Documente Academic
Documente Profesional
Documente Cultură
ROBERT A . ADAMS
Department of Mathematics
The University of British Columbia
Vancouver, British Columbia, Canada
Adams, Robert A
Sobolev spaces.
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
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
valid for functions u and u that are la1 times continuously differentiable near x .
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.
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
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].
Normed Spaces
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.
1.8 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 .
X’(Y) = 0 9 4 x 7
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.
Compact Sets
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.
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.
1.22 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).
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.
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.
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.
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
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 :
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].
p ( U ? = , A j ) = limj-,mp(Aj).
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.
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
(c) Iff(x) _< g ( x ) for all x E A, and if both integrals exist, then
16 1 INTRODUCTORY TOPICS
(4)(x) dx = c
I f(~)
dx.
A(A) =
s1f ( ~ dx
)
lim
n-+m
I f , ( x ) dx = /A
(lim f . ( x ) ) dx.
n+w
THE LEBESGUE INTEGRAL 17
/A
f ( x ) dx = I.(.) + dx i L v ( x ) dx.
A ( A )=
I f ( x ) dx.
Lf(X,Y)dx dY
or, if it is desired to have the integral extend over all of Rn+m,
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
1.52 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).
lL(4n)-T,(+)/ I supI$n(X)-4(x)I
XEK sh
The right side of the inequality above tends to zero as n
lu(x)l dx.
-+ co since +,, + +
uniformly on K .
every 4 E 9 ( Q )
for every
J)(x)D"$(x)
4 E 9(R).
dx = (- 1)1"1
I*
L),(x)@(x)dx
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
+
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
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
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
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
Completeness of LP(R)
PROOF First assume 1 I p < co and let {u,} be a Cauchy sequence in LP(Q).
COMPLETENESS OF Lp(a) 27
5
JI~,IJ~5 j = 1 (1/2j) < I , m = 1,2,....
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
-
lun(x)l 5 ll~nllm~ lun(x>--u,(x)I II~n-umllm*
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.
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.
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
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.
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.
n n
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.
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
32 11 THE SPACES LP(n)
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
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
~ 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
,.I lu(x)lpdx
.. .
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.
-=
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.
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
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
If 1 < p s 2, then
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
PROOF Let u,u E Lp(R) satisfy lIu(lp= llullp = I and I!u-ullp 2 E > 0. If
2 5 p < co, we have from (37)
1 p q ; s l-2.p.&P
In either case there exists 6 = a(&) > 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).
~ ( X ) U ( X dx,
) u E LP(R).
THE NORMED DUAL OF LP(n) 39
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)
where
1 + PIIUII; = (w+tu)+(w-tu)
(46)
(47)
Equations (46) and (47) are not possible for all t > 0 unless llullp = 0 which is
impossible. Thus L , = L,.
u E LP(Q)
L(u) =
Sn u(x)u(x)dx.
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
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
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
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
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)
for any function u for which the right side makes sense, 11 I l p being, of course,
3
44
DEFINITIONS AND BASIC PROPERTIES 45
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($) +Tu(b,) for every b, E Q((Sz) as n + co. Similarly,
for every b, E 9 (a).It follows that
TD.,,(b,)+ Tue(b,)
= lim T',,,,(b,) = lim (- l)lalT,n(Dab,)
Tu=(b,) = (- l ) W u ( ~ a b , )
n-rm n-r m
By Theorems 1.22, 2.10, 2.17, 2.25, and 2.31, LNp s a Banach space that is
separable if 1 I p c co and reflexive and uniformly convex if 1 < p c co.
Let us suppose that the N multi-indices u satisfying 0 I1011
I m are linearly
ordered in some convenient fashion so that to each U E W"sP(R) we may
associate the well-defined vector Pu in LNPgiven by
PU = ( D a u ) O < l a l < m * (3)
Since IIPu; LNp)I= I I U ~ ~ , , , ~ , P is an isometric isomorphism of W m . p ( nonto
) a
subspace W c LNp.Since Wm*p(i2) is complete, W is a closed subspace of LNP.
