (1 1)
Suppose also that (1 1) is compact. If q1 5 q < qo, then the imbedding
W"-P(R) -+ Lq(nok) (12)
(exists and) is compact.
PROOF Let 2 = 4d40 - d ) / q ( q o- 41) and P = qo(4- 41)/4(40 - 41). Clearly

I > 0 and p 2 0. By Hdder's inequality and (10) there exists a constant Ksuch
that for all u E Wm*p(R),
5 lI~ll~,ql,C)o~ll~116.qo,no~
ll~llo,q,Ro~
5 KIIUIl~,q,,Ihk Ilullkp,n. (13)
Let {ui} be a sequence bounded in W"*p(R). Since (1 1) is compact there exists
a subsequence {u;} that converges, and is therefore a Cauchy sequence in
Lq1(QOk). By (13), {ui'} is a Cauchy sequence in L q ( O t )as well. Hence (12) is
compact. 1
6.7 PROOFOF THEOREM 6.2, PARTI First we deal with the case j = 0 of
imbeddings (3). Assume, for the moment, that k = n and let qo = np/(n- mp).
In order to prove that the imbeddings
Wrn*w-2) + L4(R0), 1 Iq < 40, (14)
are compact it suffices, by Lemma 6.6, to do so only for q = 1. For
j = 1,2,3, ... let
0, = {x E Q0 : dist(x, bdryR) > 2/j}.
Let S be a set of functions bounded in W'"*P(O).We show that S (when
restricted to no)is precompact in I.' (no)by showing that S satisfies the
conditions of Theorem 2.21. Accordingly, let E > 0 be given and for each
rX)
u E W"pp(S2)set
C(x) =
ifotherwise.
xERo
By Holder's inequality and since W"*'(62) -+ Lqo(ao), we have
THE RELLICH-KONDRACHOV THEOREM 147
with K, independent of u. Since R, has finite volume, j may be selected large

enough to ensure that for every u E S,
j
no-nj
lu(x)ldx <
and also, for every h E R",
1no * nj
Iu"(X+h) - u " ( X ) l d X < E/2. (15)
Now if Ihl c l/j, then x+ th E R,, provided x E Rj and 0 S t S 1. If u E Cm(R),

it follows that
5 I4 l l ~ l l l , l , * o 5 Kz lhl I I ~ l l m , p , n , (16)
where K, is independent of u. Since P(R) is dense in Wm,p(R), (16) holds for
any u E Wm*p(sZ).Hence if Ihl is sufficiently small, we have from (15) and (16)
that
jnolu"(x+h)- u"(x)I < E .
Hence S is precompact in L'(R,) by Theorem 2.21, and imbeddings (14) are

compact.
Next suppose k < n but p > 1. Let r be chosen so that 1 c r c p and
n-mr c k. Let v be the largest integer less than mr; let s = kr/(n-mr), and
let q = nr/(n-mr). Assuming, as we may, that R, has the cone property we
obtain from inequalities (35) and (36) in the proof of Lemma 5.19,
IIUllo,I,nok I K3 IIUllo,s,nok
K4 ll4i,q,& lulA;:&
1-I.
5 K5 IIUll;,q,no IIUllm.p,n, (17)
where 1 = n (mr - v)/mr(n- v) satisfies 0 c A < 1, and where K3, K4, and K ,
are independent of u. Note that 1 c q c 4,. If {ui} is a sequence bounded in
Wm,P(n),we have shown that it must have a subsequence {u,'} which con-
verges in Lq(Ro). From (17), {ui'} must be a Cauchy sequence in L'(ROk)so
that Wm*p(R)+ L ' (nok) is compact. By Lemma 6.6 so are the imbeddings
Wm*p(R)+ LQ(0,k)for 1 I q < kp/(n-mp).
148 VI COMPACT IMBEDDINGS OF Wmn,'(n)
Finally, suppose p = 1 and 0 I n -m < k < n. Then clearly n - m 1 I +

k < n so that 2 I m I n. By Theorem 5.4, W".' (R) -+ W m - ' * r ( Q )where
r = n/(n- 1) > 1. Also, k 2 n - ( m - 1) > n - ( m - 1)r so the imbedding
Wm-l,r (0)+ I,' (a): is compact as proved above. This is sufficient to com-
plete the proof of the compactness of (3).
To show that (4) is compact we proceed as follows. If n = mp, p > 1, and
1 < q < c o , we may select r such that 1 I r n - m r > O , and
kr/(n- mr) > q. Assuming again that Ro has the cone property, we have
W"t*P(R)-+ WmJ(no) -+ Lyngk). (18)
The latter imbedding in (18) is compact by (3). If p = 1 and n = m 2 2, then,
setting r = n/(n- 1) > 1 so that n = (n- I)r, we have for 1 s q < 00,
W"J(0) 3 w"-IJ(R) + LyR:),
the latter imbedding being compact as proved in (18). Finally, if n = m = p = 1,
then of necessity k = 1. Letting qo > 1 be arbitrary chosen, we prove the com-
pactness of W'*'(R) -+ L'(Ro) exactly as in the case k = n of (3) considered
above. Since W'.'(R) -+ Lg(Sao)for 1 I q < 00 all these imbeddings are
compact by Lemma 6.6. 1
6.8 The reader may find it instructive to carry out the obvious generalization
of Theorem 6.2 to the imbeddings supplied by Theorems 5.35-5.37.
Two Counterexamples
6.9 Two obvious questions arise from consideration of the statement of the
Rellich-Kondrachov Theorem 6.2. First, can that theorem be extended to
cover unbounded Ro ? Second, can the "extreme cases"
Wj+m.p(R)+ Wj,q(Rt), 0 < n - mp < k In,
4 = kP/(n-mP) (19)
and
w i + m . ~ ( ( n -+
) cj.i(no), mp > n > ( m - l)p,
,I= m-(n/p) (20)
ever be compact ?
The first of these questions will be investigated later in this chapter. For
the moment we show that, at least for k = n, the answer is certainly "no"
unless R, is quasibounded, that is, unless
lim dist(x, bdry 0,) = 0.
xe%
Ixl-rm
TWO COUNTEREXAMPLES 149
6.10 EXAMPLE Let R be an unbounded domain in R" which is not

quasibounded. Then there exists a sequence { B i } of mutually disjoint open
balls contained in R and all having the same positive radius. Let 41E COm(Bl)
andsuppose /lcj,llk,p,s,=Ak,p>Oforeachk=0,1,2 ,... a n d p 2 1 . L e t 4i
be a translate of 41 having support in B i . Then clearly {4i} is a bounded
sequence in Wd""*p(R) for any fixed j , m,p. But for any q,
so that {d)i} cannot have a subsequence converging in WjVq(R).Thus no

imbedding of the form Wd+n',P(R)+ W i , q ( f 2 can
) be compact. The non-
compactness of the other imbeddings of Theorem 6.2 is proved similarly. I
We now show that the second question raised in Section 6.9 always has a
negative answer.
6.11 EXAMPLE Let R be any domain in R" and R, any bounded sub-
domain of 0.Let nokbe the intersection of R, with a k-dimensional plane in
R", say (without loss of generality) the plane spanned by the x l , ...,xk co-
ordinate axes. Let {al, a * , .. .} be a sequence of distinct points in Rgk, and
, a sequence of numbers such that 0 < r i I 1, such that Br,(ai)=
{ r l r r Z...}
{ x E R" : Ix-a,l< r i } c a, and such that all the balls Brt(a,)are mutually
disjoint.
Let 4 E COm(B,(O)) satisfy the following conditions:
(i) For each nonnegative integer h, each real q 2 1 , and each k,
1 I k I n, we have
l 4 l h , q. R k n B t ( 0 )
=I
I 4 l h . q . Rk =
c
,
1.1 = h = 0
ak + =...=an
IIDa4118,q,RknnBj(0)
(ii) There exists a E Bl(0),a # 0, such that for each nonnegative integer h,
)"' = Ah,q,k 'O4
1Dlh4(a)l= Bh > 0. (21)

Fix p 2 1 and integers j 2 0 and m 2 1. For each i let
4i(x) =
j + m-n/p
Ti 4 ((x-ai)/ri).
Then clearly 4i E Com(Bri(a,))and a simple computation shows that
h +k i q A
I(Pilh,q,Rk =
j + m-nip-
h,q.k' (22)
If h Ij + m , it follows from (22) and ri I 1 that
I4ilh.p.W" I Ah,p.n
so {4i}is a bounded sequence in W j ' m 9 p ( R ) .
150 VI COMPACT 1MBEDDlNGS OF W m ’(0)
Suppose mp < n and n-mp < k In. Taking q = kp/(n-mp), we obtain

from (22)
lI4iIIi,q,nok 2 1 4 i l j , q , W k = A j , q , k *
Since the functions 4i have disjoint supports, we have

II4i-4hh.q.Ibk 2 21’qAi.q.k ’
and so no subsequence of {4i} can converge in WJ*q(Rok).
Thus imbedding
(19) cannot be compact.
On the other hand, suppose mp > n > ( m- 1)p and let I = m - (n/p).
Letting 6 , = ai+r,a, we obtain from (21)
= rABj> 0.
1D1’41(bi)1= rm-(”/P)IDli4(a)l
Let ci = ai+ari/lal so that ci E bdryB,,(a,) and Ibi-cil = (1 -\ul)ri. Again
since have disjoint supports,
Thus no subsequence of {4i} can converge in C j * A ( a oand

) imbedding (20)
cannot be compact. 1
Unbounded Domains-Compact Imbeddings of W? (a)

6.12 Let R be an unbounded domain in R”. We shall be concerned below
with determining whether the imbedding
w,m*’(n) P(n) --t (23)
is compact. If (23) is compact, it will follow as in Remark 6.3(2) and the
second paragraph of Section 6.7 that the imbeddings
Wi+m*p(R)
4 Wj*q(nk), 0 < n - mp < k In, p q < kp/(n-mp),
and
Wd+m*”(&2)
+ Wi*q(Rk), n = mp, 1 I k In, p Iq < co
are also compact.
UNBOUNDED DOMAINS-COMPACT lMBEDDlNGS OF worn. (a) 151
As was shown in Example 6.10, imbedding (23) cannot be compact unless

R is quasibounded. In Theorem 6.13 we give a geometric condition on R that
is sufficient to guarantee the compactness of (23), and in Theorem 6.16 we
give an analytic condition that is necessary and sufficient for the compactness
of (23). Both theorems are from the work of Adams [2].
Let R, denote the set { x E R : 1x1 2 r } . Be it agreed that in the following
discussion any cube H referred to will have its faces parallel to the coordinate
planes.
6.13 THEOREM Let v be an integer such that 1 I v I n and mp > v (or

p = m = v = 1). Suppose that for every E > 0 there exist numbers h and r
with 0 < h I 1 and r 2 0 such that for every cube H c R” having edge length h
and having nonempty intersection with R, we have
p,,-v(H,R)/h -’2 hPIE,
where p,,-,(H, R) is the maximum, taken over all projections P onto (n-v)-
dimensional faces of H , of the area [that is, (n - v)-measure] of P ( H R).
Then imbedding (23) is compact.
-
6.14 The above theorem shows that for given quasibounded R the com-
pactness of (23) may depend in an essential way on the dimension of bdryR.
Let us consider the two extreme cases v = 1 and v = n. For v = n the condition
of the theorem places on 0 only the minimal restriction of quasiboundedness.
Thus if mp > n, then (23) is compact for any quasibounded R. It can also be
shown that if p > 1 and R is quasibounded and has boundary consisting
entirely of isolated points with no finite accumulation point, then (23) cannot
be compact unless mp > n.
If v = 1, the conditions of Theorem 6.13 make no requirement of m and p
but do require that bdryR be “essentially (n- 1) dimensional.” Any quasi-
bounded domain whose boundary consists of reasonably regular (n - 1)-
dimensional surfaces will satisfy these conditions. An example of such a
domain is the “spiny urchin” (Fig. 5), a domain in R2obtained by deleting from
the plane the union of all the sets s k ( k = 1,2, . . .) specified in polar coordinates
by
s k ={ ( r , e ) : r 2 k , e = n r r k/ 2 , n = 1 ,,...,
2 2k+1}.
Note that this domain, though quasibounded, is simply connected and has
empty exterior.
More generally, if v is the largest integer less than mp the conditions of

Theorem 6.13 require that in a certain sense the part of the boundary of R
having dimension at least n - v should bound a quasibounded domain.
152 VI COMPACT IMBEDDINGS OF wmSp(n)
FIG.5 A "spiny urchin."
6.15 Let H be a cube of side h in R" and E a closed subset of H. Given m and
p we define a functional I;*P on C m ( H by
)
I?P(U) = C ~ , ~C
h j ~ I u l g ,= hlalPJ~l ~ ~ u ( x ) l p d x .
1sj5m 1 51.1 dm H
We denote by Cm9j'(H,E) the "(m,p)-capacity" of E in H defined by
where C"(H, E) is the set of all functions u E C"(H) that vanish identically
in a neighborhood of E. Clearly C m V p ( H , E I )C ' " + l * P ( H , Eand
) for
E c F c H , Cm*P(H, E ) I Cm*P(H F).
, Theorem 6.13 will be deduced from
the following theorem characterizing in terms of the above capacity those
domains for which (23) is compact.
6.16 THEOREM Imbedding (23) is compact if and only if R satisfies the

following condition: For every E > 0 there exists h I 1 and r 2 0 such that the
-
inequality
C m S p ( HH, Q) 2 hPI&
holds for every n-cube H having edge length h and having nonempty inter-
section with n,. (This condition clearly implies that R is quasibounded.)
6.17 LEMMA There exists a constant K = K ( n , p ) such that for any

n-cube H of edge length h, any measurable subset A of H with positive volume,
UNBOUNDED DOMAINS--COMPACT IMBEDDINGS OF worn' (0) 153
and any u E C' (H), we have
PROOF Let y E A and x = (p, 4) E H, where ( p , 4) are spherical polar co-

ordinates centered at y , in terms of which the volume element is given by
dx = w(4)p"-'dp d4. Let bdryH be specified by p = f ( 4 ) , 4 E Z. Clearly
f ( 4 ) I G h . Since
we have by Holder's inequality,
lU(X)lPdX
Integration of y over A using Lemma 5.47 leads to
(volA)Ilull(,p,HI, 2p-'h"Ilu118,p,A+ Kh"+pLlgradu(x)lpdx
from which (24) follows at once. I

6.18 PROOFOF THEOREM 6.16 (Necessity) Suppose Q does not satisfy the
condition stated in the theorem. Then there exists a finite constant K , = I/&
such that for every h with 0 < h 5 1 there exists a sequence { H j } of mutually
disjoint cubes of edge length h which intersect R and for which
C m * p ( H jHi
, - 0)< K,hP.
-
By the definition of capacity, for each such cube Hi there exists a function
uj E C"(Hj, H j Q) such that IIujllg,p,Hj= h", ~~graduj~~g,p,,,j 5 K , h", and
j A j = (x E H j : luj(x)l < +}. By Lemma 6.17 we have
, p , H Let
~ ~ u j ~ ~ ~ K,(h).
2p- 'h" vol A j KK, h2,,+
h"<-.-
volAj 2p +-
154 VI COMPACT IMBEDDINGS OF W""(Q)
from which it follows that volAj I K 3 h n f P .Let us choose h so small that

K 3 hP I 3, whence vol A j I 3 vol Hi. Choose functions wj E Com(Hj)such that
wj(x) = 1 on a subset Sj of I f j having volume no less than 3 vol H j , and
such that
sup max sup lDawj(x)l = K 4 < 00.
-
i IalsrnxEHj
Then vi = uj wj E Com(HjnQ) c Com(R)and Iuj(x)l 2 4 on Sj n (Hi A j ) , a

set ofvolume not less than h"/3. Hence 2 .h"/3.2P.Onthe other hand,
provided Ic11,1/31 s m. Hence { u j } is a bounded sequence in Wg*P(R).Since

the functions uj have disjoint supports, 1 uj - uk1 g, p , 2 2h"/3.2p so imbedding
(23) cannot be compact.
(Sufficiency) Now suppose R satisfies the condition in the statement of
the theorem. Let E > 0 be given and choose r 2 0 and h I1 such that for every
cube H of edge h meeting R, we have C m s P ( HH, 0)2 hP/EP.Then for
every u E Co"(Q)we obtain
-
11 1 '0, p . H I ( E P / h p ) I;' ('> I E p 1 11 b, p . H *
Since a neighborhood of R, can be tesselated by such cubes H we have

by summation
I14 0 , P,R, I 8 lI4lm. p , n*
That any bounded S in Wg*p(R)is precompact in Lp(Q)now follows at once
from Theorems 2.22 and 6.2. I
6.19 LEMMA There is a constant K independent of h such that for any

cube H in 03" having edge length h, for every q satisfying p I q I np/(n- mp)
( o r p 5 q < co if n = mp, or p 5 q I 00 if n < mp), and for every U E C m ( H )
we have
PROOF We may suppose H to be centered at the origin and let R be the cube
of unit edge concentric with H. Inequality (25) holds for a by the Sobolev
imbedding theorem. For given u E C"(H) we define corresponding I2 E C"(k)
by Q ( y ) = u ( x ) where x = hy. Since
UNBOUNDED DOMAINS-COMPACT IMBEDDlNGS OF wr*(Q) 155
it follows by change of variable that
whence (25) follows. I

6.20 LEMMA If mp 2 n (or m = p = n = l), there exists a constant
K = K ( m , p , n ) such that for every cube H of edge length h in R" and every
u E Cm(H)vanishing in a neighborhood of some point y E H, we have
~ ~ u ~ ~ 5
~ , p , HKz;".P(u)*
PROOF The proof is somewhat similar to that of Lemma 5.15. First suppose
p In < mp. Let ( p , 4) denote polar coordinates centered at y. Then
If n > ( m - l)p, let q = np/(n-mp+p) so that q > n. Otherwise let q > n be

arbitrary. If ( p , 4) E H, then by Holder's inequality
It follows, using Lemma 6.19 with m - 1 replacing m, that
IluIl8,q,,H5 K 2 h 4 j H lgradu(x)lqdx
K2hq 1
la1 = 1
IIDuullZ,q,H
A further application of Holder's inequality yields
~ ~ up , H~ 5
~ !I l u,l l ! , q , H(vO1 H ) ( 4 - P ) ' q
If p > n (or p = n = l), the result follows directly from (26) with q = p,
llUll{,,,,~ 5 KzA'p(U) KI;;'*'(U).
156 VI COMPACT IMBEDDINGS OF Wrn"(R)
6.21 PROOFOF THEOREM 6.13 Let H be a cube for which, for m p > v (or
m = p = v = 1) and pn-,(H,R)/hn-v2 hP/&.Let P be the maximal projection
referred to in the statement of the theorem, and let E = P(H R). Without N
loss of generality we assume that the (n- v ) dimensional face F of H con-

taining E is parallel to the x V + , , ...,x, coordinate plane. For each point
,,
x = (x'"'') in E, where x' = ( x , , ..., x,) and x" = ( x v + ...,x,,), let Hx- be
-
the v-dimensional cube of edge h in which H intersects the v-plane through x
normal to F. By definition of P there existsy E Hx. $2. If u E C m ( H ,H R),
then u(.,x'') E Cm(Hxp,,y).Applying Lemma 6.20 to u(.,x''), we obtain
6.. Iu(x', x'')lCdx' 5 K,

1slals.m
hlalp/
HXl
ID"u(x',x")lpdx',
where K, is independent of h, x", and u. Integrating this inequality over E and

denoting H' = {x' :x = (x', x") E H for some x"}, we obtain
lIu118,p,H'xE - K1Z&':EE(u)5 K,Z,"*P(u).
<
Now we apply Lemma 6.17 with A = H'x E so that vol A = h'p,-,(H,O).
This yields
Ln-v
where K2 is independent of h. It follows that
Hence R satisfies the hypothesis of Theorem 6.16 if it satisfies that of Theorem

6.13. I
The following two interpolation lemmas enable us to extend Theorem

6.13 to cover imbeddings of W;*P(R) into continuous function spaces.
6.22 LEMMA Let 1 I p < 01) and 0 < p I 1. There exists a constant
K = K ( n , p , p ) such that for every u E Com(Rn)we have
where 1 = p p / ( n +pp).
PROOF We may assume

UNBOUNDED DOMAINS-COMPACT IMBEDDINGS OF wz’(a) 157
Let E satisfy 0 < E I N / 2 . There exists xo E R” such that lu(xo)12 N - E 2. N / 2 .

c (f o r all x, so
Now ~ u ( x o ) - u ( x ) ~ / ~ x o - xI~ M
lWl2 W0)I- MIX,-XI’ 2 3lu(xo)l
provided Ix-xoI I (N/4M)’l” = r. Hence
Since this holds for arbitrarily small E we have

llp 2 . 4 ~ / ~ ~ 3 ~ 1 + ~ I ~ ~ ~ - - n l ~ c (
IIuIIo,p,P” 2 (Kl I
from which (27) follows at once. I
6.23 LEMMA Let R be an arbitrary domain in R”, and let 0 < 1 < p I 1.
For every function u E C0*@((n)
we have
112.4; COq3)ll I 3’-”” Ilu; C(3Z)II l-A/c 112.4; C0qn)lp”. (28)
PROOF Let p = p/1, p’ = p / ( p - l), and let
6.24 THEOREM Let Q satisfy the hypothesis of Theorem 6.13. Then the
following imbeddings are compact :
Wd+mgp(R)+ C j ( Q if mp > n (29)
WJ+m.p(n) -, (Ti) if mp > n 2 ( m - 1 ) p and
0 < 1 < m- (n/p). (30)
158 VI COMPACT IMBEDDINGS OF wrnSp(n)
PROOF It is sufficientto deal with the casej = 0. If mp > n, letj* be the non-
negative integer satisfying ( m - j * ) p > n 2 (m-j*- 1)p. Then we have the
chain
W;sp(R) 3 W;-j*,p(,) + CO-"(Si) + c(n),
where 0 c p c m - j * - ( n / p ) . If {ui} is a sequence bounded in W;*p(R), then
{u,} is also bounded in Co.B(Q. By Theorem 6.13, {u,} has a subsequence
{u,'} converging in Lp(R).By (27), which applies by completion to the func-
tions u,, (u,'} is a Cauchy sequence in C(0)and so converges there. Hence (29)
is compact f o r j = 0. Furthermore, if mp > n 2 (m- 1)p (that is, j * = 0) and
0 < I < p, then by (28), {u;} is also a Cauchy sequence in C0*'((si)whence
(30) is also compact. I
An Equivalent Norm for W;.p(R)

6.25 Closely related to the problem of determining for which unbounded
domains R the imbedding W;sp(R) -,LP(R) is compact, is that concerned
with determining for which domains R the seminorm I I n t , p , n defined by
is actually a norm on W$.p(n), equivalent to the given norm 1 - I(m,p,n. Such

is certainly the case for any bounded domain as we now show.
6.26 A domain R c W" is said to have finite width if it lies between two
parallel hyperplanes. Let R be such a domain and suppose, without loss of
generality, that R lies between the hyperplanes x, = 0 and x, = d. Letting
x = (x',xn) where x' = ( x l , ...,x,- 1), we have for any 4 E Co"(R)
so that
and
AN EQUIVALENT NORM FOR wf"(Q) 159
Successive application of the above inequality to derivatives D a 4 , la( Im- I ,

then yields
I4lrn,p,n IIl4llrn,p.n 5 Kl4lm.p.n (32)
and by completion (32) holds for all ME Wr*p(R). Inequality (31) is often
called PoincarP's inequality.
6.27 An unbounded domain R in R" is called quasicylindrical provided

lim sup dist(x, bdry R) < 03.
X E R, 1x1-+ m
Evidently every quasibounded domain is quasicylindrical, as is every (un-

bounded) domain of finite width. We leave to the reader the construction of a
suitable counterexample to show that if R is not quasicylindrical, then
I . I r n , p , R is not an equivalent norm to 1 IIm, p , n on Wt*P(R).
a
The following theorem is clearly an analog of Theorem 6.13.
6.28 THEOREM Suppose there exist constants K , R, h, and v with

0 < K I1, 0 5 R < 03, 0 < h < 03, and 1 I;v n, v an integer, such that
either v < p or v = p = 1, and such that for every cube H in R", having edge
length h and having nonempty intersection with nR= {x E C! : 1x12 R } we
have
p n - v ( f I , n)/h"-"
2 K,
where p,,-,(H,R) is as defined in the statement of Theorem 6.13. Then
1. Im,p,R and 11 . Il,a,p,n are equivalent norms for W;,p(R).
PROOF As noted in Section 6.26 it is sufficient to prove that I I u I I ~ , ~ , I~
K , ( u (l , p , R for u E Com(R).Let H be a cube of edge length h having nonempty
intersection with nR.Since v < p (or v = p = 1) the proof of Theorem 6.13
(Section 6.21) shows that
C','(H, H N R) 2 pn-,,(H,fi)/K2h"-' 2 K / K z
for all u E Com(H),K2 being independent of u. Hence
llu118,p,H I( K 2 / K ) G * P ( u =
) K3 lUl:.,,H. (33)
By summing (33) over the cubes H comprising a tesselation of some neigh-
borhood of R, , we obtain
Il~Ilp0,P,RR5 K 3 l ~ l : , p , R * (34)
It remains to be proven that
IIUl18,P,BR 5 K414:,p,R9
160 VI COMPACT IMBEDDINGS OF wm'p(fi)
where BR = {x E R" : 1x1< R}. Let (p, 4) denote spherical polar coordinates
of the point x in R" ( p 2 0 , 4 E C) and denote the volume element by dx =
p " - ' o ( 4 ) dp d4. For any u E Cm(Rn)we have
so that
JP
Hence
Unbounded Domains-Decay at Infinity
6.29 The vanishing, in a generalized sense, on the boundary of R of elements

of Wc-p(R) played a critical role in our earlier establishment of the compact-
ness of the imbedding
W , m q 2 ) 3 LP(R) (35)
for certain unbounded domains. For elements of W"*P(R) we no longer have
this vanishing and the question remains: when, if ever, is the imbedding
wm*p(n)
-+ LP(R) (36)
compact for unbounded R, or even for bounded R which are sufficiently
irregular that no imbedding of the form
W"*P(R) -+ L4(R) (37)
exists for any q > p ? Note that if R has finite volume, the existence of (37) for
UNBOUNDED DOMAINS-DECAY AT INFINITY 161
some q > p implies the compactness of (36) by the method of the first part of
Section 6.7. By Theorem 5.30 imbedding (37) cannot, however, exist if q > p
and R is unbounded but has finite volume.
6.30 EXAMPLE F o r j = 1,2, . .. let Bj be an open ball in R" having radius

r j , and suppose Bj n Bkis empty i f j # k. Let R = lJT= Bj; R may be bounded
or unbounded. The sequence { u j } defined by
Uj(X) = { m l B j ) - l ' p if x e B j
if x $ Bj
is clearly bounded in Wm*p(R)but not precompact in LP(R) no matter how
rapidly rj + 0 as j tends to infinity. Hence (36) is not compact. [Note that
(35) is compact by Theorem 6.13 provided limj+, rj = 0.1 If R is bounded,
imbedding (37) cannot exist for any q > p .
6.31 There do exist unbounded domains R for which imbedding (36) is

compact (see Section 6.48). An example of such a domain was given by
Adams and Fournier [3] and it provided a basis for an investigation of the
general problem by the same authors [4].The approach of this latter paper is
used in the following sections. First we concern ourselves with necessary
conditions for the compactness of (37) ( q 2 p ) . These conditions involve rapid
decay at infinity for any unbounded domain (see Theorem 6.40). The techniques
involved in the proof also yield a strengthened version of Theorem 5.30 (viz.
Theorem 6.36) and a converse of the assertion [see Remark 5.5(6)] that
Wm*P(R)+ Lp(R) for 1 Iq < p if R has finite volume.
A sufficient condition for the compactness of (36) is given in Theorem 6.47.
It applies to many domains, bounded and unbounded, to which neither the
Rellich-Kondrachov theorem, nor any generalizations of that theorem
obtained by the same techniques can be applied (e.g., exponential cusps-see
Example 6.49).
6.32 Let T be a tesselation of R" by closed n-cubes of edge length h. If H i s

one of the cubes in T, let N(H) denote the cube of side 3h concentric with Hand
having faces parallel to those of H . The N ( H ) will be called the neighborhood
of H . Clearly N ( H ) is the union of the 3" cubes in T which intersect H. By the
fringe of H we shall mean the shell F ( H ) = N ( H ) H .
N
Let R be a given domain in R" and T a given tesselation as above. Let

I > 0. A cube H E Twill be called I-fat (with respect to R) if
p W n Q) > A p ( F ( H ) n Q),
where p denotes n-dimensional Lebesgue measure in R". (We use ''p" instead
of "vol" for notational simplicity in the following discussion where the symbol
must be used many times.) If H is not I-fat, it is called A-thin.
162 VI COMPACT IMBEDDINGS OF Wrn"(Q)
6.33 THEOREM Suppose that there exists a compact imbedding of the

form
W"*P(Q)+ Lq(n) (38)
for some q 2 p. Then for every I > 0 and every tesselation T of R" by cubes of
fixed size, T has only finitely many I-fat cubes.
PROOF Suppose, to the contrary, that for some I > 0 there exists a tesselation
T of R" by cubes of edge length h containing a sequence { H j } T z 1 of I-fat
cubes. Passing to a subsequence if necessary we may assume that
N(Hj)n N(H,) = 0 if j # k. For each j there exists + j E COa(N(Hj))such
that
(i) 14j(x)I 5 1 for all x E R",
(ii) d j ( x ) = 1 if x E H j ,
(iii) ID"+j(x)l I M for all x E R" and 0 I JaJ
Im,
where M = M ( n ,m,11) is independent of j. Let $j = cj 4 j , where the positive
constant cj is so chosen that
But then
) Q)
M P c j P p ( N ( H jn
c M p c j p p ( H jn Q)[1 +(l/I)] = M P [ l +(l/I)]cj'-q,
since Hi is I-fat. Now p(Hjn Q) Ip ( H j ) = h" so cj 2 h-"Iq. Since p-q I0,

{ $ j } is bounded in Wrn3"(Q).Since the functions $ j have disjoint supports,
{tjj}cannot be precompact in Lq(Q), contradicting the compactness of (38).
Thus T can possess only finitely many I-fat cubes. I
6.34 COROLLARY Suppose there exists an imbedding (38) for some

q > p . If T is a tesselation of R" by cubes of fixed edge length h, and if I > 0 is
given, then there exists E > 0 such that p ( H n Q) 2 E for every I-fat H E T.
PROOF Suppose, to the contrary, there exists a sequence { H j } of A-fat cubes

with limj+oop ( H j n n ) = 0. If cj is defined as in the above proof, we have
~ ,0, , psince
limj+, cj = 00. But then l i m j . + ~ ~ ~ $ j ~ = , n p c q. Since {t,hj} is
bounded away from 0 in Lq(Q) we have contradicted the continuity of
imbedding (38).
6.35 Let us consider the implications of the above corollary. If imbedding

(38) exists for some q > p , then one of the following alternatives must hold:
(a) There exists E > 0 and a tesselation T of R" consisting of cubes of
fixed size such that p ( H nQ) 2 E for infinitely many cubes H E T.
(b) For every I > 0 and every tesselation T of Iw" by cubes of fixed size,
T contains only finitely many A-fat cubes.
We shall show in Theorem 6.37 that (b) implies that Q has finite volume.
By Theorem 5.30, (b) is therefore inconsistent with the existence of (38) for
9 > p , On the other hand, (a) implies that p ( ( x ~ t 2 : N I I xSl N + I } ) does
not tend to zero as N tends to infinity. We have therefore proved the following
theorem strengthening Theorem 5.30.
6.36 THEOREM Let Q be an unbounded domain in R" satisfying

lim sup vol {x E Q : N I 1x1 IN + l} = 0.
N-rUJ
Then there can be no imbedding of the form (38) for any q > p .
6.37 THEOREM Suppose that imbedding (38) is compact for some q 2 p .

Then CI has finite volume.
PROOF Let T be a tesselation of R" by cubes of unit edge, and let I =

1/[2(3"- l)]. Let P be the union of the finitely many 1-fat cubes in T. Clearly
p ( P n R ) I p ( P ) c 00. Let H be a I-thin cube. Let H , be one of the (3"- 1)
cubes of T contained in the fringe F ( H ) and selected so that p ( H l n G) is
maximal. Thus
p ( H n G ) I Ap(F(H) n G) I 1(3"- l ) p ( H , n G) = + p ( H , n G).
If HI is also I-thin, we may select H , E T, H , c F ( H , ) such that p ( H , n Q) I
+P ( H 2 nQ).
Suppose an infinite chain { H , H I ,H , , ...} of I-thin cubes can be con-
structed in the above manner. Then
p ( H n Q) I + ( H I n 0)I ... I ( 1 / 2 j ) p ( H j n a) I 1/2j
for each j , since p ( H j n R ) I p ( H j ) = 1. Hence p ( H n Q) = 0. Denoting by
P, the union of I-thin cubes H E T for which such an infinite chain can be
constructed, we have p(Pm n a) = 0.
Let Pj denote the union of the I-thin cubes H E T for which some such
chain ends on thejth step (that is, H j is I-fat). Any particular I-fat cube H '
can occur as the end H j of a chain beginning at H only if H is contained in the
cube of edge 2j+ 1 centered on H ' . Hence there are at most (2j+ I)" such cubes
164 VI COMPACT IMBEDDINGS OF wm’”(n)
H c Pj having H‘ as chain endpoint. Thus
I (1/2j) 1 p ( H j n Q)
HcPj
so that CT==,
p ( P j nQ) < co. Since 08” = P v P, u Pl u Pz v ... we have
PL(f-Q<co* I
Suppose 1 I q < p. By Theorem 2.8 the imbedding
Wrn,P(Q)-+ Lq(Q2) (39)
exists if vol CI< 00. We are now in a position to prove the converse.
6.38 THEOREM Suppose imbedding (39) exists for some p,q such that
1 I q < p . Then Q has finite volume.
PROOF Let T, I be as in the proof of the previous theorem. Once again let P
denote the union of the I-fat cubes in T. If we can show that p ( P n Q) is finite,
it will follow by the same argument used in the above theorem that vol Q is
finite.
Accordingly, suppose p ( P nQ) is not finite. Then there exists a sequence
{Hj}T=l of I-fat cubes in T such that CT==, p ( H j nQ)= co. If L is the lattice
of centers of the cubes in T, we may break up L into 3” mutually disjoint
sublattices { Lj};I, each having period 3 in each coordinate direction. For each
i let Ti be the set of all cubes in T with centers in Li. For some i we must have
C1-fat,HET,p(HnQ) = co. Thus we may assume the cubes of the sequence
{ H i } all belong to Ti for some fixed i, so that N ( H j ) n N ( H , ) do not
overlap.
Let the integerj, be chosen so that
1 p ( H j n Q) -= 4.
jl
2I
j=1
Let bj be as in the proof of Theorem 6.33, and let
$l(X) = 2-l’p ci l
j= 1
C#lj(X).
We have, since the supports of the functions bj are disjoint, and since the
cubes Hi are I-fat, for la1 I m,
6.39 If there exists a compact imbedding of the form (38) for some q 2 p ,
then, as we have shown, R has finite volume. In fact, considerably more is
true: p ( { x E R :1x1 2 R } ) must tend very rapidly to zero as R -+ co,as we now
show.
If Q is a union of cubes H in some tesselation of R", we extend the notions
of neighborhood and fringe to Q in an obvious manner:
Given 6 > 0, let I = 6/3"(1 +a). If all of the cubes H c Q are I-thin, then
Q is itself S-thin in the sense that
AQn a) 5 &(F(Q) n 0). (40)
To see this note that as H runs through the cubes comprising Q,F ( H ) covers
166 VI COMPACT 1MBEDDlNGS OF W"l"(fi)
N(Q) at most 3" times. Hence
= 3"ACp(Q n Q)+p(F(Q) n a11

from which (40) follows by transposition (permissible since p(R) < co) and
since 3"L/(l- 3"A) = 6.
For any measurable set S c R" let Q be the union of all cubes H of our
tesselation whose interiors intersect S, and define F ( S ) = F(Q). If S is at a
positive distance from the finitely many A-fat cubes in the tesselation, then Q
consists of I-thin cubes and we obtain from (40),
p ( S n R) Ip(Q n R) ISp(F(S)n Q). (41)
6.40 THEOREM Suppose there exists a compact imbedding of the form

(38) for some q 2 p . For each r 2 0 let R, = {x E R : 1x1 > r } , let S, =
{x E R : 1x1 = r } , and let A , denote the surface area [(n- 1)-measure] of S,.
Then :
(a) For given E , 6 > 0 there exists R such that if r 2 R, then
p(R,) 6 p ( { x E R : r--E I1x1Ir } ) .
(b) If A , is positive and ultimately nonincreasing as r tends to infinity,
then for each E > 0
lim-A , + , -
- 0.
r-w A,
PROOF Given E > 0 let T be a tesselation of R" by cubes of edge length

-~/2$. Then any cube H E T whose interior intersects Rr is contained in
Or - e / ~and
F(R,) c {x E R : r--E I 1x15 r } .
For given 6 > 0 let I be as given in Section 6.39. Let R be large enough that the
finitely many A-fat cubes in T are all contained in the ball of radius R - E / ~
centered at the origin. Then for any r 2 R all the cubes in T whose interiors
intersect R, are I-thin and (a) follows from (41).
For (b) choose R , so that A , is nonincreasing on [R,, 00). Fix E ' , 6 0, =-
and let E = ~ ' / 2 Let
. R be as in (a). If r 2 max(R,+d, R), then
A,+,, I r + " A , ds I (l/~)p(R,+,) ( 6 / ~ ) p {ExR : r I1x1 Ir + E }

r+e
= ( 6 / ~r)+ ' A S ds I 6A,.

r
Since E' and 6 are arbitrary, (b) follows. I

6.41 COROLLARY If there exists a compact imbedding of the form (38)

for some q 2 p, then for every k we have
lim ekrp(R,)= 0.
r-rm
PROOF Fix k and let S = e--(k+l).From conclusion (a) of Theorem 6.40

there exists R such that r 2 R implies p(R,+ I Sp(R,). Thus i f j is a positive
integer and 0 5 t < 1, we have
ek(R+ j + r ) k ( R + j + 1) k(R+1) kj j
p(nR+j+r) < p(nR+j) 5 p(nR)
- ek(R+l)p(nR)e-j.
The last term tends to zero as j tends to infinity. I
6.42 REMARKS (1) The argument used in the proof of Theorem 6.40(a)
works for any norm p on R" in place of the usual norm p(x) = 1x1. The same
holds for (b) provided A, is well defined (with respect to the norm p) and
provided
p({x E R : r s p(x) I r + E } )
This is true, for example, if p(x) = max,
=
l+e A, ds.
(xiI.
(2) For the proof of (b) it is sufficient that A, have an equivalent, positive,
nonincreasing majorant, that is, there should exist a positive, nonincreasing
function f ( r )and a constant M > 0 such that for all sufficiently large r,
A , If ( r ) IM A , .
(3) Theorem 6.33 is sharper than Theorem 6.40, because the conclusions
of the latter theorem are global whereas the compactness of (38) evidently
depends on local properties of R. We illustrate this by means of two examples.
6.43 EXAMPLE Let f E C 1([0,a))be positive and nonincreasing with

bounded derivativef'. We consider the planar domain (Fig. 6a)
n = {(x,y) E R2 : x > 0, 0 < y < f ( x ) } . (42)
With respect to the supremum norm on Rz, that is, p(x,y) = max(lxl,lyl),
we have A, = f ( s ) for sufficiently large s. Hence R satisfies conclusion (b) of
Theorem 6.40 [and sincefis monotonic conclusion (a) as well] if and only if
holds for every E > 0. For example, f ( x )= exp(-x2) satisfies (43) but
168 VI COMPACT IMBEDDINGS OF W""(a)
FIG.6 Domains: (a) R of Example 6.43 and (b) fi of Example 6.44
f ( x ) = e-' does not. We shall see (Section 6.48) that the imbedding
W",P(n) + LP(R) (44)
is compact if (43) holds. Thus (43) is necessary and sufficient for the compact-
ness of (38) for domains of the type (42).
6.44 EXAMPLE Let f be as in Example 6.43 and assume also that

f ' ( 0 ) = 0. Let g be a positive, nonincreasing function in C'([O, 00)) satisfying
(9 d o ) = 3f(O), g'(0) = 0,
(ii) g ( x ) < f ( x ) for all x 2 0,
(iii) g(x) is constant on infinitely many disjoint intervals of unit length.
Let h ( x ) = f ( x ) - g ( x ) and consider the domain (Fig. 6b)
?i = { ( x , y ) E IW' : 0 cy c g ( x ) i f x 2 0, 0 < y < h ( - x ) i f x < 01.
UNBOUNDED DOMAINS--COMPACT IMBEDDINGS OF W m S(a)
P 169
Again we have A, =f(s) for sufficiently large s so fi satisfies the conclusions

of Theorem 6.40 if (43) holds.
+
If, however, T is a tesselation of R2 by squares of edge having edges
parallel to the coordinate axes, and if one square of T has center at the origin,
then T has infinitely many +-fat squares with centers on the positive x-axis.
By Theorem 6.33, imbedding (44) cannot be compact.
Unbounded Domains-Compact Imbeddings of W"."(a)
6.45 The above examples suggests that any sufficient condition for the
compactness of the imbedding
W"*P(R) 4 LP(R) (45)
for unbounded domains R must involve the rapid decay of volume locally in
each branch of n, as r tends to infinity. A convenient way to express such local
decay is in terms of flows on R.
By a flow on R we mean a continuously differentiable map CP: U -+ R,
where Uis an open set in R x R containing R x {0}, and where @(x,0) = x for
every x E R.
For fixed x E R the curve t 4CP(x, t ) is called a streamline of the flow.
For fixed t the map CPt: x + @(x,t ) sends a subset of R into R. We shall be
concerned with the Jacobian of this map:
It is sometimes required of a flow CP that as+,

= CPs 0 0,but we do not need
this property and so do not assume it.
6.46 EXAMPLE Let R be the domain given by (42). Define the flow
@(X,Y,f) = (X--t,Cf(x--))/f(X)IY), 0 < t < x.
The flow is toward the line x = 0 and the streamlines diverge as the domain
widens (see Fig. 7). CP, is a local magnification for t > 0:
det @,'(x,Y ) = f(x - t)/f(x).
Note that lim,,,det CPt'(x,y) = co iff satisfies (43).
For N = 1 , 2 , ... let R , * = ( ( x , y ) ~ Q : o < x < N } . QN* is bounded and
has the cone property, so the imbedding
W'.P(aN*)--t LP(R,*)
is compact. This compactness, together with properties of the flow CP are
sufficient to force the compactness of (45) as we now show.
170 VI COMPACT IMBEDDINGS OF Wrnsp(R)
f'
0 x-t X
*
X
FIG.7 Streamlines of the flow UJgiven in Example 6.46
6.47 THEOREM Let R be an open set in R" having the following

properties :
(a) There exists a sequence {R,*}~,l of open subsets of R such that
R,* c R$+l and such that for each N the imbedding
W'.P(R,*) -+ Y(R,*)
- a,*,
is compact.
(b) There exists a flow @: U + R, such that if R, = R then
(i) RNx [O, 13 c U for each N,
(ii) Qt is one-to-one for all t ,
(iii) I(d/at)@(x, r)ls M (const) for all ( x , t ) E U.
(c) The functions d,(t) = sup,,,,ldet Or'(x)1-' satisfy
(i) lim,+m d,(l) = 0,
(ii) lim,+w Jh d,(t) dt = 0.
Then imbedding (45) is compact.
PROOF Since we have Wm,p(f2) -+ W'*p(R)+ LP(R)it is sufficient to prove
that the latter imbedding is compact. Let u E C' (R). For each x E RN we have
Now
UNBOUNDED DOMAINS-COMPACT IMBEDDINGS OF Wm'(n) 171
Also
f l r
Putting 6, = max(d,(l), M j i d N ( t )dr), we have

n
and limN+a BN = 0.
Now suppose u is real valued and belongs to C1($2)n W 1 * p ( n )By
.
Holder's inequality the distributional derivatives of lulp,
Dj(lulp)= p.lulP-' asgnu- Dju,
satisfy
JnlDi(lu(x)l')ldx ~ ~P I~I ~u I~I ; ~

P~~Djullo,p,5 , ~~, ~~. ~ *
Thus lulp E W ' * ' ( n )and by Theorem 3.16 there is a sequence c#Jj of functions
in C'(n) n W ' . ' ( n ) such that limj,a - [ u ~ P ~ ~ l ,=l , 0.
n Thus by (46)
lu(x)JPdx = lim 4j(x)dx s lim ~up~Nll~jll1,1,~

Jn, j+a InN j-4)
= ~NIII~IPIIl,l,nK6NIlullP,p,n, (47)
where K = K(n,p). Inequality (47) holds for arbitrary (complex-valued)
u E C'(n) n W1*P(L2)by virtue of its separate application to the real and
imaginary parts of u. (The constant K may be changed.)
) E > 0, we may, by (47), select N so
If S is a bounded set in W 1 * p ( Qand
that for all u E S
j-Nlu(x)lpdx < &*
Since W'*P(CI- nN) -

+ LP(n RN) is compact, the precompactness of S in
LP(R)follows by Theorem 2.22. Hence W'sp(CI)-+ Lp(f2) is compact. I
6.48 EXAMPLE Consider again the domain R of Examples 6.43 and 6.46
and the flow Q, given in the latter example. We have
172 VI COMPACT IMBEDDINGS OF wmpp(n)
and by (43)
lim d N ( t )= 0 if t > 0.
N-rm
Thus by dominated convergence
lim l d N ( t ) dt = 0.
N+m
The assumption thatf' is bounded guarantees that the speed I(a/at)@ ( x , y ,f ) l

is bounded on U . Thus R satisfies the hypotheses of Theorem 6.47 and (45) is
compact for this domain.
6.49 EXAMPLE Theorem 6.47 can also be used to show the compactness
of (45) for some bounded domains to which neither the Rellich-Kondrachov
theorem nor the techniques used in its proof can be applied. For example, we
consider
R = {(XJ) E RZ : 0 < x < 2, 0 < y <f ( x ) } ,
where f~ C1(0,2) is positive, nondecreasing, has bounded derivative f', and
satisfies 1irnx+,+ f ( x ) = 0. Let u = { ( x ,y , r ) E I W :~( x ,y ) E R, - x < t < 2 - x }
and define the flow #: U + R by
@(x,y,t) = x+t, ( fZ:,".>*

-
Thus det mt'(x,y) = f ( x + t ) / f ( x ) . If R N * = { ( x , ~E)R : x > l/N}, then
satisfies &(I) 5 1 for 0 I t 5 I , and limN-rm d N ( t )= 0 if t > 0. Hence also

limN+m!,j d N ( t )dt = 0 by dominated convergence. Since RN* is bounded and
has the cone property, and since the boundedness of a#/at is assured by that
off', we have by Theorem 6.47 the compactness of
W",P(R) --t LP(R). (48)
Suppose lirn,,,, f'(x)/xk = 0 for every k. [For example, let f ( x ) = e-'/,.]
Then R has an exponential cusp at the origin, and so by Theorem 5.32 there
exists no imbedding of the form
W"*P(R) + Lq(n)
for any q > p. Hence the method of Section 6.7 cannot be used to show the
compactness of (48).
HILBERT-SCHMIDT IMBEDDINGS 173
6.50 REMARKS (1) It is easy to imagine domains more general than

those in the above examples to which Theorem 6.47 can be applied, although
it may be difficult to specify a suitable flow. A domain with many (perhaps
infinitely many) unbounded branches can, if connected, admit a suitable flow
provided the volume decays sufficiently rapidly and regularly in each branch,
a condition not fulfilled by the domain d of Example 6.44. For unbounded
domains in which the volume decays monotonically in each branch Theorem
6.40 is essentially a converse of Theorem 6.47 in that the proof of Theorem
6.40 can be applied separately to show the volume decays in each branch in
the required way.
(2) Since the only unbounded domains R for which Wm*”(R)imbeds
compactly into L”(R) have finite volume there can be no extension of Theorem
6.47 to give compact imbeddings into Lq(R)( q > p ) , CB(R),etc.-there do not
exist such imbeddings.
Hilbert-Schmidt Imbeddings
6.51 A complete orthonormal system in a separable Hilbert space X is a

sequence { e i } s , of elements satisfying
[where (-, .)x is the inner product in X I , and such that for each x E X we have
Thus x = C ;, (x, ei)xei,the series converging with respect to the norm in X.

It is well known that every separable Hilbert space possesses such complete
orthonormal systems. There follows from (49) the Parseval identity
Let X and Y be two separable Hilbert spaces and let {ei>E1and {fi}glbe
given complete orthonormal systems in X and Y, respectively. Let A be a
bounded linear operator with domain X taking values in Y, and let A* be the
adjoint of A taking Y into X and defined by
(X,A*Y)X = (AX,Y),, x E X , Y E y.
Define
174 VI COMPACT IMBEDDINGS OF w"'p(!d)
If JIIA111 is finite, A is called a Hilbert-Schmidt operator and we call IIIAIII

its Hilbert-Schmidt norm. (Recall that the operator norm of A is given by
We must justify the above definition.
6.52 LEMMA The norms lllAlll and (IJA*JI(are independent of the

particular orthonormal systems {e,}, {A}used. Moreover,
IllAlll = IlIA*lIl 2 IIAII.