By Theorem 1.21, W is separable if 1 Ip c co and reflexive and uniformly
convex if 1 < p c co. The same conclusions must therefore hold for
W"*P(Q)= P-'(W).
DUALITY, THE SPACES W-””‘(a) 47
3.6 In the following sections we shall take, for fixed Q, m,andp, the number
N, the spaces L N pand W, and the operator P to be as specified in Section 3.4.
We also define
<u,u> = j-nu(x)u(x)dx
for any functions u, u for which the right side makes sense. For given p, p‘
shall always designate the conjugate exponent :
if p = l
if 1 < p c co
if p = 00.
L(u) = 2 (Uj,Uj).
N
j= 1
PROOF If 1 < j < N and W E LP(a), let w ( ~=) (0,..., 0,w,O, ...,0) be that
element of LNp whosejth component is w, all other components being zero.
Setting Lj(w) = L ( W ( ~ )we
) , see that Lj E (LP(Q))’ and so by Theorems 2.33
and 2.34 there exists (unique) uj E LP’(Q)such that for every w E LP(Q)
L(w,j,) = L j ( W ) = {W,Uj).
If u E LNp,then
48 111 THE SPACES WrnSp(R)
so that IIL;(LNP)'IIs IIu; L$ 11. We show that these norms are in fact equal as
follows: if 1 < p < 00 and 1 s j s N, let
It is easily checked that IL(u,, ,. .,uN)I= IIo; L$ I p ' = Ilu; LNpIIIIu; .L$II. If p = 1,
we choose k so that Iluklla, = maxl,jsNIIuj~~m.For any E > 0 there is a measur-
able subset A c R having finite, nonzero volume, such that Iuk(x)I 2 IIukIla, - E
{r
for x E A. Set
uk(x)/uk(x) if x E A and uk(x) # 0
u(x) =
otherwise.
Then
L(u(k)) = (%uk) =
s* l'k(x)ldx 2 (\\Uk\\a,-E)lIu\ll
= (II~;LNDDII-~)II~~~);LNIII.
Since E is arbitrary, (4) must follow in this case also. I
3.8 THEOREM Let 1 5 p < co. For every L E ( Wrn*p(R))' there exists an
element u E L$ 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).
z(u) = 1
0 < ( a ( srn
<ua,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.
T(4) =
0s
c
Jalsrn
T,.(Da4) = L ( 4 )
5 c
O i l a l srn
II'"(4k-4n)IIp IIuallp'
~ ~ ~ k - ~ n ~ ~ 0,
m p as
~ ~ k,n
~ ;--t~ 00-
$ ~ ~
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 < p < co.
IIUllm, P 41
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
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)
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.
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
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.
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.
-
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
k
II+-uIIm,p.R 5 C I I ~ j - ~ j I I m , p , n< E *
j=o
-
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)
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.
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).
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").
3.24 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
L L
= u,(x,)***uj-,(xj-,)
=0, 2 ~ j ~ n - 1 ,
APPROXIMATION BY FUNCTIONS IN Com(R); (I)I,P')-POLAR SETS 59
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
Hence
~ u ( x ) 9 ( x ) d x= TU(9) = k
Hence u ( x ) - k = 0 a.e. in B. I
SB 9(x,dx.
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
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. [
3.32 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
Transformation of Coordinates
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.
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)
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
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:
,
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].
68 Iv INTERPOLATION AND EXTENSION THEOREMS
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
t
@
P, aiid P,
f'
FIG.2
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
~ 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
+
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.
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
Hence,
+ Kp(b-a)-pllf(r)lpdr.
b
lblf’(r)lpdrI K p ( b - a ) P b If”(t)lPdt (2)
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
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)
m ’I
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)
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
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,.
FIG.3
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.
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
( ( 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.
4.19 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
If(x)l I 3 I;rlf(v)l + l1
0 If’(Oldt,
equation continues
INTERPOLATION INEQUALITIES INVOLVING COMPACT SUBDOMAINS 81
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
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
lu(x)lpds K1eP +
L n Q o I D , u ( ~ ) ( ~ d s K1 LnQql~(x)lpds,
LnQo
82 Iv INTERPOLATION AND EXTENSION THEOREMS
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.