PROOF By Parseval's identity
=
m m
1 1 I(%A*fi)XIZ
j = 1 i=1
= 2 IIA*fi;XIIZ
j= 1
= IlIA*lll.
Hence each expression is independent of {ei}and {A}. For any x E X we have
We leave to the reader the task of verifying the following assertions:

(a) If X, Y, and Z are separable Hilbert spaces and A, B bounded linear
operators from X into Y and Y into Z, respectively, then B 0 A, which takes X
into 2, is a Hilbert-Schmidt operator if either A or B is. (If A is Hilbert-
Schmidt, then I I l B o Allls IlBllIllAlll.)
(b) Every Hilbert-Schmidt operator is compact.
The following theorem, due to Maurin [43] has far-reaching implications

for eigenfunction expansions corresponding to differential operators.
6.53 THEOREM Let R be a bounded set having the cone property in R".
Let m, k be nonnegative integers with k > 4 2 . Then the imbedding mappings
W"+k*2(R)+ wm*2(R) (50)
HILEERT-SCHMIDT IMBEDDINGS 175
are Hilbert-Schmidt operators. Similarly, the imbeddings

W;+k*2(,) + W:*2(Q) (51)
are Hilbert-Schmidt operators for any bounded domain R.
PROOF Given y ER and 2 with la1 s m we define a linear functional T: on

wm+k.2
(a)by
T,”(u) = D”u(y).
Since k > 4 2 the Sobolev imbedding Theorem 5.4 implies that 7’’’is bounded
2 and has norm bounded by a constant K independent of y and
on W m + k .(0)
U:
IT,”(u)l max suPIDau(x)l 5 KIIullm+k,2,R. (52)

0s;lalsmxeR
By the Riesz representation theorem for Hilbert spaces there exists

u,” E W m + k * 2 (such
R ) that
Duu(y) = T,” (u) = (u, V,”)m + k 9 (53)
where (., is the inner product in W m + k . 2 ( Qand
) , moreover
llu,”llm+k.2,R = IIT,”;CWm+k’2(R)1’! 5 K* (54)
If { e i } s , is a complete orthonormal system in Wm+k*2(R),
then
Consequently,
Hence imbedding (SO) is Hilbert-Schmidt. The same proof holds for (51)
without regularity needed for the application of Theorem 5.4. I
The following generalization of Maurin’s theorem is due to Clark [17].
6.54 THEOREM Let p be a nonnegative, measurable function defined on

the domain R in R”. Let W:*2;P(Q) be the Hilbert space obtained by com-
pleting Com(R)with respect to the weighted norm
For y ER let 7(y) = dist ( y , bdry R). Let v be a nonnegative integer such that
176 VI COMPACT IMBEDDINGS OF wm”.(R)
If k > v+n/2, then the imbedding

w$+k*2(Q) -+ wz.2;” (a) (57)
(exists and) is Hilbert-Schmidt.
PROOF Let {e,}, T;, and uyb be defined as in the previous theorem. If y E R,
let yo be chosen in bdry R such that z (y) = I y -yo I. If v is a positive integer
and u E Com(R), we have by Taylor’s formula with remainder
D’uW = c (1/B!)Da+8u(YB)(Y-Y8)b
181= v
for some points y8 satisfying Jy-ysI 5 ~ ( y )If. la1 < m and k > v+n/2, we
obtain from the Sobolev imbedding theorem [as in (52)]
Ipu(y)l 5 Kllullm+k,2[z(y)lv* (58)
Inequality (58) holds, by completion, for any u E W$+k*Z(R),
and also holds
if v = 0, directly from (52). Hence by (53) and (54)
Iltl~lln+k.Z = IDau(y)l 5 K [ z ( y ) l v *
11 Ull m + k.1= 1
It finally follows as in (55) that
K 21.1 c
sm j*Cr(Y)l’vP(Y)dv < 00
by (56). Hence imbedding (57) is Hilbert-Schmidt. I

6.55 REMARK Various choices of p and v lead to generalizations of
Maurin’s theorem for imbeddings of the sort (51). If p ( x ) = 1 and v = 0, we
obtain the obvious generalization to unbounded domains of finite volume. If
p ( x ) = 1 and v > 0, R may be unbounded and even have infinite volume, but
by (56) it must be quasibounded. [Of course quasiboundedness may not be
sufficient to guarantee (56).] If p is the characteristic function of a bounded
subdomain R, of R, and v = 0, we obtain the Hilbert-Schmidt imbedding
Wc+k*2(f2) -+ Wm*z(Ro), k > 42.
F'ractional Order Spaces
Outline
7.1 In this chapter we are concerned with the problem of extending the notion
of Sobolev space to allow for nonintegral values of m. There is no unique
method for doing this, and different approaches may or may not lead to
different families of spaces. The major families of spaces that arise are the
following:
(i) The spaces W"*P(R)-which can be defined by a "real interpolation"

method but can also be characterized in terms of an intrinsically defined norm
involving first-order differences of the highest-order derivatives involved.
(ii) The spaces Ls*p(R)-which can be defined by a "complex inter-
polation" method but, if R = R", can also be characterized in terms of Fourier
transforms.
(iii) The Besov spaces B'*'(R)-defined in terms of an intrinsic norm
similar to that of Ws*P(12)but involving second rather than first differences.
(iv) The Nikol'skii spaces HsiP(R)-having norm involving Holder
conditions in the LP-metric.
Only the spaces WssP(R) and Ls*p(R)coincide with W"'n,p(R)when s = rn,

an integer. B'~p(f2)coincides with Ws*p(R)for all s ifp = 2 but otherwise only
for nonintegral s. The space Hs*p(f2)is always larger than (but close to)
177
178 VII FRACTIONAL ORDER SPACES
Ws*p((n). From the point of view of imbeddings, the simplest and most com-
plete results obtain for the spaces BSsPand H S i P (see Theorem 7.70 and
Section 7.73). However, it is in terms of the W s v pspaces that the problem,
mentioned in Section 5.20, of characterizing the traces on smooth manifolds
of functions in W’”9p((n)has its solution (Theorem 7.53). For this reason we
concentrate our effort in this chapter on elucidating the properties of the
spaces Ws*P(Cl) and give only brief descriptions of the other classes.
About half of the chapter is concerned with developing the “trace inter-
polation method” of Lions, on which we base our study of the spaces Ws*p(f2).
These latter spaces are introduced in Section 7.36. The trace interpolation
method is one of several essentially equivalent real’interpolation methods for
Banach spaces concerning which there is now a considerable literature.
Descriptions of these methods may be found in the work of Butzer and Berens
[13] and Stein and Weiss [65], and the interested reader is referred to the
work of Peetre [56] and Grisvard [28] for some applications in the direction
of fractional order Sobolev spaces. A treatment of these spaces is also given
in Stein [64a]. Most of the material in this chapter follows that of Lions [37,
381 and Lions and Magenes “1.
The Bochner Integral
7.2 In this chapter we shall need to make frequent use of the notion of
integral of a Banach space-valued functionfdefined on an interval of R. We
begin therefore with a brief discussion of the Bochner integral, referring the
reader to the text by Yosida [69], for instance, for further details and proofs
of our assertions.
Let B be a Banach space with norm denoted by 11. JIB. Let {Al, A,, ..., A,,,}
be a finite collection of mutually disjoint, measurable subsets of R, each having
finite measure, and let {bl, b,, ..., b,,,} be a corresponding collection of points
of B. The function f on R into B defined by
xA being the characteristic function of A, is called a simple function. For

simple functions we define, obviously,
where p ( A ) denotes the (Lebesgue) measure of A.

Let A be a measurable set in R and f an arbitrary function defined a.e. on
A into B. The function f is called (strongly) measurable on A if there exists a
THE BOCHNER INTEGRAL 179
sequence {f,}of simple functions with supports in A such that

lim [ l f n ( t ) - f ( t ) l l B = 0 a.e. in A. (1)
n-rm
Let ( ., .) denote the pairing between B and its dual space B’ (i.e., b’(b) =
( b , b‘), b E B, b’ E B‘). It can be shown that any function f whose range is
separable is measurable provided the scalar-valued function (f(.), b’) is
measurable on A for each b E B‘.
Suppose that a sequence of simple functionsf, satisfying (1) can be chosen
in such a way that
lim
n-m
I[.& ( t )- f ( t ) 11 B dl = O.
Thenfis called (Bochner) integrable on A and we define
lf(t) dt = lim
n-m SR f , ( t ) dt. (2)
[The integrals on the right side of (2) do converge in (the norm topology of)
B to a limit which is independent of the choice of approximating sequence
{f,I*l
A measurable function f is integrable on A if and only if Ilf(.)llBis
(Lebesgue) integrable on A. In fact,
7.3 Let - co Ia < b S co. We denote by Lp(a,b; B) the vector space of

(equivalence classes of) functions f measurable on (a, b) into B such that
{
I l f ( . ) l l B E LP(a,b). The space LP(a,b; B) is a Banach space with respect to the
norm
{ ~ ~ ~ f ( r ) \ ~ ; d t } ifl ’ p1 I p < GO
Ilf; LP(a,b; B)II =

ess IIf(t)IIB if p = c o .
te(a,b)
Similarly, i f f € Lp(c,d; B) for every c, d with a < c < d < b, then we write
(a, b; B), and, in case p = 1, call f locally integrable.
f E Lf”,
A locally integrable function g on (a, b) is called the j t h distributional
derivative of the locally integrable function,fprovided
[ @ j ) ( t ) f ( t ) dt = (- 1)’ [ + ( t ) g ( r ) dt
b b
Ja Ja
for every (scalar-valued) testing function 4 E 9 ( u , b) = Com(u,b).

Semigroups of Operators and the Abstract Cauchy Problem
7.4 The following sections present a discussion of those aspects of the theory
of semigroups of operators in Banach spaces which we shall require in the
subsequent development of “trace spaces” and the fractional order spaces
WSip(Q). In this treatment we follow the work of Zaidman [70].
7.5 Let B be a Banach space, and L(B) the Banach space of bounded linear
operators with domain B and range in B. We denote the norms n B and
L(B) by I\ llB and I1 * llL(B), respectively.
A function G with domain the interval [0, co) and range in L(B is called
a (strongly) continuous semigroup on B provided
(i) G(0) = I, the identity operator on B,
(ii) G(s)G(t) = G ( s + t ) for all s, t 2 0,
(iii) for each b E B the function G ( . ) b is continuous from [0, co) to
(the norm topology of) B.
We note that (ii) implies that the operators G(s) and G ( t ) commute. Also,
(iii) implies that for each to 2 0 the set {t : 1IG(t)llL(B)> t o } is open in R and
hence measurable. If 0 I to c t , c co,then for each b E B, G ( .)b is uniformly
continuous on [ t o , t l ] and hence there exists a constant Kb such that
IIG(t)bllBIKb if to I t I t , . By the uniform boundedness principle of
functional analysis, there exists a constant K such that IIG(t)llL(B)I K for
all b E B and all t E [to,t l ] . Thus IIG(*)IIL(B)E L;,(O, GO). We amplify this
result in the following lemma.
7.6 LEMMA (a) The limit

lim(llt) log\~G(r)~~L(B)
=
t-tm
exists and is finite.
(b) For each 6 > do there exists a constant M , such that for every t 2 0
IIG(t)IlL(B)5 Maed‘-
PROOF Let N ( t ) = l O g \ l G ( t ) l l L ( B ) . Since

I1G(s+ t ) IIL(B) IIlG(4 I I L I~G ( t )IIL ( B )
we have that N is subadditive;
N ( s + t ) IN(s) + N(t).
Let So = inf,,, N ( t ) / t . Clearly 0 I 6 , c 00. Let E > 0 be given and choose
r > 0 so that N ( r ) / r c SO+&. If t 2 2r, let k be the integer such that
SEMIGROUPS OF OPERATORS AND THE ABSTRACT CAUCHY PROBLEM 181
( k + 1)r I t < ( k + 2 ) r . Then
Now t- kr E [r, 2r) so, as noted above, N ( t - kr) is bounded, say by K. Thus
N ( t ) kr
6 , 1 -1 -(do+&)
t t
+t K
+
The right side tends to a0 8 as t + 03, and (a) follows since E is arbitrary.
If 6 > do, there exists f d such that if t 2 t d , then N ( t ) 1 6t, or equivalently
1 ear.Conclusion (b) now follows with
IIG(r)llL(B)
Md = max(l~supO,r,r,~~G(t)~~L(B,). 1
7.7 For given b E B the quotient ( G ( t ) b - b ) / t may or may not converge
(strongly) in B as t -,O + . Let D(A)be the set of all those elements b for which
the limit exists, and for b E D ( A ) set
G (t )b - b G ( t ) b- G(O)b
Ab = lim = lim
r+o+ t t+O+ t
Clearly D(A)is a linear subspace of B and A a linear operator from D ( A ) into
B. We call A the injinitesimal generator of the semigroup G. Note that A
commutes with G ( t ) (t 2 0) on D ( A ) .
7.8 LEMMA (a) For each b e B
sb
lim ( l / t ) G ( t ) b d t = b.
r+O+
(b) For each b E B and t > 0 we have
Sbc(r)b dt E D ( A ) and A
sb G ( t ) bdt = G ( t ) b - b.
(c) For each b E D ( A ) and t > 0 we have
L G ( ~ ) dt
A ~= ~ ( t ) -b b.
(d) D ( A ) is dense in B.
(e) A is a closed operator in B, that is, the graph {(b, Ab) : b E D(A)} is a
closed subspace of B x B.
PROOF Let b E B. By the continuity of G ( . ) bwe have lim,+o+~ ~ G ( t ) b - b=~ ~ B

lim,,,+IIG(t)b-G(O)bll, = 0. Conclusion (a) follows since b = ( I / t ) & b dt.
For fixed t we have
lim G ( s ) - G ( o ) ~ G ( z ) bdz = lim -

s+o+ S s+o+s 's'0
[ G ( s + z ) - G ( z ) ] bdz
lim ( ~ [ + ' G ( z ) b d r - Isf 'oG ( r ) b d r )

= s+o+
G ( z ) G ( t ) bdr - b = G ( t ) b - 6 .
This proves (b). If b E D(A), then
+O as s + O + .
Thus
G(s)b - b
G ( r ) b d z = lirn L G ( z )
s+o + S
which proves (c). Part (d) is an immediate consequence of (a) and (b).
If b, E D(A), b, + b and Ab, + bo in B, then by (c)
G ( t ) b , - b, = k ( z ) A b ndr.
We may let n + co,justifying the interchange of limit and integral in the same
way as was done in the proof of (c), and obtain
G ( r ) b - b = k ( z ) b o dz.
Thus by (a)
G ( t ) b- b
lirn
t+o+ t
whence b E D(A) and Ab = boaThis proves (e). I
7.9 REMARK It follows from the closedness of A that D ( A ) is a Banach
space with respect to the norm Ilb; D(A)I\= (Ibl\B+ I\Ab\\,,. We have, more-
over, the obvious imbedding D ( A ) --+ B.
SEMIGROUPS OF OPERATORS AND THE ABSTRACT CAUCHY PROBLEM 183
The following theorem is concerned with an abstract setting of a typical

Cauchy initial value problem for a first-order differential operator.
7.10 THEOREM Let A be the infinitesimal generator of a continuous

semigroup G on the Banach space B. Let a E D(A) and letfbe a conti,nuously
differentiable function on [0, co)into B. Then there exists a unique function
u, continuous on [0, 00) into D(A), and having derivative u' continuous on
[0, 00) into B, such that
u'(t) - h ( t ) = f ( t ) , t20
(3)
u(0) = a.
In fact, u is given by
u(t) = G ( t ) a +p(t-t)f(t)
dt. (4)
PROOF (Uniqueness) We must show that the only solution of (3) for
f ( r ) 3 0 and a = 0 is u ( t ) E 0. If u is any solution and t > t , we have
a G ( t - t - S ) u ( t + S ) - G(t - t )U ( T )
- G ( t - T ) u ( t ) = lim
at S-rO S
G(t-2-s) - G(t-t)
= lim u (4
s-rO S
u(t+s) -u(t)
+ limG(t-r-s)
s+o S
= -G(t-t)Au(t) + G ( t - z ) u ' ( r ) = 0.
Thus G ( t - t ) ~ ( t )= G(t)u(O) = 0 for all t > T. Letting t + t + , we obtain
U ( T ) = G(O)u(z) = 0 for all T 2 0.
(Existence) We verify that u given by (4) satisfies (3). First note that
u(0) = a and (d/dt)G(t)a= AG(t)a. Hence it is sufficient to show that the
function
is continuously differentiable on [0, co) into B, takes values in D(A), and

satisfies g'(t) = Ag(f)+f(t).
Now
g(t+s) - g ( t )
S
= ‘1
s o
t+s
G ( t + s - z ) f ( z ) dz - -
s’
s o
G(t-r)f(z) d7
s )- -
G ( ~ - T ) ~ ( T +d~
s’
s o
G ( t - ~ ) f ( 7 d7
)
Sincefis continuously differentiable on [0, 00) into B we have that
g’(t) =
sb G ( t - ~ ) f ’ ( rdt
exists and is continuous on [0, 00) into B. On the other hand,

) + G(t)f(O)
By Lemma 7.8(a) and the continuity off the latter integral converges tof(t)
as s - t O + . This, together with the existence of g ’ ( t ) guarantees that
lim,,o+[G(s)-G(0)]g(t)/s exists, that is, g ( t ) E D(A). Thus g ’ ( t ) =
Ag(t)+f(t) as required. I
The Trace Spaces of Lions
7.11 Let B, and B2 be two Banach spaces with norms II.(JB, and 1). Ile2,
respectively, and let X be a topological vector space in which B , and B2 are
each imbedded continuously (i.e., B, n U is open in B i , i = 1,2, for every U
open in X). Then the vector sum of B1 and B2
B1 + B , = {b1+b2 € X : b i E B,, b2 E B2)
is itself a Banach space with respect to the norm:
IIU;BI+B211 = inf
brsBi
(IlblllBl +~ ~ b 2 ~ ~ B ~ ) *
bi + b 2 = u
Let 1 < p I CQ, and for each real number v let t v denote the real-valued
function defined on [0, 00) by t ” ( t )= t v , 0 I t < 00.
We designate by W ( p ,v ; B 1 ,B2),or when confusion is not likely to occur
simply by W, the vector space of (equivalence classes of) measurable functions
THE TRACE SPACES OF LIONS 185
f on [0, 00) into B, + B, such that

r”fE Lp(O,@;B,) and r’f’ E Lp(O,co;B,),
f’denoting the distributional derivative off. The space W is a Banach space
with respect to the norm
llfll w = Ilf; W(P,00 ;B,, B,) I
= max(lltvf;p(O,co;B1)II, Ilt’f’;LP(o, w;B2)11).
As an example of this construction the reader may verify that
W(p,O;W 1 . P ( R nLp(R”))
), is isomorphic to the Sobolev space W 1 * p ( Q )
whereQ= ( ( x , t ) = ( x , ,...,x , , r ) ~ R ” + ~ : t > O } .
We shall show that for certain values of p and v , functions f in W possess
“traces”f(0) in B, B,. +
7.12 LEMMA Let f E W. There exists b E B, + B, such that
f(t)= b + l f ’ ( 7 ) dz a.e. in (0,co).
Hence f is almost everywhere equal to a continuous function on (0, co) into

(5)
+
B, Bz.
PROOF Since t ” f ~Lp(O,co;Bl), therefore f E t f o c ( O , 00; Bl). Similarly f ’ E

Lfoc(O,co;B,), and the function u defined a.e. on (0, co) into B , 4-B, by
u(t) =f ( r ) dz
1
+
belongs to LP,,(O, 00; B, B,). It follows that the scalar-valued function
( u ( . ) , b ’ ) belongs to Lfoc(O,00) for each b‘ E (B, + B,)’. Thus for every
$ E Com(O,00) we have
= -( [ u ( r ) # ( r ) dr, b’)
= - (6 f(r) Q’(t) dt - Sm$’(t) dr ff’(z) dz, b’)

0 1
= ([f’(t) $ ( r ) dr - S m f ’ ( r ) $ ( r ) dr, b‘
0 ) = 0.
(The change of order of integration and pairing ( ., .) above is justified since

u is integrable on the support of and so can be approximated there by simple
functions for which the interchange is clearly valid.) By Corollary 3.27,

+
) each b’ E (B, B,)’. Thus v ( t ) = b, a
( v ( r ) , b’) is constant a.e. on ( 0 , ~for
fixed vector in E , + E , a.e. on ( 0 , ~ and
) ( 5 ) follows at once. Clearly the integral
in (5) is continuous on (0,00) into B,, and hence into B, B,. +
7.13 LEMMA Suppose ( l / p ) + v < 1. Then the right side of ( 5 ) converges
in B, + B, as r -+ O+. The limit is defined to be the trace,f(O), offat t = 0.
PROOF If 0 < s < t, we have, for 1 < p < 00,
+
The last factor tends to zero as t + 0 since ( l / p )+ v < 1. (The same argument
works with the obvious modifications if p = 1 or p = 00.) Thus ji f ’ ( r ) dz
converges in B, as t O+ , which proves the lemma.
+
7.14 Given real p and v with 1 s p 00 and 0 = ( l / p ) v < 1 we denote by
T ( p ,v ; B,, B2), or simply by T, the space consisting of all traces f ( 0 ) of
functionsfin W = W ( p ,v ; B,, B,). Called the trace space of W, T is a Banach
space with respect to the norm
ll4T = inf
u=f(O)
Ilfllw.
f 6W
T is a subspace of B, + B, and lies topologically “between” B, and B, in a
sense which will become more apparent later.
Before developing some properties of the trace space T we prepare one
further lemma concerning W which will be needed later.
7.15 LEMMA The subspace of W consisting of those functions f e W

which are infinitely differentiable on (0,00)into B, is dense in W if 1 ~p < 00.
PROOFUnder the transformation

t = e‘, f(e*) = f(z),
we have that fE W if and only if
+
where 8 = ( l / p ) v. If J, is the mollifier of Section 2.17, then, just as in Lemma
2.18, J,*fis infinitely differentiable on R into B1 and
+
m
2?+ (eeprIIJ,*f(~)-f(~)llg, e('-l)prll(J~*f)'(r)-~'(~)ll~~) dt = 0.
Thus f n ( t ) = J,,,*f(logt) is infinitely differentiable for 0 -= t < co with

values in B, . Sincef, +fin W the lemma is proved. I
We now investigate interpolation properties of the space T,
7.16 LEMMA Let 8 = ( l / p ) + v satisfy 0 < 8 < 1. Then

(a) every u E T satisfies
llullT = inf I l t ' f ; ~ p ( o ,c o ; ~ , ) I J ~ - ~ l l f ~ f ; ~ ( o , c o ; ~ ~ ) 1 1 (6)
~;
f EW
f(O)=u
(b) if u E B, n B,, then u E T and

ll4lT 5 ~II~11~;'II~lIBg2, (7)
where K is a constant independent of u.
PROOF (a) Fix U E T and E > 0 and let f~ W satisfy f ( 0 ) = u and

IIJ'llw I; IlUllT+E. Let
R = IltXL"(0, CO; B1)ll, S = Ilt"f';L"(O,m;B2)I1.
For A > 0 the function f A defined by f A ( t ) = f ( l t ) also belongs to W and
satisfiesfl(0)= u. Moreover,
1lrY~;Lp(O,0 0 ; B1)ll = A-'R, Ilt'(fJ;L"(O, CO; B J ) ) = A'-'S.
These two expressions are both equal to R'-'S' provided we choose A = R/S.
Hence
max(R,S) = llfllw 5 ll4lT +E
5 inf
120
IlfAllW +E
5 inf max(I-'R,ll-'S)
120
+ E s R1-'S' + E g max(R,S) + E .
Since E is arbitrary, (6) follows at once.
(b) Let 4 E Cm([O,co)) satisfy 4(0)= 1, 4 ( t )= 0 if t 2 1, I $ ( t ) l ~1,
and I$'(t)l5K1for all t 2 0. If u E B , n B,, letf(t) = $ ( t ) u so that u =f(O).
Now
Ilt'f; L"(0, B1)ll K2 IIuIIB19
where K, = {JA tVpdr}'~P.

Similarly,
Hence YE W and (7) follows from (6). I

The above lemma suggests the sense in which T lies "between" B, and B,
provided 0 .c 8 < 1. The spaces T corresponding to all such values of 8 are
sometimes described as constituting a scale of Banach spaces interpolated
between B , and B,. Many properties of Tcan be deduced from corresponding
properties of B, and B, via the following interpolation theorem.
7.17 THEOREM Suppose that B, n B, is dense in B, and B,. Let B,,

B,,and r? be three spaces having the same properties specified for B,, B,, and
X in Section 7.1 1. Let L be a linear operator defined on B1 n 8, into B, n l ,
and suppose that for every u E B , n B, we have
Thus L possesses unique continuous extensions (also denoted L) to B , and

B, satisfying (8) and (9), respectively, and hence to B, + B, satisfying
II Lull8, +
max(Ki K d llullBI +Bz.
B ~5 9
If 0 < 8 = (l/p)+ v < 1, then for every u E T = T(p, v; B,, B,) we have
Lucy= T(p,v;Bl,B",) and
/ILuIIT Kl!-'K2' II'IIT* (10)
PROOF L is defined on B, + B, and hence on T. By (6) we have for u E T

(1 LU\( = inf (1 t"J; ~ p ( 0 , c o ;B,)((-'(1 fy';LP(O,00; 8, (1'.
s EW
I(O)= Lu
Also, there exists, for any given E =- 0, an elementfe W withf(0) = u such that
< l l ~ l l T+ 6 .
IltvS~p(O,~;B,)II'-ellr'f';LP(O,co;B,)~~e
For t 2 0 let f ( r ) = Lf(t)so that f(0) = Lu and
l\t"J;~ ( 0co;B,)II'-'
, llry';LP(O,co;B,)lle
I K: -'1 t ' f ; Lp(O,co;Bl)ll -'K,' 1 t 'f'; Lp(O,a;B,)Ile.
Hence
IILullr < K:-e~,e(II~I17+E>.
Since E is arbitrary (10) follows. I
If we denote by ~ ~ L ~ ~ theLnorm
~ s , of
s )L as an element of the Banach space
L ( S , s”) of continuous linear operators on S into s” we may write (10) in the
form
7.18 LEMMA Suppose that B, n B, I s dense in B, and B,, and that there
exists a sequence {Pi}?= of linear operators belonging simultaneously to
L ( B , ) and L(B,) and having range in B, n B,. Suppose also that for each
b,eBi,i=1,2,
lim llPj bi - billB,= 0.
j-+m
Then for every u E T we have

lim IIPju-ullT = 0.
j+ca
In particular, B , n B, is dense in T.
PROOF Fix u E T and choose f~ W such that f ( 0 ) = u. Let f j ( t ) = P j f ( t ) .

If bi E B,, i = 1,2, there
exists an integer j , = j , (bi)such that i f j 2 j,, then
IIPjb,--biIJBiI1 .
Hence { P j b i }is bounded in B i , i = 1,2, independently of j . By the uniform
boundedness principle there exist constants K, and K2 such that for every j
IIpjIIL(B,) 5 Ki*
It follows that
Ilf;.(t)llB~ 5 K1 ~ ~ f ( t ) ~ \ B ~~ ~ ~ ’ ( I
t )K 2 ~ i l~f ’ (Bt > l ~l B , *
Since for almost all t > O , f j ( t ) -+f(t) in B, andfj’(t) -.f’(t) in B, a s j a, -+
we have by dominated convergence that t x . + t y f in LP(O,co;Bl)and

tx.‘+ ty‘ in Lp(O,m; 8,).Hence4 + j i n Wand so Pju = P j f ( 0 )-f(O) = u
in T. Since ty, and t’fj‘ take values in B, n B,, P j u belongs to B, n B,. I
We quote now a theorem characterizing the dual of a trace space as another

trace space. The proof is rather long and will not be given here-the interested
reader may find it in the work of Lions [37] where trace spaces somewhat more
general than those introduced above are studied. (The following theorem is
a special case of Theorem 1.1 of Ref. [37, Chap. 113.)
7.19 THEOREM Suppose that B, and B, are reflexive and also satisfy the
conditions of Lemma 7.18. If 1 < p < a and (l/p)+v = 9 satisfies 0 < 9 < 1,
then(I/p’)-v= 1-(l/p)-v= 1-8and
In particular, T @ , v; B,, B,) is reflexive.
We now prove an imbedding theorem for trace spaces between two Lp-
spaces which will play a vital role in extending certain aspects of the Sobolev
imbedding theorem to fractional order spaces (see Theorem 7.57). If R is a
domain in R”, then Bl = Lq(R) and B2 = Lp(R) are both continuously
imbedded in the topological vector space X = L:oc(R). (A subset U t L.ioC(f2)
is open if for every U E U there exists E > 0 and K c c R such that
Ilu-ullo, < e, u E L:,,(R) implies u E U.)
7.20 THEOREM Letp,q,8 satisfy 1 s p I; q < CO, 0 < 8 < 1, 8 = (I/p)+v.

Then
PROOF Suppose thatfE Cm([O,a)).From the identity
we may readily obtain
I; {( l vp If(?) I dt >”” + ([ t vp If I( t )I p dt >””} (6; - vp‘ dt >,”

= Kl(lltYfIlo,p.~o,m) + IltY ’ l l o , p , ( o , m ~ ) ~
where K, < 00 since 8 = (l/p)+v < 1. By a homogeneity argument similar to
that used in the proof of Lemma 7.16(a), we may now obtain
If(O)l 5 2K1 IltYfII~,-p!(o.m) llt f

V I B
IIo,p,(o,m)* (12)
Now suppose thatfe W ( p ,v; Lq(R), Lp(R)) and, for the moment, that f is
infinitely differentiable on (0,00) into Lq(sl). Let f ( x , t) = f ( t ) ( x ) for
0 5 t < 00, x E R. From (12) we have thatf(x,O) = lim,,o+ f ( x , t) satisfies
SEMIGROUP CHARACTERIZATION OF TRACE SPACES 191
for almost all x E $2.Thus, by Holder's inequality,
x (I(
n So" r ' ~ I ~ f ~ x , t ) ~ d t ) 8 1 1 1 1 p d x ) l i s ' ,
where (l/s)+(l/s')= 1. If we choose s so that (1 -8)rs = q, then also

8rs' = p and we have
)(1 -em
Ilt"f';LP(O, CO;LP(n))lle
= K3 Ilt'f; Lp(O,oo;L~(R))II -e )Ity-';LP(O,CO; L P ( Q ) ) ~ ~ ~ .

By the density of infinitely differentiable functions in W (Lemma 7.15) the
above inequality holds for any f E W.
It now follows that if u E T ( p ,v; Lq(R), LP(R)), then
= K3llUllT
by Lemma 7.16(a). This establishes imbedding (1 1).
7.21 REMARK With minor modifications in the proof, the above

theorem extends to cover q = 00 provided we use in place of Lq(R)the closure
of LP(R) n La@) in L"(R).
Semigroup Characterization of Trace Spaces
7.22 Let B be a Banach space and G a continuous semigroup on B which is

uniformly bounded, that is, for which there exists a constant M such that
IIG(t)llL(B)I M , 0I t -= CO.
Let A be the infinitesimal generator of G so that (see Remark 7.9) D ( A ) , the

domain of A in B, is a Banach space with respect to the graph-norm
Ilu;D(A)II = IlullB+ IIAullB,
and is also a dense vector subspace of B. The spaces B , = D ( A ) and
B, = X = B satisfy the conditions of Section 7.11, and we may accordingly
construct the trace space T = T ( p ,v;D(A), B) provided 8 = (l/p)+v < 1
Theorem 7.24 characterizes T in terms of an explicit norm involving the
semigroup G. First, however, we obtain an inequality of Hardy, Littlewood,
and P6lya [28] which will be needed.
7.23 LEMMA Letfbe a scalar-valued function defined a.e. on [O, co) and
let
= (l/t) /k)
dt.
0
If 1 Ip c co and (l/p)+v = 8 c 1, then
PROOF We may certainly suppose that the right side of (13) is finite. Under
the transformation t = e',f(er)= f ( ~ )(, = ee, g(e') = #(T), (13) becomes
Note that
J._mm eePrIg(.r)lpd~I[1/(1 - 8 ) q
1-: eeprIf(z)lpd.r. (14)
g(7) = e-' f(a)e'do.
Let E(T)= eer and

e(e-l)r if T >o
F(z) =
if 7 I0.
Then E-g"= F* ( E a f ) , and so by Young's theorem 4.30,
IIE.9"110, p , w 5 IIFllo, 1, w llJ?fllo, p , R.
This inequality is precisely (14) since fEWIF(7)l dr = 1/(1-8).
7.24 THEOREM Let A be the infinitesimal generator of a uniformly

bounded, continuous semigroup G on the Banach space B. If 1 Ip < co and
0 c (l/p)+v < 1, then T = T ( p , v ; D ( A ) ,B) coincides with the space T o of
all u E B for which the norm
is finite. The norms 11 and 1 To are equivalent.
PROOF First let u E T and choose f E W such thatf(0) = u. For the moment
let us assume that f is infinitely differentiable on (0,co) into D ( A ) , Let
f ’ ( t ) - A f ( t ) = h(t). If t 2 E > 0, we obtain by Theorem 7.10,
f ( t ) = G(t-E)f(E) +StG(t-7)h(7) dz.

e
Hence
G(t-E)f(E) - f ( ~ )=
Letting E -,O + in this identity and noting that f ( ~+f(O)

) by definition, we
obtain
G(t)f(O) -f(O) = (16)
Now (16) holds for anyfE W since, by Lemma 7.15, f is the limit in W of a
sequence {f,} of infinitely differentiable functions on (0,co) with values in
D(A). Hence for u E T and anyfE W withf(0) = u, we have
G(t-z)h(t) d7, h(z) = f ‘ ( z ) - Af(z).

Thus
where we have used the uniform boundedness of G. Applying Lemma 7.23,

we obtain, putting 0 = (l/p)+v,
194 VJI FRACTIONAL ORDER SPACES
Since this holds for anyfE W withf(0) = u we have
Also, the identity operator on B provides imbeddings of B into B (trivially)

and D(A)into B, each imbedding having unit imbedding constant. By Theorem
7.17 we have as well, T+ B and llullSI IIuIIT. Hence u E T implies u E T o and
Conversely, suppose that u E T o . Let 4 E Ca([O, a))satisfy 4(0) = 1,

4(t) = 0 for t 2 1, I ~ ( z ) Is 1 and I4'(t)lIK , for t 2 0. Let
f(0 = 4(t)s(t),
where
g ( t ) = (l/t)
so' G(~)udr, t > 0.
In order to show that u E Tand lJuIITIK2 IIu)ITo,it is sufficient to prove that

(17)
f E W and
llfllW 5 K211ullTo (18)
[sincef(O) = lim,+,+ 4 ( t ) g ( t ) = u by Lemma 7.8(a)] and this can be done by
showing that t'g E Lp(O,1; D(A)) and r'g' E Lp(O,1; B) with appropriate
norms bounded by K3 llull
By Lemma 7.8(b), j k G(r)u d7 E D(A) and
G(7)udr = G(r)u - U.
Thus
g'(t) = (l/t)G(t)u - (l/t2)

so'G(T)u d~
= ( l / t ) ( G ( t ) u - u ) - (l/t2)
6 (G(T)u-u)&,
and since
and, by Lemma 7.23 with v - 1 replacing v ,
I Cl/(2-e)p1 IIuIIpTo
we therefore have
Thus g E W, (18) holds, and the proof is complete. I

7.25 For our purposes we require a slight generalization of Theorem 7.24.
Let A * , Az, ..., An be a finite family of infinitesimal generators of commuting,
uniformly bounded, continuous semigroups G , , Gz, ..., G, on B. Thus
Ilcj(t)llL(B) Mj; l l j l n , tro;
G j ( s ) G k ( t )= Gk(t)Gj(S); 1 Ij , k I n, s,? 2 0.
Let B" denote the product space B x B x ... x B (n factors) which is a Banach
space with norm
n
II(bl,b2, ..*,bn)llBn = IlbjllB*
j=1
The operator A is defined on D(A) = D(Aj) into B" by

AU = (Al u,AZu,..., A,,u).
We leave it to the reader to generalize Lemma 7.8 to show that D ( A ) is dense
in B, that A is a closed operator, and hence that D ( A ) is a Banach space with
respect to the norm
c
n
Ilu;D(A)II = ~ ~ u ~ ~ B + ~ ~=A IIuIIB+
u ~ ~ B n
j=1
IIAjullB*
7.26 THEOREM If 0 < (l/p)+ v < 1, 1 I p c oc), then T = T ( p ,v; D(A),B)

coincides with the space T o of all u E B for which the norm
is finite. The norms 1) and 11 To are equivalent.

1% VII FRACTIONAL ORDER SPACES
PROOF The proof is nearly identical to that of Theorem 7.24 except that in
place of the function g ( t ) given by (17) we use
( ) . . . G,(T,,)Udz,dz,
g ( t ) = ( l / t n ) S ' S ' . . . S ' G l ( z ~ ) G ~z ~ dz,,.
0 0 0
The details are left to the reader. I

7.27 EXAMPLE Let B = LP(Rn),1 I p < 00, and for u E B set
(Gj(t)u)(x) = u(xI,..., x j + t , ...,x,); j = 1,2,...,n.
Clearly the Gj are commuting, uniformly bounded (Mj = I), continuous
semigroups on Lp(Rn).(In fact they are groups if we allow t < 0.) The corre-
sponding infinitesimal generators satisfy
D(Aj) = {U E Lp(R") : Dju E Lp(R")},
A~=
u Dju, u E D(Aj).
Accordingly, D ( A ) = n;=,D ( A j )= W1*p(R").By Theorem 7.26 the norm
(Ilullg, p. R" + i S't(v-1).
j=l 0
J-"Iu(x1, ..., + t , ...,x,,) - ...,x,,)IPdx dt
xj 4x1,
is equivalent to llullT on the space T = T ( p ,v; W1.p(Rn), LP(R")),

0 < (l/p)+v < 1.
Higher-Order Traces
7.28 Up to this point we have considered only tracesf(0) of functions satis-

fying, with their first derivativesf', suitable integrability conditions on [0,03)
into various Banach spaces. We now extend the notion of trace to obtain
values for f ( j ) ( O ) , 0 5,j I m - 1, provided f,f ', . ..,f(")
satisfy such inte-
grability conditions. As a result of this extension we will be able later to
characterize the traces on the boundary of a regular domain Q, of functions
belonging to W"*p(Q).
7.29 Let B be a Banach space and A,, ...,A,, infinitesimal generators of

commuting, uniformly bounded, continuous semigroups GI,..., G,, on B. For
each multi-index a we define a subspace D(A")of B and a corresponding linear
operator A" on D(A")into B by induction on la1 as follows.
If a = (O,O, ..., 0), then D(A") = B and A" = I, the identity on B.
If a = (0,..., 1, ..., 0) (1 in thejth place), then D(A") = D(Aj)and A" = A j .
If D(AB)and ABhave been defined for all with IpI I r, and if la1 = r 1 , +
HIGHER-ORDER TRACES 197
then
D(A") = { u : u E D(AB)and A a - hE D(A8)for all < u},
I\"= A;i ...A?.
If k is a positive integer, let D(Ak)= n l a l s k D ( A a Once
). again we leave to
the reader the task of verifying (say by induction on k) that D(Ak)is dense in
B, that Ak = is a closed operator on D(Ak) into nIals
k B , and
hence that D ( A k )is a Banach space with respect to the norm
7.30 For positive integers m and real p , 1 s p I 00, let W" = W m ( pv, ; A; B )
denote the space of (equivalence classes of) measurable functionsf on [0, 00)
into B such that
t v f ( k )E Lp(O,~ o ; D ( A " - ~ ) ) 0 I k < m,
f ( k )being the distributional derivative dy/dtk.The space W" is a Banach space
with respect to the norm
IlfIIWm = max \ J C ' ~ ( ~ ) ;CO;D(A"-~))\\.
L~(O,
Osksm
Note that W' = W ( p ,v ; D(A),B) with D ( A ) as in Section 7.25.
7.31 LEMMA Let f E W", m 2 1. If 0 Ik Im- 1 and u is a multi-index

such that lul+ k < m- 1, then the functionf, = A\"f(k)E W' and
Ilfak)lW1 5 IlfIIWm.
whence the lemma follows.

-=
7.32 Let us assume hereafter that 0 < (l/p)+v 1 and denote by To the
trace space T(p,v ; D(A),B) corresponding to W ' . By Theorem 7.26 we may
take the norm in Toto be
llUllT0 = (Ilulli + IIullP,)l'p,
where
Higher-order trace spaces may now be defined as follows. For k = 0, 1,2, .. .

we define Tk= Tk(p, v ; A; B) to be the space consisting of those elements
u E D(Ak) for which A"u E To whenever la1 5 k. The space Tkis a Banach
space with respect to the norm
It follows from Lemmas 7.31 and 7.13 that i f f € W" and (a15 m - k - 1,
then Aaf(k)(0) exists in T o and
5 K& I I f l I W " ,
Il~Yk'(o)llTO
where Kakis a constant depending on a and k . Hencef(k)(0)E Tm-k-l

and
5 KkIIfIIWm.
It follows that the linear mapping
f -b (f(O),f'(O), ...,f'"-l YO)) (19)
is continuous on W minto Tm-'x Tm-2x ... x To= np;: ; that is,
Tm-k-l
We shall prove that this mapping is onto (see Lions [38]).
nF:J
n?::
7.33 THEOREM The range of the mapping (19) is Tm-k-l. If
(uo,u l , ..., u",-~)E T m - k - l ,there exists YE W" such that f k ) ( 0 =
) uk,
0 5 k 5 m - 1, and
PROOF The proof is similar to the second part of that of Theorem 7.24, but
is rather more complicated. To achieve some simplification we shall deal with
the special case n = I so that Aj becomes just A and A" becomes Ak (la1 = k ) .
Suppose we have constructed for each k, 0 I k I m- 1, a function

fk E wmsuch that f$"(o) = uk and
llfkllWym Kk\\Uk\\Tm-"-I*
Let L r , k , 0 s r I m- 1, satisfy the nonsingular system of equations
c riArBk
m-1
r=O
=
1
0
if j = k
if O I j s m - 1 , j#k.
Then the function
satisfies
gik)(0)= uk, g$j'(O) = 0, O Ij I m - 1, j # k.
Moreover, it is easily checked that
IIgkIIW"' 5 RkII&llWm.
Hence the functionf(t) = g k ( t ) has the properties required in the state-

ment of the theorem. Thus we need only construct&.
In the rest of the proof we shall make extensive use of the convolution
product of operator-valued functions on [O, co). If for t 2 0, Fl(t) and F 2 ( t )
belong to L(B), we define Fl * F, on [0, co) into L(B) by
Fl * F,(t)b =
l F1(t-7)F2(z)bdz, b E B.
(All the operators we use will commute.) If Fl is continuously differentiable

on [O,oo) into L ( B ) so that F,'(t) E L(B), we have evidently
(Pi * F2)'(t) = F,' * F,( t ) .
We denote by F(("'))the convolution product F* F* ... * F having m factors;
this is well defined since * is associative for mutually commuting factors. If
I ( t ) = I denotes the identity on B, we have clearly
I""))(t) = [ t m - */(m- I)!] I.
If G is the continuous semigroup whose infinitesimal generator is A, then
by Lemma 7.8 we ,have
A(I*G) = G - I
or, when both sides are restricted to elements of D(A),
I * AG = G - I.
Given u = u k E T m - k - lwe define

h(t)= C(m+k)!/k!l4 ( t ) g ( t )
where 4 E C([O,00)) satisfies 4 ( t )= 1 if t I 3, 4 ( t )= 0 if t 2 1, and
I#j)(t)lI; Kl, 0 I;j I;m, and where
g ( t ) = t -mZ((k+ l)) * G(("))(t)u.(20)
[Note that this is the same function g as defined in (1 7 )in the proof of Theorem
7.24 if m = 1 and k = 0.1 We must verify that
fk'k'(0) = u (21)
and
llfkllwm 5 K? IIUIITm-k-1. (22)
In view of the constancy of 4 for t I 3, (21) will follow if we can show that
+
g(k)(0)= [ k ! / ( m k ) ! ]u.
However,
*
g ( t ) = t-mZ((k+l))(G-I+Z)""))(t)u
since - m l ( ( k + 1+ m - i ) ) ( t ) = ( t k - j / ( k + m - j ) ! ) Z has vanishing kth derivative

if j > 0 we have
kf
I * (G-Z)"o"(0)~= -
(k m)! +
In order to establish (22) it is clearly sufficient to show that
PI
holds for every i , j such that 0 I j 4 m and 0 5 i Im - j . We distinguish three

cases.
1 wr(t),
i
Ak”’(t) =
r=O
IK5tm-i-j-2 [ l I G ( z ) w - dz.
w(JB
Since i Im - j we therefore obtain for 0 < t I1
llA~(j)(t)llBI &t-’
sb IIG(z)w-~llt,d~.
It follows by Lemma 7.23 (with v- 1 in place of v) that

202 VJI FRACTIONAL ORDER SPACES
CASE 3 Suppose k + l 4 j 4 m and 1 Ii 5 m-j. Then i s m - k - 1 and

, i E A i U E T m - k - l - i with
5
~ ~ U " ~ ~ T r n - k - 1 - IIUIITm-k-1.
i
Let h ( t ) = Aig(?),Thus
h(r) = t - m I ( ( k + 1)) * G((m))(t)G.
In order to prove (23) in this case it is sufficient to show that
For purposes of proving (25) we may omit the term

t - ' " P k + l +"))(t)ii = [tk/(k+m)!]ii
since the j t h derivative of this term vanishes for t > 0. Accordingly we consider
We repeat m - k - i - 2 more times the preceding argument used in

deriving (26) from (24), each time discarding terms which are polynomials of
degree sj- 1 and so contribute nothing to fz(j). This leads to
h(t) C
m-1
t-m~((k+2+(m-k-f-2)+l)) * G((m-I)) (0Am - k - 1 - i u- .
I=O
k t =Am-k- 1- i
u" = Am-k-lu. The terms of the above sum are of the
form
w,(t) = t-mz((m+l-l)) * G((rn-1))( t ) w9
where 0 5 I Im - 1. Note that m + I - i 2 j . In order to prove (25) it is suffi-

cient to show that
At this point we must distinguish two subcases, i s m - j - 1 and

i = m-j.
If i Irn-j- 1, then w { j ) is a linear combination of terms of the form
-m-r I( ( m + I - i - j + r ) ) * G(W-O)(t),,,
whose norms in B are bounded by
K -m-r+(m+l-i-j+r-l)+m-I
12 IIWllr, s K12tm-j-*-i llWllT0
and (27) follows at once.

If i = m-j, we have
wI(t)= t - m p + O ) * GKm-1-1)) * ( G - Z + I ) ( t ) w
- t - m p + O ) , G((m-1-1)) * ( G - I ) ( ~ ) ~
-
+t -mj((j+l+W ,, G((m-1- 1 ) )(0w*
We repeat this procedure on the last term m- I- 1 more times to obtain
c t - mz ( ( j + I + s ) ) * * (G-Z)(t) w + t - ' " Z ( ( j + " ) ) ( t ) ~ .

m-1-1
Q(m-I-s-1))
Wl(t) =
s=o
Again we may discard the last term which makes no contribution to (27). It is
therefore sufficient to establish (27) with w Ireplaced by
w,*(t) = t - m I ( ( j + l + S ) ) * G ( ( m - i - S - 1 ) ) * (G- I )( t )W.
However, w { i ) ( t )is a linear combination of terms of the form
-m-rI((I+s+r)) * G((m-I-s- 1)) * (G-Z)(t)w
for 0 Ir Ij . It follows just as in Case 1 that for 0 < t I1,
Ilw{$(t)IIB IKi3tV2
sb IIG(T)W-WIIB~~
and hence, using Lemma 7.23 again, that (27) is satisfied. This completes the
proof.
We remark that the proof for general n is essentially similar to that given
for n = 1 above. In place of (20) one uses (a suitable multiple of)
g(t) = t-mnl((k+l)) * G((m))
1 * ... * GiO"))(t)u.
7.34 EXAMPLE Let B = Lp(Rn)and G,, 1 < j < n, be as in Example

7.27 so that Aj = Di. Evidently
D(Ak) = LP(Rn)
{U E : D"u E LP(Rn), IQI I k } = Wk~P(Rn).
For each u E Lp(rWn,'l) define ii a.e. on [0, m) into LP(Rn)by
ii(t)( X i , ..., X") = u(x,, ...,x,, i).
Then u E Wmpp(Rnfl) provided ii(k)E Lp(O,GO; Wm-kip(Rn)),
0 I k I m.
Accordingly,
Wmvp(Rn,'l)r W m ( p0;
, A; LP(Rn))
with A = (D1,...,D,), If 1 < p < 00, the mapping y
7: u + (u(., .., .,0),D,+ 1U( a, ..., -,0), ...,D~,;'u(*, ..., a, 0))
is an isomorphism and a homeomorphism of WmiP(R~+l)/kery
onto the
product n;:; Tm-k-lwhere
Tk= Tk(p,O;A,LP(R"))
= {V E : D'v E TO, lor1 < k }
Wk'P(Rn)
and
+ l #Cl = k j=1 ~ t - p ~ R n I D ',...,

u ( x~ jl+ t ,..., x,>
- D'u(x1, ..., x,)lpdxdt
The Spaces Ws*p(R)
7.35 We now define spaces Ws*P(Q) for arbitrary domains R in R", arbitrary
values of s, and 1 < p < m. These spaces coincide for integer values of s with
the spaces Wm.p(Q)and W-mn,p(R)defined in Chapter 111. For s T 0 the
definitions can be extended to p = 1 and p = 00, but for the time being we
ignore these limiting values.
The spaces B, = W ' . p ( Q ) and B, = X = LP(R) clearly satisfy the con-
ditions laid down in Section 7.11. For 0 < 8 < 1 let
TOnP(Q) = T ( p ,v; W'-P(R), LP(Q)),
where v+(l/p) = 0. Denoting W = W ( p , v ;W'*p(R),LP(R)), we write the
THE SPACES ws"(Q) 205
norm of u in P,p(Q) as
7.36 Let s 2 0 be arbitrary. If s = m, an integer, we define W s * p ( Q=)

WmpP +
(Q). If s is not an integer, we write s = m CT where m is an integer and
0 < c < 1. The space Ws+p(Q)is, in this case, defined to consist of those
p whose distributional derivatives
(equivalence classes of) functions u E W m"(0)
D'u, IuJ= m, belong to T 1 - u , p ( Q )Then
. Wsip((n)is a Banach space with
respect to the norm
7.37 The operator P given by

Pu = (u, (D'u)lal = m)
(the multi-indices u with = rn being ordered in some conveient way) is an

isometric isomorphism of W s * P ( Qonto
) a closed subspace of the (product)
Banach space
S = Wm*p(Q)
x n
(aJ=m
T1-u3p(Cl)
having norm
Il(u,(VPr)lal=m);SII = { II~II;,~,~+
1 la1 = rn I I ~ ~ ~ ~ ~ - ~ ~ ~ ( Q ) I I ~ } " ~
Since Wm*p((n)
and T1-u3P(Q)are reflexive (Theorems 3.5 and 7.19) it follows
by Theorems 1.21 and 1.22 that W s S p ( Qis) reflexive.
7.38 THEOREM For any s 2 0, Com(Rn)is dense in Ws~P(R").
PROOFThis result has already been proved for s = 0, 1,2, ... (Theorems 2.19
and 3.18) and in particular WlVp([W") is a dense subset of Lp(R") [i.e., dense
with respect to the topology of LP(R")].If s = m + u > O , where m is an
integer and 0 < u < 1 , the theorem may be proved as follows.
Let t,b E Cm(R) satisfy $ ( t ) = 1 if t I0 and $ ( t ) = 0 if t 2 1 . For
j = 1,2,3, .. . let $ j E Com(R")be defined by
$j(x) = q(1xl-j).
Let J, be the mollifier introduced in Section 2.17. If u is a function defined

(a.e.) on R", set
Pju = J i , j * ( $ j * U ) , t~= 1,2, ....
Evidently Pj is a bounded linear operator on Wm*p(Rn) into Wm*p(R") for
rn = 0, 1,2, ..., and has range in COm(R")in each case. We can deduce from
Lemmas 2.18 and 3.15 that if u E Wm*p(Rn), then
lim I I P I ~ - ~ l J m , p=, R0.n
j+m
It follows by Lemma 7.18 that if 0 < 8 < 1 and u E TesP(Rn),
then
lim IIPju-u; P*P(R")II = 0.
1-m
Since
D"Pju = J , l j * D " ( $ i j . u ) = P j D % + J , l j * ~ j ,
where
and since
it follows that for any u E Ws*p(Rn)

provided u E W1al*P(Rn), we have, taking
la1= m,
lim IIDaPju-D"u; T1-a*P(Rn)l\= 0.
j+m
Hence
and the proof is complete. I

7.39 Let W $ p ( n ) denote the closure of Com(n)in the space W S i p ( Q(s) 2 0).
By the above theorem W$p(R") = Ws*p(R"). For s < 0 we define
W S q 2 )= [W;J*P'(n)]', (l/p) + (l/p') = 1.
It follows by reflexivity that for every real s
r W--"P'(W").
[WJ*"(R")]'
We shall not comment further on the structure of the spaces Ws*p(ln)for
s < 0 except to note that, being duals of spaces having 9@)= C O W )as
dense subsets, they are spaces of distributions on n.
THE SPACES ws''(a) 207
Many properties of the spaces W s ~ p ( R are