Also by (21) withj and m replaced byj- 1 andj, respectively (the case already
proved), we have
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
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.
EXTENSION THEOREMS 85
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
&<x> = 4 O W +j = 1 cN
ej<x>
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”
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.
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)”
EXTENSION THEOREMS 87
and
PROOF For fixed N , let a k , N , k = 0,1,2, . .., N, be the solution of the system
of linear equations
N
xo x 1 *-. XN
v(xO,x1, ..., X N ) = Xo2 XI2 ... XNZ
V(1,2,..., 2 k - 1 , - 1 , 2 k + l , ...)2 N )
=
ak,N
V(1,2,...,2")
= AkBk,N
where
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
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)
EXTENSION THEOREMS 89
= KllUk12km(1/2k)
{ 1pIsm L + n
5 Kllak12k" Il4llrn,p,W+"*
u * u(x) =
s.. u ( x - y ) u ( y ) dy,
E
=
s..
u ( x - y ) u ( y ) dy
= 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
s.. 1(x) w ( x ) dx =
s.. u * u(x) w ( x ) dx
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
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).
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,
92 Iv INTERPOLATION AND EXTENSION THEOREMS
-
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 [$j(pa)u(y-pg) dp do (35)
where Js. da denotes integration over the unit sphere, and the constant Kj
will be determined shortly. If y E Uj n 0, then, assuming for the moment that
u E Cm(0),we have by condition (iii) of Section 4.4 that u ( y - p a ) is infinitely
differentiable for pa E supp 4 j . Now integration by parts rn times yields
Hence
EXTENSION THEOREMS 93
= C 5, * D'u,
-
where & = (- l)lal(m!/a!)aaC$j
1.1 =m
-
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.
94 Iv INTERPOLATION AND EXTENSION THEOREMS
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)
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
PART I If R has the cone property, then there exist the following imbeddings :
PART II If R has the strong local Lipschitz property, then Case C of Part I
can be refined as follows :
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.
I s .+ '
Iu(x', ()IP = u(x', 0 ) D,,u(x', t ) dt
100 v IMBEDDINGS OF w""(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
Hence
+ 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
PROOF OF THE IMBEDDING THEOREM 101
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.
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)
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
PROOF OF THE IMBEDDING THEOREM 103
61 lu(x + (1 -t)e,)lYdt
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.
PROOF OF THE IMBEDDING THEOREM 105
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.
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
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
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)
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
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 OF THE IMBEDDING THEOREM 109
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
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
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.
PROOF OF THE IMBEDDING THEOREM 111
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.
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
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)
1bdry n
~ ( xdo
) = c1 j bdryn
f ( x ) uj ( x ) do,
I I E ~ l l m , p , ~5
n K1IIullm,p,n*
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
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
Similarly, if la - PI I k, then
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
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
D"u ( x ) = y
j= 1
Pa, ( x ) p - ("I [log(4R/p)] -j , (43)
so that
D"u(x) = y Pa,j(x)pl-2~"~[log(4R/p)]-j,
j=l
120 v IMBEDDINGS OF W*"(R)
ID"u(x)lP I f K,,jpP"-'~')[log(4R/p)]-jP.
j=l
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.
p ( p ) d p = 1.
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,
COUNTEREXAMPLES AND NONIMBEDDING THEOREMS 121
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
~ *<
(ii) / m l u ( j ) ( r ) l r ~ ( r , dr ) co if o < j 5 m,
1
where the coefficients cli depend only on 1. Now u ( x ) vanishes for 1x12 1
and so
IMBEDDING THEOREMS FOR DOMAINS WITH CUSPS 123
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
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)
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
IMBEDDING THEOREMS FOR DOMAINS WITH CUSPS 125
t y2
1”
(b)
xn = Yn,
which has Jacobian determinant
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
IMBEDDING THEOREMS FOR DOMAINS WITH CUSPS 127
Then
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 ,... ,
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.