) conveniently proven only for
R = R", and must then be deduced for more general R with the aid of an
extension operator extending functions defined on R to R" with preservation
of differential properties (see Section 4.24). For fractional s = m+a suitable
extensions are obtained by interpolation. Thus the existence of a strong
(m+1)-extension operator for R will normally be required. Such is, for
example, assured if R satisfies the hypotheses of Theorem 4.26 (see also
Section 4.29).
7.40 THEOREM If s = m+ u where m is an integer and 0 < t~ < 1, and

if there exists a strong (m+ 1)-extension operator E for R, a domain in R",
then the set of restrictions to R of functions in Com(Rn)
is dense in Ws.p(R).
PROOF (Recall that the conclusion holds for Wm*p(R)under the assumption
only that R has the segment property.) The proof follows the same lines as
that of Theorem 7.38 except that in place of the operator Piwe use the operator
pi defined by
piu = R, PiEu, u defined on R,
where R, is the operator restricting to R functions defined on R". The details
are left to the reader.
The following localization theorem requires, in addition to the existence

of a strong (m+ 1)-extension operator E for Q, a representation for the
derivatives DaEu(x)in terms of the derivatives of u such as is provided by
Theorem 4.26. Thus the hypotheses of the theorem below will certainly be
satisfied by any domain satisfying the hypotheses of Theorem 4.26.
7.41 THEOREM Let R be a domain in R" for which there exists a strong
(m+ I)-extension operator E and, for IyI 5 la[ = m, linear operators E,,
continuous from W'*p(R) into W'*P(R") and from Lp(R) into LP(R")such
that if u E Wm,p(R),then
D"Eu(x) = 1 EayDyu(x) a.e. in R"
lylsm
If s = m+a > 0,O I a < I , then W s * p ( Rcoincides
) with the set of restrictions
to R of funcitons in Ws*p(R").
PROOF If a = 0, the result is an immediate consequence of the existence of

a strong m-extension operator for R. Suppose 0 < a < 1. If u E Ws.P(R), then
u E Wmn,p(R)and Eu E Wm*p(R,); also
IIEullm.p,Rn 5 KIII~llrn.p,~ K1II~lls.p.n. (31)
If 171< m, then Dyu E T'-"*p(R) and

IIDYu;T1-u*p(Q)ll5 K, IIUlla,p,n. (32)
(This holds by definition if IyI = m and via Lemma 7.16 if 171< m.)Since Eay
is linear and continuous on W'.p(R) into W'*P(Rn)and on Lp(R)into LP(Rn),
by Theorem 7.17 it is also continuous on T'-"*P(R) into T1-uip(Rn). By (30)
and (32) we have for la/ = m
IID"Eu; T1-YRn)Il 4 K3 IIUIIs,p.n.
Combining this with (31), we obtain
IIE~lla,p,w5 K4 IIUll.,p.n
and u is the restriction to R of Eu E Ws*p(R").
Conversely, the operator R, of restriction to 0, being continuous from
Wm*p(Rn) into W " * p ( 0for
) any m,is also continuous by Theorem 7.17 from
Ws*p(Rn)into Ws*p(R) so the restriction R,u of U E WsvP(R") belongs to
WS*"R). I
We remark that under the conditions of the theorem the extension

operator E is continuous from Ws*P(R)into WS*p(Rn)for any s, 0 < s 5 m + 1.
An Intrinsic Norm for Ws*p(R)
7.42 We now investigate the possibility of constructing a new norm for

Ws*p(R),s 2 0, which is equivalent to the "trace norm" (29) (s not an integer)
but which is expressed "intrinsically" in terms of properties of the element
involved. In view of Example 7.27 it is most convenient to begin with the case
a = R". Following Lions and Magenes [34] we define new spaces pSp(fl)
with intrinsically defined norm and then show, at least for suitably regular
domains R, p i p ( R )coincides with W s * p ( f l ) .
7.43 For 0 < 8 < 1 and 1 5 p < w let Te*p(R)denote the space of (equiv-
alence classes of) functions u E Lp(R) for which the norm
is finite, where v+(l/p) = 8.
7.44 LEMMA The space Te.P(Rn)coincides with the Banach space

the norms in the two spaces being equivalent.
Te*P(Rn),
AN INTRINSIC NORM FOR ws'p(n) 209
PROOF The norm of an element u in Te*P(Rn) was defined [in (28)] to be its
LP(R")).By Example 7.27 we may
norm in the trace space T ( p , v ; W'pP(Rn),
take the norm to be
+
j= 1 )"'.
f i ( v - l ) p ~ n I u ( x l,...,xj+t, ..., x n ) - u ( x l , ...,xn)lpdxdt
where
since A > 0 and n- 1 -2A c 0. Setting

y , = zi, 1 I i I j , xj = t + y j , xi = zi, .i+ 1 5 i S n
in ( 3 9 , we obtain, using (34),
) llullt I K4 IIuIIT.
Thus U E ~ p p ( R "and
Conversely, suppose u E TB*p(Oarn). Let x' = (xz, ...,x,) and z' = (zz, ..., z,)
and integrate the inequality
lu(x,+t,x') - u(x,x')IP I; K,(lu(x,+t,x')- u(x,+3t,z')lP
+lu(xl+*t,z') - u(x,x')IP)
with respect to z' over the disk D ( t , x ' ) centered at x' E R"-' and having
radius +t, thus obtaining
Iu (XI + 1, x') - u (Xl ,x')IP I (K6/tn- l ) [I, ( 2 , x) +I, (0,x)],
where
for s = t or s = 0. Now
- dt
rt(' knt"-'1
I, (t, x) dx
where we have put z1 = x1-32, dz = -+

dt and used the fact that in the inner
integralinthesecondlast line Ixl-zllrIx'-z'/ so that Ix,-z,Irlx-zl/fl.
AN INTRINSIC NORM FOR ws"(n) 211
A similar inequality holds for I,(0,x). Thus
Similar inequalities hold for differences in the other variables x2, ..., x,, and
combining these we obtain llullT I K8 IIullr. I
In order to extend the above lemma to more general domains R we require

the following extension lemma.
7.45 LEMMA Let R be a half-space in R" or a domain in R" having the

uniform 1-smooth regularity property and a bounded boundary. Then there
exists a linear operator E mapping Lp(R) into Lp(R")such that if u E LP(R),
then
Eu(x) = u ( x ) a.e. in R,
and if 0 < 0 < 1 and u E TeiP(R),then Eu E p p p ( R nand
)
IIEu; Te*p(R")II< Kllu; Te3p(R)ll
with K independent of u.
PROOF The proof is quite similar to that of Theorem 4.26. We begin with the
c a s e R = R + " = { x ~ [ W " : x , , > 0 }Letusdenotebyx'=(x,,
. ...,x,-,)andfor
u E Lp(i2) set
Eu(x)=
u(x', - x n )
a.e. in R+"
-
a.e. in R" R+".
Then
= 2 lI~IIPo,p,R~+*
+
Also, setting 21 = n- 1 (1 - v)p = n + (1 - O)p > 0, we have
where
[since (x,,+y,)' 2 (xn-y,JZ when x, 2 0 and yn 2 01,and similar inequalities

hold for 1- + and I - - . Thus
IIEu;T@*'(R")Il5 41'PIIu;P*P(R+")l).
Now suppose that R is uniformly C'-regular and has a bounded boundary.
Then the open cover { Uj} of bdryR and the corresponding collection {Oj}of
1-smooth maps of Uj onto B = {y E R" :lyl< I} referred to in Section 4.6
are both finite collections, say 1 5j 5 N. We may also assume that the sets
Uj are bounded. Let Uo be an open subset of R, bounded away from bdryn,
such that R c uy=o U j . Let { ~ ~ } j be
m , a~ C"-partition of unity for R sub-
ordinate to { U j } . Given u E Lp(R) let uj be defined a.e. in R by u j ( x ) =
o j ( x ) u ( x ) . Clearly uj E Lp(R) and llujllo,p,ns ( I u I I ~ , ~ , ~If. u E TeSP(R),then
for 1 S j s N
But since Uj is bounded we have for y ER n Uj by Lemma 5.47,
and K 3 may be chosen independent of the finitely many values o f j involved.

Thus u, E T@*'(R)and
5 K4 IIu; p*p(Q)II*
Tfe*p(R)II
Iluj;
Since w o ( x ) = 1 for all x E R lying outside the bounded set u;=,Uj,the

above inequality also holds for uo.
AN INTRINSIC NORM FOR wS'"(a) 213
For 1 2 j 5 N let u, be defined on R+" by

if y E B n R , "
if Y E R+" B, -
where Y j = @;I. Then u, E TeVp(R+").In fact, putting y = aj(x),q = Qj(5),
we have
if x E Uj
if X E R" N U,.
Then clearly wj ( x ) = u j ( x ) a.e. in R, supp wj c c U j , and by a calculation
similar to the one carried on in (36),
Ilwj; T'B,p(Rn)II 5 K* IIEuj;T'B*p(Rn)I).
Finally, we set
N
E * U ( X ) = u,-,(x) + j= 1
wj(x).
It is clear that E* has all the properties required of u in the statement of the
lemma. I
It should be remarked that the comments made in Section 4.29 concerning

weakening the hypotheses for the extension theorems 4.26 and 4.28 apply as
well to the above lemma.
7.46 COROLLARY Under the conditions of Lemma 7.45 the spaces

Te*p(R)and TesP(R) coincide, and their norms are equivalent.
PROOF The coincidence of the two vector spaces follows from the fact that
they coincide with restrictions to R of functions in the coincident spaces
Te*p(R")and TeSP(R"). If u E TepP(R)and E is the extension operator con-
structed in the above lemma, we have
IIu; Tesp(R)Z)I)
5 IIEu; TeiP(R")II5 K,IIEu; Tegp(R")(IIK2Ilu; Terp(f2)II.
The reverse inequality follows in the same way, using instead of E the strong
1-extension operator constructed in Theorem 4.6 (the case m = 1) which, as is
implicit in the proof of Theorem 7.41, is an extension operator for
TeSp(f2).I
7.47 For s 2 0 let p i P ( R ) be the space constructed in exactly the same way
that WSqp(R)was constructed in Section 7.36 except using the spaces
cfl-u*p(R)in place of T'-"vP(R).In view of Corollary 7.46 we have proved
the following theorem.
7.48 THEOREM Let R be R", or a half-space in R", or a domain in R"

which is uniformly C'-regular and has a bounded boundary. Then the spaces
@"(R) and Wssp(R) coincide algebraically and topologically for each s 2 0.
+
In particular, if s = m Q where m is an integer and 0 < Q < 1, then the norm
II * I L p , n given by
is equivalent to the original norm 11 - Ils,p,n on Ws*p(R).
7.49 REMARK It is the spaces which we have above denoted p * p ( R )

which one encounters most frequently in the literature, and which are usually
designated Ws*p(R).The space WSm(SZ) may obviously be defined in an
analagous way. It consists of those u E W"3m(R)for which the norm
Imbedding Theorems
7.50 As we have already seen (in Example 7.34), if 1 < p < co, the linear
mapping
u + yu = (you,...,ym-lu); yju = Dn.IU(.,..., *,O)
IMBEDDING THEOREMS 215
establishes an isomorphism and a homeomorphism of W"sP(R+")/ker y onto

n?: T"-k-'(p,O;A;Lp(R"-')),
W k * p ( R " - land
where A = ( D l ,...,D,,-'). Since D(Ak)=
) since To(p,O;A ; L p ( R n - ' ) )= T1/p,p(Rn-l),we have
T k ( p ,0;A; LP(R"-')) = W k + l - l l p , p ( [ W n - 1 ) .
Thus y is in fact an isomorphism and homeomorphism of W".p(R+")/kery

onto nrZ: Wmn-k-1'p7p(Rn-1). In particular, the traces on R"-' = bdry R,"
of functions in Wmn.p(R+n) belong to, and constitute the whole of the space
W"- ''P*P(bdryR+n). [This phenomenon is sometimes described as the loss
of (l/p)th of a derivative on the boundary.] The result can be extended to
smoothly bounded domains.
7.51 If 0 is a domain in R" having the uniform C"-regularity property and

a bounded boundary, then the open cover { U j } of bdryR and the associated
collection {Vj} of rn-smooth maps from B = {y E R" : lyl< l } onto the sets
U j (referred t o in Section 4.6) are finite collections, say 1 I j I r. If {aj}is a
partition of unity for bdryn subordinate to { V j }and if u is a function defined
on bdryn, we define eju on R"" by
where y' = (yl, ...,y,- l).

For s 2 0 and 1 < p < oc) we define W"*P(bdryf2)to be the class of
functions u E LP(bdryR) (see Section 5.21) such that Bju belongs to W"*j"R"-')
for 1 I j 5 r. The space Ws*p(bdry51)is a Banach space with respect to the
norm
As defined above, the space WSqP(bdryR)appears to depend on the

particular cover { Uj}, the mappings {Y,}, and the partition of unity {aj}used
in the definition. It can be checked that the same space, with an equivalent
norm, results if we carry out the construction for a different collection { o,},
{ukj}, and {Gj}.(We omit the details; see Lions and Magenes [MI.)It can also
be checked that C"(bdryn) is dense in W"*P(bdryQ).
7.52 Let u E COm(R"). [The restrictions of such functions to R are dense in

Wm3P(Q).] Let y denote the linear mapping
u -b yu = (you,...)y m - l u ) ; yju = './

ani bdryR
, (37)
where ai/anJ denotes the j t h directional derivative in the direction of the in-
ward normal to bdryf2. Using a partition of unity for a neighborhood of
bdryn subordinate to the open cover {ui}, we can prove the following
generalization (to f2) of the result of Section 7.50. (The details are left to the
reader.)
7.53 THEOREM Let 1 < p < 00 and let f2 satisfy the conditions
prescribed above. Then the mapping y given by (37) extends by continuity
to an isomorphism and homeomorphism of W",P(f2)/kery onto
n
m- 1
k=O
Wm k-
"- l/p*p (bdry f2).
7.54 It is an immediate consequence of the following theorem that the kernel

kery of the mapping y, that is, the class of u E Wm*P(f2)
for which yu = 0, is
precisely the space Wr*p(f2). We adopt again the notations of Section 7.30.
Let Worndenote the closure in W" = W"(p,v;A;B) of the set of functions
f E W", each of which vanishes identically on an interval [0, E ) for some
positive E (which may depend on f ).
7.55 THEOREM If YE W" and if f ( k ) ( 0=) 0 for 0 Ik Im- 1, then

f E Worn.Thus Wornis the kernel of the mapping
f (f(O),f'(O), ...,f(m-l)(o))
n;:;
+
of W" onto Tm-k-',
PROOF Let YE W" satisfy f ( 0 ) = ... =f("-')(o) = 0. Let $ E Cm(R) satisfy

$ ( t ) = 0 for t 51, $ ( t ) = 1 for t 2 2, 0 I$ ( t ) I 1 and l$(k)(t)lIK, for all
t, 0 Ik Im. Let f , ( t ) = $ ( n t ) f ( t ) . Clearly f,E Worn.We must show that
f ( t ) - f , ( t ) = (1 - $ ( n t ) ) f ( t ) + 0 in W" as n -+ 00, that is, we must show
that for each k, 0 Ik Im, and each multi-index u, lul Im-k, we have
Now
as n -,00 sincef E W". Hence we need only show that if 1 s j s k, then

IMBEDDING THEOREMS 217
as n + 03. But the left side of (38) does not exceed a constant times
(39)
--
Since f ( 0 ) =f'(O) = . =f("-')(O) = 0, and since k - j 5 rn- 1 we have
[wherep-'+@'I-' = 11
5 KzfjP-vP-' ltvp IIA"f(k)(z)IIgP

dt.
It follows that (39) does not exceed a constant times
as n + 03 since f E W". I
7.56 The characterization of traces on bdryR of functions in Wm>p(R)has
important applications in the study of nonhomogeneous boundary value
problems for differential operators defined on R. Theorem 7.53 contains
both "direct" and "converse" imbedding theorems for W"'*p(R) in the
following sense: If u E W",P(R), then the trace u = UlbdryR belongs to
W"-"P*P(bdryR) and
IIullm-l/p,p,bdryIl I KIIIullm,p,R;
and conversely, if u E W"-'/P*P(bdryR), then there exists u E W"*p(R) with
= I bdry R and
IIUllm,p,R 5 K2 IIUllm-l/p,p,bdryR.
Before stating a very general imbedding theorem for the spaces Ws*p(R)
we show how some (but not all) imbeddings for these spaces can be obtained
from known cases for integral s by the interpolation Theorem 7.17.
7.57 THEOREM Let R be a domain having the cone property in R". Let
s>Oandl sp, then Ws.P((n)+ Lr(R) for p Ir np/(n-sp).
(b) If n = sp, then WsSp(R)+ L'(R) for p I r < 03.
(c) If n < (s-j)p for some nonnegative integerj, then Ws9"(R)+ CBj(i2).
PROOF The results are already known for integer s so we may assume s is
not an integer and write s = m + a where m is an integer and 0 < u < 1. First
let us suppose m = 0. Then
W'*p(a) = L!'(a)n T'-'*p(R)
= LP(n)n T ( p , l-u-(l/p); W1*p(a),LP(a)).
Now the identity operator is continuous from W1,P(sl)into L"P/("-P)(a)and
(trivially) from LP(Q)into Lp(Q).By Theorem 7.17 it is also continuous from
T(p, 1 -u-(l/p); W1.p(a),Lp(a)) into ~ ( p1 -o-(~/~);L~P/(~-P)(~),LP(R)).
,
By Theorem 7.20 this latter space is imbedded in L"p/("-"p) provided n > ap.
Hence
~ P
0 (a)
, + Lnp/(n - UP)
0.
For general m, we argue as follows. Let U E Wm+u.p(fl).If la1 = m ,
then D"u E Wu3p(f2) + Lnp/(n-up)(a). If la1 I, m- 1, then Dau E W'*P(n)+
Hence Wm+uvp(S2)
Lnp/(n-up)(S2). + Wm~np/(n-up)(R). If n > sp, we have by
Theorem 5.4 that W m ~ n p ~ ( n - u+ p )Lnpi(n-sp)(Q).
(Q) Hence (a) is proved. If
n = sp, then Wm*np~("-up)(R) --* L'(R) for any r such that p I r < coy so (b) is
proved. If (s--j)p > n, then (rn-j)np/(n-ap) > n and so W m * n p ' ( n - a p ) (a)-b

C,'(n) and (c) is proved. I
The restriction p < n in the above theorem is unnatural and was placed
only for the purpose of achieving a very simple proof.
The following theorem contains all the imbedding results cited above as
special cases. It comprises results obtained by several writers, in particular,
Besov [9, lo], Uspenskii [67,68], and Lizorkin [41]. The theorem is stated
for R" but can obviously be extended to domains with sufficient regularily,
such as those satisfying the conditions of Theorem 7.41. We shall not attempt
any proof.
7.58 THEOREM Let s > O , I 0 and x is not an integer, or
(iii) x 2 0 and 1 2 and x a nonnegative
integer. (In particular, one cannot in general strengthen Case A of Part I
of Theorem 5.4 to allow k = n -mp.)
BESSEL POTENTIALS-THE SPACES L".'(a) 219
Conversely, if p = q and if either

(iv) x = s - ( n - k ) / p > 0 and is not an integer, or
(v) x 2 0 and p 2 2,
then we have the reverse imbedding
W y R k ) + WS*P(R")
in the sense that each U E W X S P ( Risk )the trace on Rk (i.e., u = WIWk) of a
function w E Ws*p(Rn) satisfying
llwll~.P.W" s KIIullx,P. Wk
with K independent of u. (The trace is understood in the sense of Section 5.2.)
Bessel Potentials-The Spaces Ls."(R)
7.59 We shall outline here, without proofs, another method of constructing

fractional order spaces which originates in studies of Bessel potentials by
Aronszajn and Smith [7] (and their collaborators-Adams et al. [5] and
Aronszajn et al. [S]) and which is presented by Calderbn [13] and Lions and
Magenes [MI. The resulting spaces, denoted Ls,p(Q)(or HSsP(Q)by Lions
and Magenes-but not to be confused with the H-spaces of Nikol'skii defined
in Section 7.73) coincide with the spaces W s P p ( Qfor
) integer values of s if
1 < p c co, and for all s when p = 2.
The space Ls*p(R") is constructed directly in terms of Fourier transforms
of tempered distributions. It is shown then that for 1 < p < a,L s ~ p ( R "and
)
Ws*p(R") are isomorphic and homeomorphic when s is an integer. For any
values of s,,s,s2 with s, Is 5 s2, the space Ls9p(Rn) can be identified as an
intermediate space interpolated between L"'*P(W) and Lsz~P(R")by a "com-
plex" interpolation method (see Calderbn [IS] or Lions [36]) which is not
identical to the trace method of Lions described earlier. This interpolation
method then provides a means of defining Ls*p(Q)for domain R c R" as an
intermediate space between spaces of the form WmSp(Q) for integer values
of m.
Proofs of assertions made in the discussion of the spaces Ls*p(Q) and their
relationship to the spaces Ws*p(Q)can be found in one or another of the
papers by Calderbn, Lions, and Lions and Magenes cited above.
7.60 First we introduce the notion of tempered distribution. We denote by

9'(Rn) the space of rapidly decreasing functions in R", that is, functions 4
satisfying
sup Ix"DB4(x)I < co
X€R"
for all multi-indices ct and 8. The space Y ( R n )carries a locally convex topology
characterized by the following notion of convergence : The sequence {+,}
converges to 0 in Y (R") if for all ct and 8
lim Xd0@4~(x)= o uniformly on R".
j-rcu
It may be readily verified that the Fourier transform
and the inverse Fourier transform
are each continuous on Y ( R " ) into Y ( R " ) , and, since .F-'F4 =

FF-'+ = 4, each is in fact an isomorphism and a homeomorphism of
Y ( R " ) onto Y ( R " ) .
It is clear from the definition of 9'(Rn) that Q(Rn) -+ Y ( R " ) . Hence the
dual space Y f ( R " ) consists of those distributions T E !9'(R") which possess
continuous extensions to Y ( R " ) . For instance, if 1 ~p I03 andfE Lp(R"),
then
defines Tf E Yf(R"). The same holds for any function f of "slow growth" at
infinity, that is, for which for some finite k we have If(x)l I const lxlk a.e. in
some neighborhood of infinity. The elements of Y ' ( R " ) are therefore called
tempered distributions. Y ' ( R " ) is given the weak-star topology as dual of
Y ( R " ) and is a locally convex topological vector space with this topology.
The direct and inverse Fourier transformations are extended to Y ' ( R " ) by
.FT(4) = T(.F4), 9-l T ( 4 ) = T(.F-'4).

Once again, each is an isomorphism and a homeomorphism of Y ' ( R " ) onto
Y ' ( R " ) and .F6-'.FT = F.F-'T = T.
7.61 Given a tempered distribution u on R" and a complex number z the

Bessel potential of order z of u is denoted J'u and defined by
J'u = .F-'((l +I * 12)-='2.Fu).
Evidently J' is one-to-one on Y'(R") into Y'(R"). If Rez > 0 and 1 Ip I 03,
BESSEL POTENTIALS-THE SPACES c' (a) 221
or if Rez 2 0 and 1 < p < co, then J' transforms Lp(R") continuously into
Lp(R"), and DaJz+laldoes likewise.
7.62 For real s and 1 s p I co let Ls3p(R")denote the image of LP(Rn)

under the linear mapping J'. Thus L s ~ p ( R " ) - i Y ' ( Rfor
n ) every s, and
Ls*p(Rn)+ Lp(R") for s 2 0. If u E Ls7p(R"),
then there exists unique ti E Lp(R")
with u = JSC.We define
I u; L
S v 0It = IIu" It 0,p . W" .
With respect to this norm, Ls*p(R")
is a Banach space. We summarize some
of its properties.
7.63 THEOREM (a) If s 2 0 and 1 Ip < 00, then 9(R") is dense in

LS*p (R").
(b) If 1 < p < co and p' = p / ( p - l), then [LssP(Rn)]' r L-s*p'(R").
(c) It t c s, then Ls.p(R")4 L'*p(R").
(d) If i Is and if either 1 c p I q In p / [ n - ( s - t ) p ] < co o r p = 1 and
1 I q < n/(n-s+t), then Ls~p(R")+ P q ( R " ) .
-,Co9"(Rn).
(e) If 0 5 p 5 s - ( n / p ) < 1, then Ls3p(R")
(f) If s is a nonnegative integer and 1 0, then for every s we have
,+,P(R") + WS,"R") + Ls-"p(R").
7.64 We now describe a complex interpolation method of Calder6n [lS]

and Lions [36] in which setting the spaces L s * p ( R ncan
) also arise.
Let B, and B , be Banach spaces both imbedded in a topological vector
space X as in Section 7.1 1, and let the Banach space B,+B, be defined as in
that section. We denote by F ( B , , B , ) the space of functionsf of a complex
variable z = ~7 + iT taking values in B, + B, and satisfying
(i) f is holomorphic on the strip 0 < o < 1,
(ii) f i s continuous and bounded on the strip 0 5 ~7 I1,
(iii) f ( i r ) E B, for 7 E R, the map 7 +f(ir) is continuous on R into Bo,
and limlrl+iu f(i.r) = 0, and
(iv) f(l + i ~ ) E B , for T E R, the map T -f(l +i7 is continuous on R
into B , , and l i m l r l - , m ~+(i Tl ) = 0.
F(B,, B , ) is a Banach space with respect to the norm
Ilf;F(BO, B1)ll = max Ilf(

~ ~ f ( i 7 ) ~ ~ B ~ ~+ i T ) l l B ~ } .
IEB
For 0 5 u I 1 set
B, = +
[B,; B , ] , = { u E Bo B, : u =f(a) for some f E F(B,, B , ) } .
With respect to the norm
IIuIIB., = [Iu; CBO; B1ltrll = inf
/eF(Bo, Bi)
Ilf;F(BO, B1)ll,
/(m)=u
B, is a Banach space imbedded in Bo B,. +

The intermediate spaces B, possess interpolation characteristics similar to
those of the trace spaces of Lions. If C,, C,, and Yare spaces having properties
similar to those specified for Bo, B,, and X,and if L is a linear mapping from
+ +
Bo B, into C, C, satisfying
IILUllCo 5 KOIIullB~5 IILuIIC, I KlllUIIBI,
then for each u E B, we have Lu E C, and

IILullC, 5 K t - a K l u IlullB,.
The following theorem may be found in the papers by Calderbn [l5] or

Lions [36].
7.65 THEOREM For any real so and s,, and for 0 I u I 1 we have
[Lso.p(Rn);L S I , P ( R ~ )=
] , ~ ( - +1 J + ~ ~ I , P (R").
We remark that the corresponding statement for intermediate spaces

between Wso*P(Rn)and WSI~P(Rn) obtained by trace interpolation is not
valid for all so and s, though it is for certain values, in particular if so and s1
are consecutive nonnegative integers.
7.66 The above theorem suggests how the spaces Ls*P(R)may be defined for
arbitrary domains R c R". If s 2 0, let m be the integer satisfying
s Im < s+ 1 and define LsSp(R) = [Wm.p(R);Lp(R)]~m-,,,,. If R is suffi-
ciently regular to possess a strong m-extension operator, then an interpolation
argument shows that Ls*p(R)coincides with the space of restrictions to R of
functions in Ls*p(R"). Also, Theorem 7.65 is valid for the spaces LS>"(R)
provided 0 5 so, s, < m.
The definition of LsVp(R) for negative s is carried out in the same manner
as for the spaces Ws*p(Cl). One denotes by L$P(R)(where s > 0) the closure
of 9 ( R ) in Lsip(R) and defines, for 1 c p < 00 and s c 0, the space Ls*P(R)to-
be [L;s.p'(R)]', where (l/p)+(l/p') = 1.
All the properties stated for LsVp(R") in Theorem 7.63 possess analogs for
LsJ'(R) provided R is suitably regular.
OTHER FRACTIONAL. ORDER SPACES 223
Other Fractional Order Spaces
7.67 Certain gaps in the general imbedding theorem for the spaces
Ws*p(Rn) (see Theorem 7.58) led to the construction by Besov [9, lo] of a
family of spaces Bs*P(Rn)which differ from Ws*p(Rn) when s is a positive
integer and which naturally supplement these latter spaces in a sense to be
made precise below.
BS*P(R")is defined for s > 0 and 1 s p 5 as follows. Let s = m + a
where m is a nonnegative integer and 0 c (r II . The space Bs*p(Rn) consists
of those functions u in Wm*P(Rn)for which the norm Ilu;Bsvp(R")I/ is finite.
If 1 s p < Go,
llu; BS*P(R")II
IDau(x)- 2D"u((x+y)/2) D"u(y)lP +
dx dy}"'.
Ix -y 1" + u p
I ) U ] ) ~ , ~ , max
~ ~ , ess sup
+
p y x ) - 2D'u((x+JJ)/2) DQU(Y)(
I a l = m x,yoR" Ix-Yl"
X*Y
is a Banach space with respect to the above norm. If 1 I p c 00,

Bs*p(Rn)
is dense in Bs,P(R").
COrn(R")
7.68 LEMMA If 1 5 p c m and s > 0 is not an integer, then the spaces

Ws*P(R")and Bs*P(Rn)coincide, and have equivalent norms.
PROOF For functions u defined on R" we define the difference operator Az by

A,u(x) = U ( X + Z ) - u(x).
The second difference operator Az2 is then given by
) ~ u ( x + z+
A='U(X) = AZAzu(x) = u ( x + ~ z - ) u(x).
The identity
k-1
Az u = (1/2') A p Zu - -5 C (1/2') A$jzu
j=O
may readily be verified by expanding the sum on the right side.

Evidently, the norm of a function u in Bspp(R") is equivalent to
while by Theorem 7.48 the norm of u in Wsip(Rn)can be expressed in the form
It is clear that (42) is bounded by a constant times (43); we must prove the
reverse assertion.
Suppose, therefore, that u E Co"(R")). We have, using (41),
(We have substituted p = 2 k in ~ the first integral, p = 2'2 in the second.)

Since s is not an integer, we have CT < 1 and so k may be chosen large enough
that k(1-0) 2 1 . It then follows that (43) is bounded by a constant times
(42). Since Com(rW")is dense in B"*p(R") the lemma follows. I
7.69 If s is a positive integer and p = 2, the spaces Ws*2(R") and BS*'(R")

coincide. For p # 2, s an integer they are distinct but for any E > 0 we have
Ws+e*P(rWn) + Ws*P(R")
-+ BS*P(R") if 1 <p I 2
-+ w"p(R")
BS+E.P(Rn) -+ B"P(R") if p 2 2.
The spaces BSsP(R") are of interest primarily for their imbedding character-
istics. They possess a "closed system" of imbeddings and at the same time fill
gaps in the system of imbeddings for the spaces Wssp(R").
7.70 THEOREM Let s > 0, 1 I p I q I 00, and 1 I k I: n, k an integer.

Suppose
r = s - (nip) + (k/q)=- 0.
Then
B"P(W") + B'JyRk).
OTHER FRACTIONAL ORDER SPACES 225
Conversely, if p = q and r = [ s - ( n - k ) ] / p > 0, then the reverse imbedding

B"P(Rk) + BS*P(R")
holds, in the sense that each element u in B'*p(Rk)is the trace u = ulpr of some
element u in BS.P(R")satisfying
Ilv;BS*P(R")II s Kllu;B'*P(Rk)II,
where K is independent of u.
7.71 THEOREM If s > 0, 1 5 p 5 co, and 1 5 k 5 n, and if r =

[s - (n-k)]/p, then
--* B'*P(R")
WS*P(R")
and conversely
7.72 The definitions and theorems above can be extended to suitable domains
R c R" and smooth manifolds Qk of dimension k contained in a. For
1 5 p < co the norm in Bs*p(i2)is
11% BS*W)II
where R, = { y ~ a : ( x + y ) / 2 ~ R } .
7.73 A different class of spaces having imbedding properties similar to the

Besov spaces are the spaces H s * P ( Q introduced
) by Nikol'skii [49-511. These
spaces, having norms involving Holder conditions in the LP-metric, were
studied earlier than the (fractional order) W- or B-spaces and provided
impetus for the latter.
Again we set s = m + a where m 2 0 is an integer and 0 < CT I1. For
1 5 p < co and R c R" a function u belongs to H s * p ( f 2provided
) the norm
is finite, where Q, = { x E Q; dist ( x , bdry Q) 2 2~1).The obvious modification

is made f o r p = co so that in fact H"*"(Q) = B","(Q). An argument similar to
that of Lemma 7.68 shows that if s is not an integer, the second difference A*
in the norm of HSaP(R) can be replaced by the first difference A without

changing the space.
are larger than the corresponding spaces Ws.p(i2);
The spaces HssP(SZ)
but if E > 0, we have
The spaces Hs*p(IW") possess a closed system of imbeddings identical to those

of the Besov spaces, that is, Theorem 7.70 holds with B everywhere replaced
by H.Strong extension theorems can be proved for H-spaces over smoothly
bounded domains so that the imbedding theorem can be extended to such
domains and traces on smooth manifolds in them.
The imbedding theorems for the spaces Hs*P(R") and BSpP(R") are proved
by a technique involving approximation of functions in these spaces by entire
functions of exponential type in several complex variables (see Nikol'skii
[49], for example).
7.74 Numerous generalizations of the above spaces have been made, partly
for their own sake and partly to facilitate the solution of other problems in
analysis. We mention two such directions of generalization. The first involves
replacing ordinary LP-norms by weighted norms. The second involves the use
of different values of s and p in terms of the norm involving integration in
different coordinate directions (anisotropic spaces). The interested reader is
referred to two excellent survey articles (Nikol'skii [52] and Sobolev and
Nikol'skii [64]), and their bibliographies for further information on the whole
spectrum of spaces of differentiable functions of several real variables.
VIII
Orhz and Orlicz-Sobolev Spaces
Introduction
8.1 In this final chapter we present some recent results involving replacement
of the spaces LP(n)with more general spaces L A @ )in which the role usually
played by the convex function t P is assumed by more general convex functions
A ( t ) . The spaces LA@), called Orlicz spaces, are studied in depth in the
monograph by Krasnosel’skii and Rutickii [34] and also in the doctoral
thesis by Luxemburg [42] to either of which the reader is referred for a more
complete development of the material outlined below. The former also
contains examples of applications of Orlicz spaces to certain problems in
nonlinear analysis.
Following Krasnosel’skii and Rutickii [34] we use the class of “N-
functions” as defining functions A for Orlicz spaces. This class is not as wide
as the class of Young’s functions used by Luxemburg [42] (see also O’Neill
[55]); for instance, it excludes L’ (a)and L“(R) from the class of Orlicz spaces
However, N-functions are simpler to deal with and are adequate for our
purposes. Only once, in the proof of Theorem 8.35, is it necessary to refer to a
more general Young’s function.
If the role played by Lp(R) in the definition of the Sobolev space WMn,”(l2)
is assigned instead to an Orlicz space L A @ ) ,the resulting space is denoted by
W”LA(a)and called an Orlicz-Sobolev space. Many properties of Sobolev
227
228 VIII ORLlCZ AND ORLICZ-SOBOLEV SPACES
spaces have been extended to Orlicz-Sobolev spaces, mainly by Donaldson

and Trudinger [22]. We present some of these results in this chapter.
It is also of some interest to note that a gap in the Sobolev imbedding
theorem 5.4 can be filled by consideration of Orlicz spaces. Specifically, Case
B of that theorem provides no “best” target space for imbeddings of WmsP(f2)
with Q a “regular” domain in 08” and mp = n. We have Wmip(Q)-P Lq(R)for
p Iq < co but Wm~P(f2)t,Lm(Q).In Theorem 8.25 an optimal imbedding of
Wmsp(f2) into an Orlicz space is constructed. This result is due to Trudinger
C661.
N-Functions
8.2 Let a be a real valued function defined on [O, co) and having the
following properties :
(a) a(0) = 0, a ( t )> 0 if t > 0, lim,+wa ( t )= co;
a is nondecreasing, that is, s > t r: 0 implies a ( s ) 2 a ( t ) ;
(b)
(c) a is right continuous, that is, if t 2 0, then lim,,,, a ( s ) = &(I).
Then the real valued function A defined on [0, 00) by
is called an N-jiunction.
It is not difficult to verify that any such N-function A has the following
properties :
(i) A is continuous on [0, 0 0 ) ;
(ii) A is strictly increasing, that is, s > t 2 0 implies A(s) > A ( ? ) ;
(iii) A is convex, that is, if s, t 2 0 and 0 < A < 1, then
A ( h + ( l - A ) t ) IAA(s)+ ( l - A ) A ( t ) ;
(iv) lim,,,, A ( t ) / t = 0, lim,,,A(t)/t = co;
(v) if s > t > 0, then A (s)/s > A ( t ) / t .
Properties (i), (iii), and (iv) could have been used to define N-function since
they imply the existence of a representation of A in the form (1) with a having
the required properties (a)-(c).
The following are examples of N-functions :
A(t) = tp, 1<p<co,
A ( t ) = e‘ - t - 1,
A ( t ) = e(rp)- 19 l<P<co,
A @ ) = (1 + t ) log(l+t) - t.
N-FUNCTIONS 229
Evidently A ( t ) is represented by the area under the graph a = a ( ~from

)
7 = 0 to 7 =t as shown (Fig. 8). Rectilinear segments in the graph of A corre-
spond to intervals of constancy of a,and angular points in the graph of A
correspond to discontinuities (i.e., vertical jumps) in the graph of a .
1
5 s
/
a =47)
s=A(t)
8.3 Given LZ satisfying (a)-(c), we define

Z(s) = sup t . (2)
a(?)<s
It is readily checked that the function a so defined also satisfies (a)-(c) and
that a can be recovered from 2 via
a(t)= SUP S. (3)
;(s)st
(If a is strictly increasing, then Z = u - '.) The N-functions A and A" given by
are said to be complementary; each is the complement of the other. Examples

of such complementary pairs are:
-40)= WP, Z(s) = sP'/p', I <p < 0, (l/p) + (l/p') = 1;
~ ( t =) e' - t - I, A"(s)= ( I +s) log(1 +s) - s.

A"@) is represented by the area'to the left of the graph a = a ( t ) [or more
correctly 7 = Z(a)] from a = 0 to a = s as shown in Fig. 9. Evidently we have
st I A(?) + A"@), (5)
230 VIII ORLICZ AND ORLICZ-SOBOLEV SPACES
which is known as Young's inequality. Equality holds in ( 5 ) if and only if

either t = Z(s) or s = a ( t ) .Writing ( 5 ) in the form
K(s) 2 st -A(I)
and noting that equality occurs when t = S(s), we have
K(s) = max(sr-A(t)).
120
This relationship could have been used as the definition of the N-function 2
complementary to A .
t 7
Fra. 9
Since A and 2 are strictly increasing they have inverses and ( 5 ) implies
that for every t 2 0
A - l ( t ) P ( t ) s A(A-'(t)) + K ( K - ' ( t ) ) = 2t.
On the other hand, A (I) s t a ( t ) so that, considering Fig. 9 again, we have for
t>O
K ( A ( t ) / t ) < ( A ( r ) / t ) t= A @ ) . (6)
Replacing A (t) by t in (6), we obtain
K ( r / A - y f ) )< t.
Hence, for any t > 0 we have
a-yt) 5 21.
t < /I-'(?) (7)
N-FUNCTIONS 231
8.4 We shall require certain partial-ordering relationships among N-

functions. If A and B are two N-functions, we say that B dominates A globally
provided there exists a positive constant k such that
A ( ? ) IB ( k t ) (8)
holds for all t 2 0. Similarly B dominates A near infinity if there exist positive
constants k and to such that (8) holds for all t 2 to. The two N-functions A
and B are equivalent globally (or near infinity) if each dominates the other
globally (or near infinity). Thus A and B are equivalent near infinity if and
only if there exist positive constants k,, k,, and to such that if t 2 to, then
B(k, t) 5 A ( t ) 5 B ( k , t ) . Such will certainly be the case if
If A and B have respective complementary N-functions 2 and 4, then B

dominates A globally (or near infinity) if and only if A" dominates i[ globally
(or near infinity). Similarly A and B are equivalent if and only if A" and i[ are.
8.5 If B dominates A near infinity and A and B are not equivalent near
infinity, then we say A increases essentially more slowly than B near infinity.
This is the case if and only if for every k > 0
The reader may verify that (9) is in turn equivalent to the condition
B-'(t)
lim -= 0.
t+m A d ' ( ? )
Let 1 < p < 00. We shall hereafter denote by A, the N-function

A , ( t ) = t"/p, 0 It < CO. (10)
If 1 < p < q < 00, then A, increases essentially more slowly than A, near
infinity. However, A, does not dominate A , globally.
8.6 An N-function A is said to satisfy a global A,-condition if there exists

a positive constant k such that for every t 2 0,
A(2t) 5 kA(t). (1 1)
It is readily seen that this will be the case if and only if for every r > 1 there
exists a positive constant k = k ( r ) such that for all t 2 0
A(rr) I;kA(r). (12)
Similarly A is said to satisfy a A,-condition near infinity if there exists t o > 0

such that (1 1) [or equivalently (12) with r > 11 holds for all t 2 to. Evidently
to may be replaced by any smaller positive number t l , for if tl I t I to, then
If A satisfies a A,-condition globally (or near infinity) and if B is equivalent

to A globally (or near infinity), then B also satisfies such a A,-condition.
Clearly the N-function A , @ ) = t P / p , I < p < 00, satisfies a global A,-
condition. It may be verified that A satisfies a A,-condition globally (or near
infinity) if and only if there exists a finite constant c such that
holds for all t 2 0 (or for all t 2 to > 0) where A is given by (1).
Orlicz Spaces
8.7 Let R be a domain in R" and let A be an N-function. The Orlicz class
KA(R)is the set of all (equivalence classes modulo equality a.e. in R of)
measurable functions u defined on R and satisfying
Since A is convex KA($2) is always a convex set of functions but it may not be
a vector space; for instance there may exist u E KA(R) and 1 > 0 such that
h $ K A (a).
We call the pair (A, R) A-regular if either
(a) A satisfies a global A,-condition, or
(b) A satisfies a A,-condition near infinity and R has finite volume.
8.8 LEMMA KA (R) is a vector space (under pointwise addition and scalar
multiplication) if and only if (A, R) is A-regular.
PROOF Since A is convex we have:

(i) 1u E KA(R) provided u E KA(R)and Ills 1, and
(ii) if u E KA(R)implies lu E KA (R) for every complex A, then u, u E KA (R)
implies u + u E KA (a).
It follows that KA(R)is a vector space if and only if u E K A(0) and 111 > 1
implies lu E KA(R).
If A satisfies a global A,-condition and 111 > 1, then we have by (12) for
ORLICZ SPACES 233
U E KA(Q>
/nA(IAU(x)l)dX 5 k(lll)/ n A ( l W l ) .€ 03.
Similarly, if A satisfies a A,-condition near infinity and volR < co, we have
for 1A1> 1, u E KA(Q),and some to > 0
In either case KA(Q)is seen to be a vector space.

Now suppose that (A,R) is not A-regular and, if volR < co, that ro > 0
is given. There exists a sequence { t i } of positive numbers such that
(i) A (2t,) 2 2jA (tj),and
(5) ti 2 to > 0 if volR < 00.
Let { Q j }be a sequence of mutually disjoint, measurable subsets of Q such that
if volR = 00
V0lQj =
Let
Then
if volR = 00
A(t,)volR if volQ < 03.
But
m
j-nA(12u(X)I) 2 1 2 J A ( f j V0lQj
j=l ) = O3.
Thus KA(a)is not a vector space. I

8.9 The Orlicz space L A @ )is defined to be the linear hull of the Orlicz class
K, (Q), that is, the smallest vector space (under pointwise addition and scalar
multiplication) containing K A(Q). Evidently LA(Q) consists of all scalar
multiples Au of elements u E KA(Q). Thus KA(Q)c LA(Q);these sets being

equal if and only if (A, Q) is A-regular.
The reader may verify that the functional
is a norm on LA@),(This norm is due to Luxemburg [42].) For I J u J ) ~ > 0 the

infimum in (13) is attained; in fact, letting k decrease toward llull,, in the
inequality
J * A ( Y ) d X I1,
we obtain by monotone convergence
J-*A(-)dx S 1.
Equality may fail to hold in (15) but if equality holds in (14), then k = IIuII,,.
8.10 THEOREM L A @ )is a Banach space with respect to the norm (13).
The completeness proof is quite similar to that for the space Lp(Q)given
in Theorem 2.10. The details are left to the reader. [We remark that if
1 < p < 00 and A, is given by (lo), then
Lp(Q)= LA,(Q)= KAp(C2).
Moreover, IIUIIA,,* = OP
Ilullp,*.l
8.11 If A and A” are complementary N-functions, a generalized version of

Halder’s inequality
can be obtained by applying Young’s inequality ( 5 ) to lu(x)l/llullA and

Iu(x)I/IIulln and integrating over Q.
The following elementary imbedding theorem is an analog for Orlicz

spaces of Lemma 2.8.
8.12 THEOREM The imbedding &(a)+ L A @ ) holds if and only if

either
(a) B dominates A globally, or
(b) B dominates A near infinity and vol Q < 00.
ORLICZ SPACES 235
PROOF If A ( t ) IB(kr) for all t 2 0, and if u E Lfl(Cl),then
Thus u E LA(R) and lIulJA5 k IIullfl.

If volR < 00, let t , = A - ' ( ( 2 vol a)-'). If E dominates A near infinity,
then there exist positive to and k such that A (t) 5 B(kt) for t 2 to. Evidently
we have for t 2 tl
A(t) Kl B W ) ,
-
where K1 = max(l,A(r,)/B(kr,)). If U E LB(f2) is given, let U ( u ) =
{ x E R : Ju(x)/2K1k llullfl< t i } and W(u)= R R'(u). Then
Thus u E LA(R) and [lullA5 2kK1 IIullfl.

Conversely, suppose neither hypothesis (a) nor (b) holds. Then there
exist points rj > 0 such that
A ( t j ) 2 B(jtj), j = 1,2, ... ,
If volR < 00, we may assume in addition that
tj 2 (l/j)B-1(l/volR).
Let f2, be a subdomain of R having volume l / B ( j t j ) ,and let
8.13 A sequence {uj} of functions in LA@) is said to converge in mean to

u E L A@ )if
jlim
+m /nA(luj(x)-u(x)l)cix = 0.
Convexity of A implies that for 0 c E 5 1 we have
from which it is clear that norm convergence in L A (R) implies mean conver-
gence. The converse holds, that is, mean convergence implies norm con-
vergence if and only if the pair (A, R) is A-regular. The proof is similar to that
of Lemma 8.8 and is left to the reader.
8.14 Let E A (R) denote the closure in LA($2)of the space of functions u which
n,
are bounded on R and have bounded support in If u E KA(R), the sequence
{Uj) given by
converges a.e. on R to u. Since A ( l u ( x ) - u j ( x ) l ) s A ( l u ( x ) l ) , we have by

dominated convergence that uj converges to u in mean in LA(a).Therefore if
(A, R) is A-regular, then EA(R)= KA(R)= LA@). If (A, R) is not A-regular,
we have
EA KA(n) s LA (1 8)
so that EA(R) is a proper closed subspace of L A @ )in this case. To verify the
first inclusion of (18) let u E EA(R)be given, Let u be a bounded function with
bounded support such that I I u - u I ~ ~ < 3. Using the convexity of A and (15),
we obtain
whence 224-20 E KA(R).Since 2u clearly belongs to KA(R) and KA(R)is

convex we have u = 3(2u - 2u) +3(2u) belongs to KA (a).
8.15 LEMMA EA (a)is the maximal linear subspace of KA(R).
PROOF Let S be a linear subspace of KA(a)and let u E S. Then l u E K A(Q)

=-
for every scalar 1.If E 0 and uj is given by (17), then uj/&converges to U / E in
mean in LA(R) as noted in Section 8.14. Hence for sufficiently large values o f j
n
and therefore uj converges to u in L A @ ) .Thus S c EA(R). I

DUALITY IN ORLICZ SPACES 237
8.16 THEOREM Let R have finite volume and suppose that the N-function
A increases essentially more slowly than B near infinity. Then
LB(R) EA(n)*
PROOF Since LB(R)+ L A @ ) is already established we need only show that

LB(R)c EA(R). Since LB(R)is the linear hull of KB(R)and EA(R) is the
maximal linear subspace of KA(R), it is sufficient to show that lu E KA(R)
whenever u E KB(R)and l is a scalar. But there exists a positive number to
such that A(llzlt) I B ( t ) for all t 2 to. Thus
$* A(lnu(x)l) dx = ($ (x: lu(x)l 5 t o )

+$
(x: lu(x)l > t o )
} A ( l W x > l ) dx
I A(IIIto) volR + B(Iu(x)l)dx < 03
whence the theorem follows. I
Duality in Orlicz Spaces
sn
8.17 LEMMA For fixed u E Ln(R) the linear functional L, defined by
L,(u) = U(X)V(X) dx (19)
belongs to [LA(R)]’. Denoting by IILvll its norm in that space, we have

IlUlln 5 I l L v l l 5 2IIUlln. (20)
PROOF It follows by Holder’s inequality (16) that if u E LA@), then

ILv(u)l 5 211ullA IIVII1*
Thus L, is bounded on LA (0)and the second inequality of (20) holds.

To establish the first inequality we may assume that u # 0 in Ln(R) so
that llLvll = K > 0. Let
lo if u ( x ) = 0.
If llullA > 1, then for sufficiently small E > 0 we have
Letting E + O+ we obtain, using inequality (6),
This contradiction shows that I(u1IAs 1. Now
so that
We remark that the lemma holds also when L, is restricted to act on EA(a).
To obtain the first inequality of (20) in this case take I!L,II to be the norm of
L, in [EA(R)]’ and replace u in the above proof by x,,u where 2. is the character-
istic function of R,, = { x E R :1x1 s,n and lu(x)l S n } . Evidently ~ , , ubelongs
to EA(R),I(xnuIJA 5 1, and (21) becomes
Since x,(x) increases to unity a.e. in R we obtain by monotone convergence
8.18 THEOREM The dual space [EA(sZ)]’

of EA(R) is isomorphic and
homeomorphic to LA@).
PROOF We have already shown that any element u E LA(sZ)determines via

(19) a bounded linear functional on LA@),and so also on EA(R), having in
either case norm differing from 1 v 1 by at most a factor of 2. It remains to be
shown that every bounded linear functional on EA(R)is of the form L, for
some such u.
DUALITY IN ORLICZ SPACES 239
Let L E [EA(R)]’ be given. We define a complex measure rl on the measur-

able subsets of R having finite volume by setting
= L(xs),
xs being the characteristic function of S. Since
~na(~~s(x)~a-’(l/volS))dx
= (22)
Since the right side tends to zero with vol S, the measure rl is absolutely con-
tinuous with respect to Lebesgue measure and by the Radon-Nikodym
theorem 1.47, A may be expressed in the form
P
A(S)= J D(X) dx,

S
where v is integrable on R. Thus
L(u) =
holds for measurable, simple functions u.

I* u ( x ) v ( x ) dx
If u E EA(Q) a sequence of measurable, simple fuqctions ui converging a.e.

to u can be found such that lui(x)l 5 lu(x)l on R. Sin;ce luj(x)u(x)l converges
a.e. to lu(x)v(x)l, Fatou’s Lemma 1.44 yields
s llLll IIujIIA 5 llLll llullA

i
It follows that the linear functional
L,(u) =
I* U(X)O(X) dx
is bounded on EA(Q) whence D E LZ(R) by the remark following Lemma 8.17.