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
so that
Hence limf,o+tv~u(t)~=O.I
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
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.
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.
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
IMBEDDING INEQUALITIES INVOLVING WEIGHTED NORMS 131
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
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
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,
+
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 )
i=2
$, lulna/(n - 1)rkn v / ( n - 1) dx
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
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
~ R ~ ~ u ( x ) ~ p [ r k ( x<
) l vad x
Ix-z1-"dx I; K(volR)l-vlk,
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
IMBEDDING INEQUALITIES INVOLVING WEIGHTED NORMS 137
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.
so that
(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
Similarly,
PROOF The proof is the same as that given for inequality (28) in Lemma 5.17,
except that the inequality
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
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
sup
XEQk Iv(x)l + lgradv(y)lpl[rk(y)lvd~}l'p. (gl)
5 K(V){ik[lv(Y)p
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 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
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
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
(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
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)
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
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
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
j
no-nj
lu(x)ldx <
1no * nj
Iu"(X+h) - u " ( X ) l d X < E/2. (15)
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 .
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)
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
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
Note that this domain, though quasibounded, is simply connected and has
empty exterior.
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.
-
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.)
lU(X)lPdX
-
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)
-
i IalsrnxEHj
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
~ ~ 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
IluIl8,q,,H5 K 2 h 4 j H lgradu(x)lqdx
K2hq 1
la1 = 1
IIDuullZ,q,H
~ ~ 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
-
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
6.22 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).
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
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
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
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.
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,
Then there can be no imbedding of the form (38) for any q > p .
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
$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
UNBOUNDED DOMAINS-DECAY AT INFINITY 165
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)
- 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.
holds for every E > 0. For example, f ( x )= exp(-x2) satisfies (43) but
168 VI COMPACT IMBEDDINGS OF W""(a)
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).
6.45 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:
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
- 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
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
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)
= ~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
lim l d N ( t ) dt = 0.
N+m
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
Hilbert-Schmidt Imbeddings
[where (-, .)x is the inner product in X I , and such that for each x E X we have
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)
=
m m
1 1 I(%A*fi)XIZ
j = 1 i=1
= 2 IIA*fi;XIIZ
j= 1
= IlIA*lll.
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
Consequently,
Hence imbedding (SO) is Hilbert-Schmidt. The same proof holds for (51)
without regularity needed for the application of Theorem 5.4. I
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)
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
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:
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.
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
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.
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,
{
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
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
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.
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 ) .
sb
lim ( l / t ) G ( t ) b d t = b.
r+O+
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.
G ( z ) G ( t ) bdr - b = G ( t ) b - 6 .
+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
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
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
)
g’(t) =
sb G ( t - ~ ) f ’ ( rdt
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
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
+
B, Bz.
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’)
= ([f’(t) $ ( r ) dr - S m f ’ ( r ) $ ( r ) dr, b‘
0 ) = 0.
+
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.
+
where 8 = ( l / p ) v. If J, is the mollifier of Section 2.17, then, just as in Lemma
THE TRACE SPACES OF LIONS 187
+
m
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
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
In particular, B , n B, is dense in T.
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,
190 VII FRACTIONAL ORDER SPACES
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.)
x (I(
n So" r ' ~ I ~ f ~ x , t ) ~ d t ) 8 1 1 1 1 p d x ) l i s ' ,
)(1 -em
Ilt"f';LP(O, CO;LP(n))lle
= K3llUllT
by Lemma 7.16(a). This establishes imbedding (1 1).
7.23 LEMMA Letfbe a scalar-valued function defined a.e. on [O, co) and
let
= (l/t) /k)
dt.
0
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)
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,
Hence
G(t-E)f(E) - f ( ~ )=
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
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
= ( l / t ) ( G ( t ) u - u ) - (l/t2)
6 (G(T)u-u)&,
SEMIGROUP CHARACTERIZATION OF TRACE SPACES 195
and since
I Cl/(2-e)p1 IIuIIpTo
we therefore have
c
n
Ilu;D(A)II = ~ ~ u ~ ~ B + ~ ~=A IIuIIB+
u ~ ~ B n
j=1
IIAjullB*
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
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,
Higher-Order Traces
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
-=
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
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
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
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 ) .