Since L, and L assume the same values on the measurable simple functions,
a set which [see the proof of Theorem 8.20(a)] is dense in EA(Q),they agree
on E A (R) and the theorem is proved. 1
A simple application of the Hahn-Banach extension theorem shows that

if EA(SZ) is a proper subspace of L A (R) [that is, if (A, R) is not A-regular], then
there exists a bounded linear functional L on LA(R) not given by (19) for any
D E LA@). We have as an immediate consequence the following theorem.
8.19 THEOREM LA@)is reflexive if and only if both (A, 0) and (1R)
,
are A-regular.
We omit any discussion of uniform convexity for Orlicz spaces. This

subject is treated in Luxemburg's thesis [42].
Separability and Compactness Theorems
We next generalize the approximation Theorems 2.13,2.15, and 2.19.
8.20 THEOREM (a) Co(R) is dense in EA(0).

(b) EA(0)is separable.
(c) If J, is the mollifier introduced in Section 2.17, then for each
u E E A ( 0 )we have lime,,+ J, * u = u in EA(R).
(d) COm(SZ)is dense in EA(fZ).
PROOFPart (a) is proved by the same method used in Theorem 2.13. In

approximating u E EA(0) first by simple functions we may assume that u is
bounded on R and has bounded support. This is required for the dominated
convergence argument used to show that the simple functions converge in
norm to u in EA(a).(The details are left to the reader.)
Part (b) follows from part (a) by the same proof as given for Theorem 2.15.
Consider (c). If u E E A (SZ), let u be extended to R" so as to vanish identically
outside R. Let u E LA@). Then
by Holder's inequality (16), where u,,(x) = u ( x - t y ) . Thus by (20) and

Theorem 8.18,
* - u\IA,R =
II ull Z,ns 1
IIn(J& * u(x) - u(x>)u(x>dx
5 1 IYlSl
-
J ( Y ) Iluey u I I A , R dy.
Given 6 > 0 we can find u" E C, (a) such that I\u - fi 1 A , < 6/6. Evidently
< 6/6 and for sufficiently small E , IIfi,y-U"IIA,n < 6/6 for every y
(Iu,y-fiey/IA,n
with lyl I 1. Thus I I J , * U - - U I I ~<, ~6 and (c) is established.
Part (d) is an immediate consequence of (a) and (c). I
SEPARABILITY AND COMPACTNESS THEOREMS 241
We remark that LA(R)is not separable unless LA@) = EA(R), that is,
unless (A,R) is A-regular. A proof of this fact may be found in the work of
Krasnosel'skii and Rutickii [34, Chapter 11, Theorem 10.21.
8.21 A sequence uj of measurable functions is said to conuerge in measure on

R to the function u provided that for each E > 0 and 6 > 0 there exists an
integer M such that i f j > M , then
vol{xER:~ui(x)-u(x)~>&}
I d .
Clearly, in this case there also exists an integer N such that if j , k 2 N, then
vol{x E : Iuj(x)-uk(x)I 2 &} 5 6 .
8.22 THEOREM Let R have finite volume and suppose that the N-
function Bincreases essentially more slowly than A near infinity. If the sequence
{u]} is bounded in LA(R)and convergent in measure on R, then it is convergent
in L@).
PROOF Fix E > 0 and let Uj,k(X) = [uj(X)-Uk(X)]/&. Clearly {uj,k} is

bounded in LA(R);say I K. Now there exists a positive number to
such that if t > to, then
B(r) I *A(t/K).
Let 6 = 1/4B(t0) and set
Rj,k = {xER:Iuj,k(X)IrB-1(1/2volR)}.
Since {uj} converges in measure there exists an integer N such that if j , k 2 N ,
then vdRj,k I 6. Set
0S.k = {x E Rj,k : luj,k(X)I 2 t o } , fir!
1.k = n.
j.k '>,k*
For j , k 2 M we have
/nB(luj,k(x)I)dx = (/ n-ni,k +InI,.)B(Iuj,k(x)l)dx
I-
vol R
2 vol R
+!I,
4 n,,k
) 1.
A(!y)dX+6B(fOI
Hence IIUj-UkIIB,Q IE and so {uj} converges in L,(R). I

The following theorem will prove useful when we wish to extend the
Rellich-Kondrachov Theorem 6.2 to imbeddings of Orlicz-Sobolev spaces.
8.23 THEOREM Let R have finite volume and supposethat the N-function
B increases essentially more slowly than A near infinity. Then any bounded
subset S of LA (0) which is precompact in L
' (R) is also precompact in LB(R).
PROOF Evidently LA@)+ L' (R) since R has finite volume. If {uj*} is a
sequence in S,it has a subsequence {u]} convergent in L1(a);say u, + u in
L
' (0).Let e, S > 0. Then there exists an integer N such that if j 2 N,then
IIuj-uIIl,n4&S. It follows that vol{xER:lu,(x)-u(x)lr&} 1 6 so {uj}
converges in measure on R and hence also in LB(R). I
A Limiting Case of the Sobolev Imbedding Theorem
8.24 If mp = n and p > 1, the Sobolev imbedding Theorem 5.4 provides no

best (i.e., smallest) target space into which Wm*p(R)can be imbedded. In
fact we have in this case, for suitably regular R,
W"*P(R) + LyR), p 5q< 00,
but (see Example 5.26)
W"*P(R) c$ L"(0).
If the class of target spaces for the imbedding is enlarged to include Orlicz
spaces, then a best target space can be found. We consider first bounded
domains R. The case' m = 1 of the following theorem was established by
Trudinger [66].

property. Let mp = n and p > 1. Set
A ( t ) = exp[t"/("-")] - 1 = exp[tP/(P-')] - 1. (23)
Then there exists the imbedding
W""(R) LA(n).
PROOF Let x E R and let C be a finite cone contained in R having vertex at

x. Let u E C"(C). Applying Taylor's formula
to the functionf(t) = u ( y + t ( x - y ) ) , and noting that
f(')(r) = 2
lal=i
a!
b u ( y + t(x-y))(x-y)',
A LIMITING CASE OF THE SOBOLEV IMBEDDING THEOREM 243
we obtain
Let Y be the volume and h the height of C. Let (p, 8) denote spherical polar
coordinates of y E C referred to x as origin so that C is specified by
0 < p < h, 8 E Z, and the volume element dy can be written in the form
p n - ' o ( 8 ) dp do. Then
By density the above inequality holds for all u E Wmil(C).In particular, for
any u E Wmip(R),and for almost all x E R, we have
where K2 depends on m,n and the height h and volume V of the cone deter-
mining the cone property for R.
We wish to estimate I I u ~ ~for
~ , arbitrary
~ s > 1. Accordingly, if u E L"@)
where s' = s/(s- l), then
Now we have from Lemma 5.47 that if 0 5 v < n, then

1
dx 5 K3(v,n) (vol R)' -@I").
In fact a review of the proof of that lemma shows that K3(v, n) = K,/(n- v)
with K4 depending only on n. Hence
Hence
The constant K7 depends only on m,n and the cone determining the cone
property for R. Setting s = nk/(n-m) = p k / ( p- I), we obtain
A LIMITING CASE OF THE SOBOLEV IMBEDDING THEOREM 245
Since eP/(P-’)> e, the series C;P= (l/k!)(k/eP/(P-l))k converges to a finite

sum, say KE. Let K9 = max(1, KEvolR)and put
KIo = eK9K, [ p / ( p - l)](p-l)/pIIullm,p = K11IIUllm,p*

Then
since K9 2 1 and p k / ( p - 1) > 1. Expanding A ( ? ) in power series, we now

obtain
Hence u E LA(a) and

I l ~ l l A5 K1o = K11IIUllm.p,
where K, depends on n,m yvolR, and the cone determining the cone property
forR.
The imbedding established in the above theorem is “best possible” in the

sense that if there exists any imbedding of the form
) MR),
w7p(Q+
then A dominates B near infinity. A proof of this fact for the case m = 1,
n = p > 1 can be found in the notes of Hempel and co-workers [30]. The
general case is left to the reader as an exercise.
Theorem 8.25 can be generalized to fractional-order spaces. For results
in this direction the reader is referred to Grisvard [28] and Peetre [56].
8.26 If R is unbounded and so (having the cone property) has infinite volume,
then the N-function A given by (23) may not decrease rapidly enough at zero
to allow membership in L A @ ) of every u E Wm*p( f2) (where mp = n). Let k ,
be the smallest integer such that ko 2 p - 1 and define a modified N-function
A0 by
246 VIII ORLICZ AND ORLICZ--SOBOLEV SPACES
Evidently A . is equivalent to A near infinity so for any domain R having

finite volume, LA@) and LAo(R) coincide and have equivalent norms.
However, A . enjoys the further property that for 0 < r s 1,
Ao(rt) 5 rkoP/(P-l)AO(t)
5 rPAo(t). (24)
We show that if mp = n and R has the cone property (but may be unbounded),
then
W"*P(R) -b LAo(R).
As in the proof of Lemma 5.14 we may write R as a union of countably
many subdomains R, each having the cone property determined by some fixed
cone independent of j, satisfying for some constants K , and K2
0 < K1 5 v01Rj S K,,
and finally such that for some positive integer R any R + 1 of the subdomains
a, have empty intersection. It follows by Theorem 8.25 that if u E Wm*p(R),
then
I I ~ I I A ~5, ~K3
, IIullm,p,n,~
where K3 does not depend on j. Using (24) with r = RIIp(1 u )I,',, 11 u 11 ,, p ,
and the finite intersection property of the domains Q, , we have
Hence I ~ U ~ J , , ~ 5
, ~ R1IPK3~ \ u ~ ~ as, , required.
,,~,~
We remark that if ko > p- 1, the above result can be improved slightly
by using in place of A,, the N-function max(tP,A,@)).
Orlicz-Sobolev Spaces
8.27 For a given domain R in R" and a given defining N-function A the
Orlicz-Sobolev space W"LA (a) consists of those (equivalence classes of)
functions u in LA (R) whose distributional derivatives D'u also belong to
L A @ )for all c1 with 1alI;m. The space W"EA(a)is defined in analogous
fashion. It may be checked by the same method used in the proof of Theorem
3.2 that WmLA(R)is a Banach space with respect to the norm
IIUllm.A = IIUllm,A,n = max IID'uIIA,~, (25)
Oslalsm
IMBEDDING THEOREMS FOR ORLICZ-SOBOLEV SPACES 247
and that WmEA(0)is a closed subspace of W"LA(Q) and hence also a Banach
space under (25). It should be kept in mind that W"E,(Q) coincides with
WmL,(R)if and only if ( A , Q ) is A-regular. If 1 < p < CCI and A p ( t ) = t P ,
then W"'LAp(Q)= W"'EAp(R) = W"'*p(i2),the latter space having norm
equivalent to that of the former.
As in the case of ordinary Sobolev spaces, W,"L,(R) is taken to be the
closure of Com(Q)in WmLA(Q). [An analogous definition for WOmEA(R)
clearly leads to the same space in all cases.]
Many properties of Orlicz-Sobolev spaces are obtained by very straight-
forward generalization of the proofs of the same properties for ordinary
Sobolev spaces. We summarize some of these in the following theorem and
refer the reader to the corresponding results in Chapter 111for method of proof.
The details can also be found in the article by Donaldson and Trudinger [22].
8.28 THEOREM (a) WmEA(R) is separable (Theorem 3.5).

(b) If ( A , Q ) and (2,Q)are A-regular, then W"'EA(!2)= W"'LA(Q) is
reflexive (Theorem 3.5).
(c) Each element L of the dual space [W"EA(Q)]' is given by
L(u) = c
Oslo1S m
~ n o % ( x ) u , ( x )dx
for some functions u, E LA@), 0 s1.1 Im (Theorem 3.8).

(d) Cm(Q)n W"E,(Q) is dense in W"'E,(Q) (Theorem 3.16).
(e) If 0 has the segment property, then Cm@) is dense in W"EA(Q)
(Theorem 3.18).
(f) Com(Rn) is dense in W"'E,(R"). Thus WomLA(Rn) = W"'EA(R")
(Theorem 3.18).
Imbedding Theorems for Orlicz-Sobolev Spaces
8.29 Imbedding results analogous to those obtained for the spaces W m V p ( R )

in Chapters V and VI can also be formulated for the Orlicz-Sobolev spaces
WmLA (Q) and W"EA(a).The first results in this direction were obtained by
Dankert [20] and by Donaldson. A fairly general imbedding theorem along
the lines of Theorems 5.4 and 6.2 was presented by Donaldson and Trudinger
[22] and we develop it below.
As was the case with ordinary Sobolev spaces, most of the imbedding
results are obtained for domains having the cone property. Exceptions are
those yielding (generalized) Holder continuity estimates; these require the
strong local Lipschitz property. Some results below are obtained only for
bounded domains. The method used in extending the analogous results for
ordinary Sobolev spaces to unbounded domains (see Lemma 5.14) does not
appear to extend in a straightforawrd manner when general Orlicz spaces are
involved. In this sense the imbedding picture is still incomplete.
8.30 We concern ourselves for the time being with imbeddings of W' LA(R) ;
theimbeddings of WmLA(R) are summarized in Theorem 8.40. In the following
R is always understood to be a domain in R".
Let A be a given N-function. We shall always suppose that
replacing, if necessary, A by another N-function equivalent to A near infinity.

[If R has finite volume, (26) places no restrictions on A from the point of view
of imbedding theory since N-functions equivalent near infinity determine
identical Orlicz spaces.]
Suppose also that
For instance, if A = A, given by (lo), then (27) holds precisely when p 5 n.

With (27) satisfied we define the Sobolev conjugare A, of A by setting
It may readily be checked that A, is an N-function. If 1 < p < n, we have,

setting q = np/(n-p),
A,&) = ql-qp-q/PAq(r),
It is also readily seen for the case p = n that A,,(t) is equivalent near infinity
to the N-function e'- t - 1.
Before stating the first imbedding theorem we prepare the following

technical lemma that will be needed in the proof.
8.31 LEMMA Let u E W,:;' (R) and let f satisfy a Lipschitz condition on
R. Then g E W&'(R) where g ( x ) =f(lu(x)l), and
D,s<x> =f'(l.(x>I) s g n w * D,u(x).
-
PROOF Since Iu I E W,:;' (R) and 0,Iu ( x )I = sgn u ( x ) Dj u (x) it is sufficient
to establish the lemma for positive, real-valued functions u so that
g ( x ) = f ( u ( x ) ) . Let 4 E 9(0).
Letting {e,}y=l be the usual basis in R", we
obtain
u(x+hej) - u ( x )
h-0 h d (4 dx,
where, since f is Lipschitz, for each h the function Q(-,h) is defined a.e. on
0 by
otherwise.
Moreover, IIQ(. ,h)((,,n 5 K for some constant K independent of h. A well-
known theorem in functional analysis assures us that for some sequence of
values of h tending t o zero, Q ( . , h ) converges tof'(u(.)) in the weak-star
topology of L"(f2). On the other hand, since u E W ' . (supp 4) we have
which evidently implies the lemma. I

8.32 THEOREM Let R be bounded and have the cone property in R". If
(26) and (27) hold, then
W I L ( W LA.(Q).
+
Moreover, if B is any N-function increasing essentially more slowly than A*

near infinity, then the imbedding
is compact.
PROOF The function s = A , ( t ) as defined by (28) satisfies the differential

equation
and hence, by virtue of the left inequality of (7),

ds
-
dt
< s""K-'(s).
Therefore a ( ! ) = [A*(!)]("- satisfies the differential inequality
da n-1
- < -K-l((a(t)y/("-l)).
dt n
(30)
Let U E W'.L,(Q) and suppose for the moment that u is bounded on Q

and is not zero in LA@).Then j n A * ( l u ( x ) l / l ) dx decreases continuously
from infinity to zero as I increases from zero to infinity, and accordingly
assumes the value unity for some positive value of A. Thus
Let f ( x ) = a ( l u ( x ) l / K ) . Evidently u E W1*'

(0)and a is Lipschitz on the
range of lul/Kso that, by Lemma 8.31,fbelongs to W'*'(Q).By Theorem 5.4
we have W'*'(Q)4C/("-')(R) and so
By (31) and Holder's inequality (16),we obtain
Making use of (30), we have
Suppose 1> 1. Then

Thus
Let g ( t ) = A&)/? and h ( t ) = c ( t ) / t .It is readily checked that h is bounded

on finite intervals and limr-rmg ( t ) / h ( t ) = 03. Thus there exists a constant to
such that if t 3 to, then h ( t ) G g ( t ) / 2 K 1 .Putting K2 = K , ~ u p ~ ~ ~ ~ we ,~h(t),
have, for all t 3 0,
a ( t ) G (1/2K,)A*(t) + (K2IKl)t.
Hence
where K3 = 2K2 11 1lld c co since R has finite volume.

Combining (33)-(35), we obtain
+ 3 + (KJ/K)lI4lA,
1 G (2~l/K)(~-1)ll4ll,"
so that
l141Al = K-G K4II4I1,A. (36)
We note that K4can depend on n,A, volR, and the cone determining the cone
property for R.
To extend (36) to arbitrary u E W'L,(R) let
Evidently #k is bounded and belongs to W' LA(Q) by Lemma 8.31. Moreover,

Il#kllA, increases with k but is bounded by K4 \lull I , A . Thus limk-rmII#kllA, = K
exists and K < K4 llull 1 , A . By Fatou's lemma
whence u E LA*@)and (36) holds.

Since fl has finite volume we have
W'L"(Q) + W'J(R) + L'(R),
the latter imbedding being compact by Theorem 6.2. A bounded subset S of
W I L A ( Q is
) bounded in LA@) and precompact in L'(R), and hence pre-
compact in LB(R) by Theorem 8.23 whenever B increases essentially more

slowly than A near infinity. I
Theorem 8.32 extends to arbitrary (even unbounded) domains R provided

W is replaced by W,.
8.33 THEOREM Let R be any domain in R". If the N-function A satisfies

(26) and (27), then
WO'LAR) + LA.(Q.
Moreover, if R, is a bounded subdomain of R, then the imbeddings

W0lLA + LB(RO)
exist and are compact for any N-function B increasing essentially more
slowly than A near infinity.
PROOF If u E W,'L,(Q), then the function f in the above proof can be

approximated in W'.'(R) by elements of Com(Q).By Sobolev's inequality
(Section 5.11), (32) holds with the term 1 f [lo, absent from the right side.
Therefore (35) is not needed and the proof does not require that R has finite
volume. The cone property is not required either since Sobolev's inequality
holds for all u E Cooo(R").The compactness arguments are similar to those
above. I
8.34 REMARK Theorem 8.32 is not optimal in the sense that for some A,
LA. is not necessarily the smallest Orlicz space into which W'L,(R) can be
imbedded. For instance if A ( t ) = A,(?) = t"/n, then, as noted earlier, A , ( t )
is equivalent near infinity to e'- t- 1. However this N-function increases
essentially more slowly near infinity than does exp(t"'("-'))- 1 so that
Theorem 8.25 gives a sharper result than does Theorem 8.32. Donaldson and
Trudinger [22] assert that Theorem 8.32 can be improved by the methods of
Theorem 8.25 provided A dominates near infinity every A, with p < n, but
that Theorem 8.32 gives optimal results if for some p < n, A, dominates A
near infinity.
8.35 THEOREM Let Q have the cone property in 88". Let A be an N-

function for which
d7 c co
Then
PROOF Let C be a finite cone contained in R. We shall show that there exists
a constant K, depending on n, A, and the dimensions of C such that
IIUllm,c G KI IIuIl1,A,C* (39)
In so doing, we may assume without loss of generality that A satisfies (26),
for if not, and if B is an N-function satisfying (26) and equivalent to A near
infinity, then W ’ L , ( C ) -+ W’ LB(C)with‘imbedding constant depending on
A, B, and vol C by Theorem 8.12. Since B satisfies (38) we would have
IIUllm.C G K2 ~ ~ t ~ ~ ~ G IIuII1,A,C*
i , KB3 , C
Now R can be expressed as a union of congruent copies of some such

finite cone C so that (39) clearly implies
II&J,n G (40)
r$2
KlIlulll,A,R*
Since A is assumed to satisfy (26) and (38) we have
dt = K4 < a.
Let
Then A-’ maps [O,oo) in a one-to-one way onto [0, K4) and has convex
inverse A. We extend the domain of definition of A to [0, co) by setting
A ( t ) = co for t 2 K4. The function A, which is a Young’s function (see
Luxemburg [42] or O’Neill [SS]), is not an N-function as defined in Section
8.2 but nevertheless the Luxemburg norm
is easily seen to be a norm on Lm(C)equivalent to the usual norm; in fact,

(~/K4)ll~llm,c
G Il ~l Ih ,cG [ l / A - l ( l / ~ ~ l C I)I I4 m , c . (41)
Moreover, s = A ( t ) satisfies the differential equation (29) so that the proof
of Theorem 8.32 can be carried over in this case to yield, for u E W’L,(C),
IIuII*,c G K5 I I 4 I l , A , C * (42)
Inequality (39) now follows from (41) and (42).
By Theorem 8.28(d) an element U E W’E,(Q) can be approximated in
norm by functions continuous on R.It follows from (40) that u must coincide
a.e. in R with a continuous function. (See the first paragraph of the proof of
Lemma 5.15.)
Suppose that an N-function B can be constructed so that B ( t ) = A (t) near

zero, B increases essentially more slowly than A near infinity, and
Then by Theorem 8.16, U E W'L,(C) implies U E W'E,(C) so that

W'L,(R) c C(R) as required.
It remains, therefore, to construct such an N-function B. Let
I < t , < 1, < be such that
We define a sequence {sk} with sk 2 tk, and the function B - ' ( t ) , inductively
as follows.
Let s1 = t , and B - ' ( f ) = A - ' ( t ) for 0 d t d sl. Having chosen
'
sl,s,, ...,s,- and defined B-' ( t ) for 0 d t Q sk- we continue B - ( 1 ) to the
right of 9 - 1 along a straight line with slope ( ~ - ' ) ' ( s k - ~ -(which
) always
exists since A-' is concave) until a point tk' is reached where B-'(tk') =
2 k - ' ~ - '(tk'). Such tk' exists because lim,+,, A - ' ( t ) / t = 0. If tk' > t,, let
sk = fk'. Otherwise let = tk and extend B-' from tk' to sk by setting B-' ( t ) =
2 k - 1 A - 1 ( t ) .Evidently B-' is concave and B is an N-function. Moreover,
B ( t ) = A ( t ) near zero and since
B-'(t)
lim- = a,
,+a0 A-'(t)
B increases essentially more slowly than A near infinity. Finally,
dt
as required.
8.36 THEOREM Let R be a domain in Rnhaving the strong local Lipschitz

property. If the N-function A satisfies
then there exists a constant K such that for any u E WILA(R)(which may be
assumed continuous by the previous theorem) and all x , y E R we have
PROOF We establish (44)for the case when R is a cube of unit edge; the
extension to more general strongly Lipschitz domains can be carried out just
as in the proof of Lemma 5.17. As in that lemma we let R, denote a parallel
subcube of C2 and obtain for x E Iz,
BY W), 1 1 I ~ Z , ~ ~ , =
, l/X-' (t -"o-"). It follows by Holder's inequality and (7)
that
Hence
If x , y ~ Rand o = I x - y l < 1, there exists such a subcube R, with

x,y E nu
c i2. Using (45) for x and y, we obtain
For I x - y l > 1, (44)follows from (40) and (43). 1

8.37 Let M denote the class of positive, continuous, increasing functions of
t > 0 which tend to zero as t decreases to zero. If p E M ,the space C,,(i?),
consisting of functions u E C(i?) for which the norm
is finite, is a Banach space under that norm. The theorem above asserts that if
(43) holds, then
If p, v E M are such that p / v E M ,then for bounded R we have, as in

Theorem 1.31, that the imbedding
c,(m -b C"(Q
is compact. Hence so also is
W'UQ -b C"(IZ>,
if p is given by (46).
The following is a trace imbedding theorem which generalizes (the case

rn = 1 of) Lemma 5.19.

property, and let Qk denote the intersection of R with a k-dimensional plane
in R". Let A be an N-function for which (26) and (27) hold, and let A , be given
by (28). Let 1 1 and C is an N-function increasing essentially more
slowly than A:/" near infinity, then the imbedding
W'L*(n) + Lc(nk) (47)
is compact.
PROOF The problem of verifying that A:/" is an N-function is left to the

reader. Let u E W'L,(Q) be a bounded function. Then
We wish to show that

K 6 K1IIullt,*,n (49)
with Kl independent of u. Since (49) is known to hold for the special case
k = n (Theorem 8.32) we may assume without loss of generality that

K 3 II41'4.,n = llullA:/",n". (50)
Let w ( t ) = [ A * ( t ) ] l / qwhere q = np/(n-p). By Lemma 5.19 (the case m = 1)
Now set g ( r ) = A * ( r ) / t P and h ( t ) = ( m ( r ) / r ) p . It is readily checked that

lim,-+mg ( t ) / h ( r )= a.In order to see that h ( r ) is bounded near zero let
s = A , ( t ) and consider
Since B is an N-function limt-,m+B-'(7)/7 = 00. Hence for sufficiently small

values o f t we have
Therefore there exists a constant K3 such that for t 2 0

( ~ ( t ) ) 'd (1/2K2)A*(t) + K3 t P .
Using (50), we now obtain
From (51)-(53) there follows the inequality

2nK2 1 1 K4K2
1G-e-K P qpI I ~ I I ? . A . * + j ++uIIPA,n
and hence (49). The extension of (49) to arbitrary u E WILA(Sl)now follows

as in the proof of Theorem 8.32.
Since B ( t ) = A ( t l / P ) is an N-function and R is bounded we have
W'LA(R)+ W1iP(Sl) -,L ' (p), the latter imbedding being compact by
Theorem 6.2 (provided p > 1). The compactness of (47) now follows by
Theorem 8.23. I
8.39 We conclude this chapter with the general Orlicz-Sobolev imbedding

theorem of Donaldson and Trudinger [22]. For a given N-function A we
define a sequence of N-functions Bo, B l ,B2,... as follows:
BOW = A @ )
At each stage we assume that

replacing Bk, if necessary, by another N-function equivalent to it near infinity

and satisfying (54). Let J = J ( A ) be the smallest nonnegative integer such that
Evidently J ( A ) 6 n.
If p belongs to the class M defined in Section 8.37, we define the space
C,,"(n) to consist of those functions u E C"(n)for which D"u E C,,(n). The
space C,,"(n) is a Banach space with respect to the norm
Ilu; c,"'(n)II = max IID"u;C,,(X'BII.
I4sm
8.40 THEOREM Let R be a bounded domain in Rn having the cone

property. Let A be an N-function.
(a) If m 6 J(A), then W"LA (R) + L,,,, (R) and the imbedding
WmLA(R)+ Lc(R) is compact for any N-function C increasing essentially
more slowly than B,,, near infinity.
(b) If m > J ( A ) , then W'"LA(R)--+ C,(R) = C(R) n Lm(R).
(c) If also R has the strong local Lipschitz property and if m > J = J(A),
then WmLA(In) + CF-J-l (a) where
Moreover, the imbeddings W"'LA(R)+ Cm-J-l(a) and W"L, (a)+

C:-J-' (32) are compact provided v E M and p/v E M.
8.41 REMARK The above theorem follows in a straightforward way

from Theorems 8.32,8.35, and 8.36. Moreover, if we replace LA by E A in part
(a), we get W"EA(R)+ EB,(R) since the sequence {u,} defined by (37) in the
proof of Theorem 8.32 converges to u if u E W'EA(R).Theorem 8.40 holds
without restriction on R if W"LA(R)is everywhere replaced by Wo"LA(R).
References
1. R. A. Adams, Some integral inequalities with applications to the imbedding of Sobolev

spaces defined over irregular domains, Trans. Amer. Math. SOC. 178 (1973). 401429.
2. R. A. Adams, Capacity and compact imbedding. J. Math. Mech., 19 (19701,923-929.
3. R. A. Adams and J. Fournier, Compact imbedding theorems for functions without
compact support, Canad. Math. Bull., 14 (1971), 305-309.
4. R. A. Adams and J. Fournier, Some imbedding theorems for Sobolev spaces, Canad.
J . Math. 23 (1971), 517-530.
5. R. D. Adams, N. Aronszajn, and K.T. Smith, Theory of Bessel potentials, part 11, Ann.
Inst. Fourier (Grenoble), 17 (No.2) (1967), 1-135.
6. S.Agmon, “Lectures on Elliptic Boundary Value Problems,” Van Nostrand-Reinhold,
Princeton, New Jersey, 1965.
7. N. Aronszajn and K. T. Smith, Theory of Bessel potentials, Part 1, Ann. Inst. Fourier
(Grenoble) 11 (1961), 385475.
8. N. AronsaJn, F. Mulla and P. Szeptycki, On spaces of potentials connected with Lp-
classes, Ann. Inst. Fourier (Grenoble) 13 (No.2) (1963), 21 1-306.
9. 0. \I. Besov, On a certain family of functionel spaces. Imbedding and continuation
theorems, Dokl. Akad. Nauk SSSR, 126 (1959), 1163-1 165.
10. 0.V. Besov, OnsomeconditionsofmembershipinLPforderivativesofperiodic functions
NauEn. Dokl. Vyss. Skoly. Fit.-Mat. Nauki (1959), 13-17.
1 1. F. E.Browder, On the eigenfunctions and eigenvalues of general linear elliptic operators,
Proc. Nat. Acad. Sci. USA 39 (1953), 433439.
12. F. E. Browder, On the spectral theory of elliptic differential operators I, Math. Ann.
142(1961), 22-130.
13. P. L. Butzer and H. Berms, “Semigroups of Operators and Approximation,”
Springer-Verlag, Berlin, 1967.
260
REFERENCES 261
14. A. P. Calderbn, Lebesgue spaces of differentiable functions and distributions, Proc.

Sym. Pure Math., 4 (1961), 3349.Amer. Math. Soc.,Providence, Rhode Island.
15. A. P. Calderh, Intermediate spaces and interpolation, the complex method, Studia
Math. 24 (1 964), 1 13-1 90.
16. A. P. Calder6n and A. Zygmund, On the existence of certain singular integrals, Acta
Math. 88 (1952),85-139.
17. C. W.Clark, The Hilbert-Schmidt property for embedding maps between Sobolev
spaces, Canud. J. Math. 18 (1%6), 1079-1084.
18. C. W. Clark, “Introduction to Sobolev spaces,” (Seminar notes) Univ. of British
Columbia, Vancouver, 1968.
19. J. A. Clarkson, Uniformly convex spaces, Trans. Amer. Math. SOC.40 (1936),396-414.
20. G.Dankert, “Sobolev imbedding theorems in Orlicz spaces,” (Thesis). Univ. of
Cologne, 1966.
21. J. Deny and J. L.Lions, Les espaces du type de Beppo Levi, Ann. Inst. Fourier (Grenoble)
5 (1955),305-370.
22. T . K.Donaldson and N. S . Trudinger, Orlicz-Sobolev spaces and imbedding theorems,
J. Functional Analysis, 8 (1971), 52-75.
23. G. Ehrling, On a type of eigenvalue problem for certain elliptic differential operators,
Math. Scand. 2 (1954),267-285.
24. E. Gagliardo, Properieta di alcune classi di funzioni in piu variabili, Ricerche Mat. 7
(1958), 102-137.
25. E.Gagliardo, Ulteriori properieth di alcuneclassi di funzioni in piu variabili, Ricerche
Mat. 8 (1 959). 24-5 1 .
26. I. G.Globenko, Embedding theorems for a region with null salient points, Dokl. Akud.
Nauk SSSR132(1960),251-253English trans.: SovietMath. Dokl. 1 (1960),517-519
27. I. G. Globenko, Some questions in the theory of imbedding for domains with singularities
on the boundary, Mat. Sb., 57 (99)(1962),201-224.
28. P. Grisvard, Commutativite de deux foncteurs d’interpolation et applications, J . Math.
Pures Appl. 45 (1966)143-290.
29. G.H.Hardy, J. E. Littlewood, and G. Polya, “Inequalities,” Cambridge Univ. Press,
London and New York, 1943.
30. J. A. Hempel, G. R. Morris, and N. S.Trudinger, On the sharpness of a limiting case of
the Sobolev imbedding theorem, Bull. Austral. Math. SOC.3 (1970),369-373.
31. M.Hestenes, Extension of the range of a differentiable function, Duke J. Math. 8 (1941)
183-192.
32. E. Hewitt and K. Stromberg, “Real and Abstract Analysis,” Springer-Verlag, New
York, 1969.
33. V. I. Kondrachov, Certain properties of functions in the spaces Lp, Dokl. Akad. Nauk
SSSR 48 (1945),535-538.
34. M. A. Kranosel’skii and Ya. B. Rutickii, “Convex Functions and Orlicz Spaces,”
Noordhoff, Groningen, The Netherlands, 1961.
35. L. Lichenstein, Eine elementare Bemerkung zur reelen Analysis, Math. Z. 30 (1929),
794-795.
36. J. L. Lions, Une construction d’espaces d’interpolation, C.R. Acad. Sci. Puris 83 (1960),
1853-1 855.
37. J. L. Lions, Sur lesespaces d’interpolation; dualite, Math. Scand. 9 (1961),147-1 57.
38. J. L. Lions, Theorimes de trace et d’interpolation (IV), Math. Ann. 151 (1963),42-56.
39. J. L. Lions, “Probl6mes aux Limites dans les equations aux DerivCes Partielles,”
(Seminar notes), Univ. of Montreal Press, 1965.
262 REFERENCES
40. J. L. Lions and E. Magenes. Problemi ai limiti non omogenei (111). Ann. Scuola Norm.
Sup. Pisa 15 (1961), 41-103.
41. P. I. Lizorkin, Boundary properties of functions from “weight” classes, Dokl. Akad.
Nauk SSSR 132 (1960), 514-517 English trans.: Sot’ier Math. Doki1(1960), 589-593.
42. W . Luxemburg, “Banach function spaces,” (Thesis), Technische Hogeschool te Delft,
The Netherlands, 1955.
43. K. Maurin, Abbildungen vom Hilbert-Schmidtschen Typus und ihre Anwendungen.
Math. &and. 9 (1961). 359-371.
44. V. G. Maz’ja, Classes of domains and imbedding theorems for function spaces, Dokl.
Akad. Nauk SSSR 133 (1960), 527-530 English trans].: Soviet Math. Dokl. 1 (1960),
882-885.
45. V. G. Maz’ja,p-conductance and theorems of imbedding certain function spaces into the
space C, Dokl. Akad. Nauk SSSR 140 (1961), 299-302 [English transl.: Soviet Marh.
Dokl. 2 (1961), 1200-1203].
46. N. Meyers and J. Serrin, H = W ,Proc. Nar. Acad. Sci. USA 51 (1964), 1055-1056.
47. C. B. Morrey, Functions of several variables and absolute continuity, 11, Duke J. Math. 6
(1940), 187-215.
48. M. E. Munroe, “Measwe and Integration,” 2nd ed. Addison-Wesley, Reading, Mass-
achusetts 1971.
49. S. M. Nikol’skii, Inequalities for entire functions of exponential type and their application
to the theory of differentiablefunctions of several variables, Trudy Mar. Inst. Steklou 38
(1951), 244-278 [English transl.: Amer. Math. SOC.Transl. (2). 80 (1969), 1-38].
50. S. M. Nikol’skii, Properties of certain classes of functions of several variables on differ-
entiable manifolds, Mar. Sb. 33 (75) (1953). 261-326 [English trans].: Amer. Math. SOC.
Transl. (2). 80 (1969). 39-1 181.
51. S. M. Nikol’skii, Extension of functions of several variables preserving differential
properties, Mar. Sb. 40 (82) (1956), 243-268 [English trans].: Amer. Marh. Soc. Transl.
(2). 83 (1969). 159-1883.
52. S . M. Nikol’skii, On imbedding, continuation and approximation theorems for differ-
entiable functions of several variables, Russian Math. Surveys 16 (1961) 55-104.
53. L. Nirenberg, Remarks on strongly elliptic partial differential equations, Comm. Pure
Appl. Marh. 8 (1955). 649-675.
54. L. Nirenberg, An extended interpolation inequality, Ann. Scuola Norm. Sup. Pisa 20
(1966) 733-737.
55. R. 0”eill. Fractional intergration in Orlicz spaces, Trans. Amer. Math. SOC.115 (1965),
300-328.
56. J. Peetre, Espaces d’interpolation et thCor6rne de Soboleff,Ann Inst. Fourier (Grenoble 16
(1966), 279-317.
57. F. Rellich. Ein Satz tiber mittlere Konvergenz, Gortingen Nachr. (1930), 30-35.
58. W. Rudin, “Real and Complex Analysis,” McGraw-Hill, New York, 1966.
59. W. Rudin, “Functional Analysis,” McGraw-Hill, New York, 1973.
60. L. Schwartz, “ThCorie des Distributions,” Hermann, Paris, 1966.
61, R. T.Seeley, Extension of Cm-functionsdefined in a half-space, Proc. Amer. Marh. SOC.
15 (1964), 625-626.
62. S. L. Sobolev, On a theorem of functional analysis, Mar. Sb. 46 (1938), 471-496.
63. S.L. Sobolev,“Applicationsof Functional Analysisin MathematicalPhysics,” Leningrad
1950 Dnglish transl.: Amer. Mafh. Soc., Transl..,Math Mono. 7 (1963).]
64. S. L. Sobolev and S. M. Nikol’skii, Imbedding theorems, Izdat. Akad. Nauk SSSR,
Leningrad (1%3), 227-242 [English trans].: Amer. Math. SOC.Transl. (2). 87 (1970).
147-1 73.
REFERENCES 263
64a. E. M. Stein, “Singular Integrals and Differentiability Properties of Functions,”

(Princeton Math. Series, Vol. 30), Chapt. V, Princeton Univ. Press, Princeton, New
Jersey, 1970.
65. E. M. Stein and G. Weiss, “Introduction to Fourier Analysis on Euclidean Spaces”
(Princeton Math. Series, Vol. 32), Princeton Univ. Press,Princeton, New Jersey, 1972.
66. N. S. Trundinger, O n imbeddings into Orlicz spaces and some applications, J . Math.
Mech. 17 (1967), 473-483.
67. S. V. Uspenskii, An imbedding theorem for S.L. Sobolev’sclasses W/of fractional order,
Dokl. Akad. Nauk SSSR, 130 (1960) 992-993 [English transl.: Soviet Math. Dokl. 1
(1960), 132-1331. ’
68. S . V. Uspenskii, Properties of the classes W; with fractional derivatives on differentiable

manifolds, Dokl. Akad. Nauk SSSR 132 (1960), 40-62 [English trans.: Soviet Math.
, Dokl. 1 (1960), 495-4973.
69. K.Yosida, “Functional Analysis,” Springer-Verlag, Berlin, 1965.
70. S. Zaidman, “gquations Diffkrentielles Abstraites,” (Seminar notes), Univ. of Montrkal
Press, 1966.
Index
A Capacity, 152
Almost everywhere, 14 Cartesian product, 8
Approximation theorem, 14 Cauchy sequence, 4
for Lebesgue spaces, 28,31 Characteristic function, 14
for Orlicz spaces, 240 Clark, C. W.,175
for Orlicz-Sobolev spaces, 247 Clarkson’s inequalities, 37
for Sobolev spaces, 52,54,57,205, ,207 Cm-regularity,67
Aronszajn, N., 219 Compact imbedding, 9
Ascoli-Arzela theorem, 10 Compact imbedding theorem
for continuous function spaces, 11
B for Orlicz-Sobolev spaces, 249, 252, 256,
Banach algebra, 1 15 259
Banach space, 4 forSobolevspaces, 144,151,157,162,163,
Besov, 0. V., 218,223 166,167,170
Besov space, 177,223 Compact set, 6
Bessel potential, 219, 220 Compact support, 2
Bochner integral, 178 Complementary N-function, 229
Boundary, 2 Complete orthonormal system, 173
Bounded operator, 9 Completeness, 4
Browder, F. E., 70,79 of Lebesgue spaces, 26
of Orlicz spaces, 234
C of Orlicz-Sobolev spaces, 246
Calderbn, A. P., 83, 89,91,219, 221 of Sobolev spaces, 45
Calderbn extension theorem, 91 Completion, 4
Calderbn-Zygmund inequality, 9 1 Complex interpolation method, 177,221
265
266 INDEX
Cone, 65.96 Essentially bounded, 24

standard, 124 Extension operator, 83
Cone property, 66 Extension theorem, 84,88,91,211
uniform, 66
Conjugate exponent, 23 F
Conjugate N-function, 248 Fatou’s lemma, 17
Continuous functional, 3 Finite cone, 65
Convergence Finite width, 158
in the sense of 9 (n),19 Flow, 169
norm, 4 Fourier transform, 220
weak, 6 Fractional order spaces, 177
Convexity Besov spaces, 223
local, 3 Nikol’skii spaces, 225
uniform, 7,34,46 Sobolev spaces, 204,206
Convolution, 29,90, 199 spaces of Bessel potentials, 219.221
Countably additive, 13 Fringe, 161, 165
Cusp, 122,123 Fubini’s theorem, 18
exponential, 122 Functional, 3
standard, 124 G
Cylindrical coordinates, 124
Gagliardo, E.. 67.68, 70,75,95,99, 101
D
A-regular, 232 H
A2tondition, 23 1,232 Hahn-Banach extension theorem, 6
Dense, 4 Hestenes, M.,83
Derivative Hilbert space, 5
of a distribution, 20 Hilbert-Schmidt imbedding, 175, 176
distributional, 21, 179 Hilbert-Schmidt operator, 174
weak, 21 Halder condition, 10
Distribution, 19 Holder’s inequality, 23,234
tempered, 220 reverse form, 24
Distributional derivative, 21, 179 Homogeneous function, 91
Domain, 1 I
Dominated convergence theorem, 17
Dominating N-function, 231 Imbedding, 9,96
Donaldson, T. K., 247 compact, 9, see also Compact imbedding
Dual space, 3 theorem
of a Lebesgue space, 40,41 Hilbert-Schmidt, 175,176
of an Orlicz space, 238 trace, 96
of an Orlicz-Sobolev space, 247 Imbedding constant, 97
of a Sobolev space, 48-5 1 Imbedding inequality, weighted norms, 128,
133, 136, 138
E Imbedding theorem
Ehrling, G., 70, 75, 79 for Besov and Nikol’skii spaces, 224-226
Embedding, see Imbedding for continuous function spaces, 11
met, 7 for cusp domains, 126,127
Equivalent N-function, 231 direct and converse, 217-219
Equivalent norm, 4, 8 for Lebesgue spaces, 25
for Sobolev spaces, 79, 158,214,223 for Orlicz spaces, 234,237
Essential rate of increase, 231 for Orlicz-Sobolev spaces, 249, 252, 255,
Essential supremum, 24 256,259
INDEX 267
for Sobolev space Wm*p,97,242 M

for Sobolev space WsSp, 216219
Infinitesimal generator, 181 Magenes, E., 178,219
Inner product, 5 Maurin, K., 174
Integrability Measurable function, 14, 17, 178
Bochner, 179 Measurable set, 13
Measure, 13
Lebesgue, 15, 17
local, 19, 179 complex, 13
Integral Lebesgue, 14
positive, 13
Bochner, 178, 179
Lebesgue, 15, 17 Meyers, N., 45,52
Minkowski's inequality, 23
Interpolation inequality, 70, 74, 75, 78, 79,
81,82
reverse form, 24-25
Interpolation method Mollilier, 29
complex, 177,221 Monotone convergence theorem, 16
real, 177 Morrey, C.B., 95,98
trace, 178,186 Multi-index, 1
Interpolation space, 188
Interpolation theorem N
between Banach spaces, 188,222 N-function, 227, 228
involving compact subdomains, 81,82 Nikol'skii, S. M.,177, 225,226
between spaces L..P, 222 Nirenberg, L., 70,75
between spaces W m *,*74,75,78,79 Norm, 3
Isometric isomorphism, 4 equivalence, 4,8,79, 158,214,223
Normable space, 4
K Normed dual, 5
Kernel, 204,214-218 Normed space, 4
Kondrachov, V.I., 143
Rellich-Kondrachov theorem, 144 0
Operator, 8
L bounded, 9
I-fat cube, 161 compact, 8
Lebesgue dominated convergence theorem, completely continuous, 9
17
extension, 83
imbedding, 9
Lebesgue integral, 15, 17
Lebesgue measure, 13, 14 Orlicz class, 232
Lebesgue spaces, 22,24 Orlicz space, 227, 233,236
Leibniz formula, 2,21 Orlicz-Sobolev space, 246-247
Lichenstein, L., 83 Orthonormal system, 173
Linear functional, 3
Lions, J. L., 56, 178,184,219 P
Lipschitz property, 66 Parallelepiped, 65
Lizorkin, P. I., 218 Parallelogram law, 5
Local integrability, 19, 179 Partition of unity, 51
Localization theorem, 207 Polar set, 56
Locally convex TVS,3 Precompact set, 7, 31,33, 242
Locally finite, 65 Product
Lusin's theorem, 14 Cartesian, 8
Luxemburg norm, 234,253 inner, 5
INDEX
Q Spherical coordinates, 130

Quasibounded domain, 148 -
Spiny urchin, 15 1 152
Quasicylindrical domain, 159 Stone-Weierstrass theorem, 10
Streamline, 169
R Strong extension operator, 83,84
Radon-Nikodym theorem, 17 Strong local Lipschitz property, 66
Reflexivity, 6 Subspace, 5
of Lebesgue spaces, 42 support, 2
of Orlicz spaces, 240
T
of OrliczSobolev spaces, 247
of Sobolev spaces, 47,205 Tempered distribution, 220
of trace interpolation spaces, 189-190 Tesselation, 106, 161
Regularization, 29 Testing function] 19
Rellich, F.. 143 Topological vector space (TVS), 2
Rellich-Kondrachov theorem, 144 Total extension operator, 83,88
Riesz representation theorem Trace, 96,185, 186
for Hilbert space, 5 on the boundary, 113,114.215-217
for Lebesgue spaces, 40,41 higher-order, 196, 198
interpolation method, 178, 186
S interpolation space, 186, 198
Scale of spaces, 188 on a subspace, 96,112
Schwartz distribution, 18, 19 Transformation of coordinates, 63-64
Second dual, 6 Trudinger, N. S.,228,242,247
Seeley, R., 83, 86
U
Segment property, 54,66
Semigroup of operators, 180 Uniform Cm-regularity,67
Separability, 4 Uniform cone property, 66
of Lebesgue spaces, 28 Uniform convexity, 7
of Orlin spaces, 240 of Lebesgue spaces, 34,38
of Orlicz-Sobolev spaces, 247 of Sobolev spaces, 47
of Sobolev spaces, 47 Uspenskii, S. V.,218
Serrin, J., 45, 52
0-algebra, 13 V
Simp$ extension operator, 83,91 Vector sum of Banach spaces, 184
Simple function, 14, 178 Volume, 14
Smith, K. T., 219
m-Smooth transformation, 63 W
Sobolev, S. L., 45,95 Weak convergence, 6
Sobolev conjugate N-function, 248 Weak derivative, 21
Sobolev imbedding theorem, 95,97 Weak sequential compactness, 7
limiting case, 242 Weak-star topology, 3, 19
Sobolev inequality, 104 Weak topology, 6
Sobolev space, 44
Y A 5
equivalent definitions, 52 B 6
fractional order, 177,204,205,206 Young’s function, 227, 253 c 7
integral order, 44 Young’s inequality, 230 D ~ I
E 9
negative order, 50,5 1 Young’s theorem, 90 F O
G 1
H ?
1 3
J 4

(Pure and Applied Mathematics 65) Robert A. Adams-Sobolev Spaces-Academic Press, Elsevier (1975)

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

(Pure and Applied Mathematics 65) Robert A. Adams-Sobolev Spaces-Academic Press, Elsevier (1975)

Diunggah oleh

Hak Cipta:

Format Tersedia

SOBOLEV SPACES

A CA DEM I C PR ESS New York San Francisco London 1975

ACADEMIC PRESS, INC.

United Kingdom Edition published by

Library of Congress Cataloging in Publication Data

(Pure and applied mathematics series; v. 65)

AMS (MOS)1970 Subject Classifications: 46E30,46E35

PRINTED IN THE UNITED STATES OF AMERICA

This monograph is devoted to a study of properties of certain Banach

We acknowledge with deep gratitude the considerable assistance we have

If a and fl are two multi-indices, we say P I c1 provided /3, I aj for

The reader may wish to verify the Leibniz formula

1.2 If G c R", we denote by G the closure of G in R". We shall write G c c R

{ x E R" :x 4 G} is the complement of G.

Topological Vector Spaces

1.4 A topological vector space, hereafter abbreviated TVS, is a Hausdorff

from the topological product spaces X x X and @ x X , respectively, into X

1.5 By a functional on a vector space X we mean a scalar-valued function f

1.6 A norm on a vector space X is a real-valued functional f on X which

1.7 A sequence {x,} in a normed space X is convergent to the limit xo if and

1.9 If X is a vector space, a functional (., . ) x defined on X x X is called an

inner product on X provided that for every x, y, z E X and a, b E 62

The Normed Dual

1.10 A norm on the dual X‘ of a normed space X can be defined by setting,

Since @ is complete, X ’ , with the topology induced by this norm, is a Banach

1.11 THEOREM (Riesz representation theorem) Let X be a Hilbert

and in this case 1 1 ~ ’ ; X’I(= IIx; XII. Moreover, x is uniquely determined by

A vector subspace M of a normed space Xis itself a normed space under

1.12 THEOREM (Hahn-Banach extension theorem) Let Mbe a subspace

1.14 THEOREM Let X be a normed space. X is reflexive if and only if

Weak Topology and Weak Convergence

1.15 The weak topology of a normed space X is the weakest topology on X

1.16 A subset A of a normed space X i s called compact if every sequence of

1.18 THEOREM A set A is precompact in a Banach space X if and only

A set Ne with this property is called afinife &-netfor A.

1.20 THEOREM A uniformly convex Banach space is reflexive.

The following two theorems will be used to establish the separability,

1.21 THEOREM Let X be a Banach space and M a subspace of X closed

The completeness, separability, and uniform convexity of M follow easily

Operators and Imbeddings

Spaces of Continuous Functions

1.26 Since R is open, functions in C"(R) need not be bounded on R. If

1.27 If 0 < A S 1, we define CmiA(n) to be the subspace of Cm@)consisting

It should be noted that for 0 < v < L s 1,

1.28 THEOREM (Stone-Weierstruss theorem) Let R be a bounded

The following is a straightforward imbedding theorem for the spaces

1.31 THEOREM Let m be a nonnegative integer and let 0 < v < 15 1.

from which we conclude that

I const I Da+ (x) -Da4(JJ)I -v'a (9)

The Lebesgue Measure in R"

1.32 Many of the spaces of functions considered in this monograph consist

1.33 A collection C of subsets of R" is called a a-algebra if the following

1.34 By a measure p on C we mean a function on Z taking values in either

whenever A j E Z, j = 1,2, ..., and A j n A, = 0 for j # k. (For a complex

1.35 THEOREM There exists a a-algebra C of subsets of R" and a positive

The elements of C are called (Lebesgue) measurable subsets of R"; p is

1.36 If B c A c R" and p ( B ) = 0, then any condition that holds at every

1.37 THEOREM (a) Iffis measurable, so is I f l .

is called the characteristic function of A. A real-valued function s on R" is

1.39 THEOREM Given a real-valued function f with domain A c [w"

The Lebesgue Integral

1.40 We are now in a position to define the (Lebesgue) integral of a measur-

Iff is measurable and nonnegative valued, we define