HIGHER-ORDER TRACES 199
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$j'(O) = 0, O Ij I m - 1, j # k.
Moreover, it is easily checked that
IIgkIIW"' 5 RkII&llWm.
Fl * F,(t)b =
l F1(t-7)F2(z)bdz, b E B.
1 wr(t),
i
Ak”’(t) =
r=O
IK5tm-i-j-2 [ l I G ( z ) w - dz.
w(JB
llA~(j)(t)llBI &t-’
sb IIG(z)w-~llt,d~.
Let h ( t ) = Aig(?),Thus
h(r) = t - m I ( ( k + 1)) * G((m))(t)G.
In order to prove (23) in this case it is sufficient to show that
Again we may discard the last term which makes no contribution to (27). It is
therefore sufficient to establish (27) with w Ireplaced by
w,*(t) = t - m I ( ( j + l + S ) ) * G ( ( m - i - S - 1 ) ) * (G- I )( t )W.
However, w { i ) ( t )is a linear combination of terms of the form
-m-rI((I+s+r)) * G((m-I-s- 1)) * (G-Z)(t)w
for 0 Ir Ij . It follows just as in Case 1 that for 0 < t I1,
Ilw{$(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.
204 VII FRACTIONAL ORDER SPACES
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
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
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.
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).
206 VII FRACTIONAL ORDER SPACES
and since
Hence
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.
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").
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
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
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
210 VII FRACTIONAL ORDER SPACES
) 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
Similar inequalities hold for differences in the other variables x2, ..., x,, and
combining these we obtain llullT I K8 IIullr. I
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
212 VII FRACTIONAL ORDER SPACES
where
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
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.
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
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).
Now
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
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 < p < n .
(a) If n > 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).
218 VII FRACTIONAL ORDER SPACES
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
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.
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
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
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.
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
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").
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
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
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),
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 } .
7.74 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
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
1
5 s
/
a =47)
s=A(t)
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
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
8.5 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 ' ( ? )
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.
U E KA(Q>
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
Let
Then
if volR = 00
But
m
j-nA(12u(X)I) 2 1 2 J A ( f j V0lQj
j=l ) = O3.
J * A ( Y ) d X I1,
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
-
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
jlim
+m /nA(luj(x)-u(x)l)cix = 0.
236 VIII ORLICZ AND ORLICZ-SOBOLEV SPACES
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
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)*
sn
8.17 LEMMA For fixed u E Ln(R) the linear functional L, defined by
lo if u ( x ) = 0.
If llullA > 1, then for sufficiently small E > 0 we have
238 VIII ORLICZ AND ORLICZ-SOBOLEV SPACES
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 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
L(u) =
L,(u) =
I* U(X)O(X) dx
8.19 THEOREM LA@)is reflexive if and only if both (A, 0) and (1R)
,
are A-regular.
* - 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.
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@).
For j , k 2 M we have
I-
vol R
2 vol R
+!I,
4 n,,k
) 1.
A(!y)dX+6B(fOI
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
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
244 VIII ORLICZ AND ORLICZ-SOBOLEV SPACES
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
) 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
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].
L(u) = c
Oslo1S m
~ n o % ( x ) u , ( x )dx
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
It is also readily seen for the case p = n that A,,(t) is equivalent near infinity
to the N-function e'- t - 1.
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
IMBEDDING THEOREMS FOR ORLICZ-SOBOLEV SPACES 249
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
is compact.
Thus
exist and are compact for any N-function B increasing essentially more
slowly than A near infinity.
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.
d7 c co
Then
IMBEDDING THEOREMS FOR ORLICZ-SOBOLEV SPACES 253
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
r$2
KlIlulll,A,R*
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
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.
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
is finite, is a Banach space under that norm. The theorem above asserts that if
(43) holds, then
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
260
REFERENCES 261
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
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